Anthropic on Claude's global workspace

[Alan Downing]

An Anthropic paper + video summary describes how Claude uses an internal reasoning workspace that is similar in some ways to the internal processing in human consciousness:

https://www.anthropic.com/research/global-workspace

Claude's global workspace has the following properties:

• Claude tell you what it is thinking about by looking in its workspace
• Claude can put things into its workspace upon request
• Claude uses the workspace for internal reasoning
• Representations in the workspace can be flexibly used for many tasks
• The workspace is not involved in most of what the language model does. Speaking fluidly, recalling simple facts, etc. are lower-level cognitive functions.

If this type of stuff is of interest to you, at least take a look at their video.

Enjoy,

Alan D.

[Alan Downing]

Oh, I forgot to mention something important: The workspace wasn't programmed by Anthropic, but instead it emerged on its own during Claude's training process.