Events

For reinforcement learning, the method behind AlphaGo, RLHF and reasoning models.

OpenAI, SoftBank and Oracle announce $500 billion of US data centres. Compute as national infrastructure.

An open-weights reasoning model at o1 level from a Chinese lab with a fraction of the compute. Nvidia lost $590 billion in market value in one day a week later.

Anthropic open-sources a standard for tools and data. The USB-C of agents.

Half for computational protein design, half for AlphaFold.

For foundational discoveries that enable machine learning with artificial neural networks.

The first reasoning model. Thinks before it answers, and spends more compute on harder problems.

The first broad law on AI. Risk tiers, bans on some uses, duties for general-purpose models. In force from August 2024.

Fired on a Friday, reinstated the next Wednesday after most staff threatened to leave for Microsoft. The board was replaced.

The Future of Life Institute asks labs to pause training beyond GPT-4 for six months. Thirty thousand signatures. No pause.

Multimodal, passes the bar exam, and the first model whose makers withheld all technical detail.

Meta's research-only model appears on 4chan a week after release. Within weeks: llama.cpp, Alpaca, and the open model ecosystem.

A chat interface on GPT-3.5. One hundred million users in two months, the fastest adoption of any consumer product.

Open weights for a text-to-image model that runs on a gaming GPU. Image generation leaves the labs.

Median accuracy at the level of experiment. A fifty-year grand challenge in biology closed.

The first frontier model sold as an API. Demos of code and copy generation go viral. The business model of the field since.

The ACM's highest honour for deep learning, announced March 2019. Recognition for a bet that took thirty years.

OpenAI publishes GPT-2 but withholds the full model as too dangerous. The weights came out in November. The argument about open weights starts here.

Four games to one in Seoul. Go was thought to be a decade away. Move 37 in game two surprised every professional watching.

A non-profit with $1 billion pledged by Musk, Altman, Thiel, Hoffman and others, to build AI that benefits humanity.

About $500 million for a London lab with no product. The deal that set the price of AI talent.

A deep CNN trained on GPUs cuts the top-5 error from 26% to 15%. Every vision group switched to deep learning within a year.

IBM's chess machine wins the rematch 3.5 to 2.5. Search and hardware, not learning, but the first world champion beaten by a computer.

The Lisp machine market collapsed and expert systems proved costly to maintain. AI became a word to avoid in funding proposals.

XCON saved DEC millions a year. Corporations bought Lisp machines and rule engines. Japan's Fifth Generation project.

Funding cuts after the Lighthill report and DARPA's retreat. Neural networks had already been sidelined by Perceptrons (1969).

A UK government review that found AI had not delivered. Funding was cut in Britain and the first AI winter followed.

Weizenbaum's pattern-matching therapist. People confided in it. He spent the rest of his career warning about that.

Eight weeks at Dartmouth College. McCarthy's proposal named the field. The founders of symbolic AI met here.