The history of AI in 6 documents: from Dartmouth 1955 to $581.7B

Look up the history of AI and you’ll get a timeline: Turing in 1950, the Dartmouth proposal in 1955, ImageNet in 2012, the transformer in 2017. The dates are right. The shape is wrong, because a list of dates hides the only part that’s useful to you now. Here’s the short answer. AI has run the same cycle at least twice, and it’s running again now. The cycle is this: a general method starts working on a measurable task, money arrives, the claims outrun the method, and then somebody discovers what the method doesn’t generalise to, long after the money has committed. What survives each round is the method, because the method keeps working. What dies is the promise bolted onto it.

That’s a claim about repetition, so the rest of this piece traces it through documents we opened rather than through a story anyone remembers. The 1955 funding proposal. The 1972 review commissioned by the UK’s funding council. A National Academies accounting of what the US spent in the 1980s. The 2012 benchmark table. Four papers between 2017 and 2022. Then the current compute and investment figures, which is where the pattern either holds or breaks.

The founding claim was written down in 1955, and nobody narrowed it since

Alan Turing got there first in print. His paper “Computing Machinery and Intelligence” ran in Mind, volume LIX, issue 236, pages 433 to 460, published 1 October 1950. It opens flatly: “I propose to consider the question, ‘Can machines think?'” But Turing then made a dated, checkable prediction, and that is the part worth holding onto. Within fifty years, he expected machines with storage around 10⁹ to play the imitation game well enough that an average interrogator would have no more than a 70% chance of a correct identification after five minutes of questioning.

The field got its name and its budget five years later. The proposal for the Dartmouth Summer Research Project on Artificial Intelligence is dated 31 August 1955, and it was signed by John McCarthy at Dartmouth, Marvin Minsky at Harvard, Nathaniel Rochester at IBM and Claude Shannon at Bell Telephone Laboratories. They asked the Rockefeller Foundation for $13,500. The plan was ten people, two months, one summer, so the ambition and the budget were never close.

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

McCarthy, Minsky, Rochester and Shannon, Dartmouth proposal, 31 August 1955

Read that conjecture next to the budget and you have the field’s permanent problem in one page. The scope is everything a mind does. The resource was ten people for a summer. Nobody has narrowed the conjecture since, which means every later era gets measured against a target set before anyone knew what it cost. That’s why “AI” keeps sounding like a broken promise even in years when the engineering works. The vocabulary was fixed at the maximum, and it has stayed there. We’ve written separately about what the words mean when a vendor uses them today, and the drift starts here.

The 1972 review asked the question every later reversal turned on

Seventeen years after Dartmouth, the Science Research Council asked Sir James Lighthill, then Lucasian Professor of Applied Mathematics at Cambridge, to survey the field. Its chairman wanted an assessment from someone outside AI, because applications for research support in the field kept arriving and the Council had no lay view of them. “Artificial Intelligence: A General Survey” is dated July 1972. Its first move is the one that did the damage, because it didn’t come from a critic. Instead it came from the practitioners themselves, which meant nobody could dismiss it.

Lighthill reported that “most workers in AI research and in related fields confess to a pronounced feeling of disappointment in what has been achieved in the past twenty-five years”. He then split the field into three categories. Category A was advanced automation: pattern recognition, speech, machine translation, theorem proving, all aimed at replacing human effort in industrial and scientific work. Category C was computer-based study of the central nervous system. Category B was the robot-building work in between, meant to bridge the two.

Category B is where his verdict landed. The attempt to build a real bridge between A and C, he wrote, was not succeeding, and he expected the field to undergo fission as A and C bonded with their own application areas instead. That’s a structural criticism, not a performance one. He wasn’t saying the demos were bad. He was saying the general middle didn’t exist, so the funding case for a unified field didn’t either. His own forecast for category B was that it would “cause progressive loss of prestige, from which a diminution of funding will ultimately follow”.

The lesson that carries forward isn’t that Lighthill was right about everything, because he wasn’t. It’s that the review asked a question the demos couldn’t answer: what generalises? Every subsequent contraction has turned on the same question.

