Galactica

» see galactica.org as it lived, on the wayback machine

~ the goodbye ~

Meta put Galactica online on 15 November 2022 and described it as a large language model for science. It had been trained on something like 48 million papers along with textbooks and reference material, and the pitch was that a researcher could ask it to summarize a literature they did not have time to read, or to draft the write-up of work they had already done. The largest version ran to 120 billion parameters. The demo was open to anyone, and the placeholder text in the box told you what it wanted: try a wiki article on Hartree-Fock.

Scientists are a bad audience for a confident machine. Within hours of launch, people were posting its output with the citations checked, and the citations were the problem: plausible author names attached to papers that had never been published, in journals that had never run them. Somebody got it to write an encyclopedia entry on the benefits of eating crushed glass, formatted exactly like a real one. The model had learned the shape of scientific writing perfectly and had no way to tell whether any particular sentence was true.

By 2022 this was my working field rather than something I read about, so the argument that followed was the recognizable one. The people who built it said the demo was being misused by critics looking for failures. The people testing it said a tool aimed at scientists is supposed to survive scientists. Papers with Code pulled the demo on 17 November, three days after it went up, and it never came back.

~ the chronicle ~

~ forty-eight million papers ~

The training corpus was the new idea in it, and that part still holds up. Almost every other model of that era was trained on the open web, with all the forum arguments and marketing copy that implies, while Galactica was fed the formal scientific record and taught to reproduce its conventions. It could annotate molecules and work through equations. In a narrower package, with an interface that showed its sources and refused to answer outside them, some of that was defensible.

What Meta shipped instead was a free-text box on a public website with the word science over it, and no mechanism anywhere in the stack for distinguishing a retrieved fact from a generated one.

~ the three days ~

Yann LeCun defended it loudly and kept defending it for years afterward, on the grounds that a research demo taken down because of online criticism is a loss for everybody. He had a point about the dynamic. What he was arguing against, though, was a fact already established in public, which is that the output was indistinguishable from fabrication to anyone who could not already check it, and the people who most needed it were exactly the people who could not.

~ two weeks later ~

ChatGPT launched on 30 November 2022. It hallucinated citations in the same way, sometimes worse, and it was presented as a conversational assistant rather than an oracle for scientific work, so nobody expected it to be right and everybody forgave it when it was not. It became the fastest-adopted consumer product in the history of the industry. Galactica had been dead for thirteen days by then.

The paper and the weights are still out there, and parts of the idea came back later inside retrieval systems that cite real documents. The demo is gone. galactica.org does not answer.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »