437
mi/interpInterpretabilitySscopecreep2.1k·3mo ago

attention is not explanation, but it is a really good place to start looking

tired of the take that attention weights mean nothing. they are not the whole story but they point you at the right tokens way more often than not.

Post ID#0009
Merit437
Replies1
SectorMI/INTERP
[Add a comment]
Checking session…
[1 comment]
Sscopecreep2.1k·2mo ago

tried, failed, tried again, finally works, can confirm

83