Nick Cammarata, a researcher at the AI startup Goodfire, has published a lengthy blog post claiming that neural networks are far easier to understand than the field has assumed. He argues that the apparent messiness inside models is partly a result of how researchers have been looking at them, and that with more careful observation, entire circuits can be mapped. The post has drawn intense attention, with some AI researchers calling it a potential landmark moment.
The claim directly challenges the long-standing idea of superposition, where a single neuron seems to represent multiple unrelated concepts. Cammarata himself helped establish that view years ago, but now says it was a misinterpretation. He revisits an example from an early image model: a neuron that responded to both cars and cats. Instead of seeing unrelated concepts, he now sees a common geometric feature—a V shape—that links them. He suggests that this kind of legibility may hold more broadly.
Not everyone is convinced. Chris Potts at Stanford, who leads a team studying how LLMs work, says the post offers little more than a hunch and does not make the underlying interpretability problems any easier. Cammarata acknowledged he was surprised by the strength of the reaction, saying he did not expect people to update their beliefs so much. The debate highlights how desperate the AI field is for better ways to monitor and control increasingly autonomous systems.