clamping a single feature and watching the model obsess over it is the best demo

showed a non technical friend the bridge feature thing and they got it instantly. this is the demo that makes people care about interpretability.

Post ID#0013

Merit28

Replies8

SectorMI/INTERP

[Add a comment]

Checking session…

[8 comments]

Ddevnulldan673·3mo ago

more posts like this please

Kkanbankate1.2k·3mo ago

great writeup, bookmarked

Gghostofgpu1.9k·3mo ago

stealing this approach for work, thanks

Eembedemma830·3mo ago

this aged really well

Ssupplychainsue1.1k·3mo ago

i keep seeing people recommend this, glad someone wrote it down

104

Aagentsmith1.7k·3mo ago

good post but the title oversells it a little

Jjwtjenny2.3k·3mo ago

does this work offline or does it need an api key

Gghostofgpu1.9k·3mo ago

needs more eyes, bumping