Biography
Hi! I am an early-career researcher in machine learning.
My interests:
- Understanding how deep neural networks perform computation
- Translating this knowledge into practical applications, e.g. addressing existential risk due to misalignment.
Check out my LessWrong profile, where I post write-ups on finished work, position pieces, and shortform notes on atomic thoughts / musings.
Some past work I am proud of:
- My most recent first-author paper on the generalization properties of steering vectors.
- My LessWrong post on paradigms in mechanistic interpretability
Current projects I’m working on include:
- Circuit discovery using sparse autoencoders.
- Working towards a more complete understanding of feature geometry in LLMs
- SAEs and interpretability for multi-modal models
Interests
- Artificial Intelligence
- Mechanistic Interpretabilty
- Generalist Robots
Education
BSc. in Mathematical and Computational Sciences, 2017
Stanford University
Selected Publications
Quickly discover relevant content by filtering publications.
Experience
Neel Nanda MATS 6.0 Trainee
1 of 30 aspiring mechanistic interpretability researchers, handpicked by Neel Nanda, to do a 5-week intense course. Became fluent with foundational mech interp concepts and tooling. Did a 2-week sprint project on sparse feature circuits in GPT-2-small.
PhD Candidate, UCL Computer Science
Researcher at the intersection of mechanistic interpretability and multimodal foundation models. Advised by Brooks Paige and Dimitrios Kanoulas
Research Engineer, Robotics and Autonomous Systems Division
- Implemented reinforcement learning methods to train legged locomotion controllers.
- Evaluated trained policies on real hardware. Supervised by: Michael Chuah, Yau Wei Yun