Building an AI Physicist: ChatGPT Co-Creator’s Next Venture

Periodic’s approach treats nature as the RL environment: simulations can guide search, but experimental measurements ultimately correct simulator deficiencies and provide hard-to-game reward signals.

They distinguish the scaling curve from its evaluation distribution: in-domain internet/test benchmarks may improve predictably, while out-of-domain physics tasks can have such a shallow improvement slope that progress would be impractically slow without targeted data and objectives.

They argue discovery requires repeated cycles of hypothesize → simulate → test → update; without the ability to act and learn from outcomes (including failures), models can at best reproduce literature distributions rather than reduce epistemic uncertainty.

Reported properties can vary by orders of magnitude, negative results are rarely published, and context dependencies are high—so training purely on papers can yield weak predictors; a dedicated lab can systematically generate cleaner, more informative datasets.

As a phase transition, superconducting behavior is comparatively robust to hard-to-simulate microstructural details; it also provides a crisp success metric (raising Tc) and motivates building end-to-end autonomy across synthesis, characterization, and simulation.