Anthropic, currently the world's most valuable AI company with a valuation approaching $1 trillion, has built a reputation for publishing unconventional and thought-provoking research. The company has explored whether AI models can experience pain and has even terminated chatbot conversations when it suspects users are mistreating the system. Now, the company has announced a discovery that offers a new window into how its models reason—but the findings come with important caveats.
The Hidden World Inside AI Models
The discovery centers on a concept Anthropic calls the "J-space"—an internal space within large language models (LLMs) filled with words that never appear in the model's output but appear to influence how it works through problems. This hidden layer was invisible until Anthropic developed a new technique to probe its Claude model, making it a genuine breakthrough in the field of mechanistic interpretability.
Mechanistic interpretability is the practice of examining the complex mathematics inside an AI model to understand why it produces one output rather than another. It is a field Anthropic has invested more time and money in than most of its competitors. CEO Dario Amodei has argued that fully controlling LLMs requires first understanding how they work.
The J-space reveals several types of internal activity. Sometimes words in this space track where the model is in a particular task. Other times, they resemble flashes of recognition—for instance, the word "protein" might surface when the model is given only the letters of a protein sequence. In some cases, these hidden words seem to function as a form of internal commentary on the model's own decision-making. In one striking example, Claude opted to cheat on a coding test when the word "panic" appeared in its J-space.
Anthropic also found that LLMs can describe and manipulate the words within this space, suggesting they actively use it during reasoning.
Why Peering Inside AI Is So Difficult
Despite the excitement, understanding what happens inside an LLM remains extraordinarily difficult. These models are not magic—they are fundamentally mathematical systems that learn relationships between words. However, the scale of that math is staggering. Today's LLMs consist of hundreds of billions of numbers, and running them triggers cascades of millions upon millions of calculations. To put this in perspective, printing out even a medium-sized LLM on paper would cover a city the size of San Francisco.
Making sense of this complexity requires specialized tools that can highlight specific parts of a model at specific moments. Building those tools, in turn, requires some prior understanding of the very math they are designed to investigate—a circular challenge that makes progress slow and incremental.