Physix FrontierPhysix Frontier · News Briefing Card (TMTPost · Oct 13, 2026)
OpenAI Retracts 3 of 722 AI Math Papers; Microsoft Builds on Alibaba Qwen
KEY FACTS
- On October 6, OpenAI published 722 AI-generated mathematics manuscripts on GitHub, including 377 new proofs.
- Within two days of release, 3 papers were retracted because sign errors caused their arguments to collapse, and another 14 were corrected.
- Currently only about 42% of the main conclusions have passed Lean formal verification.
- In October, Microsoft released a decision model trained on Alibaba's open-source model Qwen3.5-9B.
- The model achieved an average accuracy of 83.5% on 36 blind-test benchmarks, with a P50 latency of 85 milliseconds.
KEY DATA
722Published math manuscripts
377New proofs
about 42%Lean formal verification pass rate
83.5%Microsoft decision model average accuracy
PHYSIX OBSERVATION
OpenAI used an unreleased model to mass-produce mathematical proofs, yet it did not even filter out sign errors; a 42% verification rate shows a clear gap between benchmark scores and real academic standards. Microsoft turning to Alibaba's Qwen as its base releases another signal: open-source small models are becoming big tech's first choice for low-cost trial and error, and exclusive closed-source lock-in is no longer solid.
Source: TMTPost report