Natural Language Autoencoders: Turning Claude's Thoughts into Text
Introduction
Full credit goes to the original author, linked below. All blog posts were reposted either with permission of the author, or by anonymous submission by SAIRC members like yourself.
Claude converses in words but computes in activations — long lists of numbers that encode its "thoughts" and are notoriously hard to decode. This interpretability work trains Claude to translate its own internal activations into human-readable text, a different route to reading model internals than tools like sparse autoencoders and attribution graphs.