Import AI 464: Fables writes GPU kernels; AI automation; and analog computation
Summary
The latest Import AI brief highlights significant advancements in AI capabilities across several domains. Fable, an AI system, demonstrated an 18.71X speedup on KernelBench-Mega by writing highly optimized Cuda GPU kernels, outperforming other models like Claude Opus 4.8 (14.4X) and GPT 5.5 (4.34X). This indicates AI's growing proficiency in fundamental R&D tasks. Concurrently, AI systems are rapidly improving at online freelance projects, with success rates on the Remote Labor Index (RLI) rising from 2.5% in October 2025 to 16.1% in July 2026, driven by models like Fable 5. Furthermore, the OSWORLD 2.0 benchmark reveals AI's progress in complex, multi-hour computer-use tasks, though current agents like Claude Opus 4.8 achieve only 20.6% accuracy. JD, China's e-commerce giant, also detailed its Oxygen AI Item Center, an LLM/VLM-centric solution managing tens of billions of SKUs on Huawei Ascend NPUs, showcasing AI's role in large-scale, self-updating business operations.
Key takeaway
For AI Engineers and Directors of AI/ML evaluating strategic investments, these advancements indicate a critical inflection point. You should prioritize exploring AI-driven automation for core R&D tasks like kernel optimization and for complex, multi-step computer operations, as these areas show rapid capability expansion. Additionally, consider how "person-light, AI-heavy" organizational models, exemplified by JD's Oxygen AIIC, could reshape your operational efficiency and competitive landscape, necessitating a re-evaluation of your long-term workforce planning and AI integration strategies.
Key insights
AI systems are rapidly advancing in core R&D, online labor automation, and complex computer interaction, signaling potential for recursive self-improvement and economic shifts.
Principles
- AI proficiency in kernel design accelerates R&D.
- Economically capable AI agents are advancing rapidly.
- Computer-use benchmarks proxy AI task competence.
Method
JD's Oxygen AIIC employs ontology engineering, "Semantic Search then Discrimination," self-evolving LLMs/VLMs with expert modules, and a "Unified item tunnel" for inventory management.
In practice
- Evaluate AI for GPU kernel optimization.
- Track AI agent performance on digital labor tasks.
- Assess AI for multi-program computer automation.
Topics
- GPU Kernel Optimization
- AI R&D Automation
- Digital Labor Automation
- Remote Labor Index
- OSWORLD 2.0
- Inventory Management AI
- Huawei Ascend NPUs
Code references
Best for: Machine Learning Engineer, Research Scientist, AI Scientist, AI Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Import AI.