AI Moves From Mathematics to the Frontiers of Theoretical Science

Researchers at MIT, led by Neil Thompson, tracked the performance of successive OpenAI models — from o1 through GPT‑5.5 — on problems taken from papers published after each model’s training cutoff. By denying the models internet access, the team ensured that solutions had to be generated from first principles rather than retrieved from existing literature.

Between mid‑2024 and mid‑2026 the success rate on theoretical tasks in probability, statistical theory, economic theory, high‑energy theory, gravitation and cosmology, and quantum physics jumped from about 2 % to roughly 35 %. The study focused on domains where reasoning alone suffices, without the need for simulations or laboratory experiments.

The rapid ascent mirrors the upheaval already seen in pure mathematics, where an unreleased model recently solved the long‑standing Navier‑Stokes problem. As AI begins to automate the derivation of theoretical results, the traditional workflow of conjecture, proof, and publication faces a fundamental shift.

Why it matters for GPU / AI infrastructure

Training and serving the massive models that achieve these scores demand dense GPU clusters, high‑bandwidth interconnects, and efficient memory hierarchies. Cloud providers that can deliver scalable, low‑latency GPU capacity will become the backbone of scientific discovery, enabling researchers to iterate on AI‑generated hypotheses at unprecedented speed.