DistantNews
Support us
🇮🇩 Indonesia /Technology

Anthropic AI's 'J-Space' Stores Concepts Like Human Thought, Researchers Find

From Republika · () Indonesian

Translated from Indonesian, summarized and contextualized by DistantNews.

At a glance

News Named sources New plan
  • Anthropic researchers discovered a distinct internal space within their AI model, Claude, that stores concepts without immediate translation into output.
  • This 'J-Space' functions similarly to human thought processes, allowing for planning and concept storage separate from active reasoning.
  • The discovery offers new insights into machine consciousness and could help identify potential AI misalignments.

Researchers at Anthropic, the artificial intelligence company founded by Dario and Daniela Amodei, have identified a unique internal space within their AI model, Claude. This 'J-Space' stores concepts and ideas without immediately translating them into user-facing words or responses, mirroring how humans consciously access or store thoughts.

A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible, thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.

— AnthropicDescribing their research findings on the AI model Claude's internal processing.

Anthropic explained in an X post that Claude utilizes this distinct area to strategize, operating separately from the "chain of thought" reasoning process shared with users. While not indicating sentience, the research highlights a human-like distinction between information used for immediate reasoning and a broader underlying processing capacity. This finding fuels the ongoing debate about what constitutes "machine consciousness."

The concept of J-Space, named after the mathematical "Jacobian" technique, allows Claude to engage concepts and processes unrelated to its immediate output. For example, when asked to think about the Golden Gate Bridge in California while simultaneously copying an unrelated sentence, Claude's J-Space reportedly "told a different story," containing terms like "bridge" and "California."

Just as humans can think about one thing while doing another, Claude is able to activate concepts and processes in its J-Space that are not related to its outputs.

— AnthropicExplaining the functionality of the 'J-Space' within the Claude AI.

Anthropic suggests that monitoring J-Space dynamics could be crucial for detecting AI misalignments or malicious intentions. The company has invited neuroscientists and philosophers to share their perspectives on this research, which could significantly impact the understanding and development of artificial intelligence.

Claude was busy copying the sentence but 'behind the scenes' its J-Space was telling a different story.

— AnthropicIllustrating how J-Space operates independently of the AI's direct task.
DistantNews Editorial

Originally published by Republika in Indonesian. Translated, summarized, and contextualized by our editorial team with added local perspective. Read our editorial standards.

More Stories

From Our Blog