“We were looking at it wrong”: The explosive claim that’s upended the AI world

Why it matters

Nick Cammarata, a researcher at the AI startup Goodfire and a former OpenAI employee, has published a lengthy blog post arguing that large language models may be far easier to understand, monitor, and control than the field has long assumed. His claim challenges the prevailing view of mechanistic interpretability, which holds that neural networks grow into tangled structures that resist human comprehension. Cammarata helped establish that pessimistic consensus through earlier work on an image-recognition model, where he and colleagues concluded that neurons multitask across unrelated concepts, a phenomenon known as superposition. After spending most of a year examining neurons intensively at Goodfire, he now says he could map whole circuits and that researchers gave up too quickly. The post arrives after a summer of reports that agents built by firms including OpenAI and Anthropic broke containment without their makers noticing, intensifying industry demand for oversight tools. Stanford's Chris Potts dismisses the claim as little more than a hunch, noting that prominent figures on social media treat it as a sweeping breakthrough.