How we got here

Alan Turing asks what a machine can compute at all — and finds limits.

McCulloch and Pitts model a brain cell as simple arithmetic.

Shannon measures information in bits. This guide starts here.

Turing proposes the test: if you cannot tell the difference, does it matter?

the term "artificial intelligence" is coined

Rosenblatt builds a machine that learns to tell two classes apart by changing its weights.

first expert system, for chemistry

Weizenbaum's chatbot reflects your words back — and people confide in it anyway.

saves DEC tens of millions a year

Japan bets big public money on parallel logic machines.

Rumelhart, Hinton and Williams show networks can learn their own hidden features.

LeCun's convolutional net reads handwritten digits.

a network learns backgammon by playing itself, given no human strategy

Vapnik's maximum-margin method makes small-data learning rigorous.

A gated loop remembers over long stretches, easing the vanishing gradient.

IBM's machine beats world chess champion Garry Kasparov.

Hinton's greedy layer-wise training restarts the neural-network field.

NVIDIA opens general-purpose programming on graphics chips — the accident that trains everything since.

Fei-Fei Li's labelled image set gives vision a shared yardstick.

the thaw, and the boom that never stopped

Mikolov turns words into vectors whose arithmetic carries meaning.

Bahdanau gives translation a way to look back at the input — the seed of the transformer.

Goodfellow pits a generator against a discriminator until the fake looks real.

Skip connections let networks get very deep without falling apart.

One network learns dozens of Atari games from the pixels alone.

A search-plus-network agent beats Go champion Lee Sedol.

"Attention Is All You Need" replaces recurrence with attention.

Pre-train on everything, then fine-tune — the default recipe for years.

A large language model writes fluent text; the release is staged.

Loss falls as a smooth power law in compute, data and parameters — so you can plan ahead.

175 billion parameters, and prompting alone works.

Protein folding is solved well enough to matter for real biology.

Images and text land in one shared space, so you can search one with the other.

Learn to undo noise, and you can make an image from scratch.

Most giant models were badly under-trained on data. Fix that, and smaller wins.

Image generation runs on a home GPU and spreads everywhere.

Preference tuning turns a text model into a product. The public arrives.

Frontier models pass many human exams, and the race goes public.

Llama 2 and Mistral make good models runnable and modifiable by anyone.

An open protocol for plugging tools into a model.

Hopfield and Hinton share the physics prize; Hassabis and Jumper the chemistry prize.

A strong open-weight model trained at a fraction of the assumed cost.

Open reasoning weights land, with a cost shock and a new rivalry in the open.

Qwen, GLM, Kimi and others ship frontier-adjacent models anyone can download.

Tool use, sandboxes and long-running coding agents move from demo to daily work.

The US moves advanced-chip licences for China to case-by-case review, easing the squeeze that shaped the efficiency wave.

A one-million-token context arrives with hybrid attention, replacing the very MLA trick the V2 line was built on.

Scale, reasoning, and the open question of what is left to invent.