Superposition is a feature, not a bug: it does not affect interpretability, and it increases memory capacity. If each neuron in an LLM corresponds to a single feature, that's a problem, and we need to fix it.

→ View original post on X — @pmddomingos