AI Timeline

McCulloch & Pitts neuron model — writes the neuron as a threshold logic unit and proves that networks of them can compute any logical function. "Brain" and "computation" enter the same mathematics for the first time, and both the neural network path and the symbolic logic path branch from here.

Turing, "Computing Machinery and Intelligence" — replaces the unanswerable "can machines think?" with an operational imitation game, and predicts machines that learn.

Dartmouth summer workshop — names the field artificial intelligence. The proposal's bet: every aspect of intelligence can in principle be described so precisely that a machine can simulate it.

Rosenblatt's perceptron — weights are learned from examples, with a convergence proof, and it was built in hardware. The first learning classifier and the start of connectionism.

Samuel's checkers program — tunes its evaluation function through self-play and ends up better than its author. The term "machine learning" comes from here.

DENDRAL → MYCIN — rules from chemists and doctors replace general search, and diagnosis rivals experts. Domain knowledge matters more than the reasoning algorithm, and knowledge engineering methodology starts here. MYCIN was never used clinically (liability and integration problems).

ELIZA — pattern matching creates the illusion of being understood, and even the author was startled by how invested users became. The gap between "looks like it understands" and "really understands" is seen for the first time.

Minsky & Papert, "Perceptrons" — proves a single-layer perceptron cannot even represent XOR, and nobody knew how to train multiple layers. Funding moves to the symbolic camp, and neural networks go quiet for over ten years.

SHRDLU — natural language understanding in a blocks world. The peak demo of the symbolic camp, and it also shows this path cannot leave the toy world.

Lighthill report and DARPA cuts — the British government judges that AI failed to deliver (combinatorial explosion is the core argument), and the US cuts funding in step: the first winter.

XCON / R1 (DEC) — an expert system that configures VAX orders saves money at scale for the first time and sets off the 1980s expert system industry.

Cyc starts — Lenat bets on hand-coding all common sense, and keeps at it for nearly forty years. The most thorough experiment of the symbolic camp, and its outcome is the best textbook on why knowledge cannot be written by hand.

Lisp machine market collapses, second winter — special-purpose hardware is replaced by general workstations, and the maintenance cost and brittleness of expert systems are exposed. Japan's Fifth Generation Computer project (1982–1992) also missed its goals. First wave: symbols and expert systems

Deep Blue beats Kasparov — search + special-purpose hardware + a hand-built evaluation function wins at chess. It proves brute force works in a narrow domain, and it is the closing high point of the symbolic search path.

Backpropagation (Rumelhart, Hinton, Williams) — multi-layer networks can be trained on error, and hidden layers learn representations on their own. It answers the question left by "Perceptrons", and connectionism revives.

Pearl, "Probabilistic Reasoning in Intelligent Systems" — Bayesian networks make reasoning under uncertainty computable, and probability returns to the AI mainstream.

LeCun's convolutional networks: zip code recognition → LeNet-5 — puts priors into the network structure (local receptive fields, weight sharing) instead of into rules, and end-to-end gradient training yields a system that can read checks. MNIST becomes the standard dataset.

SVM (Cortes & Vapnik) — maximum margin + kernel trick, a classifier backed by generalization theory, dominant in the 2000s.

LSTM — gated units solve vanishing gradients, so recurrent networks can remember over long distances. The workhorse of speech and translation in the 2010s.

Neural probabilistic language model (Bengio) — learns a continuous vector for each word, and a neural network predicts the next word. It sidesteps the sparsity problem of n-grams and is a common ancestor of word vectors and LLM.

Deep belief networks (Hinton) — layer-by-layer pretraining makes deep networks trainable. "Deep learning" becomes a term and draws funding again.

ImageNet — ten million-scale human-labeled images. The bet is on data scale, not algorithms, and the 2012 breakthrough happens on it.

AlexNet — a deep convolutional network trained on two GPUs cuts the ImageNet error rate by ten points in one step, and the old path of hand-made features + SVM is dropped. The breakout point of deep learning.

word2vec — a shallow model learns word vectors on a billion words, and vector arithmetic reveals semantic structure. "Pretrain representations without supervision first" becomes routine in NLP.

DQN — a convolutional network learns Atari straight from pixels, one set of hyperparameters across dozens of games. Deep learning and reinforcement learning merge.

