Theodore Chapman

Theodore Chapman is an independent AI safety researcher focused on the nature and limits of capability elicitation in large language models. He holds degrees in data science and physics from the University of Rochester, where he also built machine learning pipelines for NASA satellite imagery analysis. He participated in the ML Alignment & Theory Scholars (MATS) Winter 2023-24 cohort under the supervision of Evan Hubinger, producing research on fine-tuning-based capability elicitation in GPT-3.5. His key finding was that the performance achieved by fine-tuning an LLM on a task using one prompt format does not reliably bound the performance achievable with a different prompt format, complicating safety evaluations that rely on fine-tuning to elicit hidden capabilities. He subsequently received a 6-month researcher stipend to continue this line of work, exploring how chat fine-tuning affects LLM capability elicitation, and has published related work on LessWrong and the Alignment Forum.

No endorsements yet.

Personal Website

-

LinkedIn

LinkedIn

Twitter / X

-

LessWrong

Theodore Chapman

EA Forum

-