ChatGPT

ChatGPT was released on 30 November 2022 as a free research preview, a text box and a model, and its makers did not expect much. The model was GPT-3.5, an updated GPT-3 fine-tuned with human feedback on conversations the way InstructGPT had been on instructions. It remembered what had been said earlier in the exchange, refused some requests, admitted some mistakes, and answered nearly anything in fluent, organised prose. It reached a million users in five days and an estimated hundred million in two months, faster than any consumer application in history.

Nothing in it was new to researchers. GPT-3 had been available for two years; the alignment method was ten months old. What was new was the door. A conversation is the interface Turing had proposed in 1950 and Weizenbaum had accidentally tested in 1966, and a model that could hold one turned a research API into a thing every person could try, in their own words, about their own work. Students, lawyers, programmers and doctors did, and the ELIZA effect operated on a billion people.

The consequences fill the rest of this timeline: the race between laboratories, the investment, the regulation, the disputes over training data and jobs, and the agents. The organism on this site is at its densest here.