FrontierScience & the state of AI research capability · May 2026

Where AI Stands vs Human Researchers

AI now beats expert humans on closed-form science exams, competitive math, competitive coding, and protein-structure prediction. It still can't do real research — open-ended problems with novel reasoning, niche depth, and physical experiments. The gap is closing exponentially, but it is still a gap.

The 5-second read.
AI's ability to do a task doubles every ~4 months (METR, 2026). It already dominates on speed, scale, and standardized exams.
Humans still own open-ended, multi-step, hypothesis-driven research — the work that actually moves a field.
Capability landscape — AI score vs human-expert score (0–100, normalized)
hover any bar for source & insight
capability / research task
score 0 — 100  ·  human (base) · AI ahead · AI behind