Nick Cammarata

Nick Cammarata published a 26,000-word blog post arguing that our understanding of how neural networks work internally may be fundamentally wrong.

Cammarata's 26,000-word article is currently still at the "claim" stage, and the strength of its evidence, methodological details and reproducibility have not yet been systematically tested.

Cammarata's article has an impactful title and opening, but the source material provides limited information and does not elaborate on key details such as specific evidence, methodological differences and community doubts.

Nick Cammarata is a researcher at Goodfire.

In 2020, Cammarata and Chris Olah provided evidence in the paper "Zoom In" showing that the network hides car feature signals inside a group of dog neurons, reinforcing the superposition hypothesis.

In the large-scale automated neuron interpretation project Cammarata conducted at OpenAI in 2023, the tools available at the time could explain only about 2% of neurons.

In the modern vision models Cammarata tested, 99%–100% of the neurons in the last layer could be given fairly clear explanations.

Cammarata proposes the theory of recursive generalization, arguing that neurons gradually expand their responses along continuous dimensions of semantic or perceptual features during training, leading to apparent polysemanticity.

Cammarata's extreme stress experiment showed that after compressing 480 curve-detecting neurons in a certain layer of InceptionV1 down to 10 and retraining, these neurons were still clearly interpretable rather than chaotically superposed.

Cammarata admits that his article is only a working hypothesis, does not yet cover multimodal models, and that language models lack training checkpoint data.

Nick Cammarata's analysis points out that InceptionV1's 4e:55 neuron actually detects a V-shaped geometric form, rather than cat faces or car fronts.

Researcher Nick Cammarata proposed that agent-based methods can effectively solve the problem of polysemanticity in neural networks and significantly improve model interpretability and alignment.

Cammarata's claim spread mainly through social media, and no complete peer-reviewed paper or reproducible experimental details have yet been seen.

In October 2025, Nick Cammarata published a non-peer-reviewed long article through the interpretability company Goodfire, proposing that the "cat-car neuron" is actually distinguishing "pointy versus round" rather than encoding cats and cars simultaneously.

Cammarata's article triggered compatibly polarized reactions in the AI community: some see it as a breakthrough, while others remain wary.

If Cammarata's results can be independently reproduced, it may mark a paradigm shift in interpretability research from static analysis to dynamic interaction.