Riemann Hypothesis claimed twice in one day: a Claude-driven Lean proof and a 940-page human proof
Latent Space · rss · 2026-11-21
Latent Space chronicles mathematics' most dramatic day: two independent Riemann Hypothesis proofs announced within hours.
The Claude-driven proof (Levent Alpöge, Anthropic)
- Started Oct 6; final "GRH" entry Nov 16. Released as Lean code only (250,000 pages in prose, 1.3GB)
- Elliot Glazer confirms the formalized statement is GRH and puts Lean-bug odds well under 1% — probably correct if it compiles, but no human will read it all
- Claude handled higher-rank automorphic forms, the parts Alpöge says he doesn't master; model and steering level undisclosed
- ANT stock fell 10%; unconfirmed DeepMind "math horizon" repo rumors
The human proof (Alessandro Masullo)
- A 940-page GRH draft, 12 years in the making, under secret manual review — probably technically wrong but understandable
Also covered
- Third in the frontier-math trilogy; OpenAI's September Navier-Stokes claim remains the cautionary tale; both AI Millennium claims need years of verification
- Thematic threads: whether verifiably-correct-but-unreadable Lean counts as "slop," and Weak-to-Strong alignment as lived experience
- Plus Thinking Machines' open-weights looped Transformer Inkling 2 Mini: 26.4B params (3.4B active), Apache 2.0, AA Index 20.1, DeepSWE v1.1 48.2%, ARC-AGI-2 49.3% at $0.14/task, beating Inkling 1.0 at 2.7% of the size; expert pruning + logit distillation recipe, variable-loop research model, Goodfire analysis, day-0 vLLM/Ollama/llama.cpp support
More from Models
- Google AI Mode now enforces usage limits, users surprised by rate-limit popup — elvano124 · 2026-11-22
- Sol 6.2 now refuses reverse engineering that Sol 6.1 happily did for a month — cardak98 · 2026-11-22
- Anthropic's record ~$81B IPO, GPT-6.2 pricing, Qwen4-Max and Copilot Autopilot: week in review — ksraj1001 · 2026-11-22
- Reddit rant: Frontier AI labs are training models to imitate lazy engineers — RossPeili · 2026-11-22
- MBZUAI open-sources PAN 2, a 29B-parameter general-purpose world model — LinusEkenstam · 2026-11-22
- Inkling Mini turns out surprisingly fluent in Russian, sparking data recipe questions — teortaxesTex · 2026-11-22