Bio
I'm a Research Engineer on the language model interpretability team at Google DeepMind, working on trying to understand frontier AI model's cognition. I also spend a lot of time thinking about how to improve the efficiency and productivity of safety researchers. More about me here: bilalchughtai.co.uk.
Projects
Research Direction: Interpretability
We’re primarily interested in more pragmatic interpretability ideas these days, although we have also listed some basic science questions we would be interested in supervising projects on as well.
- Applied Interp:
- Build probes that generalize better. Possible methods: Data pruning, Synthetic data, Asking an LLM to apply distribution shifts to existing data
- Build a probe creation agent: Given a target behavior, the agent iterates on a probe to detect that behavior
- Determine to what extent LLMs know when you mess with their CoT
- Model Psychology:
- Trawl arxiv / lesswrong / etc for weird model behaviors. Auto replicate these behaviors and possibly investigate weird behaviors in more depth
- Take a weird behavior and study it in depth, e.g. introspection, Neural Chameleons, roleplaying, uncertainty probes
- To what extent do models have consistent values? Consistent preferences? How can we find these?
- Model representations:
- Study in context learned representations
What I'm looking for in a Mentee
We like to iterate fast on different ideas. We think it's great if we spend each of the first few weeks trying a different idea and then choose one to dive into in more depth.
- We haven't mentored for Pivotal before but think that groups of 1-3 mentees are probably best.
- We like to meet with our mentees once a week, and also try to be responsive on slack to research updates.
- Prior interpretability research experience isn't required, but we do think a strong research (e.g. current pursuing or have completed a PhD) or engineering (software, quant finance) background will result in the highest probability of a successful project