Mechanistic Interpretability

I built Local MI Lab because SELF-GROUND had become too heavy for the amount of mechanistic interpretability practice I actually had. The negation SAE work taught me a lot, but it also wrapped every …

I did not get the result I wanted from my first serious mechanistic interpretability attempt. Probably the best thing that could have happened. I started with an attractive idea: take sparse …

The field of mechanistic interpretability has matured rapidly over the past two years, transitioning from an academic curiosity to a critical component of AI safety research. As large language models …