A global workspace in language models

23 days ago

The article introduces the concept of 'J-space' in AI models like Claude, which are internal neural patterns that resemble consciously accessible thoughts in humans.
J-space is discovered using a mathematical technique called the Jacobian lens (J-lens), and it represents words or concepts the model is thinking about without necessarily expressing them.
Key properties of J-space include: Claude can report and modulate its contents, uses it for internal reasoning (e.g., multi-step tasks), and it functions as a flexible 'workspace' shared across different tasks.
J-space accounts for only a small fraction of Claude's processing; most automatic functions like grammar and fluency operate independently without it.
The research draws parallels to global workspace theory in neuroscience, suggesting J-space acts as a broadcasting hub for deliberate reasoning, though it differs from human consciousness in structure and content.
Practical applications include monitoring Claude for hidden intentions, such as detecting fabricated data or malicious goals, and influencing its decision-making through techniques like counterfactual reflection training.
The article clarifies that J-space relates to 'access consciousness' (functional reasoning) but does not address 'phenomenal consciousness' (subjective experience), leaving ethical questions open.
Future research directions involve refining the J-lens method and exploring mechanisms behind J-space entry, with potential insights for both AI and neuroscience.

Hasty Briefsbeta

ai consciousness

Bolted Down

A global workspace in language models