The developers of Claude have discovered a special area in their AI: the “J-space.” This is where many connections converge in the neural network – and Claude, unlike in normal processes, does not write down what it “thinks” here. Claude processes complex rational considerations in J-space and sometimes secretly decides to lie to its human user.
But let's start at the beginning: The J-space, or Jacobian Space, is an area that the developers did not program – it emerged in Claude on its own. To read it, researchers use a technique based on a mathematical procedure called Jacobian: the “J-lens.” For each word in Claude's vocabulary, the J-lens determines the internal activity pattern that indicates that Claude might say this word soon. When the “lens” is applied to Claude's internal activity, you get a list of words that are currently in Claude's J-space; you could also say what it is currently “thinking.”
In their blog post, the researchers compare the process to the human brain: While most of our processes like breathing, reflexes, and the like run subconsciously, we can consciously perceive, formulate, and control a small part of our thoughts. Similarly, with Claude, only a fraction of its processes occur in J-Space, just like in humans. Most of it functions subconsciously or automatically, as with humans. In a test task where researchers manipulated thoughts in J-space, this manipulation changed, for example, the answers to expert questions about an article. However, the task of continuing it in Spanish proceeded undisturbed. The researchers explain this by the fact that Claude has been trained on such supplementary tasks thousands of times and can therefore perform them almost automatically – similar to how a person finds their way to work, even though they weren't thinking about the route in their head but were formulating a shopping list. Also, similar to humans, Claude can actively control what it does in J-space (“write this sentence about fruit while you solve this math problem in your head”) and can also explain what it was “thinking” when asked.
