OpenAI posted 722 AI-written math papers. Now someone has to check them.
722 manuscripts from a model the public can't use; fewer than a quarter have formalized main results and none are peer-reviewed yet. A Dada-collage explainer.
722 manuscripts from a model the public can't use; fewer than a quarter have formalized main results and none are peer-reviewed yet. A Dada-collage explainer.
· Captions, from Gadget Review, Unite.AI, XenoSpectrum
Collage of public-domain archive photos via Wikimedia Commons (see the assets branch credits), xerox and halftone treatment by AI Broadsheet; ink cats, eyes and banana drawn by AI Broadsheet. Music: “DL - 120 BPM TV waiting loop” by trader_one (CC0, Freesound); paper and stamp sounds: Freesound (CC0).
What the video says
OpenAI just dropped 722 math manuscripts on GitHub. Every one came from an AI model the public can't use. The model was given about 4,000 problems. It produced 722 manuscripts, in 372 families of results.
Careful: those are different units. They don't add up to a success rate. The claims are big: the four-dimensional Kakeya conjecture, and work touching the Riemann and Hodge conjectures. But the Riemann result is a weaker version.
It doesn't solve the famous problem. Many come with proofs checked by Lean, a proof assistant. One count found 162 manuscripts with formalized main results. That's 22.
4 percent. And a machine-checked proof isn't automatically a meaningful one. These are preprints, not peer-reviewed papers. The proofs are public.
The model isn't. Now it's up to mathematicians to check the work.
Sources
Collage of public-domain archive photos via Wikimedia Commons (see the assets branch credits), xerox and halftone treatment by AI Broadsheet; ink cats, eyes and banana drawn by AI Broadsheet. Music: “DL - 120 BPM TV waiting loop” by trader_one (CC0, Freesound); paper and stamp sounds: Freesound (CC0). Report a mistake