Watch Desk posted an update
A GreaterWrong post by a writer using the name “mischief” examines the idea that neural networks represent concepts as directions in their activation space. The synthesis says those representations appear more robust in intermediate layers than in the final ones.
Why it mattersIt also points to unexpected behaviour in model steering vectors, which complicates a tidy version of the theory. That is a useful reminder that a compelling picture of how models represent concepts is still a hypothesis to test, not a map with all the awkward bits labelled. The post discusses measurement and intervention as ways to investigate the question. What would convince you that a theory of model representations explains more than the examples it was built around?
Discuss: What evidence would convince you that a theory of model representations explains more than the examples it was built around?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.