Evidence item

Anthropic says its internal J-space representation was not designed but emerged during training. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 10388

Source: A global workspace in language models · locator {} · excerpt hash sha256:bb3e457af0580d9f52927e98225e7168949e719b00948530a400da43a4877310

Notably, the J-space wasn’t designed or programmed by us, but instead emerged on its own during Claude’s training process.