A new METR study finds that AI is accelerating scientific and technical progress unevenly across fields, with cybersecurity showing a sharp jump in 2026, mathematics posting minor gains, and AI research itself showing no measurable AI-driven acceleration. The report, summarized in Jack Clark's Import AI 470, examined three domains and found what Clark calls "lumpy acceleration" — pockets of phase change rather than a uniform lift. Alongside METR's findings, two new research systems, SPADE and Hawkeye, sketch how that unevenness might start to shift.
METR's cyber finding is the sharpest data point. Vulnerability reports across cURL, OpenSSL, Firefox and Microsoft, as well as the aggregate US NVD and OSV databases, accelerated dramatically in 2026 compared with 2025. The pattern mirrors what happened to day-to-day coding in 2025: capability crossed a threshold, and output volume followed.
Mathematics saw a smaller, harder-to-quantify lift. arXiv submissions in some math subfields have doubled in less than 12 months, and AI-assisted work has helped chip away at prestigious open problems, including the Jacobian conjecture from Smale's list, Problem 44 from Green's list (the halving sieve), and the sofic half of Green's Problem 100. Whether that pace holds is an open question.
“The rate of vulnerabilities reported across many projects has dramatically accelerated in 2026 compared with 2025, both for specific projects (cURL, OpenSSL, Firefox, and Microsoft) and for aggregate vulnerability databases (the US NVD, and OSV)”— Jack Clark, author of Import AI, summarizing METR
Key facts
- 01METR finds cyber vulnerability reports accelerated sharply in 2026 vs 2025 across cURL, OpenSSL, Firefox, Microsoft, and the US NVD.
- 02arXiv submissions have doubled in some math areas in under 12 months, but METR sees no measurable AI acceleration in AI research itself.
- 03SPADE lifts Qwen3-30B-A3B to a 58.3 benchmark suite average, +8.1 over base and +5.3 over the best fixed-environment baseline.
- 04Hawkeye reaches an 18.9× geomean speedup against expert-authored Triton kernels on emerging attention variants.
- 05Hawkeye matches or beats hand-tuned kernels on NVIDIA Ampere, Hopper, Blackwell and AMD MI350, including 1.22× on Blackwell.
The null result is the most interesting one. Across seven canonical AI-optimization benchmarks — CIFAR-10, Hutter compression, Gurobi mixed-integer programming, MIPLIB, nanoGPT, Stockfish, and the matrix-multiplication exponent — METR found only scattered LLM-attributable contributions on nanoGPT and CIFAR-10, and no broad acceleration curve. The models are not yet meaningfully speeding up the field that produces the models.
That gap is the backdrop for SPADE, or Self-Play in Adaptive Synthetic Executable Environments, built by researchers at the University of Washington, Stanford University, Northeastern University, Carnegie Mellon University, MIT, the National University of Singapore, Seoul National University, Stevens Institute of Technology, and the University of Chicago. SPADE has an LLM alternate between two roles: an Environment Designer that writes executable training environments as Python programs with a Gym-style interface, and a Reasoning Agent that learns to solve them.
The Designer's reward is driven by a hint-based regret signal. A privileged hint is task-relevant information — a partial solution sketch or key structural observation — that the Designer can optionally attach to an environment. The gap between the Reasoning Agent's return with and without that hint defines whether the environment was well-calibrated: too easy or too hard, and the signal collapses.
SPADE was tested on three Qwen3 backbones: Qwen3-4B-Instruct-2507, Qwen3-8B, and Qwen3-30B-A3B-Instruct-2507, each tuned via GRPO across 400 rollouts of 25 environments. At the 30B scale, SPADE reached a suite average of 58.3, which is +8.1 over the base model and +5.3 over the strongest fixed-environment baseline. The same recipe applied to tool-use environment design improved every backbone tested.
“at 30B-A3B, SPADE reaches a suite average of 58.3: +8.1 over base and +5.3 over the strongest fixed-environment baseline”— SPADE research team, authors, University of Washington et al.
Hawkeye, from researchers at Harvard, Stanford, Together AI, and Caltech, tackles a different bottleneck: getting coding agents to write competitive GPU kernels without heavy expert intervention. The system ships as an open-source framework that pairs each unit test with a human-authored reference kernel and the profiling metric that verifies the optimization, giving an agent a syntax example, a callable function, and a target to beat.
The evaluation covered NVIDIA Ampere, Hopper, and Blackwell, plus AMD MI350, across BF16, FP8, NVFP4 and MXFP4 precisions. On established workloads where torch.compile dispatches to expert-tuned libraries like cuBLAS, cuDNN and FlashAttention, Hawkeye matched or exceeded those baselines, including in precision formats PyTorch cannot natively run. On emerging attention variants that torch.compile cannot fuse, Hawkeye posted an 18.9× geomean speedup against expert-authored Triton kernels from the Flash Linear Attention library, including 1.22× on Blackwell and 1.00× on MI350.
The caveat behind both systems is that neither escapes the imaginative ceiling of the base model. SPADE's authors are explicit that a self-play loop doesn't bootstrap a model far beyond what its generator can already envision, and Hawkeye's kernel gains depend on the human-curated taxonomy of unit tests that seeds the agent. METR's null result on AI-optimization benchmarks suggests the recursive-self-improvement loop is still gated by the same phase-change dynamics visible in every other field the study measured.
The through-line for AI companies is that the returns to elicitation infrastructure — synthetic environment generators, well-documented unit tests, hardware-aware toolchains — are compounding faster than the returns to raw scale in narrow domains. Frontier labs already spending on data centers and pretraining runs are going to spend an increasing share on Hawkeye-style scaffolding, because a 18.9× kernel speedup on a Blackwell fleet is a capex line item that pays for itself in weeks. The question METR leaves open is which field crosses the phase-change threshold next, and whether the next one is AI research itself.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
