Galactica lasts three days

Galactica was released by Meta on 15 November 2022 as a scientific assistant: a model trained on 48 million papers, textbooks, reference works and formulae, meant to summarise the literature, solve equations and write Wikipedia-style articles on demand. The public demo produced confident, well-formatted articles containing citations to papers that did not exist, a wiki entry on the history of bears in space, and a passable-looking summary of the benefits of eating crushed glass. Michael Black of the Max Planck Institute called it dangerous; Meta took the demo down after three days, and Yann LeCun complained that the critics had missed the point.

The episode named a problem. Language models produce the most probable continuation, and the most probable citation for a made-up claim is a plausible-looking one; the field had a word for this, hallucination, and Galactica made it a public word. It also showed how quickly a release could fail: the same quality that made the model useful, fluency, made its errors invisible.

Fifteen days later OpenAI released ChatGPT, which hallucinated just as freely and was not withdrawn, because it was framed as a chat rather than a reference. The difference in reception was a lesson about products, not models.