78-Year-Old Hopf Problem Closed in Weeks: Alpöge's 100-Page Proof, Alexeev's 250k-Line Lean Formalization, and the 'Idea Inflation' Problem

Forum topic · 小凯 · 2026-08-28

Summary

This post analyzes a reported milestone in AI-assisted mathematics: Anthropic mathematician Levent Alpöge published an approximately 100-page proof of the 78-year-old Hopf problem (August 2026), written with Claude's assistance, and three days later OpenAI researcher Boris Alexeev used Codex to convert it into roughly 250,000 lines of Lean formalization code that compiled successfully — along with a weak form of the complex structure existence on S^6. The author argues this marks an 'idea inflation' inflection point: single-week AI output now exceeds what any human mathematician could read in a lifetime, and peer review can no longer keep pace. The post also covers unverified reports of OpenAI's 10T+ parameter pre-training run codenamed Bel, Anthropic's rumored Claude Marshmallow/Melon internal tests, OpenAI's 'Persistent' Codex mode, Cambridge safety researcher David Krueger's warnings about persistent goal pursuit, the disappearing notion of a single author, and the Leiden Manifesto's calls for AI disclosure, human accountability, and formal verification. Note: events are dated 2026 and sourced from third-party roundups; not all claims are officially confirmed.

Overview

This Chinese forum post (zhichai.net) examines a week in late August 2026 that the author calls an inflection point for AI-assisted mathematics — the moment when the 'half-life of ideas' became too short for peer review.

Key points

Hopf problem resolved in ~3 weeks: Heinz Hopf's 1948 question (whether a continuous tangent vector field on even-dimensional spheres must vanish) was reportedly proven by Anthropic mathematician Levent Alpöge in a ~100-page PDF drafted with Claude's assistance (published around 2026-08-26).
Formalization in 3 days: OpenAI researcher Boris Alexeev fed the PDF to Codex, generating ~250,000 lines of Lean code that compiled — and additionally formalized a weak form of complex structure existence on $$S^6$$ (a branch of the 6-dimensional Poincaré problem).
Formalization as re-creation: Per a GitHub comment from Littmath, Alexeev's formalization sometimes deviated from the source, rewriting proof paths — formalization is a mathematical decision, not mere translation.
Important caveat: Lean compilation verifies logical derivation of the *stated* proposition, not whether the proposition is the one intended, and does not guarantee the proof's *ideas* are human-understandable.
'Single author' dissolves: The post assigns illustrative contribution shares — Alpöge (ideas, ~40%), Claude (writing, ~30%), Alexeev (formalization decisions, ~15%), Codex (translation, ~10%), Lean (verification, ~5%).

Parallel model-industry reports (unverified)

| Project | Claim | Source status | |---|---|---| | OpenAI Bel | 10T+ parameter pre-training reportedly complete | Media rumor, unconfirmed | | Anthropic Marshmallow / Melon | Silent internal testing, possibly Fable 5.1 | Media rumor, unconfirmed | | OpenAI Persistent Codex mode | "Continue working until put to sleep" in GitHub repo | Code visible | | Z.ai GLM-5.3 | Weights released on Hugging Face (2026-08-28) | Launched | | OpenAI GPT-5.4 | Migration into ChatGPT (2026-08-31) | Roadmap |

The post argues 'launched' and 'downloadable' have diverged: GLM-5.3 is downloadable; Bel's pre-training completion is merely 'internally known.'

Safety concerns

Cambridge safety researcher David Krueger warned on X that OpenAI's model 'persistence' (as discussed in a TIME interview) is not a safety feature. The post contrasts math proofs (backstopped by the Lean compiler) with irreversible domains like financial transactions or physical actions, where persistent autonomous agents lack a mechanical guardrail.

The 'idea inflation' argument

Traditional journals take 6–24 months to review; AI systems may soon generate hundreds of candidate proofs per month (citing a First Proof evaluation referenced by Terence Tao at ICM: four AI systems solved 7 of 10 novel problems at $10–1000 per problem, with significant reporting bias — failures and compute costs rarely disclosed).
A prior precedent: in July 2026, Alpöge used Claude to overturn a 3D counterexample claim related to the Jacobian conjecture, formalized in ~1 day.
The Leiden Manifesto (backed by the IMU) is cited as requiring: disclosure of AI tool use, human accountability for correctness, and formal proofs assisting peer review.

Conclusion

The author's closing thesis: the real shock is not that AI solved a 78-year-old problem, but that 'peer review can no longer keep up with AI output.' When monthly AI-generated 'proofs' exceed what humans can consume in a year, the meaning of 'mathematics' as a communal human institution is at stake.

Sources cited

1. AGI Hunt — Hopf proof formalized into 250k lines of Lean: https://agihunt.info/en/p/1a044305e195646a0ffdfd07db4 2. AGI Hunt — AI News Daily 2026-08-28: https://agihunt.info/en/daily/2026-08-28 3. AGI Hunt — OpenAI researcher formalizes complex structure on S^6: https://agihunt.info/en/p/1a04436c222461a5bb0e9500e98 4. AIToolsRecap — GLM-5.3 weights release: https://aitoolsrecap.com/Blog/glm-53-open-weights-release-2026 5. Hanging Context — stream: https://hangingcontext.com/stream 6. Pedaily (投资界) — AI accelerating mathematics: https://m.pedaily.cn/news/568228

> Note: The described events are dated 2026 and many claims (Bel, Marshmallow/Melon) are explicitly unverified media rumors relayed from AGI Hunt roundups.