Francis Rhys Ward
I'm open to quite a broad range of work. By default I expect scholars to work on reasonably conceptually-simple empirical projects, such as building capability or propensity evals.
Stream overview
Alignment-science style work eg:
- Train model organisms of scheming.
- Study how character training influences reward hacking rates and how models reason about reward hacking.
Tech-gov style work:
- Write a report detailing public information on frontier AI company monitoring and control practices.
- Write a report detailing public information on training episode 'length' of frontier AIs.
Other:
- Build agentic propensity evaluations for reward hacking, apparent-success seeking, grader-approval, etc.
- Work on measuring AI R&D speed-up / indicators of RSI.
- Conceptual work, e.g., criticising existing threat models, or on the deal-making agenda.
Standard (1-2 hours of weekly 1:1s)
London
Indifferent
Indifferent
Mentors
Rhys is the research director of Arrow: a new AI safety non-profit based in London.
He is currently most excited about working on alignment science (e.g., model organisms research) and technical governance style work which reduces public uncertainty regarding important questions (e.g., his recent work on no-CoT time-horizons).
Rhys started working on AGI risk in 2019. He has previously worked at: Redwood Research, LawZero, UK AISI, GovAI, CLR, and the Centre for Assuring Autonomy. He did his PhD in AI Deception at Imperial College London.
Fellows we are looking for
Technical skills, experience building evals with inspect, or working on agents or fine-tuning models with tinker.
It's good if you have at least one completed AI safety project.