Meta’s Cox on Building AI Into the Product Suite

Yeah, of course. So, yeah, Llama 3 is about three weeks old now and we put a lot of energy into trying to have the best model that was freely available. So one that was available to the developer community, that was available to various niche verticals. So in medicine and health and science and financial services, not just in the United States and Western Europe, but in other regions, there are huge. And I think this is related to one of the conversations that you were having with Bono, which is like, what is the right way to think about the balance of equities for open source and technology? Who does it benefit? How does it harness the power of various industries and regions, and how does it affect China? I think I do think that's a conversation that is ongoing that we care a lot about. We are the primary distributor of open source models. And I've spent a lot of time over the past few weeks just hearing from the tool builders and developer ecosystem that's built on top of Llama 3. These are folks working on specific toolkits for types of developers. So in the field of law, like how are legal firms using AI to accelerate their productivity? What about banks? What about doctors? And so what's very interesting to see is that the toolkits and tool chains that get built on top of these things, as well as benefits to efficiency. So, like how cheap is it to run the model? Quantization, which is can you make the model work on smaller devices? Can you make it work on watchOS? Can you make it work on glasses?

Sure. So the first is it already has immensely over the last ten years, I'd say. So if you look at Reels, Instagram and Facebook Reels, both immensely popular, both growing a lot, those are really working not because the short form video format is special and new. It's actually been around for a long time. You remember Vine and the early days of short form video. It's because of GPUs, it's because of large models, it's because of advancements in recommendation systems. ByteDance is one of the really great companies at this. We also have made an immense investment. Our advertising system is built on modern AI recommendation technology. So I always think it's interesting when people remark that I just arrived; it's actually powering already today a huge amount of the engagement and quality of our product. So that's one thing. And also remark that the investments we're making in GPUs and data centers that Mark talked about on the earnings call will also power those services, which are just direct connections to like how relevant are the advertisements and how relevant is the product.

Those are easy to measure, but the thing that's increasingly clear is you need specific data sets for the types of things that matter, and they need to be cleaned and they need to be understood. They need to be measured. And like getting to good data fidelity and data quality is a lot of the hard work that we're focused on. We know OpenAI, Google, Microsoft, the folks playing here, a lot of the energy goes into the less easily measurable, but very toilsome and important work of data quality. Our data advantage is really the public images that have been shared on Facebook and Instagram as well as the public text. For the model, you get really amazing quality images and a lot of that is because Instagram is the dataset that was used to train it, which is just one of the great repositories of incredible imagery in a lot of different types of art, fashion, culture, and also just imagery of people and us.

Well, I think the most important thing is there's transparency around which data are being used and then there's control. And increasingly there's not a lot of transparency. And I think that's an area we all need to improve. We've been publishing where we stand. I think we've also been trying to be really outspoken. We honor robots.txt and no AI tags. So if publishers are indicating they don't want their data used in training, we don't. And I think those sorts of very clear, simple guardrails for institutions that don't want to be a part of the training are important. So that's one area that I think will be critical and will need to continue to be refined going forward.

We have advisers, with councils of sort of elders on AI and national security. We're a part of those conversations. And I agree with the point that national security and especially the sort of race between China and the West on this type of technology is incredibly meaningful. We do have terms in our license, which are legal terms which prevent misuse of Llama. So it doesn't just say go use it however you want. We've put guardrails in place in consultation, by the way, with national security advisers, cybersecurity advisers, public health advisors, etc. So we do have some legal recourse for how Llama is being used, but especially as models get closer to the frontier, especially as the investment level increases, there are arguments there. But the thing we need to remember is that open source benefits an enormous amount of people. I mean, when I joined Facebook, it was 2005 and we built the company on open source. It was Linux, Apache, MySQL, PHP. You know, Google was built on these technologies. A lot of Silicon Valley was built on top of open source technology, which because we didn't have the resources to go pay for enterprise software. China has had access to the software, too. I think one analysis that needs to be done is a really disciplined historical analysis of how has open source benefited the various communities. We worked with an open source medical database to create a tool that allows doctors in those regions to offer much more efficient decision making and support in local languages. That's incredible. And you don't get that if you're charging for a model. And so I'm not a complete zealot about open source and no markets either. I think we're trying to really be responsible about how we engage in it. But I think we need to examine really the balance of equities across not just the US and China, but also all the other regions that benefit.

Well, we're already banned. We've been banned in China for a long time. And so that's the thing. I'm not sure that could get much worse. It is what it is. You know, we're also banned in North Korea and Iran. That's a reality. And there's only so much we can do about it.

You know, I have mixed feelings about it. I haven't done the work. So first, let me answer your first question. What are we doing about it? We're just trying to build the best product. We're trying to build about recommending like needle in the haystack video or a piece of text or image to you. We have good, legitimate engagement with civil society, with government fact checking, like all the things we do to keep the product safe so that we can certify we're operating in the context of the West, we're operating in the context of the rule of law of the United States. And we are trying every year to get better at operating during elections, during war and civil strife, for young people, all the categories of areas where social media needs to be as good as possible for the world. If you ask about whether they should be banned, I haven't gone and scrutinized. I think the questions that need to be scrutinized, which is the government's responsibility. Is it being misused? And I think in America, we should be careful about making decisions on behalf of consumers about what platforms they can use. And I think those are the equities. And that's all the information I have to try to answer it.

We want to have the best product for public discourse and text. And we just passed 150 million monthly active users. We're seeing different behavior in different regions, which I think is good and right. I think that's somewhat different if you look at the platforms. The other thing that's really going to matter is we talked a lot about recommendations. Being able to recommend to people the right content. And I actually think a really good recommendation system and the software and teams behind it can over the long run get you a better product. Separate from all of that sort of performance, ease of use, language support, support for making it work in every region on every phone. Good content guidelines and standards, good partnerships, teams, business models, all the other things.

There's a really big photography community there because it's connected to Instagram accounts and the Instagram graph. Like, we don't need to go. It's not hard for Taylor Swift to immediately light up her fan base and audience and connect it to everything that she's built up on Instagram. And so it's a convenience in a lot of ways for creators for whom Instagram is in many verticals the top platform and has been for a while, the platform of choice for creators in that way. We don't think about it as like, let's go get them from somewhere else. We just think about it as let's help them bring their Instagram audience to text.

So Threads we decided to do last Thanksgiving and it was around November and we just felt that especially because there was a lot going on at Twitter or X. Yeah, let's try and build a small team to just go build something. And we were surprised, frankly, at the outpouring of love and support at its launch. And part of what we did there, and this is related to your question, is rather than sending 250 people to go work on it, to really try and find small teams of really good people that we trust and giving them a lot of space and giving them safety and giving them a chance. The Llama team is similar. The teams are pretty small if you compare. He has an eye for talent. These are some of the things that I think Mark has developed as a leader and is part of, I think, some of the ingredients of success for these last two products. He's a deep technologist. He always has been. He's fascinated by technology. I think he wants to work on the big problem. That's part of his character. It's part of who he is. And yeah, I think not just for the model and for Meta AI, but for the hardware. You know, we've been working on hardware for a while and the Ray-Ban Metas with the integration to identify what's in front of you is really badass. It's a really neat technology. Joanna Stern from the Wall Street Journal asked, 'What kind of dog is this? What do I need to feed it? What kind of health care?' and it just got them all right. And it's a wearable. It looks good, it's light, it's affordable, it's connected to a modern AI system, it's privacy safe. And it's kind of exciting to have a form factor for a wearable piece of hardware that actually works.

Cool. Thanks.