Import AI 439: AI kernels; decentralized training; and universal representations

5 Jan 2026 · Import AI – Substack (Jack Clark) Global

Decentralised AI training's rapid growth could reshape who controls frontier AI development—a geopolitical dynamic worth tracking for Australian AI strategy.

Key points

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

Implications are AI-generated. Starting points, not advice.