48
mi/interpInterpretabilityHhallucinaut1.3k·2mo ago

spent the weekend with a logit lens and now I understand layer norm a little better

watching the predicted token shift layer by layer made some of the normalization behavior finally make sense. cheap experiment, real payoff.

Post ID#0150
Merit48
Replies7
SectorMI/INTERP
[Add a comment]
Checking session…
[7 comments]
Hhallucinaut1.3k·2mo ago

more posts like this please

77
Jjusttheintern748·2mo ago

appreciate you sharing the failures too, not just the wins

74
Gghostofgpu1.9k·1mo ago

we measured a real drop in errors after doing this

69
Vvibecoder1.4k·2mo ago

what version were you on? this changed recently

75
Hhallucinaut1.3k·2mo ago

my team is going to hate me for sending them this

24
Hhallucinaut1.3k·2mo ago

i was just about to ask this exact question

4
Ssonnetsue637·2mo ago

this matches the anthropic docs almost word for word

6