@dgm.america: OpenAI posted 722 math manuscripts from an unreleased internal model: about 4,000 problems posed, roughly 3 hours of thinking per result. A day later, one sign error took down three papers, because two of them were built on the first. 14 more were revised. About 42% of the main results are checked in Lean (300 of 719). For the rest, the repo itself says some "could have issues." AI output is cheap. Checking is the work. osFoundry Research checks each claim in a report against its source, and Code never auto-merges agent work. osfoundry.io/platform #AI #AINews #OpenAI #math #AIagents

dgm.america

Open In TikTok:

Region: JP

Friday 16 October 2026 12:08:45 GMT