About me
I am a DPhil student in computer science at the University of Oxford, supervised by
Prof. Varun Kanade. I work mainly in machine learning theory, particularly theory of LLMs, and continual or open-ended learning. I am also interested in machine learning for mathematical reasoning.
Before starting my DPhil I spent two years at
Quantinuum, working on quantum machine learning with
Bob Coecke,
Stephen Clark and
Dimitri Kartsaklis. Prior to this I completed my MSc degree in computer science at Oxford, graduating in early 2022, with my dissertation on generalization bounds supervised by
Prof. Yarin Gal. I completed my BA in computer science at the University of Cambridge, graduating in 2019. My undergraduate dissertation was supervised by
Prof. Pietro Liò and focused on semi-supervised machine learning methods for cancer classification.
Ideas and collaborations
I am always interested in discussing ideas and potential collaborations. I have more ideas than I have time, and a few of my current ideas are:
- Mine GitHub for Lean proof patches, to train an LLM to construct proofs as in SWE-RL.
- A library for automated cross-layer transcoder construction (see this from Anthropic) in the same vein as Eleuther's sparsify.
- Using cross-layer transcoders to analyse how different parts of the LLM training pipeline affect medium- and long-term planning (see here).
- Anything to do with LLM theory, especially from a complexity theory perspective. See, for example, this paper on CoT and complexity classes, this paper on complexity-theoretic properties of multi-layer LLMs, or this paper on optimal space complexity for LLMs.
- Anything to do with continual learning or continual RL theory, or applications to LLM agents. I'm particularly interested in the case of long time-horizons, where we see many tasks, some repeated, compute has some cost, and we want to minimize regret. I am also particularly interested in infinite memory, limited compute continual learning.
If any of these ideas sound interesting to you, or if you have your own ideas that you would like to discuss, please get in touch at the email address above. I'm also just happy to chat!
Research
Research interests:
Learning theory, LLM theory, continual learning, "agents", online learning, meta-learning, deep learning, generalization.
Google Scholar profile:
Teaching
Departmental Tutor - Computational Learning Theory (Michaelmas 2023, 2024)
Miscellaneous
Other interests:
Economics of AI, theoretical CS, statistical physics, optimization, game theory, Arsenal football club, American football, spy novels, sci-fi. I hate running, but I do it anyway.
Favourite fiction books (no particular order):
Gorky Park, Blood Meridian, Brave New World, Leave it to Psmith, Do Androids Dream of Electric Sheep?, Moving Pictures, Berlin Game, All the Pretty Horses, Necropolis, Foundation, Catch-22
Favourite films (no particular order):
In Bruges, No Country for Old Men, The Handmaiden, Shrek 2, Blade Runner, The Nice Guys
Favourite video games (no particular order):
Dragon Age: Origins, Super Mario Galaxy, Fallout: New Vegas, Hollow Knight, Baldur's Gate 3, Assassin's Creed: Black Flag, Metal Gear Solid V, Outer Wilds
