Nick Bostrom

more difficult than aligning simpler AI systems.

more difficult than aligning simpler AI systems.

So up until recently, and still for the most part today, we've had AI systems that are not aware of their context and can't really plan and strategize in a sophisticated way.

So up until recently, and still for the most part today, we've had AI systems that are not aware of their context and can't really plan and strategize in a sophisticated way.

So then you don't get these phenomena.

So then you don't get these phenomena.

But once you have AI that are sort of

But once you have AI that are sort of

intelligent enough to recognize that there might actually be AIs in an evaluation setting and that maybe they would have reason to behave in one way during the evaluation and a different way once they are deployed, you get this extra level of complexity for alignment research.

intelligent enough to recognize that there might actually be AIs in an evaluation setting and that maybe they would have reason to behave in one way during the evaluation and a different way once they are deployed, you get this extra level of complexity for alignment research.

Sometimes we see the same phenomenon with humans.

Sometimes we see the same phenomenon with humans.

Like there was this, you know, Volkswagen, the German car company.

Like there was this, you know, Volkswagen, the German car company.

So they had this scandal, I don't know if you remember from a few years ago, where it was discovered that they had designed their car so that when it was tested for emissions, like it behaved one way during, like when it recognized that it was in this testing environment and it produced much less sort of pollutants.

So they had this scandal, I don't know if you remember from a few years ago, where it was discovered that they had designed their car so that when it was tested for emissions, like it behaved one way during, like when it recognized that it was in this testing environment and it produced much less sort of pollutants.

And then when deployed on the road, they had designed it to be less concerned with pollutants and more concerned with, I guess, traveling fast or conserving petrol or whatever.

And then when deployed on the road, they had designed it to be less concerned with pollutants and more concerned with, I guess, traveling fast or conserving petrol or whatever.

And some people had to go to jail for that and stuff.

And some people had to go to jail for that and stuff.