What's this about?
New analysis plotting discovery rates across fields to test whether AI is producing measurable acceleration. The tentative conclusions, presented as tentative: cyber shows sharp acceleration, math shows some acceleration, and algorithms show no clear acceleration. Ethan Mollick's reading is that this is early evidence of acceleration appearing exactly where you'd expect it first.
What exactly makes this hard to measure?
Discovery has no natural counter. You can count papers, but publication volume responds to incentives more than to insight. You can count citations, but they lag by years. And you face a baseline problem: fields accelerate and decelerate for many reasons, so attributing a change to AI requires knowing what would have happened otherwise. That's why the field-by-field split is the interesting move — if AI were causing the acceleration, you'd expect it in domains where AI is strongest and verification is cheap, and not in domains where it isn't. Cyber, where an exploit either works or doesn't, is exactly the first case.
Why is this worth noticing?
- The negative result is as valuable as the positives. No clear acceleration in algorithms, despite enormous attention to AI for algorithm discovery, is a real data point against the strongest claims.
- The pattern is internally consistent, which is what makes it more than noise: acceleration tracks verifiability, being sharpest where results are automatically checkable.
- The honest hedging is the right posture — the authors call the conclusions tentative, and this is one of the first attempts to measure rather than assert the acceleration everyone assumes.