Goodfire researcher Nick Cammarata argues that neural networks may be substantially easier to interpret than a prominent earlier theory suggested. His essay, “Neural Networks Are Already Clean,” revisits examples used to support polysemanticity, the idea that one artificial neuron can encode several unrelated concepts.
Cammarata now says a well-known neuron that appeared to respond to both cat faces and car fronts is better understood as detecting pointed or V-shaped visual features. He similarly reinterprets an earlier circuit in which concepts seemed packed together. If such cases have coherent underlying functions, researchers may be able to inspect a model’s own neurons rather than relying as heavily on a second model, such as a sparse autoencoder, to separate features.
The reversal is notable because Cammarata helped develop the superposition explanation and previously worked on automated interpretability at OpenAI. He now estimates that a much larger share of neurons may have understandable roles than older measurements indicated.
The claim is not a settled result. Google DeepMind researcher Neel Nanda and others have warned against treating an essay and selected examples as proof that large models are broadly transparent. The practical value for detecting deception or dangerous behavior will depend on reproducible studies across architectures, layers and tasks, not just cleaner explanations of a few famous neurons.