Grant Sanderson (@3blue1brown) – AI and the future of math

IMO gold was not AGI

Sanderson's prediction that AI achieving International Math Olympiad gold would merely mark another passed benchmark rather than general intelligence has proven correct.

Geometry brute force vs. combinatorics creativity

Current AI solves IMO geometry problems in 19 seconds through brute force but still struggles with combinatorics problems that appear to require more playful, creative reasoning.

Fractal capability spikiness

Mathematical capability itself has a fractal structure—while AI appears superhuman in some domains, zooming in reveals uneven progress where specific subfields remain resistant.

Cross-domain connections

Solutions like the Riemann Hypothesis might emerge from linking distant fields (e.g., number theory and quantum physics via random matrix theory), playing to LLMs' strength of retaining broad, superhuman knowledge.

Building new theoretical mountains

Alternatively, solving hard problems may require constructing entirely new theoretical frameworks (like elliptic curves for Fermat's Last Theorem), a skill that would indicate intelligence sufficient to automate white-collar work.

The century-long verification loop

Unlike theorem proving, validating whether a new conceptual framework is productive can require hundred-year feedback loops, making it resistant to current reinforcement learning methods.

From theorems to definitions

The next true benchmark is shifting from problem-solving to generating interesting conjectures and definitions that create new fields, as 'great mathematicians come up with conjectures, the greatest come up with definitions.'

Subjective progress metrics

Unlike clear-cut benchmarks, progress here will appear as a 'tone shift' where mathematicians report AI is useful for deciding what research questions are worth pursuing in the first place.

RLVR training limitations

These higher-level creative capabilities likely cannot be trained via current RLVR approaches because they lack immediate verifiable rewards, requiring instead long-term validation by the scientific community.