Import AI 464 reports two thresholds crossed in seven days: a first of its kind 'megakernel' (a single fused GPU compute routine) on the KernelBench Mega AI kernel leaderboard, and a Remote Labor Index climb from 2.5% to 16.1% on real freelance work.
Fable wrote a GPU kernel that runs 18.71 times faster than the optimized PyTorch baseline on the same hardware. The same week, the CAIS/Scale Remote Labor Index crossed 16.1 percent on real paid freelance tasks. Both numbers landed in Import AI issue 464, Jack Clark's weekly research digest. Read together, they describe a single transition: AI is authoring the low-level compute infrastructure future AI systems depend on, while clearing a measurable slice of paid skilled work in parallel.
The Fable submission is the headline. The benchmark is KernelBench-Mega, a leaderboard for AI-generated GPU kernels: low-level code, typically written in CUDA or Triton, that controls how a graphics processor runs a specific computation. Kernel design is a foundational task inside AI research and development, because the speed of any neural network is bounded by how well its kernels exploit the underlying hardware. Per Import AI, KernelBench-Mega is a meaningful signal of "recursive self-improvement": the better AI gets at writing kernels, the faster it can run experiments, train models, and ship the next generation of systems.
The Fable run is unusual not for the speedup alone but for the architecture. Per benchmark maintainer Elliot Arledge's analysis and follow-on commentary from Scaling01, Fable's solution launched exactly one cooperative kernel per decoded token, where every other high-scoring entry decomposed the same workload into four to fourteen separate kernel launches per token. On a Kimi-Linear W4A16 batch-1 decode task targeting an RTX PRO 6000 Blackwell, that architectural choice translated into an 18.71× speedup over the optimized PyTorch baseline. Comparison attempts landed at 14.4× for Claude Opus 4.8 in Triton, 11.14× for GLM-5.2 in Triton, and 4.34× for GPT 5.5 in Triton, all as reported in Import AI 464. Arledge called it "the first genuine (and fastest) megakernel ever submitted to KernelBench-Mega." The official leaderboard corroborates the ranking, and the underlying benchmark code lives in the Infatoshi kernelbench repository.
The second number is an economic one. The Remote Labor Index, a joint project of the Center for AI Safety and Scale Labs, measures how often current AI systems can complete real online freelance projects end to end. The metric climbed from 2.5 percent in October 2025 to 16.1 percent in July 2026, a 6.4× rise over nine months. Import AI describes the curve as a contest between AI capability expansion and human comparative advantage, and treats the index as one of the cleanest public measurements of AI moving from assisted to autonomous work on revenue-bearing tasks.
The two numbers are easy to file under separate desks: a benchmark story for the AI engineering beat, a labor story for the economics beat. Import AI treats them as one event: kernel authorship is one of the foundational tasks inside AI R&D itself, so the Fable result and the RLI climb draw from the same capability curve. Given that both thresholds crossed in the same seven-day window, the Fable megakernel matters because it ties AI's first-of-kind systems work to the curve that just cleared 16 percent of paid freelance contracts.
The 18.71× speedup is one benchmark task on one Blackwell GPU, not a general claim about AI replacing kernel engineers. The RLI number is a researcher-issued metric with its own methodology choices, and the comparison models in the Fable story are described in Import AI's own coverage rather than independently re-benchmarked across systems. The "first genuine megakernel" attribution traces to one maintainer, Arledge, and one newsletter, not a peer-reviewed verdict.
The simultaneity still establishes something concrete. The expansion that lets an AI top the KernelBench-Mega leaderboard is now scoring 16 percent on paid freelance work. The wire will report both numbers; the question worth tracking is whether next quarter's KernelBench-Mega leaderboard and next quarter's RLI print keep climbing in step.