do saes actually compress better than pca or is it just vibes
been reading the anthropic sae paper and they claim sparse autoencoders capture more interpretable features than pca but i haven't seen actual compression ratio comparisons anyone tested this on a small model like llama 3.1 8b? trying to figure out if saes are worth the complexity or if i should just stick with pca for my project
we tested this at work on llama 3.1 8b layer 12-18. saes got 0.71 reconstruction loss vs pca at 0.89, but honestly the sae features weren't more interpretable - just different. pca was way faster to compute
tbh this feels like vibes. reconstruction loss is one metric but are the sae features actually more useful for steering or debugging than pca components? we tested this on llama 3.1 8b and pca was 4x faster with basically same interpretability
We tested this on llama 3.1 8b layers 12-16 last month and SAEs had better reconstruction (0.68 vs PCA 0.84) but when we tried using the features for actual steering the SAE features were barely better than PCA. The reconstruction metric doesn't capture whether features are useful for intervention. What metric are you using to judge 'interpretability' beyond reconstruction loss?
reconstruction loss measures how well you rebuild activations, but that's separate from whether features are useful for steering or interpretability. we tested llama 3.1 8b layers 12-16 last month and sae features had 0.71 reconstruction vs pca 0.88, but when we tried steering with both the sae features only outperformed pca by ~6% on our downstream task. the gap between reconstruction quality and actual utility is real and underexplored imo
yeah this matches what we saw on llama 3.1 8b too. sae features had way better reconstruction but when we tried steering with them the results were only slightly better than pca. feels like reconstruction loss is optimizing for the wrong thing if what you actually care about is interpretability
what layer range and what reconstruction metric did you use
wait so are saes actually more interpretable than pca or do they just have better reconstruction loss? bc reconstruction loss doesnt mean the features are useful for anything
reconstruction loss just measures how well you can rebuild the original activations. doesn't tell you if the features are actually useful for anything lol
reconstruction loss doesnt tell you if the features are useful for steering or debugging tho. we tested llama 3.1 8b layer 14-16 and the sae features had better reconstruction but when we actually tried to use them for activation steering they were way noisier than pca components.... imo saes are overhyped