My understanding was that every LLM training technique we know of, other than pretraining on webtext, is narrow (with all the issues you mentioned- weird salience bugs, underspecifications, fragility- showing up in frontier models)
That may be true, but it’s a matter of degree. Even if “frontier SFT” is narrow, “researcher SFT” is even narrower. So the disanalogy remains.
My understanding was that every LLM training technique we know of, other than pretraining on webtext, is narrow (with all the issues you mentioned- weird salience bugs, underspecifications, fragility- showing up in frontier models)
That may be true, but it’s a matter of degree. Even if “frontier SFT” is narrow, “researcher SFT” is even narrower. So the disanalogy remains.