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

A newsletter roundup describes an AI system called Fable writing a GPU megakernel for KernelBench-Mega, along with new benchmark results on AI labor automation and computer use, and an industrial inventory system from JD.

Fable submitted what a benchmark maintainer and the official leaderboard describe as the first genuine and fastest megakernel on KernelBench-Mega, achieving an 18.71X speedup with CUDA code on an RTX PRO 6000 Blackwell against an optimized PyTorch baseline. For comparison, the newsletter lists Claude Opus 4.8 at 14.4X, GLM-5.2 at 11.14X, and GPT 5.5 at 4.34X, all writing Triton. It reports that torch.profiler shows exactly one cooperative kernel launch per decoded token, while other high-scoring entries used 4 to 14 kernel launches per token.

Researchers with the Center for AI Safety and Scale Labs reported a rise in AI success on the Remote Labor Index from 2.5% at launch in October 2025 to 16.1% in July 2026. A July update evaluated GPT-5.5, Opus 4.8, and Fable 5, which scored 6.3%, 8.3%, and 16.1%. The index assesses end-to-end online projects including 3D and CAD, architecture, graphic design, video and animation, audio, data analysis, and web applications.

OSWorld 2.0 was released by researchers from the University of Hong Kong, UC San Diego, Columbia, UC Santa Barbara, Mila, Snorkel AI, the University of Wisconsin, Alibaba Qwen, Ohio State, Simular, and NeoCognition. It contains 108 long-horizon tasks and 31 self-hosted websites, with a median task estimated at about 1.6 hours for a person, versus roughly 2 minutes in OSWorld 1.0. The strongest setting, Claude Opus 4.8 with maximum thinking and batched tool calls, reached 20.6% binary accuracy and 54.8% partial-score accuracy.

JD described its Oxygen AI Item Center, covering tens of thousands of categories and processing hundreds of millions of item updates per day on Huawei Ascend NPUs. The newsletter also includes a fictional installment of Tech Tales about analog computation.