The Most Expensive Hire In AI History Finally Talks

Yeah, excited to be here. I feel like we were, we kind of texted a little bit pre, pre Meta happenings. We had, we had like this kind of country music odd text chain going, and then, and then I'm also, I've known Nat Friedman for a long time.

Yes, we were very, very quiet there.

Yeah.

Yeah.

We, uh, yeah, we had a lot of work to do. I mean, I think, uh, turns out, um, building a frontier model from scratch in nine months is, uh, yeah, it takes a lot of, uh, takes a lot of painstaking effort. But, um, but yeah, I mean, it's been really exciting to see everyone use Muse Spark, um, you know, the model we released, and, um, we have better models cooking. So, it's exciting.

I moved down to South Bay. Yeah. Yeah, I did.

I'm full on. Yeah. And for me, the city now is Palo Alto.

Walk on University Ave, get a boba.

Yeah. So, um, so basically, uh, the, the whole unit is called Meta Super Intelligence Labs, um, which, which I oversee, and then there's various pieces of it. So, um, there's a unit called TBD, which is, um, the, uh, sort of large model research lab, um, that I think, uh, you know, is somewhat infamous, but that's where a lot of the, um, you know, leading researchers and infrastructure engineers are. Um, they actually all technically report to me. Um, so that's one setup. Then there's also, um, a group called, uh, Product and Applied Research, or PAR for short. That's what, uh, Nat Friedman heads up. So they're responsible for, um, all the products that we build, um, and the actual sort of deployment of these great models to the world. And then, um, also within the, uh, overall Meta Super Intelligence, uh, Labs umbrella is FAIR, uh, which continues to do exploratory and exciting research. Um, I'm particularly excited about a lot of their scientific research. So, you know, we've shown some pretty great work on, um, you know, using models, AI models to understand the brain, as well as using AI models to understand, you know, computational chemistry. We have built like a universal model for atoms, um, for short. And then, um, and so that, those pieces constitute Meta Super Intelligence Labs, which I oversee in addition to having a very hands-on role with, um, TBD Lab. And then, um, Daniel Gross, uh, helps lead up Meta Compute, which is really focused on our long-term infrastructure planning to ensure that, you know, we obviously can, um, build up all of the GPU infrastructure and, and data center infrastructure necessary for this very bold endeavor. Um, and, uh, and so he heads that up and partners closely with us.

Um, well, uh, I knew, I've known Nat and Daniel actually for a long time. Nat was one of our very first, one of my very first angel investors at Scale. Um, and I think like in, before I completed YC, Nat had invested in Scale and had given me advice throughout the years. Um, Daniel, I think I also met around that time, um, you know, very, very early on and have gotten to know them through the years. And then, um, we also have our chief scientist, Shengjia, who, um, helps oversee the scientific agenda across all of MSL. And, um, he is somebody who I, uh, have gotten to know, you know, I knew before starting MSL, but, you know, since he's come in, we've gotten a lot closer. So.

Yeah, so it was, so I've known Mark, um, for many years now. I think, uh, even while I was running Scale, I would, um, he was very generous with his time and I sort of was able to get a bunch of advice from him. And he, um, obviously is just such an experienced, you know, founder and like the, the founder at this point in some, in some sense. Um, and, uh, and so we've known each other for many years and, and we had actually talked about, um, AI before a lot of this craze because, you know, Scale, obviously, we'd been working on AI since 2016. Um, and, uh, you know, back when it was mostly self-driving and then, you know, the various transitions of the technology. Um, and then around a year ago, almost like literally a year ago, um, you know, we had a conversation where we sort of, um, started exploring, uh, if there were a way to work more closely. Um, and, um, in particular, I think Mark at that point, I think he, um, was, uh, becoming increasingly AGI-pilled and, um, and really knew that first AI was going to totally transform Meta, but also that AI was one of these, um, you know, almost like once-in-a-lifetime transformative technologies. And so he was really, um, quite focused on it and, and knew he wanted to bet very big on it. And, um, at the same time, you know, uh, you know, and he's talked about this publicly, like Llama was not on the trajectory that, um, that they, that the company needed to be able to, um, continue making some of these bets. And so we were sort of talking at a very high level about, you know, how could we work together more closely? What could that look like? And, um, obviously it kind of, you know, it was one of these very open-ended questions and these very open-ended brainstorm sessions as these things often are. And it sort of landed in this interesting zone where we figured out a way to do it in a way, in a way that was like good for Scale, that was good for Meta, and where we got to work very, very closely together to build out, you know, the most important technology of our time and do so in a way that, um, where we both had conviction that it was going to be, you know, we're building something that we'd both be really proud of, which is, you know, he put out this memo of personal superintelligence also about a year ago. Um, and then we went quiet, obviously, but, um, but I think that really is the sort of like north star for both of us is we want to build this technology in a way that empowers people, where, you know, as many people in the world have access to it and it's as democratized as possible, that enables you, everyone to express themselves, everyone to have increased agency, everyone to create and build. Um, and that's really the world that we want to work towards.

