Import AI 439: AI kernels; decentralized training; and universal representations
Decentralised AI training's rapid growth could reshape who controls frontier AI development—a geopolitical dynamic worth tracking for Australian AI strategy.
Key points
- Meta's KernelEvolve uses LLMs to auto-generate optimised AI kernels, cutting development time from weeks to hours.
- Epoch AI analysis shows decentralised AI training compute growing 20x per year, with major governance implications for who controls frontier AI.
- Both developments are primarily of academic and technical interest; limited direct APS operational relevance at this stage.
Summary
This edition of Import AI covers two research developments. First, Meta's KernelEvolve system demonstrates AI-driven automation of hardware kernel design, achieving significant infrastructure efficiencies at hyperscale and illustrating how AI is increasingly used to accelerate its own development pipeline. Second, an Epoch AI analysis of decentralised AI training finds compute scale growing at 20x per year—far outpacing frontier centralised training's 5x growth—though decentralised runs remain roughly 1000x smaller than frontier models. The newsletter frames decentralised training as a political technology that could broaden access to frontier-scale AI beyond the current small cluster of US and Chinese tech companies, with potential implications for academic, government, and non-commercial AI development globally.
Implications for Australian agencies
- Monitor DISR, AISI, and strategy teams may want to monitor decentralised AI training trends, as rapid compute growth could affect assumptions about who develops frontier-capable models and how Australian policy responds.
- Monitor Agencies tracking AI supply chain and infrastructure risk may want to watch whether AI-automated kernel development accelerates capability timelines in ways that affect existing risk assessments.
Implications are AI-generated. Starting points, not advice.
"Import AI 439: AI kernels; decentralized training; and universal representations" Source: Import AI – Substack (Jack Clark) Published: 5 January 2026 URL: https://importai.substack.com/p/import-ai-439-ai-kernels-decentralized This edition of Import AI covers two research developments. First, Meta's KernelEvolve system demonstrates AI-driven automation of hardware kernel design, achieving significant infrastructure efficiencies at hyperscale and illustrating how AI is increasingly used to accelerate its own development pipeline. Second, an Epoch AI analysis of decentralised AI training finds compute scale growing at 20x per year—far outpacing frontier centralised training's 5x growth—though decentralised runs remain roughly 1000x smaller than frontier models. The newsletter frames decentralised training as a political technology that could broaden access to frontier-scale AI beyond the current small cluster of US and Chinese tech companies, with potential implications for academic, government, and non-commercial AI development globally. Implications for Australian agencies: - [Monitor] DISR, AISI, and strategy teams may want to monitor decentralised AI training trends, as rapid compute growth could affect assumptions about who develops frontier-capable models and how Australian policy responds. - [Monitor] Agencies tracking AI supply chain and infrastructure risk may want to watch whether AI-automated kernel development accelerates capability timelines in ways that affect existing risk assessments. Retrieved from SIMS, 18 May 2026.