Goodfire essay claims neural networks are "already clean"; researchers disagree on the assessment
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
If models' neurons really are simpler than thought, understanding SI from the inside would get easier — but experts do not agree yet.
In the essay "Neural Networks Are Already Clean," Nick Cammarata writes that neural networks are much easier to understand than believed: the famous "polysemantic" neurons were, in his account, misinterpreted (the "cats and cars" neuron responds to V-shaped objects), and by his estimate 90% of neuron poles in Qwen3-32B have one main function. Google DeepMind's interpretability lead Neel Nanda is skeptical: he believes superposition is real, it is unclear whether findings carry over from image models to language models, and the essay "doesn't really change the big picture."
Sources 1