Artificial Intelligence: Where It Came From and Where It Stands

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

1. The dream before the machine: automata, Leibniz, Babbage, Lovelace, Turing and Dartmouth
2. Two roads: hand-written rules, two winters, and neural networks (1956–1997)
3. The deep-learning turn and the ChatGPT moment (2012–2022)
4. The state of affairs, late 2026: capabilities, compute, cost and who builds it
5. 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:

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:

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.

1. 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.
2. Compute. Graphics cards (GPUs), built for video games, turned out to be very good at the multiplications neural networks need.
3. 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:

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:

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).