The field closest to the models is the one with the least lift

Differential acceleration is the rule for how AI enters science. METR's study, which measured AI's contribution across three domains, found the impact is a scatter, not a wave. The pattern is the news.

METR looked at cyber, mathematics, and AI research itself. Cybersecurity is mid-phase-change: vulnerability discovery in cURL, OpenSSL, Firefox, and Microsoft accelerated sharply in 2026 versus 2025, with Aisle disclosures and US NVD aggregates behind the curve. Mathematics shows modest movement: more arXiv submissions in some subfields and named problems cracked, including Smale's Jacobian conjecture and parts of Green's problem list. AI research, the domain closest to the models, shows no measurable acceleration on the algorithmic benchmarks METR tracked. CIFAR-10, Hutter compression, Gurobi MIP, MIPLIB, nanoGPT, Stockfish, and the matrix-multiplication exponent are flat across the study window.

The tentative mechanism: AI's lift appears to depend on which side of the training-data distribution a domain sits — where the corpus is well-represented in pre-training and answers are checkable, AI acts as a force multiplier; where the frontier is the model itself, the gains are zero. Jack Clark's Import AI 470 newsletter coined the term differential acceleration to capture this pattern, and the framing is the takeaway, not the cyber-versus-math headline.

Apply the test. Identify the field, ask how much of its working corpus lives in the next training run, and read METR's next update to see where the lift lands. Jack Clark's Import AI 470 framed this dynamic as a phase change — real and selective. The closer the work sits to the model, the less AI seems to do for it.