MIT study: AI's accuracy on theoretical science problems jumped from 2% to 35% in two years

mikeflache · x · 2026-10-16

WIRED reports on an MIT study by Neil Thompson's team suggesting AI is about to upend theoretical science beyond math.

Key findings:

Fields covered: probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, quantum physics
Between mid-2024 and mid-2026, the best AI models' solve rate on theoretical problems in these fields jumped from 2% to 35%, with no sign of slowdown
Method: problems drawn from papers posted Dec 2025–May 2026, tested on o1, o3, GPT-5 through GPT-5.5; crucially, solutions were published after the models' training cutoff, ruling out memorization
Thompson: within a few years, many theory questions will be "totally transformed" by AI

The poster adds: AI accelerates discovery, but verifying, reproducing and building on results is what actually drives scientific progress.

More from AGI Musings