MIT study: best AI models went from solving 2% to 35% of theory problems in two years
nordicinst · x · 2026-10-15
WIRED covers an MIT study by Neil Thompson et al.: between mid-2024 and mid-2026, top AI models went from solving 2% to 35% of theoretical problems in probability, statistics, economic theory, high-energy physics, gravitation/cosmology and quantum physics. The team tested OpenAI models (o1 through GPT-5.5) on problems from papers published after training cutoffs, ruling out memorization. Thompson expects many theory fields to be transformed within years.
More from AGI Musings
- adam_karvonen revises robot mechanic odds to 10% in five years, citing possible ASI — adam_karvonen · 2026-10-15
- AI skeptic: trust controlled experiments, not AI labs' self-reporting — gerardsans · 2026-10-15
- Why Are Robots Writing Emails Instead of Mining? The Anthropomorphism Trap — bennash · 2026-10-15
- Scott Aaronson: Claiming AI Can't Do Math Is Now 'Completely Untenable' — haider1 · 2026-10-15
- Lancet Commission flags malicious AI among 17 catastrophic health threats — msharmas · 2026-10-15
- AI Is a Multi-Decade Optimization Through Bits Cashing Out in Atoms — herbiebradley · 2026-10-15