DeepMind researcher argues AI math breakthroughs demand new progress model

Kasirzadeh distinguishes two models of mathematical progress. The checklist model treats progress as ticking off predefined problems (Olympiad, Millennium Prizes), which AI now rapidly solves. The generative encounter model sees progress as open-ended, arising from messy real-world encounters like Dantzig's invention of linear programming from military logistics or Wiener's cybernetics from anti-aircraft targeting. She argues the pure-applied hierarchy that elevated checklist problems is a cultural prejudice AI exposes.

Four urgent problem clusters need human-AI collaboration: provably safe cyber infrastructure, mathematical science of deep learning, multi-agent control mathematics, and epistemology of machine-generated proofs. Kasirzadeh contends the future depends on redesigning credit, education, and institutions around human purpose, not proof discovery alone.