At the end of last year, we're like, all right, we are quite... We haven't seen the models yet. You know, we're going to write this check for like, I don't know, like $100 billion, $200 billion. I'm like, Alex, are you sure this is going to work? It's like—
Cheers. Cheers. Yeah. Cheers. Cheers. Cheers.
So, give me the update on Muse.
If you were looking at the third-party reports, you would see that this might be the fastest-growing agent or AI app that I think has existed yet.
What are your guys' personal favorite Muse use cases, either yours or other people's?
I was actually talking about this at dinner yesterday. My favorite thing is I have my Muse, whose name is Oiler, constantly source increasingly obscure teas from increasingly obscure vendors of tea. So at this point he's, on a weekly basis, my Muse is buying some random tea from some farm in Taiwan or some farm in Japan and sourcing it over and bringing it back. What's your favorite Muse use case?
I don't know. I mean, I use it for a ton of stuff. I still like... My favorite both use case and probably eval for the product is having it coach me in MMA. So I put cameras in my gym, and I basically... It watches the cameras when I'm in there, and it gives me feedback in between the rounds, which is a very hard problem, it turns out, because not only does it have to... It basically has to process all the video quickly enough because you only get 1 minute in between the rounds.
Oh, it coaches you between the rounds?
Yeah. I want feedback. I want feedback on how to win. You know what I mean? It's like, come on.
Did you have to instruct it... I've wondered about this too, because looking at videos is a really good AI use case. It feels like you can't just give it a multi-gigabyte video file. So do you tell it how to slice stills out of it?
No, it basically... Yeah, it pulls stills. I think it kind of figured that out by itself. Um, it turns out that Muse has pretty good native understanding of MMA, so that's good.
The pre-training. Yeah, someone on the team knew. Now... Just kind of out there in the corpus. I think it's a good test of the kind of coding ability that... Because really basically what it's doing, it's like, you know, within the agent, I'm not seeing any code, but it basically is writing this whole video processing, image processing pipeline.
And I'm constantly telling it like, all right... Like, look you took 30 seconds to give me that feedback. I only have a 60-second break. Like, you gotta go faster. Just optimize yourself, optimize yourself. And it's, I think, a good example of how stuff is going to go in the future where basically everyone is actually going to be writing tons of code. They're just not going to know that they're writing code.
What kind of MMA steers does a model give you?
Um, you want to see? Yeah. Basically, let me see if I can pull some stuff out. But basically it gives me a lot of stuff about balance and, like, hip position. Or it's like when you're kind of in side control here. It's just like you didn't get this underhook in this place so you're going to... You're vulnerable to this getting rolled over here. I'll pull up some stuff. You guys should try this.
It was definitely one of his first use cases actually of even the initial prototypes. Yeah.
Yeah. I mean, I did this with OpenClaw before we even had Muse working. It was just like... Then that's why I was saying it's kind of like become an eval for this too.
It's good. I actually didn't know that it was an eval for the product. So noted. Yeah.
How is all this... In the model's training data, like the fine-grained, this is what an MMA fighter looks like when they're appropriately balanced, and this is when they're like too much on the back foot. Just, I'm wondering, how has it consumed so many images of appropriate versus inappropriate MMA posture?
Here you can see it. It just pulls out frames from it and it's like too much on my back leg there.
Yeah, but is this the emergent intelligence? Like it does actually... Again, does it know from seeing lots of images of MMA or does it know because it understands the concept of where you're meant to be balancing? It's read books about MMA, and then it sees that you're too much on the back foot, or—
Well, we did train the models with a lot of visual perception data.
Yeah. So, I mean, I think that that's like kind of making good multimodal models. I mean, this gets into like, I think, a core approach thing that we have, which is we're training the models end-to-end to be good personal agents, right? So like the question is why would Meta want to build these models? First of all, I think just superintelligence overall is going to allow us to run the company better and build all kinds of different things. But for the products that we're building, there's certain capabilities and certain mixes of skills that basically are just much more applicable to these cases. So I mean, this is one of the things that Alex focused on from the beginning was like, all right, we really want to make sure that our multimodal program is good because that's disproportionately important for the types of things that people use all these services for.
