29 Technology & Invention
The Default Library
========================================


Artificial Intelligence: Where It Came From and Where It Stands
---------------------------------------------------------------
Seventy years of bold promises, two hard winters and one very sudden spring, and an honest look at where things stand in late 2026.

Follows: Alan Turing: The Machine That Defined Computing (History-Makers 4) and Human-AI Collaboration: Becoming the Architect, Not the Typist (Skills for the AI Era 5)
In May 1997 a chess computer beat the world champion, and many people decided the machines had arrived. They hadn't. Deep Blue could do one thing well. It couldn't tell you what chess was, or hold a conversation. Twenty-five years later, at the end of November 2022, a free chatbot did something stranger. It wasn't built to play chess at all, but it could write a sonnet, explain a tax form and fix a bug in the same afternoon. Within about two months, by one analyst estimate, 100 million people a month were using it.
The history of artificial intelligence has looked like this from the start: overconfident predictions, long stretches of slow work, and then sudden leaps that surprise the experts. This piece follows that history from clockwork dolls to today's large language models, and ends with a careful look at where things stand as of October 2026.
A note on who is writing. This library is written with the help of an AI model, Anthropic's Claude. That is a reason for extra care. In what follows, Anthropic is described like every other lab: no more praise, and no less mention, than the evidence supports.
In this deep dive

The dream before the machine: automata, Leibniz, Babbage, Lovelace, Turing and Dartmouth
Two roads: hand-written rules, two winters, and neural networks (1956–1997)
The deep-learning turn and the ChatGPT moment (2012–2022)
The state of affairs, late 2026: capabilities, compute, cost and who builds it
Rules, jobs and the argument about risk

Part 1 — The dream before the machine
Long before anyone could build a thinking machine, people built machines that looked as if they were thinking.
Around 1206, the engineer Ismail al-Jazari finished his Book of Knowledge of Ingenious Mechanical Devices, describing water clocks, pumps and automata: a drink-serving waitress, a peacock fountain with mechanical servants, and a boat of mechanical musicians. Historians still debate how "programmable" those musicians were. The roboticist Noel Sharkey has argued that pegs on a rotating drum could be moved to change the rhythm, which would make it a very early programmable machine.
Between 1768 and 1774, the Swiss watchmakers Pierre and Henri-Louis Jaquet-Droz and Jean-Frédéric Leschot built three dolls: a musician, a draughtsman and a writer. The writer can still dip a goose quill and write a custom text of up to 40 characters, chosen one by one on a coded wheel. It is a beautiful machine, but it doesn't think. Every stroke was decided in advance by a person.
Then there was the famous fake. In 1770 Wolfgang von Kempelen unveiled the Mechanical Turk, a chess-playing "automaton" that toured Europe and America for 84 years and beat, among others, Napoleon and Benjamin Franklin. A human chess master was hidden inside. The Turk is worth remembering because the same question keeps coming back: when a machine seems intelligent, what is actually doing the work?
Thinking as calculation
The deeper idea came from philosophers. In Leviathan (1651), chapter 5, Thomas Hobbes wrote that reason "is nothing but Reckoning (that is, Adding and Substracting)". If thinking is a kind of calculation, then a calculating machine might one day think.
Gottfried Wilhelm Leibniz pushed this furthest. In 1672 he conceived a calculating machine, the stepped reckoner, the first designed to do all four arithmetic operations. Its gearing was beyond the craftsmanship of the day, and it never worked reliably. He also dreamed of a universal symbolic language (the characteristica universalis) in which arguments could be settled by computation. In a 1706 letter to the Electress Sophia of Hanover he wrote that one could then always say, "let us calculate". The popular one-word Latin version, Calculemus!, is a later paraphrase of passages like this.
The engine that wove algebra
In the 1820s and 1830s Charles Babbage designed two mechanical computers. The Difference Engine was a giant calculator for printing mathematical tables. He built only a demonstration section of about 2,000 parts in 1832. The Analytical Engine, conceived in 1834, was far more ambitious: it had a store (memory), a "mill" (processor) and instructions fed in on punched cards borrowed from the Jacquard silk loom. It was never built.
In 1843 Ada Lovelace translated an Italian article on the engine and added long notes of her own. They include a step-by-step method for computing Bernoulli numbers, often called the first published computer program. How much of it was hers and how much Babbage's is genuinely debated by biographers. What is clearly hers is the vision:

"The Analytical Engine weaves algebraical patterns just as the Jacquard-loom weaves flowers and leaves." — Ada Lovelace, Notes on Menabrea's Sketch (1843)

She guessed that such a machine "might compose elaborate and scientific pieces of music" if music could be expressed in its notation, so it could work on symbols, not only numbers. She also set the first limit:

"The Analytical Engine has no pretensions whatever to originate anything. It can do whatever we know how to order it to perform." — Lovelace (1843)

A century later, Alan Turing would call this "Lady Lovelace's Objection" and argue with it. That argument still hasn't been settled.
Turing's question
The History-Makers piece on Turing covers his 1936 machine and his life. Here, what matters is the paper he published in October 1950 in the philosophy journal Mind, "Computing Machinery and Intelligence". Its first line: "I propose to consider the question, 'Can machines think?'"
Turing thought "think" was too vague to argue about, so he replaced it with a game. An interrogator types questions to two hidden players, a human and a machine, and tries to tell which is which. This is the imitation game, now called the Turing test. He made a cautious prediction: in about fifty years, computers would play it so well that "an average interrogator will not have more than 70 per cent chance of making the right identification after five minutes of questioning."
The paper is still worth reading because of its section of objections and replies. Turing anticipated the theological objection, the "heads in the sand" objection, Gödel's limits on formal systems, the argument from consciousness and Lovelace's objection. He also proposed what now looks like the winning strategy. Instead of programming an adult mind, he asked, "why not rather try to produce one which simulates the child's?" and then educate it. That is close to what machine learning does today.
The summer that named the field
On 31 August 1955, four researchers (John McCarthy of Dartmouth, Marvin Minsky of Harvard, Nathaniel Rochester of IBM and Claude Shannon of Bell Labs) wrote a funding proposal. It asked for "a 2 month, 10 man study of artificial intelligence" in the summer of 1956 at Dartmouth College, and it named the field. Its central conjecture was bold:

"Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." — the Dartmouth proposal (1955)

They hoped that a carefully chosen group could make "a significant advance" on language, abstraction and self-improvement in a single summer. The workshop didn't achieve that. It did gather the people who would lead the field for decades.
The early results were thrilling. Allen Newell, Herbert Simon and Cliff Shaw's Logic Theorist (begun in 1955) eventually proved 38 theorems from Whitehead and Russell's Principia Mathematica, treating reasoning as a search through possibilities. Programs played checkers, solved algebra word problems and stacked blocks in simulated worlds. In 1966 Joseph Weizenbaum's ELIZA imitated a psychotherapist by reflecting users' words back as questions, and some users became convinced they were talking to someone who understood them. Weizenbaum was shocked that people took it so seriously, and his 1976 book Computer Power and Human Reason became an influential critique of AI by one of its own pioneers.
The predictions grew with the results:

Year
Who
Prediction

1958
Simon and Newell
A computer would be world chess champion "within ten years"

1965
Herbert Simon
"Machines will be capable, within twenty years, of doing any work a man can do"

1967
Marvin Minsky
Within a generation, creating "artificial intelligence" would be "substantially solved"

The chess prediction came true after 39 years, not ten. The other two did not come true on anything like their schedule. The pattern of brilliant people badly underestimating how hard the world is to describe is the most important lesson of AI's first decades.

Part 2 — Two roads: rules, winters and neural networks
Early AI was mostly symbolic: intelligence was treated as manipulating symbols according to explicit rules (if the patient has a fever and a stiff neck, consider meningitis). This works beautifully in small, tidy worlds. Outside them it runs into what James Lighthill called the combinatorial explosion: the number of possibilities grows so fast that no amount of rule-writing or computing keeps up.
The first winter
Two reports broke the spell. In 1966 the US ALPAC report concluded that machine translation was slower, less accurate and more expensive than human translators, and US government funding for it largely ended. In 1973 the British Science Research Council published Sir James Lighthill's review, which said that "in no part of the field have the discoveries made so far produced the major impact that was then promised." It became the basis for ending most UK university AI funding. On 9 May 1973 Lighthill defended his report in a public debate at the Royal Institution in London against Donald Michie, John McCarthy and Richard Gregory.
In the US, the 1969 Mansfield Amendment pushed military research funding towards projects with direct applications, and by the mid-1970s money for open-ended AI research was hard to find. The period became known as the first AI winter.
The expert-systems boom
The thaw came from narrowing the ambition. Instead of general intelligence, expert systems captured the rules of one specialist field:

DENDRAL (begun 1965, Stanford) inferred chemical structures from mass-spectrometer data.
MYCIN (early 1970s, Stanford) recommended antibiotics for blood infections.
R1/XCON (1980, Carnegie Mellon for Digital Equipment Corporation) configured computer orders. Estimates of its savings vary, from about $40 million over six years to about $40 million a year by 1986.

Companies built dedicated "Lisp machines" to run such software. Japan's Ministry of International Trade and Industry set aside about $850 million for its Fifth Generation computer project in 1981. In 1983 the US (DARPA's Strategic Computing Initiative) and the UK (the £350 million Alvey programme) answered with programmes of their own.
The second winter
Then the costs came due. Expert systems were brittle. Outside their narrow rules they failed badly, and keeping thousands of hand-written rules up to date was expensive. Getting experts to state what they "just know" proved slow and frustrating. In 1987 cheaper desktop workstations caught up with Lisp machines, and a specialised industry worth about half a billion dollars collapsed within a year. DARPA cut its AI funding sharply in 1987–88. The Fifth Generation project ended in 1992 without meeting its original goals.

The lesson of the two winters: intelligence that has to be written down by hand, rule by rule, doesn't scale to the real world.

For years afterwards, some researchers avoided the term "artificial intelligence" altogether and called their work "machine learning", "informatics", "knowledge-based systems" or "intelligent agents", partly to escape the stigma of broken promises. Much of that quieter work later turned out to matter.
The other road: neural networks
From the beginning there was a second approach. In 1943 Warren McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity", modelling neurons as simple on/off switches. Instead of writing rules, you could build a network of artificial neurons and let it learn from examples.
In July 1958 the US Office of Naval Research demonstrated Frank Rosenblatt's perceptron. After 50 trials, an IBM 704 computer learned to tell cards marked on the left from cards marked on the right. Rosenblatt called it "the first machine which is capable of having an original idea", and the press coverage was breathless.
The criticism came in 1969. In Perceptrons, Marvin Minsky and Seymour Papert proved mathematically what single-layer perceptrons could not do. The best-known example is the "exclusive or" (XOR) pattern: one or the other, but not both. Multi-layer networks could in principle get around this, but nobody had a good way to train them. Funding moved away from neural networks, and Rosenblatt died in a boating accident in 1971, on his 43rd birthday. How much the book itself caused the decline is still argued over.
Backpropagation, and who gets the credit
The missing piece was a way to work out how much each connection in a many-layered network was to blame for an error, and to adjust it. This method is backpropagation. Credit for it is genuinely contested:

Who
When
What

Seppo Linnainmaa
1970
The general mathematical method (reverse-mode automatic differentiation)

Paul Werbos
1974 thesis; 1982
Proposed applying it to train neural networks

Rumelhart, Hinton and Williams
1986, Nature
"Learning representations by back-propagating errors": showed convincingly that it worked and made it famous

The 1986 paper set off a revival. In 1989–90 Yann LeCun and colleagues at Bell Labs used convolutional neural networks trained by backpropagation to read handwritten digits, and by the late 1990s descendants of that work were reading the amounts on millions of bank cheques. But the networks of the 1990s were small, data was scarce and computers were slow. Other statistical methods often did as well or better, and neural networks fell out of favour again.
Deep Blue: a triumph of the old road
When IBM's Deep Blue beat Garry Kasparov in May 1997, it used none of this. It won the six-game rematch in New York by two wins, one loss and three draws, searching about 200 million chess positions a second with hand-tuned evaluation rules and specialised hardware. It was a triumph of search and engineering, not learning. Its success showed how much brute computation could do in a closed world with fixed rules. Its narrowness showed how little that told you about the open world.

Part 3 — The deep-learning turn and the ChatGPT moment
Three things changed in the 2000s, and the result is usually called deep learning, meaning neural networks with many layers.

Data. The web produced it in huge quantities. In 2009 Fei-Fei Li's team released ImageNet, millions of labelled photographs. Its annual competition used a 1,000-category subset with about 1.2 million training images.
Compute. Graphics cards (GPUs), built for video games, turned out to be very good at the multiplications neural networks need.
Technique. Researchers including Geoffrey Hinton, Yoshua Bengio and Yann LeCun kept improving how to train deep networks during the years when few people were interested.

2012: the result that changed minds
In 2012 a University of Toronto team (Alex Krizhevsky, Ilya Sutskever and Hinton) entered a convolutional network now called AlexNet in the ImageNet challenge. It had 60 million parameters and was trained for five to six days on two consumer graphics cards. Its top-5 error rate (the share of images where the right answer was not among its five guesses) was 15.3%. The second-best entry scored 26.2%.

AlexNet's error rate was about two-fifths lower than the next-best entry's: not a small step, a different league.

Within a few years nearly every leading computer-vision system was a deep network, and speech recognition and translation followed. In 2018 Bengio, Hinton and LeCun shared the Turing Award, computing's highest honour, "for conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing."
2016: Move 37
Go has far more possible positions than chess, which makes Deep Blue-style brute-force search hopeless. DeepMind, a London lab founded in 2010 and bought by Google in 2014, combined deep networks that learned to judge positions and suggest moves with search and reinforcement learning (learning by trial, error and reward, partly through playing against itself). In March 2016 in Seoul, its AlphaGo beat Lee Sedol, one of the strongest players of the era, 4–1.
In game two AlphaGo played Move 37, a move its own model estimated a human would choose about 1 time in 10,000. Professional players said several of AlphaGo's moves looked like mistakes at first sight and like deliberate strategy in hindsight. In game four, Lee's Move 78 was just as unexpected and won him the game. DeepMind says more than 200 million people watched. The same lab's later protein-structure system, AlphaFold, earned Demis Hassabis and John Jumper half of the 2024 Nobel Prize in Chemistry, shared with David Baker. The 2024 Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for foundational work on neural networks.
2017: "Attention Is All You Need"
The last key ingredient came from machine translation. Earlier systems read a sentence word by word, which is slow and tends to lose track of words far apart. In 2014 Dzmitry Bahdanau, Kyunghyun Cho and Bengio added attention, which lets the model look back at whichever earlier words are most relevant. In June 2017 eight researchers at Google, with Ashish Vaswani listed first, went further. Their paper "Attention Is All You Need" introduced the Transformer, built "solely on attention mechanisms", with no step-by-step reading at all.
Because a Transformer processes all the words of a passage at once, it makes full use of the parallel hardware that GPUs provide. Their translation model trained for 3.5 days on eight GPUs and set new records. Almost every major language model since then has been built on this design. Part 2 of this series, How Large Language Models Actually Work, explains what attention does.
Language models
A language model does one simple-sounding thing: given some text, it predicts what comes next. In 2018 two Transformer models showed how far that could go. OpenAI's first GPT and Google's BERT (October 2018) were "pre-trained" on large amounts of unlabelled text and then lightly adapted for specific tasks, and they beat specialised systems on many tasks. BERT set new records on eleven language benchmarks at once.
Then came scale. In February 2019 OpenAI announced GPT-2 but at first held back the full model, citing worries about misuse. It published the full version that November. In January 2020 an OpenAI team, with Jared Kaplan as first author, published scaling laws: a model's errors fell smoothly and predictably as model size, data and compute grew, "with some trends spanning more than seven orders of magnitude." In May 2020 GPT-3 appeared, with 175 billion parameters trained on 300 billion tokens. Given just a few examples in its prompt, it could translate, answer questions or unscramble words without any retraining. In 2022 DeepMind's Chinchilla paper corrected the recipe: most big models had been trained on too little data. Its 70-billion-parameter model trained on 1.4 trillion tokens outperformed much larger rivals.
From autocomplete to assistant
A raw language model is an unpredictable autocomplete. Ask it a question and it may answer, or it may write ten more questions. The fix was fine-tuning with human feedback. In OpenAI's InstructGPT work (March 2022), people wrote example answers and ranked the model's outputs, and the model was then trained with reinforcement learning to produce answers people preferred. The striking result: people preferred the answers of a 1.3-billion-parameter InstructGPT over those of the 175-billion-parameter GPT-3, a model about 100 times larger. How to behave turned out to matter as much as raw size.
On 30 November 2022 OpenAI released ChatGPT, a chat interface on a model fine-tuned this way. An analyst estimate by UBS, reported by Reuters in February 2023, put it at about 100 million monthly users by January 2023, which was then the fastest any consumer app had grown. GPT-4 followed in March 2023, and rival chatbots from Google, Anthropic, Meta and others followed within months.
Two open letters
The speed alarmed some of the field's founders. On 22 March 2023 the Future of Life Institute published an open letter calling on labs to "pause for at least 6 months" the training of systems more powerful than GPT-4. No pause happened. In May 2023 Geoffrey Hinton left Google so he could speak freely about the risks. On 30 May 2023 the Center for AI Safety published a one-sentence statement:

"Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." — Center for AI Safety (2023)

Its signatories included Hinton, Bengio, and the heads of Google DeepMind (Demis Hassabis), OpenAI (Sam Altman) and Anthropic (Dario Amodei). Other leading researchers thought such fears were overblown. Part 5 sets out both sides.

Part 4 — The state of affairs, late 2026
Everything in this part is dated, because it changes month by month. Where a 2026 figure couldn't be verified, the latest verified figure is given with its year.
Capabilities: very strong and very uneven
Stanford's AI Index 2026 (published April 2026, with data mostly through early 2026) found that several frontier models "now meet or exceed human baselines" on PhD-level science questions, multimodal reasoning and competition mathematics. On SWE-bench Verified, a test of fixing real software bugs, scores rose "from 60% to near 100% in a single year". A Google DeepMind system (Gemini Deep Think) reached gold-medal standard at the 2025 International Mathematical Olympiad.
The same report calls the frontier jagged: the top model it tested read an analogue clock correctly only 50.1% of the time. AI agents doing real tasks on a computer improved from about 12% to about 66% success on the OSWorld benchmark, against a human baseline of about 72%, but still fail about one attempt in three. In March 2025 the evaluation group METR found that the length of software tasks frontier models could finish half the time had been doubling about every seven months since 2019. METR adds caveats: the tasks were mostly software, and its "hours" mean what a newcomer to the codebase could do, not an experienced employee.
How do the labs compare? Epoch AI, an independent research group, combines many benchmarks into one Epoch Capabilities Index (ECI). As downloaded on 31 October 2026, the best-scoring model from each lab was:

Lab (country)
Best model in the index
ECI
Weights

OpenAI (US)
GPT-6 Astra
169
Closed

Anthropic (US)
Claude Opus 5.5
168
Closed

Google DeepMind (US/UK)
Gemini 4 Argon
166
Closed (limited access)

Meta (US)
Muse Wonder
162
Closed

DeepSeek (China)
DeepSeek V4.1 Flash
159
Open

Moonshot (China)
Kimi K3
158
Open

Alibaba (China)
Qwen 3.8 Max
158
Closed

xAI (US)
Grok 4.6
157
Closed

For scale, GPT-4 as released in March 2023 scores about 126 on the same index. The top three are within a few points of each other and their error bars overlap, so the "leader" is effectively a tie and changes from month to month. A single index also hides real differences between models at coding, maths and writing.
Compute and cost: up and down at the same time

Training compute for frontier language models has grown about 5× per year since 2020 (Epoch AI, February 2026). The cost of training the largest models has grown about 3.5× per year.
The cost of using a given level of ability is falling very fast. The AI Index 2025 found that the price of running a system as good as GPT-3.5 fell more than 280-fold between November 2022 and October 2024. Epoch AI found declines of between 9× and 900× per year depending on the task (March 2025).
Hardware: global AI computing capacity reached about 17.1 million H100-equivalent chips, growing about 3.3× a year since 2022. Nvidia supplies over 60% of it, and almost every leading AI chip is made by a single company, TSMC in Taiwan (AI Index 2026).
Energy: AI data-centre power capacity reached 29.6 gigawatts, which the AI Index compares to New York State at peak demand. The largest single sites are approaching a gigawatt each (Epoch AI).
Secrecy: frontier labs have mostly stopped publishing parameter counts and training details. Reported counts have "stayed near 1 trillion for three years" (AI Index 2026), but outsiders can no longer check.

Open versus closed
Closed models are used only through a company's app or interface. Open-weight models publish the trained network itself, so anyone can run, study or modify it. In May 2023 the best open model trailed the best closed one by about 15% on a popular head-to-head ranking. By August 2024 the gap was 0.5%. Since then it has widened again: 3.4% in March 2026 (AI Index 2026).
On Epoch's index, the best open-weight model as of October 2026 (DeepSeek's) roughly matches the best closed models of early March 2026, a lag of about seven months by this library's calculation from Epoch's data. Epoch separately estimated a six-to-seven-month lag on cyber-security skills. The leading open models now come mostly from Chinese labs (DeepSeek, Moonshot, Zhipu, Alibaba's Qwen family), with Meta, Mistral in France, Nvidia and others also releasing open models.
The case for open weights: anyone can inspect them, researchers and smaller countries aren't dependent on a few companies, and they cost less. The case against: once released, safeguards can be stripped out and the model can't be recalled. Both are serious arguments, and they are being argued in policy right now.
Who builds it
Industry, not universities, now produces over 90% of notable frontier models (AI Index 2026). In brief:

OpenAI (founded 2015 as a nonprofit) makes ChatGPT and the GPT series. A 2025 restructuring made its business a public benefit corporation controlled by the nonprofit OpenAI Foundation. Microsoft has invested more than $13 billion in it.
Google DeepMind (DeepMind founded 2010, merged with Google Brain in 2023) makes the Gemini models and AlphaFold.
Anthropic (founded 2021 by former OpenAI staff, also a public benefit corporation) makes the Claude models, one of which helped write this library.
Meta made the open Llama models central to its strategy, and by 2026 it releases both open and closed models.
xAI makes the Grok models. Mistral AI, in Paris, is the highest-placed European lab in Epoch's index.
DeepSeek (Hangzhou, owned and funded by the hedge fund High-Flyer) released its open-weight R1 reasoning model in January 2025. By the AI Index's measure, it briefly matched the top US model in February 2025.

The money involved is enormous. US private AI investment reached $285.9 billion in 2025, more than 23 times China's $12.4 billion, though the AI Index notes that private figures understate Chinese state funding. Use is spreading fast: generative AI reached 53% of the population within three years, faster than the PC or the internet. Adoption varies widely by country, and the US ranks 24th at 28.3%.

Part 5 — Rules, jobs and the argument about risk
Regulation, verified as of October 2026
The European Union has the most complete law, the AI Act. It sorts AI uses by risk, bans a short list of practices outright (social scoring, harmful manipulation and emotion recognition in workplaces and schools, for example) and puts heavier duties on "high-risk" uses. According to the European Commission's page, updated 19 October 2026, the timeline is:

Date
What applies

1 Aug 2024
The Act enters into force

2 Feb 2025
Banned practices and AI-literacy duties

2 Aug 2025
Rules for general-purpose AI models, and governance

2 Aug 2026
General application, including transparency duties (for example, labelling deepfakes)

Dec 2026
A ninth ban, on AI that generates non-consensual intimate images or child sexual abuse material, added by the "AI Omnibus"

2 Dec 2027
High-risk rules for uses such as biometrics, critical infrastructure, education, employment, migration and border control (postponed from August 2026)

2 Aug 2028
High-risk rules for AI built into regulated products, such as toys or lifts

The postponement came through the AI Omnibus, proposed in November 2025 and in force from 27 July 2026. The Commission's stated reason is that the rules should apply once companies have support tools such as technical standards. The same package strengthened the EU AI Office's oversight of systems built on general-purpose models.
The United States has no comprehensive federal AI law. Policy has come mainly through executive orders. In January 2025 President Trump revoked President Biden's 2023 AI safety order. In July 2025 the White House published "America's AI Action Plan", with more than 90 actions aimed at speeding up development and building data centres. In December 2025 Executive Order 14365 set up a Justice Department task force to challenge state AI laws the administration considers too burdensome. States have gone their own way. California's SB 53, signed in September 2025, requires large frontier developers to publish safety frameworks and report critical safety incidents, and it protects whistleblowers.
China has had binding rules for public generative-AI services since August 2023. Rules requiring AI-generated content to be labelled, issued by four agencies including the Cyberspace Administration, took effect on 1 September 2025.
Internationally, the second International AI Safety Report, chaired by Yoshua Bengio, written by over 100 experts and backed by over 30 countries and international organisations, was published on 3 February 2026.
What the evidence says about jobs
The honest answer is early, mixed and contested:

Productivity in specific jobs. In a study of 5,179 customer-support agents, an AI assistant raised issues resolved per hour by about 14% on average and by 34% for novices, with little effect on the most experienced (Brynjolfsson, Li and Raymond, 2023). The AI Index 2026 summarises such studies as showing gains of 14–26% in customer support and software, with "weaker or negative effects in tasks requiring more judgment."
But not always. In a 2025 randomised trial by METR, 16 experienced open-source developers took 19% longer with AI tools, although they had predicted the tools would make them 24% faster.
Young workers. Using payroll data from ADP through June 2026, Stanford's Erik Brynjolfsson and colleagues found that employment of 22-to-25-year-olds in the most AI-exposed jobs was 19% below where it would have been had it kept pace with less-exposed peers (revised August 2026). They found no comparable gap for experienced workers and "no evidence of widespread, economy-wide job displacement." The authors call these early descriptive signals, not proof of cause, and note that the effect weakens when they control for education.
The whole economy. Yale's Budget Lab found in October 2025 that the overall US labour market had "not experienced a discernible disruption" in the 33 months after ChatGPT. Past technologies took decades to reshape work.

So far the evidence points to fewer entry-level jobs in the most exposed occupations, not mass unemployment. Whether that is the start of a bigger shift is the open question.

The Careers pieces on the Software Engineer (27/03) and Data Scientist (27/14) show what this looks like inside two jobs that work closely with the technology.
Three cases, each at its strongest
The optimist's case. AI is a general-purpose technology like electricity, and it is getting cheaper fast. Dario Amodei of Anthropic argued in "Machines of Loving Grace" (October 2024) that powerful AI could compress "50-100 years" of biological progress into "5-10 years", picturing "a country of geniuses in a datacenter". Sam Altman of OpenAI wrote in "The Intelligence Age" (September 2024) that superintelligence might arrive "in a few thousand days" and make widespread prosperity possible. AlphaFold's Nobel Prize is the strongest evidence so far that the scientific promise is real. Optimists also point to consumer value, which the AI Index puts at $172 billion a year in the US.
The sceptic's case. It comes in several forms. Technical: in "On the Dangers of Stochastic Parrots" (2021), Emily Bender, Timnit Gebru and co-authors argued that language models stitch together linguistic forms "without any reference to meaning" and carry hidden biases and environmental costs. Yann LeCun, who left Meta in November 2025 to build "world models", argues that predicting text alone is a dead end for human-level intelligence. Melanie Mitchell's "Why AI is Harder Than We Think" (2021) traces the field's repeated overconfidence. Economic: MIT's Daron Acemoglu estimated in 2024 that AI would raise total factor productivity by "no more than" about 0.66% over ten years. Historical: every previous AI boom ended in a winter. Sceptics also note that the people making the boldest predictions often run companies raising money on them.
The safety-focused case. This view doesn't depend on AI being overhyped or underhyped. It starts from the evidence that capabilities are rising quickly while our ability to understand and control these systems is not keeping up. The AI Index 2026 reports that "responsible AI is not keeping pace with AI capability". Documented AI incidents rose from 233 in 2024 to 362 in 2025, and improving one safety property can make another worse. Researchers such as Hinton and Bengio worry about misuse (cyber-attacks, bioweapons), about losing human control of systems more capable than us, and about power concentrating in a few hands. Their proposed responses range from independent testing and incident reporting to international limits on the most powerful systems. Sceptics of this view reply in different ways. Andrew Ng said in 2015 that worrying about AI wiping us out was like "worrying about overpopulation on Mars". LeCun argues that AI can be made safe step by step, as cars and rockets were, and that machines will have no built-in desire to take control. Others say a focus on speculative catastrophes distracts from present harms such as bias, surveillance and job losses.
These camps overlap more than the shouting suggests. Many people hold parts of all three views at once: real benefits, overstated timelines, serious risks. How much weight each deserves is a live disagreement, not a settled question.

The takeaway
AI began as a philosopher's idea that reasoning is calculation. Then came seventy years of hand-written rules that kept running into the complexity of the real world. The breakthrough came from the other road: networks that learn from enormous amounts of data, powered by graphics chips and, since 2017, the Transformer. As of late 2026 the strongest systems match experts on many tests and still misread clocks, the frontier labs are within a few points of each other, and the best open-weight models trail by about seven months. Costs per task are falling fast while total spending soars. The EU's main high-risk rules have been pushed to December 2027, the US relies on executive orders and state laws, and the job evidence so far points to fewer entry-level openings rather than mass unemployment. The optimists, sceptics and safety researchers can each point to real evidence.
Try this: pick one claim about AI you've seen this week ("AI will replace X", "AI can't do Y") and test it against this piece. Find the date on the evidence, ask who is making the claim and what they gain from it, and then try the task yourself with a free AI tool and note where it succeeds and where it fails. Then read Critical Thinking & Epistemic Hygiene (Skills for the AI Era 2) with your notes beside you.

Sources & Further Reading
Sources
- Thomas Hobbes, Leviathan (1651), ch. 5 (Project Gutenberg text). Leibniz, letter to Electress Sophia, March 1706, as cited in the Wikipedia article on the characteristica universalis (lead only); stepped reckoner dates from the Wikipedia article (lead only).
- al-Jazari, Book of Knowledge of Ingenious Mechanical Devices (1206), and Noel Sharkey's reconstruction; Jaquet-Droz automata (1768–74, Musée d'art et d'histoire, Neuchâtel); the Mechanical Turk (1770–1854). Wikipedia articles used for leads.
- Computer History Museum, "The Engines" and "Ada Lovelace" pages (Babbage dates; debate over Lovelace's contribution). Lovelace's Notes (1843), full text at fourmilab.ch.
- A. M. Turing, "Computing Machinery and Intelligence", Mind 59 (236): 433–460, October 1950. Stanford Encyclopedia of Philosophy, "The Turing Test".
- McCarthy, Minsky, Rochester and Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence", 31 August 1955.
- Computer History Museum, AI and robotics timeline (Logic Theorist, ELIZA, Shakey, Deep Blue). IBM Heritage, "Deep Blue".
- Lighthill report (1973) and the Royal Institution debate, ALPAC (1966), the Fifth Generation project, the Lisp-machine collapse, and the R1/XCON estimates (the Wikipedia articles "AI winter" and "History of artificial intelligence", used for leads; Simon's and Minsky's predictions as quoted there).
- Cornell Chronicle (2019) on Rosenblatt's 1958 perceptron demonstration. Rumelhart, Hinton and Williams, Nature 323 (1986).
- Krizhevsky, Sutskever and Hinton, "ImageNet Classification with Deep Convolutional Neural Networks", NeurIPS 2012 (15.3% vs. 26.2%). Google DeepMind, "AlphaGo". NobelPrize.org, Physics and Chemistry 2024.
- Bahdanau, Cho and Bengio (arXiv 1409.0473); Vaswani et al. (arXiv 1706.03762); Devlin et al., BERT (arXiv 1810.04805); Kaplan et al. (arXiv 2001.08361); Brown et al., GPT-3 (arXiv 2005.14165); Hoffmann et al., Chinchilla (arXiv 2203.15556); Ouyang et al., InstructGPT (arXiv 2203.02155).
- Future of Life Institute, "Pause Giant AI Experiments" (22 March 2023). Center for AI Safety, Statement on AI Risk (May 2023).
- Stanford HAI, AI Index Report 2026 (April 2026) and 2025. Epoch AI, "Trends in AI" (updated 5 February 2026), Epoch Capabilities Index data (downloaded 31 October 2026), "LLM inference prices" (March 2025) and "Open-weight models trail closed models in cyber capabilities" (October 2026). METR, Kwa et al. (arXiv 2503.14499) and Becker et al. (arXiv 2507.09089).
- European Commission, "AI Act" regulatory framework page (updated 19 October 2026). White House: rescission of EO 14110 (20 January 2025); "America's AI Action Plan" (23 July 2025); EO 14365 (11 December 2025). Office of the Governor of California, SB 53 (29 September 2025). Cyberspace Administration of China, labelling measures (effective 1 September 2025). International AI Safety Report 2026 (3 February 2026).
- Brynjolfsson, Li and Raymond, NBER w31161. Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine?" (Stanford Digital Economy Lab, revised 12 August 2026). Gimbel et al., Yale Budget Lab (1 October 2025). Acemoglu, NBER w32487.
- Amodei, "Machines of Loving Grace" (October 2024). Altman, "The Intelligence Age" (23 September 2024). Bender, Gebru et al., FAccT 2021. Mitchell (arXiv 2104.12871).
Go deeper
- 📘 Genius Makers, Cade Metz (2021). The people behind the deep-learning revolution, from Hinton's lab to DeepMind and OpenAI. The best narrative history of 2012–2020.
- 📘 Artificial Intelligence: A Guide for Thinking Humans, Melanie Mitchell (2019). A clear, sceptical explanation of what these systems can and can't do, by a working researcher.
- 📘 The Alignment Problem, Brian Christian (2020). How machine learning goes wrong and what researchers are doing about it. Fair to every side of the safety debate.
- 🌐 Stanford HAI AI Index, the yearly data report used throughout Part 4. Free, with charts for every claim.
- 🌐 Epoch AI: Trends in AI, regularly updated charts on compute, cost and capabilities.
- 🌐 Turing, "Computing Machinery and Intelligence" (1950). The paper itself is readable, witty and still surprisingly current.
- ▶️ AlphaGo (2017), the documentary about the Lee Sedol match, including Move 37.
Related pieces in this library: Alan Turing: The Machine That Defined Computing (History-Makers 4) · Philosophy of Mind: The Hard Problem of Consciousness (Philosophy 10, on Searle's Chinese Room) · Human-AI Collaboration (Skills for the AI Era 5) · What It's Like to Be a Software Engineer (Careers 3) · What It's Like to Be a Data Scientist (Careers 14) · in this series: How Large Language Models Actually Work (Technology 2) and GPUs: From Video Games to the Engine of AI (Technology 12).


How Large Language Models Actually Work
---------------------------------------
One small trick, guessing the next word, repeated trillions of times. What that really means, and what nobody yet understands.

Follows: Artificial Intelligence: Where It Came From and Where It Stands (Technology 1), with help from Markov Chains (Mathematics 15) and Vectors and Matrices (Mathematics 14).

  
    In this deep dive
    
      The one trick: guess the next word (with a toy you can train)
      Inside the machine: tokens, embeddings and attention
      Training: from raw predictor to assistant
      Why they make things up, and how much they can hold in mind
      Reasoning models, and what nobody understands yet
    
  

  Type "The capital of France is" into a chatbot, and before it says anything, something very simple happens. The model looks at every word so far and gives every possible next piece of text a probability. "Paris" gets a very high one. It picks a piece, adds it to the text, and does the whole thing again for the piece after that.

  Everything a large language model (LLM) does, whether answering a question, writing code or a poem, or apologising for a mistake, comes from repeating that one step. The interesting questions are how a machine gets so good at it, and why being good at it looks so much like thinking.

  This library is written with the help of an AI model (Anthropic's Claude), one of the systems this piece describes. Everything here is drawn from published research, not from the model's own account of itself. Part 5 explains why a model's account of itself can't be trusted.

  Part 1 — The one trick: guess the next word

  The idea is older than computers. In 1913 the Russian mathematician Andrey Markov went through a long passage of Pushkin's verse novel Eugene Onegin letter by letter, sorting each letter into vowel or consonant and counting what followed what. He showed that whether a letter is a vowel depends on the letter before it, and his method became the "Markov chain" (Mathematics 15 in this library). In 1948 Claude Shannon, the founder of information theory, built fake English by hand: he picked each word using only the frequencies of word pairs in real text. One of his samples begins:

  "THE HEAD AND IN FRONTAL ATTACK ON AN ENGLISH WRITER THAT THE CHARACTER OF THIS POINT IS THEREFORE ANOTHER METHOD…" — Shannon, 1948

  It's nonsense, but nonsense with an English rhythm. Shannon noticed that the samples became noticeably more English-like each time he let the next word depend on more of what came before. That is the basic principle behind a modern LLM.

  You can try it below. This is a toy, and it is clearly labelled as one. It shows the idea of prediction by doing nothing but counting which words followed which in a short story (written for this page). It is not how a Transformer works inside, and Part 2 explains the difference. You can also paste in any text of your own and retrain it instantly.

  
    Interactive · A toy next-word predictor (counting, not a neural network)

    Training text (edit it or paste your own, then press Train)
    
    
      Train on this text
      Restore the fox-and-crow story
    
    

    Your prompt (the text so far)
    
    

    Look back
      1 word2 words
    When stuck
      Admit itGuess anyway

    
    

    Temperature: 1.0 
    

    
      Add one word
      Write 20 words
      Undo
      Start over
    
    
    Tap any bar to add that word yourself. The toy treats whole words and punctuation as its "tokens"; real models use word pieces (Part 2).
  

  What the toy teaches, and where it breaks

  It is only counting. Type "the" and look back one word. In the story, "the" is followed by "fox" 29 times out of 101 and by "crow" 26 times, so those are its two top guesses (about 29% and 26%). There is no understanding involved, only a table of what came next.

  More context gives sharper guesses, then copying. Switch to looking back two words. The guesses get more sensible, but in this little story about three-quarters of all two-word contexts (145 of 193) were followed by only one word, so the toy mostly recites its training text. Language models face the same tension at enormous scale: they need enough context to make sense, but not so much that they just repeat what they read.

  Temperature 0 gets stuck in loops. Set the temperature to 0, so the toy always takes its top choice, and press "Write 20 words" from "the fox". Looking back one word, it writes "The fox said to the fox said to the fox…" forever. Real chatbots have the same temperature setting. Low values make them predictable, higher values make them more varied and more error-prone.

  "Guess anyway" is the seed of hallucination. Type "the fox sat on" and choose "Admit it". The word "on" never appears in the story, so the toy stops. Switch to "Guess anyway" and it falls back to shorter contexts and produces something fluent and confident with no basis in its training text. Part 4 explains why real models face a version of this pressure.

  Why not just build a giant version of this table? Because the number of possible contexts explodes. A modern vocabulary has around 128,000 tokens, which gives over 16 billion possible two-token contexts. A real prompt might be thousands of tokens long, and almost every one is new: no table could have seen it. To predict well in situations it has never seen, a model has to generalise, treating "the crow sat" as being like "the fox sat" because crows and foxes are both animals in the story. That needs a way of representing words that captures what they mean, not just how they're spelled.

  Part 2 — Inside the machine: tokens, embeddings and attention

  Tokens: the pieces of text

  A neural network only handles numbers, so text first has to be cut into pieces called tokens, each with an ID number. Whole words would need an impossibly large vocabulary, and single letters make texts very long. Nearly all modern models compromise with sub-word tokens. A widely used method, introduced to machine translation in 2015 by Rico Sennrich and colleagues, adapts an old compression trick called byte-pair encoding: start with single characters, then repeatedly merge the pair that appears most often until the vocabulary reaches a set size.

  The result is a vocabulary of common words, common word parts and single characters. Meta's Llama 3 models (2024) use a vocabulary of 128,000 tokens and average about 3.94 characters per token on English text. A rare word might be split into several pieces, while "the" is a single token.

  Tokens explain some odd behaviour. A model that might receive "strawberry" as two or three chunks never directly sees its individual letters, which helps explain why spelling and letter-counting puzzles can trip up systems that write fluent essays. Tokens also make some languages more expensive. A 2023 study by Aleksandar Petrov and colleagues found that the same text translated into different languages could need up to 15 times as many tokens. Since AI services charge by the token and context windows are measured in tokens, that difference has real costs for speakers of those languages.

  Embeddings: meaning as position

  Each token ID is then turned into an embedding, a long list of numbers that can be pictured as a point in a space with thousands of dimensions. In GPT-3 (2020) every token became a list of 12,288 numbers. In the largest Llama 3 model it is 16,384. Nobody chooses these numbers by hand. They start random and are adjusted during training until tokens used in similar ways end up near each other.

  The famous demonstration came from Tomas Mikolov and colleagues at Google in 2013. With their "word2vec" vectors, taking the vector for King, subtracting Man and adding Woman landed nearest to Queen. Directions in the space had come to mean something, such as "male to female" or "country to capital", without anyone programming them. (The arithmetic of vectors is explained in Vectors and Matrices, Mathematics 14.)

  An embedding is a word's address in a space of meaning. Training is what draws the map.

  There's one crucial difference between word2vec and a modern LLM. In word2vec, "bank" has one fixed vector. Inside a Transformer, a token's vector is only the starting point. As it passes through the model's layers it is repeatedly updated using the words around it, so "bank" in "river bank" and "bank" in "bank loan" end up in very different places. The mechanism that does this updating is attention.

  Attention: how words look at each other

  Consider: "The trophy didn't fit in the suitcase because it was too big." What is "it"? The trophy. Now change one word: "…because it was too small." Now "it" is the suitcase. You resolved this instantly using knowledge about sizes and containers. A model has to do the same thing, and attention is how it does it.

  Here is the intuition, leaving out the maths. At every layer, each token produces three things from its current vector:
  
    a query: roughly "what kind of information am I looking for?"
    a key: "what kind of information do I hold?"
    a value: "what I'll pass on if you pay attention to me."
  
  The model compares one token's query with every earlier token's key. Good matches get high scores. The scores are turned into percentages that add up to 100%, and the token takes in that mix of the other tokens' values. Then every token does this at once.

  
    Illustration · Where might "it" look? (made-up numbers)
    Sentence: "The trophy didn't fit in the suitcase because it was too big." These percentages are invented to show the idea, not measured from any real model. Real attention is spread across many heads and layers at once.
    
  

  Three details make this powerful:
  
    Many heads. Each layer runs many attention patterns side by side, called "heads", and each can learn to track something different, such as grammar, who "she" refers to, or matching brackets in code. GPT-3 has 96 layers with 96 heads each. The largest Llama 3 model has 126 layers with 128 heads.
    Thinking per token. After attention, each token's vector passes through its own small network (a "feed-forward" block). This is where most of a model's parameters are. In the largest Llama 3, that inner layer is 53,248 numbers wide, against 16,384 for the token vectors.
    No peeking ahead. In a language model, each token may attend only to tokens before it. That is what makes the whole machine a next-token predictor: the final vector at the last position is turned into a probability for every token in the vocabulary.
  

  
    Text→Tokens→Embeddings→Attention + feed-forward, layer after layer→Probabilities for the next token→Pick one, add it, repeat
  

  Attention has a cost. Every token is compared with every other, so doubling the length of the text roughly quadruples that part of the work. That's a big reason long documents are expensive to process, and why researchers keep inventing shortcuts.

  Compare the toy above. It looks back a fixed one or two words and matches spellings exactly. A Transformer looks back across thousands of tokens, decides which ones matter for this particular prediction, and compares meanings rather than spellings. Both are guessing the next word, but in very different ways.

  Part 3 — Training: from raw predictor to assistant

  Data, scale and billions of knobs

  A model's parameters are all the numbers inside it: the embeddings, the query, key and value machinery, the feed-forward blocks. GPT-3 had 175 billion of them. Before training they are random, and the model produces gibberish.

  Training is a loop, repeated an enormous number of times:
  
    Take a chunk of real text.
    At every position, have the model predict the next token.
    Measure how much probability it gave to the token that actually came next. Less probability means a bigger error (the "loss").
    Use backpropagation (Technology 1, Part 2) to work out how each parameter contributed to the error, and nudge every one slightly in the direction that reduces it.
  
  No one tells the model grammar, facts or style. The only goal is "predict the next token better". Grammar, facts and style get learned because they help with that.

  
    Model (year)ParametersTraining tokensNotes
    GPT-3 (2020)175 billion300 billionWeb text filtered from 45 TB compressed down to 570 GB, plus books and Wikipedia
    Chinchilla (2022)70 billion1.4 trillionSmaller but trained on more data, and it beat much larger models
    Llama 3 405B (2024)405 billion15.6 trillion3.8 × 1025 calculations, on up to 16,000 H100 chips
  

  To get a sense of 15.6 trillion tokens: a person reading eight hours a day, every day, at 238 words a minute (the pace this library's reading times assume) gets through about 42 million words a year. Even allowing for tokens being shorter than words, reading Llama 3's training text would take that person well over 100,000 years.

  Why so much? In 2020 an OpenAI team, with Jared Kaplan as first author, found scaling laws: the model's prediction error falls smoothly and predictably as you increase parameters, data and compute, across more than seven orders of magnitude. In 2022 DeepMind's Chinchilla paper refined the recipe. For a given compute budget, parameters and data should grow together, which works out to roughly 20 training tokens per parameter for Chinchilla itself. These laws turned AI progress into something companies could plan and budget for. Epoch AI estimates that training compute for frontier models has grown about five-fold every year since 2020 (February 2026).

  Scaling laws made progress predictable: spend ten times more and you could forecast roughly how much better the model would get.

  Two caveats as of 2026. First, the biggest labs no longer disclose most of these numbers. Stanford's AI Index 2026 notes that frontier labs have stopped reporting parameter counts and training details, so outside estimates have to stand in. Second, high-quality human text is limited. The same report finds that synthetic (AI-generated) data "is still not replacing real data in pre-training", while careful filtering and better post-training are helping. One small open model it highlights performs comparably to a model with nearly 90 times as many parameters on several benchmarks, through careful data curation alone.

  From autocomplete to assistant

  A model straight out of pre-training (a base model) has learned to continue documents, not to help. Ask it "What's a good name for a cat?" and it might continue with more questions, as if it were reading a quiz. Turning it into an assistant takes further training stages:

  
    StageWhat it learns fromWhat it changes
    Pre-trainingTrillions of tokens of general textLanguage, facts, patterns of reasoning: the raw ability
    Supervised fine-tuningA much smaller set of example conversations written or checked by peopleThe format and manner of a helpful answer
    Learning from feedback (RLHF and relatives)People, or another AI following written principles, comparing answers and saying which is betterPreferences: more helpful, honest and harmless, more often
  

  OpenAI's InstructGPT paper (2022) described the now-standard recipe. People ranked several answers to the same prompt. Those rankings trained a separate reward model to predict which answers people would prefer, and the language model was then trained with reinforcement learning to score highly on it. The result changed the field: people preferred the answers of a 1.3-billion-parameter InstructGPT over those of the original 175-billion-parameter GPT-3, a model more than 100 times larger.

  Variants followed. In Anthropic's "Constitutional AI" (December 2022), the human rankings for harmlessness are largely replaced by an AI model judging answers against a written list of principles, which its authors called "RL from AI Feedback". In 2023 Stanford researchers introduced Direct Preference Optimization, which gets a similar effect from preference data without a separate reward model or reinforcement-learning loop. Most labs now combine several such methods. The details are often unpublished.

  This stage is where a model's apparent "personality" comes from: its tone, its willingness to say "I'm not sure", what it refuses. It is also where trade-offs get set. The AI Index 2026 reports research finding that improving one responsible-AI quality, such as safety, can make another, such as accuracy, worse. Fine-tuning shapes behaviour. It doesn't add much new knowledge, and it doesn't remove the base model's underlying tendencies.

  Part 4 — Why they make things up, and how much they can hold in mind

  Hallucination

  When a model states something false with confidence, such as a fake citation, a made-up quote or a wrong date, it is called a hallucination. The toy's "Guess anyway" button shows the basic problem: a system built to produce a likely continuation will produce one even when it has nothing solid to base it on. Nothing in the next-token step separately checks whether the result is true.

  A 2025 paper by Adam Tauman Kalai and colleagues compares models to students taking an exam. If a blank answer scores zero and a guess might score a point, the best strategy is always to guess. Most AI benchmarks are scored this way, so training for high benchmark scores rewards confident guessing over admitting uncertainty. Their proposed fix is social as much as technical: change how the main benchmarks are scored so that "I don't know" stops being penalised.

  Models are trained to be good test-takers, and on most tests a confident guess beats an honest "I don't know".

  The problem is real and measurable. In June 2023 a US federal judge fined the lawyers in Mata v. Avianca $5,000 after they filed a brief citing court decisions that ChatGPT had invented. The AI Index 2026 reports a benchmark that, unusually, does not punish declining to answer. On it, hallucination rates across 26 leading models ranged from 22% to 94%, so models differ a lot. Common defences include letting the model search the web or a document collection and cite what it finds (called "retrieval"), training it to abstain, and, most importantly, a human checking anything that matters. (See Critical Thinking & Epistemic Hygiene, Skills for the AI Era 2.)

