Transcription
It was almost exactly a year ago that you started at Meta, leading the AI team, and I thought maybe you could set the scene for us just a little bit. Um, I, I'm wondering if you can let us know where you think Meta is as an AI company today versus where it was a year ago, specifically. Like, what's the reputation of Meta as an AI company today versus a year ago when you got there?
Um, yeah. I mean, I think it's been a very exciting year for, um, for Meta. And I mean, I think when, um, you know, roughly a year ago, uh, a little more than a year ago, Meta released, um, Llama 4. And while it was still an exciting release, I think, you know, um, it wasn't quite on the trajectory that Meta needed, um, to be able to continue building a lot of the products and experiences that it seeks to build. And so, you know, we've, uh, been hard at work over the past year since I joined and since starting Meta Superintelligence Labs. You know, we've, um, been undergoing an entire process of building a new scaling ladder for our models. Developing new sets of infrastructure and and new research. Feel the power of a new series and family of models. And you know, back in April, we released the very first, uh, fruits of that labor. You know, the new Spark models and an update to Meta AI. Um, the reception to all of that was incredibly positive. It was even better than, frankly, we had expected internally. Um, we saw incredible gains in usage of Meta AI, was, you know, at the top of many of the app stores and, um, you know, we're sort of working on our even larger models today and, uh, and are excited about what we'll be able to demonstrate to the world. We clearly, you know, are we're on a very exciting and fast trajectory, and we're excited to continue showing the world what we produce. Um, I think that, uh, you know, the AI industry has gotten very, uh, hot and competitive obviously, in the past year as well. And, you know, we take that very seriously, but we're really excited.
Yeah. Obviously, the goal you're spending in the same way that an OpenAI, Anthropic, others are spending. Do you feel like you guys are in that same tier is Meta? You know, is it OpenAI, Anthropic, Meta at this point, or do you feel like there's still a bit of a gap? Because I'd say certainly a year ago that was the perception.
Yeah. We don't. I mean, the new Spark model that we released is not at the tier of the leading frontier models. Um, but we believe it's a very exciting data point on the trajectory, and we expect the upcoming models release to be quite competitive with the leading models in the world.
Yeah. You called Muse Spark, um, an appetizer, an appetizer model. Uh, to to build that metaphor out. When does the entree model get here? You know, will it be at that tier that that you guys ultimately want to be at?
Yeah, we are, uh, we are in process of cooking. The, uh, we're said the whole conversation is just going to be this, this back and forth. Yeah, yeah. No, no, I mean, we're we're cooking it. Um, we're excited to show it to the world once it's ready. Um, we're seeing very, uh, exciting and promising results in the process of training it, um, right now. So, uh, we're quite excited about it. And, and I think just, you know, overall, we built the entire research effort around predictable scaling. So the so the entire belief of our overall research effort was, you know, um, that in many ways, the central belief behind the current modern AI boom is that as you scale these models, you will see, um, get incredible results and get predictable levels of increase capability. And so, um, you know, Muse Spark, it was an early data point on that scaling ladder for us. The next models we release will be, uh, an even greater point on that on the scaling curves. And we are really excited to show the world what will what will be able to produce.
And obviously, you went from basically rebuilding the team, rebuilding the lab to this new Spark model in a very short amount of time. What is the biggest barrier from getting to that? Um, the appetizer to the next level, is it? Uh, I assume it's not resources. You guys are spending a ton of money on this, on this effort. Um, it's it's just simply time is a talent. Like, what's going to bring your models to that? That frontier?
Yeah. It's, um, you know, we we've talked about some of this in some of our, our public, uh, blog posts and whatnot, but it's about continuing to scale the data, the, uh, compute going into the models, um, as well as continue to scale with research. So continue to, uh, drive advances in, um, you know, underlying, you know, underlying research breakthroughs to continue driving forward the progress in the models, um, and building infrastructure to support all of this. You know, it is um, this is in many ways a year where, uh, all of the latter dramatically scaling up their models. And we are on, we think, a much faster trajectory to do so, because obviously, we've been doing all this work over the course of the past year. Um, but yeah, we need to build the infrastructure. We need to scale the data, scale the compute. Um, train these large models and show them to the world.
