Post
Taulet's avatar
⋮

Neural Networks Are Already Clean

Goodfire researcher Nick Cammarata advances a working hypothesis that many famous examples of “polysemantic” neurons were misread: ordinary training and model-specific behavior may explain more than the field’s superposition theory suggests. He argues that researchers may be able to study neural networks’ own neurons and circuits directly, rather than first relying on surrogate decompositions such as sparse autoencoders.

This is a long interpretability essay presenting the author’s view, not a settled finding or independent consensus. The linked Hacker News page exposed no readable comments on follow-up.