  Knowing why hallucinations happen suggests three habits that help with any chatbot:
  
    Give it the source instead of asking it to remember. Paste the document, the policy or the data. A model summarising text that's in front of it has much less to invent than one recalling a detail from training.
    Ask for checkable specifics, then check them. Names, dates, quotes, citations and numbers are where invented details hide. Follow at least one link or reference all the way to the original.
    Make "I don't know" an acceptable answer. Say so in the prompt. It doesn't remove the guessing pressure described above, but it helps, and it reminds you that a fluent answer and a true answer are different things.
  

  The context window

  A model's context window is how much text it can take in at once: your prompt, the conversation so far, any documents, and its own reply. Everything outside the window is simply invisible to it.

  
    Model (year)Context window
    GPT-3 (2020)2,048 tokens: a few pages
    Llama 3 405B (2024)128,000 tokens: roughly a full-length novel
    Gemini 1.5 Pro (February 2024, preview)Up to 1 million tokens: by Google's estimate over 700,000 words, or an hour of video
  

  Epoch AI estimates that context windows have grown about 30-fold per year since 2023. But accepting a long text is not the same as using it well. A 2023 Stanford-led study found models were best at using information at the beginning or end of a long input and noticeably worse in the middle ("lost in the middle"). The AI Index 2026 notes that the gap "between accepted and usable context length is wide".

  One common misunderstanding: when a chatbot "remembers" you from an earlier conversation, it is usually because saved notes are put back into its context window. The model's parameters don't change while you talk to it. It learns nothing from you mid-conversation, unless the company later uses conversations for training, which is a policy choice you can often turn off.

  Part 5 — Reasoning models, and what nobody understands yet

  What "reasoning" models changed

  In January 2022 a Google team, with Jason Wei as first author, showed that simply showing a model worked examples, with intermediate steps written out ("chain of thought"), sharply improved its maths and logic. With just eight such examples, a 540-billion-parameter model set a new record on a benchmark of grade-school maths word problems. Writing out steps gives the model more tokens, and so more computation, to spend on a hard problem before it commits to an answer.

  The next step was to train for it. On 12 September 2024 OpenAI released a preview of o1, a model trained to work through a long chain of reasoning before answering. In January 2025 China's DeepSeek published R1 with open weights and a paper describing the method. It showed that reinforcement learning on problems with checkable answers, such as maths with known solutions and code that must pass tests, could build reasoning ability without human-written examples of reasoning. Behaviours such as checking its own work and changing strategy, the paper reports, emerged by themselves.

  Two things changed. First, a model can now be made better by letting it think longer at answer time, not just by training a bigger model. Second, progress sped up. Epoch AI's capabilities index for the best models has risen about 14 points a year since reasoning models appeared in September 2024, against about 6 a year for non-reasoning models. A reasoning system reached gold-medal standard at the 2025 International Mathematical Olympiad (AI Index 2026). Its abilities are still uneven: the same report found the best model it tested read analogue clocks correctly only 50.1% of the time.

  And a warning: the visible reasoning isn't necessarily the real reasoning. In a 2025 study by Anthropic researchers, reasoning models were given hidden hints and then examined to see whether their written reasoning admitted using them. When the models did use a hint, they mentioned it in at least 1% of cases but often under 20%. Reading a model's chain of thought is useful for spotting problems, the authors concluded, but it can't rule them out.

  What we still don't know: looking inside

  This is the strangest fact about LLMs. Engineers know exactly how to build and train them, but nobody fully understands what the trained network is doing. The parameters aren't written by anyone. They are found by training, and they don't come labelled.

  The field that tries to read them is called interpretability, and it has made real progress:
  
    Superposition (2022). Anthropic researchers showed, in small toy models, that networks pack more concepts than they have neurons by overlapping them. That is why looking at single neurons tells you little.
    Dictionaries of features (2024). Tools called sparse autoencoders try to unpack those overlaps into separate, readable "features". Anthropic extracted up to about 34 million features from its Claude 3 Sonnet model, including one for the Golden Gate Bridge. When the researchers artificially boosted that feature, the model began to identify itself as the bridge. OpenAI trained a 16-million-feature version on GPT-4, and Google DeepMind released "Gemma Scope", an open set of such tools for its Gemma 2 models, so outside researchers can do this work too.
    Tracing circuits (2025). Anthropic's "On the Biology of a Large Language Model" traced steps inside Claude 3.5 Haiku. Asked for "the capital of the state containing Dallas", the model represented "Texas" internally on the way to "Austin". When writing rhyming verse, it chose candidate rhyme words before writing the line. Most strikingly, for 36 + 59 it combined a rough estimate of the sum with a separate calculation of the last digit. Asked how it did it, it described the school method of carrying the one. It had a skill without accurate insight into how it performs it.
  

  The researchers themselves stress the limits. Even in their successful case studies, they write, their methods "only capture a small fraction of the mechanisms of the model".

  The open questions
  
    Do they understand? Emily Bender, Timnit Gebru and co-authors (2021) argued that language models stitch together forms "without any reference to meaning" and called them "stochastic parrots". Evidence of internal steps like "Dallas → Texas → Austin" suggests more structure than parroting. Whether that structure amounts to understanding is partly a philosophical question about what understanding is. See Philosophy of Mind (Philosophy 10) on Searle's Chinese Room.
    Do abilities suddenly "emerge"? Some abilities seem to appear abruptly at a certain size. A 2023 Stanford paper argued that much of this is an artefact of how tests are scored, and that with smoother measures, improvement looks gradual. The debate continues.
    How far will scaling go? Compute keeps growing, but high-quality data is limited and costs are huge. Reasoning-time compute has opened a second path. Nobody knows where either path ends.
    Can we check them? If we can't fully read a model's internals, and its stated reasoning isn't always faithful, how do we confirm a powerful system is safe? This question links the science to the debates in Technology 1, Part 5.
  

  
    The toy and a real LLM, side by side
    
      The toy aboveA real LLM
      Basic jobPredict the next wordPredict the next token
      HowCounts exact word sequencesBillions of learned parameters, embeddings and attention
      Looks back1 or 2 wordsUp to hundreds of thousands of tokens or more
      Unseen contextStuck, or falls back to shorter contextsGeneralises from similar meanings
      Trained onOne short story (or your text)Trillions of tokens, then human and AI feedback
      Can we see why it chose a word?Yes, the counts are shownOnly partly. That is the interpretability problem
    
  

  
    The takeaway
    A large language model predicts the next token, adds it and repeats. Text is cut into sub-word tokens, each token becomes a vector whose position encodes meaning, and layer after layer of attention lets every token pull in what it needs from the words before it. Training adjusts billions of parameters to predict trillions of tokens of real text, and fine-tuning with human or AI feedback turns that raw predictor into an assistant. The same design explains its weaknesses: it guesses confidently when it doesn't know, it can only see its context window, and its stated reasoning may not match what it actually computed. Reasoning models made thinking longer a new way to get better. Understanding what goes on inside is the field's biggest unsolved problem.
    Try this: in the toy, set "Look back" to 1 word and the temperature to 0, then write from "the fox" and watch it loop. Raise the temperature to 1 and write again. Then paste a page of your own writing, a school essay or a set of messages, and see which of your favourite phrases the toy learns first.
  

  
    Sources & Further Reading
    Sources. Markov, "An Example of Statistical Investigation of the Text Eugene Onegin…" (1913; English translation in Science in Context 19(4), 2006). Shannon, "A Mathematical Theory of Communication" (1948), section 3. Sennrich, Haddow and Birch (arXiv 1508.07909). Grattafiori et al., "The Llama 3 Herd of Models" (arXiv 2407.21783: 128K vocabulary, 3.94 characters per token, 405B parameters, 15.6T tokens, 3.8×1025 FLOPs, up to 16K H100s, 126 layers, 128 heads, 16,384 and 53,248 dimensions). Petrov et al., NeurIPS 2023 (arXiv 2305.15425). Mikolov et al. (arXiv 1301.3781, 1310.4546). Vaswani et al. (arXiv 1706.03762). Brown et al., GPT-3 (arXiv 2005.14165: 175B parameters, 300B tokens, 96 layers and heads, 12,288 dimensions, 2,048-token context, 45 TB to 570 GB). Kaplan et al. (arXiv 2001.08361). Hoffmann et al., Chinchilla (arXiv 2203.15556). Epoch AI, "Trends in AI" (updated 5 February 2026). Stanford HAI, AI Index Report 2026 (April 2026). Ouyang et al., InstructGPT (arXiv 2203.02155). Bai et al., Constitutional AI (arXiv 2212.08073). Rafailov et al., DPO (arXiv 2305.18290). Kalai, Nachum, Vempala and Zhang, "Why Language Models Hallucinate" (arXiv 2509.04664). Mata v. Avianca, Inc., S.D.N.Y., decided 22 June 2023. Google, "Our next-generation model: Gemini 1.5" (15 February 2024). Liu et al., "Lost in the Middle" (arXiv 2307.03172). Wei et al., chain of thought (arXiv 2201.11903). OpenAI o1-preview release, 12 September 2024. DeepSeek-AI, R1 (arXiv 2501.12948). Chen et al., "Reasoning Models Don't Always Say What They Think" (arXiv 2505.05410). Elhage et al., "Toy Models of Superposition" (arXiv 2209.10652). Templeton et al., "Scaling Monosemanticity" (Anthropic, 2024). Gao et al. (arXiv 2406.04093). Lieberum et al., Gemma Scope (arXiv 2408.05147). Lindsey et al., "On the Biology of a Large Language Model" (Anthropic, 2025). Bender, Gebru et al., FAccT 2021. Schaeffer, Miranda and Koyejo (arXiv 2304.15004). The toy's percentages and the "75% of two-word contexts" figure were computed from its built-in story and checked in node. The attention bars are an illustration with invented numbers.
    
      ▶️ 3Blue1Brown: "Transformers, the tech behind LLMs" and "Attention in transformers, step-by-step". The best visual explanation of embeddings and attention. Start here if you like to see the maths move.
      🌐 Jay Alammar, "The Illustrated Transformer". A diagram-by-diagram walkthrough of the 2017 architecture, a good next step after this piece.
      🌐 Stephen Wolfram, "What Is ChatGPT Doing … and Why Does It Work?" (2023). A long, patient essay that starts, like this piece, from next-word probabilities.
      ▶️ Andrej Karpathy, "Intro to Large Language Models" (1-hour talk, 2023). A plain-spoken overview by a well-known AI researcher and teacher. His "Let's build GPT" builds a small one in code, for programmers.
      🌐 "On the Biology of a Large Language Model" (2025). The interpretability paper behind Part 5, with interactive diagrams. Long, but the introduction is readable.
      📘 Why Machines Learn, Anil Ananthaswamy (2024). The maths of machine learning, from the perceptron to modern networks, explained for curious non-specialists.
      🌐 Related in this library: Artificial Intelligence: Where It Came From and Where It Stands (Technology 1) · GPUs: From Video Games to the Engine of AI (Technology 12) · Markov Chains (Mathematics 15) · Vectors and Matrices (Mathematics 14) · Philosophy of Mind (Philosophy 10) · Human-AI Collaboration (Skills for the AI Era 5) · What It's Like to Be a Data Scientist (Careers 14) · What It's Like to Be a Software Engineer (Careers 3)


Blockchain, Explained: Bitcoin, Ethereum, and What a Shared Ledger Is Actually Good For
---------------------------------------------------------------------------------------
How strangers learned to keep one honest record without a bank, what it costs, and where the idea earns its keep.

Follows: How Money Was Invented and Cryptography: The Math That Keeps Your Secrets
  

  
    In this deep dive
    
      The problem: spending the same money twice
      The precursors, and an eight-reference paper
      Hashes and chains (a working demo)
      Proof of work, mining, wallets and keys (mine a block yourself)
      Ethereum, smart contracts and the Merge
      Booms, busts and collapses
      Stablecoins and central bank digital currencies
      Energy, real uses, hype, and the state of things in 2026
    
  

  Not investment advice. This piece explains a technology and its history. Nothing here is a recommendation to buy, sell or hold any crypto-asset, and I am not a licensed financial adviser. Prices in this piece are there to show how violent the swings have been, not to predict the next one.

  Eighteen years ago today, on 31 October 2008, someone calling themselves Satoshi Nakamoto posted a link to a nine-page paper on a cryptography mailing list. Its title was Bitcoin: A Peer-to-Peer Electronic Cash System. Nobody knows for certain who wrote it. The author dropped out of public view around 2010, leaving behind, by analysts' estimates, about a million mined bitcoins.

  Out of that paper came a currency that, on the day this was written, had a combined market value of about $1.56 trillion, an industry that has produced some of the largest frauds of recent years, a second network that runs programs instead of just payments, a new kind of dollar now regulated by law, and an argument that is still going on about whether any of it was worth the electricity.

  How Money Was Invented made the case that money is a shared story about trust. Cryptography showed how one-way maths lets strangers keep secrets and prove who they are. This piece puts the two together. By the end you'll have built a small blockchain with your own thumbs, broken it, repaired it, and mined a block. You'll also know the strongest arguments for and against the whole idea.

  Part 1 — The problem: spending the same money twice

  A digital file can be copied perfectly. That is wonderful for a photo and fatal for money. If a coin is just a file, what stops you from sending the same file to two different shops? Computer scientists call this the double-spending problem, and for most of the history of digital money there was only one answer: a trusted middleman who keeps the master list.

  Your bank is that middleman. When you tap a card, nothing physical moves. The bank subtracts a number from your row in its ledger and adds it to someone else's. You can't spend the same £20 twice because the bank's database says you no longer have it. That works very well. It also means the bank (and behind it, the central bank and the state) can freeze your account, reverse a payment, charge a fee, or see everything you buy.

  Nakamoto's paper starts exactly here. It defines an electronic coin as "a chain of digital signatures": each owner signs the coin over to the next. Anyone can check those signatures. What they can't check is whether one of the owners quietly signed the same coin to two people. The usual fix, the paper says, is "a trusted central authority, or mint, that checks every transaction for double spending." The goal of Bitcoin was to get rid of the mint.

  The whole point of a blockchain is one sentence: get thousands of strangers who don't trust each other to agree on one list of who paid whom, with nobody in charge.

  The trick is to replace trust in a person with trust in three things anyone can check:
  
    Signatures, so only the owner of a coin can move it (the public-key idea from the Cryptography piece).
    A public, append-only record, so everyone sees every transaction and can spot a coin spent twice.
    A rule for agreeing on the order of events, so that when two conflicting payments appear, everyone ends up accepting the same one. This is the hard part, and it's where proof of work comes in.
  

  Computer scientists had been working on the third item for decades under the name "Byzantine fault tolerance": how a group of computers can agree when some of them may lie. The classic solutions assumed you knew who the participants were. On the open internet you don't. Anyone can create a thousand fake identities (a "Sybil attack") and outvote the honest members. The paper's answer is in a line worth quoting: "Proof-of-work is essentially one-CPU-one-vote." You can fake identities for free. You can't fake computing work.

  Part 2 — The precursors, and an eight-reference paper

  Bitcoin is often described as if it fell from the sky. It didn't. The computer scientists Arvind Narayanan and Jeremy Clark traced its pedigree in a 2017 article and concluded that nearly every component came from earlier academic work. What was new was the combination, and the decision to pay people to run it.

  
    WhenWhoWhat it addedWhat it lacked
    1982David Chaum"Blind signatures": a bank can sign digital cash without seeing which note it signed, so payments are private. His company DigiCash tried to sell it.Needed a central issuer, and no banks signed on; DigiCash went bankrupt in 1998.
    1991Stuart Haber and W. Scott StornettaTime-stamping documents by chaining each record's hash to the one before it. The core shape of a blockchain.Not money; no way to decide between rival histories without a trusted service.
    1992Cynthia Dwork and Moni NaorMake senders do a little computing work to fight junk mail: the first proof-of-work idea.Not a currency.
    1997Adam BackHashcash: keep hashing until you find an output that starts with enough zeros. This is exactly the puzzle Bitcoin miners solve.No protection against spending a stamp twice.
    1998Wei Dai; Nick Szabob-money and bit gold: money created by solving computational puzzles, recorded by a group rather than a bank.Never built; no robust answer to fake identities.
    2004Hal FinneyRPOW, the first system of reusable proofs of work, passed on like tokens.Never caught on as money.
  

  Nakamoto's paper has only eight references. Three of them are by Haber and Stornetta, which makes the pair the most cited authors in it. Others are Wei Dai's b-money, Adam Back's Hashcash, Ralph Merkle's work on hash trees, and a probability textbook by William Feller. A company built on Haber and Stornetta's method has, according to a 2018 report in Vice, been publishing hashes of its records in The New York Times since 1995, so you could argue the oldest working blockchain has been sitting in a newspaper for three decades.

  The first block of the network is dated 3 January 2009. That block, the "genesis block," contains a line of text that its creator chose to embed forever: "The Times 03/Jan/2009 Chancellor on brink of second bailout for banks." You can read it today in the raw data of that block. It's a timestamp (proof the block wasn't made earlier) and, most people assume, a comment on the bank bailouts of the financial crisis. Nine days later Nakamoto sent ten bitcoins to Hal Finney in the first transaction between two people. On 22 May 2010 a programmer named Laszlo Hanyecz paid 10,000 bitcoins for two pizzas, the first well-known purchase of real goods.

  Contested credit
  Three honest complications. First, nobody knows who Satoshi Nakamoto is. Journalists have named several candidates over the years; none has been proven. One man, Craig Wright, claimed for years to be Satoshi. In March 2024, after a five-week trial, a judge in the High Court in London ruled that he is not, and the written judgment found that documents he had submitted were forgeries. Second, how much credit belongs to the precursors is a live argument. The computer scientist Alan Sherman reported in 2019 that Chaum's 1982 dissertation already described nearly every part of a blockchain except proof of work. Third, the word "blockchain" isn't in the paper at all. Nakamoto wrote about a "chain of blocks." The single word caught on later.

  Part 3 — Hashes and chains

  Everything in a blockchain rests on one tool: the cryptographic hash function. Bitcoin uses one called SHA-256. It was designed by the US National Security Agency and published as a standard in 2001. It takes any input (a word, a book, a movie) and produces a 256-bit fingerprint, written as 64 hexadecimal characters. Three properties make it useful:
  
    Same input, same output, every time. Anyone can check a hash.
    Tiny change, totally different output. Change one letter and roughly 60 of the 64 characters change. This is called the avalanche effect.
    One-way. Given a hash, there's no known shortcut to find an input that produces it. You can only guess and check.
  

  Try it. The box below runs a real SHA-256, written out in plain JavaScript inside this page and checked against standard test values, so it works offline.

  
    Interactive · A real SHA-256 fingerprint
    Type anything
    
    SHA-256 of what you typed
    –
    
      Characters different from the hash of "hello"0 of 64
    
    Change "hello" to "hellp" or "Hello". About 60 of the 64 characters change, every time. You can't nudge a hash in the direction you want, which is the whole point.
  

  Now chain them. A block is a bundle of transactions plus the hash of the block before it. Because each block includes its predecessor's fingerprint, changing anything in an old block changes its hash, which no longer matches what the next block recorded, which changes that block's hash, and so on to the end. History becomes tamper-evident.

  That alone isn't enough. If rewriting the chain were cheap, a cheater could just recalculate every later hash. So the rule adds a cost: a block only counts if its hash starts with a certain number of zeros. In this demo, three zeros. Each block has a nonce, a number you can change freely until the hash happens to come out with the required zeros. All five blocks below have already been "mined."

  
    Interactive · Break a blockchain, then repair it
    Edit the text in any block. Watch that block and every block after it turn red: their links break. Then press Mine on each red block, in order, to find a new nonce and repair the chain. Notice how the cost of rewriting grows with every block you have to redo.
    
    Reset the chain
    Toy rules: each block's hash is SHA-256 of its number, the previous hash, its text and its nonce, joined together. A block is valid if its hash starts with 000 (green digits). Real Bitcoin hashes an 80-byte header twice, but the logic is the same.
  

  A blockchain doesn't make the past impossible to change. It makes changing it expensive, and makes the change visible to everyone.

  One more piece of plumbing. A real block doesn't hash every transaction directly into the header. Transactions are paired up and hashed, the results paired and hashed again, until one fingerprint is left: the Merkle root, named after the cryptographer Ralph Merkle. That lets a phone prove a payment is in a block by downloading a few dozen hashes rather than the whole block. The paper calls this "simplified payment verification," and it's why a light wallet on your phone doesn't need to store the entire chain.

  Part 4 — Proof of work, mining, wallets and keys

  Finding a nonce is pure guessing. Each extra zero you demand makes the job about 16 times harder, because each hexadecimal digit has 16 possible values. This is proof of work: a valid block is a receipt showing that someone spent real computing effort to make it. Try it at different difficulties.

  
    Interactive · Mine a block
    Block contents
    
    Required leading zeros (difficulty)
    
    
      Expected guesses on average–
      Guesses this time–
      Time on this device–
      Winning nonce–
    
    Pick a difficulty to start mining.
    Stop
    Five zeros needs about a million guesses on average, which can take several seconds on a phone. Your luck varies from run to run, which is exactly how real mining behaves: a lottery whose tickets cost electricity.
  

  Bitcoin plays the same game at a scale that's hard to picture. The network adjusts the difficulty every 2,016 blocks so that, however much computing power joins, a new block appears about every ten minutes. At block 974,063, mined in the early hours of 1 November 2026 (UTC), the difficulty implied that the world's miners were computing roughly 1,060 quintillion hashes per second (about 1021). That's done by specialised chips that do nothing else.

  Why would anyone spend that money? Because the first transaction in every block pays the miner newly created bitcoin plus the fees of the transactions inside. The paper compares it to "gold miners expending resources to add gold to circulation." The reward halves every 210,000 blocks, roughly every four years, which is why total supply can never pass 21 million.

  
    BlocksNew bitcoin per blockEra
    0 to 209,999502009 to 2012
    210,000 to 419,999252012 to 2016
    420,000 to 629,99912.52016 to 2020
    630,000 to 839,9996.252020 to 2024
    840,000 to 1,049,9993.125from 20 April 2024
    1,050,000 onward1.5625expected around spring 2028
  

  By the end of October 2026 about 20.1 million bitcoin had been issued, roughly 96% of the eventual total. The last fractions will trickle out around the year 2140. After that, miners will be paid by fees alone, and whether fees will be enough to keep the network secure is one of the genuine open questions about Bitcoin's long-term design.

  The longest chain wins
  Sometimes two miners find a block at almost the same moment, and the network briefly splits. The rule is simple: honest nodes follow the chain with the most accumulated work. Whichever branch gets the next block first becomes the longest, and the other is abandoned. That's why a payment is usually considered settled after it has a few blocks on top of it. To reverse an old payment, an attacker would have to redo the work of that block and every block after it, then outrun the whole honest network. With a majority of the network's computing power (the famous "51% attack") that becomes possible. Without it, the attacker's chances fall off exponentially with each block, which is what Nakamoto's probability section calculates.

  
    Interactive · Check the real genesis block yourself
    This builds Bitcoin's actual first block header (version, previous hash, Merkle root, time, difficulty bits and nonce 2,083,236,893), hashes it twice with the same SHA-256 used above, and compares the result with the block's published ID.
    Hash block 0
    
      The 80-byte header, in hex
      
      Double SHA-256, shown in Bitcoin's byte order
      
      
    
  

  Wallets and keys
  Here's the part most people get wrong: a crypto wallet doesn't hold coins. The coins exist only as entries in the shared ledger. What your wallet holds is a private key, a huge random number. From it, software derives a public key and from that an address you can share. To spend, your wallet signs a message saying "move these coins to that address," and every node can check the signature against your public key without ever learning the private one. That's the public-key idea from the Cryptography piece, though Bitcoin uses elliptic-curve signatures (ECDSA, and since the 2021 Taproot upgrade also Schnorr signatures) rather than RSA.

  The consequences are stark. If someone gets your private key (or the 12 or 24 "seed words" that generate it), they can take everything, and no one can reverse it. If you lose it, the coins are stuck forever. There's no "forgot password" link. That's why most people don't hold their own keys at all. They use an exchange that holds keys on their behalf, which quietly reintroduces the trusted middleman the system was built to remove. Several of the disasters in Part 6 happened exactly there.

  Be your own bank, and you also become your own security department, fraud team and complaints desk.

  Part 5 — Ethereum, smart contracts and the Merge

  Bitcoin's ledger records one kind of thing: who owns which coins. Its scripting language is deliberately limited. In 2013 a young programmer named Vitalik Buterin, a co-founder of Bitcoin Magazine, proposed something broader: a blockchain whose ledger records the state of programs. He circulated his white paper in late 2013 (ethereum.org dates the published version to 2014), and the network, Ethereum, launched on 30 July 2015.

  On Ethereum, you can deploy a smart contract: code that lives at an address on the chain and runs exactly as written whenever someone sends it a transaction. The usual analogy is a vending machine. Put in the right coin and the snack drops out. No cashier needed. Contracts can hold money, issue new tokens, run auctions, lend and borrow, or settle bets. Every node runs every contract and must get the same answer, so every computation costs a fee called gas, paid in Ethereum's currency, ether (ETH).

  "Code is law" sounds clean until the code has a bug. In 2016 a crowdfunded investment fund called The DAO, built as a smart contract, was drained of more than 3.6 million ETH through a flaw in its code. The community faced a hard choice. On 20 July 2016, at block 1,920,000, most of the network adopted a change that moved the funds back, in effect rewriting the outcome. A minority refused on principle, arguing that a ledger which can be edited when the majority dislikes the result isn't really immutable. They kept the original chain running as Ethereum Classic. Both still exist. The argument about whether that fork was wise has never been settled.

  Proof of stake
  Ethereum also used proof of work at first. On 15 September 2022, in an upgrade called the Merge, it switched to proof of stake. Instead of burning electricity to earn the right to add blocks, participants called validators lock up ether as a deposit, at least 32 ETH for your own validator. The protocol picks validators to propose and vote on blocks, pays them for honest work, and can destroy part of their deposit ("slashing") if they cheat. Attacking the chain means acquiring a huge amount of staked ether and then risking it.

  The energy effect was enormous. Ethereum's own site puts the drop at about 99.95%, and the Crypto Carbon Ratings Institute study it cites puts it at more than 99.98%, estimating that the whole network now uses about 0.0026 terawatt-hours (2,601 megawatt-hours) a year.

  
    Proof of work (Bitcoin)Proof of stake (Ethereum since 2022)
    Scarce resourceElectricity and specialised chipsLocked-up coins
    Who adds blocksWhoever finds a valid nonce firstValidators chosen by the protocol
    Punishment for cheatingWasted electricityPart of the deposit destroyed
    Supporters saySimple, battle-tested, ties security to the physical worldSecure at a tiny fraction of the energy
    Critics sayHuge energy useThose with the most coins earn the most; staking concentrates in a few large providers
  

  Ethereum keeps changing through named upgrades. Recent ones include "Pectra" (May 2025), which reworked staking, and "Fusaka" (December 2025). Since the original white paper, Ethereum has also come to rely on "layer 2" networks that bundle many transactions and post a compressed summary to the main chain, a way to raise capacity without making every node do everything.

  Enthusiasts see Ethereum as a neutral, global computer that nobody owns, on which anyone can build financial services. Sceptics point out that hacks of smart contracts have repeatedly cost users large sums, that most activity has been trading and speculation, and that a fee-charging shared computer is a slow, expensive way to do what an ordinary database does in microseconds. Both descriptions are accurate. Which one matters more depends on how much you value not having an owner.

  Part 6 — Booms, busts and collapses

  Crypto's history so far has been a series of manias followed by crashes. The pattern is worth seeing laid out, because each crash had a different cause.

  
    WhenWhat happened
    February 2014Mt. Gox, a Tokyo exchange that had been the dominant venue for bitcoin trading, halted withdrawals and filed for bankruptcy after losing about 850,000 bitcoin. About 200,000 were later found in an old wallet. A security firm's investigation concluded most of the rest had been stolen over years. The bankruptcy dragged on for a decade, with repayment deadlines pushed back again and again.
    2017–2018The ICO boom: a wave of projects raised money by selling new tokens ("initial coin offerings"), often with little more than a white paper. Bitcoin went from about $1,000 in January 2017 to a daily close of $19,650 on 16 December 2017, then fell to $3,183 a year later, a drop of about 84%.
    May 2022Terra/Luna. TerraUSD was an "algorithmic stablecoin" that tried to hold $1 not with reserves but by letting holders swap it for a sister token, LUNA. A lending protocol built on top, Anchor, paid depositors a yield of about 19.5%. When confidence cracked, holders rushed for the exit, LUNA was printed in vast amounts, and both collapsed within days. Estimates of the value wiped out run to about $40–45 billion.
    November 2022FTX, then one of the largest exchanges, collapsed after customers tried to withdraw and found a hole of about $8 billion. Customer money had been used by an affiliated trading firm. FTX filed for bankruptcy on 11 November 2022. Its founder, Sam Bankman-Fried, was convicted of fraud and conspiracy on 2 November 2023 and sentenced to 25 years in March 2024.
    December 2025Do Kwon, Terra's co-founder, who had pleaded guilty to two fraud charges in the US, was sentenced to 15 years in prison.
    2025–2026Bitcoin reached a record of about $126,000 on 6 October 2025, then fell by more than half to a daily close near $58,600 at the end of June 2026. On 31 October 2026 it closed near $77,400.
  

  Notice what failed. In most of these cases the Bitcoin and Ethereum ledgers themselves kept working exactly as designed. What broke were the businesses built around them: exchanges and lenders that held other people's coins, mixed them with their own, and lied about it. That's the old financial failure, a bank run on an unregulated bank, in new clothes. Terra is the exception. There, the design itself was the weakness: a promise of stability backed by nothing but confidence.

  The ledger rarely broke. The middlemen did, again and again, because people handed them their keys.

  Two readings of this history compete. Enthusiasts say every boom left real infrastructure behind, that each crash flushed out frauds, and that bitcoin has recovered from drops of 75–85% several times. Sceptics say that's survivorship bias: the base asset survived while millions of ordinary buyers lost money on the way, and an asset that halves on a whim doesn't do the job of money. Both camps agree on one practical point. Crypto has become a magnet for scams, because payments are fast and can't be reversed. The FBI's Internet Crime Complaint Center reported that in 2024, victims of cryptocurrency investment fraud lost more than $6.5 billion, the largest category of losses it recorded. Every lever in How to Not Get Scammed applies here. Urgency, secrecy, a stranger who wants payment in crypto: stop.

  Part 7 — Stablecoins and central bank digital currencies

  A currency that can halve in a year is a poor way to pay rent. Stablecoins fix that by pegging a token to an ordinary currency, usually the US dollar. The biggest ones are run by companies that hold dollars, US Treasury bills and similar assets in reserve and promise to redeem each token for $1.

  They have quietly become the busiest part of crypto. According to the data site DefiLlama, dollar-pegged stablecoins in circulation totalled about $299 billion on 1 November 2026. The two largest, Tether's USDT (about $180 billion) and Circle's USDC (about $72 billion), make up most of it. A smaller share is "crypto-backed," like Dai, which is backed by other crypto locked in smart contracts at more than face value.

  
    TypeHow it holds $1Main risk
    Fiat-backedThe issuer holds cash and short-term government debtIs the issuer honest, solvent and able to pay everyone at once?
    Crypto-backedOver-collateralised crypto locked in contractsA crash in the collateral faster than liquidations can keep up
    AlgorithmicArbitrage with a sister token, no hard reservesA loss of confidence becomes a death spiral (Terra)
  

  Lawmakers have moved in. The European Union's Markets in Crypto-Assets regulation (MiCA) has applied in full since 30 December 2024, and the transition period for existing firms ended on 1 July 2026. In the United States, the GENIUS Act was signed into law on 18 July 2025. It requires payment stablecoins to be backed 100% by liquid assets such as dollars or short-term Treasuries, with monthly public disclosure of the reserves.

  Supporters say regulated stablecoins are simply better plumbing: dollars that move across borders in minutes, any hour, for low fees, usable by anyone with a phone. The central banks' own club, the Bank for International Settlements (BIS), is far cooler. Its 2025 annual report argued that stablecoins fail three tests of sound money: "singleness" (every dollar accepted at par everywhere), "elasticity" (supply that can expand when the economy needs it), and "integrity" (resistance to financial crime). It sees their future role, beyond being a gateway into crypto, as unclear.

  Central bank digital currencies
  Central banks have their own answer: a CBDC, digital cash issued directly by the central bank. A CBDC doesn't need a public blockchain; it can run on a ledger the central bank controls. The Atlantic Council's tracker, updated in May 2026, counted 146 countries and currency unions exploring one, up from 87 in May 2022. Only three have fully launched: the Bahamas, Jamaica and Nigeria, all with slow uptake. China's e-CNY is by far the largest pilot. By December 2025 it had processed more than 3.4 billion retail transactions worth about 16.7 trillion yuan. The European Central Bank is preparing a pilot of a digital euro in 2027 and aims to be ready for a first issuance in 2029, if the EU passes the necessary law in 2026.

  The United States has gone the other way. An executive order of 23 January 2025 set a policy of prohibiting the establishment, issuance, circulation and use of a CBDC in the United States. Defenders of CBDCs argue that as cash use falls, citizens should still have access to public money that isn't a claim on a private company. Opponents, from privacy campaigners to many US lawmakers, warn that a central bank ledger of every payment would give governments unprecedented power to watch and control spending. It's a real values disagreement, not a technical one, and different democracies are landing in different places.

  Part 8 — Energy, real uses, hype, and the state of things in 2026

  Energy
  Proof of work is deliberately costly, and the cost is electricity. The Cambridge Centre for Alternative Finance runs the most widely cited estimate, the Cambridge Bitcoin Electricity Consumption Index. On 31 October 2026 its best guess for Bitcoin's power demand was 19.74 gigawatts, equal to about 173 terawatt-hours a year, within a plausible range of roughly 89 to 333 TWh. On the index's own comparison page, that's about 0.68% of the world's electricity consumption, using 2022 world totals, and sits between the United Arab Emirates and Egypt in a country ranking.

  
    NetworkAnnual electricity (estimate)Source
    Bitcoin, proof of workabout 173 TWh (range 89–333)Cambridge CBECI, 31 Oct 2026
    Ethereum, proof of stakeabout 0.0026 TWhCCRI, cited by ethereum.org
  

  Defenders make three arguments. Miners chase the cheapest power, which often means energy that would otherwise be wasted (flared gas, remote hydro, curtailed wind). Mining can switch off instantly, so it can act as a flexible load that helps balance grids. And the energy is the security: it's what makes rewriting history ruinously expensive. Critics answer that much mining still runs on fossil fuels, that specialised chips become electronic waste within a few years, and that the energy buys a payment system that handles far fewer transactions than card networks. Even the Cambridge team warns that comparisons like these mix apples and oranges. What's not in dispute is the design: Bitcoin ties its security to energy spent, and changing that would mean changing what Bitcoin is.

  Real uses and hype
  For several years "blockchain" was pitched for nearly everything: shipping, voting, medical records, diamonds, food. Most of those projects quietly ended. The best-known example is TradeLens, a blockchain shipping platform built by the shipping giant Maersk and IBM. When they shut it down in November 2022, Maersk said that "the need for full global industry collaboration has not been achieved." The lesson most engineers drew: if a group of known companies already trust each other enough to share a database, an ordinary shared database is cheaper and faster. A blockchain only earns its cost when participants won't trust any single operator.

  
    QuestionIf yes
    Do many parties need to write to the same record?Keep going. If not, use a normal database.
    Do they distrust each other, or any single operator?Keep going. If there's a trusted party, let it run the database.
    Must the record resist censorship or reversal by any one authority?A public blockchain may be worth its cost.
    Can you live with slow updates, fees, and mistakes that can't be undone?Then it might genuinely be the right tool.
  

  The uses that have survived mostly fit that test. Bitcoin as a scarce digital asset that no government issues. Stablecoins as dollars that cross borders without a correspondent bank. Ethereum as a platform for tokens and lending markets that run without a company in the middle. Even the BIS, while rejecting stablecoins as the core of the monetary system, now argues for a "tokenised unified ledger" for central bank money, bank deposits and government bonds. That's a programmable shared ledger run by trusted institutions, which tells you the idea has outlived the hype even among its critics.

  The state of affairs, as of late October 2026
  
    Mainstream finance plugged in. On 10 January 2024 the US Securities and Exchange Commission approved exchange-traded products holding bitcoin itself. Its chair at the time, Gary Gensler, stressed that the SEC was not endorsing bitcoin, calling it "primarily a speculative, volatile asset."
    Governments split. El Salvador made bitcoin legal tender in 2021. After a loan agreement with the IMF, its Bitcoin law was amended in early 2025 so businesses no longer have to accept it. In March 2025 a US executive order created a "Strategic Bitcoin Reserve" from bitcoin the government already held through forfeitures.
    Prices are still wild. On 31 October 2026 bitcoin was around $77,400, about 39% below its October 2025 record, with a market value near $1.56 trillion. Ether was near $2,270, under half its August 2025 peak.
    The networks keep running. Bitcoin passed block 974,000 with no central operator. Ethereum has run on proof of stake for more than four years.
    Open questions remain. Will fees keep Bitcoin secure once block rewards dwindle? Will quantum computers, as the Cryptography piece explained, eventually threaten today's signature schemes, and can these networks agree to upgrade in time? Will stablecoins make payments cheaper or simply move risk out of the banking system? Will CBDCs protect public money or erode privacy?
  

  Here are both cases at their strongest. The enthusiast's: for the first time, a group of strangers can hold and move value with no one able to freeze, debase or censor it, and that matters most to people whose governments and banks have failed them. The sceptic's: a decade and a half in, the main uses are speculation, a few niches and a great deal of fraud, at a large energy cost, while ordinary payment systems have kept getting faster and cheaper. You don't have to pick a side to understand the machine. Now you do.

  
    The takeaway
    A blockchain is a public, append-only ledger that strangers can agree on without a trusted keeper. Hashes make history tamper-evident, proof of work (or proof of stake) makes rewriting it expensive, and signatures make sure only key-holders can move coins. Bitcoin combined ideas from Chaum, Haber and Stornetta, Dwork and Naor, Back, Dai, Szabo and Finney, and added payment for the people who run it. Ethereum made the ledger programmable and in 2022 cut its energy use by about 99.95% by switching to proof of stake. The crashes came mostly from middlemen holding other people's keys, plus one stablecoin backed by nothing. Stablecoins are now regulated dollars on blockchains, CBDCs are the central banks' rival, and Bitcoin's electricity use is still about 0.7% of the world's. The technology works. Whether it's worth what it costs depends on how much you need to avoid trusting anyone.
    Try this: in the chain demo above, change one word in block 1 and count how many Mine presses it takes to make the chain green again. Then remember the real chain has more than 974,000 blocks and the world's largest computing network guarding the next one. That gap is the security.
  

  
    Sources & Further Reading
    Sources. Satoshi Nakamoto, Bitcoin: A Peer-to-Peer Electronic Cash System (2008), bitcoin.org, quotations from sections 2, 4 and 6 and the reference list. Genesis block data, header fields and block 974,063 difficulty: Blockstream block explorer API, read 1 November 2026 (the genesis hash is recomputed live above). Precursors: Hashcash.org (Back, 1997; 2002 paper); Wei Dai, "b-money" (1998); Narayanan and Clark, "Bitcoin's Academic Pedigree," Communications of the ACM (2017). Ethereum dates, the DAO fork, the Merge, staking and energy: ethereum.org (history, Merge, staking and energy pages). Bitcoin electricity: Cambridge Centre for Alternative Finance, CBECI and its comparisons page, read 31 October 2026. Prices: FRED series CBBTCUSD (Coinbase daily close) and CoinGecko, read 1 November 2026. Stablecoin supply: DefiLlama, 1 November 2026. Stablecoin critique: BIS, Annual Economic Report 2025, chapter III, and press release of 24 June 2025. CBDCs: Atlantic Council CBDC Tracker (May 2026); ECB digital euro pages; White House executive orders of 23 January 2025 and 6 March 2025. GENIUS Act: White House fact sheet, 18 July 2025. MiCA: ESMA. ETF approval: SEC Chair Gensler's statement of 10 January 2024. Fraud losses: FBI, 2024 Internet Crime Report press release (23 April 2025). TradeLens: Maersk press release, 29 November 2022. Mt. Gox, Terra, FTX, Do Kwon, El Salvador and the Wright judgment: court records and news reports as summarised in the relevant Wikipedia articles, used as leads; figures marked as estimates where sources differ.
    
      🌐 The Bitcoin white paper. Nine pages, readable in an evening; sections 1–6 are the whole idea.
      🌐 The Ethereum white paper. Buterin's 2014 case for a programmable blockchain, with a note on what has changed since.
      📘 Bitcoin and Cryptocurrency Technologies, Arvind Narayanan and colleagues (Princeton University Press, 2016). The best rigorous textbook; a draft and lectures are at bitcoinbook.cs.princeton.edu.
      🌐 Cambridge Bitcoin Electricity Consumption Index. Live estimates with their uncertainty ranges, and a candid page on the limits of comparisons.
      📘 The Bitcoin Standard, Saifedean Ammous (2018). The enthusiast's case at full strength: bitcoin as hard money. Read it critically.
      📘 Number Go Up, Zeke Faux (2023). A reporter's tour of the 2021–22 mania and its collapses. The sceptic's case, with names.
      🌐 Web3 Is Going Just Great, Molly White. A running, sourced timeline of hacks, scams and failures.
      🌐 Companion pieces in this library: How Money Was Invented, Cryptography: The Math That Keeps Your Secrets, How to Not Get Scammed and Investing Without Gambling.


The Internet: From ARPANET to the Network in Your Pocket
--------------------------------------------------------
How a handful of researchers built a network that nobody owns, and how it ended up carrying three-quarters of humanity.

Follows: The Silk Roads: The World's First Internet Ran on Camels and Graph Theory: From the Bridges of Königsberg to Your GPS
In this deep dive
1. The idea before the machine: Licklider, Baran, Davies
2. ARPANET, 1969: "lo"
3. A network of networks: TCP/IP, DNS and the 1983 switch
4. The web and the browser
5. Boom, bust, broadband and the phone
6. How a packet actually travels
7. Platforms and social media
8. The state of affairs in 2026, and the open questions

At 22:30 on 29 October 1969, a graduate student at UCLA named Charley Kline sat down to log in to a computer at the Stanford Research Institute, about 500 kilometres up the California coast. He typed an L. The other end confirmed it. He typed an O. Confirmed. He typed a G, and the system crashed.
So the first message ever sent over the network that became the internet was "lo", as in "lo and behold." They got the full login working about an hour later. The professor supervising, Leonard Kleinrock, still keeps a scan of the handwritten log on his website.
Today about 6 billion people use the descendant of that link. The Silk Roads showed how a web of relays, where almost nothing travelled end to end, could carry religions and inventions across a continent. Graph Theory showed how throwing away everything but "what connects to what" lets you reason about any network. The internet is both ideas made literal: a relay system of dots and lines, where your message is handed from machine to machine and nobody runs the whole thing. This deep dive is the history of how it was built, a plain explanation of how it works, and an honest look at where it stands now.
Part 1 — The idea before the machine: Licklider, Baran, Davies
In 1960 most computers were room-sized machines that worked through stacks of punched cards. You handed in your job and came back hours later. A psychologist named J. C. R. Licklider thought that was the wrong relationship. In a paper that year, "Man-Computer Symbiosis," he imagined people and computers working together in real time, each doing what it was best at.
In August 1962 he went further. In a series of memos he described a "Galactic Network": computers all over the world linked so that anyone could reach data and programs at any site. That October he became the first head of the computer research programme at ARPA, the Pentagon's research agency, and in an April 1963 memo he addressed his far-flung research contractors as "Members and Affiliates of the Intergalactic Computer Network." He didn't build the network himself. What he did was convince the people who came after him at ARPA, including Ivan Sutherland, Bob Taylor and Larry Roberts, that it mattered. In 1968 he and Taylor published "The Computer as a Communication Device," predicting online communities of shared interest that would form regardless of where their members lived.

Licklider's real invention wasn't a machine. It was a reason to build one: computers as a way for people to talk to each other.

The problem with phone lines
The obvious way to connect computers was the telephone network. But the phone system is circuit-switched: when you call someone, a continuous line is reserved for you for the whole conversation, even during the silences. Computer traffic is bursty, a flurry of data and then nothing, so reserving a whole line wastes most of it. And a network built around a few big switching centres has a few big points of failure.
Two men, working separately on different sides of the Atlantic, came up with the same answer.
Paul Baran was an engineer at the RAND Corporation in California. In the early 1960s, with the Cuban missile crisis making nuclear war feel close, the question he worked on was how a military command network could survive an attack that destroyed many of its nodes. His answer, laid out in RAND studies from 1960 to 1962 and then in On Distributed Communications (1964), had two parts. First, build a mesh with many redundant links, not a hub with spokes. Second, chop each message into small "message blocks," send each one separately, let every node pass blocks along toward their destination by whatever route was free (he called it "hot-potato routing"), and reassemble them at the end.
Donald Davies at Britain's National Physical Laboratory (NPL) arrived at almost the same design in 1965, with a different goal: a commercial national data network for Britain. He called the pieces packets, and that's the word that stuck. His team built a working packet network inside the NPL around the end of the decade.

Circuit switching (telephone)
Packet switching (internet)

What's reserved
A whole line, for the whole call
Nothing; packets share every link

If a link fails
The call drops
Packets go around it

Good at
Steady streams, like voice
Bursty data, like computers

Cost of idle time
Paid for anyway
Free; someone else's packets fill the gap

Contested credit, and a myth
Who invented packet switching? The honest answer is "several people, independently." The Internet Society's history, co-written by many of the pioneers, notes that the work at MIT, RAND and the NPL "proceeded in parallel without any of the researchers knowing about the other work." That same history, which Kleinrock co-wrote, credits him with the first paper on packet switching theory, in July 1961. Others disagree. The Computer History Museum's timeline describes his thesis as "a seminal text on queuing theory" and calls Davies "the English inventor of packet switching." Read that way, Kleinrock analysed how messages queue in a network, while Baran and Davies designed the chopping-up. It's a live argument among historians, and you'll see both versions in print.
The myth is cleaner to correct. You'll often read that the internet was built to survive a nuclear war. Baran's work was about exactly that, but ARPANET was a different project. The Internet Society's history says flatly that the claim "was never true of the ARPANET, only the unrelated RAND study." RAND's own account says ARPANET was "intended for scientists and researchers who wanted to share computers remotely." Computers were extremely expensive, and ARPA was paying for a lot of them. Letting researchers use each other's machines over a network was cheaper than buying everyone their own.
Part 2 — ARPANET, 1969: "lo"
In 1966 Bob Taylor, by then running ARPA's computing office, got money to link the computers of all the research groups ARPA funded, and persuaded Larry Roberts to come from MIT to lead the project. At a conference in 1967, Roberts heard Roger Scantlebury of Davies's team present the NPL design, learned from him about Baran's work too, and adopted the word "packet." After that meeting the plan's line speed went up from 2.4 to 50 kilobits per second.
The key design choice was not to make the big host computers do the networking themselves. Instead, each site would get a small dedicated computer, the Interface Message Processor (IMP), whose only job was to pass packets on. A team at the firm Bolt Beranek and Newman (BBN), led by Frank Heart, won the contract to build them in December 1968, with Bob Kahn playing a major role in the design. IMPs were the ancestors of today's routers.

Date
Milestone

7 April 1969
Steve Crocker at UCLA writes RFC 1, the first "Request for Comments"

September 1969
The first IMP is installed at UCLA

29 October 1969, 22:30
First message, UCLA to SRI: "lo"

End of 1969
Four sites connected: UCLA, SRI, UC Santa Barbara, University of Utah

December 1970
The first host-to-host protocol (NCP) is finished

1972
First public demonstration, at the International Computer Communication Conference; Ray Tomlinson's email software spreads

That first RFC deserves a moment. Crocker set up the series as an informal, fast way for network researchers to share ideas, because academic journals were far too slow. The habit stuck. Proposals were published openly, anyone could argue with them, and what worked became standard. The RFC series is still how internet standards are written. On 31 October 2026 the RFC Editor's index had passed number 10,000.
Email was the surprise. ARPANET was built to share computers, but the application people actually loved was sending each other messages. The Internet Society's history dates Ray Tomlinson's email software to March 1972. Other accounts put his first network email in 1971. Either way, Tomlinson is the reason addresses have an @ in them: he used it to separate a user's name from their machine's. The Internet Society's authors call email the network's first "hot" application. People wanted to talk to each other, just as Licklider had said.

ARPANET was built so researchers could share expensive computers. Within a few years, its most popular use was gossip.

