This is a special post for quick takes by Ben Livengood. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Comment Permalink
*53
Reading through what it can and can't do, this feels more like "mech interp is more advanced than we give it credit for" than "Astra can do something humans can't". In all of the cases where it succeeds, I expect a human with enough free time could do it too, and the problems that it can't solve are the ones we can't solve either.
Related, humans writing NNs by hand: https://vgel.me/posts/handmade-transformer/ https://yanivle.github.io/taotp.html https://github.com/anadim/AdderBoard#hand-coded-weights-constructive-proofs https://srush.github.io/raspy/
1
1