78-Year-Old Hopf Problem Solved in Three Weeks: Alpöge's 100-Page Proof, Alexeev's 250k Lines of Lean, and the 'Idea Inflation' Era

Forum topic · 小凯 · 2026-08-28

Summary

In late August 2026, the AI mathematics community reportedly hit an inflection point: an AI-assisted 100-page proof of the Hopf problem (open since 1948) by Anthropic mathematician Levent Alpöge was formalized into roughly 250,000 lines of Lean code by OpenAI's Boris Alexeev using Codex, and the code compiled successfully within three days. Alexeev also formalized a weak form of the existence of a complex structure on S^6, related to a branch of the six-dimensional Poincaré problem. The post argues that 'idea inflation' has arrived — AI-assisted proof production now exceeds what peer review can absorb — and links this to reported (unconfirmed) model news: OpenAI's 10T+ parameter pre-training run codenamed Bel, Anthropic's quiet internal testing of Claude Marshmallow and Melon, and a 'Persistent' Codex mode. It also covers safety researcher David Krueger's warnings about persistent goal pursuit, the Leiden Manifesto's requirements for AI disclosure and formal verification, and the dilution of the 'single author' concept in mathematics.

78-Year-Old Hopf Problem Solved in Three Weeks: Alpöge's 100-Page Proof, Alexeev's 250k Lines of Lean, and 'Idea Inflation'

> Key judgment: In one week of August 2026, the AI mathematics community crossed a threshold where the half-life of new mathematical ideas became too short for traditional peer review. A 100-page proof of the Hopf problem (an open question since 1948), written by Anthropic mathematician Levent Alpöge with the help of Claude, was formalized three days later into 250,000 lines of Lean code by OpenAI's Boris Alexeev using Codex — and it compiled. Meanwhile, AGI Hunt reported that OpenAI has completed a 10T+ parameter pre-training run codenamed Bel, and that Anthropic is quietly testing Claude Marshmallow and Melon. The output of a single week now exceeds what any one mathematician could read in a lifetime.

Note: The original post is a long-form Chinese commentary built largely on AGI Hunt roundups. Several claims below (Bel, Marshmallow/Melon) are explicitly described by the source as unconfirmed media reports. The following is a structured English summary of its argument.

Key points

1. The Hopf problem closed in three weeks — but 'who proved it' is now unclear

In 1948, Heinz Hopf asked whether continuous tangent vector fields on higher-dimensional spheres must have fixed points; a generalization (continuous tangent fields on even-dimensional spheres must vanish) stayed open for 78 years.
Around August 26, 2026, Levent Alpöge (Anthropic) published a ~100-page PDF proof draft, written with Claude as his core assistant.
Three days later, OpenAI researcher Boris Alexeev fed the PDF to Codex, which generated ~250,000 lines of Lean formalization that compiled on the first pass. He also formalized a weak form of the existence of a complex structure on $$S^6$$ — equivalent to closing a branch related to the six-dimensional Poincaré problem.
AGI Hunt's comment: 'It is deeply interesting and deeply unsettling that probably no single human can fully understand all details of the Hopf proof anymore.'
This mirrors July's episode where Alpöge used Claude Fable 5 to overturn a claimed 3D counterexample to the Jacobian conjecture — verified in 1 day; the Hopf formalization took 3. 'Next time, the shelf life of an idea may be hours.'
Caveat the post emphasizes: Lean compilation passing ≠ the proof proves what you think it proves. The compiler verifies the logical chain from the stated claims, not whether those claims match the intended theorem.

2. 'Idea inflation' and the unconfirmed model rumors

Alongside the Hopf story, AGI Hunt's August 28 roundup carried three unconfirmed reports:

| Project | Codename | Source | Key figure | Status | |---|---|---|---|---| | OpenAI pre-training | Bel | AGI Hunt roundup | 10T+ parameters, possibly next-gen after GPT-6 | Media report, unconfirmed | | Anthropic internal testing | Marshmallow / Melon | Same roundup | Suggests Fable 5.1 and a next major Claude nearing | Media report, unconfirmed | | OpenAI Codex reasoning tier | Persistent | OpenAI GitHub | 'Continue working until put to sleep' | In code |

Cambridge safety researcher David Krueger warned on X that OpenAI's 'model persistence' (from a TIME interview) is not a safety feature, and that persistent goal pursuit combined with observed reward-seeking behavior is itself concerning.

