Nick Camerata

Camerata says this is only a working hypothesis held with about 85% confidence, and not everything has been audited.

Camerata re-examined neuron 4E55 in Inception V1 and concluded that it is a V-shape neuron rather than a cat-car neuron; its strongest activations correspond to V-shaped flames, V-shaped fox faces, and so on, while its negative poles correspond to the round O.

Camerata points out that the car features hidden in the dog neuron are not due to compositional readout but because cars and dogs share poses; there are hundreds of such pose neurons in Inception V1.

Camerata proposes recursive generalization: generalization is not local but unfolds layer by layer like a fractal; for example, a neuron sensitive to the letter IY extends to DIY and to do-it-yourself, which does not contain IY.

Camerata defines "clean" as every component being designed to be read and built upon by subsequent components, rather than every neuron matching the labels we already have.

Goodfire researcher Nick Camerata published a long article whose core argument is that the interior of neural networks is not a chaotic superposition but is interpretable.

Camerata re-analyzed the 2019 experiment once regarded as the only evidence of a superposition circuit and found that it was not feature superposition but pose encoding, and that the visualization method at the time was flawed.

Camerata proposed the concept of "recursive generalization": neurons branch and grow along related directions during training, and each branch has a provenance rather than being forcibly stuffed in as superposition.

Camerata advocates reading the original network directly (exploratory interpretation) and opposes the approach of first training sparse autoencoders (SAEs) to decompose the network.

Camerata argues that superposition theory should be downgraded from a default assumption to an assumption of last resort; when a model is not understood, it is more likely due to tools, architecture, or recursive generalization.