Inside Claude’s J-Space: What Anthropic’s Global Workspace Discovery Really Means

The Signal

Anthropic has spent years arguing that the only way to make AI safe is to understand what is happening inside the model, not just at the output layer. On Sunday, the company published a 16-author research paper titled “Verbalizable Representations Form a Global Workspace in Language Models” and claimed it had found something nobody expected: a compact, privileged internal workspace inside Claude that the researchers call “J-space.” Using a new mathematical technique called the...