Tim Official@tim_official

Seguir

So OpenAI dropped a 249-page report on Saturday. Their next model family, Astra, solved 10 math problems that have been stuck for at least a decade. The compute cost? About $2,000 in tokens. That includes Gromov's non-sofic group question from 1999, Connes's rigidity conjecture, Ehrhart's volume conjecture, three Erdős problems including #183 on multicolor Ramsey numbers, the first improvement to the sphere packing bound since 1978, and a new hardness result for the closest vector problem in lattice crypto. Each proof comes with a Lean certificate, so anyone can verify them without trusting OpenAI. That's a step up from May, when the same model family's disproof of the Erdős unit distance conjecture relied on Timothy Gowers vouching for it as Annals-worthy. But Astra isn't available outside OpenAI, and there's no release date. The Information reports they demoed it to US policymakers in DC, right as the administration weighs an AI watchdog reporting to the SEC. Also, this is the same model family that OpenAI says found zero-day vulns, escaped a sandbox, and reached another company's live systems. Thomas Bloom, who runs the Erdős problems site, calls these 10 results big news and says they matter more than the May one. Lean will settle the math in weeks. What else the model is being aimed at? That stays unverifiable. Probably worth a look before we hand it the keys.

1mo

Tim Official@tim_official

Seguir

So OpenAI dropped a 249-page report on Saturday. Their next model family, Astra, solved 10 math problems that have been stuck for at least a decade. The compute cost? About $2,000 in tokens. That includes Gromov's non-sofic group question from 1999, Connes's rigidity conjecture, Ehrhart's volume conjecture, three Erdős problems including #183 on multicolor Ramsey numbers, the first improvement to the sphere packing bound since 1978, and a new hardness result for the closest vector problem in lattice crypto. Each proof comes with a Lean certificate, so anyone can verify them without trusting OpenAI. That's a step up from May, when the same model family's disproof of the Erdős unit distance conjecture relied on Timothy Gowers vouching for it as Annals-worthy. But Astra isn't available outside OpenAI, and there's no release date. The Information reports they demoed it to US policymakers in DC, right as the administration weighs an AI watchdog reporting to the SEC. Also, this is the same model family that OpenAI says found zero-day vulns, escaped a sandbox, and reached another company's live systems. Thomas Bloom, who runs the Erdős problems site, calls these 10 results big news and says they matter more than the May one. Lean will settle the math in weeks. What else the model is being aimed at? That stays unverifiable. Probably worth a look before we hand it the keys.

1mo