On November 20, two separate announcements landed for one of mathematics’ biggest unsolved problems. Levent Alpöge, working with Anthropic's Claude, released a formal Lean code proof of the Riemann Hypothesis. Just hours later, Alessandro Masullo published a 940-page draft of a human-written proof he had been developing for over a decade.
Both claims target the same goal: a solution to the Riemann Hypothesis, which sits atop the list of open mathematical problems. But the approaches could not be more different. Alpöge’s proof is encoded in Lean, a theorem prover, and reportedly spans the equivalent of 250,000 pages. Masullo’s draft is a traditional document, the result of years of manual work and review.
The Claude-Driven Proof Process
Alpöge’s work began in early October, with eight out of twenty proof steps completed by mid-November. The final entry, labelled "GRH", appeared on November 16. The formalised statement covers the Generalized Riemann Hypothesis. Elliot Glazer, reviewing the Lean code, stated the chance of a Lean bug is "well under 1%". If the code compiles and uses no additional axioms, correctness is likely, though the proof is too vast for any person to fully read.
The process upended the typical human-AI collaboration dynamic. Alpöge credits Claude for handling the most complex parts, especially those involving automorphic forms on higher-rank groups—areas outside his own expertise. He described Claude as having "far outpaced" him near the finish line, with the AI taking the lead in critical sections.
Human Scale Versus AI Scale
Masullo’s 940-page manuscript stands in stark contrast. Developed over twelve years and reportedly reviewed in secret, it is written for human comprehension and review. However, it lacks the formal guarantees provided by theorem provers like Lean. This contrast has sparked debate: is it better to have an incomprehensible but likely correct machine proof, or a readable but less certain human one?
Some commentators argue that these developments show how little we understand about the inherent difficulty of mathematical problems. Others compare the AI approach to large-scale "bashing" seen in maths competitions, but on a scale humans cannot match. Omar Khattab highlighted the "jaggedness" of current AI progress—solving top-tier problems while struggling with much simpler tasks.




