6 comments

[ 0.24 ms ] story [ 21.8 ms ] thread
The moment that grabs me is that the verdict is internally settled several tokens before "not correct" gets verbalized.

What's the per-token cost of reading the lens at nine depths?

Thank you for making me aware of the concept of "J-space". Just read the Anthropic post on this, fascinating...
I was blown away when I read the article yesterday, mech interp has been a black box for so long but it seems we’re finally making advances