Greg Burnham

Before Epoch AI, Burnham worked at Elemental Cognition and Bridgewater Associates. He holds a BA in mathematics from Princeton. His current work brings that mathematical background to the design and interpretation of AI evaluations.[1]

Evaluating work by its own standards

With Kelly Hong, he co-authored Epoch’s automation report, which gives models tasks drawn from the organization’s actual work. Employees grade outputs against task-specific rubrics, combining objective checks with judgments about clarity and research quality. Examining the agents’ trajectories helps explain why an output failed. The study distinguishes reliable completion of well-defined steps from the judgment needed to choose topics, design experiments and follow through on an open-ended task.[3]

Spatial reasoning in a practical setting

His co-authored Furniture Assembly Benchmark work asks models to find mistakes in photos of partially assembled furniture. The task tests visual and spatial reasoning in a setting connected to physical repair and assembly. It provides a concrete domain to track over time, while the connection to useful real-world assistance remains a question to investigate rather than something established by a score alone.[4]

Burnham argues that a benchmark should help answer a larger question than whether a model can complete one narrow task. His examples include whether systems improve through repeated experience, how performance generalizes across domains and whether an agent can complete a messy job rather than a self-contained assignment. He treats measurement as a way to investigate the forces behind AI’s impact, with realism and evaluation cost constraining what can be tested.[2]