Power, Windows, and Dependencies

Exercise: Map it out

Contested Questions in AI Governance

How should we handle open weights as models get more capable?

When, if ever, should frontier development be slowed - and by whom?

Finding Your Leverage Point

More Career Resources

Next Steps: Fellowships

Next steps: Programs

State of Play at the Frontier

Independent Evaluations

The Hugging Face Incident and the System Card

Exercise: The Briefing

Grappling with the Intelligence Explosion

Exercise: Governance Under Acceleration

Visions of Victory

Face off: Plan S vs Plan A

Argue it out (with yourself)

Now the stress test: what happens to the strategies you compared in Unit 3 if AI itself substantially accelerates research and technological change?

Forethought sets out the scenario. CSET gives you competing models of what research automation could look like; OpenAI's account of its own internal workflows is the closest thing we have to evidence. More code, more experiments, and more agents doesn't automatically mean proportionately faster progress - look for where humans and compute still bottleneck, how the tasks were chosen, and whether capability research and safety research would accelerate equally.

Most of what you've read assumes the problem is control: who does what with AI. Nielsen asks whether that's even the right frame. Superintelligence Strategy asks what survives if states race regardless. Ngo's efficiency-vs-robustness frame asks what the reflex to centralize would cost.


**Scenario assumption:** An intelligence explosion has begun. AI capabilities are accelerating faster than institutions can track, competitive dynamics between states and labs are intensifying, and power is concentrating.

Pick the governance approach you have found most compelling across Units 1–3: Plan S, Plan A, Radical Optionality, another IAPS strategic vision, a specific mechanism, or a position you have been developing.

Walk through what breaks in this world. Identify the first point of failure and the mechanism by which it fails.

Then try to fix it. What design choices would help your proposal survive: adaptive mechanisms, sunset clauses, distribution of power, or triggers for escalation? Specify the relevant authority, information, and capacity. You are identifying the features a framework needs, rather than drafting legislation.

Look at your redesign through Ngo’s lens. Does it centralize power in ways that create new risks? If you strengthen oversight, who oversees the overseers? If you distribute power, can the system still coordinate against catastrophic risks?

Use the CSET/OpenAI comparison to state your assumptions about research acceleration. Which bottlenecks remain, and what would you change if progress proved slower, faster, or more uneven?

Some approaches and tools you have encountered:
- **Unit 1:** evaluations and their limits; developer disclosures and deployment controls; independent incident investigation and evidence gaps.
- **Unit 2:** international coordination; frontier safety policies; reporting duties, investigative authority, and institutional capacity; China’s approach and UK leadership.
- **Unit 3:** Plan S and Plan A; Radical Optionality; the IAPS strategic visions and CSET foundational assumptions.
- **Optional routes:** compute governance; private governance, auditing, and liability; constitutional constraints on government control; defensive AI deployment.


Finally, produce a one-page brief titled "What [your chosen proposal] needs to become to survive the intelligence explosion - and what that costs." 


One-page brief

Frontier AI Governance

Race Conditions, Power Concentration and Governance in the Limit

The Proposal Landscape

Frontier AI Governance

Exercise: Governance Under Acceleration

Resources (15 mins)

Exercises