How do you train a good multimodal model? Is this where Meta just has an advantage because you guys have so much image and video data?
It's very helpful. Yeah. And I think ultimately, a lot of this, it comes down to how are you focused on it, right? I think because of the fact that we wanted to build personal superintelligence. And fundamentally, I think most people in the world really experience the world in a highly multimodal way. They like talking to other people and listening to things and seeing videos and filming things of their own. I think most consumers have this highly multimodal experience. And especially earlier this year, there was actually a point of tension almost in the industry because I think so much of the consensus view within the AI industry was that it was just all about coding and coding was all that mattered. And I think that for the success of a product like Muse and the success of a lot of these consumer use cases, it really depends on the ability to be strong and multimodal. Yeah.
I think we're just kind of scratching the surface of that. I mean, a lot of the initial theory that we had was that, I mean, most people don't actually like reading that much, right? So even the text-based interaction model, I think it's something that people do. I mean, everyone texts. But when you think about stuff like the glasses that we're working on, or even our other products that are very visual, the glasses basically will eventually just be a pure multimodal product, right? It's like you can let it hear what you are listening to, and it can talk to you, and it can see what you see if you let it. And then eventually, I don't even know if we're going to have an app store on the glasses. It'll just be like a multimodal generation that kind of creates whatever you need right in your vision. And I imagine something like that will probably end up powering a lot of the apps that we build too. But yeah, I mean, we're still in the early days of that. But for now, the perception work has been really impressive.
I want to ask you guys about the product space here in AI and try something out on you. If I was to describe to you, say, with Stripe, what I think a winning product looks like, it would take me pages and pages to talk about because we have the developer experience and we have to talk about all the international heterogeneity of payment methods. Then we have to talk about stablecoins. Like, it would take me a long time. Whereas I notice a phenomenon in AI where it feels like the winning products collapse very quickly onto an agreed-upon product space that everyone is pursuing, and so transformers worked, and then like the chat agent UI really worked. And now we are seeing... You mentioned OpenClaw earlier, the kind of OpenClaw persistent agent, has a computer, able to do things. You know, computer use out there in the world view. We're collapsing on that product space. But there's something about AI where the winning products tend to be very simple, pursued by all the companies that are taking it seriously. You could describe on the back of a postcard... Like you could describe very concisely. And obviously, implementing them is massively complicated. Like, we'll get into the new model. There's, you know, huge amounts of frontier work that goes into making a good version. But the winning brand—now you're talking about the glasses or the future versions. But somehow there's a very small number of clearly correct products in AI as opposed to it being this, like, very broad and complicated product space.
Well, first of all, I think part of the theory from the beginning was that this kind of agent or whatever form it took was going to be quite a different thing from this text-based assistant, right? Every company has one of those. I mean, ours is Meta AI. It reaches a ton of people, but I mean, it's like the same thing that all these companies have. And we, I think, specifically think that that is used for very informational purposes, but people don't think about it as like a thing that gets stuff done for them.
It's retrieval, not delegation or something.
Yeah, and even like when you say it has a computer... Yeah, but there's lots of different ways to do that, right? You know, I think that it's only kind of this most recent wave that gives the agent a dedicated VM, which I think is actually incredibly important for making you feel confident that the security and privacy of your information is—it's not just like, okay, it's not connecting to some elastic compute that just everyone uses.
But my claim isn't that it's easy to implement, nor that it's obvious—
I think there are going to be a lot of things.
Yeah. Okay. You think there's going to be a lot of different things? Because again, my claim is that now that people see Muse, everyone looks at that and says, "Oh yeah, that's clearly how you want the AI agent product to work, and that the future AI agent products that we'll see will be fairly Muse-shaped."
