The researchers of the Anthropic company announced that during the study of the Claude language model, they discovered a special internal structure - a "hidden thinking space", which was named J-space. It allows the model to "keep in mind" and develop ideas without immediately translating them into words or displaying them to the user.
According to the researchers, the J-space can be imagined as a small internal workspace where Claude can simultaneously work with different ideas, compare them and only then form the final answer. Unlike famouschain of thought(chain of thought), which in some cases is presented in text form, J-space operates entirely at the level of neural activities, and the user does not see the processes taking place there.
Anthropic notes that J-space was not specifically created by developers. It was formed independently during Claude's training. Each such internal activity is associated with a word or concept, but this does not mean that the model is going to write that very word, it simply shows what idea the internal calculation is going on at that moment.
During the research, scientists have shown that it is possible to follow how Claude performs some logical operations at the internal level using J-space. For example, it can detect errors in software code, recognize images, plan next steps, or understand that it is being misled by specially worded instructions, even if all of this is not reflected in the text of the final response.
The researchers also noticed that when Claude was given conflicting tasks at the same time, the model could internally maintain both different hypotheses and only later choose one of them.
At the same time, the results of the research showed that J-space is functionally similar to some properties of the "Global Workspace" of the human brain. That system makes various information available for conscious control and further processing. Each internal state of J-space is associated with a certain concept or word, but its activation does not mean that the model is going to pronounce or write that very word. Rather, it is an internal state within which information becomes available to the various processes of the neural network and can be used in further computations.
However, Anthropic specifically emphasizes that this finding does not mean that Claude has consciousness or subjective feelings. According to the authors of the study, modern large language models can independently form internal information processing structures that differ from basic neural computations and are used to perform more complex cognitive operations.
According to the company, the study of J-space could be important in the field of artificial intelligence security, helping to better understand how language models make decisions and how to more effectively detect potential errors and dangerous behavior.