Claude makes a statistical leap on the Riemann hypothesis

Anthropic’s unreleased chatbot Claude has raised the proven share of non‑trivial zeros of the Riemann‑zeta function that lie on the critical line from 41.6 % to an impressive 67.2 %. The advance does not constitute a proof, but it establishes a new lower bound that eclipses the previous record by more than 25 percentage points.

Why the Riemann hypothesis matters

The hypothesis, proposed by Bernhard Riemann in 1859, links the distribution of prime numbers to the zeros of a complex function called the zeta function. If every non‑trivial zero falls on the vertical line Re(s)=½ (the “critical line”), primes would follow a far more regular pattern than the apparent randomness suggests. A proof would earn a million‑dollar prize from the Clay Mathematics Institute.

Historical benchmarks

Mathematicians have long tried to show that at least a certain percentage of zeros sit on the critical line. In 1974, Norman Levinson proved that more than one‑third do. Fifteen years later, Brian Conrey pushed the figure past two‑fifths. The next incremental gain—just 1.7 %—took another three decades.

Claude’s unconventional strategy

The breakthrough emerged when Anthropic engineer Jarred Sumner asked Claude to “seriously try to solve the hypothesis.” After an initial failure, the model spawned 60 parallel instances that worked for a day and a half, executing roughly 2,400 distinct commands, writing hundreds of programs, and performing thousands of checks on known zeros. Some clones were even tasked with probing gaps left by their peers, creating a collaborative “army” of AI agents.

Collectively the clones processed about 31 million tokens, synthesising recent research, older studies, and a wealth of numerical data. Rather than locating each zero individually, Claude built a statistical framework that weighs points on the line against those off it, yielding a rigorous lower bound without pinpointing every single zero.

What the result means for mathematics

While the 67.2 % figure still falls short of the 100 % required for a full proof, it dramatically narrows the gap and demonstrates that large‑scale language models can contribute to deep, abstract problems. The episode also highlights a new research paradigm: AI systems acting as distributed research assistants, generating code, testing conjectures, and cross‑checking literature at a speed no human team could match.

Future work will likely involve refining Claude’s statistical model, extending the token budget, and perhaps integrating more specialized mathematical reasoning modules. Whether an AI will ever deliver a complete proof remains uncertain, but the current milestone suggests that machine‑assisted exploration is becoming a serious contender in the quest to resolve one of mathematics’ oldest enigmas.

Source: https://scientias.nl/ai-model-claude-maakt-vooruitgang-met-riemann-hypothese-maar-lost-hem-niet-op/