-
22 June 2026
A field guide to one of the oldest tools in interpretability: the idea, what it found in BERT
and Othello-GPT, and why probe accuracy is not the same thing as understanding a model.
-
22 June 2026
The technique that turns "this information is in here" into "the model actually uses it."
How it works, what it found in ROME and GPT-2, and where it can mislead you.
-
23 July 2026
The tool that tries to break a model's tangled neurons into one concept at a time:
the superposition problem it solves, how the dictionary works, the Golden Gate feature,
and why a good reconstruction is not the same as understanding.
-
11 June 2026
How transformers encode bracket depth, indentation, table position,
and magnitude, and why some of these tasks get a clean curved
manifold while others don't.