Why do you care about AI Existential Safety?

In the short term, advanced AI systems are already shaping political and economic outcomes, and they increasingly mediate the information environment: either by amplifying polarization by making targeted persuasion, synthetic media, and narrative manipulation cheaper and easier to scale; through economic automation that advises or blindly performs operations in the market; or by being integrated into national security projects. I believe the current misuse of these systems leads to a higher chance of existential risks to humanity caused by AIs.

In the long term, once these systems automate a large share of the job market, the transition to this new economy could reshape the global economy faster than institutions can adapt. Without safeguards, inequality could deepen (being especially devastating to the poor), and the world could see heightened instability and conflict caused by new superpowers that might arise, empowered by AIs.

Please give at least one example of your research interests related to AI existential safety:

My research explores how multi-agent LLMs behave when they interact in environments that actually require coordination. I’m especially interested in identifying the key factors that make these agents cooperate reliably and then intentionally building on those factors to create agents whose behavior stays aligned with human goals.This is a broad research agenda. In the past, it has taken the shape of studying coordination games and heterogeneous LLMs interacting in cooperative settings, as well as developing a deterministic, realistic research environment for web agents so their behavior can be evaluated reproducibly.

More broadly, I want to understand what’s happening inside these systems: how they represent concepts internally and whether insights from biology, teleology, theoretical neuroscience and representation learning could be used in order make AI systems aligned more robustly with human preferences.