AI Workload Verification

A structured reading course on AI workload verification and assurance, the field working on mechanisms that would let labs and states make agreements about AI development and check that no one is secretly defecting. By the end, our hope is you will be able to place different verification papers and setups by the claim they check, the evidence they use, and the institution that would act on it, and to have a project sketch of your own.

4units
3-5ha week
Syllabus

Course content

1 What is verification trying to do? 8 core + 8 optional
Read 1.1-1.4The risks Read 1.5-1.9What verification is for Read 1.10-1.13The research agendas Read 1.14-1.16Further agenda comparisons
2 Mechanisms by which it may be achieved 10 core readings, about five hours
Read 2.1System overview Read 2.2-2.3Hardware-rooted Read 2.4Inference verification Read 2.5-2.6Zero-knowledge proofs Read 2.7-2.8Telemetry and detection Read 2.9-2.10Attestation and audit
3 Dive deeper 1 primary track, three to six readings
Exercise100 experts Read 3.1-3.4Hardware-rooted Read 3.5-3.7Inference verification Read 3.8-3.11Zero-knowledge proofs Read 3.12-3.17Telemetry and detection Read 3.18-3.20Attestation and audit
4 Project ideation 2 readings + your sketch, due 12 Aug
Read 4.1Theory of victory Read 4.2Research mindsets WatchTheory of change video ExerciseWorksheet PromptsScoping a project Due 12 AugSubmit your sketch
Give us feedback on this course

Many of these readings are drawn from Will Hodgkins’s public reading list.