Fable Writes GPU Cores, AI Automation and Analog Computing: Import AI 464 Digest

Fable has written the first true mega-kernel for KernelBench-Mega, achieving an 18.71x speedup on the RTX PRO 6000 Blackwell compared to the PyTorch baseline. This is the best result in the benchmark, while Claude Opus 4.8 shows 14.4x, GLM-5.2 shows 11.14x, and GPT 5.5 shows 4.34x. Fable's solution uses a single cooperative kernel launch per token, unlike the 4-14 launches by competitors. According to the Remote Labor Index, the success rate of AI systems on online freelancing has risen from 2.5% to 16.1% over 8 months; the best models as of July 2026 are GPT-5.5 (6.3%), Opus 4.8 (8.3%), and Fable 5 (16.1%). The evaluation included tasks in 3D, design, video, audio, data analysis, and web applications. OSWORLD 2.0 contains 108 long-duration tasks (median 1.6 hours) using 31 websites and software (Slack, LinkedIn, REAPER, AWS, etc.). The best agent, Claude Opus 4.8, achieved 20.6% accuracy, which is far from human level. JD has introduced the Oxygen AI Item Center for inventory management (700 million users, billions of SKUs). The system uses ontological engineering, semantic search, self-learning LLM/VLM, and a unified data tunnel. Everything runs on Huawei Ascend NPUs.