“New research from Anthropic claims to have found a division inside its Claude model between a small layer of reportable thoughts and a larger volume of automatic processing, which it calls a "global workspace."”