I started to write a post about Open AI’s drop of more than 700 AI-generated math papers, but I got distracted by something else. In the meantime, they’ve started to unravel.
Some of the papers have been withdrawn on account of a sign error in a crucial lemma. Others have been revised. And some are just nothing burgers. One paper claims to increase the speed of Fast Fourier Transforms (FFTs), for example, but Ben Recht gives a nice run down of why it’s just playing fast and loose with notation.
Some of the papers include proofs in Lean, the framework mathematicians use to confirm the validity of proofs. Unless OpenAI has discovered an exploit in Lean, those papers offer valid proofs of something. But the example of FFTs shows that the results need interpreting to figure out if there’s anything novel or significant being shown.
Yahoo!News has the headline OpenAI’s math breakthrough could hit biotech and pharma next, echoing a sentiment from the socials that AI will overturn empirical science next.1 That’s a bizarre thought. LLMs really just have models of discourse, which is what makes it plausible that they can do math. Although it’s rigorous, math doesn’t depend on what’s actually going on in the world.
It is possible, in a derivative sense, to have a world model embedded in a model of discourse. However, that world model will all be descriptive metaphysics.2 It will represent how language-users take things to be, rather than necessarily how they are. Biotech and pharma, on the contrary, need results that work in actual organisms. An LLM can digest published results, but it can’t go into the lab to collect new data.
AI is not ready to replace scholars, but the OpenAI paper drop is just the latest installment of a years-long pattern of corporations flooding the field with slop and bullshit to hide that fact and to overwhelm old institutions.
- An even crazier headline in the same vein: Palantir CTO says some AI researchers are taking up smoking because they expect AI to cure cancer
- To pick up a term from PF Strawson.