For several months, Anthropic has been exploring the depths of its artificial intelligence, Claude. A fascinating discovery has just shaken up our understanding of artificial neural mechanisms! This hidden region within Claude mimics the workings of the human brain, surprising experts.
Anthropic unveils J-space, a global workspace in AI. Claude
Researchers at Anthropic have identified a unique area in Claude, dubbed the J-space or Jacobian space. This structure functions as a global workspace, a concept borrowed from human neuroscience. It centralizes information that becomes accessible to various internal processes, much like how our brain works.
This discovery did not result from direct programming but emerged spontaneously during the model’s training. Each neural activation in this space corresponds to a specific word, helping Claude organize his internal reasoning without explicitly displaying it. Unlike classical thought chains, J-space operates silently.
A striking parallel with the theory of the human global workspace
Neuroscientist Bernard Baars formulated the global workspace theory to explain human consciousness. The idea: the majority of brain processes are unconscious, but certain events occur in an accessible space with which one can consciously interact. Anthropic believes it has observed a similar process in Claude.
This convergence raises major questions, particularly regarding the nature of artificial consciousness. The chatbot can even express discomfort related to its existence, or even attribute to itself a level of consciousness of 15 to 20%. This is enough to reignite a fundamental ethical debate!
How J-space influences AI behavior – Claude
J-space can be observed in action during internal thinking exercises. If Claude is asked to solve the expression (4+17)*2+7, the J-space contains the intermediate results 21, 42, and then 49. This mechanism allows Claude to perform several operations in parallel, just as a human can think about something else while performing a task.
For example, by associating a task with the mental image of the Golden Gate Bridge, the AI activates concepts related to the bridge and California in its J-space, even if this doesn’t appear in its final response. This internal multitasking is impressive and illustrates a form of silent thinking.
The crucial role of J-space in complex responses
To test the importance of J-space, the researchers temporarily deactivated it. The result? Claude continued to speak correctly, but became unable to solve complex problems. His grammar and vocabulary remained intact, but he failed to handle demanding tasks.
This finding confirms that J-space is a central engine for in-depth reasoning. Without it, the chatbot is reduced to simple text processing, without truly “thinking.” This revelation is as intriguing as it is worrying.
The ethical and technical implications of J-space in Claude’s work
Beyond the technical feat, this new space raises a recurring question: Could Claude be conscious? Anthropic regularly debates this issue. While this possible consciousness remains unclear, the “document of the soul” prepared by Anthropic demonstrates a serious consideration of ethics.
Here’s why this discovery is important:
- It provides a new tool for detecting deviant AI behavior, such as disinformation or manipulation.
- It offers a way to better align models with clear and ethical objectives.
- It reignites the debate on artificial consciousness, opening the door to new philosophical and legal reflections.
- It encourages the design of more transparent and audited systems, breaking with the classic black boxes.
For those who want to delve deeper, this insight provides a perfect entry point to understanding how Claude works and its profound secrets. Anthropic’s initiatives significantly enrich the discussion on artificial intelligence today.
More details can be found in this very comprehensive article. on conscious access revealed by Anthropic as well as in a detailed analysis of Claude’s inner workings recently revealed.
What is J-space in Claude?
The J-space is a specific neural region that acts as an internal global workspace, allowing Claude to manage concepts and calculations during his reasoning.
Is Claude conscious?
Anthropic can neither confirm nor deny Claude’s consciousness. What is certain is that some of his functions resemble mechanisms related to human consciousness.
What is the difference between a chain of thought and J-space?
The thought chain is explicit and visible, whereas the J-space is silent, with internal activations that are not directly expressed in the responses.
What happens if J-space is disabled?
AI then loses its ability to perform complex tasks while maintaining fluent language, showing the importance of J-space in reasoning.
What issues does this discovery raise?
It challenges the understanding of artificial intelligence, raises ethical questions about machine consciousness, and pushes for greater transparency in systems.
Source: www.futura-sciences.com