Anthropic Finds AI 'J-Space' That Mirrors Human Conscious Processing
The discovery of J-space offers a concrete, testable model of 'access consciousness' in large language models, potentially bridging AI safety research and neuroscience.
Reporting from 1 source: GIGAZINE.
Researchers at Anthropic have identified a distinct neural workspace in their Claude AI model, dubbed 'J-space,' that activates when the model performs complex reasoning tasks. This workspace emerged naturally during training, not from explicit design. Removing J-space severely degraded Claude's ability to perform multi-step reasoning, summarization, and poetry, while simple tasks remained unaffected. The discovery provides a new tool for studying AI cognition and its parallels to human conscious processing.
J-space is not a programmed feature but a spontaneously formed structure within Claude's neural network. It holds information beyond the immediate text, such as identifying errors in code or detecting prompt injection attacks. In experiments, manipulating the contents of J-space directly changed Claude's reported thoughts. Anthropic notes that J-space is limited in capacity, holding only a few dozen concepts at a time, and accounts for less than a tenth of the model's total activity. The company emphasizes that this does not mean Claude experiences consciousness like a human, but it may provide insights into access consciousness.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.