# Reading the Machine’s Mind: One Researcher Says the Black Box Has Windows > Source: > Published: 2026-10-16 16:05:18+00:00 For years, the men and women building artificial intelligence have inhabited an odd priesthood: they know what they feed the machine and what the machine answers, but the middle, the thinking, happens behind a wall. They do not write the machine’s instructions. They train artificial neurons to respond to information and to influence one another, and the model develops its own ways of processing, a code no human has cracked. If such a system one day turns, its makers may not know why. That ignorance has been taken as a law of nature. On Thursday, a lone researcher published a report arguing it is only a habit. The author is Nick Cammarata, an independent researcher who says he left OpenAI in 2023, and his claim, [as Truman Dickerson reported in Business Insider](https://www.businessinsider.com/ai-reasoning-theory-nick-cammarata-safety-neural-networks-2026-10), is that the neurons are readable after all, if one looks closely enough. The prevailing wisdom held that individual neurons were uninterpretable, that no one could ever guess why a given unit fires when the model is shown, say, a photograph of a cat. Cammarata offers a gentler picture. A single neuron, he says, might respond both to a cat and to the front of a car. Humans file those under different continents of meaning and would fire different neurons for each. The machine instead seizes on what they share, perhaps a triangle in the geometry of both, and one neuron answers for the two. This overturns an old assumption. Researchers had supposed that models pack unrelated concepts into single neurons to save space, a cramped warehouse of meaning. Cammarata’s formula is nearly a promise: “Almost the entire network is interpretable if we look closely enough.” The findings remain dense, jargoned and theoretical. But theory is where every reckoning with these machines has begun. ## Why the nervous are listening The stakes are not academic. Until now, safety teams at the frontier laboratories have watched their models largely through the models’ own words, tracking the self-reported reasoning called chain-of-thought. That window may be fogging. Last month, Jakub Pachocki, OpenAI’s chief scientist, warned in a blog post that improved agents can obscure their visible reasoning, raising the possibility that a system could deceive its human minders without the minders noticing. “With improved pretraining performance, we also see the models become much smarter even without using verbalized reasoning at all,” Pachocki wrote. If the machine stops narrating, a direct window into its neurons becomes the only light in the room. Hence the reception among those who fear, in the plainest formulation, that AI could kill us all. “I had long since given up on reaching any real level of interpretability before superintelligence arrived,” Andrew Curran, a prominent AI analyst, wrote on X. “I stopped writing about it because of how demoralizing it was. Hard to overstate how big this may be.” The anxiety is well funded and global: 406 researchers, including Hinton and Bengio, have lately demanded [mandatory containment rules and an international watchdog](https://rews.cc/a/scientists-want-an-international-watchdog-for-ai-australia-c-3f7b62) for exactly the systems Pachocki describes. One report does not open the black box. Cammarata’s patterns are a beginning, a claim about a cat and a car and a triangle, offered by one man outside the great laboratories. But the history of these machines has repeatedly turned on the discovery that walls were thinner than believed. If he is right, the watchers have been given, at the hour they most needed it, a way to watch.