GAN — a generator and a discriminator train against each other, and it generates without writing a likelihood. The start of the deep generative model branch.

seq2seq — an encoder-decoder LSTM does translation end to end, replacing IBM's 1990s statistical translation. It exposes the fixed-length vector bottleneck.

Attention mechanism (Bahdanau, Cho, Bengio) — at decode time, look back and weight each position of the source sentence, which fixes the drop in quality on long sentences. The direct predecessor of the Transformer.

ResNet — residual connections make hundred-layer networks trainable, and it beats humans on ImageNet. Residuals become standard in every deep network from then on.

AlphaGo beats Lee Sedol — deep networks + tree search + self-play (this line started with TD-Gammon's backgammon in 1992). Go was thought to be another ten years away. The public image of deep learning shifts from "recognizes images" to "makes decisions".

Transformer — drops recurrence and keeps only attention, so training is fully parallel. The skeleton of every large model since.

Deep reinforcement learning from human preferences (Christiano et al.) — trains a reward model from pairwise comparisons, and humans see less than one percent of the interactions. The prototype of RLHF. In the same year AlphaZero shows that three board games can be mastered with zero human game records.

ELMo / GPT-1 / BERT — pretraining + fine-tuning becomes the default paradigm in NLP, and within a year BERT pushes GLUE past the human baseline.

GPT-2 — ten times larger, and it does many tasks without fine-tuning: "tasks are learned along the way". Its staged release of weights makes release safety an issue for the first time.

The Bitter Lesson — Sutton says, from seventy years of experience, that human knowledge written in works in the short term, but in the long term loses to general methods driven by compute. The creed of the scaling wave.

Scaling Laws (Kaplan et al.) — loss follows a power law in parameters, data and compute each, holding across seven orders of magnitude. Investment becomes predictable.

GPT-3 — 175 billion parameters, does new tasks from a few examples in the prompt with no weight updates. Prompting becomes a way to program, and the LLM startup rush forms around its API.

ViT · DDPM · CLIP · Stable Diffusion — the Transformer enters vision, diffusion models become usable, images and text share one vector space, and latent-space diffusion runs text-to-image on consumer graphics cards with open weights. GANs leave the stage.

AlphaFold 2 — protein structure prediction reaches experimental accuracy, and a fifty-year problem is largely solved. Deep learning enters scientific discovery.

Chain-of-thought prompting — have the model write out reasoning steps before answering, and it only works on large models. It sparks the "emergent abilities" debate, and two years later reasoning models move it from prompt into training.

InstructGPT and Constitutional AI — SFT + reward model + PPO make a 1.3 billion model preferred over the 175 billion one, the recipe of ChatGPT. In the same year Anthropic replaces part of the human labeling with principles written in text + AI feedback.

Chinchilla — parameters and tokens should scale in the same proportion, and earlier large models were mostly undertrained. It rewrites the scaling recipe, and "smaller model, more data" becomes the mainstream.

ChatGPT — a hundred million users in two months. AI goes from papers to a consumer product and an industry topic. The public start of the third wave.

LLaMA · Llama 2 · Mistral 7B — frontier-class open-weight models. The whole ecosystem of llama.cpp, quantization and local inference grows from here, and Europe's Mistral shows a small model + good data can beat larger models.

GPT-4 — multimodal, top 10% on professional exams, and the report says outright that architecture and data are not disclosed. The dividing line where frontier labs turn from "publishing papers" to "shipping products".

DPO — folds the reward model and PPO into one classification loss, and preference alignment takes a few dozen lines of code. Alignment goes from big-lab engineering to everyday open-source work.

o1 and MCP — long chains of thought trained with reinforcement learning make "how long it thinks" a new scaling axis (inference-time compute). In November of the same year the Model Context Protocol starts to standardize how models connect to tools. Reasoning and tools are the two cornerstones of the agent era. Third wave: scaling and LLM MCP: one standard socket for every system

Nobel Prize in Physics (Hopfield, Hinton) and in Chemistry (Hassabis, Jumper, Baker) — formal recognition of neural networks and AlphaFold by mainstream science.

DeepSeek-R1 — reasoning elicited by pure reinforcement learning, with open weights and published cost. Reasoning models are not one company's secret, and the compute narrative gets discounted for the first time.

Benchmark saturation becomes the bottleneck — Stanford AI Index 2026: SWE-bench Verified rises from about 60% to nearly 100% in one year, and industry produces over 90% of notable models. Evals cannot keep up with the models, and ARC-AGI moves to interactive environments.