Konstantin Pilz describes research into the AI supply chainAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models. and estimates that leading Chinese AI companies have substantially less compute than US frontier developers. He presents large-scale distillationKnowledge distillation trains a smaller or different AI model to reproduce useful behavior learned from a stronger teacher model. as his explanation for how Chinese models can remain competitive despite that estimated gap.
The discussion considers how a pause or slowdown in frontier training could interact with continued compute production, including the risk of a large compute overhang. Pilz suggests directing some capacity to inferenceInference is the use of a trained model to produce an output from supplied input. as one possible policy approach.
Pilz discusses the role of deep ultraviolet lithography equipment and argues that export controls could affect China’s ability to manufacture AI chips. His assessments are presented as estimates and policy arguments, not established measurements.
The speakers also consider possible international agreementsGovernance sets responsibilities, rules and oversight for decisions about how AI systems are used., how verification might work, automation’s effects on human work and the need for more research into robotics supply chains.
Watch on YouTube
