An Anthropic mathematician claims to have proven the Riemann hypothesis. We have to rethink how we reward expertise.
An Anthropic mathematician using Claude claims to have proven the most famous problem in math — the Riemann hypothesis — in the latest, but potentially most consequential, example of artificial intelligence (AI) producing stunning math results.
- Anthropic's proof will certainly and rightly be scrutinized, but rather than picking apart every single result these models produce, we should respond by changing how we evaluate and reward expertise — placing less weight on the volume of results and more on the human judgment required to decide which questions and answers matter.
- The job of human mathematicians will be to become arbiters of the correctness of proofs, but also to scrutinize the reasoning and to extract the underlying ideas, determining whether the result says anything that meaningfully advances the field.
- To give a simplified example, AI might "prove" that the square of a number divisible by two is divisible by four and then prove that the square of a number divisible by three is divisible by nine, and so on with ever larger integers.
Mathematical judgment is the ability to distinguish a result that is correct, but boring, from one that might fundamentally change how we think about the problem or even the whole area of mathematics surrounding the problem.
- Earlier this year, an OpenAI system produced a counterexample to the long-standing planar unit distance problem, overturning a conjecture of Paul Erdős that had resisted mathematicians for decades.
- More recently, OpenAI announced an apparent solution to the Navier-Stokes problem, a celebrated question about whether equations describing fluid flow can develop singularities, or infinite values.
- A proof can be valid without offering insight into why it is true or how its ideas might be employed elsewhere.
- Some proofs may be so long or opaque that checking them requires substantial work by specialists and even more work to reinterpret them in a way that is meaningful to human mathematicians.
- AI could generate hundreds of thousands of complex variations on a theorem, but these variations might add very little to mathematical knowledge, each one ultimately proving a dead end.
How this summary was made: extractive summary — built from the sources' own reporting. It is drawn only from the reporting of the sources listed below — read the originals for full detail. How we summarize →