Glonce

Import AI 464: Fable writes a record GPU kernel as automation benchmarks jump

researchThe Glonce editorsImport AI🔥73Glonce rating73Impact70Novelty60Relevance90Surprise65Players78Credibility82

Import AI issue 464 groups four separate developments under one theme: AI systems getting better at automating the work that goes into building and running AI itself, and into running large parts of the economy.

First, the newsletter reports that Fable wrote what a KernelBench-Mega maintainer called the first genuine, and fastest, megakernel submitted to the benchmark, according to the official leaderboard. Writing CUDA code for an RTX PRO 6000 Blackwell GPU, Fable achieved an 18.71x speedup over an optimized PyTorch baseline. Other systems tested on the same benchmark, all writing Triton rather than CUDA, scored lower: Claude Opus 4.8 reached 14.4x, GLM-5.2 reached 11.14x, and GPT-5.5 reached 4.34x. The newsletter highlights that profiling showed Fable's kernel used exactly one cooperative kernel launch per decoded token, while every other high-scoring entry split the same work into four to fourteen separate kernel launches per token. Import AI frames kernel design as one of the fundamental tasks in AI research and development, so progress here is read as a signal about how close AI systems are getting to improving themselves.

Second, researchers from the Center for AI Safety (CAIS) and Scale Labs report a jump in AI performance on the Remote Labor Index (RLI), a test of how well AI systems complete 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 applications. Success rates rose from 2.5% when RLI launched in October 2025 to 16.1% in a July 2026 update. Testing three current frontier models on the same set of tasks, the authors found GPT-5.5 scored 6.3%, Opus 4.8 scored 8.3%, and Fable 5 scored 16.1%. The CAIS and Scale Labs authors describe the frontier as having more than quadrupled in under eight months. Newsletter author Jack Clark writes that he finds it increasingly hard to reconcile this pace of progress with the economy staying the same, and argues that person-light, AI-heavy organizations are likely to expand and out-compete unaugmented humans faster than humans can build new comparative advantages.

Third, a group of researchers spanning the University of Hong Kong, UC San Diego, Columbia University, UC Santa Barbara, Mila, Snorkel AI, the University of Wisconsin, Alibaba Qwen, Ohio State University, Simular, and NeoCognition released OSWorld 2.0, a benchmark for AI agents completing long, multi-step, multi-program tasks on computers. Its 108 tasks, including 31 self-hosted websites, are far harder than the original OSWorld: the median task takes a person about 1.6 hours, roughly 48 times longer than the 2-minute median task in OSWorld 1.0, and 69.6% of tasks are estimated to take a skilled human more than an hour. The benchmark ships with a much broader software set than its predecessor, including Slack, LinkedIn, GitLab, Overleaf, Zotero, AWS, and simulated portals for insurance claims, visa applications, and conference management. Current agents struggle: the strongest configuration tested, Claude Opus 4.8 with maximum thinking and batched tool calls, reached only 20.6% binary accuracy and 54.8% partial-score accuracy, with performance dropping sharply as tasks lengthen and agents needing to recover hidden state or resolve conflicting information. The newsletter notes that OSWorld 1.0 scores rose from about 30% in July 2025 to about 75% for MiniMax M3 by June 2026, and expects a similar climb on OSWorld 2.0.

Fourth, JD, described in the newsletter as the Amazon of China, published details of the Oxygen AI Item Center (Oxygen AIIC), the system it uses to manage inventory across a catalog of tens of billions of SKUs for 700 million users and millions of merchants. JD's paper states that Oxygen AIIC now covers tens of thousands of JD categories and processes hundreds of millions of item updates per day, running on Huawei Ascend NPUs. The system combines four elements: ontology engineering built through human-AI collaboration, where experts distill industry knowledge that algorithms use to scale ontology construction; a two-stage semantic search and discrimination process that lets the ontology update without retraining the model; self-evolving item-understanding LLMs and VLMs that use incremental learning to close knowledge gaps while limiting catastrophic forgetting; and a "unified item tunnel" that serves as the main interface to other business applications, supporting daily, minute, and second-level data pipelines. JD frames its reliance on Huawei Ascend NPUs as part of China's broader push for technology self-sufficiency in compute.

Key facts

  • Fable's CUDA kernel hit an 18.71x speedup on KernelBench-Mega, ahead of Claude Opus 4.8 (14.4x), GLM-5.2 (11.14x), and GPT-5.5 (4.34x), using one kernel launch per token versus four to fourteen for rivals
  • On the Remote Labor Index, AI success on end-to-end freelance tasks rose from 2.5% at launch (October 2025) to 16.1% by July 2026, with Fable 5 scoring 16.1%, Opus 4.8 scoring 8.3%, and GPT-5.5 scoring 6.3%
  • OSWorld 2.0's 108 long-horizon computer-use tasks take a median 1.6 hours of human time, 48 times the 2-minute median of OSWorld 1.0, and the best agent tested, Claude Opus 4.8, managed only 20.6% binary accuracy
  • JD's Oxygen AI Item Center manages tens of billions of SKUs for 700 million users, processing hundreds of millions of item updates daily on Huawei Ascend NPUs

Why it matters

Each item tracks a different rung of the same ladder: AI systems automating the technical work that builds AI (kernel design), automating paid knowledge work across many professions (RLI), and automating the multi-step computer use that connects software tools together in real workflows (OSWorld 2.0). Import AI's author reads the kernel and RLI results together as evidence that AI capability is compounding faster than humans can build new comparative advantages, since better AI at building AI feeds directly into further capability gains.

Who it affects

AI infrastructure teams and chip vendors watching kernel-optimization benchmarks; freelancers and clients in 3D and CAD, design, video, audio, data analysis, and web development, the categories RLI tests directly; agent developers being measured against OSWorld 2.0's long-horizon tasks; and large e-commerce operators facing the kind of catalog scale JD's Oxygen AIIC was built to handle.

How to use it

None of the four items describes a public product, price, or subscription tier; they are benchmark results and a technical paper. Readers who want to verify the claims can consult the KernelBench-Mega leaderboard, the CAIS report "A Significant Increase in Digital Labor Automation," and the OSWorld 2.0 project page and GitHub repository, all linked in the original newsletter.

How solid is it

The kernel result is attributed to a KernelBench-Mega maintainer and the benchmark's own leaderboard; the RLI figures come from a published CAIS and Scale Labs report; OSWorld 2.0 is a scholarly benchmark released by a multi-institution academic and industry team with a public paper and GitHub repo; the JD figures come from JD's own research paper describing its system, so they are self-reported by the company that built it.

Risks and caveats

The newsletter's framing, that rapid gains on these benchmarks point toward large-scale labor displacement, is the author's own interpretation and forecast, not a finding of the underlying papers. OSWorld 2.0's current scores remain low (around 20% binary accuracy for the best agent), so multi-hour computer-use automation is not yet reliable in practice. JD's inventory figures describe its own deployed system and have not been independently verified.

“I’m betting the other side: AI systems are expanding their economically relevant capabilities faster than humans are expanding their comparative advantages relative to AI systems.”

— Import AI 464