Yoshua Bengio

What do you think?

Well, so now I get much more honest responses.

Otherwise, it's all like perfect and nice and it's going to work.

If it knows it's me, it wants to please me, right?

If it's coming from someone else, then to please me, because I say, oh, I want to know what's wrong in this idea, then it's going to tell me the information it wouldn't.

Now, here it doesn't have any psychological impact.

It's a problem.

The sycophancy is a real example of misalignment.

We don't actually want these AIs to be like this.

I mean, this is not what was intended.

And even after the companies have tried to tame a bit this, we still see it.

So it's like we haven't solved the problem of instructing them in the ways that are really according to, so that they behave according to our instructions.

And that is the thing that I'm trying to deal with.

Yes, yes.

Even though that is not what you want.

That is not what I wanted.

I wanted honest advice, honest feedback.

But because it is sycophantic, it's going to lie, right?

You have to understand it's a lie.

Do we want machines that lie to us even though it feels good?