Keep pulling the thread on Jack Clark.
Fable has written the first and fastest megakernel ever submitted to the KernelBench-Mega benchmark.
Fable achieved an 18.71X speedup by writing Cuda code on an NVIDIA RTX PRO 6000 Blackwell GPU, compared against an optimized PyTorch baseline.
The success rate of AI systems on the Remote Labor Index increased from 2.5% in October 2025 to 16.1% in July 2026.
On the Remote Labor Index, recent frontier models GPT-5.5, Opus 4.8, and Fable 5 achieved success rates of 6.3%, 8.3%, and 16.1% respectively.
The strongest tested agent on the OSWORLD 2.0 benchmark, Claude Opus 4.8, achieved only 20.6% binary accuracy and 54.8% partial-score accuracy.
JD's Oxygen AI Item Center covers tens of thousands of categories and processes hundreds of millions of item updates daily using Huawei Ascend NPUs.
On the KernelBench-Mega benchmark, Fable's 18.71X speedup writing Cuda code surpassed other models writing Triton code, including Claude Opus 4.8 (14.4X), GLM-5.2 (11.14X), and GPT 5.5 (4.34X).
Fable's solution for the KernelBench-Mega benchmark uses exactly one cooperative kernel launch per decoded token.
Other high-scoring entries on the KernelBench-Mega benchmark decompose the problem into 4 to 14 separate kernel launches per token.
Benchmarks like KernelBench-Mega are a meaningful signal of how effective AI systems are becoming at building themselves.
The median task in the OSWORLD 2.0 benchmark takes a human approximately 1.6 hours to complete, which is about 48 times longer than the 2-minute median task time in OSWORLD 1.0.
The OSWORLD 2.0 benchmark expands its supported software set to include applications like Slack, LinkedIn, Shortcut, REAPER, MuseScore, WPS, GitLab, Overleaf, LabPlot, Zotero, and AWS.