Exercises

Allocate a team of 10 researchers across five safety agendas, once per scenario, across four scenarios.

Respond to criticisms of AI control, then see one response defenders give. Ends by building your own argument.

Five deployment setups, the same question each time: assuming the untrusted model is scheming, does the setup hold the control property?

Two reasoning prompts on the game theory of control: why a catch changes the game, and where capability evaluations stop being informative.

Pick one area of control work and construct an explicit causal chain from work starting today to reduced existential risk.

Find the link in your theory-of-change chain you trust least, and name the observations that would raise or lower your confidence.

Reconstruct the password-locked model construction as three drag-and-drop flow charts: strong policy, weak policy, locked model.

Draft the terms of a deal a lab could offer a model it suspects of sabotage: task, compensation, delivery, and verification.