AI‑proved theorems reshape research

Updated: 2026.11.20 22H ago 1 sources
AI‑assisted formal proofs (a Claude‑helped Lean proof claiming Riemann, competing human drafts, and an AI‑related Navier–Stokes preprint) plus benchmarked capability growth suggest that models can generate checkable mathematical results faster than communities can digest them. This creates new bottlenecks in human verification, shifts reward and citation norms, and forces policy questions about how to treat AI‑produced 'discoveries' and computational certificates. — If models routinely produce formal, machine‑checkable breakthroughs, academic credit, reproducibility norms, and AI‑safety oversight all need rapid revision.

Sources

Links for 2026-11-20
Alexander Kruel 2026.11.20 100% relevant
Levent Alpöge’s Lean package (Claude‑assisted) accepted by independent checkers; METR report showing task‑duration doubling every ~3.6 months; and the public Navier–Stokes preprint missing its numerical certificate.
← Back to all ideas