2026-10-07 / Signal #1
OpenAI’s unreleased model dumped 722 math manuscripts (372 result families) on GitHub
“A model you’ve never been allowed to talk to just published more math papers in one drop than most departments produce in a decade.”
The Story
OpenAI released 722 manuscripts grouped into 372 result families across 17 fields of mathematics, produced by an unreleased internal frontier model from roughly 4,000 problems. Most came from a single-agent procedure averaging three hours of ChatGPT Pro thinking compute; 162 have Lean-formalized main results. Some claims include a quasi-Riemann zero-free region and various conjectures and counterexamples; OpenAI notes some unformalized results could have issues.
Why It Matters
AI is no longer assisting math—it is flooding the field with papers faster than humans can read or referee them, turning research into an industrial output with versioned GitHub artifacts. Builder utility for anyone tracking how science actually gets done now; future-shock that the queue of human insight just got machine-width.
Evidence
OpenAI announcement and GitHub repo; Scientific American, CellCog, and independent indexes of the collection.
Sources
Daily scan: 2026-10-07