Import AI 439: AI kernels; decentralized training; and universal representations
Decentralised AI training's rapid growth challenges the assumption that frontier AI remains concentrated in a handful of US companies—relevant to sovereign AI strategy thinking.
Key points
- Meta's KernelEvolve uses LLMs to automate AI kernel design, cutting development time from weeks to hours.
- Epoch AI analysis finds decentralised AI training growing at 20x per year, raising governance and sovereignty implications.
- Item is a technical research newsletter; policy implications are present but require significant extrapolation for APS use.
Implications for Australian agencies
- Monitor Strategy and policy teams tracking sovereign AI capability may want to monitor the Epoch AI decentralised training analysis for implications on who can develop competitive AI outside major tech companies.
- Monitor Agencies following AI self-improvement risk may want to watch PostTrainBench-style evaluations as a leading indicator of when AI systems can autonomously conduct AI research.
Implications are AI-generated. Starting points, not advice — see methodology for how they're framed.
View original source
Copied.
Appeared in:
Weekly digest, 5 January 2026
"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 issue of Import AI covers four research developments. Meta's KernelEvolve system uses LLMs to automate low-level AI kernel design at scale, achieving dramatic inference speedups across heterogeneous hardware. An Epoch AI analysis of 100+ papers finds decentralised AI training runs are growing compute at 20x per year—far faster than frontier centralised runs—though still roughly 1,000x smaller, with implications for who can access frontier-class AI. The University of Tübingen's PostTrainBench tests whether frontier LLMs can autonomously fine-tune other models, finding current systems approach but do not match human researcher performance. MIT research shows that as AI scientific models scale, their internal representations of physical reality converge toward a common structure regardless of architecture or training domain.
Implications for Australian agencies:
- [Monitor] Strategy and policy teams tracking sovereign AI capability may want to monitor the Epoch AI decentralised training analysis for implications on who can develop competitive AI outside major tech companies.
- [Monitor] Agencies following AI self-improvement risk may want to watch PostTrainBench-style evaluations as a leading indicator of when AI systems can autonomously conduct AI research.
Retrieved from SIMS, 18 July 2026.