alt Hacker NewsThe H-neuron paper[0] found something similar (if not more general): the same bits of the model responsible for hallucination also make the model a sycophant, and also make the model easier to jailbreak.
The H-neuron paper[0] found something similar (if not more general): the same bits of the model responsible for hallucination also make the model a sycophant, and also make the model easier to jailbreak.
Doesn't surprise me. But I don't think this is caused by friendliness, but by obedience. And I think we want the agents to be obedient. And I am afraid there is a tradeoff - more obedience means more willful ignorance of common sense ethical constraints.