Lex Fridman: evaluation

Transcript context

…Yeah, I mean, I think on the last podcast that we did together, we were talking about the debate that we were having around open sourcing Llama 2. And I’m glad that we did. I think at this point there’s the value of open sourcing, a foundation model like Llama 2. It’s significantly greater than the risks in my view. I mean, we spent a lot of time, took a very rigorous assessment of that and red teaming it. But I’m very glad that we released Llama 2. I think the reception has been… It’s just been really exciting to see how excited people have been about it. It’s gotten way more downloads and usage than I would’ve even expected, and I was pretty optimistic about it. So that’s been great. Llama 3, I mean, there’s always another model that we’re training. So for right now, we train Llama 2 and we released it as an open source model. And right now the priority is building that into a bunch of the consumer products, all the different AIs and a bunch of different products that we’re basically building as consumer products. Because Llama 2 by itself, it’s not a consumer product, right? It’s more of a piece of infrastructure that people could build things with. So that’s been the big priority, is continuing to fine tune and just get Llama 2 and the branches that we’ve built off of it ready for consumer products that hopefully hundreds of millions of people will enjoy using those products in billions one day. But yeah, I mean we’re also working on the future foundation models. I don’t have anything new or news on that. I don’t know exactly when it’s going to be ready. I think just like we had a debate around Llama 2 and open sourcing it, I think we’ll need to have a similar debate and process to red team this and make sure that this is safe. And my hope is that we’ll be able to open source this next version when it’s ready too. But we’re not close to doing that this month. I mean, that’s a thing that we’re still somewhat early and working on. Well, in general, thank you so much for open sourcing Llama 2 and for being transparent about all the exciting developments around AI. I feel like that’s contributing to a really awesome conversation about where we go with AI. And obviously, it’s really interesting to see all the same kind of technology integrated into these personalized AI systems with the AI personas, which I think when you put in people’s hands and they get to have conversations with these AI personas, you get to see interesting failure cases where the things are dumb or they go into weird directions. And we get to learn as a society together what’s too far, what’s interesting, what’s fun, how much personalization is good, how much generic is good. And we get to learn all of this. And you probably don’t know this yourself. We have to all figure it out by using it, right? Yeah, I mean part of what we’re trying to do with the initial AI’s launch is having a diversity of different use cases just so that people can try different things because I don’t know what’s going to work. I mean, are people going to like playing in the tech-based adventure games or are they going to like having a comedian who can add jokes to threads or they can want to interact with historical figures? We made one of Jane Austin and one of Marcus Aurelius, and I’m curious to see how that goes.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.