In this module, you will tackle a comprehensive technical verification system proposal, synthesizing the verification mechanisms you learned in Module 2, actors and incentive structures you learned in Module 1, and verification intuitions you gained in Module 0.
By the end of this module, you will be able to:
- Explain how the requirements of an international AI agreement might be translated into a technical verification system.
- Understand and critique the assumptions, design choices, and unresolved problems involved in verifying AI compute use under conditions of limited trust between the parties.
- Apply an existing compute verification regime to a realistic facility simulation, identifying weak points, evasion possibilities, and testable assumptions.
Assignment
Read Naci Cankaya’s working paper A System Overview for Near-Term, Low-Trust AI Compute Verification.
Answer Questions 1, 2 and 5 together with one of Questions 3, 4, 7 or 9. Support each answer with page or section references to the paper.
Question 8 is an optional research task and may require consultation of an external technical source.
In your answers:
- distinguish the author’s claims from your own conclusions.
- reconstruct arguments by identifying their premises and conclusions.
- state explicitly when the available evidence does not justify a definite conclusion.
- formulate counterarguments in their strongest plausible form.
Questions
Written output · 3.x
Written Output
The track’s module anatomy gives every module one written output. Module 3 is the one module for which the outline never says which. Draft into it anyway if you like — the brief will appear here once it exists.
- Budget
- about 800 words