Industrial Policy for Advanced AI: Compute Pricing and the Safety Tax

TL;DR

This paper models how compute pricing influences AI safety investments and strategic behavior, revealing conditions under which subsidies and pricing policies can effectively reduce disaster risks in competitive AI development.

Contribution

It introduces a game-theoretic model analyzing the impact of compute pricing and subsidies on safety investments in AI development, providing new insights into policy design.

Findings

Compute price increases promote safety if performance scales faster than safety.

Subsidies to a single agent can arbitrarily reduce disaster probability.

Targeted subsidies to productive or safety-conscious agents often improve safety.

Abstract

Using a model in which agents compete to develop a potentially dangerous new technology (AI), we study how changes in the pricing of factors of production (computational resources) affect agents' strategies, particularly their spending on safety meant to reduce the danger from the new technology. In the model, agents split spending between safety and performance, with safety determining the probability of a ``disaster" outcome, and performance determining the agents' competitiveness relative to their peers. For given parameterizations, we determine the theoretically optimal spending strategies by numerically computing Nash equilibria. Using this approach we find that (1) in symmetric scenarios, compute price increases are safety-promoting if and only if the production of performance scales faster than the production of safety; (2) the probability of a disaster can be made arbitrarily low by providing a sufficiently large subsidy to a single agent; (3) when agents differ in productivity, providing a subsidy to the more productive agent is often better for aggregate safety than providing the same subsidy to other agent(s) (with some qualifications, which we discuss); (4) when one agent is much more safety-conscious, in the sense of believing that safety is more difficult to achieve, relative to his competitors, subsidizing that agent is typically better for aggregate safety than subsidizing its competitors; however, subsidizing an agent that is only somewhat more safety-conscious often decreases safety. Thus, although subsidizing a much more safety-conscious, or productive, agent often improves safety as intuition suggests, subsidizing a somewhat more safety-conscious or productive agent can often be harmful.