An Anthropic study says that a structure emerged during training that allows the model to reason without turning its thoughts into text.
Anthropic announced this week that Claude, its artificial intelligence (AI) language model, has an operating mechanism that resembles human consciousness.
According to a study published on Monday, the 6th, Claude developed a small set of internal patterns that function as a private "mental space," where it thinks without writing—what the company's researchers call "J-Space."
The research also states that this space was neither designed nor programmed by anyone, but instead emerged spontaneously during the model's training. Furthermore, the company says the evidence does not indicate "whether Claude is conscious in the same way humans are, or whether it feels anything"...
J-Space:
Each pattern within J-Space is associated with a word, but that does not mean the model is actually saying that word—it simply means the concept is "in its mind."
This differs from the so-called draft, or "chain of thought," which is the text models write to themselves while reasoning, according to the study.
Anthropic says J-Space operates silently within the neural network's internal activations, allowing Claude to think about a concept without putting it into words.
According to the study, what appears there goes far beyond the text Claude is reading or writing. When the model reads code containing an unnoticed bug, its J-Space activates the word "error."
When it reads a protein sequence, the space contains its biological function. And when it receives search results that are actually an attempt to manipulate it—a type of attack known as prompt injection—J-Space activates the words "injection" and "false."
The tests also showed that J-Space has unusual properties compared to the rest of the model's processing.
The first is that Claude can report what is present there. When asked what it is thinking about, it describes the contents of J-Space.
The second is that it can control the space on demand—if asked to think about citrus fruits while copying a sentence about a painting, the space activates "orange" and "fruits," even though none of those words appear in the final text.
To confirm that this space actually causes the behavior rather than merely reflecting it, the researchers performed direct interventions. When Claude was asked to think of a sport, "football" appeared at the top of the list; after replacing the "football" pattern with the "rugby" pattern inside the network, the model began answering "rugby." The response followed the modification, indicating that it is indeed read from that space.
How AI Thinks Without Speaking:
The study shows that Claude uses J-Space to reason.
When presented with the question, "The number of legs of the animal that spins webs is," the model must first infer that the animal is a spider and then "remember" how many legs it has.
The word "spider" appears neither in the question nor in the answer—it is an internal reasoning step. When the researchers replaced "spider" with "ant" in J-Space, the model answered six instead of eight.
The same space is used for multiple tasks simultaneously.
By replacing "France" with "China" in this mental space, Claude correctly answered Beijing as the capital, Chinese as the language, Asia as the continent, and yuan as the currency—all from a single modification. It is this shared representation that led researchers to compare the finding to a well-known theory in neuroscience.
0 Comments