OpenAI has dumped 722 maths papers – now it must clean up the mess
OpenAI is using mathematics as a test site for its most capable AI models, and that means it has a responsibility to deal with the fallout, rather than just leaving it to mathematicians, says Jacob Aron
OpenAI has released 722 mathematical papers onto the code-sharing website GitHub, in what appears to be a campaign against open problems in mathematics. The papers were generated using an unnamed internal model, with each result requiring the equivalent of three hours of ChatGPT Pro thinking compute. This is not the first time OpenAI has made such a release, as the company previously generated a Navier-Stokes result that took around 88 hours and cost $15 million to produce.
The release of these papers confirms that leading AI models can produce seemingly research-level mathematics at the push of a button. However, this does not necessarily mean that OpenAI is close to achieving artificial general intelligence, which is an AI model that can do anything a human can. Mathematics is different from other areas in that it has a crucial component of verifiability, as a proof is either true or it isn't.
This provides a loop for improving an AI's mathematical ability by having it produce a proof, formalise it, and reward the model for accuracy.
OpenAI's release is incomplete, as only 162 of the 722 papers have had their main results formalised in the Lean proof format. The remaining papers are left for mathematicians to formalise, which is akin to dumping "carcasses of raw meat onto our communal village table". It is unclear why OpenAI has not finished the job of formalisation, but it could be because the work has raced ahead of the current state of the art of formalisation, or because the results require new formalisation bricks that are not yet available.
Written by urgent.news from New Scientist's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
