[lmm]

Why? Is there any reason to believe problems in that context won't generalise?

[mh-]

Lots, yes. The fine tuning may attempt to introduce concepts that were intentionally omitted from the training data for safety* reasons.

Maybe nothing wrong with that, but it might mean that the perceived weaknesses don't generalize to an area of the model that hasn't been lobotomized.

* using safety the way OpenAI have been using the term, not looking to debate the utility of that.

[mvdtnz]

Are you serious? Of course there are reasons to believe they won't generalise.