The 1980s bought handcrafted knowledge, and it turned out brittle

The next era rebuilt the case on expert systems: rules, written by hand, encoding what a specialist knows. DARPA describes this as the field’s first wave in its own 2018 account, dating it to the 1960s and calling it “handcrafted knowledge, or rule-based systems capable of narrowly defined tasks”. That wave stayed narrow, but the commercial and government money peaked two decades later.

The National Academies put numbers on it in “Funding a Revolution: Government Support for Computing Research”, published in 1999. The Strategic Computing Program, announced in 1983, was funded at $1 billion over a planned ten-year course. Total federal AI funding rose from $105.6 million in 1984 to $274 million in 1988. Basic research inside that went from $44.1 million to $86 million, and applied research from $12.5 million to $73 million, so by our arithmetic the applied line grew almost six-fold in four years. About 60% of Strategic Computing money went to industry contractors over that period.

Some corporations reported substantial returns. Others found what the same account records, quoting John McCarthy’s 1990 argument: the expert systems were extremely “brittle”. A rule base does exactly what it was told and nothing next to it. Meanwhile the specialised hardware built for the era lost its reason to exist, because the Lisp machine market collapsed under competition from ordinary engineering workstations. So the era ran into two ceilings at once, one in capability and one in hardware economics, and neither was visible in the 1984 budget line.

In every era the limit was a discovery about what does not generalise, and it arrived after the money did.

Rundowns AI

ImageNet settled the argument about handcrafted features

DARPA dates its second wave to the 1990s, describing “machine learning technologies that created statistical pattern recognizers from large amounts of data”. The moment that decided the argument in public came later, and it’s unusually easy to check, because the scores are still posted.

The ILSVRC-2012 classification results table is still online. A team called SuperVision, listed as Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton of the University of Toronto, took first place with an error rate of 0.15315 and second place with its own other submission at 0.16422. The best entry from any other team, ISI, scored 0.26172, using a weighted sum of classifiers over SIFT, LBP, GIST and CSIFT features. By our subtraction that’s a gap of about 10.9 percentage points between a learned representation and the best hand-engineered feature pipeline in the competition, on the same data, in the same year.

They published the system as “ImageNet Classification with Deep Convolutional Neural Networks” at NIPS the same year. It reports 60 million parameters, trained on 1.3 million high-resolution images across 1,000 categories, with GPU-optimised training and a regularisation method to control overfitting. The challenge entry describes the same system as five convolutional layers, some followed by max-pooling, then three globally-connected layers and a 1000-way softmax, trained on raw RGB pixel values.

Still, nothing in that description is a new theory of intelligence. It’s a known architecture, more data than anyone had used, and hardware that made the training tractable. That’s the finding, and it repeats: the general method plus compute beat the encoded expertise, and it beat it by a margin nobody could argue with.

The transformer turned scale into a strategy, then scale got corrected

The architecture running almost everything you use today was published on 12 June 2017. “Attention Is All You Need”, by Vaswani and seven co-authors, dispensed with recurrence and convolutions entirely. It reported 28.4 BLEU on WMT 2014 English-to-German and 41.8 BLEU on English-to-French, the second after training for 3.5 days on eight GPUs, which the paper notes is a small fraction of the training cost of the best prior models.

Three years later the same architecture produced the result that reset expectations. “Language Models are Few-Shot Learners”, submitted 28 May 2020, introduced GPT-3 at 175 billion parameters, which the paper calls 10 times more than any previous non-sparse language model. The claim that mattered wasn’t the size. It was that tasks were specified purely in text, with no gradient updates and no fine-tuning dataset of thousands of examples.

Four months before that, “Scaling Laws for Neural Language Models” had given the strategy its arithmetic. Submitted 23 January 2020, it found loss scaling as a power law with model size, dataset size and compute, across trends spanning more than seven orders of magnitude. Its recommendation was specific: train very large models on relatively modest data, and stop significantly before convergence.

Then it got revised. “Training Compute-Optimal Large Language Models”, submitted 29 March 2022, concluded that the large models of the day were significantly undertrained. Its rule was that for every doubling of model size, training tokens should double too. The demonstration was Chinchilla at 70 billion parameters, trained on the same compute budget as the 280-billion-parameter Gopher but with four times the data, reaching 67.5% on MMLU and beating Gopher by more than seven points.

