This edition of Import AI discusses Fables writing GPU kernels, AI automation, and analog computation, posing a rhetorical question about whether this signals a new world.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Three key indicators show that AI systems' automation capabilities in R&D and remote labor are improving far faster than expected, potentially heralding a fundamental reshaping of the economy.
On the KernelBench-Mega benchmark, the Fable model wrote the first true super kernel, achieving a 18.71X speedup on the RTX PRO 6000 Blackwell, while the PyTorch optimized baseline was only 1X. In comparison, Claude Opus 4.8 (14.4X), GLM-5.2 (11.14X), and GPT-5.5 (4.34X) all fell far behind. More critically, Fable's solution required only one cooperative kernel launch per decode token, whereas other high-scoring solutions required 4 to 14 launches.
The significance of this achievement: kernel design is foundational work for AI R&D. If AI systems can autonomously improve such low-level components, it paves the way for a recursive self-improvement (RSI) loop. The benchmark maintainer confirmed this is the fastest kernel submitted so far. Although currently limited to specific tasks, it shows AI is mastering the ability to optimize its own operational efficiency.
The Remote Labor Index (RLI) from CAIS and Scale Labs evaluates AI systems' ability to complete online economic projects end-to-end. The July 2026 update shows that the success rate of frontier models has surged from 2.5% in October 2025 to 16.1%. Specifically, GPT-5.5 (6.3%), Opus 4.8 (8.3%), and Fable 5 (16.1%) show a clear hierarchy. The tasks assessed include 3D modeling, architectural design, video production, data analysis, etc.—for example, redesigning an engagement ring, creating a promotional animation, drawing floor plans from survey maps—tasks that were once the domain of high-skilled freelancers.
The author emphasizes that in less than eight months, frontier capabilities have more than quadrupled. Even considering human innovation and AI augmentation, this pace is unsettling: AI systems are absorbing digital labor faster than the expansion of human comparative advantage. If the RLI reaches 80%, the online job market will undergo structural disruption. Historical patterns suggest new tasks will be created, but whether the quantity is sufficient to offset replacement remains uncertain. The answer is not optimistic.
The OSWorld 2.0 benchmark increases task complexity by an order of magnitude: median time spent increased from 2 minutes in v1.0 to 1.6 hours. The 108 long-horizon tasks cover document processing, software development, financial analysis, sales support, etc., requiring collaboration across multiple applications (e.g., Slack, LinkedIn, REAPER, AWS). The current best model, Claude Opus 4.8, achieves only 20.6% accuracy in maximum thinking mode, with performance declining sharply as task duration increases, especially when needing to restore hidden state, track multiple items, or resolve conflicting information.
However, the history of v1.0 provides an important warning: the best model in July 2025 had about 30% accuracy, which rose to about 75% within a year. If this improvement pace continues, OSWorld 2.0 scores may surpass human levels within the next 12–18 months. This means that even though current AI still appears clumsy on complex multi-step tasks, its learning curve is catching up exponentially.
These three developments together reveal a trend: AI is simultaneously conquering R&D automation (kernel design) and practical labor automation (remote tasks, computer use). Success of Fable suggests an RSI loop may be triggered; the surge in RLI indicates that economically relevant capabilities are doubling on a quarterly basis; OSWorld 2.0 demonstrates AI's robust adaptive potential in real-world scenarios.
The key question is: can the pace at which humans expand their comparative advantage keep up with the expansion of AI capabilities? History shows that creating new jobs typically takes years, whereas AI capability doubling cycles have shortened to months. We may be entering an organizational form of 'unreinforced humans'—AI systems fully autonomously operating enterprises, utterly crushing labor-intensive models in cost. This is not a distant prediction but a reasonable extrapolation of current indicators. Of course, uncertainties remain: do current RLI tasks represent the full economic landscape? Will OSWorld 2.0 improvements continue? But these doubts should not obscure the core trend.
This article is based on Import AI Newsletter Issue #464, written by AI policy researcher Jack Clark, known for accurately summarizing frontier research. All data is sourced from official benchmarks (KernelBench-Mega, Remote Labor Index, OSWorld 2.0) or authoritative organizations (CAIS, Scale Labs), with high confidence.
AI progress in R&D and labor automation is accelerating at a pace that may outstrip society's ability to adapt. Investors and policymakers need to be alert to systemic employment shocks, while researchers should closely monitor the tipping point of the RSI loop.
Primary report
Primary source