The Anthropic Riemann 'proof' was hype: a 61GB logic dump, not a proof
gklambauer · x · 2026-11-21
A developer who has worked on the Riemann Hypothesis reviewed the Lean Zulip chat after reports that an Anthropic researcher "solved" it, and concluded the hype was misplaced:
- A logic verifier, not a proof: the AI produced a 61GB obfuscated code dump. Mathematician Joël Riou called it "messy" and nonstandard versus Mathlib conventions — a machine-verified logic tree, not readable mathematics.
- Lean deserves the credit: as Max Nowak put it, "This is not thanks to AI, this is thanks to interactive theorem provers like Lean." Human-built Lean kernels verified the dump in 7 hours.
- Market hype: formal proof expert Assia Mahboubi called for quarantining such announcements: "what we are contributing to is not science but market and finance."
The attached screenshot shows a Claude-generated "Riemann Hypothesis Verification App" with meaningless parameter sliders and a "97.2% support level," underlining how absurd the original claim was.
Related event: Mathematicians Question Anthropic's Riemann Proof as File Dump(2 posts)→
More from Models
- "Lean verified" doesn't mean Riemann solved: 17M-line formal proof awaits audit — gklambauer · 2026-11-21
- Mystery benchmark model Gestalt rumored to be Zhipu's long-awaited GLM-5.5 — VraserX · 2026-11-21
- Google's Chronic Pattern: Months of Wins (Gemini 2.5, Nano Banana, Gemini 3) Then It Stops Shipping — Angaisb_ · 2026-11-21
- Raschka: US and Chinese open-weight models made each other better — rasbt · 2026-11-21
- Bio researcher: cancel your $100/mo Claude sub, Gemini 4 has no biosafety theater — DeryaTR_ · 2026-11-21
- Intel researcher: open-source models are good enough for most AI tasks — JFPuget · 2026-11-21