Sit those two papers side by side and you get the most useful correction in this whole history. A quantitative law, published by a frontier lab, held for roughly two years before another frontier lab showed the industry had been reading it wrong and burning compute on the wrong axis. That’s the current state of certainty about scale, and it’s worth remembering whenever someone quotes a curve at you. We’ve gone through what the scaling laws do and don’t establish in more detail.

What the current numbers actually show

Compute is the input that changed, and it’s the one with the cleanest measurement. Epoch AI’s dataset of models trained after 2010 covers 428 systems, and its analysis, last updated 24 November 2025, puts training compute growth at 4.7 times per year, with a 90% confidence interval of 4.3 to 5.2. That’s a doubling roughly every six months. Epoch is careful about the fit, reporting an R² of 0.60 and noting that different subsets of models show somewhat different trends.

The money moved with it, though the two series measure different things. Stanford HAI’s 2026 AI Index takeaways put global corporate AI investment at $581.7 billion in 2025, up 130% on the prior year, with private investment at $344.7 billion, up 127.5% from 2024. US investment of $285.9 billion ran 23.1 times China’s $12.4 billion. AI data centre power capacity reached 29.6 GW.

Era	The document that marks it	The number in it	What broke
1955 founding	Dartmouth proposal, 31 Aug 1955	$13,500, 10 people, 2 months	Scope set to everything a mind does
1972 review	Lighthill, general survey, July 1972	25 years of “disappointment”	No bridge between categories A and C
1980s expert systems	Funding a Revolution, NAP 1999	$105.6M in 1984 to $274M in 1988	Brittle rule bases, Lisp hardware displaced
2012 deep learning	ILSVRC-2012 results table	0.15315 error vs 0.26172	Hand-engineered features stopped competing
2020 to 2022 scaling	Kaplan 2020, Hoffmann 2022	70B model beats a 280B model	The compute-optimal rule itself
2025 to 2026 buildout	AI Index 2026, Epoch AI	$581.7B invested, 4.7x compute a year	Not yet established

The last row is the honest one, because the thing that would confirm or break the pattern hasn’t happened. Investment growth of 130% in a year tells you about expectations, not returns. Nobody outside the labs can audit the training runs behind Epoch’s estimates, which is why Epoch flags the uncertainty itself. If you want the specific readings that would settle the question, we’ve laid out which numbers would actually answer it.

Four lessons that survived every cycle

Strip out the personalities and the same four findings show up in each era, so the pattern is documentary rather than a matter of hindsight.

Lesson	Where it showed up first	Where it showed up again
General methods that absorb compute beat encoded expertise	ILSVRC-2012: 0.15315 against 0.26172	GPT-3 replacing task-specific fine-tuning sets
Funding tracks demonstrations, and reviews price the gap	Lighthill, July 1972	Strategic Computing, $1B announced 1983
Brittleness shows up in deployment, not in the demo	Expert systems, per McCarthy 1990	The gap between benchmark scores and production
Quantitative rules about scale have a short shelf life	Kaplan et al, January 2020	Hoffmann et al, March 2022, two years later

There’s a fair objection to all of this, and it deserves its strongest form. The cycle framing can be made to fit anything, because any period of growth is followed by a period of slower growth, and calling that a winter is a choice about vocabulary. The 1970s and 1980s contractions also happened in a field with a handful of customers and one or two funders. Today’s spending is spread across thousands of paying enterprises and hundreds of millions of consumers, which is a different failure surface. No committee holds the leverage the Science Research Council held over Lighthill’s field in 1972, so a single review can’t set the direction now.

That objection is largely right, and it changes the mechanism rather than the question. Lighthill’s actual test wasn’t about budgets. He asked what generalises beyond the demo, and every contraction since has turned on the same thing. The current answer is being written in deployment data rather than in a survey, which makes it slower to read and harder to fake. It also makes benchmark scores a weaker signal than they look, a problem we’ve taken apart in why AI benchmarks keep lying to you.

So here’s what would change the reading. If compute growth stays near 4.7 times a year while measured production deployment stays flat, that’s the brittleness pattern repeating with better graphics. If the two move together, this era is doing something the earlier ones didn’t, and the cycle framing stops earning its keep. Turing set a five-minute test and a date. Seventy-one years after Dartmouth, the useful habit he left behind isn’t optimism about machines. It’s writing down in advance what would count as failing.