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

August 24, 2026 | Source: Import AI (Jack Clark)

Tags: METR, AI acceleration, cybersecurity, SPADE, GPU kernels, Hawkeye, Jack Clark, Import AI

Import AI #470 centers on a METR study finding that AI has dramatically accelerated cybersecurity vulnerability discovery in 2026—CVE rates spiking for cURL, OpenSSL, Firefox, and Microsoft—while contributing only modestly to mathematics and showing no measurable acceleration in AI research itself. Also covered: SPADE for auto-generating RL training environments and Hawkeye for LLM-assisted GPU kernel optimization.

Details

This edition of Jack Clark's Import AI newsletter leads with a METR analysis of where AI is actually compounding scientific and technical progress. The findings are field-specific rather than uniform. Cybersecurity sees major acceleration: reported vulnerabilities in 2026 have increased dramatically year-over-year for specific high-profile projects (cURL, OpenSSL, Firefox, Microsoft) and across aggregate databases including the US NVD and OSV. This is the clearest documented case of AI creating a meaningful inflection in a technical field. Mathematics sees minor acceleration: arXiv submissions have roughly doubled in some subfields, and several problems from prestigious open lists have been resolved (the Jacobian conjecture from Smale's list, Problem 44 from Green's list, the sofic half of Green's Problem 100). But attribution remains uncertain and the sustained trend is unclear. AI research itself shows no measurable self-acceleration across seven tracked problem areas including CIFAR-10, nanoGPT speed, and Gurobi MIP, with only scattered attributable contributions. Clark frames this as differential acceleration—the idea that AI crosses a phase threshold field by field, with coding (2025) and cybersecurity (2026) already past that point. Other areas remain uncertain. Also in this issue: SPADE, a multi-university system for automatically generating diverse executable training environments via self-play (a crude but scalable form of recursive data diversity), and Hawkeye, which applies LLM-assisted search to GPU kernel optimization.