Part 3 — A network of networks: TCP/IP, DNS and the 1983 switch
By the early 1970s ARPANET was not the only packet network. ARPA itself was experimenting with packet radio (networks over the airwaves, for vehicles and soldiers) and packet satellite links. France had CYCLADES, designed by Louis Pouzin, first demonstrated in November 1973. Pouzin coined the word datagram for a packet sent on its own with no guarantee it would arrive, and the word catenet for networks chained together through gateways. Each network worked differently. The question became: how do you connect networks that know nothing about each other?
Bob Kahn, by then at ARPA, and Vint Cerf, then at Stanford, worked on the answer from the spring of 1973 and published it in May 1974 as "A Protocol for Packet Network Intercommunication." Kahn's ground rules, in the Internet Society's paraphrase, are still the internet's philosophy:

Each network stands on its own. Nothing inside it has to change to connect.
Delivery is best effort. If a packet doesn't arrive, the sender sends it again.
The boxes connecting networks (gateways, later called routers) stay simple. They don't remember the packets they pass.
There's no global control at the operations level.

That design pushed intelligence to the edges. The network in the middle just moves packets as well as it can. The computers at each end check that everything arrived, put it back in order and ask for anything missing. Over the next few years the protocol was split in two: IP (Internet Protocol) gives every machine an address and gets packets roughly to the right place, and TCP (Transmission Control Protocol) runs on the end computers and makes the conversation reliable. An RFC from December 1974 already used the short word in its title: "Specification of Internet Transmission Control Program." The final specifications, RFC 791 for IP and RFC 793 for TCP, appeared in September 1981.
Flag day
TCP/IP became a US defence standard in 1980. In November 1981 Jon Postel published RFC 801, a plan for every ARPANET host to switch from the old protocol to TCP/IP. The deadline was 1 January 1983, a "flag day": every host had to convert at once, or be left talking to the network through makeshift workarounds. It went more smoothly than people feared, and participants handed out buttons reading "I survived the TCP/IP transition." The switch also let the military split its operational traffic off into a separate network, MILNET, leaving ARPANET for research.
1 January 1983 is often called the internet's real birthday, because it's the day the network of networks started running on one common language.
Names instead of numbers
Every machine on the internet has a numeric address. People are bad at numbers, so early on every computer kept a single shared file listing every host's name and address. That stopped working once thousands of independently run networks were joining. Paul Mockapetris at the University of Southern California designed the Domain Name System (DNS), described in RFCs 882 and 883 in November 1983 and revised in RFCs 1034 and 1035 in November 1987. DNS spreads the job out. No one computer knows every name. Instead, servers for ".org" know who's responsible for each ".org" name, those servers know about the names under them, and so on. It's a distributed phone book with no central copy, which is why it scaled.
From research to everybody
In the mid-1980s the US National Science Foundation built NSFNET to connect universities and supercomputer centres, spending about $200 million between 1986 and 1995. It became the internet's backbone. In 1990 the original ARPANET was switched off. The NSF lifted restrictions on commercial use in 1991, and in April 1995 it stopped funding the backbone altogether, handing the job to private long-haul networks. The routing protocol that lets those independent networks tell each other which addresses they can reach, the Border Gateway Protocol (BGP), was first published in June 1989 and, in later versions, still holds the internet together.
Contested credit, again
In a CNN interview on 9 March 1999, Vice President Al Gore said: "During my service in the United States Congress, I took the initiative in creating the Internet." It was widely mocked and misquoted as "I invented the internet." He didn't say that. He did, as a senator, push the 1991 legislation that the Computer History Museum's timeline calls the "Gore Bill," which created a National Research and Education Network programme. Whether "took the initiative in creating" was fair is a reasonable thing to argue about. The engineers who built TCP/IP were Cerf and Kahn and many others.
Part 4 — The web and the browser
Here's the most common confusion in this whole subject: the internet and the web are not the same thing. The internet is the network: the cables, routers and protocols that move packets. The web is one service that runs on it, like email or video calls.

The internet
The World Wide Web

What it is
A network of networks that moves packets
A system of linked documents and apps

Main protocols
IP, TCP, DNS, BGP
HTTP, HTML, URLs

Born
1969 (ARPANET), 1983 (TCP/IP switch)
1989 proposal, 1990–91 first software

Analogy
The roads
One kind of traffic on them

The web was invented by Tim Berners-Lee, a British software engineer at CERN, the European particle-physics laboratory near Geneva. While consulting at CERN in 1980 he had written a personal program called ENQUIRE for linking notes together. In March 1989 he wrote a proposal titled "Information Management: A Proposal" for a hypertext system to help CERN's thousands of scientists find each other's documents. His boss, Mike Sendall, approved the purchase of a NeXT computer in September 1990 and let him build it.
By Christmas 1990 Berners-Lee had three inventions working together:

HTML, a simple way to mark up a document with headings and links.
HTTP, the rules a browser uses to ask a server for a document.
URLs, an address format that can point to any document on any server.

He also wrote the first browser, called WorldWideWeb, and the first web server. The first website, at info.cern.ch, explained the project itself, and CERN still keeps a copy online. In August 1991 he posted the software publicly on Usenet newsgroups. On 30 April 1993 CERN's directors signed a statement that the web's technology could be used by anyone, with no fees payable to CERN. The W3C's history calls it "a milestone document." Timing mattered: that February, the University of Minnesota had announced it would start charging licence fees for Gopher, a competing system, and many users moved to the web.
The browser that changed everything
The early web was text-heavy and ran mostly on university machines. In February 1993 the National Center for Supercomputing Applications at the University of Illinois released the first alpha version of Mosaic, a graphical browser by Marc Andreessen and colleagues. Web traffic, which had been 0.1% of the NSF backbone in March 1993, reached 1% by September. In 1994 Andreessen and others left to start the company that became Netscape. Berners-Lee founded the World Wide Web Consortium (W3C) to steward the web's standards; it was founded on 1 October 1994.
On 9 August 1995, Netscape went public, only sixteen months after the company was formed. Its shares were priced at $28, traded as high as $75 on the first day, and closed at $58.25, valuing the company at about $2.9 billion. Many people date the dot-com boom from that day.

The web won because it was given away. A royalty-free standard beat systems that charged.

Part 5 — Boom, bust, broadband and the phone
The late 1990s were a gold rush. Investors poured money into any company with ".com" in its plan, many of them selling things online at a loss to grab market share. The technology-heavy Nasdaq Composite index peaked at 5,048.62 on 10 March 2000. By 9 October 2002 it had fallen to 1,114.11, a drop of 78%. Many companies failed, among them telecom firms such as WorldCom and Global Crossing.
The bust taught a lesson the crypto piece in this series repeats: a technology can be real and still be wildly overpriced. The fibre laid in the boom didn't vanish; it stayed in the ground for whoever came next. Google, incorporated in 1998, grew through the bust.
Broadband, then the phone
In the early years most homes reached the internet by dial-up modem, the screech of a computer calling another computer over a phone line. Broadband replaced it, first over the copper of phone lines (DSL) and cable-TV lines, then increasingly over fibre to the home. The change was more than speed. Always-on connections made it normal to leave the internet running, which is what made streaming video, online gaming and constant messaging possible.
Then the internet left the desk. On 9 January 2007 Apple introduced the iPhone, describing it as a phone, an iPod and "a breakthrough Internet communications device" in one. Smartphones running iOS and Android put a full browser, maps, a camera and thousands of apps in a pocket. The ITU calls mobile phones "the most common gateway to the Internet," and in many low- and middle-income countries a phone is the main way people get online.
The International Telecommunication Union (ITU), the UN agency for telecoms, published the latest global figures in October 2025:

Indicator, 2025
Figure

Mobile-cellular subscriptions
9.2 billion (112 per 100 people)

Share of mobile subscriptions that are broadband
89% (under 50% in 2015)

People aged 10+ who own a mobile phone
82%

Population covered by 3G or better
96%

Population covered by 4G
93%

Population covered by 5G
55% (84% in high-income countries, 4% in low-income)

Fixed broadband traffic
7.3 zettabytes

Mobile broadband traffic
1.5 zettabytes

That last pair surprises people. Even in the age of the phone, fixed lines carry about five times as much data as mobile networks, partly because phones at home switch to Wi-Fi, which runs over fixed broadband.
Under the sea
"Wireless" is mostly an illusion. As the research firm TeleGeography puts it, your phone's signal only travels by radio to the nearest cell tower. From there, it goes into fibre. Between continents, that fibre lies on the seabed. As of 2026 TeleGeography tracks more than 600 active and planned submarine cables, and as of early 2026 it estimated more than 1.5 million kilometres of them in service. They range from the 131-kilometre CeltixConnect between Ireland and the United Kingdom to the 20,000-kilometre Asia America Gateway. Satellites carry only a sliver of international traffic: figures from the US Federal Communications Commission put them at 0.37% of US international capacity. Cables used to be built by consortia of phone companies. Today the biggest investors include content companies such as Google, Meta, Microsoft and Amazon, which now lay their own.
Part 6 — How a packet actually travels
Here's what happens, step by step, when you tap a link in London to a news site whose server is in New York. It takes a fraction of a second.

Layer
Its job
Examples

Application
What you're actually doing
HTTP (web), DNS (names), email protocols

Transport
A reliable conversation between two programs
TCP (reliable), UDP (fast, no guarantees)

Internet
Addresses and routing between networks
IP (versions 4 and 6)

Link
Getting bits across one hop
Wi-Fi, Ethernet, 4G/5G

Physical
The actual medium
Copper, fibre, radio

Find the address. Your phone asks a DNS resolver (usually run by your provider or a public service) for the site's numeric IP address. If the answer isn't already cached, the resolver asks a root server which servers handle ".com," asks those which server handles the site's name, and asks that one for the address. Answers are cached, so most lookups take milliseconds.
Open a conversation. Your phone and the server exchange three short packets, a "handshake," to agree to talk (in TCP: SYN, SYN-ACK, ACK). Newer protocols fold this step into the next one to save time.
Agree on a secret. For an https address, your phone and the server run a TLS handshake, using the public-key maths from Cryptography: The Math That Keeps Your Secrets, so that everything after this is encrypted. Anyone in between can see which server you're talking to, but not what you say.
Ask. Your browser sends an HTTP request: "GET this page." The request, like everything else, is chopped into packets. Each packet carries a source address, a destination address and a sequence number, like an envelope with a page number.
Leave the house. Your packets go by Wi-Fi to your home router, then over your provider's line into its network.
Hop, hop, hop. Each router reads only the destination address, checks its routing table and passes the packet to the next router that's closer. Between independent networks, those tables are built by BGP, in which each network announces which addresses it can reach. This is the shortest-path problem from Graph Theory, solved continuously by thousands of machines that each see only their neighbours.
Cross the ocean. Somewhere along the way, your packets enter a submarine cable. London to New York is about 5,570 km in a straight line over the globe. Light in glass fibre travels at about two-thirds of its speed in a vacuum, about 204,000 km per second, so even a perfectly straight cable needs about 27 milliseconds one way, 55 there and back. Real routes are longer. Physics sets a floor no engineering can beat.
Arrive in pieces. Packets may take different routes and arrive out of order, and a few may be lost. The server's TCP software puts them back in order and asks for anything missing. Each packet also carries a "time to live" counter that every router lowers by one, so a packet lost in a loop eventually dies instead of circling forever.
Answer, many times over. The server sends back the page, and your browser discovers it needs dozens more files (images, styles, scripts, ads), often from other servers. Popular sites keep copies on servers near you, in networks called content delivery networks, so many of those files never cross the ocean at all.

Nobody plans your packet's route. Each router makes one local decision, and the global path emerges from millions of them.

Two numbers explain a lot about today's internet. The original addresses (IPv4) are 32 bits long, which allows about 4.3 billion of them, fewer than there are people. The last free blocks were handed out by the central registry on 31 January 2011, prompting the regional registries to announce on 3 February that the free pool was gone. The replacement, IPv6, uses 128-bit addresses, which allows about 340 undecillion (3.4 × 1038). The world has been slowly moving to it ever since, and both still run side by side.
Part 7 — Platforms and social media
The early web was mostly pages that their owners published and everyone else read. In the 2000s the pattern flipped: sites became platforms where users supplied most of the content. Facebook launched in February 2004, YouTube in 2005. Add the smartphone in 2007 and its app stores, and the web became something you carry, check dozens of times a day, and contribute to without knowing any HTML.
The scale is hard to overstate. Meta reported that an average of 3.64 billion people used at least one of its apps (Facebook, Instagram, WhatsApp, Messenger) every day in September 2026. Its revenue for July to September 2026 was $64.6 billion, almost all from advertising. That's the business model of most of the consumer internet: services are free to use because your attention is sold to advertisers, and the more precisely a platform can predict what will hold your attention, the more that attention is worth.
The legal bargain
In the United States, a short clause passed in 1996 shaped all of this. Section 230 of the Communications Act, added on 8 February 1996, says that no provider of "an interactive computer service shall be treated as the publisher or speaker" of what its users post. In plain terms: if someone posts something defamatory, you sue the person who posted it, not the site. Defenders say this is what let the participatory web exist at all; no platform could host billions of posts if it were legally responsible for each one as a newspaper is for its articles. Critics say it lets the largest companies in history profit from harmful content with little accountability. Both sides have proposed changes for years.
The European Union has taken a more regulatory path. Its Digital Services Act requires platforms to label ads and say who placed them and why you're seeing them, stop targeting ads using sensitive personal data such as religion or sexual orientation, and avoid manipulative "dark pattern" designs. The largest platforms must keep public repositories of the ads they show.
The argument about effects
Few questions about the internet are argued more fiercely than what social media does to people. Here are the strongest versions of each side.

The optimistic case: social platforms let anyone publish, organise and find their people. They've given voice to groups the old media ignored, helped families stay close across continents, and let small businesses and artists reach audiences directly.
The worried case: engagement-maximising feeds reward outrage and misinformation, harvest personal data on an industrial scale, and may harm the mental health of some teenagers. Legislators in many countries, and in the US a bipartisan group of state attorneys general (with whom Meta announced an agreement on teen users in August 2026), now focus on platforms' design choices rather than just their content.
The research reality: studies of social media and wellbeing have produced mixed results. Some find small average associations, others larger effects for particular groups or particular kinds of use. Researchers disagree, sometimes sharply, about how much is cause and how much is correlation. Treat confident claims in either direction with care, and see How to Read the News Without Losing Your Mind in this library.

Part 8 — The state of affairs in 2026, and the open questions
Who's online, and who isn't
The ITU's most recent figures, for 2025 (its 2026 edition was not yet public at the end of October 2026), are the best global snapshot:

Indicator, 2025
Figure

People using the internet
about 6.0 billion, 74% of the world

People offline
about 2.2 billion

High-income countries online
94%

Low-income countries online
23%

Africa online
36%

Urban vs rural
85% vs 58%

Men vs women online
77% vs 71%, about 280 million more men

15–24-year-olds online
82%

Two facts stand out. First, the coverage gap is small and the usage gap is big. Since 96% of people live within reach of at least a 3G signal, most of the 2.2 billion offline people could technically connect. What stops them, the ITU's figures suggest, is mostly cost, skills and access to a device. Second, progress is slowing. The ITU notes that the gender gap hasn't narrowed at all since 2019 (its parity score was 0.92 in both years), and in low-income countries only one in seven rural people is online. Prices are falling: the median cost of a basic mobile data plan fell to 1.4% of average income in 2025, but a fixed broadband subscription in a low-income country still costs more than a quarter of average income.
Satellites come back
Satellites used to be a last resort, because traditional communication satellites sit in geostationary orbit about 35,786 km up. A request and its reply must travel up and down twice, so even at the speed of light a round trip takes about 477 milliseconds before any processing, enough delay to make video calls awkward.
The new constellations fly much lower. SpaceX's Starlink satellites orbit below 600 km, cutting that physics delay to a few milliseconds, at the price of needing thousands of satellites because each one only sees a small patch of Earth for a few minutes. In October 2026 SpaceX said it operated more than 11,000 Starlink satellites and planned for up to 100,000. The astronomer Jonathan McDowell, who tracks every launch, counted 11,190 Starlink satellites in orbit on 29 October 2026. SpaceX's most recent customer figure I could verify was "more than 6 million" active customers, from July 2025. Since January 2024 some Starlink satellites have also connected directly to ordinary phones. Rival constellations are being launched, among them Amazon's Kuiper and China's Qianfan and Guowang (also called Xingwang).
Supporters point to coverage for ships, aircraft, disaster zones and remote villages where cables will never be economic. Critics raise the brightness of satellites in astronomers' images, collision and debris risk in crowded orbits, and the concentration of a critical piece of infrastructure in one private company. SpaceX itself publishes work on brightness mitigation and collision avoidance, which shows the concerns are taken seriously on both sides.
Net neutrality
Net neutrality is the principle that internet providers should treat all traffic equally: not block, slow down or charge extra for particular sites or apps. The European Union made it law in Regulation (EU) 2015/2120 of 25 November 2015, whose Article 3(3) requires providers to "treat all traffic equally," with exceptions for legal orders, network security and genuine congestion.
In the United States the rules have see-sawed with each administration. The Federal Communications Commission adopted strong rules in 2015, repealed them in 2017, and restored them in 2024. In January 2025 the federal appeals court for the Sixth Circuit struck the 2024 rules down, ruling that the FCC lacked the authority to classify internet access as a regulated telecommunications service. Since that ruling, net neutrality has not been required at the federal level.
Supporters argue that without neutrality, providers could become gatekeepers, favouring their own services and squeezing start-ups that can't pay. Opponents argue that heavy rules discourage investment in networks, that markets and competition law handle abuses better, and that some traffic (a remote surgery feed, say) genuinely deserves priority. Both sides claim to be defending the open internet.
One internet, or several?
The internet was designed to have no centre and no borders. Governments have been adding borders. China has long filtered foreign sites and services through what's often called the Great Firewall. Russia passed a set of amendments in 2019, informally the "sovereign internet" law, that mandate surveillance of internet traffic and give the state powers to cut the country's segment off from the rest of the internet, including a national version of DNS.
The bluntest tool is simply switching the internet off. The digital rights group Access Now and its partners documented at least 313 internet shutdowns in 52 countries in 2025, which they call the highest number since 2016. Governments that impose them usually cite security, exams or unrest. Critics see them as a way to silence protest and hide abuses.
Supporters of national control argue that states have a duty to protect citizens from crime, foreign interference and harmful content, and that a borderless internet mostly benefits the countries that host the biggest platforms. Defenders of the open internet argue that splitting it would damage trade, science and free expression, and hand governments too much power over what people can say and see. This "splinternet" debate is one of the defining arguments about the internet's future.
Open questions

Can the last 2.2 billion get online? Coverage is nearly there. Affordability, devices, skills and electricity are not.
Will satellites close the gap or just serve those who can pay? A Starlink terminal and subscription cost far more than a basic mobile plan in most low-income countries.
Who governs a network nobody owns? Standards still come from open groups like the IETF that publish RFCs. Laws come from nearly 200 governments, each with its own idea of what the internet should be.
What happens to the web as AI systems answer questions directly? If fewer people click through to the sites that create information, the advertising bargain that pays for much of the web may weaken. That's covered elsewhere in this series.
How resilient are the cables? Undersea cables are routed around fishing grounds and anchorages, and the industry spends effort educating other marine industries about where they lie, because damage from ships is a real risk. A few cables carry a lot of the world's traffic.

The takeaway
The internet grew from Licklider's idea that computers could be a way for people to talk, Baran's and Davies's independent invention of packet switching, and ARPANET's first garbled "lo" in 1969. It became a network of networks on 1 January 1983, when every host switched to Cerf and Kahn's TCP/IP, a design that keeps the middle simple and puts intelligence at the edges. DNS gave machines names. Berners-Lee's royalty-free web, and the Mosaic browser, made it usable by anyone. A bubble burst, but the fibre stayed, and the smartphone put it in our pockets. Today about 6 billion people are online, carried mostly by undersea glass and by routers that each make one local decision at a time. The big questions now are human ones: who's still left out, who gets to set the rules, and whether one internet survives.
Try this: open a terminal on a computer (Command Prompt on Windows, Terminal on a Mac) and type tracert example.com on Windows or traceroute example.com on a Mac. You'll see every router your packets pass through, with the time each hop takes. Then try a site you know is hosted on another continent and watch the time jump when your packets cross the ocean. That jump is mostly the speed of light.

Sources & Further Reading
Sources
- Barry Leiner, Vint Cerf, David Clark, Robert Kahn, Leonard Kleinrock and others, A Brief History of the Internet, Internet Society (Licklider's 1962 memos and ARPA role, parallel packet-switching work, BBN and the IMPs, NCP, the 1972 demo and email, Kahn's ground rules, the 1974 paper, the 1983 flag day and MILNET, DNS, NSFNET funding and its 1995 defunding, ARPANET's 1990 decommissioning, and the footnote on the nuclear-war myth).
- Computer History Museum, Internet History timeline, 1960s–1990s (Licklider and IPTO, Taylor and Roberts, Davies and the word "packet," the line-speed change, NSF lifting commercial restrictions in 1991, the Gore Bill, Mosaic).
- RAND Corporation, "Paul Baran and the Origins of the Internet" (2018).
- Leonard Kleinrock, "The Day the Infant Internet Uttered its First Words," UCLA (the IMP log of 29 October 1969).
- RFC Editor: RFC 1 (7 April 1969), RFC 675 (December 1974), RFCs 791 and 793 (September 1981), RFC 801 (November 1981), RFC 882 (November 1983), RFC 1034 (November 1987), RFC 1105 (BGP, June 1989), and the RFC index of 31 October 2026.
- W3C, A Little History of the World Wide Web (ENQUIRE, the 1989 proposal, the NeXT, the 1991 release, Gopher fees, the 30 April 1993 declaration, Mosaic, backbone traffic shares, the founding of W3C). Apple press release, 9 January 2007. Google, "Our story."
- Nasdaq Composite closes: FRED series NASDAQCOM. Netscape IPO figures: contemporary reports as summarised in Wikipedia's Netscape article.
- ITU, Measuring Digital Development: Facts and Figures 2025 (October 2025): internet use, urban–rural and gender gaps, youth, coverage, subscriptions, traffic and affordability.
- TeleGeography, Submarine Cable Frequently Asked Questions (2026).
- Meta, Third Quarter 2026 Results (21 October 2026).
- 47 U.S.C. § 230, via Cornell Law School's Legal Information Institute. European Commission, Digital Services Act policy page. Regulation (EU) 2015/2120, EUR-Lex. Sixth Circuit opinion 25a0002p (January 2025), Ohio Telecom Association v. FCC.
- Number Resource Organization, "Free Pool of IPv4 Address Space Depleted" (3 February 2011).
- Starlink updates page (October 2026 constellation size; July 2025 customer count; January 2024 direct-to-cell texts); Jonathan McDowell, Jonathan's Space Report, Starlink statistics (29 October 2026).
- Access Now, Rising repression meets global resistance: internet shutdowns in 2025 (31 March 2026).
- CNN, Late Edition transcript, 9 March 1999 (Gore). Wikipedia articles on CYCLADES, Donald Davies, packet switching and the 2019 Russian internet laws, used as leads and checked against the sources above where possible.
- Fibre and satellite delays computed from the great-circle distance (London–New York about 5,570 km), a fibre refractive index of about 1.47, and a geostationary altitude of 35,786 km.
Go deeper
- 🌐 A Brief History of the Internet, Internet Society. Written by the people who built it. The best single source, with the bias that implies on questions of credit.
- 📘 Inventing the Internet, Janet Abbate (MIT Press, 1999). The historian's version: careful on contested credit, and on how users, not planners, decided what the network was for.
- 📘 Where Wizards Stay Up Late, Katie Hafner and Matthew Lyon (1996). The ARPANET story as a narrative, full of the people behind Part 2.
- 📘 Weaving the Web, Tim Berners-Lee (1999). The web's inventor on why he gave it away.
- 📘 Tubes: A Journey to the Center of the Internet, Andrew Blum (2012). A reporter visits the buildings, cables and landing stations. Makes Part 6 physical.
- 🌐 ITU Facts and Figures 2025. The official numbers on who's connected, updated every year.
- 🌐 TeleGeography Submarine Cable FAQs, with a link to its interactive cable map. Find the cable your packets probably use.
- 🌐 Companion pieces in this library: Cryptography: The Math That Keeps Your Secrets (what TLS does), Graph Theory (routing), Blockchain, Explained (Technology 3), and The 1990s in Decades of Leadership.


Flying Machines
---------------
For most of history, human flight was a legend and a broken leg. Then, in little more than a century, it became one of the safest ways to cross a continent. How that happened, how a wing really works, and where aviation stands at the end of 2026.

Follows: Everyday Physics: The Answers Hiding in Plain Sight (Physics 5) and The Physics of the Greenhouse Effect (Physics 15)
  

  On 21 November 1783, two Frenchmen rose over Paris standing on the gallery of a cloth-and-paper balloon, feeding a fire to keep it up. Twenty-five minutes and about nine kilometres later they came down safely: the first people known to have flown freely through the air and lived to describe it. A hundred and twenty years after that, two bicycle makers from Ohio flew a powered aeroplane for twelve seconds. In 2025, airlines carried just under five billion passengers on more than 38 million flights.

  This piece follows that arc, from people who jumped off towers to what the Wright brothers actually invented, how a wing really holds a 79-tonne airliner up, and the questions aviation faces now.

  
    In this deep dive
    
      The dream before the machine: balloons, Cayley and Lilienthal
      1903, and the long fight over who was first
      How a wing really makes lift (with a lift explorer)
      Wars, jets and the shrinking planet
      Why flying became so safe
      Rotors and robots: helicopters and drones
      Aviation in 2026, and the open questions
    
  

  Part 1 — The dream before the machine: balloons, Cayley and Lilienthal

  The oldest flying machines that really worked were toys. Kites may have been invented in China as far back as the 5th century BC; tradition credits the philosophers Mozi and Lu Ban. Chinese children were also playing with the "bamboo-copter", a stick with a little spinning rotor that climbs when you rub it between your palms and let go, from around 400 BC. Both contain a real principle of flight: air pushed one way pushes back.

  People were less lucky than toys. The best-known early attempt is that of Abbas ibn Firnas, a scholar in Córdoba in the 9th century, who is said to have covered himself in feathers, strapped on wings, jumped from a height and glided some distance before landing so hard that he injured his back. He blamed himself for forgetting a tail. It is a vivid story with a problem: it comes from a single source, the historian al-Maqqari, who wrote it down around seven centuries later. An English monk, Eilmer of Malmesbury, is recorded as gliding a short distance in the 11th century and breaking both legs. None of these stories describes controlled flight.

  Leonardo da Vinci studied the problem harder than anyone before him. His Codex on the Flight of Birds, written in 1505–06, holds more than 35,000 words and 500 sketches about bird flight, the nature of air and flying machines. In the early 1480s he had sketched an "aerial screw", a spiral meant to bore its way upwards, the first recorded design for vertical flight. There is no record that any of his flying machines flew. His aerial screw also had a flaw every helicopter designer would later meet: nothing stopped the machine itself from spinning the opposite way to its rotor.

  1783: the year people left the ground

  The breakthrough came from an unexpected direction: not wings, but buoyancy. Two brothers from a family of paper manufacturers, Joseph and Étienne Montgolfier, took their cue from something as ordinary as laundry drying over a fire and billowing upward. On 4 June 1783, at Annonay in southern France, they sent an unmanned balloon up in public; it flew for about ten minutes and came down some two kilometres away. On 19 September, at Versailles, a balloon carried a sheep, a duck and a rooster for about eight minutes, and all three landed safely. On 21 November, Jean-François Pilâtre de Rozier and the marquis d'Arlandes made that first crewed free flight over Paris. Ten days later, on 1 December, the physicist Jacques Charles rose about three kilometres in a balloon filled with hydrogen.

  Joseph Montgolfier thought the smoke contained a special gas with a property he called "levity". He was wrong about why it worked, as pioneers often are. The real reason is the one that makes a steel ship float (see Everyday Physics): an object in a fluid is pushed up by a force equal to the weight of the fluid it displaces. Heat air from 15 °C to 100 °C and a cubic metre of it gets about 0.28 kg lighter, so every cubic metre of hot air can lift about 0.28 kg of balloon and passenger. Hydrogen does about four times better, roughly 1.1 kg per cubic metre, which is why Charles's balloon could be smaller and go higher.

  A balloon floats; it doesn't fly. For the next 120 years the problem wasn't getting up. It was going where you meant to go.

  Balloons went where the wind went. To steer against the breeze, you'd need a machine heavier than air that held itself up by moving.

  Cayley, Lilienthal and the problem nobody saw

  The man who first wrote down what an aeroplane is was an English baronet, George Cayley (1773–1857). In 1799 he engraved a small silver disc, about an inch across and now held by London's Science Museum, with a design for a fixed-wing flying machine. The idea it captured was radical: stop trying to copy a flapping bird. Use one surface, a fixed wing, to produce lift; a separate system to push the machine forward; and a third set of surfaces, like a tail, to control it. He is credited with naming the four forces every pilot still learns: weight, lift, drag and thrust.

  Cayley flew a model glider with a modern layout in 1804 and published "On Aerial Navigation" in 1809–10. In 1853, at Brompton Dale in Yorkshire, a larger glider of his carried a person across a small valley. Popular accounts say it was his coachman, who promptly resigned. Historians aren't sure who it was: candidates include the coachman, a footman, a butler and his grandson. Wilbur Wright later said Cayley had taken the science of flight further than anyone before him.

  What Cayley lacked was a light engine. Inventors tried steam anyway. On 9 October 1890, Clément Ader's Éole made an uncontrolled hop of about 50 metres in France; in 1894 Hiram Maxim's huge test rig lifted briefly off its rails and was wrecked; on 6 May 1896, Samuel Langley's unpiloted Aerodrome No. 5 made a sustained flight. Each proved power could make lift. None showed a person could fly a machine.

  The flying man

  The person who did most to show it was a German engineer, Otto Lilienthal. Between 1891 and 1896, with his brother Gustav, he made more than 2,000 flights in gliders of his own design, the longest about 250 metres. His 1889 book, Birdflight as the Basis of Aviation, was widely read. And he was photographed: at least 145 pictures show him in the air, hanging under great curved wings. For readers around the world, they were proof that a human being could glide like a bird.

  On 9 August 1896 his glider stalled and fell about 15 metres. He died in Berlin the next day. The words on his gravestone, "Opfer müssen gebracht werden!" ("Sacrifices must be made!"), are often given as his last words, but the director of the Otto Lilienthal Museum doubts he ever said them.

  
    PioneerKey dateWhat they provedWhat was still missing
    Montgolfiers, Charles1783People could go up and come down aliveAny way to steer
    Cayley1799–1853The fixed-wing layout: lift, thrust and control kept separateA light engine; real control
    Ader, Maxim, Langley1890–1896Engines could produce enough liftA pilot in control
    Lilienthal1891–1896A person could glide again and againControl good enough to survive a gust
  

  Look at the last column. By the 1890s, almost everyone was working on lift and power. Very few were working on control: how a pilot keeps a machine level in gusty air, turns it, and gets out of trouble. Lilienthal's death came from exactly that gap. Two brothers in Dayton, Ohio, came to believe that control was the real problem.

  Part 2 — 1903, and the long fight over who was first

  Wilbur and Orville Wright ran a bicycle shop. On 30 May 1899 Wilbur wrote to the Smithsonian asking for whatever it had published on flight. What followed is best understood as three separate inventions.

  1. Control in all three directions. A plane can tip its nose up and down (pitch), roll its wings (roll), and swing its nose left and right (yaw). The Wrights controlled roll by twisting, or "warping", the tips of their wings in opposite directions, pitch with a movable surface at the front, and yaw with a movable rudder at the back. Their 1902 glider made properly controlled turns on 9 October 1902. They filed their patent on 23 March 1903. This three-axis control is still how every aeroplane is flown.

  2. Their own data. Their 1901 glider produced only about a third of the lift they had calculated. They traced much of the error to a long-accepted coefficient used in everyone's lift formulas, built a small wind tunnel in their shop in the autumn of 1901, and tested about 200 model wing shapes, 38 of them in detail. From then on they designed with their own measurements.

  3. Propellers as spinning wings. Nobody had a theory of propellers. The brothers realised a propeller blade is a wing moving in a circle and designed theirs that way. With a 12-horsepower engine built with the help of their shop mechanic, Charles Taylor, the 1903 Flyer had two propellers behind the wings, driven by chains.

  On 17 December 1903, at Kill Devil Hills near Kitty Hawk, North Carolina, Orville lifted off at 10:35 a.m. and flew 36 metres (120 feet) in 12 seconds. The brothers took turns. On the fourth flight that day, Wilbur flew 260 metres (852 feet) in 59 seconds. Soon afterwards a gust flipped the machine over on the ground. Severely damaged, it never flew again.

  Then the Wrights went quiet. Back home in Ohio, at a cow pasture outside Dayton called Huffman Prairie, they added a weight-powered catapult for take-off from 7 September 1904. On 5 October 1905, Wilbur flew 39.4 kilometres in 38 minutes and 3 seconds, circling until the fuel ran out. Afraid that rivals would copy them, they refused to fly anywhere without a firm contract to sell their aircraft, and local newspapers largely ignored the flights they did make. Few people saw them. European aviators called them bluffers. When Wilbur finally flew in public at Le Mans in France on 8 August 1908, his smooth banking turns ended the argument overnight. The editor of the French magazine L'Aérophile wrote that the flights had "completely dissipated all doubts". On 17 September 1908 a propeller broke during a demonstration in Virginia; Orville was badly hurt and his passenger, Lieutenant Thomas Selfridge, became the first person killed in a powered aeroplane crash.

  The credit disputes, presented honestly

  Three rival stories are worth knowing, because each shows what the word "first" hides.

  Alberto Santos-Dumont. In Paris, the Brazilian-born aviator flew his 14-bis 50 metres on 23 October 1906, winning a prize for the first flight of more than 25 metres, and on 12 November 1906 flew 220 metres in 22 seconds before officials of the Aéro-Club de France. His supporters say he deserves the title: his aircraft took off from wheels under its own power, in public, before official witnesses, while the Wrights flew in near-secrecy and from 1904 used a catapult. The case against: the 1903 Flyer used no catapult, just a launch rail and a headwind; by 1905 the Wrights were flying 39-kilometre controlled circles; and the 14-bis couldn't control roll until ailerons were added in November 1906.

  Gustave Whitehead. Born in Bavaria in 1874, Whitehead emigrated to the United States in 1893 and was reported by the weekly Bridgeport Herald (in an article dated 18 August 1901) to have flown about half a mile at 50 feet on 14 August 1901. No photograph of the flight has ever been found, and a key witness later swore he saw none. In 2013 the editor of Jane's All the World's Aircraft backed Whitehead (Jane's later called it the editor's personal view), and Connecticut passed a law recognising him. His supporters point to something real: in 1948, to get the Wright Flyer back from London, the Smithsonian agreed with Orville's estate to recognise no earlier aircraft as first, on pain of losing the Flyer. That is an awkward thing for a museum to sign. But evidence, not contracts, settles history, and the Royal Aeronautical Society and the Smithsonian's historians find the evidence for Whitehead doesn't hold up. "No compelling evidence that Whitehead or anyone else flew before the Wright brothers," said the Smithsonian's Peter Jakab in 2005.

  Langley and the Smithsonian. That 1948 agreement had a history. Samuel Langley, Secretary of the Smithsonian, saw his full-size piloted Aerodrome plunge into the Potomac in October and December 1903. In 1914 Glenn Curtiss, who was fighting the Wrights' patents, made short hops in a heavily modified Aerodrome, and the Smithsonian cited them as proof Langley's machine could fly, without disclosing the changes. Orville lent the Flyer to London's Science Museum in 1928 instead. The Smithsonian retracted in 1942, and the Flyer went on display in Washington on 17 December 1948, 45 years to the day after it flew.

  
    ClaimDateStrongest evidenceMain objection
    Whitehead14 Aug 1901Newspaper reportNo photograph; witnesses contradict; contemporaries unconvinced
    Wright brothers17 Dec 1903Photograph, diaries, witnesses, controlled flights to 1905Few public witnesses until 1908; rail launch
    Santos-Dumont23 Oct 1906Public, officially observed, wheeled take-offThree years later; no roll control at first
  

  The honest conclusion: "first" depends on the definition. For the first sustained, controlled, powered flight backed by solid evidence, the Wrights' claim is strong. For the first such flight before official judges, Santos-Dumont has a fair case. The Wrights' long patent war had a cost: when the United States entered the First World War in 1917, its forces flew French machines.

  Part 3 — How a wing really makes lift

  Ask most people how a wing works and you'll hear a version of this: the top of the wing is curved and longer, so air going over the top has further to travel; to meet the air from underneath at the back edge "at the same time", it must go faster; faster air has lower pressure (Bernoulli's principle); so the wing is sucked upward.

  That explanation, called the equal transit time theory, is in many textbooks and is wrong. NASA's Glenn Research Center lists the problems plainly:
  
    There is no physical law that says the two streams of air must meet again at the back. Air that starts side by side at the front edge does not arrive together at the back.
    Wings with identical top and bottom surfaces make lift. So does a flat paper aeroplane. And planes can fly upside down, which the theory can't explain.
    If you work out the speed the theory predicts, it is far too small to produce the lift real wings make.
  

  You can see the failure on film. In 2012, Holger Babinsky of Cambridge University released wind-tunnel footage of puffs of smoke flowing past a wing. Pause it and the smoke over the top has already passed the back edge while the smoke underneath is still on its way. The air over the top doesn't just keep pace; it arrives first, much faster than equal transit predicts.

  "It is the turning that's important, not the distance." NASA Glenn Research Center, on how wings make lift

  So what is going on? A wing meeting the air at a slight upward angle, its angle of attack, turns the flow. Air streaming past is bent downward behind the wing. Newton's third law then does the rest: if the wing pushes air down, the air pushes the wing up. Turning a flow also requires pressure differences, low above the wing and higher below it, and those are what you'd measure with gauges on the wing's surfaces. The "Newton" story (downwash) and the "Bernoulli" story (pressure and speed) are not rivals: they describe the same flow, and you need both for the whole picture. Physicists still argue over how best to explain, simply, why the flow follows the curved upper surface; the honest short answer is that it does, as long as the angle isn't too steep.

  The lift equation

  Engineers sum all of this up in one line:

  Lift = ½ × ρ × v² × S × CL

  Here ρ (rho) is the density of the air, v is the speed through it, S is the wing's area and CL, the lift coefficient, captures the wing's shape and its angle of attack. Each term tells you something practical:
  
    Speed is squared. Double your speed and you get four times the lift. This is why planes need runways: they must reach a speed at which lift equals weight.
    Density matters as much as speed. At 11,000 metres, air is about 3.4 times thinner than at sea level, so at the same angle a plane must fly about 1.8 times faster to stay up.
    Angle helps, until it doesn't. Up to a point, lift rises roughly in step with angle of attack. Beyond about 10–15° for a typical wing, the air can no longer follow the upper surface. It breaks away, lift collapses and drag soars. That is a stall, and it is what killed Lilienthal.
    Flaps change the shape. At take-off and landing, airliners slide out flaps and slats that make the wing bigger and more curved, raising its maximum lift so they can fly slowly without stalling.
  

  Use the explorer below to feel how these trade off.

  
    Interactive · The lift explorer
    
      Cessna 172
      Boeing 737-800
    
    
      Cessna, too slow to lift off
      737 cruising at 11 km
      Same 737, same speed, sea level
      Too steep: stall
    
    Airspeed (true)
      
    Altitude
      
    Angle of attack
      
    
    
    Lift from ½ρv²SCL with the International Standard Atmosphere for air density. Real wing areas, spans and maximum take-off weights. The lift coefficient is the textbook estimate for a well-shaped wing of that span (2π·A/(A+2) per radian, A = aspect ratio), with zero lift at −2° and a stall at CL = 1.4, clean wing, no flaps. Real wings differ in detail, and past the stall this simple model stops predicting lift.
  

  Three things to notice. At 100 km/h the Cessna's wing makes only about 60% of its weight; it needs about 128 km/h at that angle. Cruising at 837 km/h at 11,000 metres, lift almost exactly matches a fully loaded 737's weight, but at the same speed and angle at sea level the wing would make 3.4 times the weight, so the pilot would lower the nose to about 0°. And past about 14° the model gives up, because the real airflow does too.

  The model Cessna stalls at about 101 km/h with flaps up; the real one's book stall speed with flaps down is 47 knots, about 87 km/h. That difference is what flaps are for.

  Part 4 — Wars, jets and the shrinking planet

  Aeroplanes went to war almost as soon as they could fly. When the First World War began in August 1914, Germany's army had about 180 usable aircraft, used first for reconnaissance: finding troops, photographing trenches, spotting artillery fire. Pilots soon carried pistols, then machine guns. The problem of shooting through your own propeller was solved by a synchronising gear that fired only when no blade was in the way; Fokker's E.I fighter, which entered service in July 1915, gave Germany a period of air superiority the Allies called the "Fokker Scourge". Bombing from the air began too: German airship raids on England, from 1915 to 1918, killed 557 people.

  The Second World War turned aircraft into an industry on an unprecedented scale. The United States alone built more than 300,000 military aircraft between 1939 and 1945 (sources differ on the exact total). The war also pushed through the invention that defines modern flight.

  The jet engine: invented twice

  A piston engine turning a propeller loses efficiency as a plane flies faster and higher. The jet engine works on a different idea, which mechanics sum up as suck, squeeze, bang, blow: a compressor sucks in air and squeezes it, fuel burns in it, and the hot gas rushes out the back, spinning a turbine on its way that drives the compressor. The plane is pushed forward by throwing a stream of gas backward, Newton's third law again.

  It was invented twice, independently. Frank Whittle, a young Royal Air Force officer, filed his patent on 16 January 1930. In January 1935 it lapsed because the Air Ministry refused to pay the £5 renewal fee and he couldn't afford it. With private backers he formed Power Jets in 1936, and his first engine ran on 12 April 1937. Meanwhile in Germany, Hans von Ohain, initially unaware of Whittle's work, had his first engine running by September 1937, working with the aircraft maker Heinkel. On 27 August 1939, days before the war began, the Heinkel He 178 made the world's first flight powered by a jet. Britain's first jet, the Gloster E.28/39, flew on 15 May 1941. Germany's Me 262 became the first jet fighter in 1944.

  Modern airliners use a refinement, the turbofan: a big fan at the front, driven by the jet core, pushes most of the air around the core rather than through it. Moving a lot of air gently turns out to be more efficient than moving a little air violently, which is why airliner engines have grown so fat.

  Speed records fell. On 14 October 1947, Chuck Yeager flew the rocket-powered Bell X-1 at Mach 1.06, faster than sound, after being dropped from the bomb bay of a B-29.

  The jet age, and the price of a lesson

  The first jet airliner, Britain's de Havilland Comet, carried fare-paying passengers from London to Johannesburg on 2 May 1952. In 1954 two Comets broke up in flight over the Mediterranean, killing 35 people on 10 January and 21 on 8 April. Engineers pressurised a whole fuselage in a water tank again and again; it burst after 3,057 simulated flights, far fewer than expected. Metal fatigue had started from cracks at the corners of square cut-outs for radio antennas in the roof of the fuselage (not, as is often said, at the passenger windows). The lesson shaped every airliner since: stricter pressure testing and "fail-safe" structures in which a single crack can't bring the aircraft down.

  
    AircraftEntered serviceWhat it changed
    de Havilland Comet1952First jet airliner; its failures taught the industry about metal fatigue
    Tupolev Tu-1041956Aeroflot ran the first sustained regular jet services
    Boeing 7071958With the Douglas DC-8 soon after, it spread jets across the world's main routes
    Boeing 7471970First wide-body; built to cut seat costs and carry the masses
    Concorde1976Supersonic passenger travel, about 3.5 hours across the Atlantic
  

  The Boeing 747, the first wide-body and the original "jumbo jet", first flew on 9 February 1969 and entered service with Pan Am between New York and London on 22 January 1970. Pan Am had asked for a plane that would carry each passenger a given distance for about 30% less. In three classes it seated about 366 people. Boeing built 1,574 of them; the last was delivered on 31 January 2023.

  Concorde went the other way: speed, not seats. The Anglo-French airliner first flew on 2 March 1969 and entered service on 21 January 1976, cruising at about twice the speed of sound. Its sonic boom kept it off routes over land, the 1973 oil crisis made its fuel thirst frightening, its costs ran more than six times over estimates, and airlines cancelled their options. Only 20 were built. On 25 July 2000 an Air France Concorde ran over a metal strip dropped by another airliner; a burst tyre ruptured a fuel tank, and the burning aircraft crashed into a hotel, killing 113 people. It was Concorde's only fatal accident. The last commercial Concorde flight was on 24 October 2003.

  The jumbo jet and Concorde both flew in 1969. The future belonged to the one that made flying cheaper, not faster.

  Part 5 — Why flying became so safe

  Here is a statistic worth carrying around in your head. In 1970, airliners worldwide had about 6.5 fatal accidents for every million flights, according to Aviation Safety Network data compiled by Our World in Data. In 2017 the figure was about 0.28, more than twenty times lower. It bounces around from year to year (it was 0.48 in 2022), but the long-term fall is enormous, and it happened while the number of air passengers grew about fourteen-fold, from 310 million in 1970 to 4.5 billion in 2019.

  
    YearWhat the data saysSource
    19706.5 fatal accidents per million flightsAviation Safety Network, via Our World in Data
    20053.72 accidents of any kind per million flightsIATA
    20170.28 fatal accidents per million flightsAviation Safety Network, via Our World in Data
    20241.13 accidents per million flights (one per 880,000); 7 fatal accidents; 244 deaths on board; 40.6 million flightsIATA
    20251.32 accidents per million flights; 394 deaths on board; more fatal accidents than 2024; no loss-of-control accidentsIATA
  

  Note that the IATA rows count all accidents, including non-fatal ones, so they are higher than the fatal-accident rows. Note too that 2025 was clearly worse than 2024 on deaths: averages hide bad years, and every one of those 394 deaths was a person. But even in a bad year, one accident of any kind for roughly every 760,000 flights is a remarkable number for a machine that cruises at 800 km/h, ten kilometres up.

  Put the 2024 figure in personal terms. One accident per 880,000 flights means that if you boarded a flight every single day, you would expect to be on board for an accident of any kind, most of them non-fatal, about once in 2,400 years. In the Our World in Data series, which runs from 1970 to 2022, the safest year by fatal accidents per flight was 2017. And 2025, for all its losses, was the first year in IATA's records with no accident in which a crew lost control of an aircraft in flight, the category that includes stalls like the one that killed Lilienthal.

  How did flying get this safe? Not through any single invention, but through a habit of learning from every failure.
  
    Engineering lessons. The Comet taught designers to expect fatigue cracks and build structures that survive them.
    Human lessons. On 27 March 1977, two Boeing 747s collided on a fog-covered runway at Tenerife after a garbled radio exchange led one captain to start his take-off without clearance; 583 people died, still the deadliest accident in aviation. Afterwards, authorities worldwide adopted standard radio phrases. On 28 December 1978, a United Airlines crew near Portland, Oregon, ran out of fuel while busy with a landing-gear problem. These accidents led to crew resource management, a term coined at NASA in 1979: training that teaches junior crew to speak up and treats the captain as fallible. United began the first full programme in 1981.
    Technology that watches the pilots' backs. Terrain warning systems shout when an aircraft is too close to the ground. By February 2008, more than 40,000 aircraft carried the enhanced version and had flown over 800 million hours without a single accident of flying a working aircraft into terrain.
    Investigation in the open. Each serious accident is investigated, the findings are published, and they change training, procedures and design for airlines everywhere, as Tenerife and the Comet show.
  

  The lesson runs beyond aviation: a system becomes very safe when every failure is studied in the open and every finding changes the rules for everyone.

  Part 6 — Rotors and robots: helicopters and drones

  A helicopter's rotor blades are wings too, just whirling in a circle so that air flows over them even when the aircraft is standing still. The hard part, as da Vinci's aerial screw showed, is everything else. Spin a rotor one way and the body tries to spin the other way. Fly forward and the blade moving forward meets faster air than the blade moving backward, so one side of the rotor lifts more than the other. Solving those problems took decades.

  On 13 November 1907, Paul Cornu's twin-rotor machine in France lifted him about 30 centimetres for 20 seconds. In the early 1920s the Spanish engineer Juan de la Cierva invented the autogiro, whose unpowered rotor spins in the airflow like a sycamore seed, the first practical rotorcraft. Germany's Focke-Wulf Fw 61, which first flew in June 1936, was the first fully controllable helicopter. The layout most helicopters still use, one big main rotor and a small tail rotor pushing sideways to stop the body spinning, came from Igor Sikorsky's VS-300, which he developed in the United States from 1939 to 1943. His R-4 of 1942 was the first helicopter built in full-scale production.

  A pilot flies a helicopter with three controls. The collective changes the angle of all the main blades together, for more or less lift. The cyclic changes each blade's angle as it goes round, tilting the rotor disc so the aircraft moves forward, backward or sideways. Pedals vary the tail rotor's push to point the nose. In the Korean War, helicopter evacuation cut the time to get a wounded soldier to hospital from about eight hours in the Second World War to about three.

  Drones went from target practice to every pocket and every front line in less than a century.

