A Timeline of AI: From Dartmouth to Now

Most AI timelines are lists of demonstrations. This one lists documents, because the interesting turns in this history were caused by papers, proposals and — twice — commissioned funding reviews, and because the documents are usually more measured than the stories told about them.

Founding: 1943–1958

• 1943 — McCulloch & Pitts, “A Logical Calculus of the Ideas Immanent in Nervous Activity”. Neurons as logical units. The idea that a network of simple threshold elements can compute, which is where the whole neural line begins.
• 1950 — Turing, “Computing Machinery and Intelligence”. Replaces “can machines think” with an operational test, then spends most of the paper on objections. Also contains the suggestion of building a child machine and educating it, which is the ancestor of learning-based approaches.
• 1955 — the Dartmouth proposal. McCarthy, Minsky, Rochester and Shannon’s funding application for a summer workshop held in 1956. It coins “artificial intelligence” and states the founding conjecture: that every aspect of learning or intelligence can in principle be described precisely enough for a machine to simulate it. The proposal is short and worth reading for how confidently it estimates the timescale.
• 1958 — Rosenblatt, the perceptron. A learning rule for a linear classifier, with a convergence guarantee, implemented in hardware. The first system that learned from data in a way that attracted general attention — and press coverage that promised far more than the model could deliver.

The two winters: 1966–1993

The first, 1966–1980

Two commissioned reviews, not two technical failures, ended the first funding boom. Both are readable and both are more specific than their reputations.

• 1966 — the ALPAC report. A US National Academy of Sciences committee assessed machine translation and concluded the results did not justify continued funding at the existing level. Machine-translation support was cut for years afterwards.
• 1966 — Weizenbaum’s ELIZA. A pattern-matching script that people treated as understanding them. Weizenbaum’s own reaction — that the readiness to attribute comprehension was the alarming finding — is the durable part, and it is directly relevant to reading model self-reports today.
• 1969 — Minsky & Papert, Perceptrons. Proved limits of single-layer networks, notably on linearly inseparable functions. Often blamed for killing neural network research; the book itself discusses multi-layer networks and is more careful than the legend.
• 1973 — the Lighthill report. Commissioned by the UK Science Research Council. Its central charge was the combinatorial explosion: methods that worked on toy problems did not scale to realistic ones. UK AI funding was substantially withdrawn.

Expert systems, and the second

The 1980s recovery was commercial rather than scientific. Rule-based expert systems encoding specialist knowledge found real industrial deployments — configuration, diagnosis, scheduling — and a hardware industry grew around specialised LISP machines. Japan’s Fifth Generation Computer Systems project, launched in 1982, triggered matching national programmes elsewhere.

The collapse in the late 1980s and early 1990s had causes that are worth separating. Expert systems were brittle at the edges of their rule sets and expensive to maintain as knowledge changed — the knowledge acquisition bottleneck, which was a research problem nobody solved. And general-purpose workstations overtook specialised hardware, which destroyed the LISP machine market for reasons having nothing to do with AI at all.

Statistical learning: 1986–2011

• 1986 — Rumelhart, Hinton & Williams in Nature. Backpropagation applied to multi-layer networks, in the paper that made it widely known. The parallel distributed processing volumes published the same year set out the broader programme.
• 1997 — Hochreiter & Schmidhuber, LSTM. An architecture that kept gradients from vanishing over long sequences. The workhorse for sequence modelling until attention displaced it.
• 1997 — Deep Blue defeats Kasparov. A landmark in public perception and a search-and-evaluation system rather than a learning one. Its main lesson is about the moving definition of AI: once solved, chess stopped counting.
• 1998 — LeCun et al., LeNet-5. Convolutional networks working in production on handwritten digits, years before the hardware existed to scale the idea.
• 2006 — Hinton et al. on deep belief networks. A layer-wise training scheme that made deeper networks trainable and restarted serious interest in depth.
• 2009 — Deng et al., ImageNet. A dataset, not a model, and arguably the more important contribution. It made the decisive comparison possible.

Deep learning: 2012–2017

• 2012 — Krizhevsky, Sutskever & Hinton, AlexNet. A convolutional network trained on GPUs won the ImageNet competition by a margin large enough to end the argument. The components were mostly not new; the combination of data, GPUs and depth was.
• 2014–2015 — generative adversarial networks; residual networks. A new generative approach, and a trick that made very deep networks trainable by adding identity shortcuts.
• 2016 — AlphaGo defeats Lee Sedol. Deep networks plus tree search on a game long treated as out of reach. Significant for combining learned evaluation with search, a pairing that has returned in current work on inference-time compute.
• 2017 — Vaswani et al., “Attention Is All You Need”. Removed recurrence, made sequence models parallelisable across positions, and thereby made training at current scales possible. Everything after this is downstream of it.

Scale: 2018 onward

• 2018 — BERT and GPT-1. Two directions from one architecture: bidirectional encoders fine-tuned per task, and decoder-only generative pretraining. The second turned out to scale further, which was not obvious at the time.
• 2020 — GPT-3, and Kaplan et al. on scaling laws. In-context learning as a phenomenon, and the empirical relationships that made capability a function of budget. This is the point at which model development became a resource-allocation problem.
• 2022 — Hoffmann et al. (Chinchilla) and Ouyang et al. (InstructGPT). The compute-optimal correction to the earlier scaling recommendations, and instruction tuning with human feedback — the technique that made models usable by people who were not prompt-engineering specialists.
• 2022 — the public chat release. No new capability; an interface. Its significance is in adoption and in the policy attention that followed, which is a fact about deployment rather than about research.
• 2023 onward — multimodality, long context, and reasoning models trained to spend inference-time compute on intermediate steps. Covered in what a reasoning model is. Dating this period precisely is not yet possible, and any timeline that does so with confidence is guessing.

The pattern

Three things recur, and they are the reason to read a timeline at all.

Both winters followed assessments, not failures. ALPAC and Lighthill were commissioned reviews that compared delivered results against promises made to funders. The technical work did not stop being interesting; the gap between claim and demonstration became officially documented. That is a mechanism worth remembering, because it is available again.

Ideas arrive before the resources that make them work. Backpropagation, convolutional networks and attention-like mechanisms all predate their successful application by years or decades. This cuts against confident claims that a current limitation is fundamental, and equally against confident claims that a promising idea is close.

The definition moves. Chess, then Go, then translation, then fluent conversation were each treated as requiring real intelligence until achieved, at which point each was reclassified as mere computation. That is worth holding on to when reading any argument about what a system “really” understands — the target has moved every previous time, and the reasons it moved were not always bad ones.