Cavendish Labs

Cavendish Labs is a 501(c)(3) nonprofit research organization founded in January 2023 by Andrew Gritsevskiy and Derik Kauffman, with Joseph Cavanagh also serving as a director. The institute is based in Cavendish, Vermont, where it operates as a residential research community with researchers living and working on different floors of the same building, described as a combination of a small liberal arts college and research lab. The organization's research is concentrated in two primary areas. In AI safety, the team works on model-based reinforcement learning frameworks, provable behavioral guarantees, model interpretability, and related alignment research. Their published work includes contributions on unelicitable backdoors in language models, dictionary learning interpretability, and inverse scaling in large language models. In pandemic prevention, they focus on Far-UVC light technology for ambient disinfection and the development of low-cost diagnostic platforms for viral infections. All research is conducted computationally. Prior to founding Cavendish Labs, the core team won First Place at the Prometheus ELK (Eliciting Latent Knowledge) Competition and two Third Prizes at the Inverse Scaling Prize. The team has published research in venues including Nature and arXiv, with contributions from researchers including Andrew Gritsevskiy, Derik Kauffman, Joseph Cavanagh, Hans Gundlach, and Aaron Kirtland. Cavendish Labs offers research fellowships for early-career researchers in AI safety and bio-risk mitigation, providing a $1,500 per month stipend plus food and lodging. The community visits Boston at least once a month, hosts rotating visiting scholars, and maintains collaborations with researchers around the world. In 2025, co-founders Andrew Gritsevskiy and Derik Kauffman also co-founded RunRL, a Y Combinator-backed reinforcement learning startup based in San Francisco.