@keshavchan : when we use SAEs to understand models, we’re trading off fidelity for insight. we make the model’s thoughts legible by compressing them into human friendly pieces, but we inevitably leave behind some of the richness that made those thoughts possible in the first place. i wonder • TwiCopy

keshav

@keshavchan

+ Follow

data + design @ sundial / ex @every

ID: 1324599051170373632

calendar_today06-11-2020 06:27:01

14,14K Tweet

18,18K Followers

1,1K Following

keshav

@keshavchan

2 months ago

when we use SAEs to understand models, we’re trading off fidelity for insight. we make the model’s thoughts legible by compressing them into human friendly pieces, but we inevitably leave behind some of the richness that made those thoughts possible in the first place. i wonder

thumb_up_off_alt2

chat_bubble_outline0

repeat0

shareShare