CS at Cambridge.
Research
- Learning to Read Out: Unembedding Dynamics in Language Model Pretraining (NeurIPS 2026) arXiv · code · project page · feature atlas · 🤗 crosscoders
- Sparse Readout Prism: Explaining Logit-Lens Scores in Features Instead of Tokens arXiv · code · project page · 🤗 dictionaries
Tools and explorables
- atlas-of-loss-surfaces: forty loss surfaces from the optimization literature, explored by twelve optimizers in WebGL2
- vigil: real-time TUI for monitoring cloud GPU training instances