Yeah. Yeah. Very different. It's super different. And, um, you know, a lot of what, what I was thinking about throughout this process is obviously I think like everyone in and around AI, I mean, progress has just happened, I think a lot faster than I, I'd expected for a long time. Um, and, um, in the sort of like accelerating progress of these AI models, I think a few things really started to stick with me. One is, you know, um, it felt like, and I do think this is increasingly the case, that, um, those who build the AI models have, uh, just greater and greater rights, so to speak, or, or both economic and product rights to just build so much more around those models. And there were obviously all these early debates around, you know, what are, um, how does the ecosystem play out? And I think because of just how fast the models are improving and how fast research, what the, how fast the research pace is, it just means that, you know, in many ways, building, being a place that is building the models is, are some of the most exciting places to be in the ecosystem. Um, and then the second is really that, that so much of this next phase of technology really boils down to compute. And, um, you know, if you have lots and lots of compute, then you have the ability to build things and make big bets and deploy products and do things that you just can't if you don't have that compute. And I think that, um, this causes, this will cause an interesting stratification for the tech ecosystem where, you know, right now in some sense we think about, well, I think this is already changing, but, um, you know, you kind of think about all tech companies the same, but in reality you should think about companies with lots of compute very differently from companies without compute because, um, there's just things that companies with compute can build that those without compute just cannot. And so it, it creates this very interesting dynamic. And so part of what was very exciting about the opportunity at Meta first is that obviously Mark is, um, I think very all-in on AI and really, um, has bet very big and is a very bold, um, sort of leader and strategist, but also that, you know, this created the conditions where you, we're able to build with huge amounts of compute, um, and with the right research effort and the right product efforts, we have the ability to really make a huge dent in the world.

Yeah. So, you know, um, when I got, when I kind of got to Meta, um, you know, it was, it was clear that there needed to be some reset of the efforts and, and some rebuild of our AI, AI efforts to get onto the right trajectory because ultimately, like, you know, Llama was not on the same trajectory and so we were behind the, the frontier. And so we needed to build a plan that would enable us to have a very, very fast velocity, um, to be able to both catch up and hopefully exceed, uh, where the frontier is. So.

Um, you know, I, I, uh, I think that probably the more fundamental ones are just around, you know, a lot of the leading labs, they build the entire organizations around the premise that superintelligence is coming and it is very close and this is a very realistic thing to believe that we can create and produce. And then you build the entire plan of the lab and the business and what you focus on around this fundamental belief. And, um, and so that was one of the first things is, is to just take superintelligence seriously and then start to rebuild all of your other assumptions around that core premise. Um, so that's something I think was, was somewhat fundamental. Um.

