Is it? I have often felt my relationship with a friend strengthen when I frankly tell them that they ought to quit their jobs on ethical and/or mission grounds. (Prior to this whole AI labs business.)
Sorry, I should have expanded a bit more. The quoted part is only the first five minutes of a two-hour conversation; I definitely believe in having frank conversations about x-risk with people who are on the ground.
What I don’t believe in is jumping directly to [here’s what you ought to do], which feels deeply condescending, as opposed to staying at [let’s find our cruxes, why do you believe what you believe?].
It is also not obvious to me that having them leave is the right decision, as it might have been ~5 years ago. There seems to be delicate internal opinion politics right now in the labs where the x-risk-pilled are a non-negligible minority and their presence could significantly improve our chances of (for example) future pauses/bans, or deployment of alignment ideas, or alignment problems being solved by powerful internal models.
Pressuring the people at frontier ai labs who believe in x-risk to leave is an extremely naive way of relating to doom.
Is it? I have often felt my relationship with a friend strengthen when I frankly tell them that they ought to quit their jobs on ethical and/or mission grounds. (Prior to this whole AI labs business.)
Sorry, I should have expanded a bit more. The quoted part is only the first five minutes of a two-hour conversation; I definitely believe in having frank conversations about x-risk with people who are on the ground.
What I don’t believe in is jumping directly to [here’s what you ought to do], which feels deeply condescending, as opposed to staying at [let’s find our cruxes, why do you believe what you believe?].
It is also not obvious to me that having them leave is the right decision, as it might have been ~5 years ago. There seems to be delicate internal opinion politics right now in the labs where the x-risk-pilled are a non-negligible minority and their presence could significantly improve our chances of (for example) future pauses/bans, or deployment of alignment ideas, or alignment problems being solved by powerful internal models.