MIT study: frontier AI solves cutting-edge theory problems at 34.6%, up from 2.8% in 16 months
MIT_CSAIL · x · 2026-10-17
A new MIT study (Manso, Mayrink Verdun, Thompson), "AI Disruption and Falling Costs in Theory Research," systematically measures AI progress on frontier theory questions.
- Method: 254 recent arXiv papers across seven domains in math, physics, and economics; human authors' conclusions served as ground truth unavailable to the models
- Results: the share of problems where models matched human authors rose from 2.8% (o1, Dec 2024) to 34.6% (GPT-5.5, Apr 2026) — about 22 percentage points per year, with gains spread evenly across subfields
- Cost: $5 per attempt, $15 per success for GPT-5.5 (down from $170 per success for o1)
- Implication: some areas of theoretical research are becoming dramatically more cost-effective, and the new bottleneck may be verification
More from AGI Musings
- Viral take: does AI now handle most tasks software engineers used to do? — tlakomy · 2026-10-17
- Lawyer uses Grok to map the legal scaffolding behind a band's Letterman debut — jkubicki · 2026-10-17
- Lawyer pushes back on AI-replacement takes: judgment, not documents, is the job — jkubicki · 2026-10-17
- MIT Sloan framework: when can AI make the call? Ambiguity and stakes decide — moniquejmorrow · 2026-10-17
- MIT Sloan offers a framework for when AI should make decisions — moniquejmorrow · 2026-10-17
- Finance doesn't believe in the singularity; tech does — and the gap matters — shakoistsLog · 2026-10-17