Tim Lawson

I am a PhD student in the UKRI Centre for Doctoral Training in Interactive AI at the University of Bristol, currently working on mechanistic interpretability for language models. See LinkedIn (or below) for my CV, and GitHub for recent work.

Projects

Residual Stream Analysis with Multi-Layer SAEs (July 2024–)

Accepted to the Interpretable AI workshop at NeurIPS 2024
Preprint: https://arxiv.org/abs/2409.04185
Code: https://github.com/tim-lawson/mlsae

Education

University of Bristol (2023–2027)

PhD, Interactive Artificial Intelligence

Supervised by Laurence Aitchison and Conor Houghton
Foundation year (MSci) incl. natural language processing and machine learning

University of Cambridge (2012–2016)

MSci, BA Natural Sciences, Class II division 1

Theoretical physics, e.g. quantum and classical field theory, particle physics and relativity
Analysed the effect of ATLAS inner-detector deformations on the distributions of parity-violating observables for actual and Gaussian mixture models of dilepton events with C++

Experience

Graphcore (2021–2023)

Analysis Tools Frontend Engineer

Developed desktop applications to help researchers and compiler engineers to optimize AI models for IPU systems with e.g. TypeScript, Rust, SQL, React, and Electron

Techmodal (2016–2021)

Senior Developer

Data modelling, integration, analysis and visualization to support decision-making with SQL
Developed web applications with e.g. TypeScript, Node.js and React

Skills

Programming Languages: Python, JavaScript, TypeScript, SQL, Rust, HTML, CSS
Frameworks and Libraries: PyTorch, JAX, scikit-learn, NumPy, pandas, Matplotlib, React
Tools: Git, VS Code, LaTeX