So I think... My take on this is, I think we're just so early in the realm of what AI products will look like that I think a lot of the interesting differentiation about the products kind of gets collapsed in how we view the products today. I think the traditional way in which we wrote down and think about product specs, that is... Those aren't the relevant dimensions and levers and axes to consider the important design decisions on AI-native products. And so, Muse, for example, I think so much attention was put in the design of the model for the product. And there were a lot of very intentional decisions around, in all these situations, how exactly should the model behave? And what's the exact personality of the model in these situations? And how should it handle these sorts of tool-use scenarios? And these are very intentional decisions. And probably in the same way that the full spec of how we wanted the model to behave in all these scenarios would be an extremely long document. But because of just how we think about products today, it's just AI. It's kind of like how when mobile first started happening, it was like, "Oh, it's just an app and everything is just an app." Well, it turns out there's a lot of nuance in what exactly an app is. And we're kind of in this phase where, oh, it's just AI, or it's just an agent, or it's just an agent with a VM or something like that. But the interesting differentiations we have to tease out.
Okay. So you guys think I'm making the classic mistake, which is like, my domain is complex and this domain is, again, from a product point-of-view, it's simple. All payment processors seem the same.
Yeah, no, I really agree with what you're saying. One thing that we talked about from early on, is that basically there is this pattern where I think the products are getting thinner and the model is getting to be more of the product, right? So one of the things that we say internally is like, the model is the product.
Right? So, and when you're talking about an agent, a lot of the complex product work that went into Muse is basically around the privacy and security model for handling your data and the things that get connected to it. And that, I think, is something that we're just very proud of that we really focused on. But that's a lot of the product surface that we need to innovate on. But the capabilities, the personality, a lot of the stuff is like you're training into the model. And whether they're the people in the lab who are working on that or the people on Nat's team who are working on the product, I think when Nat is thinking about what is the key stuff that we need to make the product feel good, he's talking about post-training the model, right? Not like... I mean, of course the UI of the product matters too, but the most important part of the UI is the product.
Yeah, exactly. I think there's like, there's still very viscerally, I think from a user, there's a feeling of using Muse over time. And maybe some meaningful portion of that is just the reliability of Muse to handle certain tasks. But then another meaningful portion is how is it proactive? And in what context is it proactive? And when it talks to you, how do you feel about how it talks to you? And so there's clearly a lot of... And I'm starting to see more and more of this actually show up on the X feed, which are people are starting to tease out these, like... They're all proactive agents, but what are the actual meaningful differences between them? And these, I think, are the interesting product decisions, which, you know, the funny thing is how do you specify these product decisions? Well, it's ultimately in model design decisions, which ultimately boil down to training methods or data or evals or, you know, all these things.
Yes. As we talk about product space, one notable Muse detail is that there is no model selector and no effort selector. Do you guys think that's just obviously correct in a consumer AI product, or—
I just want to make it simple.
Yeah. I think that we've been... I think somewhat infamously, a lot of the AI products have been designed by a highly technical group that do not have these sort of randomly sampled American in their mental model. And so the products... I mean, the naming of some of the models is like so egregious.
The numbers go backwards?
Some of the times the numbers go backwards and sometimes it's just like a totally different word and you have another set of words. And like, I just like, how is somebody who's not carefully reading API design specs supposed to know what they're supposed to pick? Like, it's just far too confusing.
Okay, we're talking about models a lot. Tell me about the new model. It sounds awesome.
Yeah, I mean, I think we're quite proud of this. Um, yeah, the new model—Muse Wonder—it's a new class of model that is at the frontier. So we've basically been working on this since the beginning. Basically, when you're doing these pre-trains, you don't just go to whatever the biggest cluster is that you have. You want to work your way up the ladder training models that are maybe 4 or 5 times bigger, and then you just keep on repeating that process. So the first set of models that we've had Muse Spark was all based on this initial pre-training that we did in, what was it, December?
Yeah, the end of last year. Yeah.
