Neural Networks Are Already Clean

What happened

Goodfre researcher Nick Cammarata published an extensive essay arguing that neurons in neural networks are far cleaner and more interpretable than mainstream interpretability research assumes. MIT Technology Review covered the claim as a potential breakthrough.

Context and impact

If the hypothesis holds, sparse autoencoders (superposition decomposition tools) may not be the necessary foundation of interpretability research. This would significantly simplify model analysis and redirect research priorities.

Details

• Polysemanticity, per Cammarata, mostly comes from fixable errors (feature viz reward hacking, poor regularization) not superposition
• Example: the famous cat/car neuron in InceptionV1 is reinterpreted as a V-shape detector, not superposition
• Modern architectures (residual + LayerNorm): ~99% of ConvNeXt neurons are easily interpretable
• Qwen3-32B pilot audit: ~90% of neuron poles have one primary function, ~7% have two related functions, 2-4% unexplained
• Author labels this a working hypothesis, not a conclusion — companion paper in preparation