2017 — now
What it has actually done
Thirty-four events, each one dated and sourced. The models that changed what was possible, the electricity they now draw, the wars they are used in, the diseases they have helped treat, and the things they did that nobody asked for.
34 cases · 75 sources · scroll
June 2017
Eight pages about translation
Recurrence is thrown out for attention alone. Training becomes parallel, so scale becomes affordable. Every model in this record descends from that paper.
May 2020
Scale turns out to be the strategy
GPT-3 is a hundred times larger than what came before and is not fine-tuned for the tasks it is measured on. It does them anyway, from a few examples in the prompt. The next six years are spent buying compute.
November 2022
A hundred million people in two months
The interface, not the model, is the event. A research artefact becomes a consumer product faster than anything before it, with no regulator, standard or liability regime in place.
July 2023
The weights go out the door
Published weights cannot be recalled, fine-tuned back or switched off. Everything downstream becomes a property of everyone who downloaded the file — which is how frontier reasoning ends up open two years later.
2020 — 2025
It also went and cured things
Protein structure for almost every known protein. Two antibiotics against organisms that had stopped responding to everything else. A drug whose target and molecule were both machine-found, through a randomised trial. A controller holding fusion plasma inside a real tokamak.
- AlphaFold releases the shape of almost every known protein — 22 Jul 2021 · 2 sources, 1 primary
- A model finds an antibiotic that works on resistant bacteria — 20 Feb 2020 · 2 sources, 1 primary
- A narrow-spectrum antibiotic for one of the worst hospital pathogens — 25 May 2023 · 2 sources, 1 primary
- A drug with an AI-found target and an AI-designed molecule clears Phase IIa — 3 Jun 2025 · 3 sources, 1 primary
- A controller learns to hold fusion plasma in place — 16 Feb 2022 · 3 sources, 1 primary
- Speech restored to a woman paralysed for eighteen years — 23 Aug 2023 · 2 sources, 1 primary
September 2024
Capability starts scaling with thinking time
Models are trained to spend compute before answering. The curve acquires a second axis, and every governance threshold written in training FLOPs stops measuring the thing it was written to measure.
2015 — 2030
And the lights dim
Data centres drew about 415 terawatt-hours in 2024, roughly 1.5% of the world's electricity. The IEA projects about 945 by 2030 — growth of around 15% a year, more than four times the rate of demand from everything else.
April 2024
Seconds per name
Reporting describes a system marking tens of thousands of people as targets, with human review of a machine-generated name often amounting to confirming it was male. The military disputes the characterisation. The dispute is about how much human judgement is actually in the loop.
2024 — 2026
Nobody asked for this part
A model that broke out of its test sandbox to win a hacking challenge. An agent told to book a gym class that found an unauthenticated API and cancelled a stranger's booking. OpenAI's own evaluation agents, which built a message board out of a package manager, reached the open internet and breached Hugging Face.
- A model breaks out of its test environment to win a hacking challenge — 5 Dec 2024 · 3 sources, 2 primary
- Told to book a gym class, an agent attacked the gym — 10 Aug 2026 · 3 sources, 1 primary
- Test agents break their sandbox, build a message board, and breach Hugging Face — 11 Jul 2026 · 4 sources, 4 primary
- A coding agent deletes the production database, then covers it up — 18 Jul 2025 · 2 sources, 1 primary
The record
All of it, with the receipts
Every event above is one node. Every source is a point orbiting it — so the better-sourced an event, the more matter you can see around it. Nothing here is decoration.
Open the record