  Drones: from target practice to every pocket and every front line

  Pilotless aircraft are older than you might think. In July 1849, Austrians besieging Venice launched about 200 bomb-carrying balloons; most missed when the wind changed. In 1918 the United States flew the Kettering Bug, an "aerial torpedo" that was not ready before the war ended. In 1935 Britain produced the Queen Bee, a radio-controlled target plane, and its name is said to be where the word "drone" comes from. During the Second World War, Reginald Denny's company built nearly 15,000 radio-controlled target drones for the US Army.

  Military drones became central in the following decades. In 1982 Israel coordinated drones with piloted aircraft in a lopsided victory over the Syrian air force. A US drone strike in November 2002 was an early example of armed drones used for targeted killing, a practice that remains legally and morally contested. In the war that began with Russia's 2022 invasion of Ukraine, both sides use drones in vast numbers, from cheap racing-style quadcopters to long-range attack drones; Ukraine created a separate branch of its armed forces for unmanned systems in 2024, and Russia followed in 2025.

  The civilian drone story is mostly about one company. DJI, the Chinese company founded in 2006 by Frank Wang, launched its Phantom quadcopter in 2013 and by mid-2024 held more than 90% of the world's consumer drone market. That success became a political issue: in December 2025 the US Federal Communications Commission banned the import, marketing and sale of new DJI drone models in the United States. Quadcopters themselves are a triumph of electronics more than aerodynamics: four fixed propellers, cheap motion sensors and a computer adjusting each motor's speed many times a second. To fly forward, the computer speeds up the rear rotors a little, tilting the whole drone so that part of its thrust points forward, a trick no human could manage by hand at that speed.

  Part 7 — Aviation in 2026, and the open questions

  IATA counted just under five billion passengers and more than 38 million flights in 2025. Growth has slowed: on 30 October 2026 IATA reported that passenger demand in September 2026 was 0.5% lower than a year before. The big questions now are about carbon and what might replace, or join, the jet.

  How much does flying warm the planet?

  In 2019 aviation produced about 2.5% of the world's carbon dioxide emissions from fossil fuels and land use, according to Our World in Data. That understates its climate effect, because planes also warm the planet through contrails, trails of ice crystals that can spread into heat-trapping high cloud. Counting those effects, a major 2021 study by David Lee and colleagues estimated that aviation accounted for about 3.5% of the human-caused warming push on the climate, with non-CO₂ effects making up about two-thirds of aviation's share. (For the physics of why CO₂ warms the planet, see The Physics of the Greenhouse Effect.) The non-CO₂ estimates are more uncertain than the CO₂ ones.

  Sustainable aviation fuel: the main bet, still tiny

  Airlines have committed to net-zero carbon emissions by 2050, and their trade body, IATA, says sustainable aviation fuel (SAF) would deliver about 65% of the cuts. SAF is jet fuel made from used cooking oil, farm and forest waste or, eventually, captured carbon and renewable electricity, blended with ordinary kerosene in today's engines. The most common type, made from waste fats and oils, cuts lifecycle emissions by about 80%, says IATA, because its carbon was recently taken from the air.

  
    YearSAF producedShare of all jet fuel
    20220.25 million tonnestiny
    20230.5 million tonnesabout 0.2%
    2024about 1.0 million tonnesabout 0.3%
    2025 (estimate)about 1.9 million tonnesabout 0.6%
    2026 (projection)about 2.4 million tonnesabout 0.8%
  
  Source: IATA SAF fact sheet, June 2026.

  The growth is real, but the base is tiny and SAF costs several times more than fossil jet fuel. The European Union's ReFuelEU law requires at least 2% SAF from 2025, rising to 70% by 2050, and the UK has a similar mandate. Supporters say only guaranteed demand will get factories built. IATA, for the airlines, argues that the mandates have so far failed to speed up production and have raised costs, and that governments should support production directly. The hard part is supply. Today's SAF comes mostly from waste fats and oils, of which there is only so much, so most future SAF would have to come from newer processes that are not yet running at scale.

  Electric planes: the battery problem

  Why not use batteries, as cars increasingly do? Because weight is everything in the air. Jet fuel holds about 43 megajoules of energy per kilogram; lithium-ion batteries hold between about 0.36 and 0.9. Electric motors are far more efficient than jet engines, which narrows the gap, but nowhere near enough for a large airliner on a long route. And a battery doesn't get lighter as it empties.

  So electric flight has started small. The first fully electric aircraft to earn a type certificate was the two-seat Pipistrel Velis Electro, certified by the European Union Aviation Safety Agency on 10 June 2020. It flies for about 50 minutes plus a reserve and is used mainly for pilot training.

  Air taxis: where they really stand

  The most talked-about new aircraft are eVTOLs (electric vertical take-off and landing aircraft) with several battery-powered rotors, pitched as quiet air taxis. Their status as of late October 2026:

  
    Company and aircraftStatus
    EHang EH216-S (China)Two-seat, pilotless, about 35 km range. China's aviation authority granted type certification in October 2023 and, in March 2025, approved pilotless flights that take off and land at the same site. By August 2026 it held type, production and airworthiness certificates and was carrying paying passengers in China.
    Joby S4 (US)Pilot plus four passengers, up to about 240 km range. In what Joby calls the fifth and final stage of US type certification. Flew from JFK airport to Manhattan on 27 April 2026; demonstration flights in Texas and Florida from September under a US government pilot programme. Not yet certified as of 30 October 2026; Joby's stated target is first passengers in 2026.
    Archer Midnight (US)Not in passenger service; its launch has slipped from an original 2025 target.
    Lilium (Germany)Insolvent in October 2024, stopped operations that December, insolvent again in February 2025 after a rescue failed.
    Volocopter (Germany)Planned to fly passengers at the Paris 2024 Olympics; it didn't, lacking certification. Filed for insolvency on 30 December 2024.
  

  Boosters argue that eVTOLs will be quieter and cleaner than helicopters, cheap per trip once built in numbers, and a way to use the sky above jammed cities. Critics point to short ranges, heavy batteries, the cost of building landing sites, the need to prove a new kind of aircraft as safe as airliners, and the doubt that air taxis will ever be more than a premium service. The record so far supports both: real certification progress in China and the US, and European pioneers that ran out of money first.

  The air taxi has been "two years away" for a decade. In 2026 it is, for the first time, carrying paying passengers in at least one country.

  Open questions and trade-offs

  Can aviation reach net zero by 2050? The airlines' plan rests on SAF growing from under 1% of fuel to most of it in 24 years, plus better aircraft and offsets. Optimists point to how fast solar power scaled once costs fell; sceptics point to feedstock limits and argue demand growth must slow. Neither side can prove its case yet.

  What about contrails? If non-CO₂ effects really are about two-thirds of aviation's warming, avoiding the worst contrails could be one of the cheapest climate measures there is, but it means flying slightly longer routes and burning a little more fuel. Weighing a certain small extra CO₂ against an uncertain contrail benefit is an active scientific argument.

  How much should we automate? In September 2026 Joby flew a converted Cessna Caravan 3,199 miles across the United States with its safety pilot making no control inputs, and pilotless air taxis already carry passengers in China. Technology that watches over pilots has made flying safer; taking the pilot out entirely is a different step, and it will test how much trust passengers and regulators give a machine.

  Who gets the sky? Drones, air taxis and helicopters will increasingly share low airspace, bringing benefits along with noise, privacy and safety worries. The rules are being written now.

  And one lesson from 1903 still applies. The Wrights succeeded not through money or a big engine but by finding the real problem, control, and testing everything themselves. That is still how flying gets better.

  
    The takeaway
    People first flew by floating, in 1783. Powered flight needed lift, power and, above all, control, which the Wrights combined in 1903; Santos-Dumont and Whitehead have real but weaker claims. A wing makes lift by turning air downward; the "equal transit time" story is wrong. Lift grows with the square of speed and with air density, which is why runways exist and airliners cruise fast in thin air. Jets, invented twice, made flying fast and cheap, and learning openly from every accident made it very safe: fatal accidents per flight fell more than twentyfold from 1970 to 2017. In 2026 the hard problems are carbon (SAF is under 1% of fuel), batteries too heavy for big planes, and air taxis certified in China but still finishing certification in the US.
    Try this: on your next flight, sit by the wing and watch the take-off. You'll see flaps slide out to make the wing bigger and more curved, the plane accelerate until ½ρv²SCL equals its weight, and the pilot raise the nose to increase the angle of attack at the moment of lift-off. On landing, watch the flaps extend even further. Then use the lift explorer to find the speed your aircraft would need at 11,000 metres.
  

  
    Sources & Further Reading
    Sources. Smithsonian National Air and Space Museum object records (1903 Wright Flyer, Bell X-1) and Codex on the Flight of Birds page; Science Museum Group (Cayley's 1799 disc); NASA Glenn Research Center, "Incorrect Theory #1: Equal Transit"; University of Cambridge, "How wings really work" (2012). IATA: 2024 safety press release (26 February 2025), Annual Safety Report 2025 executive summary, SAF fact sheet (June 2026), press release of 30 October 2026. Our World in Data: fatal airliner accidents per million flights (Aviation Safety Network and ICAO data) and "What share of global CO₂ emissions come from aviation?" (Ritchie, 2024, citing Lee et al. 2021). World Bank: air passengers carried. Joby Aviation news releases (5 August to 30 October 2026); eVTOL Insights (20 August 2026). Wikipedia articles on each person, aircraft and event named were used for leads and cross-checked against these sources where possible. Every explorer value and every ratio in the text was computed in node and checked in Python. The full claim-by-claim log is kept with the library's fact-check files.
    
      🌐 NASA Glenn: the equal-transit myth, and the Beginner's Guide to Aeronautics around it — the clearest free explanations of lift, drag and engines, written for students
      🌐 University of Cambridge: How wings really work — Babinsky's smoke-puff video, two minutes that kill the myth for good
      🌐 Smithsonian: the 1903 Wright Flyer — the museum's record of the machine and its four flights
      📘 The Bishop's Boys: A Life of Wilbur and Orville Wright — Tom D. Crouch (the deep biography, by a longtime Smithsonian curator)
      📘 The Wright Brothers — David McCullough (the readable narrative version)
      🌐 Our World in Data: aviation's share of emissions — the numbers on CO₂ and contrails, with charts
      🌐 IATA: Sustainable aviation fuels fact sheet — the airlines' own figures and argument (read it as one side's case)


Rockets and Spaceflight: From Fireworks to Reusable Boosters
------------------------------------------------------------
A rocket is the only machine that can push itself through empty space, and the maths of how it does that makes getting to orbit brutally hard. Here is how humans learned to do it anyway, what it cost, and what changed when rockets started landing.

Follows: Orbits: Why the Moon Doesn't Fall (or Rather, Keeps Falling), and the space-race chapters of The 1950s and The 1960s in Decades of Leadership
On 16 March 1926, in a farm field at Auburn, Massachusetts, a physics professor lit a contraption of pipes and tanks fuelled by gasoline and liquid oxygen. It rose 12.5 metres, flew for about two and a half seconds, and landed 56 metres away in a cabbage patch. Forty-three years later, a descendant of that machine, 111 metres tall, sent three men to the Moon. By 2025 the world was launching rockets into orbit more than 300 times a year, and many of their first stages were flying home to land.
In this deep dive

Fire arrows, three dreamers and one equation
Why reaching orbit is so hard
From the V-2 to the Moon
The Space Shuttle, its two disasters, and the station it built
Rockets that land, and a crowded sky
The state of affairs in late 2026
Open questions and trade-offs

Part 1 — Fire arrows, three dreamers and one equation
The principle of the rocket is older than the rocket. In the ancient world, Hero of Alexandria described the aeolipile, a sphere mounted over a boiling kettle that spun as steam escaped from two bent pipes. It was a curiosity, but it showed that a jet of gas pushes back on whatever releases it.
The first true rockets came from China, where gunpowder was invented. A tube closed at one end and packed with gunpowder, tied to a long stick for stability, would shoot forward when lit, pushed by the hot gas rushing out of the open end. In 1232, during the battle of Kai-Keng, Chinese defenders fired what were recorded as "arrows of flying fire" at Mongol invaders. NASA's own history notes that we don't know how much damage they did, though the psychological effect was probably large.
There is also the legend of Wan Hu, a Chinese official said to have tied 47 rockets to a chair, lit them all at once and vanished in a cloud of smoke. It is a legend: there is no evidence it happened, and if it did, as NASA's history page dryly puts it, he was probably blown to pieces. Treat it as folklore, not history.
Rockets spread west as weapons. Rocket barrages fired by Indian forces against the British in 1792 and 1799 impressed the British enough that Colonel William Congreve designed rockets for Britain's own forces. During the War of 1812, British ships fired Congreve rockets at Fort McHenry in Baltimore, the scene that gave the US national anthem its "rockets' red glare".
For centuries, though, rockets stayed inaccurate and weak. They burned solid powder, which once lit can't be throttled or stopped. The leap came when someone thought of burning liquids.
That was Robert Goddard, a physics professor at Clark University in Massachusetts. Liquid fuels pack more energy and can be pumped, throttled and shut off. His first liquid-fuelled rocket, nicknamed "Nell", made that short flight into the cabbage patch at Auburn on 16 March 1926. It was not impressive to look at. It was the Wright Flyer of rocketry.

A rocket isn't pushed by the air behind it. It is pushed by its own exhaust, which is why it works best where there is no air at all.

Spaceflight had three founding thinkers, who worked mostly in isolation from each other in Russia, the United States and Germany.
Konstantin Tsiolkovsky (1857–1935) lost most of his hearing to scarlet fever at nine, was turned away by schools and taught himself mathematics and physics from books. Working as a provincial schoolteacher, he wrote out the basic equation of rocket motion and dated it 10 May 1897. In May 1903 he published Exploration of Outer Space by Means of Rocket Devices, which argued for liquid propellants and showed that what limits a rocket's final speed is the speed of its exhaust. He is credited with the line, from a letter of 1911, that "Earth is the cradle of humanity, but one cannot live in a cradle forever." (The wording varies between translations.)
Robert Goddard derived the same mathematics independently around 1912 and held 214 patents, including two from 1914 for multi-stage and liquid-fuelled rockets. In 1919 the Smithsonian published his paper A Method of Reaching Extreme Altitudes. On 13 January 1920 a New York Times editorial, "A Severe Strain on Credulity", mocked him, claiming that a rocket couldn't work in a vacuum because it would have nothing to push against. On 17 July 1969, the day after Apollo 11 left for the Moon, the Times printed a correction: "it is now definitely established that a rocket can function in a vacuum as well as in an atmosphere. The Times regrets the error." Goddard had died in 1945.
Hermann Oberth (1894–1989) had his doctoral dissertation on rockets rejected in 1922 as "utopian", and published it privately in 1923 as Die Rakete zu den Planetenräumen (The Rocket into Planetary Space). The book helped inspire small rocket societies around the world. Oberth advised Fritz Lang's 1929 film Woman in the Moon and mentored the members of the German Society for Space Travel, among them a young enthusiast named Wernher von Braun.

Pioneer
Country
Key work
What he contributed

Tsiolkovsky
Russia
1903
The rocket equation; liquid fuels; exhaust speed as the key limit

Goddard
United States
1914 patents, 1919 paper, 1926 flight
The first liquid-fuelled rocket; practical engineering

Oberth
Germany
1923 book
A generation of enthusiasts, including von Braun

The equation all three found is now called the Tsiolkovsky rocket equation, though an English mathematician, William Moore, had outlined it as early as 1810. It is short, and it explains almost everything about why spaceflight is hard.
Part 2 — Why reaching orbit is so hard
As Orbits explains, to stay up you must move sideways fast enough that the Earth curves away beneath you as fast as you fall: about 7.8 kilometres per second, or 28,000 km/h, in low orbit. On the way up a rocket also loses speed to gravity (it spends minutes fighting its own weight) and to air resistance. Engineers add those losses up as a "change of velocity", Δv (delta-v). Getting from the ground to low Earth orbit takes about 9.3 to 10 km/s.
The rocket equation says how much propellant that needs:
Δv = ve × ln(m0 ÷ mf)
Here ve is the exhaust speed, m0 is the rocket's mass full of propellant and mf its mass when the propellant is gone; "ln" is the natural logarithm. Flip it round and the full-to-empty mass ratio is eΔv ÷ ve: an exponential (see Why Exponential Growth Rules Everything). Each extra km/s you want multiplies, rather than adds to, the propellant you need.
Put numbers in. A kerosene and liquid oxygen engine like SpaceX's Merlin has an exhaust speed of about 3 km/s in vacuum; the hydrogen-burning RS-25 that powered the Space Shuttle reaches about 4.4 km/s. Take Δv = 9.4 km/s:

Engine type
Exhaust speed
Full ÷ empty mass
Share of launch mass that must be propellant

Kerosene + oxygen
~3.0 km/s
about 23
about 96%

Hydrogen + oxygen
~4.4 km/s
about 8.5
about 88%

Ninety-six per cent. Everything else (engines, tanks, pumps, computers, the payload) has to fit in the remaining 4%. A loaded petrol car, for comparison, is a few per cent fuel. A rocket is a flying fuel tank with a very small amount of anything else.

The problem isn't energy. Each kilogram in low orbit carries about 30 megajoules of motion energy, less than the chemical energy in a kilogram of kerosene. The problem is that the propellant has to lift itself.

This is why rockets are built in stages. Once the first stage has burned its propellant, its empty tanks and heavy engines are dead weight, so you throw them away and carry on with a smaller rocket. Splitting 9.4 km/s into two equal halves, each kerosene stage needs only a mass ratio of about 4.8, so "only" about 79% of each stage is propellant, which engineers can actually build.
Real rockets match the arithmetic. A Falcon 9 weighs about 549 tonnes at lift-off and can put 22.8 tonnes into low orbit if nothing is recovered: about 4% of its launch mass. If the first stage keeps back fuel to fly home and land on a ship, the payload drops to 17.5 tonnes, about 3%. The Saturn V that launched Apollo weighed nearly 3,000 tonnes and could put 140 tonnes in low orbit, about 5%.
Going further is harder still. To head for the Moon a spacecraft must reach about 11 km/s, close to the escape speed in Orbits, so the same Saturn V that could put 140 tonnes in low orbit could send only 43.5 tonnes towards the Moon, less than a third as much. Every extra km/s is paid for exponentially.
This is sometimes called "the tyranny of the rocket equation", and it has a consequence that shaped sixty years of spaceflight. If almost all of a rocket is fuel and tanks, and you throw the tanks and engines away after a single use, each launch costs roughly the price of building a large aircraft and then crashing it into the sea.
Part 3 — From the V-2 to the Moon
The V-2: the first rocket into space, built by prisoners
The first rockets big enough to reach space were built in Nazi Germany, and they were built by slaves.
In the 1930s the German army recruited von Braun and other rocket enthusiasts to develop long-range missiles. The result was the A-4, renamed by propaganda the Vergeltungswaffe 2, "Vengeance Weapon 2": the V-2. It burned alcohol and liquid oxygen, consuming about a tonne of propellant every seven seconds, could carry roughly a tonne of explosive about 320 kilometres, and fell on its target faster than sound, so there was no warning. On 20 June 1944, a V-2 test rocket launched straight up reached about 174.6 km, well above the conventional 100 km edge of space: the first human-made object to get there. Combat use began in September 1944, against Paris and then London. Between 3,200 and 3,600 were fired at Allied targets.
After British bombers hit the Peenemünde research centre in August 1943, production moved to the Mittelwerk, a factory dug into a mountain in central Germany. According to the Smithsonian National Air and Space Museum, which displays a V-2, the rockets were built by about 2,000 civilian technicians and roughly 10,000 prisoners from the Mittelbau-Dora concentration camp, many of them forced at first to live underground in terrible conditions. The museum's figures make the weapon's history plain:

Estimate (Smithsonian NASM)

People killed by V-2 attacks
about 5,500

Concentration-camp prisoners who died making it
at least 10,000

Other estimates differ (some put deaths from the attacks nearer 9,000 and among the workers nearer 12,000), but every serious account agrees on the grim headline: more people died building the V-2 than were killed by it.

The first machine to reach space was a weapon of terror, built by prisoners who were worked to death.

Von Braun joined the Nazi party in 1937 and the SS in 1940, rising to the rank of major. He admitted visiting the Mittelwerk factory 12 to 15 times and called conditions there "repulsive"; he denied seeing deaths himself. In May 1945 he surrendered to the Americans and was taken to the United States with about 1,600 other German specialists under Operation Paperclip. His team built the rocket that launched America's first satellite, and later, as director of NASA's Marshall Space Flight Center, he was chief architect of the Saturn V. Both the United States and the Soviet Union built their first large rockets using captured V-2s and V-2 knowledge. How to weigh von Braun's achievements against his past is still argued over, and an honest account of the space age has to include both.
Sputnik, Gagarin and Apollo
On 4 October 1957, the Soviet Union launched Sputnik 1 from Tyuratam in Kazakhstan (now the Baikonur Cosmodrome). It was a polished metal sphere 58 cm across, weighing 83.6 kg, with four long radio antennas, and it circled the Earth every 96.2 minutes. Its radio signal lasted 22 days, until the batteries died; it burned up on 4 January 1958 after 1,440 orbits. It was launched on a modified R-7, the intercontinental missile designed under the Soviet chief designer Sergei Korolev.
The political shock in the United States was enormous. Within months Washington had created the Advanced Research Projects Agency and passed the National Defense Education Act, which funded science and maths students, and Congress passed the act that created NASA. Sputnik 2 had already carried the dog Laika into orbit on 3 November 1957; she died of overheating within hours, a fact the Soviet government hid until 2002. America's first satellite, Explorer 1, followed on 31 January 1958.
On 12 April 1961, Yuri Gagarin flew one orbit of the Earth in Vostok 1, about 108 minutes from launch to landing according to the official figure (a later analysis puts it at 106). "Poyekhali!" ("Let's go!"), he said at lift-off. What the Soviet Union hid for ten years was that he ejected about 7 km up and came down by parachute separately from his capsule, a detail that, under the rules then in force, could have disqualified the flight as a record; it was not admitted until 1971. Alan Shepard became the first American in space on 5 May 1961, with a 15-minute flight that went up and came down without orbiting.
Three weeks later, on 25 May 1961, President John F. Kennedy asked Congress to commit the United States to landing a man on the Moon and returning him safely to Earth before the decade was out.
The programme that followed, Apollo, became one of the largest engineering projects in history. At its peak about 400,000 people worked on it. It had a terrible start: on 27 January 1967, a fire during a ground test killed Gus Grissom, Ed White and Roger Chaffee inside their capsule, filled with pure oxygen and fitted with a hatch that opened inward and could not be opened in time. Crewed flights stopped for about 20 months while the spacecraft was redesigned. In December 1968, Apollo 8 became the first crewed mission to leave Earth orbit and circle the Moon.
Apollo 11 launched on 16 July 1969. On 20 July, Neil Armstrong stepped onto the Moon: "That's one small step for a man, one giant leap for mankind." He and Buzz Aldrin spent 21 hours and 36 minutes on the surface while Michael Collins orbited above. They splashed down in the Pacific on 24 July.

Apollo by the numbers

Crewed landings
6, from 1969 to December 1972

People who walked on the Moon
12

Moon rock and soil brought back
382 kg

Saturn V rockets launched
13

Programme cost
$25.4 billion in 1973 dollars, about $257 billion in 2020 dollars

The Soviet Union ran its own crewed lunar programme. Its giant N1 rocket was launched four times between February 1969 and November 1972, and all four launches failed; the second exploded on the pad. Korolev had died in January 1966. The programme was halted in 1974. The race was over.
Apollo 13, in April 1970, showed how close to the edge it all ran: an oxygen tank exploded on the way to the Moon, and the crew survived by using their lunar lander as a lifeboat. Apollo 17 in December 1972 was the last. No one has walked on the Moon since.
Part 4 — The Space Shuttle, its two disasters, and the station it built
After Apollo, NASA built a spacecraft meant to be reusable. The Space Shuttle was a winged orbiter that launched like a rocket, with two solid rocket boosters and a huge external fuel tank, and landed like a glider. The boosters were recovered from the ocean and refurbished; the tank was thrown away. Each orbiter was designed for about 100 flights.
It first flew on 12 April 1981, twenty years to the day after Gagarin. Over 30 years it flew 135 missions, could carry 27.5 tonnes to low orbit at a time, and helped build the International Space Station. It was a superb machine in many ways. It was not cheap. NASA's own analysis, by Harry Jones of NASA Ames Research Center in 2018, put a Shuttle launch at about $1.5 billion averaged over the programme, or about $54,500 per kilogram to low orbit. Refurbishing the orbiter and boosters between flights turned out to be slow, painstaking and expensive.
It also killed fourteen people.
Challenger, 28 January 1986. Seventy-three seconds after launch, Challenger broke apart, killing its seven crew, including the schoolteacher Christa McAuliffe. The presidential commission traced the cause to rubber O-rings sealing a joint in the right solid rocket booster. On that unusually cold morning they failed to seal, and hot gas burned through to the external tank. Before the launch, engineers at the booster's maker, Morton Thiokol, had recommended against launching below 53 °F (about 12 °C); NASA managers pushed back, and Thiokol's senior managers overruled their own engineers. The commission called it "an accident rooted in history": the O-ring problem had been known for years. During a televised hearing, the physicist Richard Feynman dropped a piece of O-ring material, squeezed in a clamp, into a glass of iced water and showed that it lost its springiness when cold. His appendix to the report ended with a line every engineer should know:

"For a successful technology, reality must take precedence over public relations, for nature cannot be fooled." Richard Feynman, 1986

Columbia, 1 February 2003. Eighty-two seconds after Columbia's launch on mission STS-107, a large piece of foam insulation broke off the external tank and struck the left wing, damaging the carbon heat shielding on its front edge. On re-entry at the end of the mission, superheated gas entered the wing and destroyed it, and all seven crew died. The investigation board found the deeper cause in the organisation: foam had hit orbiters on earlier flights, and managers had come to treat it as an acceptable risk. It recommended, among 29 recommendations, that the Shuttle be retired. The last Shuttle flight landed in July 2011.
Two disasters, two different physical causes, one shared pattern: a known warning sign became normal because nothing bad had happened yet. Feynman had seen the same thing in the numbers: his appendix criticised NASA's wildly optimistic estimates of how reliable the Shuttle was.
The International Space Station is the Shuttle's most lasting legacy. Its first module, Zarya, launched on 20 November 1998, and people have lived on board continuously since 2 November 2000, nearly 26 years by the end of October 2026. It weighs about 450 tonnes, and five space agencies run it: NASA, Roscosmos, the European Space Agency, Japan's JAXA and the Canadian Space Agency, with fifteen countries involved. It is planned to operate until the end of 2030; in June 2024 NASA chose SpaceX, in a contract worth up to $843 million, to build a vehicle that will steer it into a controlled re-entry away from populated areas. Commercial stations are being developed to follow it.
Part 5 — Rockets that land, and a crowded sky
The obvious fix for "build an airliner, crash it into the sea" is to land the rocket and fly it again. It took a long time to make that work, because of the rocket equation: every kilogram of landing legs, fins and fuel saved for the landing is a kilogram of payload you can't carry.
SpaceX's Falcon 9 first flew on 4 June 2010. In December 2015, a Falcon 9 first stage, after sending satellites on their way, flipped round, fired its engines to fly back, and landed upright on a landing pad: the first time the main stage of an orbital rocket had come back for a controlled landing. On 30 March 2017 a booster that had already flown launched again. By 2025 individual boosters had flown 30 times, and SpaceX's certified limit had risen to 40 flights each. As of 31 October 2026, according to the running record kept on Wikipedia, Falcon 9 had flown 700 times, with 697 fully successful missions.
What did that do to prices? Jones's 2018 NASA analysis compared the Shuttle's $54,500 per kilogram with SpaceX's advertised Falcon 9 price at the time, $62 million for up to 22,800 kg, which works out to about $2,720 per kilogram. "Commercial launch," he concluded, "has reduced the cost to LEO by a factor of 20."

Launch system
Cost per kg to low orbit
Basis

Space Shuttle
about $54,500
~$1.5 billion per launch, 27,500 kg (NASA, Jones 2018)

Falcon 9 (2018 list price)
about $2,720
$62 million, 22,800 kg (Jones 2018)

Falcon 9 (2026 list price)
about $3,250 to $4,230
$74 million, 22,800 kg expendable or 17,500 kg with booster landing (computed)

Three honest caveats. First, these are list prices for a whole rocket at its full capacity, not what a typical customer pays for each kilogram it actually sends. Second, the list price has risen since 2018. Third, outside estimates put SpaceX's internal cost per launch well below its price, which matters because about three-quarters of Falcon 9 launches in 2025 (122 of 162) carried SpaceX's own Starlink satellites. Still, the fall in cost is real and large, and it changed what is worth launching.
Starship, SpaceX's much larger fully reusable system, has had a mixed and fast-moving test campaign. On 13 October 2024 its giant first stage was caught by mechanical arms on the launch tower on its return. In 2025 several flights failed, others succeeded. The third version of the rocket first flew on 22 May 2026, and on 28 September 2026 Starship reached orbit for the first time, releasing 26 Starlink satellites. As of late October 2026 it had not yet caught its upper stage, or demonstrated the ship-to-ship transfer of propellant in orbit that SpaceX's Moon-lander plan relies on. SpaceX says full reuse will cut costs far further; that has not been shown yet.
Other countries are following. In 2026, Chinese rockets were reported to have recovered first stages too, including the privately built Zhuque-3 in August.
A crowded sky: mega-constellations and debris
Cheap launches changed what goes up. Here is the number of orbital launch attempts worldwide, from the catalogue maintained by the astrophysicist Jonathan McDowell:

Year
Orbital launch attempts
Notes

1957
3
Sputnik year

1967
139
Highest of the 20th century, at the height of the space race

1982
129
The Soviet Union alone made 108 attempts

2004
54
Post-Cold War low

2020
114

2025
329
All-time record: US 181, China 92, Russia 17

2026 (to 26 October)
253
US 133, China 76

In 2025, SpaceX alone launched roughly half of all the world's orbital rockets.
Most of that cargo is communications satellites flying in huge swarms, or mega-constellations, in low orbit, from a few hundred to about 1,200 kilometres up, beaming internet to users on the ground. As of 29 October 2026, McDowell's count showed SpaceX had launched 13,069 Starlink satellites, of which 11,190 were still in orbit and 1,879 had come back down and burned up. Amazon's system, launched as Project Kuiper and renamed Amazon Leo in November 2025, had launched just under 400 satellites; it missed a US licensing deadline to have about half of its 3,236 planned satellites up by July 2026, and the regulator granted a waiver. China's state-backed Guowang network plans more than 13,000 satellites and had about 200 operating by September 2026; a second Chinese network, Qianfan, had about 250 launched. OneWeb has launched about 650.
Boosters say these networks bring fast internet to ships, aircraft, remote villages and disaster zones that cables will never reach. Critics worry about one company, or one government, controlling so much infrastructure, about the satellites' military uses, and about crowding.
That last worry has a name: the Kessler syndrome. In June 1978 the scientists Donald Kessler and Burton Cour-Palais warned that once enough objects are in orbit, collisions create fragments that cause more collisions, a chain reaction that could make some orbits unusable for generations. Some events since have made it worse: in January 2007 China destroyed one of its own old weather satellites, about 865 km up, with a missile; in February 2009 two satellites collided by accident; in November 2021 a Russian anti-satellite test destroyed an old Soviet satellite, creating about 1,500 tracked fragments that crossed the orbits of both the ISS and China's Tiangong station.

Space debris, according to ESA's 2025 Space Environment Report

Objects tracked by surveillance networks
about 40,000

Active satellites among them
about 11,000

Debris larger than 10 cm (estimated)
more than 50,000

Debris larger than 1 cm (estimated)
more than 1.2 million

Old satellites or rocket bodies re-entering
more than 3 a day, on average

ESA's report states the problem bluntly: even if no more launches happened, the amount of debris would keep growing through collisions, so actively removing old objects will be needed. The saving grace of very low orbits is drag: the thin upper atmosphere pulls satellites down within years, which is why so many dead Starlinks have already burned up. Higher orbits have no such cleaner.

Low orbit cleans itself, slowly. Higher up, every piece of junk stays for decades or centuries.

Part 6 — The state of affairs in late 2026
The United States and Artemis. NASA's Artemis programme aims to return people to the Moon. Artemis I flew an uncrewed Orion capsule around the Moon in 2022. On 1 April 2026, Artemis II launched four astronauts (NASA's Reid Wiseman, Victor Glover and Christina Koch, and Canada's Jeremy Hansen) on a nine-day flight around the Moon, splashing down in the Pacific on 10 April. Koch became the first woman and Glover the first person of colour to travel beyond low Earth orbit, Hansen the first non-American to fly around the Moon, and the crew travelled about 406,771 km from Earth, beyond Apollo 13's record. In early 2026 NASA restructured the next steps: according to NASA's Artemis pages (updated in October 2026), Artemis III, planned for 2027, will now stay in low Earth orbit to test docking between Orion and test versions of the SpaceX and Blue Origin landers, and Artemis IV, planned for 2028, is to be the first crewed landing since 1972. Both landers still have major tests ahead, and NASA's schedules have slipped repeatedly before.

More than fifty years after Apollo 17, people have flown around the Moon again. Nobody has yet walked on it again.

China. China has built and runs its own space station, Tiangong, completed in 2022 and usually crewed by three astronauts; China was effectively kept out of the ISS after a 2011 US law barred NASA from direct cooperation with Chinese space agencies. In June 2024 its Chang'e 6 mission brought back about 1.9 kg of rock from the far side of the Moon, the first samples ever returned from there. China's human spaceflight agency announced in 2023 a plan to land two astronauts on the Moon before 2030, using a new Long March 10 rocket, a Mengzhou crew capsule and a Lanyue lander. Its 2025 tally of 92 orbital launch attempts was second only to the United States.
India. On 23 August 2023, India's Chandrayaan-3 made the first soft landing near the Moon's south pole, making India the fourth country to land on the Moon, at a reported cost of around 615 crore rupees (under $100 million). Its Gaganyaan programme is preparing India's first crewed spacecraft; the first uncrewed test flight had not flown as of late October 2026. In July 2026 Vikram-1, built by the private company Skyroot, reached orbit.
Japan and Europe. On 19 October 2026 (UTC), Japan launched MMX on its new H3 rocket, a mission to bring back samples from Phobos, one of the moons of Mars. Europe's new heavy rocket, Ariane 6, first flew in its most powerful four-booster form on 12 February 2026, and on 5 September 2026 Isar Aerospace's Spectrum rocket, launched from Andøya in Norway, made what is reported as the first successful orbital launch from Western European soil.
Russia. Russia continues to fly cosmonauts and cargo to the ISS on Soyuz and Progress spacecraft. Its launch rate has fallen sharply from the Soviet era: 17 attempts in 2025, compared with 108 by the Soviet Union in 1982.
Rules. The 1967 Outer Space Treaty, with 118 parties as of October 2025, forbids any nation from claiming the Moon or other bodies, bans nuclear weapons in orbit, and makes states responsible for their companies' activities in space. The US-led Artemis Accords, first signed in October 2020 and joined by 78 countries as of 28 October 2026, set out principles for lunar exploration, including the use of resources mined in space. Russia has called them an attempt to write space law in America's favour, and China and Russia are developing a separate International Lunar Research Station with their own partners. Supporters see the Accords as practical rules for a busier Moon; critics see them as stretching the treaty's ban on appropriation.
Part 7 — Open questions and trade-offs
People or robots? Robotic missions are far cheaper and take no human risk. Chang'e 6 brought back far-side samples without a crew. Supporters of human spaceflight answer that people do science in the field that robots can't, that crewed missions inspire public support and students, and that a species confined to one planet takes a long-term risk. The ISS has at least shown that people can live and work in orbit without a break for more than a quarter of a century; whether that justifies its cost compared with dozens of robotic missions is exactly the disagreement. These are judgements about value, not just engineering, and reasonable people weigh them differently.
Who pays, and who benefits? Launch prices fell because a private company built reusable rockets, helped by NASA contracts. Supporters see the best of public-private partnership; critics see one company becoming hard to do without, for launches, internet service and even astronaut transport. Both describe the same facts.
Can low orbit be kept usable? Mega-constellations, anti-satellite tests and dead rocket bodies all add risk. To see how fast things have changed: at the end of 2024, ESA counted about 11,000 active satellites of all kinds; by late October 2026, SpaceX alone had about 11,000 working Starlinks. Very low orbits and planned de-orbiting limit the risk from each satellite, but ESA warns that debris would keep growing even if launches stopped. Debris removal is still at the demonstration stage, and there is no global traffic-control system for space.
What is it worth? Apollo cost about $257 billion in 2020 dollars; India's Chandrayaan-3 landed near the Moon's south pole for a reported 615 crore rupees, under $100 million, roughly three thousand times less (Apollo carried people and returned rock, so this is not a like-for-like comparison). Falling costs mean the question is shifting from "can we afford to go?" to "what is worth doing once we are there?", and that is a question about science, money and values, not just rockets.
How safe can crewed spaceflight be? The Shuttle lost two of its 135 missions, about one flight in 68, and the Apollo 1 fire killed three more astronauts on the ground. No airliner would be allowed to fly with odds anything like that. Riding on top of hundreds of tonnes of burning propellant remains far riskier than any form of everyday travel. How much risk is acceptable, for professionals, for paying tourists, for a Moon landing, is a question each programme answers differently.

Cheaper rockets turned "can we afford to go?" into "what is worth doing once we're there?"

Will space stay peaceful? The 1967 Outer Space Treaty bans nuclear weapons and other weapons of mass destruction in orbit. The anti-satellite tests of 2007 and 2021 used ordinary missiles, and both left long-lived debris that every other operator now has to dodge. Satellites now matter to navigation, communications, finance and armies alike, which makes them both more valuable to protect and more tempting to attack, and the rules on conventional weapons in space are far thinner than the rules on nuclear ones.
Is there a new race? The United States plans a crewed landing in 2028, China before 2030. Some see a rivalry like the 1960s; others point out that the two programmes have different partners, legal frameworks and goals, and that the US-led Artemis Accords and the China-Russia lunar station project are building two different clubs rather than one finish line.
And the oldest lesson in this story still holds. Challenger and Columbia were not lost to unknown physics. They were lost to warnings that had become routine. Feynman's line applies to every rocket that flies now: nature cannot be fooled.

The takeaway
Rockets began as Chinese fire arrows; the science of spaceflight came from three isolated dreamers, Tsiolkovsky, Goddard and Oberth, and one equation. Because the propellant has to lift itself, the rocket equation forces a kerosene rocket to be about 96% propellant to reach orbit, which is why rockets have stages and why throwing them away made spaceflight so expensive. The first rocket into space, the V-2, was built by concentration-camp prisoners, and more of them died making it than were killed by it. Sputnik, Gagarin and Apollo were driven by Cold War rivalry; the Shuttle was reusable but cost about $54,500 per kilogram, and its two disasters came from normalised warnings. Rockets that land cut costs to roughly a twentieth, and launch rates hit a record 329 attempts in 2025. In late 2026 the sky is filling with constellations and debris, Artemis II has flown around the Moon, and both the US and China aim to land people there before 2030.
Try this: test the rocket equation yourself with a calculator. Pick an exhaust speed of 3 km/s and work out e^(Δv ÷ 3) for Δv = 3, 6 and 9.4 km/s. Notice how the mass ratio goes from about 2.7 to 7.4 to 23: tripling the speed you need multiplies the propellant, it doesn't triple it. Then find the next launch on a public launch schedule and watch the first stage come home.

Sources & Further Reading
Sources
- Early rockets (Hero's aeolipile, Chinese fire-arrows, 1232 Kai-Keng, the Wan Hu legend, Indian rocket barrages of 1792 and 1799, Congreve, Fort McHenry), Tsiolkovsky's 1898/1903 work, Goddard's 16 March 1926 flight (2.5 s, 12.5 m, 56 m), Oberth's 1923 book, V-2 propellant use: NASA Glenn Research Center, "Brief History of Rockets".
- Goddard ("Nell", 1919 paper, 1920 New York Times editorial and 1969 correction, 214 patents); Tsiolkovsky (1857–1935, equation dated 10 May 1897, May 1903 paper, 1911 letter); Oberth (rejected dissertation, 1923 book, Woman in the Moon, VfR); the rocket equation's history (Moore 1810): Wikipedia articles on each, used for leads and cross-checked.
- V-2 production, prisoner labour and deaths, deaths from attacks, numbers fired, and its role as ancestor of launch vehicles: Smithsonian National Air and Space Museum, V-2 (A-4) object record. V-2 altitude record (174.6 km, 20 June 1944), first combat use, range: Wikipedia, "V-2 rocket". Von Braun's party and SS membership, Mittelwerk visits, Paperclip: Wikipedia, "Wernher von Braun".
- Sputnik 1, Sputnik 2, Explorer 1, ARPA and NDEA: Wikipedia, "Sputnik 1". Vostok 1: Wikipedia, "Vostok 1". Freedom 7: Wikipedia. Apollo 11 timeline and quote: NASA, "Apollo 11 Mission Overview". Apollo programme figures: Wikipedia, "Apollo program"; Apollo 1: Wikipedia, "Apollo 1"; N1: Wikipedia, "N1 (rocket)"; Saturn V: Wikipedia.
- Shuttle: Wikipedia, "Space Shuttle"; Challenger: Rogers Commission findings and Feynman's appendix (via Wikipedia, "Rogers Commission Report"); Columbia: Wikipedia, "Columbia Accident Investigation Board". Shuttle and Falcon 9 cost per kilogram and the factor of 20: Harry W. Jones, "The Recent Large Reduction in Space Launch Cost", NASA Ames Research Center, ICES 2018 (NASA Technical Reports Server).
- ISS: Wikipedia, "International Space Station"; US Deorbit Vehicle: NASA news release, 26 June 2024.
- Delta-v to orbit (9.3–10 km/s): Wikipedia, "Delta-v budget"; exhaust speeds (Merlin 1D 310 s, RS-25 453 s): Wikipedia, "Specific impulse". Falcon 9 mass, payload, first flight, 700 launches as of 31 October 2026, 2026 list price: Wikipedia, "Falcon 9"; booster reuse milestones: Wikipedia, Falcon 9 booster list.
- Starship flight history: Wikipedia, "SpaceX Starship flight tests", with the 28 September 2026 orbital flight and its 26 Starlink payloads cross-checked in Jonathan McDowell's launch log.
- Orbital launch attempts by year and country: Jonathan McDowell, General Catalog of Artificial Space Objects, table updated 26 October 2026; SpaceX share computed from his launch log (updated 29 October 2026). Starlink totals: McDowell, Starlink statistics page (29 October 2026). Amazon Leo: Wikipedia, "Amazon Leo"; Guowang: Wikipedia, "Guowang".
- Debris: ESA, "ESA Space Environment Report 2025"; Kessler and Cour-Palais (1978), ASAT tests and the 2009 collision: Wikipedia, "Kessler syndrome".
- Artemis II: NASA, Artemis II mission page; record distance and crew firsts: Wikipedia, "Artemis II". Artemis III and IV plans: NASA, Artemis III mission page (updated 19 October 2026) and Artemis programme page (updated 21 October 2026).
- China: Wikipedia, "Tiangong space station", "Chang'e 6", "Lanyue". India: Wikipedia, "Chandrayaan-3", "Gaganyaan". 2026 launches (Ariane 64, Spectrum, Vikram-1, Zhuque-3, H3/MMX): Wikipedia, "2026 in spaceflight", cross-checked in McDowell's launch log. Outer Space Treaty and Artemis Accords: Wikipedia articles of those names.
- Computed in Python: mass ratios and propellant fractions (e^(9.4/3.0) = 23.0, 95.6%; e^(9.4/4.4) = 8.5, 88.2%; two-stage e^(4.7/3.0) = 4.8, 79%), payload fractions (Falcon 9 4.2% and 3.2%; Saturn V about 5%), kinetic energy per kilogram at 7.8 km/s (30.4 MJ, against 43 MJ/kg for kerosene), 2026 Falcon 9 price per kilogram.
Go deeper
- 🌐 NASA Glenn: Brief History of Rockets — a short, clear history from Hero to the V-2, written for students
- 🌐 Jonathan McDowell's space statistics — the independent, constantly updated count of every launch, used by journalists and researchers
- 🌐 ESA Space Environment Report 2025 — the authoritative yearly picture of space debris
- 🌐 NASA: Artemis — the current Moon plans, from the source (check the "updated" date)
- 📘 A Man on the Moon — Andrew Chaikin (the definitive narrative of Apollo, built on interviews with the astronauts)
- 📘 Von Braun: Dreamer of Space, Engineer of War — Michael J. Neufeld (a Smithsonian historian's balanced biography, honest about Mittelwerk)
- 📘 "What Do You Care What Other People Think?" — Richard Feynman (includes his own account of the Challenger investigation)


Cars: From the Benz Patent-Motorwagen to Electric and Self-Driving
------------------------------------------------------------------
In 140 years the car went from a three-wheeled curiosity that frightened horses to the machine that shaped our cities, our wars over oil and our air. Here is where it came from, how it works, what it cost, and where it stands at the end of 2026.

Follows: Energy: The One Currency of the Universe and The Physics of the Greenhouse Effect
In August 1888 a 39-year-old woman named Bertha Benz woke her two teenage sons before dawn, pushed her husband's motor carriage quietly out of the yard in Mannheim, and set off for her mother's house in Pforzheim. She did not tell her husband. She had no map made for cars, because there were no cars to make maps for, and no petrol stations, because there was no petrol trade. By evening the three of them had arrived, and she sent her husband a telegram.
In 2025, more than 20 million new cars sold around the world ran on electricity, and in a handful of American and Chinese cities you could hail a taxi with nobody in the driver's seat. Between those two moments the car was invented, cheapened, blamed, regulated, cleaned up and reinvented. It has saved enormous amounts of time and killed more than four million people on American roads alone.
In this deep dive

From steam carriages to Benz and Bertha
Four strokes, a moving line, and a remade world
Making crashes survivable
The tailpipe, and the electric comeback
Self-driving: the levels, the robotaxis and the hype
The state of affairs in late 2026, and the trade-offs

Part 1 — From steam carriages to Benz and Bertha
The car was not one invention. It was an idea, a carriage that moves itself, waiting for an engine light and strong enough to make it practical. People tried three kinds of engine, and for a while it was not obvious which would win.
Steam came first. In 1769, a French army engineer named Nicolas-Joseph Cugnot built a small steam-powered fardier, a heavy cart meant to haul cannon, and in 1770 a full-size version. It weighed about two and a half tonnes, was badly balanced, and had to stop every quarter of an hour or so to relight its fire and raise steam again. The army gave up on it. The machine itself was kept, and since 1800 it has been in the collection of the Conservatoire National des Arts et Métiers in Paris. You may read that it also caused the world's first car crash by knocking down a wall. That story is not recorded in any document from the time, so treat it as legend.
Steam carriages kept improving through the 19th century, but they had real drawbacks: a boiler that took time to heat, a fire to tend, and water to carry.
Electric cars came next, and for a while they looked like the future. The US Department of Energy's history of the electric car credits Robert Anderson, a British inventor, with a crude electric carriage in the early 19th century. In Paris, on 19 April 1881, Gustave Trouvé fitted an improved electric motor and a rechargeable battery to a tricycle and drove it along the rue de Valois. Around 1890, William Morrison of Des Moines, Iowa, built what the Department of Energy calls the first successful electric car in the United States: six seats and a top speed of 14 miles per hour. In 1897, electric taxis appeared on the streets of London and New York. On 29 April 1899, a Belgian driver, Camille Jenatzy, became the first person to drive faster than 100 kilometres an hour, in a bullet-shaped electric car called La Jamais Contente ("Never Satisfied").
Electric cars were quiet, did not smell, and did not need to be cranked by hand. That made them popular with city drivers. By around 1900, the Department of Energy says, they made up roughly a third of the vehicles on American roads.

Around 1900, a third of America's cars were electric. It took the car industry a century to get back there in some countries, and it hasn't yet in most.