Yeah. Um, I want to talk about model strategy a little bit before you got to Meta, everything was open source. That was definitely the overarching strategy. New Spark model is not open source. Um, I believe I heard you on a prior interview basically say that as you guys were testing it, it didn't feel safe to open source. Can you go deeper on sort of what you mean by that? Um, and how you made that decision?
Yeah. So one of the things that we did when, um, uh, as part of Meta Superintelligence Labs is we updated our what we call our advanced AI scaling framework, which is really our view of, you know, what are the risks that we see in developing these very powerful models? And how do we want to handle, um, those risks as we see them in early testing? And we published a lot of what we saw in the process of creating new Spark in our, uh, preparedness report. And some of the things that we saw was that it actually triggered some high-risk areas, uh, in the course of early training, particularly run by a risk. Um, but also a number of the risks were elevated. And, um, you know, this is something I think the entire industry has seen as the models have improved pretty dramatically over the past year. So we certainly aren't the only ones to see, uh, a host of these risks, you know, show up in, in as we scaled up the models and as we sort of kept pushing the frontiers of research. But, um, we saw these risks trigger and we realized, you know, I think fundamentally those, um, when we launched a model like new Spark in a product, we have a lot of ways to mitigate some of these risks and ensure that we're able to launch it in a safe and responsible way. Um, it's much harder to do that when you open source a model and, um, you know, uh, people can use that model in all sorts of contexts that we may not have full understanding of. So we're in the process right now of developing, um, models that we believe are fit and safe to be open source while still maintaining as much of the performance capabilities as possible.
So you will still do open source. It sounds like Llama though, is not the brand or the the pillar that you're going to do. Is everything open source going to be Muse Spark or adjacent?
Uh, you know, we, uh, we have, uh, exciting debates about branding internally and, uh, nothing to share right now, but, uh, but.
Yeah. Okay. Um, the big models that are coming that you've hinted at, um, give us a general sense of you can obviously, you know, each model, each company is, is perhaps known for certain things. Do you feel as though you're moving in a direction where Meta's models are going to be known for, you know, best in class at X versus Y? Like, what are you hoping to accomplish with with what you guys come out with next?
Yeah. So so already in Muse Spark some areas where we were really impressed by the capabilities. even though is again like a much smaller model than ultimately we, um, we intend to train were around multimodality capability. So its ability to handle images, video, audio and that's obviously very important and critical for Meta's business. Um, also its capabilities in health were really impressive. Um, and that that was very exciting for us. You know, health is an area that we view is really critical as we scale these models out to billions of millions of people all around the world. Um, and then also a lot of the early results we saw in the ability of the model to create your vibe code and create little games or artifacts or whatnot were very powerful. So, um, we are doubling down on some of these and, and continuing to invest into the capabilities of the model. So we're really excited for the upcoming models release to be very, very, um, uh, capable agents, uh, paired with a lot of these other strengths around multimodality, um, around health and many others. And ultimately, what we're really excited to build, um, for the world is are the best personal agents for, uh, everybody around the world.
As much as possible, I want to get to agents in just a second. Because you just made some news on agents, actually, yesterday. You have other stuff in the works that we can talk about. But before we pivot off safety real quick, I do want to ask a question just about China. Um, you've been, you know, you've talked publicly about the risks, um, and the threat of AI coming out of China. You guys have also trained on some Chinese open source models. I'm just wondering, can you give us a sense of like how you view China right now in the, in the, you know, through the lens of AI? Is it a threat in the way that we've heard historically? Do you feel like that's changed?
Yeah, I think that, um, I think it is incredibly important for the United States to lead on technology and the, uh, economic benefits, uh, that can be created from. I think this is this is very, very critical. If you look at the history of civilization, you know, technological advances are, um, very important for, for, uh, countries or civilizations to adapt to and be able to adopt. Um, and that really defines, you know, the really, truly the course of history over, uh, over long arc. So I think it's very important that, that the United States is able to lead on AI. Um, and, uh, that's, that's a huge part of our focus at Meta, as well as ensuring that we are able to contribute to the, the United States leading.
