Verifying Frontier AI: What Can we Prove About Compute?
Join BlueDot Chicago for our October Compute Verification Discussion, the first in a three-part series taking place in October, November, and December.
How can we verify what happens inside an AI datacenter while protecting confidential information? Join Paul Wang for an introduction to compute verification and its role in making AI governance commitments credible.
We’ll explore how researchers approach distinguishing training from inference, checking compute limits, and establishing a model’s provenance. Drawing on hardware monitoring, cryptographic methods, and emerging approaches to AI training, the discussion will examine what these tools could establish, the assumptions they rely on, and the open research questions that stand between promising proposals and practical deployment.
Dr. Paul Wang is a postdoctoral researcher at Sorbonne University’s LIP6 laboratory and a contributor to GPAI Policy Lab’s verification research programme. His work explores how cryptographic methods can verify frontier AI training while protecting confidential information, connecting technical research with the practical challenges of making AI governance commitments verifiable.
Suggested reading before the discussion
The Compute Verification Post — An accessible overview of compute verification, technical approaches, and open questions.
Weight Smuggling Likely Defeats Attempts to Cap FLOPs per Training Run — A short post examining potential weaknesses in limits on compute per training run.
Evolution Strategies at the Hyperscale — A research paper introducing EGGROLL, a training method relevant to the discussion of distinguishing training from inference.
Please read the first two posts and skim the abstract, introduction, and conclusion of the third. No detailed mathematical preparation is required.