About

I’m an independent AI safety researcher based in London, focused on understanding and evaluating the behaviour of large language models. My current work explores self-attribution bias, reasoning amplification for safety auditing, and the limits of LLM-based lie detection.

I built the Factorio Learning Environment, an open-ended benchmark for evaluating LLM agents in complex, unbounded scenarios. I was an Anthropic fellow and MATS scholar (2025-26), where I studied deceptive behaviour in language models.

Before moving into safety research, I co-founded two startups: Paperplane (YC W23), a conversational intelligence platform, and Spherical Defence, an API security company. I left because startup life felt a little solipsistic (how can I build something, so that I raise money, so that I make money), and I never found an idea with enough impact to be worth doing even if it made no money. I studied Computer Science at Cambridge (MPhil, Distinction) and Royal Holloway, University of London (BSc, 1st Class).

Outside of research, I once rowed across the Atlantic Ocean in 40 days and launched a gin company (Northwest Passage Expedition Gin) during COVID. I have also been teaching a language model to translate ancient Akkadian.