A lot of work on AI and international security focuses on the uncertainty about what other actors are capable of. But AI systems can be very strong at some tasks yet surprisingly weak at others, meaning that it may not always make sense to say that one system is "more capable" than another. This possibility seems much less explored in work on AI strategy and deterrence.

This project investigates whether this kind of jagged capability could create problems for AI-era bargaining and deterrence. If two actors have very different strengths, there may be no stable answer to the question of who is more capable overall. This could make it much harder for them to understand each other's position, communicate credible threats, predict how a crisis might develop, or judge the consequences of an escalation. The project will distinguish this from ordinary capability uncertainty, where there is a well-defined answer that is just hidden from the other side, which could be addressed through promoting greater transparency and better verification methods.

What students will work on

Students will start by studying the relevant work on AI capability jaggedness, crisis bargaining, deterrence, multidimensional signaling, and how we measure national power. From there, we'll work on formalizing the idea of jagged capability and building a model to understand how it could affect bargaining between two actors.

We'll also look at whether actors can still figure out who is ahead when their AIs are better than each other at completely different things, and whether this creates problems for bargaining.

Depending on how the project develops, we may also look at real benchmark data from frontier AI systems to see how jagged their capabilities are. We may also run small language-model wargaming experiment(s) where AI agents bargain under different capability conditions.

What I'm looking for

I'm mainly looking for students who are strong in Python and comfortable with mathematical or theoretical work. They should have a good understanding of machine learning and be interested in game theory, AI strategy, international security, and/or related areas. Students who have experience with mathematical modeling, simulations, ML research, and/or AI safety would be a strong fit.

The project is fairly open-ended, so students should also be comfortable dealing with questions where the answer is unknown and making research decisions without having every step specified in advance. Strong writing skills would be valuable, since the goal is to turn the results into a research paper and potentially a shorter public write-up potentially on LessWrong.

Theory of change

Measuring AI capabilities are likely to become increasingly important for international security and strategic competition, but we still mostly think about them as something that can be compared on a single scale. If AI systems are strong at very different things, this may become much harder. Two actors might agree that both systems are highly capable while disagreeing about which one has the advantage in a particular situation.

Understanding this could help us figure out whether jagged AI capabilities create a new source of instability in AI-era bargaining, and what actors could do about it. If we found this to be real, we may need better ways of measuring and communicating AI capabilities, as well as changes to how we think about deterrence, verification, and strategic communication. If it turns out existing approaches are good enough, that would also be a useful result.

Expected output

By the end of the program, students should have a formalized definition of jagged capability, a working model or simulation of its effects on bargaining, and an analysis of the conditions under which it does or does not create additional instability. Where feasible, they will also have empirical analysis of real AI capability profiles or results from small language-model bargaining experiment(s).

The main output will be a research paper and/or a LessWrong article developing and testing the core argument, with the possibility of turning the work into a longer project or submitting it to a relevant AI safety, game theory, or security research venue if the results are promising.