Aydin Mohseni
Portrait of Aydin Mohseni

Mathematical philosopher, metascientist, AI safety researcher

I’m an assistant professor in the Department of Philosophy at Carnegie Mellon University, a core member of CMU’s Center for Conceptual Research on Safe AI and Institute for Complex Social Dynamics, and co-director of ∃xProof AI with Daniel Herrmann. I use evolutionary game theory and Bayesian epistemology to study problems in AI safety and metascience.

AI safety · Metascience · Methods · All research · CV

Dynamics

Chaos on the simplex

Figure 1The replicator dynamics for a four-strategy game. Each point of the tetrahedron is a population state. Started near the center, the orbit never settles: it folds into a strange attractor, and nearby starts drift apart. Drag to rotate.

AI safety

AI agent foundations, value learning and alignment, automated research

My work on agency focuses especially on artificial intelligence: the foundations of AI safety, agency, and alignment. With Daniel Herrmann I direct xProof AI, an AI safety research group that uses decision and game theory, mechanistic interpretability, and philosophical analysis to understand advanced AI systems as agents.

Figure 2A learner’s hypotheses about a coin’s chance of heads, each cell as tall as its probability, with a catch-all for the values it has not named. When the catch-all grows probable enough, it splits into new named hypotheses.
Selected work
xProof AI The research group I co-direct with Daniel Herrmann, searching for a path to safe AI, if one exists. exproof.ai →

Metascience

Grant funding, replication crisis, social epistemology of science

My metascience work asks how scientific communities produce knowledge, and how to do it better. I study grant funding, the replication crisis, and the social epistemology of science, using Bayesian inference, game theory, network theory, and agent-based models.

Figure 3Twelve agents try two actions, one slightly better, and share results with the agents they are linked to. A node’s shade is its credence that the better action is better. Green: the community settles on it. Red: it locks in on the worse one.
Selected work

Methods

Bayesian epistemology, decision theory, evolutionary game theory

My methods are formal: Bayesian epistemology, decision theory, and evolutionary game theory. I also study the methods themselves: when different models of evolution, such as the replicator dynamics and the Moran process, disagree about what evolves, and why; and when equilibrium concepts pick out the outcomes that matter for evolution and when they miss them.

Figure 4Possible futures of an agent’s credence that a coin is biased toward heads, simulated from its own probabilities. For a Bayesian, their average (red) stays level, as Reflection requires. The toggle shows an agent who expects to change its mind.
Selected work

Recent and upcoming talks

All talks

Contact

amohseni@cmu.edu

Review modeHover a highlight to see what needs confirming.Exit review