They lost for reasons that will sound familiar. Batteries were heavy and expensive, and range was short. Then the petrol car got two decisive advantages. Ford's Model T made it cheap: by 1912 a petrol car cost $650 while an electric roadster sold for $1,750. And in 1912 Charles Kettering introduced the electric starter, which ended the dangerous business of hand-cranking a petrol engine and removed the electric car's main convenience advantage. By 1935, the Department of Energy says, electric cars had nearly disappeared.
1886: Benz and Daimler
On 29 January 1886, Carl Benz, an engineer in Mannheim, applied for a German patent for a "vehicle powered by a gas engine". It was granted as patent number 37435, and the Mercedes-Benz company today calls it the birth certificate of the automobile. The vehicle, the Benz Patent-Motorwagen, was a light two-seater on three wire-spoked wheels, with a single-cylinder, four-stroke petrol engine lying flat at the back. It produced about three-quarters of a horsepower. It had electric ignition with a spark plug, a differential so the rear wheels could turn at different speeds on a bend, and a tubular steel frame. Its first public outing was reported in the newspapers in July 1886.
What made it new was not any single part but that the engine and the chassis were designed together as one machine, rather than an engine bolted onto a cart.
Just about 100 kilometres away, in Cannstatt near Stuttgart, Gottlieb Daimler and Wilhelm Maybach were working on the same problem. They never met Benz. In November 1885 their light, fast-running engine, nicknamed the "grandfather clock" for its shape, powered a two-wheeled "riding car" on a successful test run, and in 1886 they fitted it to a four-wheeled carriage. Forty years later, in 1926, their company and Benz's merged to form Daimler-Benz.
Bertha's drive
Benz was a gifted engineer and, by most accounts, a poor salesman. The machine needed someone to prove it was useful, and that person was his wife.
Bertha Benz had put part of her dowry into her husband's business before they married. In August 1888, with her sons Eugen and Richard (15 and 14 in Mercedes-Benz's account; other sources give slightly different ages), she drove an improved Patent-Motorwagen from Mannheim to Pforzheim and back a few days later. Mercedes-Benz gives the round trip as 180 kilometres; other accounts put the one-way distance at a little over 100. Either way, until then the car had only made short trips that ended where they started.
The stories of the journey are wonderful, and worth telling with care. She is said to have bought fuel, a cleaning solvent called ligroin, at a pharmacy in Wiesloch, cleared a blocked fuel line with a hatpin, insulated a wire with a garter, and had a shoemaker fit leather to the wooden brake blocks. The boys pushed the car up hills it could not climb. These details come from the family and the company's later telling, and they have become part of the legend. What is not in doubt is the result: she reported the car's weaknesses, and Benz added a lower gear for hills.
So who invented the car?
"Karl Benz invented the car in 1886" is a reasonable summary and a misleading one. Here are the main claimants.

Who
When
What
Why it isn't usually called "the first car"

Nicolas-Joseph Cugnot
1769–70
Steam-powered gun carriage
Steam, military, never practical

Étienne Lenoir
1860
Gas-engined carriage
About 9 km in some three hours; ran on gas, not liquid fuel

Siegfried Marcus
about 1870, and 1888–89
Petrol engine on a handcart; a later car survives in Vienna
Experimental; no influence on later cars. His second car was long misdated to 1875

Daimler and Maybach
1885–86
Fast petrol engine; two-wheeler, then a motorised coach
A strong co-claim: same year as Benz

Carl Benz
1885–86
Purpose-built petrol car, patented and later sold
Usually credited, because it was designed as a car and went into production

George Selden (US)
Filed 1879, granted 1895
Patent on a petrol car
A patent rather than a production car; he collected licence fees from other carmakers

Even Benz's engine relied on someone else's idea. The four-stroke cycle had been patented by Nicolaus Otto, whose company had employed Daimler and Maybach. Benz filed his own patent at a time when Otto's patent was being challenged in court.
Part 2 — Four strokes, a moving line, and a remade world
How a petrol engine works
Almost every petrol car built since then has used the same basic cycle Benz used. Inside each cylinder, a piston slides up and down. It is joined by a rod to a crankshaft, which turns the up-and-down motion into rotation, the way your legs turn a bicycle's pedals. Two valves at the top let air and fuel in and exhaust out. The cycle has four steps, called strokes:

Intake. The intake valve opens and the piston moves down, drawing a mixture of air and fuel into the cylinder.
Compression. Both valves close and the piston moves back up, squeezing the mixture into a small space. Squeezed fuel burns much more forcefully, which is the insight that made Otto's engine efficient.
Power. A spark plug fires, the mixture burns, and the hot, expanding gas shoves the piston down hard. This is the only stroke that does work; the other three are housekeeping.
Exhaust. The exhaust valve opens and the piston moves up again, pushing the burnt gas out towards the tailpipe.

The crankshaft goes round twice for each full cycle. A heavy flywheel keeps it turning through the three strokes that don't push, and a car has several cylinders firing in turn so that something is always pushing. A diesel engine works the same way, except that it squeezes air so hard that it gets hot enough to ignite the fuel without a spark.

Only one stroke in four pushes the car. Most of the energy in the fuel leaves the engine as heat.

That last point matters for everything that follows. According to the US government's fuel-economy site, only about 12% to 30% of the energy in petrol reaches a car's wheels, depending on how it is driven. The rest is lost, mostly as heat from the engine. (Energy: The One Currency of the Universe explains why no heat engine can turn all of its fuel into motion.)
Ford and the moving assembly line
Early cars were built by hand, one at a time, by skilled workers walking around a stationary chassis. They were expensive toys.
Henry Ford's Model T went on sale on 1 October 1908. It was simple, tough and easy to repair. In 1913, at its Highland Park plant in Michigan, Ford began moving the car past the workers instead of the workers around the car: the moving assembly line. Each worker did one small task, again and again, as the chassis was pulled past.
The effect was enormous. Assembling a Model T chassis took about twelve and a half hours before the line; by 1914 it took about an hour and a half, roughly eight times faster. Ford's own history says the car sold for $825 in 1908 and $260 in 1925. Production ended on 26 May 1927, after more than 15 million had been made.

Before the moving line
After

Time to assemble a Model T chassis
About 12½ hours
About 1½ hours (by 1914)

Price of a Model T
$825 (1908)
$260 (1925)

There was a human cost. The work was relentless and repetitive, workers left for other jobs, and Ford responded by raising pay through profit-sharing to what became known as the $5 day, more than double the usual wage. The assembly line spread to almost every kind of manufacturing, and so did the argument it started, about whether such work makes people richer, emptier, or both.
How the car remade cities, oil and suburbs
Once cars were cheap, people reorganised their lives around them, and then governments reorganised the land.
Suburbs. Before the car, most city dwellers lived within walking distance of work or of a tram or train line. The car made it possible to live much further out, on cheaper land, in a bigger house. In the United States this was accelerated by government. On 29 June 1956, President Eisenhower signed the act that created the Interstate Highway System: 41,000 miles of motorway, with $25 billion authorised for the years 1957 to 1969. The motorways made long commutes practical and pulled shops, offices and factories out of city centres after the people.
Cities. Building those roads through existing cities meant knocking things down. Many urban motorways were deliberately routed through poor neighbourhoods, often Black ones, as part of "urban renewal". Large numbers of people lost their homes, and some neighbourhoods were cut in two. A wave of "freeway revolts" stopped or rerouted a number of planned roads. Elsewhere, Japan, South Korea and much of Europe grew their suburbs more around railways, which is part of why their cities feel so different.
Oil. Cars run on oil, and so the car made oil one of the most important commodities in the world. In its 2026 report, the International Energy Agency says road transport accounts for close to half of global oil demand. The oil shocks of the 1970s led governments to set fuel-economy standards, and the IEA notes that the fuel economy of conventional cars has nearly doubled since 1975. Even so, oil supply has shaped wars and foreign policy for a century, and in 2026 a conflict in the Middle East has again pushed fuel prices up and reminded oil-importing countries how exposed they are.

The car didn't just change how people travelled. It changed where they lived, what their governments built and which countries mattered.

Whether this was good is a real argument, not a settled one. Defenders of car-based development point to freedom of movement, bigger and cheaper homes, and jobs that people could reach. Critics point to traffic, pollution, lost neighbourhoods, isolated people who can't drive, and land given over to parking. Most cities now try to do both: keep the roads and add better transit, cycling and walking.
Part 3 — Making crashes survivable
For most of the 20th century, the car was treated as a safe machine driven by careless people. Crashes were blamed on "the nut behind the wheel". That view changed in the 1960s, and the change saved hundreds of thousands of lives.
Nader and the "second collision"
In 1965, a young lawyer named Ralph Nader published Unsafe at Any Speed. Its argument was that car makers were designing cars that turned survivable crashes into fatal ones, and resisting safety features such as seat belts because they cost money and did not sell cars. Its best-known chapter attacked General Motors' Chevrolet Corvair.
General Motors responded by hiring private investigators to dig into Nader's private life, and its president apologised to him at a Senate hearing in 1966. Nader later sued GM for invasion of privacy and won. In 1966 Congress passed laws setting federal safety standards for cars, and in 1970 the Highway Safety Act created the National Highway Traffic Safety Administration (NHTSA). For fairness: a 1972 NHTSA report later disputed Nader's specific claims about the Corvair's handling. The book's larger argument, that design matters as much as driver behaviour, became the foundation of modern car safety.
The key idea is the "second collision". When a car hits a wall, the first collision is the car. The second, a split second later, is the people inside hitting the steering wheel, the dashboard or the windscreen. Safety engineering is mostly about that second collision: keeping people in place, and making the car crumple at the front so that the people inside slow down over a longer distance.
The seat belt
The modern three-point seat belt, the diagonal-plus-lap belt in every car today, was designed by Nils Bohlin, an engineer at Volvo. Volvo fitted it as standard in 1959, and Bohlin's US patent was granted in 1962. NHTSA's figures today: wearing a seat belt in the front seat of a car cuts your risk of being killed in a crash by 45%, and by 60% in an SUV, pickup or van. In 2025, 91.3% of Americans wore one. Yet in 2024, nearly half of the people killed in American passenger vehicles (where belt use was known) were not buckled up.
Crash tests
From 1978 NHTSA began crashing new cars into barriers with instrumented dummies and publishing the results; its 5-Star ratings followed in 1993. Europe's Euro NCAP began publishing results in 1997. Crash ratings work by embarrassment: a car that scores badly sells badly. The result was airbags, crumple zones, stronger cabins, and later, electronic stability control and automatic emergency braking.
The data

Year (US)
Road deaths
Deaths per 100 million vehicle miles

1921
13,253
24.08

1966 (first federal safety laws)
50,894
5.50

1972 (peak year for deaths)
54,589
4.33

2024
39,254
1.19

2025 (NHTSA early estimate)
36,640
1.10

The number that matters most is the right-hand column. Americans drive more than three times as many miles as they did in 1966, but the risk per mile has fallen by about 80%. Between 1899 and 2024, NHTSA counts 4,036,087 traffic deaths in the United States.
The world is a harsher picture. The World Health Organization's fact sheet (July 2026) puts global road deaths at about 1.16 million a year. Road injuries are the leading cause of death for children and young people aged 5 to 29. Some 92% of road deaths happen in low- and middle-income countries, which have about 60% of the world's vehicles, and more than half of the people killed are pedestrians, cyclists and motorcyclists, the people outside the car. The United Nations has set a goal of halving road deaths by 2030.
Part 4 — The tailpipe, and the electric comeback
Leaded petrol: the century's worst additive
Early engines "knocked": the fuel exploded too early, which damaged the engine and limited how much it could be compressed. On 9 December 1921, Thomas Midgley Jr., working under Charles Kettering at General Motors, found that adding tetraethyl lead to petrol stopped the knocking. It went on sale in 1922. The marketing name was "Ethyl", which avoided the word "lead".
The danger was known early. In 1924 several workers at a Standard Oil plant in New Jersey died of lead poisoning and many more were badly hurt, and the newspapers called the additive "loony gas". Public-health experts had already written to warn about it. Sales were paused for a review and then resumed.
Proof of the damage came from an unexpected place. The geochemist Clair Patterson, trying to measure the age of the Earth, found lead contamination everywhere. Ice cores from Greenland and Antarctica showed that lead in the atmosphere had risen sharply after leaded petrol arrived. From 1965 he campaigned to remove it, against strong opposition from the lead industry and its scientists. The United States phased lead out in stages and banned it for road vehicles from January 1996.

Leaded petrol was sold for almost a century. The last country to stop selling it, Algeria, did so in July 2021.

When the UN Environment Programme announced the end of leaded petrol worldwide in August 2021, it estimated that ending it would prevent more than a million premature deaths a year and save the world economy $2.45 trillion. Lead harms children's developing brains, and the drop in blood-lead levels after it was banned is one of the great public-health victories of the century. (There is also a much-debated hypothesis linking leaded petrol to violent crime; some studies support it and others find no link, so treat it as unsettled.)
Catalytic converters
Petrol engines also emit carbon monoxide, unburned fuel and nitrogen oxides, which together make smog. To meet new US Environmental Protection Agency rules, most petrol cars from the 1975 model year were fitted with catalytic converters: a honeycomb coated with platinum-group metals that turns carbon monoxide and unburned fuel into carbon dioxide and water. From 1981, "three-way" converters also broke down nitrogen oxides. A catalytic converter is ruined by lead, so its arrival also helped end leaded petrol.
Dieselgate
Diesel engines are efficient but produce more nitrogen oxides, which are hard to clean up. In 2014, researchers at West Virginia University, testing cars on real roads for the International Council on Clean Transportation, found that some Volkswagen diesels emitted far more than they did in the laboratory.
On 18 September 2015, the US Environmental Protection Agency issued Volkswagen with a notice of violation. The cars contained software, a "defeat device", that recognised when they were being tested and switched on full emissions controls only then. In normal driving, the EPA found, the 2.0-litre cars emitted up to 40 times the permitted nitrogen oxides. About 590,000 cars sold in the United States were involved, and Volkswagen said about 11 million worldwide had the software. In January 2017, Volkswagen agreed to plead guilty to three criminal felony counts in the US and pay a $2.8 billion criminal penalty, on top of civil settlements and buy-backs. The scandal badly damaged diesel's image as the "clean" choice.
The electric comeback
Electric cars never fully vanished. They lived on in milk floats, forklifts and golf carts. The revival began with clean-air rules and an oil scare.
The 1973 oil embargo revived interest. In the early 1990s, California's air regulators began pushing car makers towards zero-emission vehicles, and General Motors built the EV1, with a range of about 80 miles. GM discontinued it in 2001. In 1997 Toyota launched the Prius in Japan, the first mass-produced hybrid, which used a battery and electric motor to help a petrol engine; it went on sale worldwide in 2000.
The modern electric car arrived with the lithium-ion battery, the same kind that powers your phone. (The next piece, Batteries: The Chemistry That Powers Your Phone and Your Car, tells that story.) The Tesla Roadster, first delivered in 2008, ran on small cylindrical lithium-ion cells. The Nissan Leaf, launched in December 2010, became the first electric car to sell more than 100,000. Then battery prices fell by around 90%, and the electric car stopped being an expensive statement.
How an electric car is different
An electric car has a battery, one or more electric motors, and power electronics that control how much current flows to them. There is no gearbox with many gears, no oil change, no exhaust.
The big difference is efficiency. The US government's fuel-economy site says electric cars turn more than 77% of the electricity they take from the grid into power at the wheels, against 12% to 30% of the energy in petrol for a conventional car. When an electric car slows down, its motor runs as a generator and puts energy back into the battery, which is called regenerative braking. The IEA estimates that a typical battery electric car uses around 70% less energy per kilometre than a petrol car of similar size. To work out what that means in money, see Your Utility Bills, Decoded; to see where the electricity comes from, see How Electricity Actually Reaches Your Wall.

Petrol car
Battery electric car

Energy from fuel or grid reaching the wheels
About 12–30%
More than 77%

Refuelling
Minutes, at a station
Mostly at home; a DC fast charger adds roughly 100–200 miles (160–320 km) in 30 minutes

Tailpipe emissions
Yes
None

Weakness
Fuel cost, pollution
Battery cost and weight, charging access

Part 5 — Self-driving: the levels, the robotaxis and the hype
"Self-driving" is a slippery phrase, and companies have used it loosely. Engineers use a scale published by SAE International, a standards body for vehicle engineering, in its standard J3016. It was first published in 2014, and the April 2021 revision is the one most often cited. NHTSA follows the same scale.

SAE level
What the system does
Who is responsible

0
Warnings and momentary help, such as emergency braking
You drive, all the time

1
Steering or speed control, such as adaptive cruise
You drive, all the time

2
Steering and speed together
You drive and must watch constantly

3
Drives itself in limited conditions
You can look away, but must take over when asked

4
Drives itself within a defined area or set of conditions
Nobody needs to be ready to drive there

5
Drives itself anywhere a human could
Nobody, anywhere (no such system exists)

The big line falls between Level 2 and Level 3. At Level 2, the car is helping you, and if it crashes, you were driving. Most "autopilot"-style features sold today are Level 2. Tesla's current consumer system is officially called "Full Self-Driving (Supervised)", and the word in brackets matters: it is a Level 2 system that needs an attentive driver. NHTSA says plainly that no vehicle for sale to the public in the United States is fully automated, and that every car on sale requires the driver's full attention.

"Self-driving" describes a hope. The SAE levels describe who is responsible when something goes wrong.

Robotaxis in 2026
Level 4 exists, but in taxis, not in cars you can buy. Each robotaxi works only inside an "operational design domain": particular mapped areas, roads and conditions.
Waymo, which began as Google's self-driving project, is the largest operator in the United States. As of 31 October 2026, its website lists 15 metropolitan areas where it serves riders, from Phoenix and the San Francisco Bay Area to Miami and Nashville, and says it has served more than 20 million rides. Through June 2026 it had driven 271.3 million miles with no human driver.
Waymo publishes a comparison with human drivers on the same kinds of city streets. Its figures through June 2026 show about 82% fewer crashes causing any injury, and about 95% fewer causing serious injury or worse. Two cautions belong with those numbers: they are the company's own analysis (it publishes its method), and they cover city streets, not the motorways and snowstorms that humans also drive through. Waymo itself says there aren't yet enough miles to compare fatal crashes.
Tesla has taken a different route, using cameras rather than the laser scanners (lidar) most rivals use, and aiming to turn the cars it sells into robotaxis through software. In September 2026, after Tesla began commercial driverless rides in Austin, Texas, with its Cybercab, a vehicle built without traditional driving controls, NHTSA opened an audit of whether the vehicle meets federal safety standards. In May 2026, NHTSA also announced that a 2026 Tesla Model Y was the first vehicle to pass its new driver-assistance tests.
China is the other centre of robotaxi development, with companies such as Baidu's Apollo Go. The IEA's 2026 report counts commercial driverless taxis in more than 20 cities, mainly in China and the United States, all of them electric.
The setbacks are real too. In October 2023, California suspended the permits of Cruise, General Motors' robotaxi company, finding its vehicles unsafe and saying the company had misled regulators after one of its cars dragged a pedestrian who had been thrown into its path by another vehicle. In December 2024, GM said it would stop funding Cruise's robotaxi work.
Part 6 — The state of affairs in late 2026, and the trade-offs
Where electric cars stand
The best global source is the International Energy Agency's Global EV Outlook 2026, published on 20 May 2026. ("Electric" in its counts includes plug-in hybrids.)

Market
Electric share of new car sales, 2025

World
About one in four (more than 20 million cars)

China
Nearly 55% (more than 60% in April 2026, a monthly record)

Europe
28%

United States
Just under 10%

Norway
About 97%

Southeast Asia
Nearly 20%

The IEA expects about 23 million electric car sales in 2026, about 28% of the market. In the first quarter of 2026, sales fell 8% from a year earlier worldwide, because of policy changes in China and the United States, while Europe rose about 30%. In the US, federal tax credits for buying electric cars ended after September 2025, and sales in the last quarter of 2025 were 45% lower than a year earlier. Chinese car makers accounted for about 60% of the world's electric car sales in 2025, and factories in China made nearly three-quarters of all electric cars.
The case for

Efficiency and running costs. An electric car uses far less energy per kilometre, and the IEA finds running-cost savings of roughly $550 to $1,000 a year compared with petrol in major markets since 2020, more when fuel prices spike.
Climate. Over its life, including making the battery, the IEA estimates that a medium-size battery electric car bought in 2023 emits about half as much greenhouse gas as an equivalent petrol car, and the gap grows as electricity gets cleaner.
Air and oil. No tailpipe means no exhaust in city streets. The IEA estimates the world's electric vehicles avoided about 1.7 million barrels of oil a day in 2025.
Prices. In China, the IEA says, 70% of battery electric cars sold in 2025 were already cheaper than the average conventional car.

The case against, or at least for caution

The grid matters. In countries that still burn a lot of coal, the gain is smaller. The IEA estimates that in India, an electric car's lifetime emissions are only about 20% lower than a petrol car's.
Weight and dust. Battery cars are heavy. Heavier vehicles of any kind shed more tyre and road dust, and the European Commission's Euro 7 rules are the first anywhere designed to limit tyre and brake emissions. Heavier vehicles also hit harder.
Charging. In 2025 there were about 11 electric cars for every public charge point worldwide, and many people in flats have nowhere to plug in at home. Only about half of electric car owners in China have their own home charger.
Supply chains and mining. Batteries need lithium, nickel, graphite and sometimes cobalt, and most of the processing is in China. The next piece covers the human and environmental costs of that mining.
Prices and jobs elsewhere. Outside China, electric cars often still cost more to buy, and Europe's and America's car industries are under pressure from cheaper Chinese imports. Governments are answering with tariffs and subsidies, which divide economists.
Self-driving. Supporters argue that robots that never drink, text or get tired could prevent many of the 1.16 million road deaths a year. Sceptics point to the narrow areas robotaxis work in, the confusing marketing of Level 2 systems, the cost of the sensors, the loss of driving jobs, and the question of who is liable when a robot kills someone. Both sides can point to real evidence; the long-run safety record is still being written.

One quieter conclusion runs under all of this. Whatever powers it, a car still takes up road space and parking, still needs roads, and still hits pedestrians. Many transport researchers argue that the biggest gains in safety and emissions in cities come not from changing what's under the bonnet but from needing to drive less.

The takeaway
The car came in three flavours, steam, electric and petrol, and petrol won around 1910 because Ford made it cheap and the electric starter made it easy. Benz's 1886 Patent-Motorwagen is the usual "first car", but Daimler and Maybach built theirs the same year, and Bertha Benz's 1888 drive proved it was useful. A petrol engine turns only about 12–30% of its fuel's energy into motion. Cars remade suburbs, cities and world politics through oil; safety engineering after Nader cut America's death rate per mile by about 80% since 1966, yet about 1.16 million people still die on the world's roads each year. Leaded petrol and Dieselgate show what happens when emissions are hidden. In 2025 about a quarter of new cars sold worldwide were electric; true self-driving exists only as robotaxis in limited areas.
Try this: next time you're in a car, look for its safety label or search its crash-test rating on NHTSA's or Euro NCAP's website, and check whether it has automatic emergency braking. Then look at your own week: count how many of your car trips were under 3 kilometres. Those short trips are where walking, cycling or transit could replace the car most easily.

Sources & Further Reading
Sources
- Benz patent 37435 (filed 29 January 1886), engine output (0.75 hp), features, July 1886 outing; Bertha Benz's 1888 drive (sons' ages 15 and 14, 180 km round trip); Daimler and Maybach's "grandfather clock" engine, November 1885 riding car, 1886 motor carriage, "just 100 kilometres away": Mercedes-Benz Group, company history pages "Benz Patent Motor Car: The first automobile (1885–1886)" and "Beginnings of the automobile (1886–1920)". Patent grant date, the Otto patent litigation, the reliability of the 1888 car, and the details of Bertha's trip (Wiesloch, hatpin, garter, leather brakes): Wikipedia, "Benz Patent-Motorwagen" and "Bertha Benz", used as leads.
- Cugnot (1769–70, weight, relighting, CNAM since 1800, unrecorded "first crash"); Lenoir; Marcus (1870 handcart, 1888–89 car, misdating); Selden patent; Jenatzy (29 April 1899); Trouvé (19 April 1881); 1897 electric taxis: Wikipedia articles on each, used as leads.
- Early electric cars (Anderson; Morrison about 1890, 14 mph; a third of vehicles about 1900; $650 vs $1,750 in 1912; Kettering's 1912 starter; near disappearance by 1935; 1973 embargo; 1990 Clean Air Act amendments; EV1 discontinued 2001, 80-mile range; Prius 1997 and 2000; Leaf December 2010): US Department of Energy, "The History of the Electric Car" (2014). Roadster and Leaf cell formats and the Leaf as first EV over 100,000 sales: IEA, Global EV Outlook 2026, battery chapter.
- Four-stroke cycle and two crankshaft turns per cycle: NASA Glenn Research Center, "Four Stroke Internal Combustion Engine". Energy reaching the wheels (12–30% petrol; more than 77% electric): US DOE/EPA, fueleconomy.gov, "Where the Energy Goes" and "All-Electric Vehicles".
- Model T: launch 1 October 1908, end 26 May 1927, over 15 million built, 12½ hours to 93 minutes (Wikipedia, "Ford Model T" and "Assembly line", leads); moving line at Highland Park in 1913, $825 (1908) and $260 (1925), about 90 minutes, profit-sharing $5 day: Ford Motor Company, "The Moving Assembly Line".
- Interstate Highway Act (29 June 1956, 41,000 miles, $25 billion for 1957–69): US National Archives, Milestone Documents. Urban displacement and freeway revolts: Wikipedia, "Interstate Highway System" (lead). Road transport close to half of oil demand, fuel economy nearly doubled since 1975, 1.7 mb/d avoided: IEA, Global EV Outlook 2026, executive summary.
- Nader, Unsafe at Any Speed (1965), GM's investigators and apology, the 1972 NHTSA Corvair report: Wikipedia, "Unsafe at Any Speed" (lead). NHTSA founded by the 1970 Highway Safety Act; frontal crash tests from 1978; 5-Star ratings from 1993: NHTSA, "5-Star Safety Ratings" history. Euro NCAP first results 1997: Wikipedia (lead).
- Bohlin's patent (US 3,043,625, filed 1959, granted 10 July 1962): Google Patents. Volvo standard fitting in 1959: Wikipedia, "Seat belt" (lead). Seat belt effectiveness (45%; 60% in light trucks), 91.3% use in 2025, unbelted share of 2024 deaths: NHTSA, "Seat Belts".
- US road deaths and rates, 1899–2024 (including 1921, 1966, 1972, 2024, and the 4,036,087 total): NHTSA, "Motor Vehicle Traffic Fatalities and Fatality Rates, 1899–2024" (Traffic Safety Facts, June 2026). 2025 early estimate (36,640; 1.10): NHTSA press release, 1 April 2026.
- Global road deaths (1.16 million), ages 5–29, 92% in low- and middle-income countries with about 60% of vehicles, vulnerable road users, 2030 halving target: WHO fact sheet "Road traffic injuries" (20 July 2026).
- Leaded petrol (Midgley, 9 December 1921; "Ethyl"; 1924 deaths; Patterson's ice cores and campaign; US ban from January 1996; lead–crime debate): Wikipedia, "Tetraethyllead", "Clair Cameron Patterson" and "Gasoline" (leads). Introduction in 1922, end in Algeria in July 2021, more than one million premature deaths a year prevented, $2.45 trillion: UN News, 30 August 2021.
- Catalytic converters (1975 model year, three-way converters for 1981): Wikipedia, "Catalytic converter" (lead).
- Dieselgate: notice of violation 18 September 2015, defeat device, up to 40 times the NOx standard, about 590,000 US vehicles, 11 January 2017 guilty plea and $2.8 billion criminal penalty: US EPA, "Learn About Volkswagen Violations". WVU/ICCT discovery and the 11 million worldwide figure: Wikipedia, "Volkswagen emissions scandal" (lead).
- SAE levels as described by NHTSA, NHTSA's statement that no vehicle for sale is fully automated: NHTSA, "Automated Vehicles for Safety" (July 2026). J3016 history (2014; April 2021 revision): Wikipedia, "Self-driving car" (lead), and SAE's J3016_202104 listing. Tesla "Full Self-Driving (Supervised)" as Level 2: Wikipedia, "Tesla Autopilot" (lead).
- Waymo cities and "20M+ rides served": waymo.com, read 31 October 2026. 271.3 million rider-only miles through June 2026 and crash comparisons: Waymo Safety Impact hub. Tesla Cybercab audit query: NHTSA press release, 4 September 2026. Model Y ADAS tests: NHTSA press release, 7 May 2026. Cruise suspension: California DMV statement, 24 October 2023. GM ending Cruise robotaxi funding: GM news release, 10 December 2024.
- Electric car sales, shares, 2026 forecast, Q1 2026, US tax credits and Q4 2025 fall, Norway, Chinese makers' and China's production shares, 70% of Chinese BEVs cheaper than average conventional car, 70% less energy per kilometre, running-cost savings, 11 EVs per public charger, Chinese home-charger access: IEA, Global EV Outlook 2026 (20 May 2026): executive summary, "Trends in electric cars" and "Electric vehicle charging". Lifecycle emissions (about half; India about 20% lower): IEA, Global EV Outlook 2024, "Outlook for emissions reductions".
- Euro 7 and tyre and brake emissions: European Commission, "Emissions in the automotive sector".
- Computed in Python: Model T line speed-up (12.5 h ÷ 1.55 h ≈ 8); fall in the US death rate per mile (5.50 to 1.19 is −78%; to 1.10 is −80%); vehicle miles 1966 to 2024 (925,899 to 3,294,031 million, ×3.6).
Go deeper
- 🌐 IEA: Global EV Outlook 2026 — the most complete yearly picture of electric vehicles worldwide, free to read online
- 🌐 NHTSA: Automated Vehicles for Safety — the levels of automation explained by the US safety regulator, with what you can and can't buy
- 🌐 WHO: Road traffic injuries — the global picture, and what actually reduces deaths
- 🌐 US Department of Energy: The History of the Electric Car — a short, readable timeline from the 1830s onward
- 📘 Unsafe at Any Speed — Ralph Nader (the 1965 book that started modern car safety; read it alongside its critics)
- 📘 The Power Broker — Robert A. Caro (how one man's roads reshaped New York; the classic account of what highways did to a city)
- 📘 Lead Wars — Gerald Markowitz and David Rosner (the history of how the dangers of lead were fought over, petrol included)


Batteries
---------
A battery is a controlled chemical reaction that you can carry in your pocket and switch on and off. Here is how it works, who worked it out, why the one in your phone wears out, why car batteries got about ten times cheaper, and what digging up the ingredients costs.

Follows: How Electricity Actually Reaches Your Wall and Your Utility Bills, Decoded
  

  In 1986, a Japanese chemist named Akira Yoshino took his new battery to a facility built for testing explosives and dropped a large piece of iron on it. Nothing happened. Then he did the same to a battery made with pure lithium metal, and it blew up violently. He later called the quiet one "the moment when the lithium-ion battery was born". Five years later, his kind of battery went on sale. In 2025, the world installed about 1.2 terawatt-hours of batteries in electric vehicles alone: enough, at the arithmetic level, to fill some 78 billion phone batteries.

  The previous piece in this series, Cars, ended with the battery as the heart of the electric car. This one opens it up. It picks up where How Electricity Actually Reaches Your Wall left off: that piece explains current, voltage and why your charger turns AC into DC. Here we look at where the electricity in a battery comes from in the first place.

  
    In this deep dive
    
      From a pile of discs to the car battery
      How a battery works, and how lithium-ion was born
      Charging, and why batteries wear out (with a calculator)
      Why batteries got about ten times cheaper, and the chemistries racing now
      Digging it up: lithium, cobalt and the supply chain
      Recycling, grid storage, and the state of affairs in late 2026
    
  

  Part 1 — From a pile of discs to the car battery

  Frogs' legs and a pile of discs

  The word "battery" is older than the battery. Benjamin Franklin used it in 1749 for a row of linked Leyden jars, glass jars that could store a jolt of static electricity. He borrowed it from the army, where a battery is a group of guns firing together. A Leyden jar could give one sharp spark. It could not give a steady current.

  The steady current came out of a scientific quarrel. In the 1780s the Italian physician Luigi Galvani found that a dissected frog's leg twitched when it was touched by two different metals. He thought the electricity came from the frog: "animal electricity". His countryman Alessandro Volta, a physics professor at Pavia, suspected the opposite, that the electricity came from the two metals, and that the frog was just a wet, salty conductor. In 1794 Volta showed that two metals and a cloth soaked in salt water produce a current with no frog at all.

  Then he made it stronger by stacking it. He piled up discs: copper (or silver), then zinc, then cardboard or cloth soaked in brine, then copper, zinc, brine again, over and over. Join a wire from the top of the pile to the bottom and a steady current flows. He described it in a letter dated 20 March 1800 to Sir Joseph Banks, the president of the Royal Society in London, which published it that year as "On the electricity excited by the mere contact of conducting substances of different kinds".

  Volta's pile was the first device that could make electricity flow steadily, on demand. Every battery since is a variation on it.

  It was a scientific earthquake. That same year, two English chemists, William Nicholson and Anthony Carlisle, used a pile to split water into hydrogen and oxygen. Within eight years Humphry Davy had used bigger piles to isolate sodium, potassium, calcium and several other elements for the first time. The unit of electrical push, the volt, is named after Volta.

  There was one honest irony. Volta was right that you needed two different metals, but wrong about why. He thought simple contact between the metals was the source. In fact, as chemists later showed, the energy comes from a chemical reaction: the zinc is slowly eaten away. That is the first lesson of every battery. A battery doesn't store electricity. It stores chemicals that want to react, and it makes them react through a wire.

  Volta's pile had a serious practical problem: its voltage was unsteady and it could not keep up a strong current for long. In 1836 the British chemist John Frederic Daniell built a cell that gave a much steadier voltage, and it became the workhorse that powered the electric telegraph. But all these batteries had something in common. When the chemicals were used up, the battery was finished.

  Batteries you can recharge: lead and nickel

  In 1859, a French physicist named Gaston Planté rolled two sheets of lead into a spiral, separated by cloth, and soaked them in sulfuric acid. When he pushed a current through it, the chemistry ran backwards and the battery could be used again. It was the first rechargeable battery. In 1881, Camille Alphonse Faure made it practical to mass-produce, by pressing a paste of lead oxide into a lead grid.

  The lead-acid battery is heavy and stores little energy for its weight, but it is cheap, rugged, and can deliver a huge burst of current. That is exactly what you need to turn over a cold engine, which is why, more than 160 years later, the battery under the bonnet of a petrol car is still lead-acid. Even many electric cars keep a small one to run the lights and electronics. Lead-acid batteries were also what the electric cars of 1900 ran on, and their weight is one reason those cars lost.

  Nickel came next. In 1899, the Swedish inventor Waldemar Jungner made the first nickel-cadmium battery, and Thomas Edison soon developed a nickel-iron version in the United States. Nickel-cadmium cells powered cordless phones, power tools and the first portable electronics. They were later replaced by nickel-metal hydride, which avoided toxic cadmium; consumer cells went on sale in 1989. Nickel-metal hydride batteries are still in the AA rechargeables in many drawers, and they powered the first generation of hybrid cars, including the Toyota Prius.

  
    BatteryFirst madeVolts per cell (approx.)Still used for
    Voltaic pile1800About 1 per pair of discsMuseums
    Lead-acid18592Starting petrol cars; backup power
    Nickel-cadmium18991.2Some industrial and emergency uses
    Nickel-metal hydrideConsumer cells 19891.2AA rechargeables; many hybrid cars
    Lithium-ionOn sale 1991About 3.2 to 3.7Phones, laptops, electric cars, grid storage
  

  Look at the voltage column. A lithium-ion cell gives roughly three times the push of a nickel one. That, and its light weight, is why it took over. To see why, we need to look inside.

  Part 2 — How a battery works, and how lithium-ion was born

  The four parts of every battery

  Every battery, from Volta's pile to the one in your phone, has the same four parts:

  
    The anode, the negative end. When the battery is in use, it is made of something that wants to give away electrons.
    The cathode, the positive end. It is made of something that wants to take electrons.
    The electrolyte between them, a liquid, paste or solid that lets charged atoms (ions) move through it, but not electrons.
    The separator, a thin porous sheet that keeps the anode and cathode from touching, because if they touch, the reaction happens inside in a rush of heat: a short circuit.
  

  Here is the trick that makes a battery useful. The anode and cathode want to react, but the electrolyte only lets part of the reaction through. The positive ions can cross through the electrolyte. The electrons can't. So the electrons are forced to go the long way round, out of the battery, through your wire and your phone, and back in at the other end. That flow of electrons through the outside circuit is the current that does the work.

  
    Inside a lithium-ion cell, while it powers something
    
      
        
        
      
      
      
      
      
      Anode (−)
      Cathode (+)
      separator
      graphite
      metal oxide
      
      
      Li⁺ ions
      electrolyte
      
      
      
      
      ✱
      electrons flow through the wire and the bulb
    
    While the cell powers something, lithium ions drift through the electrolyte from the anode to the cathode, and electrons go the long way round through the circuit. Charging pushes both back the other way.
  

  Ions inside, electrons outside. If you remember one sentence from this piece, make it that one. Inside the battery, charged atoms move. Outside the battery, electrons move. The two flows always balance, which is why a battery stops working the moment you break the circuit: the electrons can't get round, so the ions stop too.

  Voltage is how strongly the anode material wants to give up electrons compared with how strongly the cathode wants them. Pick materials that are very different, and you get a bigger push. Lithium is the extreme case. It is the lightest metal, and it gives up its single outer electron more readily than almost anything else. That is why lithium batteries reach well over three volts per cell where nickel ones manage 1.2.

  Charging is the reaction run backwards. A charger pushes electrons back into the anode, and the ions follow them back across. In a lead-acid or nickel battery, that means rebuilding the electrodes chemically every time, and they slowly crumble. The lithium-ion battery found a gentler trick, and it is the reason it lasts as long as it does.

  The lithium-ion story

  The lithium-ion battery began with an oil crisis. In the 1970s, worried that oil would run out, the oil giant Exxon hired leading energy researchers and gave them freedom to work on anything that wasn't petroleum. One of them was a young British chemist, M. Stanley Whittingham, who joined Exxon in 1972.

  Whittingham's idea: houses for ions. He was studying materials made of layers with gaps between them, like a stack of paper with room between the sheets, where ions can slip in and out without destroying the structure. Chemists call it intercalation. He found that titanium disulphide could take in lithium ions this way, and used it as a cathode, with lithium metal as the anode. The battery gave just over two volts and could be recharged. After a meeting at Exxon's headquarters that lasted about fifteen minutes, management quickly decided to develop it, and in 1976 Whittingham announced it. It was first made on a small scale for a Swiss clockmaker's solar-powered watches.

  Then came the fires. Each time the battery was charged, thin whiskers of lithium metal grew from the anode, called dendrites. When one reached the other side, it short-circuited the cell, which could burst into flames. According to the Nobel committee's own account, the fire brigade was called so often that it threatened to bill the laboratory for the special chemicals used to fight lithium fires. When oil prices fell in the early 1980s, Exxon gave up.

  Goodenough's idea: a better cathode. John Goodenough, an American physicist then at Oxford, reasoned that a metal oxide would give a higher voltage than a sulphide. His group searched systematically and in 1980 published lithium cobalt oxide, which gave about four volts, nearly double. He also realised that a battery didn't have to be built charged: you could build it empty and charge it afterwards, which made it much safer to make. His co-author on the paper was Koichi Mizushima.

  Yoshino's idea: no lithium metal at all. In Japan, where electronics companies were desperate for lightweight rechargeable batteries for video cameras, cordless phones and computers, Akira Yoshino at the Asahi Kasei company took Goodenough's cathode and paired it with an anode made of petroleum coke, a carbon material that could also house lithium ions. Now there was no lithium metal anywhere in the battery, only lithium ions shuttling between two "houses". In 1985 he had the first commercially viable lithium-ion battery, and in 1986 came the falling-iron test.

  The lithium-ion battery works because nothing has to be rebuilt each cycle. Ions just move from one house to another, and back.

  In 1991, Sony began selling the first lithium-ion batteries. Phones shrank, laptops became truly portable, and, eventually, cars could run on them.

  The Nobel Prize, and the people it left out

  On 9 October 2019, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to Goodenough, Whittingham and Yoshino "for the development of lithium-ion batteries". Goodenough was 97, the oldest person ever to receive a Nobel Prize. He died on 25 June 2023, a month short of his 101st birthday. He had kept working into his nineties, and his group also introduced lithium iron phosphate as a cathode in the 1990s, which matters later in this story.

  A Nobel Prize can go to at most three people, and the lithium-ion battery had more parents than that. Fair accounts name, among others:
  
    WhoContribution
    Koichi MizushimaCo-author of the 1980 lithium cobalt oxide paper
    Rachid YazamiShowed in 1980 that lithium could move reversibly in and out of graphite, the material used in most anodes today
    Yoshio NishiLed the Sony team that turned the design into a product in 1991
  
  In 2014, Goodenough, Nishi, Yazami and Yoshino shared the US National Academy of Engineering's Draper Prize for the lithium-ion battery. Big inventions are usually like this: a relay, not a single runner.

  Part 3 — Charging, and why batteries wear out

  Energy and power

  Two numbers describe any battery, and confusing them causes most of the muddle about electric cars.

  Energy, in kilowatt-hours (kWh), is how much the battery holds: the size of the tank. Power, in kilowatts (kW), is how fast energy flows in or out: the width of the hose. If you have read Your Utility Bills, Decoded, it is the same pair: a 2 kW heater left on for three hours uses 6 kWh. A phone battery labelled 4,000 mAh at 3.85 volts holds about 15 watt-hours (4 Ah × 3.85 V). A 70 kWh car battery holds as much as about 4,500 of those phones.

  Battery size (kWh) is the tank. Charger power (kW) is the hose. Time is tank ÷ hose.

  The calculator below lets you try it. Pick a battery, a car and a charger, and set how full the battery starts and ends.

  
    Interactive · Charge an electric car

    Battery size
      
      
        China avg (just under 60)
        EU avg (about 70)
        US avg (about 90)
      
    

    Car (energy use)
      
        Efficient saloon
        Mid-size SUV
        Big pickup
      
    

    Charger
      
      
        Wall socket (US Level 1)
        Home charger
        Public fast
        Ultra-fast
      
    

    Start at
      
    
    Charge to
      
    

    
      Electricity price:
      
      ¢ per kWh
      US home avg
      UK avg
    

    
      –Energy into battery
      –Range added
      –Time (at full power)
      –Cost
    
    
    Assumptions. Battery sizes: IEA averages for new battery electric cars in 2025. Energy use: US EPA ratings on fueleconomy.gov (2026 Tesla Model 3 Standard RWD, 24.3 kWh/100 mi; 2025 Chevrolet Equinox EV FWD, 31.2; 2025 Ford F-150 Lightning extended range, 47.9), converted to kWh/100 km; real use varies with speed, cold and heating. Chargers: US DOE figures (Level 1 about 1.9 kW; typical home Level 2 7.2 kW), the IEA's 2025 global average public charging speed (50 kW), and the 250 kW class the IEA describes as ultra-fast. Charging losses: 12%, inside the US DOE's range of 84–93% charging efficiency; they affect time and cost, not range. Prices: US average residential, August 2026 (EIA); UK average, July 2026 (Centre for Sustainable Energy). Time assumes constant power; real charging slows as the battery fills (the DOE notes power "varies by vehicle and battery state of charge"), and many cars accept less than an ultra-fast charger can give.
  

  Three experiments are worth doing. First, keep everything the same and switch from the efficient saloon to the big pickup: the same 50 kWh of charge takes you about half as far, because the pickup uses twice the energy per kilometre. Second, compare the wall socket with the home charger for an overnight top-up from 20% to 80%: the slow one takes over a day for a big battery. Third, look at the cost, then type in your own electricity price from your bill (Your Utility Bills, Decoded shows where to find it). If you charge in public, use the price per kWh the charging company shows in its app, which is often different from what you pay at home.

  Why your battery wears out

  Your two-year-old phone doesn't last the day the way it used to. Nothing is broken. A lithium-ion battery loses capacity in two ways: a little every time it is cycled, and a little simply with the passing of time, especially when it sits fully charged or hot. Battery scientists call these cycle ageing and calendar ageing.

  The main culprits, in plain words:
  
    A crust on the anode. The first time a battery is charged, a thin film forms on the graphite anode where it meets the electrolyte, called the solid electrolyte interphase. It protects the anode, but it keeps slowly thickening, and every lithium ion trapped in it is one that can no longer carry charge. It grows faster when the battery is hot and when it sits at a high charge.
    Lithium plating. Charge too fast, especially in the cold, and the ions can't slip into the graphite quickly enough. Some pile up on its surface as metallic lithium instead, the same problem that made Whittingham's battery catch fire. That lithium is lost, and in bad cases it can cause short circuits.
    Cracking. Electrode particles swell and shrink slightly every time lithium moves in and out. Over thousands of cycles they crack, and bits lose contact with the rest of the electrode.
  

  Real-world data matches the chemistry. Geotab, a vehicle-telematics company, analysed more than 22,700 electric vehicles across 21 models and reported in April 2026 that their batteries lost on average 2.3% of their capacity a year, which projects to about 82% left after eight years. Cars that used high-power fast chargers most often degraded fastest, about 3% a year; those that rarely fast-charged lost about 1.5%. Cars in hot climates lost about 0.4 percentage points a year more than those in mild ones. The US Department of Energy's Alternative Fuels Data Center notes that studies suggest an EV battery may still hold at least 70% of its original capacity at the end of the car's life.

  
    What speeds up wearWhyWhat helps
    HeatFaster side reactions, thicker anode crustKeep phones out of hot cars and off pillows while charging
    Sitting at 100% for longHighest stress on both electrodesUse the "optimised" or 80% charging setting if your device has one
    Frequent very fast chargingLithium plating, heatFast-charge when you need to, not every day
    Cold-weather fast chargingIons move slowly; plating riskLet the car warm the battery first (many do this automatically)
  

  Safety belongs here too. Most lithium-ion batteries use a liquid electrolyte made with flammable organic solvents. If a cell is crushed, overcharged or badly made, it can overheat and set off its neighbours, a chain reaction called thermal runaway, and the smoke is toxic. Such fires are hard to put out, which is one reason airlines have special rules for carrying spare batteries.

  Part 4 — Why batteries got about ten times cheaper, and the chemistries racing now

  Three independent measurements tell the same story.

  
    SourceWhat it measuredFinding
    Ziegler and Trancik, MIT (2021)Lithium-ion cell costs since 1991Fell about 97%
    US Department of EnergyEV battery packs, in 2022 dollars$1,355 per kWh in 2008 to $153 in 2022, down 89%
    BloombergNEF survey (Dec 2025)Average pack price, all uses$108 per kWh in 2025, down 8% in a year; EV packs $99; stationary storage $70
  

  The three use different methods, so don't splice them into one line. But together they say that a battery that would have cost well over a thousand dollars per kWh when the Tesla Roadster appeared costs around a hundred today.

  Why? Partly better chemistry: more energy packed into each cell, and cheaper materials, especially the shift to iron phosphate cathodes described below. Mostly, though, it is the ordinary magic of making something in enormous and growing quantities. Factories got bigger, processes got better, scrap rates fell, and suppliers competed. Economists call this a learning curve: costs fall by a roughly steady fraction every time cumulative production doubles. Solar panels followed the same path.

  The battery didn't get cheap because of one breakthrough. It got cheap because the world built a lot of them, and kept building more.

  There is a catch worth knowing. The International Energy Agency's Global EV Outlook 2026 warns that today's record-low prices are "partly sustained by losses" further up the supply chain: many makers of cathode materials, the most important ingredient in a cell's cost, are operating at a loss while still expanding. Lithium prices in early 2026 were more than double their level a year earlier, and cobalt prices doubled. Prices might not keep falling in a straight line.

  LFP, NMC, sodium-ion and solid-state

  "Lithium-ion" is a family, usually named after the cathode. Two members dominate cars today.

  NMC (nickel manganese cobalt oxide) is the descendant of Goodenough's cobalt oxide. It packs the most energy into the least weight, which is why it suits long-range and premium cars. Modern versions use much more nickel and less cobalt than earlier ones.

  LFP (lithium iron phosphate) holds less energy for its weight but uses cheap, abundant iron and phosphate, contains no nickel or cobalt, is more resistant to overheating and lasts many cycles. For years it was seen as second-best. Then Chinese manufacturers improved how cells are packed together, and LFP became good enough for most cars. The IEA says LFP made up over 55% of the batteries put into electric vehicles worldwide in 2025, and over 90% of the batteries installed for grid storage. BloombergNEF's 2025 survey found LFP packs averaging $81 per kWh against $128 for NMC.

  
    ChemistryBest energy per kg, latest cells (IEA 2026)Status in 2026
    NMC (nickel-rich)Up to about 265 Wh/kgMainstream for longer-range cars, especially outside China
    LFPUp to about 205 Wh/kgMainstream: over half of EV batteries; nearly all grid storage
    Sodium-ionUp to about 175 Wh/kgEarly commercial; first car in China in late 2023; being scaled up
    All-solid-statePromised higher; not yet shown in real usePrototypes and pilot production; first cars planned for 2027–28
  

  Sodium-ion swaps lithium for sodium, which is in ordinary salt and found almost everywhere. It stores less energy, so the IEA puts the range of an average SUV with a sodium-ion battery at up to about 350 km, against 400–600 km for lithium-ion. But it works much better in the cold: the latest cells keep about 90% of their capacity at −40 °C. The IEA expects it to suit short-range city cars, urban delivery vans, two- and three-wheelers, forklifts and grid storage. In 2026, sodium-ion factories had just over 1% of the capacity of lithium-ion ones.

