Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity

URL SCAN: Risk Ceilings and Development Deadlines: Pacing AI under Uncertain Safety Productivity
FIRST LINE: # Economics > General Economics

The Dissection

This paper converts AI governance into a constrained scheduling problem: accumulate enough safety knowledge, reserve compute for safety work, then replay development under a risk ceiling and deadline. Its strongest result is conditional: if safety research is not too weak, knowledge transfers perfectly, and the regulator can control compute, pacing approaches the theoretical minimum productivity cost. The claimed 3.3% margin is therefore an internal calibration result, not evidence of systemic control.

The paper is really doing transition management. It makes an unstable political and economic transformation appear governable through sequencing, pricing, and technical guarantees.

The Core Fallacy

The central error, under the Discontinuity Thesis, is confusing local control of a development pipeline with control of the system being transformed.

A risk ceiling may constrain a defined hazard. It does not prevent AI from severing the mass employment → wage → consumption circuit. It does not restore productive participation. It does not stop competitive actors from defecting, accelerating, or shifting development beyond the regulator’s jurisdiction.

The paper also treats coordination as an available institutional input. Under P2, that is precisely the disputed premise. Compute allocation rules, enforced deadlines, reliable hazard estimates, and full safety-knowledge transfer are not neutral technical facts. They are political and competitive capacities that deteriorate as AI capability becomes economically and strategically decisive.

Hidden Assumptions

• Regulators can observe, allocate, and restrict compute across relevant actors.
• Developers accept deadlines and do not defect or relocate.
• The final model’s unprotected hazard is measurable enough to define a ceiling.
• Safety productivity is sufficiently stable to support planning.
• Safety knowledge transfers fully across development stages and model generations.
• Risk can be represented mainly as state risk rather than systemic, strategic, labor, or coordination risk.
• Development can be paused and replayed without losing irreversible competitive advantage.
• Customers can and will pay for the guarantee at a scale that sustains the regime.
• The regulatory system can enforce a common standard while capabilities continue to diffuse.
• Delaying capability deployment does not itself create unacceptable strategic or economic penalties.

These assumptions turn the hardest problems into parameters. That is mathematically clean and historically convenient.

Social Function

Primarily transition management and elite self-exoneration, with a partial truth.

The partial truth is real: sequencing safety research before capability scaling can improve risk management under favorable conditions. But the framing lets institutions claim that AI’s consequences are a pacing problem rather than a power-transfer problem. It offers a technical receipt for continued development: buy the guarantee, reserve the compute, declare the deadline.

Its customer-service framing is especially revealing. Safety becomes a premium product sold to actors who can afford protection, while the broader population remains exposed to the economic displacement that the model does not measure.

The Verdict

This is a useful conditional control model and a weak systemic defense. It may delay specific hazards, create temporary demand for safety researchers, regulators, auditors, and verification intermediaries, and generate Servitor niches around the AI owners. It does not defeat P1, P2, or P3.

The paper can pace deployment. It cannot preserve the post-WWII economic order once AI removes the need for mass human labor. Its guarantees depend on the very coordination, enforcement, and stable measurement capacities that competitive AI deployment is likely to erode. The result is not a reversal of obsolescence. It is a more orderly way to manage the approach of the wreck.