The Squeeze on Human Advantage

When we think about artificial intelligence disrupting the workforce, we usually picture it generating marketing copy or writing basic Python scripts. But recent data suggests a much deeper, dual-front shift is underway: AI is simultaneously learning to execute complex freelance contracts and rewriting the very hardware instructions that make it smart.

On the economic front, the Center for AI Safety (CAIS) and Scale Labs have been tracking a metric known as the Remote Labor Index (RLI). This index doesn't measure whether an AI can chat; it measures whether an AI can complete economically valuable, end-to-end projects online. The trajectory is striking. According to the index, the success rate of frontier AI models on these tasks jumped from just 2.5% in October 2025 to 16.1% by July 2026.

The tasks evaluated are remarkably concrete. They include taking a client’s existing 3D model of an engagement ring, swapping an emerald-cut stone for a marquise cut, and delivering photorealistic renders. They include producing a full 60-second animated 2D advertisement for a tree-care company, complete with voiceovers. They even include taking scanned cadastral plans and site photos to generate dimensioned floor plans and interior design renders. The top performer in recent tests, a model dubbed Fable 5, hit that 16.1% mark, significantly outpacing other frontier systems.

To measure this new level of competence, researchers from institutions like the University of Hong Kong, UCSD, and Columbia University recently released OSWORLD 2.0. This benchmark tests an AI's ability to navigate computers just like a human worker would—clicking, typing, and moving between multiple applications. The median task in this new benchmark takes a human 1.6 hours to complete, representing a massive leap from the 2-minute tasks of previous evaluations.

But perhaps the most profound development isn't happening on the surface of our screens, but deep within the silicon.

In a benchmark called KernelBench-Mega, an AI system (also named Fable) successfully wrote a highly complex GPU "megakernel." Writing Cuda code for an RTX PRO 6000 Blackwell, Fable achieved an astonishing 18.71X speedup compared to an optimized PyTorch baseline. What impressed engineers most was its efficiency: while other top-tier AI models had to break the problem down into 4 to 14 separate "kernel launches" per decoded token, Fable accomplished it in exactly one smooth, cooperative launch.

This is a critical milestone. Kernel design is a fundamental part of AI research and development. If an AI can autonomously write and optimize the very code that dictates how efficiently it runs on a microchip, it is taking the first tangible steps toward recursive self-improvement—building a better brain for itself.

The classic defense against automation anxiety is that humans will simply innovate, creating new jobs and finding new niches where we hold a comparative advantage. But as AI systems expand their capabilities at this breakneck pace—mastering both the digital tools we use for daily work and the fundamental engineering required to improve themselves—we have to ask: can human innovation outrun an intelligence that is learning to upgrade itself?

Sources: Import AI #464, referencing CAIS Remote Labor Index, KernelBench-Mega, and OSWORLD 2.0.

Key Points

• The Remote Labor Index shows AI success rates on complex freelance tasks rising from 2.5% to 16.1% in under a year.
• AI models are now successfully executing end-to-end projects like 3D jewelry redesigns and animated commercial production.
• The OSWORLD 2.0 benchmark demonstrates AI's growing ability to handle multi-step computer tasks that take humans over 1.5 hours.
• An AI named Fable successfully wrote highly optimized GPU kernel code, achieving an 18.71X speedup, hinting at AI's capacity for self-improvement.

Why It Matters

As AI proves capable of both automating complex commercial labor and optimizing its own fundamental architecture, the traditional economic assumption that humans will easily out-innovate automation is facing its toughest test yet.

Sources:

• Import AI 464: Fable writes GPU kernels; AI automation; and analog computation — Import AI (Jack Clark)