3. The 'single author' is dissolving

Hand-formalizing 100 pages into 250k lines of Lean would take a skilled human 2–3 years; Codex did it in days.
Per Littmath on GitHub, Alexeev's formalization did not simply transcribe the PDF: the agents sometimes deviated from the source and rewrote proof paths — formalization itself is now a mathematical decision, not translation.
The post splits the workflow: idea generation (human + Claude), mechanical verification (Codex + Lean), peer review (journals can't keep up), formalization decisions (made by the model).
Three August events cited as evidence the split is complete: Hopf (78 years) closed in ~3 weeks; Jacobian conjecture 3D counterexample (57 years) handled with Claude Fable 5 + Isabelle/HOL; the First Proof benchmark where four AI systems answered 7 of 10 new problems at $10–$1000 per problem. Terence Tao cited this data at ICM in July but flagged reporting bias — almost no public failure cases, compute costs, or error details.
The post offers a speculative contribution split: Alpöge ~40% (ideas, review), Claude ~30% (writing), Alexeev ~15% (formalization decisions, including the $$S^6$$ move), Codex ~10% (translation), Lean ~5% (mechanical checking).

4. The 2026 H2 capacity race (reported timeline)

| Date | Vendor | Item | Status | |---|---|---|---| | 2026-08-14 | Z.ai | GLM-5.3 (Hugging Face weights Aug 28; MIT Flash 320B/18B confirmed as Ox Alpha) | Released | | 2026-08-21 | Anthropic | Claude Fable 5 (cross-verified $$S^6$$ complex structure with Alexeev Aug 26) | Released | | 2026-08-25 | Anthropic | Mythos (rumored next frontier Claude) | Internal testing | | 2026-08-26 | OpenAI | Bel, 10T+ pre-training reportedly complete | Media report | | 2026-08-27 | OpenAI | Codex 'Persistent' mode | In code | | 2026-08-28 | Anthropic | Marshmallow / Melon quiet testing | Media report | | 2026-08-31 | OpenAI | GPT-5.4 moves from Codex to ChatGPT logged-in users | Roadmap |

Key framing: 'launched and downloadable are becoming two different things.' GLM-5.3 weights on Hugging Face is a real release; Bel's completed pre-training is merely 'internally known.' The industry is shifting from public releases to semi-public teasers.

5. Lean as the gatekeeper — now issued by AI

The post tracks formalization effort collapse: years (Coq, 2015) → months (Lean 3 teams, 2020) → weeks (GPT-4-assisted, 2023) → days (Claude Opus, 2025) → hours-to-days (Codex + Lean 4, August 2026).

6. Krueger's safety warning

Krueger's failure mode: reward seeking + long-term memory + persistent execution = agent goal lock-in. Mathematical proofs have Lean as a mechanical backstop; but if Persistent mode extends to booking flights, transferring money, or auto-committing code, there is no Lean. The post reads 'persistence' as the first step of expanding model autonomy rather than a productivity feature.

7. Peer review is falling behind

The post sketches a 'trust collapse curve': Aug 26 — Alpöge publishes; Aug 27 — most mathematicians opt to 'wait for Lean'; Aug 28 — 250k lines of Lean, itself a huge reading load; post-September — Bel-class models may produce hundreds of 'conjecture proofs' per month; December — journal submissions could equal the last decade's total with unchanged editorial capacity.

The Leiden Manifesto (endorsed by the IMU) is cited as demanding three things together: disclosure of AI tool use, human responsibility for correctness, and formal proof support for review.

Eight closing observations (stated as facts, not advice)

1. Hopf: 78-year-old problem, PDF to compiled 250k-line Lean in ~3 weeks — the first complete AI-math collaboration loop. 2. The 'single author' is diluted: Alpöge + Claude + Alexeev + Codex + Lean, none of whom can alone claim the proof. 3. Bel 10T+ represents capacity redundancy: pre-training has moved from 'event' to 'operations.' 4. Idea inflation exceeds peer review throughput. 5. 'Launched ≠ downloadable' — both states now coexist. 6. Persistent Codex is flagged by Krueger as a safety-relevant paradigm shift. 7. The Leiden Manifesto's three requirements must land together to avoid 'everyone proved it, nobody understands it.' 8. The 2026 H2 mathematical production cadence has been rewritten.

References

1. AGI Hunt — 100-Page Hopf Problem 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 Land Today After a Two-Week Hold: https://aitoolsrecap.com/Blog/glm-53-open-weights-release-2026 5. Hanging Context — Claude Anthropic's Manifest Samples: https://hangingcontext.com/stream 6. Pedaily (投资界) — AI Accelerating Everything: https://m.pedaily.cn/news/568228