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

Import AI’s newsletter summarizes a METR study that analyzes the impact of large language models on scientific progress. The study finds that AI has contributed significantly to cyber vulnerability detection, with the rate of reported vulnerabilities accelerating dramatically in 2026 compared to 2025 across projects such as cURL, OpenSSL, Firefox, and Microsoft, as well as aggregate databases like the US NVD and OSV. In mathematics research, the study notes a minor acceleration, citing doubled arXiv submissions in some areas and the solving of several prestigious problems, including the Jacobian conjecture from Smale’s list and problems from Green’s list. However, the study reports no measurable acceleration in AI research itself, with only limited contributions in benchmarks such as nanoGPT and CIFAR-10. The newsletter also discusses SPADE, a framework that generates synthetic training environments using large language models. SPADE employs Qwen3-30B to create game and tool-use environments, achieving an average score of 58.3, an 8.1‑point improvement over the base model and a 5.3‑point improvement over the strongest fixed-environment baseline. Finally, the newsletter touches on philosophical concerns about AI rights and consciousness, noting that these discussions are becoming increasingly prominent as AI systems grow more pervasive in the economy. The piece concludes that differential acceleration across fields highlights the need for targeted research and policy efforts.