Old questions.

What is it to represent the world, grasp a concept, understand, or deceive? Philosophy has long pursued such questions with only indirect access to the mechanisms involved. The emerging field of mechanistic interpretability changes that: in large language models, we can locate representations, trace the circuits that use them, and intervene to see what changes.

The PhilInterp Network connects researchers exploring the intersection between philosophy and mechanistic interpretability and showcases how they are putting these new instruments to work on longstanding questions in philosophy of mind, language, and ethics.