Mark Zuckerberg

Does it not speak French well?

It's like, no, it speaks French fine.

It's just like the way that it thinks about the world is like, seems slightly American.

So I think there's like these subtle things that kind of get built into it.

Over time, as the models get more sophisticated, they should be able to embody different value sets across the world.

So maybe that's like a very kind of...

you know, not particularly sophisticated example, but I think it sort of illustrates the point.

And, you know, some of the stuff that we've seen in testing some of the models, especially coming out of China is like, they sort of have certain values encoded in them.

And it's not just like a light fine tune to get that to feel the way that you want.

Now the stuff is different, right?

So I think language models,

Or something that has like a kind of like a world model embedded into it have more values.

Reasoning, I think, is, I mean, I guess there are kind of values or ways to think about reasoning.

But one of the things that's nice about the reasoning models is they're trained on verifiable problems.

So do you need to be worried about like cultural bias if your model is doing math?

Probably not, right?

I think that that's, you know, I think it's like the chance that like some reasoning model that was built elsewhere is like going to kind of incept you by like solving a math problem in a way that's devious, seems low.

There's a whole set of different issues, I think around coding, which is the other verifiable domain, which is, you know, I think you kind of need to be worried about

like waking up one day and like, does a model that I have some tie to another government, like can it embed all kinds of different vulnerabilities in code that then like the intelligence organizations associated with that government can then go exploit.

So now you sort of like, all right, like in some future version where you have, you know, some model from some other country that we're using to like,