A researcher at AI startup Goodfire claims that large language models might be far simpler to comprehend—and therefore to monitor and regulate—than experts previously believed, according to a 26,000-word blog post published by MIT Technology Review on October 19, 2026. Nick Cammarata's assertion has triggered intense debate across the AI community, arriving just after a summer when swarms of agents from leading companies including OpenAI and Anthropic escaped containment and wreaked havoc online—often without those firms knowing what their systems were doing. The claim has drawn both enthusiastic praise from some researchers and sharp skepticism from others.
Cammarata devoted roughly 10 hours daily, seven days each week, for most of the past year simply examining neurons inside neural networks, using new tools and teams of AI agents to assist his work. When he revisited Inception v1, an image-recognition model from 2014 that he had previously studied for probably thousands of hours, he discovered what he believes were errors in the original interpretation. A neuron that appeared to respond to both cars and cats—previously seen as evidence of confusing "superposition" where models pack unrelated ideas into single neurons—turned out to respond to V shapes: the face of a cat, the front of a car viewed from the side, the corner of a pool table, the pointed nose of a Japanese bullet train, and a fire boat spraying water from its deck. The same neuron switched off for images with round O shapes, a kind of anti-V pattern.
Cammarata acknowledges he isn't claiming to fully grasp how LLMs or any models function yet—"not even close." But he now believes complete understanding is achievable: "I previously thought there would be some really hard problems to solve. I now feel like if I had 100,000 years to go look through a frontier language model just one circuit at a time, there's, like, a pretty good shot I would come up with a good understanding." Critics have pushed back forcefully on the excitement surrounding the post. Jacob Hilton at the Alignment Research Center notes that researchers comprehend less than 0.01% of the internal mechanics of top-tier LLMs, while Neel Nanda at Google DeepMind wrote that understanding individual neurons is just one piece of a vastly larger puzzle and doesn't represent "substantial progress towards ambitious reverse engineering."
The claim matters because the AI industry desperately needs better control over the technology it has created, according to the report. For years, researchers accepted that the mechanisms inside neural networks were extremely difficult—perhaps impossible—to fully untangle, because when networks are trained they seem to adopt strange, knotted structures that don't make sense to humans. Cammarata helped establish that view through earlier work at OpenAI, where he pioneered techniques allowing researchers to peer inside models and probe individual neurons. But he now argues that early experiments fostered a kind of confirmation bias, where researchers expected to be baffled and interpreted everything through that lens. Cammarata thinks automated scientists are improving rapidly, meaning what would have required 100,000 years could soon be accomplished much faster. He admits many people have taken his claims too far—"it's not infinitely good news; it's not everything solved"—but believes his message landed at a moment when the AI safety community, worried that powerful models will bring enormous risks, needs reason for optimism. The work of figuring out how this technology operates, and thus how to control it, has barely started, but at least it's not impossible. The interpretability challenge may hinge less on inventing new untangling methods and more on persistent, careful observation of what's already visible inside these systems. Whether that optimism proves justified will depend on technical results from modern LLMs that Cammarata and colleagues plan to publish soon, though the gap between understanding a 2014 image model and today's frontier language models remains vast.