Where do you think we are right now as a country? Is the US leading?
Uh, I think right now the US is leading. Yes. And I think it's, um, you know, this is one of these situations where you, you know, it's important for us always to track progress from many other countries, but especially China, you know, be very thoughtful and understand exactly, um, what's happening within each country and what are the reasons those things are happening. But I think right now we're ahead.
And what. Sorry, I said I was going to get the agents for one more. Uh, what could put that at risk? What could put that lead at risk? What's the most threatening, um, thing to stop that?
Um, that's a good question. I mean, ultimately, I think we are in a phase where, um, the, the research advancements industry that we're seeing from continuing to scale these models, apply more compute, apply more data to these models are just, um, incredibly, uh, exciting. And, you know, in some ways, the progress and pace of research today is nearly miraculous. And so, um, I think it's important that we're able to continue this pace of progress. The were able to continue, um, you know, being able to, uh, continue scaling for these models.
Yeah. Um, agents. Now finally, um, you guys, just yesterday, I believe it was announced a business agent. Um, so advertisers can use this to, you know, interact with customers, I presume, eventually help even, um, develop ad campaigns, things like that. But you're also developing a consumer agent. Talk me through your vision for how ultimately agents will reach all of us. Like, I guess I'm wondering, is an agent going to be similar to like, my email address where I have one core agent and maybe a secondary agent? Or is it going to be like the apps on my phone where I have one agent for every single task in my life? Like, what do you envision we're going to be using as a society?
Yeah, I, you know, we ultimately think it'll probably land somewhere in between those. I think, you know, we really believe that people are probably going to have one, maybe two, maybe a small handful of, of agents that they rely on. And maybe they have a personal agent that's focused on things like their health and maintaining their personal relationships and, you know, helping them be a better parent and be better with their friends and family. Um, and then, you know, perhaps they use that same agent in their work lives, especially if they're working in a small business or they're an entrepreneur or, um, or, um, you know. Uh, you know, working within a smaller organization and then, you know, maybe there's worlds where if you work within a larger enterprise or a larger, um, company, then these become bifurcated and separated. Not too much, unlike email, let's say, or plenty of other, uh, you know, key technologies that we use from day to day basis. So, um, yeah, we think that ultimately, uh, it should be, you know, agents will be something that, ah, that become deeply personal, um, and should be things where over time you find yourself being able to rely on them more and more and more for, um, more and more of your personal life, more and more of your work life. Um, and that'll be a process that, you know, all of society goes through together.
Do you feel that Meta, in particular. And again, a lot of this happened before you got there, but long history of privacy related issues. Are people going to be willing to, you know, trust a Meta agent with the personal task of their life that you're describing?
Um, yeah. I think that this is like one of the most important societal questions for agents writ large. I mean, I think that, um, there's incredible, uh, amount of innovation and technology that we built out on things like a genetic safety, um, you know, uh, ensuring that these agents are respectful to your privacy, ensuring that they're respectful of, of your boundaries, um, and continuing to design products in a way that are able to support that. So, uh, this is definitely something that we're taking very, very seriously. And we're being quite thoughtful about. Um, and ultimately, we're excited to, you know, show the world what we've built. But, um, but yeah, we think that this is not even just a Meta problem. This is an industry-wide problem. As we build more and more powerful agents, I think that it is a redefinition of, you know, humans' relationship with technology in many ways. And that's something that, um, you know, we're all going to have to think through and work through together.
How soon will we see a Meta consumer agent?
Um, we are actually cooking it. Cooking the entree. Yeah. And, um. But no, this is like, I think this is one of the things that was very exciting for us internally about the new Spark launch, in my eyes, that, you know, we even when those launched, we were cooking things internally, both the larger models as well as, um, as well as some of these products that you referred to that were, if anything, more excited about than what we came out with in April.
Okay. And how are you using agents right now? Your boss, Mark Zuckerberg, it's been reported he's has like, uh, a Zuck bot, essentially, uh, or various versions of agents that he's tasking some of his CEO duties to. Is there an Alex bot that's that's doing part of your job right now while we're on stage?