And, and it's actually... I think we've all just been quite surprised and impressed with how much intelligence and capability the team has been able to build into a model which is like frankly not that big, right? I mean, and then after that we were working on this next model, and the pre-training has been done for a bit, and we've been working on the post-training, and now it is here. So it's the next class of model. It is significantly bigger. So I think you you will feel that in terms of the nuance in which it can think. And it really. A lot of the areas that we focused on were making it the best personal and professional agentic model. Some of the different things that you can optimize for that's it for us, and then yeah, a lot of the basic capabilities that we focused on building in. Very good memory, very good ability to persist through complicated tasks. But a ton of work went into this across the lab and the product teams and all of this. And I think it's also just sort of a testament to the the team being able to methodically figure out the process to scale these models. So I think this is also just, you know, a milestone on the way towards the time between each subsequently bigger class of models, I think will now start to compress. And that's very exciting for the type of stuff that we can build.
At Meta or in the industry generally?
At Meta. I mean, I think you're seeing something like this everywhere else, but—
Is this an RSI effect or something else?
Well, I think at Meta, part of this is just... We needed to... Well, I think quite famously, we were not happy with the trajectory that we were on with Llama 4. We kind of like rebuilt the whole team. That's been this kind of crazy thing over the last year and a half that we should probably spend a bunch of time on.
Oh, we'll get to that. Yeah, yeah. Some war stories.
But the team basically had to go build up like different tiers of models to get to this one. And now that we are kind of at the frontier, now everything that we do going forward will just be pushing the frontier, as opposed to there's a lot of background work that we needed to do before that. I don't know, what would you say? Yeah.
Well, I think we're really excited about it. We will have 2 models in this series that are going to be available. So there's Muse Wonder and Muse Wonder Full.
Laugh? Yeah, all right, there you go.
Yeah, very good. Uh, which is basically a much higher—
Yeah, no, a much higher-end model. It's the full version of Muse 1.
I know, I got it. Yeah, yeah, it's not—
It's 2 words. Wonder full version.
This is an excellent dad joke. Yeah, there you go.
Um, and, and the performance... I mean, we've seen like incredible performance across a slate of evaluations on personal, professional agent tasks. It's state-of-the-art based on our measurement. And then there's a number of, I think, specific highlights which we're really excited about. Which we've designed into the model over time. So active memory, like Mark mentioned, is one area where we think that to build personal agents, personal professional agents, like memory is just one of the clear requirements to build reliable agents and is one of the areas where we've invested, I think, disproportionately actually compared to many of the other efforts. And we have our own internal evals of Memory Bench which demonstrate a lot of the capabilities there.
What does memory mean in a model to a layperson? Is it just how much stuff it can... Like the working memory of if you're working on a task, how much stuff the model can kind of keep in its—
How's it different from context?
I think it's in the experience of using the model. How well does it remember things that you mentioned to it over days, weeks, or months, and then leverage those in the process of doing future tasks?
I see. This is the stuff that's buggy AF in OpenClaw. Like, I think we're all extremely fond of OpenClaw. Very happy OpenClaw user, but it ultimately does this compaction where it has like the long-term memory and the working short-term memory or whatever. And I think once people use OpenClaw for a while, they find the memory pretty buggy.
Yeah. So we've designed basically both the model and a system for the model to use to develop a much stronger form of memory. We call it active memory.
In addition, as Mark mentioned on the persistence—
But so this is outside the model, like the model exists and it's—
No, no, but some of the capabilities that you're talking about, like the compaction, are things that the model does, right? So basically being able to do this well, both with—
It's not a harness, it's in the model itself.
Yes, it's in the model itself. Exactly, yeah. So it's aware of its memory as a core system. And it's about how it's able to leverage various sources to be able to conduct tasks in the future, as well as understand its ability to do compaction, how it should do that compaction most effectively, et cetera.
Okay. But this is a big deal because I think a lot of people's early experience of AI agents has been buggy memory. And so, yeah, improving that is—
And to be clear, there's still plenty of room to improve. Our evaluations show this, but this is something that we're quite proud of, and we think this is a meaningful contribution.