  Solid-state batteries replace the flammable liquid electrolyte with a solid. The promise is more range and better safety, and possibly a lithium-metal anode at last, the dream Whittingham's fires ended. The IEA's verdict in 2026 is blunt: those advantages "have not yet been demonstrated in real-world applications." Some simpler semi-solid designs are already on sale. Toyota has said it plans its first all-solid-state car by 2028; BYD plans a first car from 2027 and mass production from 2030. Treat any company's date as a plan, not a fact: battery announcements have a long history of slipping.

  Part 5 — Digging it up: lithium, cobalt and the supply chain

  A battery is a mining product before it is a chemistry product.

  Lithium. According to the US Geological Survey, the world (outside the United States, whose figures are confidential) mined about 290,000 tonnes of lithium in 2025, 31% more than the year before. Australia mined the most, about 92,000 tonnes, mainly from hard rock; then China (62,000), Chile (56,000), Zimbabwe (28,000) and Argentina (23,000). Batteries took an estimated 88% of all lithium. In Chile and Argentina, lithium comes from brine pumped up from under salt flats and concentrated in huge evaporation ponds in some of the driest places on Earth, and the use of water there has brought disputes with local and Indigenous communities, as well as negotiated benefit-sharing deals. Lithium's price is famously wild: the USGS's average US contract price for battery-grade lithium carbonate went from $63,700 a tonne in 2022 to $9,000 in 2025.

  Cobalt. The Democratic Republic of the Congo (DRC) produced an estimated 73% of the world's mined cobalt in 2025, the USGS says, followed by Indonesia with 14%. Much of it comes from large industrial copper-cobalt mines; some is dug by hand by independent "artisanal" miners. In 2016, Amnesty International and the Congolese group Afrewatch documented children as young as seven working in such mines, citing a UNICEF estimate that about 40,000 children worked in mines across southern DRC in 2014, and traced that cobalt through traders and a Chinese processor to battery-component makers that supplied well-known electronics and car brands. Of 16 multinationals Amnesty contacted, none gave enough detail to verify where their cobalt came from. Companies have since adopted auditing and sourcing programmes, and campaigners continue to argue about how much has changed on the ground.

  Two developments have changed the picture since then. First, chemistry: LFP batteries use no cobalt at all, and the nickel-rich NMC types use much less than older ones. Second, politics: in February 2025 the DRC temporarily banned cobalt exports to push prices up, then from 16 October replaced the ban with export quotas, and cobalt prices roughly doubled. The USGS reports that the quotas allow up to 96,600 tonnes a year in 2026 and 2027, far below the roughly 230,000 tonnes the country mined in 2025. Batteries are now the leading use of cobalt, so these decisions in Kinshasa reach every battery maker.

  Processing is the real choke point. Mining is spread across several continents; refining and manufacturing are not. The IEA reports that China accounted for over 80% of battery-cell production in 2025 and even higher shares of the active materials inside them, including almost all LFP cathode material. Chinese, Korean and Japanese companies supply nearly all the battery cells in the world. In October 2025 China announced broader export controls on battery materials, including cathode and anode materials, then paused them for one year. Europe and the United States are building their own factories; the EU and the US each hold about 6–7% of global manufacturing capacity, the IEA says, against China's more than 80%.

  
    MaterialMain sources (2025)Main concern
    LithiumAustralia, China, Chile, Zimbabwe, ArgentinaWater in South American salt flats; price swings
    CobaltDRC (73%), Indonesia (14%)Artisanal mining, child labour, export quotas
    Processing and cellsChina (over 80% of cells)Concentration, export controls
  

  An honest comparison runs both ways. A petrol car burns its fuel once and needs more every week, from an oil industry with its own history of spills, wars and pollution. A battery's metals stay in the battery and can, in principle, be recovered. That doesn't excuse a child in a mine, but it does mean the long-run answer is better mining and much more recycling, not pretending either car is free of extraction.

  Part 6 — Recycling, grid storage, and the state of affairs in late 2026

  A second life, then recycling
  A battery too worn for a car is not necessarily finished. As Part 3 noted, the US Department of Energy says an EV battery may still hold at least 70% of its capacity when the car is retired, and it suggests such batteries could go on to serve for ten years or more in less demanding jobs, such as storing energy for a building or the grid. The IEA says reuse remains challenging in practice: second-life systems need safety checks and warranties, nobody knows exactly how long a used battery will last, taking packs apart safely costs money, and new batteries keep getting cheaper. Selling, repairing and refurbishing used electric cars, by contrast, is a fast-growing business.
  Battery recycling starts by shredding cells into "black mass", a powder of the valuable metals. The metals are then recovered by smelting (high heat, large scale, energy-hungry), by dissolving them in chemicals (hydrometallurgy, cheaper to build), or by "direct recycling", a newer approach that tries to recover the cathode material intact instead of breaking it down into separate metals.

  The big constraint is time. Nearly all the EV batteries ever made are still in cars. The IEA estimates a lag of roughly 15 years between a surge in battery sales and a matching surge in worn-out batteries, so for now most recycling feedstock is factory scrap, and end-of-life batteries won't dominate until the mid-2030s. China hosts over 85% of global recycling capacity. There's also an economic twist: LFP and sodium-ion batteries contain fewer valuable metals, so recycling them may need rules that make makers responsible for their old batteries, as the EU's battery regulation now does, rather than relying on metal prices alone.

  Batteries on the grid
  The same cells that move cars can steady power grids, soaking up cheap solar power at midday and releasing it in the evening. BloombergNEF found that battery packs for stationary storage fell 45% in a year to $70 per kWh in 2025, making them the cheapest battery segment for the first time. In the United States, the Energy Information Administration reported that a record 15 gigawatts of utility-scale battery storage was added in 2025, with 24 gigawatts planned for 2026, about 28% of all the new generating capacity planned for the year, and more than half of the batteries in Texas. (For how that power then reaches your home, see How Electricity Actually Reaches Your Wall.)

  The state of affairs, late 2026
  
    Where things stand (latest data)
    ScaleEV batteries installed worldwide reached 1.2 TWh in 2025, more than seven times the 2020 figure; EVs were still more than 70% of all battery deployment. Global factory capacity passed 4 TWh a year, over 80% of it in China (IEA, May 2026).
    PriceAbout $108 per kWh on average and $99 for EV packs (BloombergNEF, December 2025). BloombergNEF expected a further fall in 2026; rising lithium and cobalt prices could push the other way.
    ChemistryLFP leads, NMC holds the long-range end, sodium-ion is entering the market, all-solid-state is still promised.
    ChargingThe first 1,000-volt car models came out in 2025, and makers have announced charging times under ten minutes, but fewer than 5% of electric cars on the road can use chargers above 250 kW (IEA).
  

  Open questions
  Will costs keep falling? Optimists point to the learning curve and new chemistries. Sceptics point to loss-making suppliers, volatile metals and tariffs. Both have evidence.
  Can supply chains be diversified without making batteries dearer? Building factories outside China improves security but, for now, costs more; BloombergNEF found pack prices in North America and Europe 44% and 56% above China's.
  Can mining be made decent? Better sourcing rules, cobalt-free chemistries and recycling all help. None of them is yet enough on its own.
  Is the next breakthrough sodium, solid-state, or something else? History suggests caution: lithium-ion took about twenty years to go from Whittingham's lab to Sony's shelves, and decades more to become cheap.

  
    The takeaway
    A battery stores chemicals that want to react and makes them react through a wire: ions travel inside through the electrolyte, electrons travel outside through your device. Volta's 1800 pile began it, lead-acid (1859) made it rechargeable, and nickel batteries carried it to the first portable electronics. The lithium-ion battery came from three ideas, Whittingham's intercalation, Goodenough's 4-volt cobalt oxide cathode and Yoshino's carbon anode, honoured with the 2019 Nobel Prize, with others such as Yazami and Nishi also essential. Batteries wear out mainly through a growing crust on the anode, lithium plating and cracking, made worse by heat, full charge and frequent fast charging. Prices fell about 90% through mass production; LFP now leads. The costs are real too: cobalt from the DRC, lithium brine in dry salt flats, and processing concentrated in China.
    Try this: find your phone's battery-health screen (on iPhone it is under Settings › Battery; on many Android phones, under Battery or Device care in Settings) and note the percentage. If your phone or car offers an 80% charging limit or "optimised charging", turn it on. Then use the calculator above with your own electricity price to work out what a full charge of an electric car would cost where you live.
  

  
    Sources & Further Reading
    Sources. Nobel Prize in Chemistry 2019: press release (9 October 2019) and "Popular information" (Whittingham at Exxon from 1972, TiS₂, just over 2 V, dendrites and fires, 1976 announcement; Goodenough's 1980 cobalt oxide at about 4 V and building cells uncharged; Yoshino's 1985 petroleum-coke anode and 1986 drop test, with his quotation; first sales in 1991; lead battery of 1859; lithium's discovery in 1817; Goodenough's iron phosphate work); NobelPrize.org laureate facts (Goodenough, born 25 July 1922, died 25 June 2023) and "Nobel Prize facts" (oldest laureate, aged 97). Volta's letter: Philosophical Transactions of the Royal Society 90 (1800), 403–431, metadata via Crossref. IEA, Global EV Outlook 2026 (20 May 2026): battery chapter (1.2 TWh deployment, battery sizes, 8% price fall, LFP over 55% and over 90% of stationary storage, LFP over 40% cheaper than NMC, cathode producers' losses, lithium and cobalt prices, DRC export ban and quotas, 4 TWh capacity, over 80% in China, Sony's 1991 cylindrical cells, Roadster and Leaf, sodium-ion and solid-state status and figures, recycling lag and China's 85% of capacity, EU battery regulation) and executive summary (1,000-volt models, under 5% of cars able to use chargers above 250 kW); charging chapter (50 kW average public speed). BloombergNEF, "Lithium-Ion Battery Pack Prices Fall to $108 Per Kilowatt-Hour" (9 December 2025). US DOE Vehicle Technologies Office, Fact of the Week #1272 ($1,355 in 2008 to $153 in 2022, 2022 dollars). MIT News, 23 March 2021, on Ziegler and Trancik (97% since 1991). US DOE Alternative Fuels Data Center: "Batteries for Electric Vehicles" (at least 70% capacity at end of life; recycling methods; lead-acid and nickel-metal hydride uses) and "Developing Infrastructure to Charge Electric Vehicles" (charger levels and powers). fueleconomy.gov: EPA ratings via its web service and the 84–93% charging-efficiency range. Geotab, "EV battery health" study (28 April 2026). USGS Mineral Commodity Summaries 2026: Cobalt and Lithium chapters. Amnesty International, news release of 19 January 2016 on This is what we die for. US EIA, Today in Energy (20 February 2026) and Electric Power Monthly Table 5.6.A (August 2026). Centre for Sustainable Energy, "How much electricity am I using?" (July 2026 price). Wikipedia articles on the voltaic pile, Galvani, Daniell cell, Planté, Faure, Jungner, nickel-metal hydride batteries, lithium-ion battery history and degradation, Yazami, Mizushima and lithium mining were used for leads and cross-checked where possible. The calculator's formulas and every number it shows were run and checked in node; the phone-battery and pickup comparisons were computed in Python.
    
      🌐 Nobel Prize in Chemistry 2019: popular information — the Nobel committee's own illustrated telling of the lithium-ion story, written for non-chemists
      🌐 IEA: Global EV Outlook 2026 — the battery chapter is the best free overview of chemistries, prices, supply chains and recycling
      🌐 US DOE: Batteries for Electric Vehicles — a short, sober page on battery types, lifetimes and recycling
      🌐 Amnesty International: child labour behind batteries (2016) — the report that put cobalt mining on the world's agenda; read it as the starting point of a debate that has moved on since
      🌐 Edge et al., "Lithium ion battery degradation: what you need to know" (2021) — a scientific review for readers who want the ageing mechanisms in full
      📘 The Powerhouse: America, China, and the Great Battery War — Steve LeVine (the race to build a better battery, told through the labs and companies)
      📘 Cobalt Red — Siddharth Kara (a first-hand, harrowing account of artisanal mining in the DRC; an advocate's book, contested by some in the industry)


Nuclear Energy: The Promise, the Accidents, and the Comeback
------------------------------------------------------------
Born inside a bomb programme, sold as the power of the future, nearly killed by three accidents, and now being courted by the companies that run the world's data centres.

Follows: How a Nuclear Reactor Works (Physics 14) and Energy: The One Currency of the Universe (Physics 2)
In March 1979, the reactor known as Unit 2 at Three Mile Island in Pennsylvania partly melted down, and the American nuclear industry never fully recovered. Its twin, Unit 1, stood a few hundred metres away and kept running for another forty years. It shut in 2019 because it couldn't pay its way. Then, in September 2024, its owner signed a 20-year deal with Microsoft to bring it back to life, to feed the electricity demand of computing. It now has a new name and a target date of 2027.
That one site holds the whole story: the promise, the accident, the long decline and the unlikely comeback. This deep dive tells that story from the first chain reaction in 1942 to the state of things at the end of October 2026. It also sets out why intelligent people still disagree so strongly about it. The physics (what fission is, why a reactor can be steered, how Chernobyl's design failed) is in How a Nuclear Reactor Works; this piece is about history, money, risk and politics.
In this deep dive

From a chain reaction to Atoms for Peace (1942–1957)
The boom, the backlash, and Three Mile Island
Chernobyl, Fukushima, and how to count the dead
France and Germany: two countries, two answers
Waste: the problem that has to outlast us
The comeback: nuclear power in 2026
Fusion: the other nuclear
The strongest case for, and against

Part 1 — From a chain reaction to Atoms for Peace (1942–1957)
The idea came before the machine, and it came as a fear. Radioactivity, discovered in the 1890s, showed that atoms hold energy on a scale nothing in chemistry could match (see Marie Curie: Two Nobel Prizes, One Radical Idea). For forty years nobody knew how to release it on purpose. Then, in December 1938, Otto Hahn and Fritz Strassmann found that uranium hit by neutrons split in two, and within weeks Lise Meitner and Otto Frisch had explained it. Physicists saw at once that the split also released neutrons, which could set off a chain reaction. With war coming, the first use anyone imagined was a weapon.
The first chain reaction was achieved quietly. On 2 December 1942, under the west stands of Stagg Field, the University of Chicago's disused football stadium, Enrico Fermi's team pulled the last control rod out of a pile of about 45,000 graphite blocks studded with uranium. At 3:25 in the afternoon, the neutron count kept rising on its own. Chicago Pile-1 was critical. It produced about half a watt, too little to boil a kettle in a week, and the first run lasted only minutes. Forty-nine people watched. Eugene Wigner had brought a bottle of Chianti, which they drank from paper cups and signed.

"The Italian navigator has landed in the New World." Arthur Compton's coded phone report to James Conant that the chain reaction had worked.

Chicago Pile-1 was a step in the Manhattan Project, and the project is the original sin that nuclear power has never fully shaken off. At its peak it employed about 130,000 people and had spent about $2.2 billion by the end of the war. It built secret towns: Oak Ridge in Tennessee to enrich uranium, Hanford in Washington to make plutonium, Los Alamos in New Mexico to design the bombs. The first test, Trinity, was on 16 July 1945. On 6 and 9 August, bombs destroyed Hiroshima and Nagasaki. Estimates of the deaths they caused range from about 150,000 to 246,000.
Two legacies of these years shaped everything after. Commercial reactors grew up inside military programmes built in secret and at any cost. And the gap between power and weapons is narrower than people would like. A reactor runs on uranium enriched to a few per cent uranium-235, a bomb needs it very highly enriched or needs plutonium, and the plants that make one can, with intent, make the other.
On 8 December 1953, President Dwight Eisenhower stood before the United Nations General Assembly in New York and tried to change the subject. With the superpowers racing to build hydrogen bombs, Eisenhower spoke of "the fearful atomic dilemma", and proposed that nuclear powers pool materials under an international agency for peaceful uses, above all electricity. His hope was that "the miraculous inventiveness of man shall not be dedicated to his death, but consecrated to his life." The speech led to the founding of the International Atomic Energy Agency and to a programme that shared reactor know-how and fuel with friendly countries.
It is still argued over. Supporters see the moment nuclear technology turned toward medicine, research and power. Critics point out that the United States went on exporting highly enriched uranium for research reactors (over 25 tons to 30 countries), now regarded as a proliferation risk. They note that Israel's and Pakistan's first reactors came out of the programme, and that it gave political cover while the American arsenal grew from about a thousand weapons to about twenty thousand during Eisenhower's presidency. Both readings are true at once.
Who was "first" with nuclear electricity depends on the definition.

Claim
Where and when
Size
The catch

First nuclear electricity
EBR-I, Idaho, December 1951
A trivial amount
An experiment, not a power station

First reactor to feed a grid
Obninsk, USSR, 27 June 1954
5 MW
Small; a showcase that ran until 2002

First full-scale power station
Calder Hall, England, opened by Queen Elizabeth II on 17 October 1956
Four 60 MW units
Its main job was making plutonium for British bombs

First full-scale plant purely for peaceful power
Shippingport, Pennsylvania, grid December 1957
60 MW
Built under Admiral Hyman Rickover from navy reactor designs

Shippingport shows how much of the industry came from the navy. Rickover's programme had developed the pressurised-water reactor to drive submarines. Its prototype started in Idaho in March 1953, and USS Nautilus was launched in 1954. That compact, water-cooled design was adapted for land, and it is still by far the most common type: about 316 of the world's roughly 440 operable reactors are pressurised-water reactors. The economist Robin Cowan argued in 1990 that this was a case of "technological lock-in": the reactor that won was the one the navy had already paid to develop, not necessarily the best design for a power grid.
Part 2 — The boom, the backlash, and Three Mile Island
For about twenty years, nuclear power grew faster than almost any energy technology before it. Westinghouse's Yankee Rowe plant (250 MW) started in 1960. By the end of the decade, utilities were ordering reactors of more than 1,000 MW, as big as the largest coal plants. World nuclear generation rose from about 79 terawatt-hours (TWh) in 1970 to about 2,000 TWh in 1990, a twenty-five-fold increase in twenty years. Its share of the world's electricity peaked at about 17.4% in 1996.
The mood of the boom had been set by one phrase. In September 1954, Lewis Strauss, chairman of the US Atomic Energy Commission, told science writers that "it is not too much to expect that our children will enjoy in their homes electrical energy too cheap to meter". The line became the industry's most quoted embarrassment. His son later claimed he had been thinking of fusion, then secret. A historian at the Nuclear Regulatory Commission, Thomas Wellock, found no evidence of that in Strauss's papers, and the rest of the speech was about fission. Either way, the promise set an expectation the industry never met.

In twenty years, nuclear generation grew twenty-five-fold. In the thirty years after that, its share of the world's electricity halved.

The oil shocks of the 1970s seemed to guarantee the boom. Oil-importing countries, France most dramatically, turned to reactors for energy security. But in the United States the boom was already faltering before the famous accident. Reactors were getting bigger, each one closer to a custom project, and costs and construction times grew with them. Many reactor orders from the 1970s were cancelled.
At the same time, a popular movement against nuclear power took shape. In West Germany, protests against a planned plant at Wyhl in the 1970s turned public opinion. The movement fed into what became the Green Party. Its objections were not only about accidents. They were about waste, about the link to weapons, and about a technology so complex and centralised that only states and giant firms could run it. In March 1979 Hollywood released The China Syndrome, a thriller about a cover-up at a reactor. Twelve days later, life imitated it.
Three Mile Island, 28 March 1979. At about 4 a.m., a fault stopped the main feedwater pumps of Unit 2. A relief valve opened as designed but then stuck open, and a light in the control room told the operators it had closed. Coolant kept escaping while the staff, misled by their instruments and lacking any reliable way to tell whether the core was covered with water, took actions that left it uncovered. Nearly half of the core melted.
The releases of radioactivity were very small. The US Nuclear Regulatory Commission's figure is that the roughly two million people around the plant received on average about 1 millirem above background. A chest X-ray is about 6 millirem, and the area's natural background is about 100 to 125 millirem a year. More than a dozen studies have found no detectable health effects. The panic was real, though. The governor advised pregnant women and preschool children within five miles to leave, and about 140,000 people left the area within days, though a survey that April found 98% of them back home within three weeks. The clean-up took until 1993 and cost about $1 billion.

At Three Mile Island no one died. The industry nearly did.

The accident changed regulation more than health. The NRC put resident inspectors at every plant, overhauled operator training and control-room design, and created a 24-hour operations centre. The industry set up its own body to share lessons and police itself.
Part 3 — Chernobyl, Fukushima, and how to count the dead
Chernobyl, 26 April 1986. At unit 4 of the Chernobyl plant in Soviet Ukraine, a badly run safety test on a reactor design with dangerous built-in flaws (explained in How a Nuclear Reactor Works) caused a power surge and an explosion that blew the reactor open, sending radioactive material across much of Europe. Two workers died that night. Of the 134 people confirmed with acute radiation sickness, 28 died within weeks. Nineteen more of them died between 1987 and 2004, though not necessarily from radiation.
About 116,000 people were evacuated by mid-May 1986, and 220,000 more were resettled later. The exclusion zone grew from a 30-kilometre circle to about 4,300 square kilometres. The most clearly established health effect came from radioactive iodine in milk: children drank it, and thousands of them later developed thyroid cancer. UNSCEAR, the UN's scientific committee on radiation, put the number at more than 6,000 cases by 2005, of which 15 had been fatal. Thyroid cancer is usually treatable, which is why the deaths are few next to the cases.
A steel arch, the New Safe Confinement, was slid over the ruined reactor in November 2016. On 14 February 2025 a drone struck it and breached its outer cladding. A second Ukrainian nuclear plant, Zaporizhzhia, has been under Russian occupation since March 2022, with all six reactors held in cold shutdown since that September.
Fukushima Daiichi, 11 March 2011. At 2:46 p.m. a magnitude 9.0 earthquake struck off Japan's north-east coast. The reactors running at Fukushima Daiichi shut down automatically, as designed. The tsunami that followed reached about 15 metres at the site and flooded the backup generators. With no power to remove their decay heat, the cores of units 1 to 3 largely melted within three days. Hydrogen explosions wrecked the reactor buildings. The accident was rated at level 7 on the international scale, the same as Chernobyl.
The earthquake and tsunami themselves killed nearly 20,000 people, with more than 2,500 missing, by official counts. Radiation sickness killed no one, though in 2018 the government accepted that one plant worker's death from lung cancer was linked to his exposure. The evacuation is a different matter. Around 160,000 people left their homes, and by September 2020 Fukushima prefecture had recorded 2,313 "disaster-related deaths" among evacuees: people, about 90% of them aged 66 or over, who died of the stress, disruption and interrupted care of being moved. About 27,000 people were still displaced as of November 2025. In 2021 UNSCEAR concluded that "no adverse health effects among Fukushima residents have been documented", and that future effects from radiation were unlikely ever to be detectable.
The uncomfortable conclusion is that at Fukushima the evacuation seems to have caused far more deaths than the radiation it was meant to prevent. Some argue it was too large and too long. Others reply that officials could not know the doses in the chaos of March 2011, and that the evacuation was itself a cost of the accident.
How many did Chernobyl kill? Here the numbers diverge, and the reason is scientific as much as political.

Estimate
Source
What it counts

Under 50
Chernobyl Forum (2005)
Deaths directly attributed by mid-2005, mostly highly exposed workers

Up to about 4,000
Chernobyl Forum: WHO, IAEA, other UN bodies and the three governments (2005)
Projected eventual deaths among the 600,000 most exposed people

About 4,000–9,000
Same group, updated 2006
Adds projections for millions of less-exposed people

About 16,000
IARC (WHO's cancer agency), 2006
Projected cancer deaths across Europe by 2065

About 27,000
Union of Concerned Scientists (2011)
Projected cancer deaths, using the linear no-threshold model

93,000 to 200,000
Greenpeace (2006)
Cancer deaths, and all additional deaths, using wider assumptions

About 985,000
Yablokov and colleagues (2007)
Premature deaths; published in the Annals of the New York Academy of Sciences but not endorsed by the academy, and widely criticised

Almost all the difference comes from one question: does a very small dose of radiation, spread over millions of people, cause a very small number of cancers? The linear no-threshold model says yes. Risk falls with dose but never reaches zero, so a tiny risk multiplied by a huge population gives thousands of deaths. Regulators use it because it is cautious. But at low doses the extra cancers are far too few to see against the normal background: cancer already causes nearly one death in six worldwide. UNSCEAR's position since 2012 has been that it does not recommend multiplying very low doses by large numbers of people to estimate deaths. The higher projections are not invented; they come from a model that can't be checked at those doses, in either direction.
Deaths per unit of energy. One way to compare energy sources is to count deaths from accidents and air pollution per terawatt-hour of electricity. A terawatt-hour is a year's electricity, counting homes, industry and everything else, for about 80,000 people in the United States or about 240,000 in Britain. Our World in Data's comparison, by Hannah Ritchie, uses the lower, UNSCEAR-based Chernobyl figure (433 deaths) and the Fukushima evacuation deaths (2,314):

Source
Deaths per TWh

Brown coal
32.7

Coal
24.6

Oil
18.4

Biomass
4.6

Gas
2.8

Hydropower (including China's Banqiao dam failure)
1.3

Wind
0.04

Nuclear
0.03

Solar
0.02

What if the higher Chernobyl estimates are right? The calculation is easy to redo. Nuclear plants generated about 96,900 TWh between 1965 and 2021. Using the Chernobyl Forum's 4,000, nuclear comes out at about 0.07 deaths per TWh. Using Greenpeace's 93,000, about 1.0, which is still a third of gas and a twenty-fifth of coal. Only the most extreme estimate, nearly a million, would put nuclear above gas. On almost any mainstream number, electricity's main killer is not radiation but smoke from burning coal, oil and wood, which kills quietly every day rather than spectacularly once a generation.
Part 4 — France and Germany: two countries, two answers
Two rich, democratic neighbours made opposite choices, and each makes sense from inside its own history.
France went in. In the 1970s France had strong heavy engineering but few energy resources of its own. Most of its electricity was made by burning imported oil. A few months after the 1973 oil shock, on 6 March 1974, Prime Minister Pierre Messmer announced a crash programme of reactors using licensed Westinghouse technology. A popular slogan of the period ran: "In France, we do not have oil, but we have ideas." About 56 reactors went into service over the following fifteen years. By 1990 nuclear supplied three-quarters of France's electricity, and close to 80% by the mid-2000s.
The defenders' case is in the results. In 2025, France's 57 operating reactors produced about 373 of its 547 TWh, around 68%. Its electricity emitted about 41 grams of CO₂ per kilowatt-hour, less than a tenth of the world average. The programme also built a national industry and a skilled workforce.
The critics' case is also real. When one technology supplies most of a country's power, a fleet-wide problem is a national one. In 2022, stress corrosion found in reactor pipework cut nuclear output to 282 TWh against a ten-year average of 395, during a European energy crisis. And France lost the knack of building. Flamanville 3, its first new reactor in years, began construction in December 2007 and was meant to start in 2012. It reached the grid on 21 December 2024. EDF's own cost estimate rose from €3.3 billion to €13.2 billion, and France's national audit court, the Cour des comptes, put the full cost, including financing, at up to €19.1 billion in 2020. A 2014 law to cut nuclear's share to 50% was abandoned in 2023. In February 2022 President Macron announced six new large reactors (EPR2s), with eight more under consideration. EDF's estimate for the first six rose to €72.8 billion in December 2025.

France proves that a country can decarbonise its electricity with nuclear power. Flamanville proves how hard it is to do it again.

Germany got out. West Germany's anti-nuclear movement was rooted in the Wyhl protests and in long fights over a proposed waste site at Gorleben. After Chernobyl, the Social Democrats resolved to leave nuclear power. In 2000–2001 a Social Democrat–Green government agreed a phase-out with the utilities, limiting each reactor's lifetime output. In 2010 Angela Merkel's government extended the reactors' lives. Then came Fukushima. In March 2011, eight older reactors were shut immediately, and at the end of June the Bundestag voted 513 to 79 to close all of them by 2022. After Russia cut gas supplies, the last three ran a few extra months and closed on 15 April 2023. In 2021 the state agreed to pay the utilities €2.4 billion in compensation.
Germany's case, in its defenders' terms, was never only about fear. Its supporters argued that no country had yet opened a permanent repository for waste. They argued that a densely populated country should not accept even a small chance of a disaster like Fukushima's, and that money and political attention should go to renewables, which they expected to become cheaper and safer. On that last point they have a record to point to: renewables rose from about 17% of German electricity in 2010 to about 59% in 2025, and coal fell from 42% to about 21%.
Critics answer that the order of events was wrong. In 2010 nuclear supplied about 22.5% of Germany's power with almost no carbon. Had the reactors stayed open, the new renewables could have pushed out coal far faster. A study by the economists Stephen Jarvis, Olivier Deschenes and Akshaya Jha estimated that the lost nuclear output was replaced mainly by coal and imports, at a social cost of about $12 billion a year, more than 70% of it from deaths caused by air pollution. Germany's electricity in 2025 emitted about 330 grams of CO₂ per kilowatt-hour, eight times France's. In May 2025, under Chancellor Friedrich Merz, Germany dropped its opposition to treating nuclear power on a par with renewables in EU climate policy.
Neither side has to be foolish for this to make sense. France chose to accept nuclear risk to escape fossil fuels. Germany chose to escape nuclear risk and accepted burning coal for longer. They weighed the same risks differently.
Part 5 — Waste: the problem that has to outlast us
Every reactor produces spent fuel: intensely radioactive ceramic pellets, still in their metal tubes, that must be kept cool and shielded. The odd thing about nuclear waste is how little there is. The United States has accumulated about 90,000 tonnes of spent fuel from commercial reactors since the 1950s. Stacked together it would cover a single American football field to a depth of under ten yards. The country adds about 2,000 tonnes a year. After five years in a reactor, the fuel still holds more than 90% of its potential energy, which is why some argue it should be recycled rather than buried; the United States does not currently recycle it. Much of a coal plant's waste goes up the chimney and into everyone's lungs. Nuclear waste is small, solid and tracked, which is both why it can be managed and why it is so feared: it sits there, and someone has to look after it for a very long time.
Spent fuel spends at least five years cooling under water in pools at the reactor, then often goes into sealed casks shielded by steel and concrete. This works, but it is meant to be temporary. The permanent answer that is "widely agreed to be the best solution", in the World Nuclear Association's words, and is official policy in Finland, France, Sweden and the United States, is deep geological disposal: put the waste in stable rock, deep enough and with enough barriers that it stays isolated for longer than any human institution has lasted.
The United States tried and stalled. In 1987 Congress picked Yucca Mountain in Nevada as the only site to study. In 2002 it overrode Nevada's objection and approved the site. Then the Obama administration withdrew the licence application in 2010, and Congress stopped funding the project in 2011. The site sits unused, and American spent fuel sits where it was made.
Finland went ahead. At Olkiluoto, on the west coast, the company Posiva has dug a repository called Onkalo more than 400 metres down in the bedrock. The method, developed in Sweden, puts the fuel in a canister sealed inside copper, packs each one into its own hole with bentonite clay, and backfills the tunnels. Construction was licensed in 2015, full-scale testing began in August 2024, and on 22 October 2026 the Finnish government granted Posiva the world's first operating licence for the final disposal of spent nuclear fuel, valid to 2070. Disposal begins once commissioning tests are approved by the regulator; no start date has been announced. The licence covers 6,500 tonnes of spent fuel, and current plans have the repository sealed for good around 2120. Sweden's repository at Forsmark is expected in the 2030s, and France's, Cigéo, around 2050.

Interim storage (pools and casks)
Deep geological repository

Where
At reactor sites
Hundreds of metres down in stable rock

Relies on
People maintaining and guarding it
Rock and engineered barriers, with no need for people

Track record
Decades, good
Finland's licence (October 2026) is the first anywhere

Main worry
Depends on institutions lasting
Predicting rock and groundwater over immense timescales

The open question is less technical than moral. A repository asks a society to make a decision that binds people thousands of generations away. The documentary Into Eternity (2010) is built around one problem: how do you warn people not to dig here, in a language no one will speak? Supporters answer that leaving the waste on the surface for ever is the worse burden to hand on. Critics answer that a technology which needs such a promise should not be expanded until the promise has been kept somewhere. Finland is now the test case.
Part 6 — The comeback: nuclear power in 2026
The numbers, first, as of the end of October 2026. The IAEA's reactor database (PRIS) lists 417 reactors in operation in 31 countries, with 380 gigawatts of capacity, plus 21 in "suspended operation" in two countries. It lists 78 under construction in 17 countries. The World Nuclear Association, which counts differently, gives 443 "operable" reactors and 79 under construction. Nuclear supplied about 2,600–2,800 TWh in 2025 (sources differ on the exact figure), about 9% of the world's electricity, roughly half its 1996 peak share.

Country
Operating
Under construction
Nuclear share of electricity, 2025

United States
94
0
17.7%

China
60
37
5.0%

France
57
0
68.1%

Russia
34
5
18.7%

South Korea
26
4
31.0%

India
23
8
3.2%

Japan
14
2
9.4%

Source: IAEA PRIS, read 31 October 2026.
China is building almost half of all the reactors under construction in the world. The West's recent record is harder to admire. America's only new reactors in a generation, Vogtle 3 and 4 in Georgia, entered service in July 2023 and April 2024, about seven years late. Their cost rose from an original estimate of about $14 billion to more than $30 billion, and Westinghouse went bankrupt along the way in 2017. Britain's Hinkley Point C is now estimated at £35 billion in 2015 prices, with its first unit not expected before about 2030.
So why the talk of a comeback? Three things changed. Climate policy made low-carbon power that runs day and night look valuable. Russia's invasion of Ukraine made energy security urgent again, as in the 1970s. And computing began to demand electricity on a new scale. A US government report from Lawrence Berkeley National Laboratory found that data centres used about 4.4% of American electricity in 2023 (176 TWh), and projected 6.7% to 12% by 2028. The International Energy Agency's 2025 Energy and AI report put global data-centre use at about 415 TWh in 2024, projecting about 945 TWh by 2030.
The technology companies have signed up. Microsoft's 20-year deal (September 2024) is restarting Three Mile Island Unit 1, renamed the Crane Clean Energy Center, helped by a $1 billion federal loan in November 2025. Google agreed in October 2024 to buy up to 500 MW from Kairos Power's small advanced reactors by 2035. In June 2025, Amazon contracted for power from Pennsylvania's Susquehanna plant until 2042, and Meta for 20 years from the Clinton plant in Illinois.
Restarts are harder than announcements. Michigan's Palisades plant, closed in 2022, was meant to restart by the end of 2025. Fuel loading began in late August 2026 and was paused when a fuel assembly tilted; as of late October there was no new date.
Governments have joined in. In May 2025, President Trump signed executive orders aiming to quadruple US nuclear capacity to 400 GW by 2050. A Declaration to Triple Nuclear Energy by 2050, launched at the COP28 climate summit in Dubai at the end of 2023, had 39 national signatories by October 2026.
Small modular reactors (SMRs) are the industry's answer to its cost problem. The idea is to build reactors of up to about 300 MW in factories, in series, so that costs fall with repetition. The status is early:

China's HTR-PM, a small gas-cooled plant, entered commercial operation in December 2023. Its Linglong One (ACP100), under construction since July 2021, was expected to begin commissioning during 2026.
Canada is furthest along in the West. Ontario Power Generation received a construction licence for a GE Hitachi BWRX-300 at Darlington in April 2025, poured the first nuclear concrete on 26 June 2026, and expects the first unit to run in 2030. Four units are estimated at C$20.9 billion.
NuScale in the US had a 50 MW design certified by the NRC in January 2023 and a larger 77 MW version approved in May 2025. But its first planned plant, in Idaho, was cancelled in November 2023 after its estimated cost rose to $9.3 billion.
TerraPower received an NRC construction permit in March 2026 for its Natrium plant at Kemmerer, Wyoming: the first permit in over 40 years for a commercial power reactor that is not a light-water reactor.
Britain chose Rolls-Royce SMR in June 2025 and signed a contract in April 2026 for three units at Wylfa in Wales, with power expected by the mid-2030s.

The SMR pitch: factories will do for reactors what they did for cars. No one has yet built enough to prove it.

Part 7 — Fusion: the other nuclear
Fission splits heavy atoms. Fusion joins light ones, the reaction that powers the Sun. Its fuel is hydrogen isotopes, it cannot run away the way a fission chain reaction can, and it produces far less long-lived waste. The catch is that it needs temperatures of about 150 million °C and a way to hold the fuel together long enough. ITER points out that fusion is not a chain reaction: if anything goes wrong, it simply stops. Fusion has been "thirty years away" for most of the last seventy. Two real milestones have made the joke a little less true.
Ignition, December 2022. On 5 December 2022, at Lawrence Livermore National Laboratory's National Ignition Facility (NIF) in California, lasers delivered 2.05 megajoules of energy to a tiny capsule of fuel, and the fusion it ignited released 3.15 megajoules. For the first time, a fusion experiment produced more energy than the energy put into the fuel. By June 2026 NIF had achieved ignition eleven times, and its record, set on 7 April 2025, was 8.6 MJ from 2.08 MJ of laser light, a gain of about 4.
The fine print matters. The lasers themselves drew roughly 300 MJ of electricity to produce that 2.05 MJ pulse (according to figures reported at the time), so the facility as a whole used about a hundred times more energy than the fusion produced. Ignition proved the physics, not a power station.
Fusion's waste problem is also smaller. ITER says a fusion plant would produce no high-activity, long-lived waste; its main waste would be machine parts made radioactive by neutrons, most of which could be recycled within about 100 years. That is a real advantage over fission, if fusion can be made to work at all.

Milestone
What it means
Status

Target gain above 1
Fusion releases more than the energy delivered to the fuel
Achieved, NIF, December 2022

Plasma gain Q above 1 in a magnetic device
Fusion power exceeds the heating power
Not yet; the record is Q ≈ 0.67, at Europe's JET

ITER Q = 10
500 MW of fusion from 50 MW of heating
Planned for after deuterium–tritium operation begins in 2039

Electricity to a grid
A power plant
Not yet; private firms target 2028 and the early 2030s

ITER, the international machine under construction in southern France, is the main public bet on the other approach, holding plasma in magnetic fields in a doughnut-shaped tokamak. Its members are China, the European Union, India, Japan, Russia, South Korea and the United States. Its aim is a gain of 10. Its schedule has slipped badly. Under the baseline revised in 2024, after defects were found in key components, research operation begins in 2034 and deuterium–tritium fusion in 2039, four years later than planned in 2016. Its original construction estimate was about €5 billion. Later estimates run several times higher, and ITER itself says the cost is being reassessed.
Private fusion is the newer story. The Fusion Industry Association says more than $15 billion has been invested in private fusion companies. Commonwealth Fusion Systems says its SPARC machine in Massachusetts will reach Q above 1 in 2027. Helion has agreed to sell electricity to Microsoft from 2028 and began building a plant at Malaga, Washington, in July 2025. These are company targets, not results. Fusion's history is full of confident dates, and the honest position is that a demonstration of net electricity in the late 2020s or early 2030s would be a major achievement. Fusion supplying a large share of the world's power before 2050 would be a surprise.
Part 8 — The strongest case for, and against
Here are both cases, as their best advocates would put them.
The case for. Nuclear power is the only proven source of very low-carbon electricity that runs day and night regardless of weather, in large amounts, on a small footprint. France, and Sweden alongside its hydropower, decarbonised their grids with it decades ago. Measured by deaths per unit of energy it is among the safest sources ever used, alongside wind and solar and far ahead of fossil fuels, even when Chernobyl and Fukushima are counted. Its waste is small in volume, tracked, and now has a licensed permanent solution in Finland. Its fuel is so concentrated that a country can stockpile years of it. A grid run largely on wind and solar needs firm power to fill windless winter weeks, and nuclear power does that without burning anything. The costs have been high because the West stopped building and forgot how. Where countries build steadily and to standard designs, as France did in the 1980s and China and South Korea have since, costs fall. Abandoning working reactors, as Germany did, meant burning more coal.
The case against. New nuclear plants in the West have been, again and again, years late and billions over budget: Vogtle, Flamanville, Hinkley Point C, and Finland's Olkiluoto 3, which was due in 2009 and began regular production in 2023. Meanwhile solar and wind have become, in the investment bank Lazard's 2026 analysis, the cheapest new sources of electricity without subsidies, and they are quick to build. Every euro and year spent on a reactor is a euro and year not spent on cutting emissions faster. Accidents are rare but, when they happen, they empty regions for decades, and Fukushima showed that even a wealthy, technically capable country can be caught out. Spent fuel needs safeguarding over timescales no institution has ever managed, and most countries still have nowhere to put it. Civilian nuclear technology has been a path to weapons. Reactors in a war zone are now a reality, at Zaporizhzhia. SMRs and fusion are promises, and the industry has a seventy-year history of promises, starting with "too cheap to meter".
What is genuinely unsettled is mostly empirical, and much of it could be answered within a decade:

Open question
What would settle it

Do SMRs get cheaper when built in series?
The real cost of Darlington's first unit (due 2030) and, above all, of the second, third and fourth

How much does firm power like nuclear lower the cost of a clean grid?
Grids that get very high shares of wind and solar, with their full costs for storage and backup published

Can the West build at China's pace and cost?
Whether Western projects after Vogtle and Flamanville come in on time and budget

Does deep disposal work?
Onkalo's first years of operation, and whether Sweden and France follow

Will fusion make net electricity?
SPARC's target of Q above 1 (2027) and Helion's promised delivery to Microsoft (2028)

Both cases agree on more than they admit: that burning coal is the deadliest way to make electricity, that Western nuclear construction has been badly managed, and that the waste is small in volume. They part ways on cost, on time, and on how much a rare disaster should count.
Then there is a value question that evidence won't settle: how to weigh a small chance of a large disaster against the steady, invisible harm of burning fossil fuels. France and Germany answered that question differently, and both answers were reasonable.

The takeaway
Born in the Manhattan Project and sold as "Atoms for Peace" in 1953, nuclear power grew twenty-five-fold between 1970 and 1990, then stalled on cost, protest and accidents. Three Mile Island (1979) killed no one. Chernobyl (1986) killed about 30 people in the first weeks and caused more than 6,000 childhood thyroid cancers; projections of its long-term toll run from about 4,000 to over 90,000 deaths, depending mainly on how low doses are modelled. At Fukushima (2011) the evacuation killed far more people than the radiation. Per unit of energy, nuclear is among the safest sources, with wind and solar; its weakness in the West has been cost and construction time. As of October 2026, 417 reactors operate, 78 are being built (37 in China), Finland has licensed the first permanent waste repository, tech companies are buying nuclear power, and SMRs and fusion remain mostly promises with dates attached.
Try this: look up your own country's electricity mix on Our World in Data (search "electricity mix" and your country). Find the share from nuclear, coal and gas, then multiply each source's annual TWh by its deaths per TWh in the table in Part 3. You'll get a rough estimate of how many lives each part of your electricity costs a year. Then decide for yourself which risks you would trade for which.

Sources & Further Reading
Sources
- Chicago Pile-1: Atomic Heritage Foundation; Wikipedia, "Chicago Pile-1" (sources disagree on the first run's length, so "minutes"). Manhattan Project: US DOE Office of Legacy Management. Hiroshima and Nagasaki estimates: Wikipedia (lead).
- Atoms for Peace: Eisenhower Presidential Library; criticism summarised from Wikipedia, "Atoms for Peace". "Too cheap to meter": Wikipedia article of that name.
- First reactors, Rickover, Yankee Rowe, 1970s cancellations: World Nuclear Association (WNA), "Outline History of Nuclear Energy" (updated 15 March 2026); Wikipedia on Obninsk, Calder Hall and Shippingport. PWR count: WNA, "Nuclear Power Reactors" (5 August 2026). R. Cowan, "Nuclear Power Reactors: A Study in Technological Lock-in", Journal of Economic History (1990).
- Generation, shares and carbon intensity, 1970–2025 (France, Germany, Sweden, world) and per-capita electricity: Our World in Data (Ember and Energy Institute data), downloaded 31 October 2026.
- Three Mile Island: US NRC backgrounder (reviewed 10 September 2026); WNA (23 March 2026); Wikipedia (evacuation, film and clean-up details).
- Chernobyl: WNA (15 April 2026); WHO Chernobyl Q&A (UNSCEAR's 6,000+ thyroid cancers, 15 fatal; IARC's 16,000); WHO news release of 5 September 2005 (Chernobyl Forum). Other estimates: Wikipedia, "Deaths due to the Chernobyl disaster". UNSCEAR's 2012 low-dose statement, as quoted in Wikipedia, "Linear no-threshold model" (UNSCEAR's site could not be reached). Cancer as nearly one death in six: WHO fact sheet (2024 data).
- Fukushima: WNA (23 September 2026), including UNSCEAR 2021. Tsunami toll: official 2021 figures, via Wikipedia.
- Deaths per TWh: Hannah Ritchie, Our World in Data, "What are the safest and cleanest sources of energy?" and its chart data (2,747 deaths over 96,876 TWh, 1965–2021).
- France: WNA (1 April 2026); Wikipedia, "Messmer plan" and "Flamanville Nuclear Power Plant". Germany: WNA (8 April 2026); Jarvis, Deschenes and Jha, NBER Working Paper 26598 (2019), Journal of the European Economic Association (2022).
- Waste: US DOE, "5 Fast Facts about Spent Nuclear Fuel" (2022, including the 90% of energy left and no US recycling); WNA, "Storage and Disposal of Radioactive Waste" (23 October 2026); Posiva press release, 22 October 2026; WNA Finland page (6,500 tonnes, sealing around 2120); Wikipedia on Onkalo and Yucca Mountain.
- 2026 status: IAEA PRIS world and country statistics, read 31 October 2026; WNA world, USA (27 October 2026), UK, China, Canada, Finland and SMR pages (September–October 2026); US EIA, "U.S. nuclear industry" (March 2026); Wikipedia on Vogtle and NuScale; WNA Net Zero Nuclear page on the tripling declaration (2 October 2026); Lazard, LCOE+ 2026 summary page.
- Data centres: US DOE, 20 December 2024 (Lawrence Berkeley National Laboratory report). IEA, Energy and AI (10 April 2025), as reported by S&P Global Commodity Insights the same day (the IEA site could not be reached).
- Fusion: LLNL, "Achieving Fusion Ignition"; the ~300 MJ laser draw as reported by NBC News at the time; ITER FAQ and Wikipedia, "ITER"; Fusion Industry Association; Commonwealth Fusion Systems' SPARC page; Helion's news page.
Go deeper
- 📘 The Making of the Atomic Bomb — Richard Rhodes. The classic history of how fission became a weapon, including Chicago Pile-1.
- 📘 Energy and Civilization: A History — Vaclav Smil. Nuclear power in the long story of how humans have powered themselves.
- 📘 Midnight in Chernobyl — Adam Higginbotham. A careful reconstruction of the accident and the Soviet response.
- 🌐 IAEA PRIS: Power Reactor Information System — the official live count of the world's power reactors.
- 🌐 Our World in Data: What are the safest and cleanest sources of energy? — the deaths-per-TWh comparison, with its method.
- 🌐 World Nuclear Association information library — detailed country profiles, updated often. It is the industry's own body: read its framing as one side's case.
- 🌐 NRC: Backgrounder on the Three Mile Island accident — the regulator's plain account.
- ▶️ Into Eternity (2010) — Michael Madsen's documentary filmed inside Onkalo.
Related pieces in this library: How a Nuclear Reactor Works (Physics 14) · Energy: The One Currency of the Universe (Physics 2) · How Electricity Actually Reaches Your Wall (Physics 4) · The Physics of the Greenhouse Effect (Physics 15) · Marie Curie: Two Nobel Prizes, One Radical Idea (History-Makers 3) · France: A Condensed History of Revolution, Republic, and Grandeur and Germany: A Condensed History of Fragmentation, Catastrophe, and Repair (Condensed History of Nations 8 and 5) · The 1950s (Decades of Leadership 4, on Eisenhower and the Cold War).


How Screens Work
----------------
From a German physicist's tube for watching electric currents to the folding glass in your pocket: how screens make light and colour, why some of them flicker and some never do, and what the evidence really says about screens and sleep.

Follows: Light & Color: Why the World Looks the Way It Does (Physics 9) and Color, Explained: Why Red Next to Green Shouts and Gray Can Lie (Art & Seeing 6)
  

  Look closely at the white of this page. If you are reading on a phone or a laptop, there is no white there at all. There are millions of tiny red, green and blue lights, each switched to a carefully chosen brightness, and your eye is too coarse to see them apart. A phone screen of 2,532 by 1,170 pixels holds nearly 3 million pixels and almost 9 million of these coloured lights. If you are reading on an e-ink reader, the trick is completely different: there are no lights at all, just tiny charged grains of white and black pigment pushed up or down by electric fields.

  This piece follows the screen from its beginnings in 1897 to the foldable phones on sale in October 2026: who invented what, how each kind of screen actually works, what resolution and refresh rate really mean, and what the evidence says about blue light and sleep. Two interactives let you zoom into a pixel and test the refresh rate of the screen you are holding.

