Will Brown

Brown completed a PhD in algorithmic game theory at Columbia, advised by Christos Papadimitriou and Tim Roughgarden. He previously studied computer science, data science, and philosophy at the University of Pennsylvania. His 2024 PhD retrospective describes starting graduate school in 2019 with an interest in systems where learning algorithms interact. The central question was how to design those systems so that adaptive, self-interested behavior leads to desirable outcomes.[1][2]

Before Prime Intellect, Brown worked in Morgan Stanley’s Machine Learning Research group on language-model projects. His public biography also records research and engineering roles at AWS, Two Sigma, MongoDB, and AmFam. At Prime Intellect he works on agentic reinforcement learning: training agents from the outcomes of tasks they attempt, with infrastructure that makes those tasks repeatable and their performance measurable.[1]

The Verifiers v1 announcement, co-authored by Brown, separates an environment into a taskset, harness, and runtime. The taskset defines data, tools, and scoring. The harness is the program attempting the task, which could be a simple model loop or a full coding agent. The runtime provides its execution environment. Keeping these pieces separate lets researchers compare different agents on the same task, move execution into a sandbox, and reuse recorded interaction traces for training.[3]

Brown’s PhD retrospective explains why a mathematical guarantee that learners reach an equilibrium does not necessarily mean they reach a useful one. Several stable outcomes can exist, with very different consequences for participants. His research asks how a system designer can steer learning toward better outcomes while accounting for how participants adapt. This provides context for his later interest in agent training environments, where choosing tasks and reward signals is part of defining what improvement means.[2]