"Paul Christiano runs the Alignment Research Center, a non-profit research organization whose mission is to align future machine learning systems with human interests. Paul previously ran the language model alignment team at OpenAI, the creators of ChatGPT."
The start of this video on the Bankless channel is shocking, especially given the position that Christiano previously held:
'overall maybe you are getting like 50/50 chance of doom shortly after you have AS systems that are at human level.'
Christiano doesn't get too much more optimistic, as you can imagine from the above prediction:
'My default picture is like we have time to react in terms of the nature of AI systems changing, their capabilities changing. With luck we have some various kinds of smaller catastrophes occurring in advance, but I think that one of the bad things about the actual catastrophe we are worried about, does have these dynamics similar to like a human coup or revolution, where we don't have little baby coups and we see like the rate of how these coups occur, they may just go like straight to... the ship has sailed once AIs start taking over...I think you will probably have like years between people say 'that looks like a takeover risk' and when an actual takeover occurs. And that's pretty good and that's why I'm a lot more optimistic... I'm like, well, I think people are wrong about the rate of progress, I think people will be able to see things that can generally be recognised as pretty concerning, prior to the actual catastrophe.'
On Alignment Christiano continues:
'The problem is really hard... Three categories that I would probably think about in terms of addressing this are like technical measures that can reduce the risk of takeover. measurements that can inform us about the risk of takeover and understand what are the relevant dynamics that can also involve policy intervention.
Should the predictions prove even vaguely accurate then the questions arise about our responses. What is the state of our current systems and institutions to act decisively, in a coordinated and timely manner in proportion to the identified threats, in a short timeline?
My own considerations lead me to hypothesis that any 'intelligence' from an AGI will be an alien intelligence, it will certainly mimic a human, but emotional intelligence will be lacking, social intelligence non existent, intrapersonal, or interpersonal intelligence very limited. This is a distinct danger that we, as a species, fall for the mimicry in the desire to find a 'god'.
Comments
Post a Comment