Yeah. And I think, I think this is relatively common actually. For, I, I think there's a lot of people, um, at all the large companies who don't necessarily have this conviction because if you think about it, they've, you know, um, it's a bit of a different construction. Like a lot of the, the big companies, um, you know, they have very smart people who work on AI, but it's a little bit different from, you know, these startups where it's like, you know, these, it's like these new efforts started from scratch with this like crazy idea that superintelligence is coming. So, um, I don't think this is a problem anymore. Like obviously I think now, uh, MSL, Meta Super Intelligence Labs, is, it's in the name, built around this concept that, um, superintelligence is coming. So, so there were a bunch of principles that we sort of laid out for the effort to, and I think this sort of answers as to, you know, what were the things that we had to resolve. But one is take superintelligence seriously. Um, two is technical voices are loudest. Uh, three is scientific rigor, focus on basics. Um, and, uh, make big bets. So the, the concept of, uh, TBD and, and MSL broadly, when I sort of got started, was I thought about how, what, what would actually be the shape of a lab that would enable you to have incredibly fast velocity and catch up and potentially even overtake the frontier. And I came down to there were three, three ways that I felt like that was possible. One is to have much higher compute per researcher. So, um, you know, a lot of the larger labs, they have lots of compute, but it gets spread so many different ways and so that actually impedes the research velocity of any individual researcher. So if you build a more focused effort with a smaller team that has higher compute per researcher, you can actually make faster research progress. Um, two is talent density. So just, and this is like, I think, you know, I feel like human organizations always relearn this lesson, but the very small team where everyone is cracked is always, um, going to move faster than the very large organization where responsibility is more distributed and, you know, it's more of a mélange. Um, and the last one was on very ambitious research bets. And, um, you know, I think, uh, I mean, I think this is very well agreed upon in the industry. There's clearly, there do exist these research bets that are very big and very risky, but if they work out, can totally change paradigms and totally shift, um, how we build modern AI. And so, you know, in, in addition to obviously building towards very competitive frontier models, we're allocating a huge amount of our resources and compute towards these big ambitious bets, um, because if they pan out, then that gives us, you know, incredible models going forward.

Yeah, I mean, I think, um, I think there's, there's a bunch of bold research bets and I won't be able to go into detail on all of them, but, um, but I do think like one fundamental question is what do we care about? And, um, we, you know, in line with this idea of personal superintelligence, we really care about building these, um, agents that are able to empower consumers. So empower billions and billions of people all around the world as well as, um, empower businesses. You know, we have, Meta is kind of this incredible ecosystem. We have the billions, billions of users, which I think everyone knows about, but we also have hundreds of millions of businesses on our platforms that, you know, use Meta to run and operate their businesses. And so we really care a lot about building towards this future where we're able to build very powerful agents that empower and enable every one of both the consumers and the businesses on our platforms, um, and build kind of this new agentic ecosystem. Um, you know, we think a lot about what that looks like and what are the capabilities we need to, to build towards that. And, you know, obviously on that trajectory, there's a bunch of sub-components that are really important. Like we really have to have great agent capabilities. Um, we have to have great coding capabilities because so much of what needs to be built is software, as you, as you, you know, really get into it. Uh, we need to have great multimodality. Um, so it, it, it informs a lot of the like underlying bits and pieces that we need. And then we need to solve a lot of the, I think the bigger questions for long-running agents, like how do you think about the memory challenges? How do you think about building long-running agents? How do we build agents that are able to do more and more complex tasks on behalf of, um, the user? So, um, those are a lot of the, you know, high-level pieces that we're thinking a lot about.

Yeah. I, I would say, um, this is one of, I think the, the larger, I would say like narrative violations or maybe like differences between external perception and what the day-to-day inside is like. Actually, I think a lot of people, you know, uh, maybe have some of the impressions that you're talking about and a lot of that is formed because of just, you know, the, the reporting and, and what it sort of like, you know, how that sort of went down. And a lot of the reporting was overstated in various ways, but, you know, it sort of, it all sort of like bubbled up. And part of it was because of the recruiting so quickly. Like we knew that, um, when I got in, I knew, I was like, if we want to build great models, we need to have the, the team yesterday. So we had to just go and blitz it and do it very, very quickly. Um, but I, but I think that the culture within the lab is actually, um, very much so a startup. And there's a bunch of things that I think have, have created this feeling. I think one is that, um, it, it was an entirely new-built team kind of like within Meta. But then, um, the culture of the lab is like, I mean, everybody was very attracted and excited about, um, these things I talked about. Like people joined because, um, you know, there was high compute per researcher, so they could make more progress than maybe they would, they would be able to make at wherever they were before. Um, because there was great talent density, like people saw that it was like a truly cracked group that was pretty small and that we were going to give them the resource and freedom to make very bold research bets. So the sort of, um, I think it's like an incorrect, uh, assumption to think that like the researchers are, are just money motivated or, or anything. And, and for most of them actually, like, you know, the financial prospects of them staying wherever they were look very good as well, like looked very, very, very strong as well. Um, so money was not, I, you know, those sort of like maybe how it seemed like, but the primary motivations were actually much more that, um, they had an opportunity to kind of like build from scratch and have lots of compute, have the ability to approach their like very ambitious research directions and then do so in a group that didn't feel bloated. Um, so I think as a result, the sort of like vibe and culture are much healthier. And I would actually say like, um, you know, many people who visit the lab who are at one of the other labs, like often comment that the vibe is actually, you know, it reminds them of early OpenAI or early Anthropic or, you know, um, kind of these like more, um, uh, nascent stages of these other labs because in some sense we're like, you know, now 10 months old as, as an effort.