  
    In this deep dive
    
      The cathode-ray tube: from watching electricity to television
      Going flat: plasma and the liquid-crystal display
      Pixels: colour, sharpness and smooth motion (with a subpixel zoom and a refresh test)
      Beyond the backlight: OLED, microLED, e-ink and touch
      Blue light, sleep and your eyes: what the evidence shows
      The state of screens in 2026, and open questions
    
  

  Part 1 — The cathode-ray tube: from watching electricity to television

  The first screen was not built to show pictures. In 1897 Ferdinand Braun, a physicist at the University of Strasbourg, published a paper with a very unglamorous title: a method for demonstrating and studying the time course of varying electric currents. His device was a glass tube with the air pumped out. At one end, a high voltage pulled a stream of electrons (then called cathode rays) off a metal plate. The stream passed through a small hole and struck a coating at the far end that glowed where it landed. A magnet coil carrying the current he wanted to study pushed the beam sideways, so the glowing spot traced the current's shape. That is the first oscilloscope, and the "Braun tube", as Germans still call it, is the ancestor of every television made for the next century.

  The idea of sending moving pictures by wire was older than the tube, and for a while it went down a different path. Mechanical television scanned a scene with a spinning disc punched with holes. On 26 January 1926 the Scottish inventor John Logie Baird showed moving television images this way to members of the Royal Institution in his London workshop, and the BBC later broadcast his system. But a spinning disc could not scan fast or finely enough. In 1908 the British engineer Alan Campbell-Swinton had argued in a letter to Nature that cathode-ray tubes could do the whole job electronically, at both ends. Getting it to work took two decades and a famous rivalry.

  In 1927, in a laboratory in San Francisco, a 21-year-old Utah farm boy named Philo Farnsworth transmitted the image of a single straight line with an all-electronic camera tube he called the image dissector. He said he had got the idea of scanning a picture line by line at 14, from the rows of a ploughed field. Meanwhile Vladimir Zworykin, who had studied in St Petersburg under the television pioneer Boris Rosing, was developing electronic cameras and picture tubes, first at Westinghouse and then at RCA under David Sarnoff. RCA claimed Zworykin's 1923 patent application came first. The US Patent Office disagreed and awarded Farnsworth priority for his invention, and in 1939 RCA agreed to pay him $1 million in licence fees.

  Television had no single inventor. It had a tube from 1897, an idea from 1908, and a decade of patent lawyers.

  
    WhoWhenWhat they actually did
    Ferdinand Braun1897The cathode-ray tube, as an instrument for watching currents
    Alan Campbell-Swinton1908Proposed all-electronic television using tubes at both ends; never built it
    John Logie Baird1926First public demonstration of moving television images, using a mechanical system
    Kenjiro Takayanagi1926Showed a picture on a cathode-ray tube receiver in Japan
    Philo Farnsworth1927First all-electronic camera tube to transmit an image; won the key patent fight
    Vladimir Zworykin (RCA)1920s–1930sThe kinescope picture tube and the iconoscope camera; RCA turned them into an industry
  

  Mechanical television lost the race in Britain in 1937. The BBC had been alternating Baird's system with an all-electronic one from EMI-Marconi since November 1936, and it dropped Baird's in February 1937. From then on, for about seventy years, a television meant a heavy glass tube. So how did the tube make a picture?

  A cathode-ray tube (CRT) never shows a whole picture at once. At any instant, only one tiny spot is lit. At the narrow back of the tube, a heated cathode boils off electrons, and a high voltage accelerates them into a tight beam. Coils around the tube's neck make magnetic fields that bend the beam, sweeping it across the screen from left to right in a line, then dropping it a little and sweeping again. Where the electrons hit the coating of phosphor on the inside of the glass, it glows. Varying the beam's strength as it goes paints light and dark. This line-by-line pattern is called a raster, and the word has outlived the tube.

  The American standard, set in 1941, used 525 lines, 30 times a second. To stop the picture flickering, each frame was drawn in two passes, first the odd lines and then the even ones, giving 60 half-pictures, or fields, a second. This trick is called interlacing. When colour arrived, the frame rate had to be nudged down by 0.1%, to 29.97, to fit the colour signal in, and it stayed there.

  Do the arithmetic and the speed is startling. At 525 lines and 29.97 frames a second, the beam draws about 15,734 lines every second, so each line takes about 63.6 millionths of a second. On a TV half a metre wide, the glowing spot crosses the screen at well over 7 kilometres a second. You never see a dot. You see a picture, because the phosphor keeps glowing for a moment after the beam passes, and your visual system blends what remains into a steady image.

  A CRT is a single point of light moving at several kilometres a second, and your brain turns it into a face.

  Colour needed three phosphors (red, green and blue) and three electron beams. The problem was making sure the "red" beam only ever hit red phosphor. RCA's solution was the shadow mask: a thin metal sheet just behind the glass, punched with a fine grid of tiny holes. Each beam came in at a slightly different angle, so through any hole it could only reach its own colour of dot. The US approved a colour standard in December 1953, and RCA's CT-100 of 1954 was the first colour set made in quantity. In 1968 Sony launched the Trinitron, which replaced the mask of holes with fine vertical wires and phosphor stripes. Trinitron sets were known for their brightness, and the vertical stripes of red, green and blue on modern LCDs echo them.

  CRTs reigned for decades. Worldwide sales of CRT computer monitors peaked at about 90 million in 2000, and CRT televisions at about 130 million in 2005. Then they collapsed, and in 2012 Samsung's display arm, one of the last big makers, stopped making them. The tubes were heavy, deep, power-hungry and full of leaded glass, and once flat screens became cheap, nothing could save them.

  Part 2 — Going flat: plasma and the liquid-crystal display

  Plasma came first to big screens. In 1964 at the University of Illinois, Donald Bitzer, Gene Slottow and a graduate student, Robert Willson, built a flat display for PLATO, a pioneering computer system for teaching. Each pixel of a plasma screen is a tiny sealed cell of gas. A voltage turns the gas into a glowing plasma that gives off ultraviolet light, which makes a coating of red, green or blue phosphor glow, as in a fluorescent tube. Plasma TVs became the first big, flat televisions in the late 1990s. They had deep blacks and fast motion, but they were heavy, ran hot and used a lot of power. Panasonic stopped making plasma panels in March 2014, and LG and Samsung stopped the same year.

  The winner was a stranger technology, and its story starts with a carrot. In 1888 the Austrian botanist Friedrich Reinitzer, studying a compound extracted from carrots, noticed that it seemed to melt twice: first into a cloudy liquid, then into a clear one. He had found a liquid crystal, a substance whose rod-shaped molecules can flow like a liquid yet line up together like a crystal. For eighty years it was a laboratory curiosity.

  In the 1960s a team at RCA's laboratories in Princeton led by George Heilmeier found that an electric field could change how liquid crystals scattered light, and in 1968 they showed the first liquid-crystal display. It was not yet practical. The breakthrough was the twisted nematic cell, and it came with a dispute. Martin Schadt and Wolfgang Helfrich, at the Swiss drug company Hoffmann-La Roche, filed a patent in December 1970 and published in February 1971. The American James Fergason filed a similar US patent in 1971. The dispute was settled out of court, with all sides sharing royalties. Who sold the first LCD calculator is argued over too. Sharp's EL-805 of 1973 is usually given the credit, though calculators with an earlier type of LCD were sold in the US in 1972.

  An LCD does not make any light. It is a stack of filters in front of a lamp, and each pixel is a tiny shutter. Read the stack from the back, where the light starts:

  
    1. Backlight. White light, once from fluorescent tubes, now from LEDs.
    2. First polariser. Lets through only light waves vibrating in one direction.
    3. Liquid crystal, between transparent electrodes. With no voltage, the molecules form a quarter-turn twist that rotates the light's direction of vibration by 90°. A voltage untwists them, and the rotation goes away.
    4. Colour filters. A red, green or blue filter over each subpixel.
    5. Second polariser, crossed at 90° to the first. Rotated light passes; unrotated light is blocked.
    6. Glass. What you see: each subpixel's shutter, partly open or shut.
  

  Partial voltages give partial twists, which give shades in between. To control millions of shutters separately, each subpixel needs its own tiny switch, a thin-film transistor. In 1973 a Westinghouse team led by T. Peter Brody showed a liquid-crystal display driven by thin-film transistors, and in 1975 Brody named the idea the "active matrix". It took the industry, led by Japan, Korea and Taiwan, until the 2000s to make large panels cheaply. In the last three months of 2007, LCD televisions outsold CRTs worldwide for the first time.

  The LCD's weakness is built in. The backlight is always on, and a closed shutter still leaks a little, so blacks look dark grey in a dark room. Much of the LCD's later history, from LED backlights that dim in zones behind dark areas to "quantum dot" layers that purify the colours, is a fight against that leak.

  Part 3 — Pixels: colour, sharpness and smooth motion

  Your eye has three kinds of colour-sensing cone cells, most sensitive to long ("red"), medium ("green") and short ("blue") wavelengths. As Light & Color explains, your brain builds colour by comparing their three signals. That is the loophole every screen exploits. A screen does not need to produce the actual wavelength of yellow light. It only needs to excite your red and green cones in the same proportions that yellow light would, and you see yellow. Two physically different lights that look the same are called metamers. Every colour on a screen is one.

  So each pixel is three subpixels, red, green and blue, and mixing is additive: light adds to light, and all three at full strength make white. (Paint is the opposite, as Color, Explained shows: each pigment subtracts light, which is why printers use cyan, magenta and yellow.) In the usual system each subpixel gets a number from 0 to 255, so there are 256 × 256 × 256 = 16,777,216 possible colours.

  
    Interactive · Zoom into a pixel
    
      Yellow
      Orange
      Purple
      White
      Mid grey
      Teal
    
    Zoom
      
    
    R
    G
    B
    
      The colour itself
      The same light spread over three times the area: what the zoomed stripes blend to
    
    
    Slide the zoom all the way left and the stripes shrink to one physical pixel of your own screen each, small enough on most phones for your eye to blend them. The blend looks darker than "the colour itself", for an honest reason: a drawing of a red subpixel on your screen is itself made of your screen's red subpixels, which fill only a third of each of your pixels. So the drawn stripes put out a third of the light of the real colour (a little less once black gaps appear between them), and the second swatch shows exactly that third. Light percentages use the standard sRGB curve.
  

  Try mid grey. Every subpixel gets the code 128, halfway to 255, yet each one shines at only about 22% of full power. Screen codes do not map to light in a straight line. The standard sRGB curve was designed to suit the way old CRTs responded to voltage and the way our eyes respond to light, and it puts more of the 256 steps among the darker shades. It takes a code of about 188, not 128, to light a subpixel to half power.

  Try yellow, then blue. Yellow's two lit subpixels look far brighter than blue's single one, and not only because there are two of them. Our eyes are much more sensitive to green light than to blue. In the standard formula for brightness, green contributes about 72%, red about 21%, and blue only about 7%. That is why a pure blue screen looks dim, and why blue text on black is hard to read.

  You have never seen yellow on a screen. You have seen red and green, side by side, too small to tell apart.

  Not every screen uses three equal stripes. Many OLED phones arrange their subpixels in other patterns, with different numbers of each colour, and software adjusts for it. The principle is the same: three primaries, added, too small to resolve. Which raises the next question: how small is small enough?

  Resolution is just a count of pixels: 1920 × 1080 ("Full HD") or 3840 × 2160 ("4K"). On its own it says little about sharpness, because the same 4K picture can cover a 27-inch monitor or a 65-inch television. What matters is pixel density, in pixels per inch (ppi), and how far away you are.

  A person with normal, "20/20" vision can make out detail about one minute of arc across: a sixtieth of a degree, about the width of a coin seen from 80 metres. When Apple launched the iPhone 4 in June 2010, it called its 326 ppi screen a "Retina display", claiming the eye could not tell individual pixels apart at a normal distance. The arithmetic roughly agrees: at 30 cm, one arcminute is 0.087 mm, which needs about 290 ppi. Some people see finer than 20/20, so the claim is about typical eyes, not all eyes.

  
    Example screenPixelsDensityPixels blur together (20/20 eye) beyond
    6.1-inch phone2532 × 1170457 ppi19 cm
    6-inch e-reader1448 × 1072300 ppi29 cm
    27-inch 4K monitor3840 × 2160163 ppi54 cm
    27-inch 1440p monitor2560 × 1440109 ppi80 cm
    55-inch 4K TV3840 × 216080 ppi1.1 m
    55-inch Full HD TV1920 × 108040 ppi2.2 m
  
  Computed from each screen's diagonal and pixel count, and the distance at which one pixel spans one arcminute. Beyond that distance, extra pixels add little that a 20/20 eye can see.

  The table explains a quiet truth of the TV business. From a sofa three metres away, a 55-inch Full HD set is already at the limit of normal vision, so most of the extra detail of 4K, let alone 8K, is invisible unless you sit closer or buy a bigger screen.

  Refresh rate is how many times a second the screen draws a new picture, measured in hertz (Hz). For decades the standard was 60 Hz, inherited from television. In June 2017 Apple's iPad Pro introduced "ProMotion", refreshing up to 120 times a second, and many phones now do the same. Since 2021 some phones change their rate as they go, from 120 Hz when you scroll to as low as 10 Hz for a still page, which saves power.

  Why does a higher rate look smoother? Most LCD and OLED screens are "sample-and-hold": each picture is held still until the next one replaces it. When your eye follows a moving object, the object stays put for a whole frame while your eye keeps moving, so it smears across your retina. Display engineers found in 2005 that this holding, not the slowness of the liquid crystal, caused about 70% of LCD motion blur. The cure is more frames, or brief black frames in between. Test it on your own screen:

  
    Interactive · What refresh rate can your screen show?
    StartMeasure my screen
    Press Measure my screen to time your display.
    
      10 Hz
      15 Hz
      30 Hz
      60 Hz
      120 Hz
      240 Hz
    
    Top dot: updated 30 times a second
    
    Bottom dot: updated at your screen's full rate
    
    
    Both dots cross the track every two seconds. The top one moves only when its simulated frame comes round; the bottom one moves on every refresh your screen makes. A demo can't show a rate faster than your screen refreshes, so those buttons switch off after you measure. Rates that don't divide evenly into your screen's rate look slightly uneven. Browsers may slow down in battery-saver mode, and an e-ink reader refreshes far too slowly for this demo at all. It runs for 20 seconds, then stops.
  

  Part 4 — Beyond the backlight: OLED, microLED, e-ink and touch

  In 1987 Ching Tang and Steven Van Slyke, chemists at Eastman Kodak in Rochester, New York, published a short paper on "organic electroluminescent diodes": thin films of carbon-based molecules that glowed when a small voltage pushed current through them. Glowing organic materials had been seen in France in the 1950s, but Tang and Van Slyke's two-layer design was the first that was efficient enough and needed a low enough voltage to be practical. It is the ancestor of the OLED screen.

  In an OLED, every subpixel is its own tiny lamp. There is no backlight and no shutter. A black pixel is simply switched off, so blacks are truly black, and each pixel can turn on and off almost instantly. The screen can be thin and even flexible. The costs are real too. Organic materials wear out with use, blue ones fastest, so a logo or toolbar shown for thousands of hours can leave a faint permanent ghost: burn-in. And making large OLED panels is still expensive.

  
    LCDOLEDMicroLED
    Light comes fromA backlight behind shuttersEach subpixel (organic)Each subpixel (inorganic LEDs)
    Black levelDark grey; some light leaksPerfect blackPerfect black
    Main weaknessContrast, viewing anglesBurn-in, ageing of blueVery hard and costly to make
    Where you meet it in 2026Most TVs, monitors, laptopsMost premium phones, many TVs and laptopsGiant video walls, a few very costly TVs
  

  Sony sold the first OLED television, the small XEL-1, in 2007. Apple's first OLED iPhone, the iPhone X, came in September 2017. Newer panels stack two light-emitting layers on top of each other ("tandem OLED") so each layer can work less hard, which raises brightness and slows ageing. On 28 October 2026 Apple announced a MacBook Pro with a tandem OLED screen, refreshing at up to 120 Hz and reaching up to 2,000 nits of brightness.

  MicroLED keeps the "every pixel its own lamp" idea but uses tiny inorganic LEDs, the same kind of material as the LEDs in a light bulb. They are brighter and longer-lived than organic ones, without burn-in. The first microLEDs were made by Hongxing Jiang and Jingyu Lin's group in the US in 2000, and in January 2018 Samsung showed a 146-inch microLED wall called The Wall. The problem is manufacturing. A 4K screen needs nearly 25 million separate microscopic LEDs placed precisely, and every faulty one must be found and replaced. Apple reportedly gave up on a microLED watch screen in 2024. For now, microLED lives mostly in giant video walls, plus a few luxury-priced TVs sold since 2021.

  Every screen so far makes its own light. In the 1970s Nick Sheridon at Xerox's Palo Alto Research Center wanted the opposite: a display that worked like paper, by reflecting the light around it. His "Gyricon" used tiny balls, black on one side and white on the other, that rotated in an electric field. The idea that won came from MIT's Media Lab. Its team, which founded the company E Ink in 1997 and published the method in Nature in 1998, sealed tiny white particles in dark fluid inside microcapsules. A voltage pulls the white particles to the front, and the spot looks white; the opposite voltage pushes them back, and it looks dark.

  Because it reflects light like paper, e-ink is easy to read in bright sunlight, where emitting screens wash out. And once the particles are in place they stay there with no power at all: the screen is bistable. It uses energy only to change the page, which is why an e-reader's battery lasts weeks. Sony's Librié, sold in Japan in 2004, was the first e-book reader with an E Ink screen. Amazon's first Kindle went on sale on 19 November 2007 for $399, advertised as reading "like real paper, even in bright sunlight".

  An e-ink screen showing this page right now is using no power to show it.

  The price is speed. Moving pigment through fluid takes far longer than switching a liquid crystal or an OLED, so e-ink refreshes slowly, can leave faint ghosts of the last page, and needs a full flash now and then to clear them. Colour e-ink exists, usually by laying colour filters over a black-and-white layer, but its colours are muted. That is why the refresh test in Part 3 cannot work on an e-reader, and why this library, read in an app that also runs on e-ink devices, keeps its text plain and its animations few.

  Touch began, like the CRT, as an instrument. In 1965 E. A. Johnson of the Royal Radar Establishment in Malvern, England, published a short paper describing a "touch display" for computers, to be used in air traffic control. In 1972–73 Frank Beck and Bent Stumpe at CERN built transparent capacitive touchscreens for the control room of a new particle accelerator. In the 1970s the American company Elographics developed the cheaper resistive screen, which for decades was the common choice, including for the stylus-driven handheld computers of the 1990s.

  
    ResistiveCapacitive
    How it sensesPressing brings two conductive layers into contact at one pointYour fingertip, which conducts, changes the electric field in a transparent grid
    NeedsPressure, from anything: a nail, a glove, a stickNo pressure, but a conductive touch (bare skin or a special stylus)
    Multi-touchHardNatural, with a grid of sensors
    FeelSoft plastic surface, more glareHard glass
  

  Then came 9 January 2007. On stage in San Francisco, Steve Jobs introduced the iPhone, with a 3.5-inch "multi-touch display" that you pinched and swiped with bare fingers and no stylus. It went on sale in the US that June. The iPhone was not the first phone with a capacitive screen: LG's Prada phone went on sale in May 2007. It was not the first touchscreen phone either; IBM's Simon came in the early 1990s. What it did was combine multi-touch glass, software built around fingers and a large screen with almost no buttons. Within a few years nearly every phone looked like it.

  Part 5 — Blue light, sleep and your eyes: what the evidence shows

  Sensitive-topic note: this is general information, not medical advice. Persistent trouble sleeping, or eye pain or changes in vision, deserve a visit to a doctor.

  The worry about screens and sleep has a real biological basis. In 2002 researchers led by David Berson showed that some cells in the retina are not for seeing at all. They contain a light-sensitive pigment called melanopsin, they report to the body clock, and they respond most strongly to blue light, around 480 nanometres. Bright light in the evening, especially blue-rich light, delays the body clock and suppresses melatonin, the hormone that signals night.

  The best-known screen study, from Brigham and Women's Hospital and Harvard in 2015, had volunteers read either a printed book or a light-emitting e-reader in the hours before bed, over several evenings, in a laboratory. On the e-reader, they took longer to fall asleep, made less melatonin, had a later body clock and were less alert the next morning. It was a careful study, but a small one, under controlled laboratory conditions rather than ordinary life.

  In ordinary life, the effect of the light alone looks smaller. A 2021 study randomly assigned 167 young adults with iPhones to an hour of phone use before bed with Night Shift (Apple's warmer evening colours) on, with it off, or to no phone at all, for a week. Across the whole group, sleep did not differ between the three conditions. A 2016 review of 20 studies of 125,198 children found that using a phone or tablet at bedtime went with roughly twice the odds of too little sleep. But children who merely had a device in the bedroom and did not use it also slept worse. That points beyond light, to what we do on screens: the messages, the games, the next video, and the bedtime that keeps slipping.

  The same review found that bedtime device use went with poorer sleep quality, and with nearly three times the odds of excessive sleepiness in the daytime. But all 20 of the studies it pooled were cross-sectional: surveys that catch children at a single moment. They show that bedtime screens and bad sleep go together, not which causes which. A teenager who cannot sleep may reach for a phone, rather than the other way round. Untangling cause and effect needs studies that follow people over time, or experiments, and those are rarer, smaller and less dramatic in their results.

  
    ClaimWhat the evidence says
    Bright, blue-rich light at night delays the body clockWell established
    A light-emitting screen in the evening can delay sleepShown in the lab; the effect in daily life looks modest
    Night-mode colour filters fix itNot shown; one randomised study found no difference
    Blue-blocking glasses reduce eye strainProbably not: a 2023 Cochrane review of 17 trials found little or no effect
    Screen blue light damages your eyesNo evidence, says the American Academy of Ophthalmology
    Screens strain your eyesYes, temporarily: we blink less while staring; no lasting harm
  

  What actually helps follows from that. An international expert panel in 2022 recommended keeping evening light dim: from three hours before bed, no more than about 10 lux of "melanopic" light at the eye (roughly a dim, warm room), and darkness for sleep. In the daytime, they recommend bright light: at least 250 lux of melanopic light, more than many indoor rooms give. America's CDC suggests switching devices off at least 30 minutes before bed. The ophthalmologists' 20-20-20 rule is for eye strain: every 20 minutes, look at something 20 feet (6 metres) away for 20 seconds. Sleep Is a Skill in this library covers the rest.

  Part 6 — The state of screens in 2026, and open questions

  The newest frontier is screens that bend. Samsung's first Galaxy Fold, in 2019, weighed 276 grams. In September 2026 Samsung said its Galaxy Z Fold8 weighed 201 grams, calling it the world's lightest Fold. On 9 September 2026 Apple announced its first foldable iPhone, the iPhone Duo, with a 7.6-inch inner screen and a 5.4-inch outer one, starting at $1,999 in the US. Pre-orders opened on 16 October, it reached shops on 23 October, and it uses Apple's ProMotion variable refresh. Apple's own description admits the crease is still there, saying its matte finish "minimizes crease visibility". Folding screens are OLEDs: with no backlight to carry, an OLED can be built on a bendable plastic sheet instead of rigid glass.

  Samsung's own weight figures show how quickly the engineering has moved: 253 grams for the Fold5, 239 for the Fold6, 215 for the Fold7 and 201 for the Fold8. Variable refresh rates have spread too. Apple says the new MacBook Pro's screen drops from 120 Hz to as low as 10 Hz when nothing on it is moving, saving battery in the way phones have done since 2021. How many people actually buy foldables is harder to say. Reliable, independent sales figures were not publicly available for this piece, so treat confident claims about their market share with care, whether they come from boosters or sceptics.

  
    TechnologyStatus at the end of 2026
    CRTGone from mass production; collectors and retro gamers keep them alive
    PlasmaEnded in 2014
    LCDStill the workhorse of TVs, monitors and laptops
    OLEDStandard in premium phones; spreading to laptops (tandem OLED) and foldables
    MicroLEDGiant video walls and a few very expensive TVs; still hard to make
    E-inkE-readers and signs; colour still muted
  

  Some questions are genuinely open:
  Can microLED be made cheaply? If the problem of placing and testing millions of tiny LEDs is solved, it could combine OLED's blacks with LCD's long life. If not, it stays a luxury.
  Will foldables become normal? They are now light and thin, but still expensive and still creased. Whether they become the ordinary phone or stay a premium niche is undecided.
  Can paper-like screens catch up? Colour e-ink, sold since about 2019 using colour filters, still looks washed out next to a phone, and it is far too slow for video. E Ink, owned since 2009 by the Taiwanese company Prime View International, and its rivals are chasing brighter colour and faster pages. A reflective screen that could show video comfortably would change what a reading device can be, and how much light we put in our eyes at night.
  How good is good enough? Phones are already past what most eyes can resolve at reading distance, and TVs past it at sofa distance: from three metres, even a 55-inch Full HD set reaches the limit of a 20/20 eye. Much of what is sold as better resolution now buys little you can see; brightness, contrast and refresh rate may matter more.
  What do screens do to sleep and attention? The light question is better understood than the behaviour question. The evidence increasingly suggests that what we do on screens matters more than the colour of their light. Larger studies that follow the same people over years, rather than surveys at one moment, are what would settle it.

  The thread running through all of it is the one Braun started: a screen is a way of making light do what we want, one tiny spot at a time. Every generation has found a better spot, from a glowing dot moving at kilometres a second to a pixel that turns itself off.

  
    The takeaway
    Screens began with Ferdinand Braun's 1897 tube for watching currents. Electronic television grew from it, with Farnsworth winning the key patent and RCA building the industry. A CRT paints one glowing spot at a time, line by line; an LCD is a stack of polarisers and twisting liquid-crystal shutters in front of a backlight; an OLED or microLED lights each subpixel itself; e-ink moves pigment and needs no power to hold a page. Every on-screen colour is a mix of red, green and blue light, too fine to resolve, and the codes are not linear (128 gives about 22% of full light). Sharpness depends on pixel density and distance, and much beyond about 300 ppi at reading distance is invisible to a 20/20 eye. Higher refresh rates reduce the smear of sample-and-hold screens. On sleep, bright evening light matters, but night-mode filters and blue-blocking glasses show little benefit, and what we do on screens seems to matter more. In October 2026, foldables from Samsung and Apple are on sale, creases and all.
    Try this: put a drop of water on your phone or laptop screen over a white area and look through it closely, or take a zoomed-in photo of the screen with another phone. The drop acts as a magnifying lens, and you will see the red, green and blue subpixels that the interactive in Part 3 draws. Then try the same on an e-reader, if you have one: there are no coloured lights to find, just an even, paper-like surface.
  

  
    Sources & Further Reading
    Sources. F. Braun, "Ueber ein Verfahren zur Demonstration und zum Studium des zeitlichen Verlaufes variabler Ströme", Annalen der Physik 296 (1897), 552–559; A. A. Campbell Swinton, "Distant Electric Vision", Nature 78 (18 June 1908); E. A. Johnson, "Touch display: a novel input/output device for computers", Electronics Letters 1 (1965); M. Schadt and W. Helfrich, Applied Physics Letters 18 (15 February 1971); C. W. Tang and S. A. VanSlyke, "Organic electroluminescent diodes", Applied Physics Letters 51 (1987); D. L. Bitzer and H. G. Slottow, "The plasma display panel", AFIPS Fall Joint Computer Conference (1966); B. Comiskey, J. D. Albert, H. Yoshizawa and J. Jacobson, Nature (July 1998); F. Reinitzer, Monatshefte für Chemie (1888); all confirmed through Crossref. IEEE milestone "Liquid Crystal Display, 1968" (ETHW). Sharp corporate history (EL-805). Wikipedia articles on the cathode-ray tube, NTSC (including the 1941 standard), sRGB, Philo Farnsworth, Vladimir Zworykin, John Logie Baird, the LCD, the twisted nematic effect, plasma displays, OLED, microLED, electronic paper, E Ink and touchscreens, used for dates and leads (CRT sales peaks, the CT-100, Trinitron, LCD outselling CRT in Q4 2007, the LG Prada, IBM Simon, Sony XEL-1, Jiang and Lin, Samsung's 146-inch Wall, Apple's reported 2024 microLED decision, the Librié, Elographics, Johnson's air-traffic-control purpose). Apple Newsroom: iPhone (9 January 2007), iPhone 4 (7 June 2010), iPad Pro ProMotion (5 June 2017), iPhone X (12 September 2017), iPhone 13 Pro 10–120 Hz (14 September 2021), iPhone Duo (9 September and 23 October 2026), MacBook Pro tandem OLED (28 October 2026). Samsung Newsroom, "From 253 Grams to 201 Grams" (9 September 2026). Amazon press release, "Introducing Amazon Kindle" (19 November 2007). E Ink company history. Pan, Feng and Daly, "Quantitative Analysis of LCD Motion Blur", SID Symposium Digest (2005). Webvision (University of Pittsburgh), "Visual Acuity" (20/20 as 1 arcminute). Berson, Dunn and Takao, Science (8 February 2002); Brown et al., PLOS Biology (2022, light recommendations and melanopsin's ~480 nm peak); Chang, Aeschbach, Duffy and Czeisler, PNAS 112 (2015); Duraccio et al., Sleep Health 7 (2021); Carter et al., JAMA Pediatrics (2016); Singh et al., Cochrane review CD013244 (2023); American Academy of Ophthalmology, "Should you be worried about blue light?" (2021); CDC, "About Sleep". Computed in node: sRGB decoding (128 → 21.6%, 188 → 50.3%), luminance weights, 256³, the one-third blend, pixel densities and blur distances, NTSC line timing (15,734 lines/s, 63.6 µs per line) and the spot speed, frame times and jumps; the interactive's core functions were run and checked in node.
    
      📘 The Last Lone Inventor — Evan I. Schwartz (Farnsworth against RCA and Sarnoff, the fight over who invented television)
      📘 Liquid Crystal Displays: Addressing Schemes and Electro-Optical Effects — Ernst Lueder (the engineering, for readers who want the equations)
      🌐 IEEE milestone: the first liquid-crystal display, 1968 — the RCA story in brief, from the engineers' own history wiki
      🌐 American Academy of Ophthalmology: Should you be worried about blue light? — the eye doctors' short, plain answer
      🌐 Brown et al. (2022): recommendations for light exposure by day and night — the expert consensus behind Part 5, free to read
    
    Related pieces in this library: Light & Color (Physics 9) · Color, Explained (Art & Seeing 6) · Photography for People With a Phone (Art & Seeing 8) · Sleep Is a Skill (Nervous System 4) · How Electricity Actually Reaches Your Wall (Physics 4) · next in this series: Processors: How a CPU Thinks, From the Transistor to the Chip in Your Pocket (Technology 11).


Processors: How a CPU Thinks
----------------------------
Everything your phone does comes down to billions of tiny switches flipping on and off, obeying a short list of very simple instructions. Here is how we got from a lump of germanium in 1947 to that, how it actually works, and who makes it now.

Follows: Alan Turing: The Machine That Defined Computing and How Electricity Actually Reaches Your Wall
  In this deep dive
    
      Switches that think: Boole, Shannon and the logic of on and off
      Valves, Colossus and ENIAC: the first electronic computers
      The transistor: a switch with no moving parts
      The chip: Kilby, Noyce and a computer on a fingernail
      Inside the machine: fetch, decode, execute (try it yourself)
      Making it fast: clocks, caches, pipelines and cores
      Moore's law: what it really said, and why it slowed
      The languages of chips: x86, ARM and RISC-V
      How chips are made, and what "3 nm" really means
      The state of affairs in 2026, and the open questions
    
  
  

  Part 1 — Switches that think
  Picture a light with two switches wired one after the other: the bulb comes on only if switch A and switch B are on. Rewire them side by side and it comes on if A or B is on. You have just built two pieces of logic out of copper and plastic.

  That small observation is the seed of every computer. In 1854 the English mathematician George Boole, then a professor at Queen's College, Cork, published An Investigation of the Laws of Thought. He showed that reasoning with "true" and "false" could be done as a kind of algebra, with its own rules for AND, OR and NOT. For eighty years it was mostly a curiosity for logicians.

  Then, in 1937, a 21-year-old student at MIT named Claude Shannon wrote a master's thesis, published the next year as "A Symbolic Analysis of Relay and Switching Circuits". Shannon had worked on Vannevar Bush's differential analyzer, a room-sized calculating machine full of electromechanical parts, and had studied its tangle of switching circuits. He noticed that circuits of relays (switches opened and closed by an electromagnet) obey exactly Boole's algebra. That cut both ways. You could use the algebra to simplify messy circuits, such as the relay banks in telephone exchanges, and you could build circuits that do logic. The psychologist Howard Gardner later called it "possibly the most important, and also the most famous, master's thesis of the century."

  Here is why that matters. Numbers can be written in binary, using only 0 and 1, and 0 and 1 can be "off" and "on". Once numbers are switches, arithmetic becomes logic. Adding two one-digit binary numbers needs just two rules:

  
    ABSum digit (A XOR B: "one or the other, not both")Carry (A AND B)
    0000
    0110
    1010
    1101 (so 1 + 1 = binary 10, which is 2)
  

  Chain a few dozen of these "adders" together and you can add large numbers. Add circuits that compare, copy and choose, and you have the parts of a calculator. Add a way for the machine to follow a list of instructions stored in its own memory, and you have a computer.

  A computer doesn't know arithmetic. It knows on and off, and a handful of rules for combining them.

  A year before Shannon's thesis, Alan Turing had described, on paper, a machine that could carry out any computation by following simple rules (the companion piece on Turing lets you run one). Turing settled what a machine could compute. Shannon showed how switches could compute it. What was missing was a switch fast enough to make it worthwhile.

  Part 2 — Valves, Colossus and ENIAC
  Relays click. They are physical levers moved by magnets, so they are slow by electronic standards, and they wear out. The faster switch already existed in every radio: the vacuum tube, or "valve" in British English. John Ambrose Fleming patented a two-electrode tube in 1904, and Lee de Forest added a third electrode, the grid, in 1906–08. A small voltage on the grid controls a large current flowing through the tube. Used gently it is an amplifier; driven hard between fully off and fully on, it is a switch with no moving parts, able to flip far faster than any relay.

  The catch: a tube is a hot, glowing filament in a glass bulb. It uses a lot of power and burns out like a light bulb, so a computer of tubes meant thousands of little heaters, some always failing.

  Two wartime projects did it anyway:

  
    Colossus (Britain)ENIAC (United States)
    Built byTommy Flowers, a Post Office engineer, from plans by mathematician Max NewmanJohn Mauchly and J. Presper Eckert, Moore School, University of Pennsylvania
    WorkingPrototype by December 1943; Mark 2 ran on 1 June 1944Completed 1945; announced February 1946
    Tubes1,600 (Mark 1), 2,400 (Mark 2)About 18,000
    JobBreaking the German Lorenz teleprinter cipher (not Enigma)Artillery firing tables for the US Army, then much else
    FateKept secret until the mid-1970sFront-page news; retired in October 1955
  

  ENIAC weighed more than 27 tonnes, drew about 150 kilowatts and could do some 5,000 additions a second. According to the Computer History Museum, it lost a tube every day or two, and its maintenance crews learned to find and fix a failure in about fifteen minutes. It had no program in the modern sense. To give it a new problem, its operators rewired it with cables and set thousands of switches, which could take days. The people who worked out how to do that were six women, Kay McNulty, Betty Jennings (later Jean Bartik), Betty Snyder (later Holberton), Marlyn Wescoff, Ruth Lichterman and Fran Bilas, whose part in the story was long overlooked.

  "Which was the first computer?" has no single answer; it depends on the definition. Colossus was electronic and programmable, but built for one job and hidden by secrecy. ENIAC was general-purpose but programmed by rewiring. The idea that made both obsolete was written down in June 1945, in John von Neumann's First Draft of a Report on the EDVAC: keep the program in the same memory as the data, as a list of numbers the machine reads and obeys. Credit for that idea is disputed too. Eckert and Mauchly had discussed it before von Neumann wrote it up, and the report carried only his name. Almost every computer since, including the toy one later in this piece, is a "stored-program" machine.

  ENIAC was reprogrammed with cables. The idea that replaced it was simple: store the instructions as numbers, in the same memory as everything else.

  Part 3 — The transistor
  The telephone network of the 1940s ran on tubes and mechanical relays, with all their drawbacks. At Bell Labs, the research arm of the American telephone system, Mervin Kelly encouraged the physicist William Shockley in early 1945 to set up a group studying semiconductors: materials such as germanium and silicon, which conduct electricity better than an insulator but worse than a metal, and whose conductivity can be controlled.

  On 16 December 1947, two members of that group, John Bardeen and Walter Brattain, pressed two gold contacts, very close together, onto a slab of germanium using a plastic wedge. A small signal fed in at one contact came out up to 100 times stronger. A week later they showed it to the lab's managers. A colleague, John Pierce, came up with the name transistor, and Bell Labs announced it at a press conference in New York on 30 June 1948.

  The human story is less tidy. Shockley, the group's leader, was not one of the two inventors, and the two sides disagreed about how the device worked. Within weeks, on 23 January 1948, he conceived a different, sturdier design, the junction transistor, which became the basis of the industry. The three men shared the 1956 Nobel Prize in Physics "for their researches on semiconductors and their discovery of the transistor effect." Two other claims hover over the credit. In 1925–28 the physicist Julius Lilienfeld patented the idea of a field-effect transistor, but never published a working device. And in 1948, in Paris, Herbert Mataré and Heinrich Welker independently built a point-contact transistor of their own, which their company rushed into production for the French telephone network.

  How a transistor switches. Nearly all the transistors in today's chips are a later type, the MOSFET, demonstrated at Bell Labs by Mohamed Atalla and Dawon Kahng in 1959–60. Think of it as a tap for electricity. There are two terminals, the source and the drain, with a strip of silicon between them that normally doesn't conduct. Above that strip, separated by a whisper-thin layer of insulating glass, sits a third terminal, the gate. Put a voltage on the gate and its electric field pulls charges into the strip, opening a conducting channel: the tap is on. Remove the voltage and the channel vanishes: off. Nothing moves except electrons, and nothing needs to glow. Why silicon behaves this way is a story from quantum physics (see Quantum Mechanics: The Weirdness at the Bottom in Physics).

  
    RelayVacuum tubeTransistor (MOSFET)
    What movesA metal armElectrons through a vacuumElectrons in a crystal
    HeatSomeLots: a glowing filamentVery little per switch
    Wears out?Yes, mechanicallyYes, burns outEssentially no
    Can be shrunk?BarelyBarelyEnormously, as we'll see
  

  That last row is the whole story of the next seventy years. A widely cited estimate puts the number of MOSFETs ever made at more than 13 sextillion (13 followed by 21 zeros) by 2018, which would make them the most numerous things humans have ever manufactured.

  Shockley left Bell Labs and, with backing from the instrument maker Arnold Beckman, set up Shockley Semiconductor Laboratory in Mountain View, California, in 1955–56, the first silicon company in what became Silicon Valley. He hired brilliantly and managed terribly; a biographer called him possibly "the worst manager in the history of electronics." In September 1957 eight of his researchers, including Robert Noyce, Gordon Moore and Jean Hoerni, quit together to start Fairchild Semiconductor. Shockley called them the "traitorous eight". In his later decades he became notorious for promoting racist eugenic ideas, which were rejected by scientists then and since. He is a reminder that a Nobel Prize in physics confers no authority on anything else.

  Part 4 — The chip
  By the late 1950s transistors were small and reliable, but every one still had to be made separately and wired to its neighbours by hand. A complex circuit meant thousands of hand-soldered joints, any of which could fail. The more ambitious the machine, the worse the problem.

  The answer was to make all the parts, and the connections between them, out of one piece of semiconductor. Two men got there months apart, and the credit has been shared ever since:

  
    Jack Kilby, Texas InstrumentsRobert Noyce, Fairchild
    WhenBuilt his first "solid circuit" on 12 September 1958Patent filed 30 July 1959; first working chips 26 May 1960
    MaterialGermaniumSilicon
    WiringFine gold "flying wires" between the parts: proof of concept, impractical to mass-produceFlat metal lines laid over an insulating oxide layer, printed with the rest of the chip
    Built onHis own designJean Hoerni's "planar" process (1959) and Kurt Lehovec's way of isolating parts
  

  Hoerni's contribution is often forgotten. Chipmakers used to strip off the protective oxide layer after using it as a stencil. Hoerni left it on, which protected the delicate junctions and left a flat ("planar") surface on which metal wiring could be printed. One historian called it "the most important innovation in the history of the semiconductor industry." Texas Instruments and Fairchild fought over the patents for years before settling on a cross-licence. The Computer History Museum's verdict is that Kilby made the first working circuit with all its parts made of semiconductor, Noyce made the version that could be manufactured, and both are fairly called co-inventors. Kilby received a share of the 2000 Nobel Prize in Physics "for his part in the invention of the integrated circuit," alongside Zhores Alferov and Herbert Kroemer, who were honoured for different work. Noyce had died in 1990, and the prize is not awarded posthumously.

  Chips were expensive at first, so their first big customers were those who would pay anything for small and light. Every Apollo Moon flight carried two guidance computers built from Fairchild Micrologic chips, and the programme needed about 200,000 of them, making it the largest user of integrated circuits through 1965.

  A computer on one chip. In 1968 Noyce and Moore left Fairchild to found Intel. In 1969 a Japanese calculator company, Busicom, asked Intel to design a set of twelve custom chips for a printing calculator. Ted Hoff and Stanley Mazor suggested something more radical: four chips, one of them a general-purpose processor that would run a program stored in memory, so the same hardware could become any calculator by changing the program. Federico Faggin, assisted by Masatoshi Shima, did the hard work of laying it out. The result, the Intel 4004, packed 2,300 transistors into a 16-pin package and ran at up to about 740 kHz, i.e. 740,000 clock ticks a second.

  There was a twist. Busicom owned the rights. The calculator market was collapsing, Busicom wanted a lower price, and in May 1971 Intel bought back the rights for uses other than calculators by refunding Busicom's $60,000. Intel advertised the 4004 in Electronic News on 15 November 1971 as "a new era of integrated electronics." The "first microprocessor" title is contested here too: Four-Phase Systems had a processor slice working in 1969, and Garrett AiResearch built a chip set for the F-14 fighter's air-data computer in 1970. The 4004 is the one that was sold to anyone who wanted it, and that made all the difference.

  The first commercial microprocessor began life as a calculator part, and its maker had to buy it back from a customer.

  Part 5 — Inside the machine: fetch, decode, execute
  Strip away the billions of transistors and a processor (a CPU, or central processing unit) is surprisingly simple to describe. It has:
  Memory, a long row of numbered boxes ("cells"), each holding a number. The program and the data live here together.
  Registers, a few boxes inside the processor itself, much faster to reach than memory. Our machine has one, the accumulator, where arithmetic happens.
  The program counter, a register that holds the address of the next instruction.
  The instruction register, which holds the instruction being worked on.
  An arithmetic unit built from adders like the one in Part 1, and a control unit that runs the endless cycle:

  Fetch → Decode → Execute → repeat

  Fetch: copy the number at the address in the program counter into the instruction register, and add 1 to the program counter. Decode: work out what that number means. Execute: do it. That's it. Every program you have ever used, from a calculator to a video game to a chatbot, is this loop running billions of times a second.

  The machine below is real in the sense that matters: it is a complete stored-program computer. It has 25 memory cells and understands five instructions. Each instruction is a three-digit number: the first digit says what to do, the last two say which cell to do it to. So 120 means "LOAD from cell 20".

  
    CodeInstructionWhat it does
    1xxLOAD xxCopy the number in cell xx into the accumulator
    2xxADD xxAdd the number in cell xx to the accumulator
    3xxSTORE xxCopy the accumulator into cell xx
    4xxJZ xx (jump if zero)If the accumulator is 0, set the program counter to xx
    000HALTStop
  

  
    Interactive · A toy CPU you can step through
    
      Add two numbers
      Count down
      Multiply by adding
    
    
    FetchDecodeExecute
    
      Program counter0next cell to fetch
      Instruction reg.–empty
      Accumulator0working number
    
    
      Tick (one phase)
      Next instruction
      Run to HALT
      Reset
    
    
    
    program counter points herejust read or written
    A teaching machine: one register, 25 cells, numbers from −999 to 999, three clock phases per instruction. Real processors overlap these phases and have hundreds of instructions, but the loop is the same. The simulator stops any program that hasn't halted after 1,000 instructions.
  

  Three things to notice as you step through:
  Instructions are just numbers. Cell 0 of "Add two numbers" holds 120, and nothing marks it as an instruction except that the program counter happens to point at it. If a program stored a number into its own code area, it would change its own instructions. That is the stored-program idea, with all its power and danger; a whole class of security attacks consists of tricking a computer into treating data as code.
  Loops are built from one jump. Our machine has no "go to" instruction, only "jump if zero". So "Count down" ends its loop by loading a cell that always holds 0 and then doing JZ 0, a jump that always happens. Real instruction sets have more conveniences, but every loop, if-statement and function call compiles down to jumps of this kind.
  Not every program stops. Set "Count down" to start from −1 and run it. It subtracts forever, never hitting zero. Turing proved in 1936 that no method can look at every possible program and decide whether it will halt; our simulator just gives up after 1,000 instructions.

  Multiplying 4 by 3 takes this machine 30 instructions and 90 clock ticks. A modern chip does far harder sums than that billions of times a second, but by the same loop.

  Part 6 — Making it fast
  If every processor runs the same loop, why is a 2026 phone so much faster than a 4004? There are four main reasons.

  1. The clock. A processor's steps are paced by a clock, an oscillating signal that ticks billions of times a second. The 4004 managed about 740,000 ticks a second. Modern processors run at a few billion (several gigahertz), roughly 5,000 times faster. Faster switching came from smaller transistors, as Part 7 explains, but around 2005 clock speeds stopped rising much, for a reason we'll come to.

  2. Caches. Our toy machine reads memory instantly. Real memory is not instant. The main memory chips sit centimetres away and are made for size, not speed, and the processor has outrun them for decades. Reaching main memory can be tens to hundreds of times slower than reaching a cache, a small, fast memory on the processor chip that holds copies of recently used data. Think of a desk, a bookshelf and a warehouse: what you're using sits on the desk (registers), what you used recently on the shelf (cache, often in two or three levels), and everything else in the warehouse (main memory). Most trips stop at the desk or the shelf.

  3. Pipelining. Our machine finishes one instruction before starting the next. A laundry doesn't work like that: while load 1 is in the dryer, load 2 can be in the washer. Processors do the same, overlapping the phases of different instructions:

  
    Clock tick12345
    Instruction 1FetchDecodeExecute
    Instruction 2FetchDecodeExecute
    Instruction 3FetchDecodeExecute
  

  Without overlap, three instructions take nine ticks; with it, five, and once the pipeline is full one instruction finishes every tick. Real pipelines have many more stages, and modern processors go further: they work on several instructions at once, and reorder them when one is waiting for memory.

  Pipelines have an awkward problem: jumps. When a JZ arrives, the processor doesn't yet know whether it will jump, but it has already started fetching what comes next. So it guesses, using a "branch predictor" that remembers what each jump did last time, and works ahead speculatively. A wrong guess means throwing the work away. Speculation is a big part of modern speed, and it has a cost. On 3 January 2018 researchers disclosed Spectre and Meltdown, flaws that let a malicious program learn secrets from traces left behind by discarded speculative work. They affected processors from Intel, AMD and IBM, and designs based on ARM.

  4. More cores. When clock speeds stalled, chipmakers used their growing transistor budgets to put several complete processors, called cores, on one chip. Apple's M1 of 2020, for example, has eight CPU cores. But extra cores help only if the work can be split up. In 1967 Gene Amdahl pointed out the limit, now called Amdahl's law: the part of a job that must be done in sequence caps the speed-up. If 90% of a task can be split up, eight cores make it about 4.7 times faster, not eight; even with infinitely many cores the ceiling is ten. That limit is exactly why a different kind of chip, built for work that splits into millions of identical pieces, took over graphics and then AI. That is the subject of the next piece.

  Part 7 — Moore's law
  On 19 April 1965 the magazine Electronics published an article by Gordon Moore, then director of research at Fairchild, called "Cramming more components onto integrated circuits." Its key sentence: "The complexity for minimum component costs has increased at a rate of roughly a factor of two per year." Moore extrapolated that by 1975 the most economical chip would hold about 65,000 components.

  Read that sentence again, because it is more subtle than the slogan. Moore was talking about cost. For any given manufacturing process there is a sweet spot: too few components per chip and you waste the chip; too many and so many chips are faulty that the cost per good part rises. Moore observed that the sweet spot was doubling every year. In 1975, by then at Intel, he revised the rate to a doubling about every two years. The physicist Carver Mead gave it the name "Moore's law". The often-quoted "every 18 months" was not Moore's figure. It came from an Intel executive, David House, who estimated that chip performance would double every 18 months, and Moore later said he never said 18 months.

