Mark Zuckerberg

Okay, well just, you can distill from both of them and then build something that's better than either of them for your own use case.

So that's cool.

But you do need to solve the security problem of knowing that you can distill it

in a way that is safe and secure.

And so this is something that we've been researching and have put a lot of time into.

And what we've basically come to is like, look, anything that's kind of like language is quite fraught because there's like a lot of values embedded in that.

So unless you don't care about having the values from whatever the model is that you got, you probably don't want to like distill the straight like language world model.

On reasoning, I think you can get a lot of the way there by limiting it to verifiable domains, running kind of code cleanliness and security filters, like whether it's like the LamaGuard open source or the CodeShield open source things that we've done that basically allow you to incorporate different input into your models and make sure that both the input and the output are secure.

And then just a lot of red teaming to make sure that you're,

You just have people who are experts who are looking at this.

It's like, all right, is this model doing anything that isn't what I want after distilling from something?

And I think with a combination of those techniques,

you can probably distill on the reasoning side for verifiable domains quite securely.

That's something I'm pretty confident about.

And it's something that we've done a lot of research around.

But I think this is a very big question is like, how do you do good distillation?

Because there's just so much value to be unlocked.

But at the same time, I do just think that there is some fundamental bias in the different models.

Speaking of value to be unlocked,

I mean, like we were talking about before, there's going to be all these different applications and different applications tend towards different things.