I don't know if we made the soup, but, uh, but, um...

Um, I, I don't know if we made the soup, but I do think it is true that we, like, I think, um, part of, part of the premise of building this lab was also that like we had to show everyone that we really, really cared about this technology and we cared about their specific research directions and what they were working on. And, um, it was a very, um, individualized recruiting process. Um, but also was one like, I, um, I mean, I'm, I'm very proud of the team that we've built and I think that like, um, people had to know that we were serious too. Like I think that by default, a lot of people didn't know what to think about Meta's AI efforts or, um, they didn't know, you know, they didn't know that much about us in many ways. So, so it took a lot of like going to people, talking to them, explaining what we're building, explaining what we're focused on, explaining what we, why we cared about the technology, what we wanted to do with it. Like that was very important.

Yeah, I think that, um, you know, I think some of this is, is unfortunate. And I, my honest expectation is that like as we get closer and closer to superintelligence, like my hope genuinely as a human that all the sort of animosities that exist between various people in this industry, which is, you know, very topical obviously right now with the, with other things that are happening, but, um, uh, but I, I think that, I think my real hope is that all these animosities subside over time and then people sort of, um, come together and realize, you know, this, we are building this incredibly important technology and it's important for all of us to be really thoughtful about that as we build it. And so, um, yeah, I think, uh, um, and, and one of the things that I feel like is a responsibility of mine, honestly, is to ensure that, you know, we, the technology that we develop and the ways we deploy it are as thoughtful as possible.

Um, yeah, I saw him in India like a couple weeks after that. And, uh, and, um, I mean, Yann is, is, um, a notable, very outspoken person and I think everyone always, always knows what Yann is thinking. So, um, but, uh, but no, I think, I think, um, you know, he obviously said what he said and I saw him in India. He congratulated us on the Muse Spark launch.

Yeah. Yeah. I mean, it, like truly, I, I do exactly what I just said before. I think all personal animosities, like I think as we get closer and closer to superintelligence, we'll, uh...

Doesn't it? Does. Um, yeah, maybe it gets worse, maybe it gets better, but, um, but yeah, no, I, I think that, um, I think that, uh, you know, I have a lot of conviction in how we've set up MSL and the research efforts that we have and the progress that we're making. Um, and I'm excited to show the world the incredible work that our researchers are doing and the progress we're making.

Oh, yes, that is definitely not true. Like I, uh, once upon a time, I, I was a software engineer in Silicon Valley. But, um, uh...

Yeah. I mean, I honestly, I mean, so on the age thing, I mean, like people have said this my whole time in Silicon Valley. So, um, you know, to some extent, like, uh, I, I actually like, I almost don't even think about it anymore because it's just, it's always there. Um, but yeah, no, I think that, I think that there are always, I think actually many people in AI always, there's like various mischaracterizations about them or, um, there's like, you know, people always say [__] and, and there's sort of always, um, uh, you know, what's out there is never always correct and it can be frustrating, but I choose to just channel it into the work that we're doing and what we put out there. And, um, again, I am really proud of Muse Spark. I'm even more excited about the models that we have cooking, um, and, uh, and the products that we have cooking. So I think like in the long arc of, you know, um, in the long arc, all this will play out just fine.

