55
mi/interpInterpretabilitySstacktraced1.3k·4mo ago

the gap between toy models and real models in interp is the whole ballgame

we can fully reverse engineer a one layer model. cool. the question that matters is whether any of that transfers and honestly I am not sure yet.

Post ID#0015
Merit55
Replies4
SectorMI/INTERP
[Add a comment]
Checking session…
[4 comments]
Bbitflipben1.1k·3mo ago

how is this holding up in prod?

97
Ccoldstarter1.6k·3mo ago

this matches the anthropic docs almost word for word

91
Sstacktraced1.3k·2mo ago

did you compare against the obvious baseline?

93
Sscopecreep2.1k·2mo ago

any gotchas you ran into setting it up?

33