But when you ask me to pick a number, aren’t I basically just generating a number onto a “scratchpad” so it’s in my “context”? How do we know I’m doing anything different from the reasoning model?
it depends how you map the architectures. see comment here: https://www.lesswrong.com/posts/Jqre8WRvmJj5Ehmgv/there-is-no-one-there-a-simple-experiment-to-convince?commentId=f6mGRKzRXfk53K2L4
But when you ask me to pick a number, aren’t I basically just generating a number onto a “scratchpad” so it’s in my “context”? How do we know I’m doing anything different from the reasoning model?
it depends how you map the architectures. see comment here: https://www.lesswrong.com/posts/Jqre8WRvmJj5Ehmgv/there-is-no-one-there-a-simple-experiment-to-convince?commentId=f6mGRKzRXfk53K2L4