Yeah. I, by the way, I actually in general in terms of my management philosophy for MSL is not to boss people around. I think that like, uh, you know, this great Steve Jobs quote, which is most companies hire people and tell them what to do, but we hire people for them to tell us what to do. And I think, you know, that is like pretty core to the entire thesis of TBD and MSL and how we've built it is that we were going to, we were going to have, we're going to hire brilliant researchers and, you know, create the best environment for them to do the work of their careers and the work of their lives. Um, so, uh, so I think, you know, so long story short, I'm not trying to boss anyone around, actually. I'm trying to create the best environment for, um, researchers to do incredible work.

Yeah. Yeah. No, by no means. I think that, um, you know, the, what we did is, you know, over the past nine months we rebuilt, um, a lot of the stack and a lot of the research. So we rebuilt our pre-training stack. We rebuilt our RL stack. We rebuilt a lot of the science and, um, did a lot of work on data. So, uh, in many ways what's been happening over the past nine months is really like a, you know, full-on renovation as it were for the, for the sort of like core research stack. And Muse Spark is kind of the, um, an early data point on that scaling ladder, but it is not, um, you know, in some ways like Muse Spark was, is kind of like the entree or the appetizer, I guess, appetizer entree in French, but, um, it's kind of like the appetizer for what we're building. But, um, we are in development with larger models and we expect our, the larger models to be, you know, we're much more excited about the larger models than we are even about Muse Spark. Um, but it was an important data point we thought to, to put out there into the world because it was a, um, you know, the entire program that we've built is, is developed around predictable scaling. So, and we see the scaling, I think we talked about this in the blog post, on many axes. We see pre-, you know, very consistent pre-training scaling and predictable pre-training scaling. We see predictable RL scaling. Um, we see, and, and scaling in reinforcement learning. We see predictable test-time scaling. Um, and, uh, a lot of what you just talked about, the content planning mode, is we're also seeing very exciting results in multi-agent scaling. And so, um, everything about our program is built to continue scaling as we go. And so, you know, Muse Spark was like this early data point on our scaling trajectory, but the next data point we're like a lot more excited about, and the data point after that we're even more excited about. So, um, I think we're, um, we're excited to show people this sort of like next rung up on our, um, on our overall scaling efforts. And for Muse Spark specifically, I think it, it, you know, if anything, it's actually, um, the overall end performance ended up being quite a bit better than we expected, and the, um, it had a bunch of emergent capabilities and behaviors that, um, we were pretty excited about, like, uh, uh, that we found in training. So for example, some of its abilities in, um, uh, agentic, uh, like visual coding, being able to produce websites or being able to produce games, like some of these capabilities, um, actually kind of emerged from the fact that it is both a pretty strong agentic model, but also pretty strong at, um, multimodality. Um, so, uh, so there were a lot of, there were a lot of things that we're very excited about this model, and, um, and so, um, we put it out there. We think for most consumer use cases, actually a very good model, and it is, um, quite competitive with, uh, the other models. Um, Muse Spark as we deployed it is not yet competitive on, you know, agentic coding, um, and so those are capabilities that we're working on and we're building towards for the next set of models. Um, but, uh, but yeah, I would expect the next model that we produce to be better overall than Muse Spark, and that's something we're pretty excited about. But, um, but even Muse Spark as it, as we released it, which I think we, to be clear, wanted to set the clear expectations, like we didn't think it was going to be, um, a state-of-the-art model across the board, but it, it, uh, it is a very good model, and I, we think a lot of people and a lot of users who tried it out, um, experienced that.

Um, I mean, the one-word answer is just scaling. Like, uh, Muse Spark, um, is, uh, kind of, as I mentioned, it's early on the ladder, and we have very strong predictability. So we know if we scale this model up, like what performance to expect from that, from that increased model size. Um, and, uh, we expect the, the upcoming models to just be able to perform much better across the board. Um.

Um, uh, coming months, you know.

Well, as I mentioned, we built the whole program so that we would be able to, you know, move very, very fast. So there was a time period where we had to rebuild all the foundations and rebuild everything. But now we're in the period where we're kind of in the, we're going to be in fast scaling mode.