Got the pointer to this from Allison Parrish who says it better than I could:
it’s a very compelling paper, with a super clever methodology, and (i’m paraphrasing/extrapolating) shows that “alignment” strategies like RLHF only work to ensure that it never seems like a white person is saying something overtly racist, rather than addressing the actual prejudice baked into the model.
Interesting results, interesting insights, neat takeaways and argumentation.
It’s unfortunate they only tested models that were trained on SAE and they didn’t have a control group of language models in other dialects. Seems like a huge oversight.
I wonder how this would play out with a model that has been trained on AAE, another non-SAE dialect, or even one trained in English but optimized for a non-english language.