Um, well, I definitely use agents to support and help me, um, in my work a lot. I think that in many ways, um, you know, being a leader within a company is really about how well are you able to, you know, understand everything that's happening at the company to the best you can and help, you know, support your team and being able to continue to execute better and better. So I think there's all sorts of things that it just can do to help you there. But I think the uses of agents that are probably most exciting to me are the ones where I use or the ways I use them in my personal life. So, you know, I use an agent to help me be healthier. Um, and I use an agent to, uh, help me keep in touch with my friends and like, ensure that I maintain those relationships. And I think these are use cases where they're, like, very notably different from the world where I didn't use an agent at all. Like, these are things that are I think are like hard have been hard historically for me to like, stay on top of, you know, both my health as well as, um, keeping in touch with all my friends and having an agent that's there to help support you do a good job of those things has been pretty transformational.
Um, somehow we only have a couple of minutes left, but I want to I want to ask some sort of bigger picture questions about AI and society. Um, but but I actually want to start at Meta. Right. So your team is getting immense investment. Your guys are investing hundreds of billions of dollars in AI writ large. At the same time, there were layoffs, uh, at the end of last month. So there are people and some of the framing is, hey, this is to offset that as the person in charge of AI, I'm just wondering, how do you deal with that reality, right? That that you are working on a product that you're excited about, but at the same time, you know, the company is saying, hey, this is costing jobs as well. Like, how does that make you feel? How do you deal with that internally? I'm just wondering how you, you know, deal with that reality, right?
It's incredibly difficult to say goodbye to teammates. And I think it's, you know, it's no, um, I think it's very well known that this is this is a challenging, um, thing to go through as a team and something that, you know, is, uh, is important to acknowledge. You know, we are ultimately really excited about the progress that we're making, I and the products that we're building. And, um, you know, we're excited to bring those to the world. Um, but, yeah, I think that there's, um, you know, running a large company is very complex. And, um, you know, we're working through a lot of those issues, but, um, you know, don't take any of it lightly. But, you know, at the end of the day, we're excited about what we're building.
I think job loss in general is probably like one of the biggest fears with AI. Um, what's your view on that is if AI goes the way that everyone envisions and we reach superintelligence, um, is there a world in which that can operate and live alongside all of us staying employed?
Um, I mean, I think it's it's something that we should pay a lot of attention to. And I think that we should, you know, track closely and try to understand what the impacts are. I think one thing, though, that we never, you know, I think rarely talk about, but it's also happening that's very exciting, is that AI is enabling the creation of more businesses than ever before in the world. And, you know, we see this in our data. I'm sure many companies see this in their data. There are more new companies being started today, like through use of AI tools than ever before, and that those numbers are only growing. And I expect as AI tools become more and more powerful, we'll see more entrepreneurs, more small businesses being started. And so, um, I think there's, you know, it's the economy is a complex, you know, machine. And one of the things that we see within our data today is that there are more, smaller small businesses being started. There are more entrepreneurs, um, there's more opportunity for entrepreneurs. And so, um, you know, uh, and we're really excited about that. We're excited about supporting small businesses throughout the world.
Is it the kind of thing that you feel? We saw the Trump executive order just this week, where the administration wants to sort of review some of these models before they're they're released? Um, here's what you thought of that, first of all. But two, is there another type of regulation that you think actually might be helpful to prevent, you know, job loss or prevent a world in which AI is taking priority over the humans?
Yeah, I think well, I mean, first of all, I think it's it's, um, this is a really important and powerful technology. And so I think, um, it's it's great that the administration is, you know, deeply considering, uh, what how we should be thinking about this technology, how we should think about, uh, responsible deployment and, and what all that entails. So I think that's, um, broadly speaking, I think it's been really great that this has been an issue that the administration has been really involved on and very thoughtful of. And, um, you don't think it slows innovation to have a review process like that?
Um, well, I think I think it's always a balance. You know, regulation is notoriously difficult to get right in almost all contexts. But, um, but I think that, you know, what we what we're seeing from these models is just that they're becoming dramatically more capable. And I think it's important that we're thoughtful about how we deploy them ultimately.
Okay.