Narrow finetuning can produce broadly misaligned LLMs (opens in new tab)

Is teaching a child? Is talking with your friend? Is punishing a criminal?

Or, more pointedly, what about training the model in the first place? Why do you pretend that AI are somehow "people" with a "natural tendency" we're overriding?