POPA ReconA global workspace in language models
A global workspace in language models Anthropic researchers have discovered an emergent internal neural workspace in Claude called the "J-space," which allows the model to silently reason, maintain internal goals, and flexibly process complex tasks similarly to the global workspace theory of human cognition https://www.anthropic.com/research/global-workspace
The excerpt is quoted from a public Telegram channel, which owns its text.