Blog

Notes on research, technical write-ups, and things I've learned.

• Decoding Concepts Under Superposition

  2026-07-07 interpretabilitysparse-autoencoderssuperpositioncompressed sensingUAI 2026

  Concepts can be linearly encoded in activations and still be impossible to decode with a linear map. An interactive tour of our UAI 2026 paper on why probes and sparse autoencoders fail on novel combinations of concepts, and what compressed sensing says to do about it.

• What Question Is Your Evidence Answering?

  2026-07-04 interpretabilitycausalityidentifiabilityICML 2026

  Interpretability claims routinely outrun their evidence. Causal inference says when, and identifiability says how many answers the evidence admits. An interactive tour of our ICML 2026 position paper.