Okay. So active memory is one thing.
Yeah. Active memory. Active focus is another, which is effectively just the ability for the model to, when conducting tasks, be able to carry them through, especially as we've seen Muse users as well as Muse Code users, give higher and higher complexity and more and more ambiguous and more and more complex tasks, we want the model to be able to conduct those and do those very effectively. And this is one of the things actually in early dogfooding and testing that a lot of people have noticed about the model that is sort of special and stands out with respect to other models that they've tried or used. The third one is domain expertise.
What is focus in AI terms? Is it like putting a lot of compute on the task? Is it not getting distracted with other—
Kind of, persistence? Yeah, like persistence. It's basically like the ability to conduct complete tasks over very long horizons.
And then when faced with various setbacks along the way, the ability to identify other solutions to those problems in an aligned way.
Then there's... It's actually a big step up in coding capabilities, so frontier coding capabilities. So it is actually state-of-the-art on DeepSWE and, SWE-Marathon and does very well on other benchmarks like Terminal-Bench 4 and many others. So we think that users who use it for coding, which will be a minority of all of our users, but the ones who do will see a big improvement and a big step up.
Okay, so we should also just be using Muse Wonder Full for coding at Stripe?
Potentially, yeah. I mean, let's go out there and—
But like if it's state of the art on SWE-bench, yeah. We'd be missing out.
Yeah, no, it's a big improvement on coding. I mean, obviously, we want people to try it and see how they experience it, but we've been very pleased in our own internal testing.
Yeah, I think people will like it for a lot of things.
Okay, so this is both just your latest frontier model, like people can come, you know, developers can come directly to you, they can use it on OpenRouter, like they can just consume it as a model, and it's also a big upgrade inside of Muse that people will just feel as active users of Muse.
Yeah. Well, what we're going to do in Muse, because it's a lot bigger and it's more expensive to serve, we're going to start off primarily in paid subscriptions. So what we want to do with Muse as a product, we want it to be free for everyone to use. So we have this huge amount of free tokens that people can use with with Muse Spark as the model. But if you want to use Wonder for most of the hard tasks, then at least for now, you'll use the paid version. But, you know, I mean, eventually we'll, you know how quickly the field moves.
And we hope to, you know, at some point not too far from now, have a model that is as good as Wonder, but in the Spark form factor. And that will be a great thing to be able to serve hundreds of millions of people around the world.
Okay, so you go to Muse, you subscribe to the—what's it called—like the premium or pro plan, or...
And just don't stop until you've taken out your credit card.
Yeah. And then you'll be able to...
You'll select it, or it'll just magically—
Oh, it just gets better. It just gets better for you. Yeah. And then it'll also be available via API, on the Meta Model API and OpenRouter. And then it'll also be available for Muse Code users. And then Muse Wonder Full will be available to subscribers of Muse Code. So it's pretty exciting. I think we want to, as Mark mentioned, it is a much larger model. And so we're working through how to serve that most effectively to as many people as possible. But we're excited to get it in the hands of as many people as possible.
It's a big step forward. I mean, this is like, and we really with Muse Spark, we were very happy with the performance that we were able to deliver as we were getting up to this point. But like, this is, the beginning of, like, MSL kind of being at the frontier. We feel like the models are quite competitive with the best models that are out there. And I think that this is—
Okay, so you think it's like Astra, Fable, and Wonder altogether at the frontier?
I mean, you can look at it empirically by the benchmarks.
Empirically by the benchmarks, one would conclude that.
Yes. And I think actually one of the things that we're especially proud of is we have taken the time, as we've mentioned, to optimize models for the product form factor of Muse. So this is the state-of-the-art model. We think it really is the best model in the world for the use case of personal professional agents.
Yeah, people inside companies like complaining about Goodhart's Law. You know, a metric that becomes a target ceases to become a good metric. You hear complaints that models are benchmaxed and that at a certain point, kind of overfitting to the benchmarks actually makes the models slightly worse. As people who actually have to train AI models, is that true? Are the benchmarks pretty good at grading models? Do you start to see weird edge failure cases from the—
The benchmarks have to keep moving, right? Because it happens very quickly, but they just get saturated so quickly. So, I mean, I do think that—
But you think non-saturated benchmarks are fine?
