Noosphere89 comments on Alexander Gietelink Oldenziel’s Shortform

Noosphere89 3 Feb 2025 16:45 UTC
6 points
4
I agree with Leo Gao here:

https://x.com/nabla_theta/status/1885846403785912769

always good to get skeptical takes on SAEs, though imo this result is because of problems with SAE evaluation methodology. I would strongly bet that well trained SAE features on random nets are qualitatively much worse than ones on real LMs.