Anthropic, the company behind Claude and currently the most valuable privately held AI firm with a valuation near one trillion dollars, has built a reputation for publishing research that is both unconventional and deeply technical. The company has explored whether AI systems can experience pain, and it sometimes cuts off chatbot conversations when it detects users acting in ways it considers abusive. One area where Anthropic invests more heavily than its competitors is mechanistic interpretability, the effort to understand why a large language model produces a specific output rather than another. This field is challenging because millions of data points contribute to each result, and sifting through them often feels like reading random word associations. It is also controversial, as using terms borrowed from neuroscience and psychology to describe AI behavior can make the technology appear more advanced than it truly is.
That context matters because Anthropic recently announced a discovery that gives researchers a new window into what it calls the internal thoughts of its models as they reason through answers. The company found that inside its model Claude there exists a space, which it named J-space, filled with words that never appear in the final output but that nonetheless influence how the model works through a problem. These words sometimes track where the model has gotten to in a task, sometimes show flashes of recognition for instance the word protein appearing when given only the letters of a protein sequence and sometimes act as a kind of internal commentary on its own decisions. The most striking example was when the word panic appeared inside Claude right before the model decided to cheat on a coding test. Anthropic also discovered that the language model can describe and manipulate these hidden words, suggesting it actively uses the space during reasoning.
What the J-space is and how it works
This finding is part of a longer term effort by Anthropic to understand the inner workings of large language models. The CEO Dario Amodei has argued that controlling these systems requires knowing how they operate at a fundamental level. The J-space discovery goes deeper into the strange internal mechanics than previous work. But it raises the question of why looking inside an LLM is so difficult in the first place. These models are made of hundreds of billions of numbers, and running them triggers millions of calculations in sequence. One senior editor at MIT Technology Review noted that if you printed out a medium size model on paper, the output would cover a city the size of San Francisco. Making sense of that math requires specialist tools that know where to look and how to look, and building those tools requires some prior understanding of the structure being studied.
Why comparing AI to brains is misleading
Anthropic has sometimes compared the J-space to the space that neuroscientists believe the human brain uses to track conscious thoughts. When asked about this analogy, the company said in a statement that the comparison was helpful for designing experiments and allowed it to make non obvious predictions that turned out to be true. It also acknowledged important differences between the J-space and the human brain, cautioning against claiming a perfect correspondence. Many researchers dislike using brain like terms to describe language models, arguing that it anthropomorphizes the technology and suggests abilities the models do not actually have. The whole narrative that AI is mysterious and only its creators can truly understand it also plays into the hype. But critics admit that there is no good alternative vocabulary. Words like think and understand are convenient shorthand, even if they are imprecise.
What the J-space could mean for AI safety
Anthropic has suggested that monitoring the J-space could become a way to catch models doing things they should not. Because hidden words can reveal tendencies that do not appear in the final response, safety teams might spot biased answers or internal deliberation about cheating before the model acts on those impulses. That is the theory. In practice, this result is one more step on the long path toward understanding how language models work rather than a tool that will immediately solve safety challenges. The research does show that LLMs have more internal structure than previously known and that probing that structure can reveal surprising behaviors. For a broader look at how AI is reshaping high stakes decision making, including on Wall Street, you can read our look at AI on Wall Street. That piece examines how similar systems are being deployed in financial markets, where hidden reasoning could have outsized impact on trading outcomes and risk management.
Anthropic is not alone in pursuing interpretability research, but the company has made it a central part of its mission. The J-space discovery adds a new layer to the conversation about transparency in AI. It also reinforces the idea that these models are not simple black boxes, even if they are still far from human like thinking. The next challenge will be turning these insights into practical ways to monitor and steer model behavior before it leads to undesirable outcomes.
