Emergent Complexity via Multi-Agent Competition
Competitive self-play creates an automatic curriculum: as opponents improve, agents discover increasingly complex physical strategies.
Selected papers · 2014–2022
Since 2022, my research has shipped as models: o1 and o3 at OpenAI, and Muse Spark at Meta.
Competitive self-play creates an automatic curriculum: as opponents improve, agents discover increasingly complex physical strategies.
A meta-learning approach for agents that must adapt rapidly as their environment or their opponents change.
A meta-learning system designed to generalize across language tasks with different domains, label spaces, and few examples.
Uses unlabeled text to construct the learning experience needed to become better at adapting to new language tasks.
A spectral algorithm for recovering latent topics under interpretable assumptions, with provable error bounds.