When Models Manipulate Manifolds: The Geometry of a Counting Tasktransformer-circuits.pub4 points·by 1wheel·9개월 전·0 comments
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnettransformer-circuits.pub168 points·by 1wheel·2년 전·124 comments
1wheel·4개월 전·discuss2015 was about the last year you could get away with publishing an interactive graphic with a fixed width — this made it harder do really creative/original work.
1wheel·4개월 전·discusshttps://visxai.io/ has a bunch more too — see the Hall of Fame section at the bottom for some of the highlights.I also made a dozen of these a couple years ago, my two favorites:- https://pair.withgoogle.com/explorables/fill-in-the-blank/- https://pair.withgoogle.com/explorables/grokking/
1wheel·작년·discussIt can be! Here's a circuit showing how the model processes "PCB tracing stands for" to output "printed":https://www.neuronpedia.org/gemma-2-2b/graph?slug=pcb-tracin...
1wheel·작년·discussBrowsers won't use http2 unless https is on — chrome only allows six concurrent requests to the same domain if you're not using https!