Import AI 464: Fables writes GPU kernels; AI automation; and analog computation

Details

Fable has topped the KernelBench-Mega leaderboard with an 18.71X speedup over an optimized PyTorch baseline on an RTX PRO 6000 Blackwell GPU. The key distinction: Fable's solution uses exactly one cooperative kernel launch per decoded token, while every competing entry used 4–14 separate launches. Claude Opus 4.8 achieved 14.4X (Triton), GLM-5.2 reached 11.14X, and GPT-5.5 got 4.34X. Kernel design is a foundational skill for AI R&D, making Fable's lead a meaningful indicator of AI systems' ability to improve their own infrastructure. The second major data point comes from CAIS and Scale Labs, who track the Remote Labor Index—a benchmark testing AI on real, paid freelance projects including 3D/CAD, graphic design, video, web apps, and data analysis. AI success rates rose from 2.5% at launch in October 2025 to 16.1% by July 2026. Fable 5 leads at 16.1%, Opus 4.8 at 8.3%, GPT-5.5 at 6.3%. The authors call it 'the frontier more than quadrupling in under eight months.' Jack Clark frames the kernel result as early evidence of a recursive self-improvement loop—AI systems becoming better at the tasks required to train and optimize AI systems. The RLI result matters on a different axis: it's one of the most structured attempts to measure actual AI displacement of paid labor, not synthetic benchmark performance.