This doesn’t provide comfort that the head of the most advanced AI lab is making vague and somewhat laissez faire comments about his proposed direction (or really non-direction) of AI safety work.
“I kind of think you can’t make progress on the safety part of this, as we mentioned earlier, by sitting in a room and thinking hard. You’ve got to see where the technology goes. You’ve got to have contact with reality. ”
This almost seems like he’s implying that he thinks safety and alignment should always lag behind capabilitiies and be in catchup mode, while we watch the technology advance ahead of our scope of knowledge and understanding rather than safety measures being ahead of the game and defining what capabilities steps are safe to take next.
Also how else do you do alignment work other than sitting in a room and thinking (or coding or computing) hard? Maybe I’m just not clear what he means here.
And again with
“But it’s all like, for me, it’s all this same thing of like, we’re trying to solve the problems that emerge at each step as we get what we’re trying to go.”
Again this suggests he is endorsing a reactive approach, like put the model out, then look for problems, then fix them. I know little about AI failure modes, but most that people suggest seem like if they are already out of the box it may well be too late.
I would feel much more comfortable if he was talking along the lines that he wants to identify potential problems before they acure.
“But I think calling for a pause is like naive at best.”
“Naive at best” seems unnecessarily harsh and dismissive. Its fine to disagree with something and make object level complaints without calling it (and by proxy those who argue for it) naive. Many AI experts on the forum here and elsewhere who are unlikely to be straigh-up naive (at best, I’m not sure what he is suggesting is worse than naive) have made well reasons and thorough arguments for a pause of some sort. He could make the effort to engage with their arguments rathe
Also as a side note it’s a little slightly ironic that he followed up the accusation of “Naive at best” with the rebuttal that started with “I kind of think”. I would hope for more confidence if you are accusing others of naievity.
Thanks that’s a good point I don’t have context. Although I think the heads of these companies should be more careful then about what they wax lyrical in a podcast.
This doesn’t provide comfort that the head of the most advanced AI lab is making vague and somewhat laissez faire comments about his proposed direction (or really non-direction) of AI safety work.
“I kind of think you can’t make progress on the safety part of this, as we mentioned earlier, by sitting in a room and thinking hard. You’ve got to see where the technology goes. You’ve got to have contact with reality. ”
This almost seems like he’s implying that he thinks safety and alignment should always lag behind capabilitiies and be in catchup mode, while we watch the technology advance ahead of our scope of knowledge and understanding rather than safety measures being ahead of the game and defining what capabilities steps are safe to take next.
Also how else do you do alignment work other than sitting in a room and thinking (or coding or computing) hard? Maybe I’m just not clear what he means here.
And again with
“But it’s all like, for me, it’s all this same thing of like, we’re trying to solve the problems that emerge at each step as we get what we’re trying to go.”
Again this suggests he is endorsing a reactive approach, like put the model out, then look for problems, then fix them. I know little about AI failure modes, but most that people suggest seem like if they are already out of the box it may well be too late.
I would feel much more comfortable if he was talking along the lines that he wants to identify potential problems before they acure.
“But I think calling for a pause is like naive at best.”
“Naive at best” seems unnecessarily harsh and dismissive. Its fine to disagree with something and make object level complaints without calling it (and by proxy those who argue for it) naive. Many AI experts on the forum here and elsewhere who are unlikely to be straigh-up naive (at best, I’m not sure what he is suggesting is worse than naive) have made well reasons and thorough arguments for a pause of some sort. He could make the effort to engage with their arguments rathe
Also as a side note it’s a little slightly ironic that he followed up the accusation of “Naive at best” with the rebuttal that started with “I kind of think”. I would hope for more confidence if you are accusing others of naievity.
Have you read OpenAI’s blog post “Introducing Superalignment”? In light of that post, how would you interpret Altman’s comments?
Based on the fact that OpenAI is thinking about AGI safety and alignment, I’m inclined to give him a bit more benefit of the doubt.
Thanks that’s a good point I don’t have context. Although I think the heads of these companies should be more careful then about what they wax lyrical in a podcast.