Mark Zuckerberg

projects that are based on this i mean there's one at berkeley there's you know it's just like all over and um and people have tried a lot of different things and we've tried a bunch of stuff internally so kind of where we're we're making progress

but also we're able to learn from some of the best ideas in the community.

And, you know, I think it, you know, we want to just continue, continue pushing that forward, but I don't have any news to announce on, if that's, if that's what you're, you're asking.

I mean, this is a thing that we're, uh,

We're still kind of actively working through the right way to move forward here.

Yeah, I think that's a really interesting idea that I've talked to Jan about a bunch.

And we were talking about how do you basically train these models to be as safe and aligned and responsible as possible.

And different groups out there who are doing development test different data recipes and fine-tuning.

But this idea that you just mentioned is...

that at the end of the day, instead of having kind of one group fine-tune some stuff and another group produce a different fine-tuning recipe and then us trying to figure out which one we think works best to produce the most aligned model, I do think that it would be nice if you could get to a point where you had a Wikipedia-style collaborative

way for a kind of a broader community to, um, to, to fine tune it as well.

Now there's a lot of challenges in that, both from an infrastructure and like a community management and product perspective about how you do that.

So I, I haven't worked that out yet.

Um, but, but I, as an idea, I think it's, it's quite compelling and I think it, it goes well with the ethos of open sourcing.

The technology is also finding a way to have a, a kind of community driven, um,

community-driven training of it.

But I think that there are a lot of questions on this.

In general, these questions around what's the best way to produce aligned AI models, it's very much a research area.

And it's one that I think we will need to make as much progress on as the kind of core intelligence capability of the models themselves.

A lot of the time.