The newest issue of Import AI opens with a data point that reads like a small technical footnote but points toward something larger: an AI model has produced what one benchmark maintainer called the first genuine, and fastest, megakernel ever submitted to KernelBench-Mega, a benchmark for GPU code optimization.
The model, Fable, achieved an 18.71x speedup writing CUDA code for an RTX PRO 6000 Blackwell GPU, measured against an optimized PyTorch baseline. For comparison, the newsletter cites other top attempts on the same benchmark: Claude Opus 4.8 reached 14.4x writing Triton code, GLM-5.2 hit 11.14x also in Triton, and GPT-5.5 managed 4.34x. What made Fable's entry stand out technically, according to profiling data cited in the piece, is that it launched exactly one cooperative kernel per decoded token, where every other high-scoring submission split the same work across four to fourteen separate kernel launches.
Import AI frames this as an early sign of AI systems automating tasks that are themselves fundamental to AI research and development, since kernel design and optimization sit squarely in the toolchain researchers use to make training and inference faster. The newsletter describes this as a potential early step in a recursive self-improvement loop, where AI-written tools make the next generation of AI development faster in turn.
A second item in the same issue points to a broader automation trend outside of low-level systems work. Researchers with the Center for AI Safety and Scale AI have been tracking the Remote Labor Index, which scores how well AI systems complete real, economically valuable freelance projects end-to-end, spanning categories such as 3D and CAD work, architecture, graphic design, video and animation, audio, data analysis, and web application development. According to the newsletter, success rates on this index rose from 2.5 percent when it launched in October 2025 to 16.1 percent by July 2026. A July update evaluating three recent frontier models found GPT-5.5 at 6.3 percent, Claude Opus 4.8 at 8.3 percent, and Fable at 16.1 percent, meaning the top end of the frontier has more than quadrupled in under eight months.
Neither result on its own proves systems are close to fully automating engineering or freelance labor. A kernel benchmark win is a narrow, well-specified task, and a 16 percent success rate on end-to-end freelance projects still means most attempts fail. But Import AI's read is that the direction and pace of both curves matter more than their current absolute levels, since both are measuring capability growth on tasks that were previously assumed to require substantial human judgment.