XNews

Anthropic Claude Proves Breakthrough in Riemann Hypothesis

Anthropic Claude Proves Breakthrough in Riemann Hypothesis

Anthropic revealed that an undisclosed research version of Claude achieved a major breakthrough on a problem related to the Riemann Hypothesis, increasing the known lower bound of non-trivial zeros on the critical line from 41.6% to 67.2%. - The Riemann Hypothesis (1859) concerns the distribution of prime numbers; proving all zeros lie on the critical line is a Millennium Prize problem. - The previous best bound of 41.7% was reached in 2020 by Pratt et al., capping 80 years of incremental progress. Claude’s 26-percentage-point jump is the largest single-step improvement in the problem’s history. - Claude combined recent work by Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and a 2000 paper by Bombieri, finding a novel synthesis. - The process: Jarred Sumner (a non-mathematician) instructed Claude to "seriously attempt" the problem. Claude first generated 650 ideas, all failed, then spent 1.5 days coordinating ~60 sub-agents, running 2,400 commands, and writing hundreds of Python scripts. - Sub-agent roles: 2 core math, 13 idea contributors, 30 attempted but failed, 13 verifiers, 2 paper writers. - Anthropic mathematicians Levent Alpöge and Ralph Furman validated the proof; external experts Brian Conrey and Dan Goldston reviewed it. Claude also produced a machine-verifiable Lean formal proof. - This follows Claude Fable 5’s discovery of a counterexample to the Jacobian Conjecture, overthrowing the 87-year-old statement in its general form. - Anthropic stated the technique is unlikely to directly prove the Riemann Hypothesis, but shows rapid AI math progress. Full paper, Lean proof, notes, and Claude’s reasoning logs are publicly available.

Key insights

Unprecedented Leap in Riemann Zeta Zeros Bound

Claude raised the lower bound from 41.6% to 67.2%, a 26-point gain, while prior progress from 33% to 41.7% took nearly 50 years. This is the largest single jump in the problem’s 167-year history, achieved by combining recent works from multiple mathematicians and a 2000 paper by Bombieri. It demonstrates that AI can synthesize existing research into original breakthroughs even on iconic unsolved problems.

Multi-Agent Research Architecture Enabled the Breakthrough

Claude operated through 60 sub-agents with specialized roles: 2 core math, 13 idea generators, 30 who tried but failed, 13 verifiers, and 2 paper writers. Over 1.5 days, it ran 2,400 commands and wrote hundreds of Python scripts, iterating after 650 initial failures. This orchestration points to a scalable model for AI-driven mathematical discovery beyond simple Q&A.

Rigorous Verification by Mathematicians and Formal Proof

Two Anthropic mathematicians and two external experts (Conrey and Goldston) reviewed and validated the proof. Additionally, Claude wrote a machine-checkable Lean formalization, setting a standard for AI-generated research output by pairing human review with verifiable code.

Pattern of Original AI Mathematical Discoveries

Three weeks earlier, Claude Fable 5 found a counterexample to the Jacobian Conjecture, an 87-year-old problem. Both cases involve AI proposing novel solutions on decades-old open problems, not just solving known exercises. This pattern indicates AI is moving into creative, research-level mathematics.

Source posts