OpenAI walks back two of ten Astra math claims

OpenAI's unreleased Astra model claimed ten mathematical breakthroughs, and two of them turn out to rest on other people's published proofs, per Scientific American.

That's a fast reversal for research OpenAI touted as a landmark days earlier. Anyone citing Astra's math to gauge AI's research ability now works from a tally where two of ten claims already needed correction.

Steven Miller, a mathematician at Yeshiva University, says Astra's sphere-packing improvement for dimensions above 1,000 borrows an uncredited argument from his 2016 paper. He calls it research misconduct.

Francesco Fournier-Facio at Cambridge found the same pattern in a different result. The soficity proof Astra gets credit for combines a 2016 paper and a 2019 paper that already existed, he told Scientific American.

An OpenAI spokesperson acknowledged responsibility and said updates would follow "consistent with standard academic practice." The statement doesn't say whether the other eight claimed advances have been checked against prior work, or when the updates land.