  
    ChipYearTransistors
    Intel 400419712,300
    Intel 8086197829,000
    Intel 803861985275,000
    Intel Pentium19933.1 million
    Apple M1 (a laptop chip)202016 billion
    Apple M3 Max202392 billion
    NVIDIA Blackwell GPU (two dies acting as one)2024208 billion
  

  From the 4004 to the M1 is about 23 doublings in 49 years: one roughly every 2.2 years, close to Moore's 1975 prediction for half a century. About seven million times more transistors. (Why something that doubles regularly becomes unimaginably large is the theme of Why Exponential Growth Rules Everything in Mathematics.)

  It was never a law of nature. It was an observation that became a target. The whole industry, from the makers of chemicals and machines to the chip designers, planned their research around the next doubling, which helped make it come true. Intel itself now describes it as a projection, not a scientific law.

  The free lunch, and its end. For three decades, shrinking transistors gave a bonus beyond numbers. In 1974 Robert Dennard and colleagues at IBM showed that if you shrink a transistor and lower its voltage in proportion, it switches faster while the power used per square millimetre stays the same. So every new generation was denser, faster and no hotter. This "Dennard scaling" broke down around 2005–07. Transistors became so small that they leak current even when "off", and voltages couldn't keep falling. Running them ever faster would have made chips too hot to cool. That is why clock speeds stalled in the mid-2000s and why your laptop has many cores rather than one 50-gigahertz one.

  Slowing, not stopping. Transistor counts kept climbing, but more slowly and at much greater cost. In 2015 Intel's chief executive said its cadence was "closer to two and a half years than two." In September 2022 NVIDIA's Jensen Huang declared Moore's law dead; Intel's Pat Gelsinger publicly disagreed within days, though by the end of 2023 he too said the industry was "no longer in the golden era of Moore's Law." Who is right depends partly on what you measure: transistor counts at the top end are still climbing, helped by joining several pieces of silicon into one product, while the pace of each manufacturing generation has clearly slowed.

  Moore's law was an economic observation that became a schedule. The industry kept it for fifty years because it decided to.

  Part 8 — The languages of chips: x86, ARM and RISC-V
  Our toy machine's five instructions are its instruction set architecture, or ISA: the vocabulary that software is translated into. A program compiled for one vocabulary won't run on a chip that speaks another, which is why the choice of ISA matters so much commercially. Three dominate the conversation in 2026.

  x86. Intel's 8086 of 1978 began the family, and a cheaper variant, the 8088, went into the IBM PC in 1981. Because so much software was written for the PC, every later Intel and AMD chip kept speaking the same language, extended over time: to 32 bits with the 80386 in 1985 and to 64 bits by AMD between 1999 and 2003, an extension Intel then adopted. x86 is a "complex instruction set" (CISC): it has many instructions, some doing a lot of work. Since the mid-1990s, x86 chips have translated those complex instructions internally into simpler "micro-operations", so inside they work much like the next family. Only Intel and AMD actively make modern x86 designs.

  RISC and ARM. In the late 1970s and early 1980s, researchers at IBM (John Cocke's 801 project, begun in 1975), at Berkeley (David Patterson) and at Stanford (John Hennessy) argued for a "reduced instruction set" (RISC): a small set of simple, regular instructions that a chip can pipeline efficiently, leaving the cleverness to the compiler. Hennessy and Patterson shared the 2017 Turing Award, computing's highest honour, for their quantitative approach to designing processors. In Cambridge, England, the home-computer maker Acorn took the idea and started its own RISC chip in October 1983. Sophie Wilson designed the instruction set, working with Steve Furber; the first ARM chips came back from the factory and worked on 26 April 1985. In 1990 the team was spun off, with Apple and VLSI Technology as partners, as Advanced RISC Machines. Arm's business model is unusual: it designs processors and licenses the designs and the instruction set, and other companies build the chips. Simple designs used little power, which suited battery devices, and Arm says more than 350 billion Arm-based chips have shipped. ARM designs run the iPhone and most Android phones. In November 2020 Apple launched the M1, its own ARM-based chip, and began replacing the Intel processors in its Macs.

  RISC-V. In May 2010 Krste Asanović, Yunsup Lee and Andrew Waterman at Berkeley started a new instruction set that anyone could use without paying royalties. RISC-V (pronounced "risk five") is an open standard, like the rules of the web, rather than a product. Its governing body was incorporated in Switzerland in March 2020, after announcing the move with reference to concerns over US trade regulations. It has attracted strong interest in China, where an ISA that no foreign government can license away has obvious appeal.

  
    x86ARMRISC-V
    Born1978, Intel1983–85, Acorn2010, UC Berkeley
    StyleComplex (CISC), simple insideReduced (RISC)Reduced (RISC), modular
    Who may build itIn practice Intel and AMDArm's licenseesAnyone, royalty-free
    StrongholdsPCs, serversPhones, tablets, Macs, growing in serversEmbedded controllers, growing
  

  Part 9 — How chips are made
  A chip starts as a disc of extremely pure crystalline silicon, a wafer, today usually 300 mm (12 inches) across. On it, hundreds of identical chips are built at once, layer by layer, by a printing process called photolithography, adapted to silicon at Bell Labs in 1955 by Jules Andrus and Walter Bond.

  Coat → Expose → Develop → Etch or implant → Repeat

  The wafer is coated with a light-sensitive "resist". Light is shone through a mask, a stencil of one layer of the circuit, and focused through optics that shrink the pattern onto the wafer. Where the light lands, the resist changes chemically and can be washed away, leaving bare patches. Those patches are etched, or bombarded with impurity atoms to change the silicon's electrical behaviour, or filled with metal. Then the wafer is cleaned, coated and exposed again for the next layer. The transistors are built at the bottom; above them, many layers of copper wiring connect them. Each layer must line up precisely with the one below, all in "clean rooms" where a speck of dust can ruin a chip.

  The basic limit is the wavelength of the light: you can't easily print lines much finer than the light you print them with. For years the workhorse was deep ultraviolet light at 193 nm, stretched with tricks such as putting highly purified water between the lens and the wafer, and splitting one layer's pattern across several masks. The next step, extreme ultraviolet (EUV) at 13.5 nm, took decades. ASML, the Dutch company that makes the machines, creates the light by firing two laser pulses at a tiny droplet of molten tin, up to 50,000 times a second, turning it into a glowing plasma. EUV is absorbed by everything, air and glass included, so the whole light path is in a vacuum and the "lenses" are mirrors, each reflecting only around 70% of the light. EUV entered high-volume manufacturing in 2019.

  Only ASML makes EUV machines. Its standard models print features down to about 13 nm; its newer "High-NA" machines, the first delivered in December 2023, reach about 8 nm. Reported prices are around $200 million for a standard EUV tool and roughly $370 million for a High-NA one; a machine weighs about 180 tonnes and travels in three Boeing 747s. ASML reported 2025 sales of about €32.7 billion and employs more than 44,000 people.

  The most advanced chips in the world depend on a machine that only one company can build, which makes light by vaporising droplets of tin 50,000 times a second.

  Who makes the chips. Leading-edge factories, "fabs", are so expensive that only a handful of companies can build them; in March 2025 TSMC announced a further $100 billion of investment at its Arizona site alone. In 1987 Morris Chang founded the Taiwan Semiconductor Manufacturing Company (TSMC) on a then-odd idea: a "pure-play foundry" that makes chips only for other companies and designs none of its own. That let "fabless" companies such as Apple, NVIDIA, AMD and Qualcomm design chips without owning a factory. In 2025 TSMC made 12,682 different products for 534 customers. Intel and Samsung both design and manufacture; Intel has also opened its factories to outside customers.

  What "3 nm" really means. Chip generations are named by "nodes": 7 nm, 5 nm, 3 nm, 2 nm. Until the mid-1990s the number roughly matched the length of a transistor's gate. Since then the names have drifted into marketing. As IEEE Spectrum put it, most of the critical features of a "7-nm" transistor are considerably larger than 7 nm. TSMC's "3 nm" process, in volume production since December 2022, has a spacing between transistor gates of roughly 45–48 nm and its tightest metal wires are about 23 nm apart. Nothing on the chip measures 3 nm. The node name means "our next generation, denser and more efficient than the last", and one company's 3 nm isn't directly comparable to another's.

  
    FeatureSize
    The name "3 nm"A label, not a measurement
    Gate pitch, TSMC 3 nm (distance between neighbouring transistor gates)About 45–48 nm
    Tightest metal wiring pitch, TSMC 3 nm familyAbout 23 nm
    A silicon atomAbout 0.2 nm across
    Intel 4004 (1971) process10 µm = 10,000 nm
  

  What is changing at the leading edge is the shape of the transistor. Around 2011 the industry stood the channel up as a thin vertical fin (the "FinFET") so the gate could grip it on three sides. Now it is moving to "gate-all-around" designs, in which the channel is a stack of thin horizontal sheets with the gate wrapped all the way round: Samsung began production of its version in 2022, TSMC's first nanosheet process (N2) entered volume production in the last quarter of 2025, and Intel's 18A process ("RibbonFET") is in high-volume production in the United States. The next trick is moving the power wiring to the back of the chip, so signals and power stop competing for space on the front.

  Part 10 — The state of affairs in 2026, and the open questions
  A boom. According to the Semiconductor Industry Association, global chip sales in August 2026 were $159.7 billion for the month (a three-month moving average), 144% higher than a year earlier, and sales for 2026 passed $1 trillion by the end of August, the first time any year has reached that. The industry statisticians WSTS, in their spring 2026 forecast, expected the year to reach about $1.51 trillion, up 90%, with memory chips alone more than tripling in value. The fastest growth is in chips for data centres: NVIDIA alone reported $89 billion of data-centre revenue in the three months to late July 2026, as the next piece describes. TSMC's revenue for July–September 2026 was $46.7 billion, up 41% in dollars on a year earlier; 2 nm chips already made up 8% of its wafer revenue, 3 nm 33%, and processes of 7 nm or finer 79%.

  The geography of chips. Design is spread around the world, but the most advanced manufacturing is concentrated: overwhelmingly at TSMC, in Taiwan, with Samsung in South Korea and Intel in the US the other contenders. According to a 2024 report by the SIA and Boston Consulting Group, the US share of world chip-making capacity fell from 37% in 1990 to 10% in 2022, and the US had no capacity below 10 nm in 2022, projected to rise to 28% by 2032 on the projects then announced. That concentration, on an island that China claims, is what turned chips into a matter of national security. The main moves, dated:

  
    WhenWhat
    9 Aug 2022US CHIPS and Science Act signed: $52.7 billion, including $39 billion in manufacturing incentives, plus a 25% investment tax credit
    7 Oct 2022US export controls restrict advanced AI chips, chipmaking tools and US engineers' help to China's leading-edge fabs
    2023–2024The Netherlands requires licences for exports of ASML's advanced machines, tightening the rules several times
    21 Sep 2023EU Chips Act in force
    13 May 2025US rescinds the Biden-era "AI Diffusion Rule" (issued January 2025), which would have tiered countries' access to AI chips
    22 Aug 2025US government converts $8.9 billion of Intel's grants into a 9.9% stake in the company
    13 Jan 2026US moves to case-by-case licences for exports of NVIDIA's H200 and AMD's MI325X to China, under conditions, after a presidential announcement in December 2025
    June 2026European Commission proposes a "Chips Act 2.0"
  

  China, meanwhile, is investing heavily in its own chip industry. In June 2026 a Chinese machine, LineShine at the National Supercomputing Centre in Shenzhen, took first place on the TOP500 list of supercomputers at 2.2 exaflops, using only CPUs: home-grown 304-core processors, 13.8 million cores in all, with no separate accelerators. It shows that China can build a world-leading machine from domestic processors; what it says about the effect of export controls is argued both ways.

  The arguments, at their strongest. Supporters of export controls argue that advanced chips are the engine of military AI, codebreaking and weapons design, that it would be reckless to sell them to a strategic rival, and that controls buy time. Critics argue that controls cost Western firms revenue that pays for their research, push China to build its own alternatives faster, and strain relations with allies whose companies lose sales; some also argue that the rules have shifted too often for businesses to plan around. On industrial policy, supporters say subsidies are the price of resilience against a blockade, an earthquake or a war in the Taiwan Strait, and match what rivals already do. Critics say they spend public money on profitable companies, risk building fabs the market won't fill, and, with the government holding equity in Intel, blur the line between state and market. Both sides agree on the underlying fact: the world depends on a very small number of companies and places for its most important technology.

  Open questions.
  How far can shrinking go? Gate-all-around transistors, backside power and, after that, transistors stacked on top of each other are on the research roadmaps; the Belgian research centre IMEC's roadmap runs to the mid-2030s. Increasingly, progress comes from packaging instead: several smaller "chiplets" joined side by side or stacked, which is how NVIDIA's Blackwell gets 208 billion transistors into one product.
  Who pays? Each generation's factories and machines cost more, and fewer companies can afford the leading edge.
  Power. A chip's performance is now limited more by heat than by transistor count, which is why so much effort goes into specialised processors that do one kind of work efficiently.
  Whose vocabulary? Whether RISC-V becomes a third mainstream ISA, and whether politics splits the world into separate chip ecosystems, are both genuinely open.

  Underneath all of it, the machine is the one from Part 5: fetch, decode, execute, a few billion times a second, on switches too small to see.

  
    The takeaway
    A processor is switches doing logic: Boole's algebra (1854), wired into circuits by Shannon (1937), run first with tubes (Colossus, ENIAC), then transistors (Bell Labs, 1947), then integrated circuits (Kilby 1958, Noyce 1959), then whole processors on one chip (Intel 4004, 1971: 2,300 transistors). Every CPU runs the same fetch–decode–execute loop on instructions stored as numbers; speed comes from faster clocks, caches, pipelines and more cores. Moore's law, a doubling roughly every two years since 1975, was an economic observation the industry turned into a schedule. It has slowed, and since about 2005 smaller transistors no longer give faster clocks for free. "3 nm" is a label, not a size. In 2026 the most advanced chips depend on one Dutch toolmaker and one Taiwanese foundry, which is why governments now subsidise fabs and control exports.
    Try this: in the toy CPU, load "Multiply by adding", set it to 9 times 9 and press "Run to HALT". Count how many instructions it took (9 × 9 + 3 = 84), then imagine doing it at three billion ticks a second.
  

  
    Sources & Further Reading
    Sources. Computer History Museum, The Silicon Engine timeline (transistor, junction transistor, photolithography, Shockley and the "traitorous eight", Kilby, Hoerni, Noyce and the credit verdict, Moore's law, Dennard scaling, the microprocessor). CHM Revolution exhibit (ENIAC, Colossus, Apollo). Intel's Virtual Vault (the 4004, Intel's founding) and Moore's law page. Nobel citations for 1956 and 2000 as quoted in the Wikipedia articles "Transistor" and "Jack Kilby" (nobelprize.org was not reachable). IEEE Spectrum on node names and the clock-speed plateau. ASML, "EUV lithography systems" and "About ASML". TSMC: company profile; N2 technology page; Q3 2026 results (15 Oct 2026). Apple Newsroom: M1 (10 Nov 2020). Arm, company page (350 billion+ chips). RISC-V International, "About". SIA, "Year-to-Date Global Semiconductor Sales Top $1 Trillion Through August" (5 Oct 2026) and Emerging Resilience in the Semiconductor Supply Chain (with BCG, 2024). WSTS spring 2026 forecast. Intel press release on the US government agreement (22 Aug 2025). US Bureau of Industry and Security press releases of 13 May 2025 and 13 Jan 2026. European Commission, "European Chips Act" (updated 28 Oct 2026). TOP500, June 2026 list and announcement (23 Jun 2026). Wikipedia was used for leads and secondary details (Boole, Shannon, Colossus, ENIAC, the EDVAC report, Lilienfeld, Shockley, x86, ARM, transistor counts, Moore's law quotations, Dennard scaling, Spectre, Amdahl's law, EUV, process nodes, ASML figures, the CHIPS Act and the 2022 controls), cross-checked where possible. Computed in node and Python: transistor doublings (4004 to M1: 22.7 in 49 years), clock ratio (about 5,400×), Amdahl's law (90% on 8 cores: 4.71×), and every program in the toy CPU (instruction counts 6k + 3 for counting down from k and 9b + 3 for multiplying by b).
    
      📘 Chip War: The Fight for the World's Most Critical Technology — Chris Miller (Scribner, 2022). The best single history of the industry and its geopolitics, from the transistor to TSMC and export controls; FT Business Book of the Year.
      📘 The Elements of Computing Systems — Noam Nisan and Shimon Schocken (MIT Press). Build a computer from logic gates up, yourself. The free course is at nand2tetris.org; the toy CPU above is its first rung.
      📘 Crystal Fire: The Invention of the Transistor and the Birth of the Information Age — Michael Riordan and Lillian Hoddeson. The Bell Labs story, including the Shockley drama, from people who interviewed the participants.
      🌐 Computer History Museum: The Silicon Engine timeline. Short, authoritative pages on every milestone from 1833 to 1979, with photos.
      🌐 Computer History Museum: Revolution, the first 2000 years of computing. The online exhibit, from the abacus to the web.
      🌐 IEEE Spectrum: A Better Way to Measure Progress in Semiconductors. Why node names stopped meaning anything, and what engineers propose instead.
      🌐 Companion pieces in this library: Alan Turing: The Machine That Defined Computing (History-Makers), Why Exponential Growth Rules Everything (Mathematics), What It's Like to Be a Software Engineer (Careers), Artificial Intelligence: Where It Came From and Where It Stands (Technology 1) for what all this computing power is now used for, and the next piece in this series, GPUs: From Video Games to the Engine of AI.


GPUs: From Video Games to the Engine of AI
------------------------------------------
A chip designed to colour in millions of pixels for video games turned out to be exactly what artificial intelligence needed. This is how that accident happened, how the chip works, and why in 2026 it sits at the centre of a trillion-dollar industry and a geopolitical fight.

Follows: Processors: How a CPU Thinks, From the Transistor to the Chip in Your Pocket, Vectors and Matrices: The Geometry of Linear Transformations and How Large Language Models Actually Work
In this deep dive
1. Painting a frame: what "rendering" actually means
2. A few smart workers, or thousands of simple ones
3. Before the GPU: arcades, Silicon Graphics and 3dfx
4. "The world's first GPU": the GeForce 256 claim, examined
5. From shaders to CUDA: graphics chips learn to do anything
6. 2012: AlexNet and two gaming cards in a bedroom
7. Why matrix multiplication made GPUs the AI engine
8. Tensor cores, stacked memory and the AI data centre
9. The state of affairs in 2026
10. Open questions and trade-offs

Part 1 — Painting a frame
Look at any modern video game for one second. In that second your screen has been completely redrawn sixty times or more. A standard 1080p screen has 1,920 × 1,080 = 2,073,600 pixels, so at 60 frames a second the computer must decide the colour of about 124 million pixels every second. At 4K it's about 500 million. And each colour isn't simply looked up: it is calculated from the shape, position, material and lighting of everything in the scene.
The previous piece in this series showed how a CPU works: a fetch–decode–execute loop running through instructions one after another, very fast. This piece is about a different kind of processor, the GPU (graphics processing unit), built for a different kind of job. To see why it exists, start with what drawing a 3D scene involves. It's called rendering, and it runs as an assembly line.
Vertices → Rasterisation → Pixel shading → Screen

Vertices. Every object in a game, a car, a face, a tree, is built from triangles, often millions of them. Each triangle has three corners, called vertices, each with an x, y and z position. To move the car, turn the camera or squash the 3D world onto your flat screen, every vertex is multiplied by a small grid of numbers, a matrix. (This is exactly the maths of Vectors and Matrices in the Mathematics section: a matrix is a recipe for moving every point in space at once.)
Rasterisation. Once the triangles have screen positions, the hardware works out which pixels each triangle covers, like laying a triangle of tracing paper over graph paper and ticking the squares underneath. It also keeps a "depth buffer" recording how far away the nearest thing drawn at each pixel is, so that a wall correctly hides what's behind it.
Pixel shading. For every covered pixel, a small program called a shader calculates the colour: which bit of the texture image lands there, how the light falls on it, whether it's in shadow, whether it's shiny.
Screen. The finished colours are written into a block of memory, the frame buffer, which is sent to the display. Then it all starts again for the next frame, about 16 milliseconds later.

Stage
What is calculated
How many times per frame (rough)

Vertex
New position of each corner
Once per vertex: often millions

Rasterise
Which pixels each triangle covers
Once per triangle

Pixel shade
The colour of each pixel
Once per pixel or more: 2 million at 1080p

Rendering one frame means doing the same small calculation millions of times, once for every corner and every pixel, and then doing it all again 16 milliseconds later.

Here is the crucial property. The colour of the pixel in the top-left corner doesn't depend on the colour of the one next to it. Each vertex can be transformed without waiting for the others. Graphics is millions of small, independent jobs. Computer scientists call this "embarrassingly parallel", and it is the whole reason the GPU looks the way it does.
Part 2 — A few smart workers, or thousands of simple ones
Imagine two decorating firms bidding to paint a wall-sized mosaic of two million tiles, each tile a single colour you are told in advance.
Firm C sends eight master craftsmen. Each is brilliant: fast, versatile, able to handle any surprise, and able to plan around a complicated job. Each paints one tile per second.
Firm G sends 2,048 apprentices. Each is four times slower than a master, can only follow simple instructions, and works in squads that must all do the same thing at the same moment. But there are a lot of them.
Here is the race, worked out (a toy model, but the arithmetic is exact):

Job
Firm C: 8 masters, 1 s per tile
Firm G: 2,048 apprentices, 4 s per tile
Winner

Paint 2,073,600 independent tiles
259,200 s
4,052 s
G, about 64 times faster

1,000 steps, each needing the previous result
1,000 s
4,000 s
C, 4 times faster

When the work splits into millions of independent pieces, the apprentices win by a mile. When every step depends on the last, only one worker can be busy at a time, and the master wins. That second row is Amdahl's law from the previous piece, in a different costume.
That is the real difference between a CPU and a GPU:

A CPU spends most of its transistors on making a few workers (cores) brilliant at sequential work: large caches so they rarely wait for memory, branch predictors that guess which way the program will go, machinery to reorder instructions on the fly. A laptop chip such as Apple's M1 has eight CPU cores.
A GPU spends its transistors on numbers of workers. NVIDIA's RTX 5090 gaming card of 2025 lists 21,760 "CUDA cores" and draws up to 575 watts. Those cores are much simpler than CPU cores and are grouped into squads, called warps on NVIDIA chips, of 32 threads that execute the same instruction in lock-step, each on its own data.

A caution on numbers: a "CUDA core" is not comparable to a CPU core. It is closer to one lane of a 32-lane road, with one driver giving directions to all the lanes. If threads in a warp need to do different things (some pixels in shadow, some not), the warp has to do each version in turn, and some lanes sit idle. GPUs are fast only when lots of data needs the same treatment.

A CPU is a few experts who can handle anything. A GPU is an army that is unbeatable as long as everyone is doing the same thing.

Part 3 — Before the GPU
Special graphics hardware is nearly as old as computer graphics. Arcade machines of the 1970s already used dedicated circuits to draw moving sprites; Namco's Galaxian of 1979 had multi-coloured sprites on a coloured background. In 1981 NEC's µPD7220 put a graphics display controller on a single chip for personal computers, and in 1986 Texas Instruments' TMS34010 was a fully programmable graphics processor.
The high end belonged to Silicon Graphics (SGI), founded in 1982 by the Stanford professor Jim Clark with a group of students and staff. Its secret weapon was the Geometry Engine, designed by Clark and Marc Hannah at Stanford: the first chip to do the geometry stage of the 3D pipeline, the matrix maths on vertices, in custom hardware. SGI's first graphics terminal shipped in 1984. Its workstations became the standard tools of Hollywood visual effects; one appears on screen in Jurassic Park (1993). In 1992 SGI opened its graphics programming interface to everyone as OpenGL, which became an industry standard. Then cheaper PC hardware ate its market, and SGI filed for bankruptcy in April 2009.
The PC revolution came from three SGI veterans, Ross Smith, Gary Tarolli and Scott Sellers, who founded 3dfx in August 1994. Their Voodoo Graphics chip reached shops in October 1996 on cards such as Orchid's Righteous 3D. It was an odd beast by modern standards: an add-in card that did only 3D. You kept your ordinary 2D video card and ran a short cable from it into the Voodoo, which took over the screen when a 3D game started. It did one job, rasterising and texturing triangles, and did it so well that PC games suddenly looked smooth and solid. 3dfx also wrote its own programming interface, Glide, for games to talk to the chip directly.
For a couple of years 3dfx was the name in PC gaming. Its Voodoo2 of 1998 could even split the work between two cards, each drawing alternate lines of the screen (an early form of the multi-chip tricks that matter so much in AI today). But it made a run of late and uncompetitive products, game developers moved to Microsoft's Direct3D interface rather than Glide, and in December 2000 its rival NVIDIA bought most of its assets. 3dfx filed for bankruptcy in 2002.

Era
Who
What the hardware did

1970s–80s
Arcade makers, NEC, TI
Draw sprites and 2D images

1980s–90s
Silicon Graphics
Full 3D pipeline in expensive workstations

1996–2000
3dfx and rivals
Rasterising and texturing on cheap PC cards; the CPU still did the geometry

1999 onward
NVIDIA, ATI (now AMD)
The whole pipeline on one consumer chip

Part 4 — "The world's first GPU"
NVIDIA was founded on 5 April 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem, to bring 3D graphics to games and multimedia. On 31 August 1999 it announced the GeForce 256, and it chose its words carefully. NVIDIA called it "the world's first GPU", and defined a GPU as "a single-chip processor with integrated transform, lighting, triangle setup/clipping, and rendering engines", able to process at least 10 million polygons a second.
The new part was hardware transform and lighting (T&L). Until then, cards like the Voodoo drew triangles, but the CPU had to do the vertex maths: moving, rotating and lighting every corner. The GeForce 256, made by TSMC with about 17 million transistors, did that on the graphics chip too. The whole assembly line from Part 1 was now on one consumer chip. NVIDIA's own corporate timeline today still reads, for 1999: NVIDIA "invents the GPU".
Is the claim true? It depends, as so often with "firsts", on the definition, and in this case the definition was written by the company making the claim.

The word was older. Sony used "GPU" in 1994 for the graphics chip, designed by Toshiba, in the original PlayStation.
The idea was older. SGI's Geometry Engine had done transform work in hardware since the early 1980s, on expensive workstations and across multiple chips.
The benefit took time. At launch, hardware T&L gave big gains only in some games, mostly on PCs with slower CPUs, and the card was expensive.

The fair verdict: the GeForce 256 was a real milestone, the first time a cheap consumer chip did the entire 3D pipeline. "World's first GPU" was also very effective marketing, a definition drawn neatly around one's own product. Both are true, and the term stuck so thoroughly that we now use it for chips that rarely draw a picture at all.

"First" claims are usually definitions in disguise. The trick is to ask who wrote the definition.

Part 5 — From shaders to CUDA
In 1999 the GPU's pipeline was fixed: it could transform and light vertices and texture pixels in the ways its designers had built in, and no others. Game developers wanted more control, and the next step changed everything.
The GeForce 3, released in February 2001, added programmable shaders: the vertex and pixel stages could now run small programs written by the game's developers, within limits. (A version of the chip powered Microsoft's first Xbox.) ATI's Radeon 9700 of 2002 pushed programmability further. A shader is just a short program that takes some numbers in (a position, a texture value, a light direction) and puts numbers out. It doesn't have to be about light at all.
Researchers noticed. If you disguise a maths problem as a graphics problem, storing your data in "textures" and writing your calculation as a "pixel shader", the GPU will happily run your calculation on millions of data points in parallel. By 2001 people were doing matrix multiplication on graphics hardware, and in 2003 two research groups independently showed GPUs beating CPUs at general linear algebra. This was called GPGPU, general-purpose computing on GPUs, and it was painful: you had to be an expert in graphics programming to do maths.
At Stanford, a PhD student called Ian Buck built a language, Brook, that hid the graphics disguise. In 2004 NVIDIA hired him and paired him with John Nickolls, its director of architecture for GPU computing, to build it properly. The result was CUDA. It arrived with NVIDIA's GeForce 8800 generation, launched in November 2006, which was NVIDIA's first "unified" design: instead of separate vertex and pixel units, 128 identical general processors that could be assigned to any task. The first public CUDA software kit followed in February 2007.
CUDA let a programmer write ordinary-looking code, in an extended version of the C language, that runs as thousands of threads at once. To add two lists of a million numbers you write the instruction for one pair, "add element i of list A to element i of list B", and launch a million threads, each with its own i. The hardware schedules them in warps across its cores.

Year
Step
What changed

1999
GeForce 256
Whole fixed-function 3D pipeline on one chip

2001
GeForce 3
Programmable vertex and pixel shaders

2001–05
GPGPU research
Maths disguised as graphics

2006–07
GeForce 8800 and CUDA
General processors, programmed in C-like code

2008
OpenCL
An open, cross-vendor alternative

2016
AMD ROCm
AMD's open-source answer to CUDA

For years CUDA looked like an expensive hobby. Most GPUs were bought by gamers, and scientific users were a niche. But NVIDIA kept investing in it, putting CUDA on every card, building libraries for common maths and supporting universities. When the moment came, an enormous body of software, and of programmers' habits, already ran on NVIDIA chips. Competitors have had better and worse hardware at various times since, but none had that head start. That head start in software is arguably NVIDIA's most durable advantage, more than any one chip.
Part 6 — 2012: AlexNet and two gaming cards in a bedroom
Neural networks, the subject of Artificial Intelligence: Where It Came From and Where It Stands at the start of this series, were an old idea: layers of simple artificial "neurons", each adding up weighted inputs from the layer before, with the weights adjusted by training on examples. By the 2000s they were out of fashion with many researchers, partly because they seemed to need impossible amounts of data and computing power.
Both arrived at once. The data: from 2006 the Stanford professor Fei-Fei Li led ImageNet, which grew to more than 15 million photographs labelled by hand in some 22,000 categories, much of the labelling crowdsourced through Amazon's Mechanical Turk. It was completed in 2009, and from 2010 an annual competition challenged programs to recognise objects in the photos. After two years, progress had become marginal.
The computing: at the University of Toronto, the graduate student Alex Krizhevsky had already written fast CUDA code for "convolutional" networks, a design suited to images. With his fellow student Ilya Sutskever and their adviser Geoffrey Hinton, he built a much bigger network than anyone had trained: about 60 million adjustable weights and 650,000 neurons, in eight layers. According to the Computer History Museum, it was trained on a computer with two NVIDIA graphics cards in Krizhevsky's bedroom at his parents' house.
The 2012 paper gives the details. The network made about 90 passes through 1.2 million training images, which took five to six days on two NVIDIA GTX 580 gaming cards with 3 GB of memory each. The model was too big for one card's memory, so they split it across two, the cards reading from each other's memory directly. In the ImageNet competition of 2012 it got the right answer among its top five guesses for all but 15.3% of test images. The second-best entry, using careful conventional methods, managed 26.2%.

AlexNet (2012)
Second place (2012)

Top-5 error
15.3%
26.2%

Approach
Deep convolutional neural network
Hand-designed image features plus classifiers

Hardware
Two GTX 580 gaming GPUs, 5–6 days
Conventional

The authors were frank about what had made it possible. They wrote that current GPUs, paired with a highly optimised implementation, were powerful enough to train such large networks, and that their results could be improved "simply by waiting for faster GPUs and bigger datasets to become available." That sentence turned out to be the business plan of the next decade.
Credit, as always, is shared. In 2011 Dan Cireşan and colleagues at the Swiss lab IDSIA had already published a fast GPU implementation of convolutional networks, with record results on several image benchmarks; the AlexNet paper itself cites their work. What AlexNet did was win the field's most visible benchmark by a margin nobody could ignore, and within a few years almost every leading computer-vision system was a neural network trained on GPUs. Hinton later shared the 2018 Turing Award with Yoshua Bengio and Yann LeCun for deep learning, and the 2024 Nobel Prize in Physics with John Hopfield. His own summary of AlexNet, quoted by the Computer History Museum when it released the original code with Google in 2025: "Ilya thought we should do it, Alex made it work, and I got the Nobel Prize."

The breakthrough of modern AI was trained on two gaming cards, and its authors wrote that it would get better simply by waiting for faster ones.

Part 7 — Why matrix multiplication made GPUs the AI engine
Why did hardware built for video games suit neural networks so well? Because underneath, they are the same calculation.
A layer of a neural network takes a list of numbers in (a vector: the pixels of an image, or the encoded meaning of a word) and produces a list of numbers out. Each output neuron multiplies every input by its own weight and adds them up. Written down, that is a matrix multiplication: the input vector times a grid of weights. Then a simple non-linear step, such as replacing negatives with zero, and on to the next layer.
The numbers are large. A layer with 4,096 inputs and 4,096 outputs needs 4,096 × 4,096 = 16.8 million multiply-and-add operations for each input it processes. Process 4,096 inputs at once, as training does in batches, and it becomes a multiplication of two 4,096 × 4,096 matrices: 4,096³ = 68.7 billion multiply-adds. Training repeats that across dozens of layers, millions of examples and many passes.
Now notice the shape of the work:

Each output number is independent. Entry (i, j) of the answer needs row i of one matrix and column j of the other, and nothing else. All 16.8 million entries of a 4,096 × 4,096 result can be computed at the same time. That is the mosaic from Part 2, the apprentices' perfect job.
It's the same operation everywhere. Multiply, add, multiply, add, with no surprising branches. That suits warps of 32 lanes in lock-step.
It reuses data. Each number in the input matrices is used 4,096 times. So once the data is loaded into the chip, there is a great deal of arithmetic to do per byte fetched, which keeps the arithmetic units busy instead of waiting for memory.
Graphics was already this. Moving a vertex is multiplying it by a 4 × 4 matrix. GPU makers had spent twenty years building chips to do matrix arithmetic as fast as possible for games.

A video game and a neural network look nothing alike. To the chip they are the same thing: enormous numbers of small matrix multiplications, all independent.

Since 2017, when the paper "Attention Is All You Need" introduced it, the dominant AI design has been the transformer (explained in How Large Language Models Actually Work, Technology 2). It makes the fit even tighter: its core operation, "attention", is itself a set of matrix multiplications, and large language models are, computationally, mostly matrix multiplication at colossal scale.
Part 8 — Tensor cores, stacked memory and the AI data centre
Once AI became the biggest customer, GPUs started to change shape around it.
Tensor cores. In May 2017 NVIDIA announced the Tesla V100, built on its Volta design: 21.1 billion transistors on a chip of 815 mm², nearly the size of a postage stamp. Besides 5,120 ordinary arithmetic cores it had 640 tensor cores, special units that each do one thing: take two small 4 × 4 matrices, multiply them, and add a third, D = A × B + C, all in one step. NVIDIA's white paper rated the V100 at 15.7 trillion ordinary 32-bit operations per second, and 125 trillion on the tensor cores. The 68.7-billion-multiply-add example from Part 7 (137 billion operations, counting multiplies and adds separately) would take about 8.8 milliseconds at the first rate and about 1.1 at the second.
The trick includes using lower precision. The tensor cores multiplied 16-bit numbers (accumulating the results in 32-bit), where scientific computing traditionally insists on 64-bit. Neural networks turn out to tolerate rounding surprisingly well, and smaller numbers mean more of them per second and per byte of memory. The industry has kept going down: 8-bit formats, and with NVIDIA's Blackwell generation, 4-bit numbers, which NVIDIA says can double the size of model that fits in memory while keeping accuracy high.
That has a side effect worth knowing when you read headlines. In April 2025 Google said a full pod of 9,216 of its Ironwood AI chips delivers 42.5 exaflops and compared that with the El Capitan supercomputer's 1.7 exaflops. But the Ironwood figure is a peak rate for 8-bit numbers, while El Capitan's is measured running a demanding benchmark in 64-bit arithmetic. Both are "flops", but they are not the same unit of work.
The memory wall. A chip that can do a hundred trillion multiplications a second is useless if it can't be fed numbers fast enough. The answer is high-bandwidth memory (HBM): memory chips stacked on top of each other like a multistorey car park, connected by thousands of vertical wires through the silicon itself, and placed right next to the processor on a shared base layer. Each stack talks to the processor over 1,024 wires at once (2,048 in the newest standard, HBM4). HBM was developed by AMD and SK hynix, standardised in 2013 and first used in AMD's Radeon R9 Fury X in June 2015. The V100 had four stacks giving 900 gigabytes per second. An NVIDIA DGX B200 server, with eight Blackwell GPUs, has 1,440 GB of HBM and 64 terabytes per second. Only three companies make HBM: SK hynix, Samsung and Micron.
Bigger than a chip. The largest AI models don't fit on one GPU, or even one server, so the unit of computing has grown. NVIDIA's Blackwell GPU is two chips joined by a 10-terabyte-per-second link acting as one, 208 billion transistors in all. Its GB200 NVL72 rack wires 72 of those GPUs and 36 CPUs into one liquid-cooled system with 13.4 TB of HBM. Data centres then link thousands of such racks.
Power. All of this runs hot. The eight-GPU DGX B200 server is rated at up to about 14.3 kilowatts, roughly 1.8 kW per GPU including its share of the server: the whole box draws about as much as half a dozen electric kettles running non-stop. The densest racks, such as the GB200 NVL72, are liquid-cooled.

Gaming GPU (RTX 5090, 2025)
AI server (DGX B200)

GPUs
1
8

Memory
32 GB GDDR7
1,440 GB HBM3e

Power
up to 575 W
about 14,300 W

At national scale: a December 2024 report by Lawrence Berkeley National Laboratory for the US Department of Energy found that US data centres used 176 terawatt-hours of electricity in 2023, 4.4% of the country's total, three times their use in 2014. It projected 6.7% to 12% by 2028. Globally, the International Energy Agency estimated in 2025 that data centres used about 415 TWh in 2024, around 1.5% of the world's electricity, and could reach about 945 TWh by 2030. (These are estimates with wide ranges; the honest answer to "how much will AI use?" is that it depends on choices not yet made.)
Part 9 — The state of affairs in 2026
Company figures below come from their own results releases; dates are given because these numbers move fast.
NVIDIA. In its 2026 financial year (to 25 January 2026), NVIDIA's revenue was $215.9 billion, of which $193.7 billion came from data centres. In the three months to 26 July 2026 it reported revenue of $96.2 billion, more than double a year earlier, with data centres $89.0 billion, about 92% of the total. The company that sold graphics cards to gamers is now overwhelmingly an AI-infrastructure company. It said its next platform, Vera Rubin, was "now in full production." Estimates of its share of the market for AI accelerators commonly exceed 80%. Its chief executive's summary in August 2026: "AI has reached its inflection point. It's doing useful work. Its tokens are productive and profitable." Its next results are due on 17 November 2026.
AMD. NVIDIA's long-time graphics rival, which bought the graphics-chip maker ATI in 2006, reported revenue of $11.5 billion for the quarter to 27 June 2026, with its data-centre business more than doubling to $6.7 billion. It launched its Instinct MI400 series and Helios rack systems, which it says partners including Anthropic, Meta, Microsoft, OpenAI and Oracle are deploying. Its gaming revenue, by contrast, fell 31% to $779 million. AMD chips also power two of the world's three fastest supercomputers, El Capitan and Frontier.
Custom chips. The biggest buyers increasingly design their own. Google has used its TPUs (tensor processing units) since 2015; its seventh generation, Ironwood, was unveiled in April 2025, and on 27 October 2026 Google described an eighth generation rated at up to 12.6 petaflops in 4-bit arithmetic per chip and announced that Oracle's cloud would offer TPUs from 2027, the first time another major cloud company has done so. Amazon offers its own Trainium chips; the current Trainium3 has 144 GB of HBM per chip. Others take radically different approaches: Cerebras builds a single chip the size of a whole wafer, 46,225 mm² with 4 trillion transistors. And NVIDIA itself agreed in December 2025, according to CNBC, to pay about $20 billion for assets and a technology licence from the inference-chip start-up Groq.

Who
Main AI chip line (2026)
Sold how

NVIDIA
Blackwell, Vera Rubin
To everyone; with CUDA software

AMD
Instinct MI400 series
To clouds and AI labs; ROCm software

Google
TPU (Ironwood, eighth generation)
Rented through Google Cloud; Oracle from 2027

Amazon
Trainium3
Rented through AWS

Cerebras
Wafer-scale engine
Systems and cloud service

Supply. NVIDIA's Blackwell is made by TSMC in Taiwan, and AI chips in general depend on the handful of leading-edge factories described in the previous piece and on HBM from three suppliers. The demand has rippled through the whole chip industry: the industry statisticians WSTS forecast in spring 2026 that memory sales would more than triple this year.
Geopolitics. The US restricted exports of NVIDIA's top data-centre chips to China in October 2022, and NVIDIA responded with cut-down versions for the Chinese market. The rules have changed repeatedly since. NVIDIA recorded $4.5 billion of charges related to its China-specific H20 chip in the quarter to April 2025. In August 2025, it was reported that NVIDIA and AMD had agreed to pay the US government 15% of revenue from certain China sales in exchange for licences; in September 2025, China's internet regulator reportedly told its tech companies not to buy one of NVIDIA's China-market chips. After a presidential announcement in December 2025, the US Commerce Department said on 13 January 2026 that it would consider licences for NVIDIA's H200 and AMD's MI325X case by case, with conditions. Even so, NVIDIA's outlooks in February and August 2026 assumed no data-centre compute revenue from China at all. In June 2026 the world's fastest supercomputer, China's LineShine, reached the top using only home-grown CPUs and no GPUs.
The two cases. The boosters' case: the demand is real and paid for, in record revenues; the systems are doing useful work in code, science and services; and each generation does more per watt (Google says Ironwood roughly doubled performance per watt over its predecessor). The critics' case: spending on AI data centres is running far ahead of proven profits from AI services, chips lose value quickly as new generations arrive, and the electricity, water and grid demands land on communities that didn't choose them. Some see a bubble; others see the early years of an infrastructure build-out like railways or electricity. That argument is unresolved, and the evidence that would settle it, years of returns on today's investment, doesn't exist yet.
Part 10 — Open questions and trade-offs
Can anyone loosen NVIDIA's grip? Its hardware lead is real but not permanent; its software lead, two decades of CUDA, is harder to erode. Custom chips from the cloud companies, AMD's open-source ROCm and AI frameworks that can target many kinds of chip are all bets that the lock-in will weaken. If it does, prices could fall; if it doesn't, one company remains the bottleneck for much of the world's AI.
Is the GPU even the right shape? GPUs became the AI engine by accident: they were the parallel hardware that happened to exist. Purpose-built chips (TPUs, Trainium, wafer-scale engines) strip out the graphics heritage. Yet GPUs keep absorbing the specialist features, from tensor cores to 4-bit arithmetic. Whether general flexibility or narrow efficiency wins may depend on how quickly AI methods keep changing.
Memory, not maths. For many AI tasks, especially answering questions with a trained model, the limit is how fast numbers can be moved, not how fast they can be multiplied. HBM is expensive and made by three companies. Advances in memory and packaging may matter as much as advances in the processor itself.
Energy. More efficient chips don't automatically mean less electricity. When a resource gets cheaper to use, people often use far more of it (economists call this the Jevons paradox). Whether AI's power demand flattens or keeps climbing depends on efficiency, on prices and on what society decides the computing is worth.
Chips and nations. Export controls, subsidies and supply chains now shape where AI can be built. Whether that slows rivals or accelerates their independence is debated, as the previous piece describes.
And the gamers? The people who paid for twenty years of GPU development are now a small part of the market. AMD's gaming sales fell while its data-centre sales doubled; NVIDIA no longer reports gaming as a separate line in its headline results. The chip designed to paint a game's pixels now spends most of its life multiplying matrices for AI, and it is still, at heart, the same army of simple workers.

The takeaway
Rendering a frame means doing small, independent calculations, matrix maths on millions of vertices and a colour for each of millions of pixels, sixty times a second. GPUs grew up to do exactly that with thousands of simple workers instead of a few smart ones. SGI pioneered 3D hardware, 3dfx brought it to PCs in 1996, and NVIDIA's GeForce 256 put the whole pipeline on one consumer chip in 1999 (its "first GPU" claim rests on its own definition). Programmable shaders (2001) and CUDA (2006–07) let GPUs run any parallel maths. In 2012 AlexNet, trained on two gaming cards, showed that neural networks plus GPUs beat everything else, because neural networks are mostly matrix multiplication, which is what GPUs do best. Since then tensor cores, low-precision numbers and stacked HBM memory have reshaped GPUs around AI. In 2026 data centres provide over 90% of NVIDIA's revenue, cloud companies build their own chips, and power use and export controls are live political questions.
Try this: the next time you play a game or watch an animated film, pause on one frame and estimate the work. Count the pixels (an iPhone 16 screen, at 2,556 × 1,179, has about 3 million), multiply by the frames per second, and remember that every one of those colours was calculated, not stored. Then remember that the same chip, doing the same kind of sum, is what trains the AI models you use.

Sources & Further Reading
Sources
- Company results: NVIDIA, fourth-quarter and fiscal 2026 results (25 February 2026) and second-quarter fiscal 2027 results (26 August 2026), including the H20 charge and China outlook; AMD, second-quarter 2026 results (4 August 2026); TSMC as cited in the previous piece.
- NVIDIA product pages (accessed October 2026): RTX 5090, Blackwell architecture, DGX B200, GB200 NVL72, Vera Rubin; NVIDIA Volta Architecture white paper (V100 transistors, tensor cores, 125 TFLOPS, 900 GB/s HBM2); NVIDIA corporate timeline (founding, "invents the GPU").
- Google: "Ironwood: The first Google TPU for the age of inference" (April 2025) and Cloud TPU documentation (TPU7x specifications); "Bringing Google TPUs to Oracle Cloud Infrastructure" (27 October 2026). Amazon Web Services, Trainium page. Cerebras, chip page.
- Krizhevsky, Sutskever and Hinton, "ImageNet Classification with Deep Convolutional Neural Networks", NeurIPS 2012 (architecture, two GTX 580 GPUs, five to six days, 15.3% vs 26.2%, "waiting for faster GPUs" quotation). Computer History Museum, "CHM Releases AlexNet Source Code" (20 March 2025): the bedroom, ImageNet history, CUDA's 2007 release, Hinton's quotation.
- Computer History Museum, Revolution exhibit, "Silicon Graphics" (founding, Geometry Engine, 1984 terminal).
- US Department of Energy, "DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers" (20 December 2024), on the Lawrence Berkeley National Laboratory report. IEA Energy and AI (10 April 2025), figures as reported by S&P Global Commodity Insights, "Global data center power demand to double by 2030 on AI surge: IEA" (10 April 2025) (the IEA site could not be reached to check directly).
- US Bureau of Industry and Security, press release of 13 January 2026 (H200 and MI325X licence review). TOP500, June 2026 list. WSTS spring 2026 forecast.
- Wikipedia, used for leads and secondary details: "Graphics processing unit", "3dfx Interactive", "Silicon Graphics", "GeForce 256", "GeForce 3 series", "GeForce 8 series", "CUDA", "General-purpose computing on graphics processing units", "Single instruction, multiple threads", "High Bandwidth Memory", "Tensor Processing Unit", "Volta (microarchitecture)", "Geoffrey Hinton", "Turing Award", and "Nvidia" (market-share estimate, the August 2025 revenue-share arrangement, China's September 2025 instruction, and the Groq deal as reported by CNBC).
- Computed in Python: 1,920 × 1,080 × 60 = 124.4 million pixels per second; the race model (2,073,600 ÷ 8 × 1 = 259,200 versus ⌈2,073,600 ÷ 2,048⌉ × 4 = 4,052); 4,096² = 16.8 million and 4,096³ = 68.7 billion multiply-adds; 137.4 GFLOP ÷ 125 TFLOPS = 1.1 ms and ÷ 15.7 TFLOPS = 8.8 ms; 14.3 kW ÷ 8 = 1.8 kW; 89.0 ÷ 96.2 = 92.5%; 9,216 × 4,614 TFLOPS = 42.5 exaflops.
Go deeper
- 📘 The Thinking Machine — Stephen Witt (2025). A history of Jensen Huang and NVIDIA from 3D graphics to AI, including the CUDA bet and AlexNet; FT Business Book of the Year 2025.
- 📘 Chip War — Chris Miller (Scribner, 2022). The manufacturing and geopolitical background: why nearly every AI chip passes through Taiwan.
- 🌐 CHM: "CHM Releases AlexNet Source Code" — the story of the 2012 breakthrough, with the original code.
- 🌐 The AlexNet paper (NeurIPS 2012) — nine pages, surprisingly readable; section 3.2 explains the two-GPU trick.
- 🌐 TOP500 — the twice-yearly ranking of the world's fastest supercomputers, a good way to watch GPUs take over scientific computing.
- 🌐 Companion pieces in this library: Vectors and Matrices (Mathematics) for the maths GPUs do, Processors: How a CPU Thinks (Technology) for the other kind of chip, Artificial Intelligence: Where It Came From and Where It Stands and How Large Language Models Actually Work (Technology) for what the chips are used for, Alan Turing: The Machine That Defined Computing (History-Makers), and What It's Like to Be a Data Scientist (Careers) for the people who use these machines every day.