Import AI 470: No rights for machines; automating environment generation with SPADE; and building better GPU kernels with Hawkeye

Import AI this week circles around a simple idea: AI isn’t accelerating everything equally. A METR study breaks the picture into three areas — cyber, math, and AI research — and the results are uneven. Cyber looks like the big winner. Math shows some lift. AI research itself is much harder to pin down.

On cybersecurity, the paper says the rate of reported vulnerabilities has sharply increased in 2026 versus 2025, both in named projects like cURL, OpenSSL, Firefox, and Microsoft, and in broader databases like the US NVD and OSV. That is the kind of signal you can actually feel in the ecosystem. More bugs, more reports, more action.

Math is murkier. The study points to heavier arXiv activity in some areas, with submissions doubling in less than 12 months, and notes a few headline problems that have been solved, including the Jacobian conjecture from Smale’s list, Problem 44 from Green’s list, and the sofic half of Green’s Problem 100. But the authors are careful here: more papers does not cleanly equal more discovery, and the trend may still be too early to judge.

AI research and algorithmic progress are the least clear case. Across seven benchmark areas — CIFAR-10, Hutter compression, Gurobi mixed-integer programming, MIPLIB, nanoGPT, Stockfish, and the matrix-multiplication exponent — there are some LLM-attributable contributions, especially for nanoGPT and CIFAR-10. But the overall rate of increase is nowhere near what the study sees in cyber or math.

That theme of uneven progress carries into SPADE, a multi-university system for generating synthetic executable environments through self-play. It lets one model design training environments while another learns to solve them, using a reward based on the gap with and without privileged hints. The authors test it with Qwen3 backbones up to 30B, and the results improve both game environments and tool-use environments. In the 30B case, the suite average reaches 58.3, which is 8.1 points over base and 5.3 over the strongest fixed-environment baseline.

Then there’s Hawkeye, from Harvard, Stanford, Together AI, and Caltech, which tries to make coding agents better at writing GPU kernels for specific hardware. The trick is a minimal taxonomy of unit tests, each paired with a human-written solution kernel and profiling metric. On established workloads, it matches or beats torch.compile; on harder attention variants it reaches an 18.9× geomean speedup against expert-authored Triton kernels from Flash Linear Attention. That’s a neat reminder that AI progress often comes from the dull scaffolding around the model, not just the model itself.