Research & Writing
Publications.
Selected preprints, conference papers, and technical notes on the thermodynamics of intelligence and mechanistic interpretability.
ArXiv Preprint2025
Dataset Distillation for the Pre-Training Era
Introducing Linear Gradient Matching to condense massive datasets into a single synthetic image per class, revealing shared 'Platonic' representations across foundation models.
Research Note2025
The Platonic Representation Hypothesis
Exploring how diverse foundation models converge towards a shared representation of reality as they scale in compute and data.