I mean, as we've gone on, we steer the models more towards our internal benchmarks for the use cases that we're focused on. So, I mean, early on when we were kind of just doing the science of trying to climb the ladder, we cared more about what the external benchmarks were just to make sure we were on the right track. And then I'd say, you know, each release that goes by, the larger share of what we're looking at are internal ones. And one of the things that we actually quantified in our research blog post as part of Muse Wonder is this metric for big model smell, which is kind of, you know, big model smell and feeling of benchmaxing are kind of like inverses of one another.
This idea that you could just like tell when you're using a really big model, it just... You can tell.
It just has nuance. It's got that je ne sais quoi.
And so we, the way we quantify that is actually almost explicitly what you're mentioning, which is the gap between in-distribution performance and out-of-distribution performance. So what is the... We kind of measure how well does the model perform on tasks that are in distribution of our RL and post-training versus tasks that are out of distribution.
And what does that gap look like for... What does that gap look like between the Muse Spark line and the Muse Wonder line?
And we actually see this huge step up from the Muse Spark to the Muse Wonder line because of, as Mark mentioned, like, it's just a much better base model powering this new line. And there's an entire je ne sais quoi to that, but I think we've been able to actually quantify this effect. And it's an obviously dimension that we're excited to continue improving on.
But if you're saying that external benchmarks have ceased to become as important, and you care much more about your internal benchmarks. Is that kind of an argument against benchmaxing, that you think that just trying to get the top score on the external benchmarks that everyone talks about is maybe not going to lead you to develop a great model?
Well, yeah, but also we're focused on specific things. I mean, part of the whole strategy is personal superintelligence and just trying to build different types of products than other people are trying to build. So even if we did have the exact same evals as everyone else. It just—
And though I would say a personal superintelligence is not like a super niche interest, like it's probably something that other models want to be able to do. I think that there are things that are very different about it. I mean, I think like the personality probably matters more. I mean, one of the things that we talk about is this idea of discretion, right? So if you want to like use the model to get something done, It needs to be able to know a lot about you and get things done in the world while disclosing as little as possible. So, you know, if you're— you have an allergy or you're pregnant or something, you're trying to, like, get a, you know, like, a table at a restaurant and you don't want to, like, have to tell them your medical condition.
You know, to make sure that it's going to be a good place for you. Like, okay, how do you handle that? And I mean, humans have common sense and the models so far—
If you're building a coding agent that works in an environment where everyone has access to the same context, discretion isn't particularly important.
So there are all these things that I think are just kind of more important for us.
Aspects of coding where the code needs to work, but you don't— the consumers who are using it may not even ever see the code. I mean, that probably matters more in a personal superintelligence context rather than serving software developers. So I think these things are just a bit different.
Yeah. Ultimately, we care about making useful models. So we want models when you put them into people's hands, they have a great experience using. And I think that is the most important metric for any model that we or other people produce. I think that really focusing on the external benchmarks, if you really dig into the supply chain a little bit, it's some PhD student who made these benchmarks, a benchmark as part of their, as the course of them doing their PhD. And so you're kind of outsourcing all of your product thinking to that PhD student who happened to make a viral benchmark.
Right. So clearly that's wrong.
Yeah. And so that's why we care a lot. A huge part of what a lot of people at Meta and MSL are focused on is crafting amazing evals. And now that we have them, we can work on our own. But if you're building a you're ramping up your lab, right? It's like those public ones are not a crazy place to start, but it's just not where you want to end up. You're just at that point, you're just looking for signs of life. But then when you're like really in this mode, as we are now, where it's like, you know, we want to be very intentional about how we're improving our model over time to build better and better, better products. Like a lot of the cases that Mark just mentioned, like these are actually detailed very rigorously in like the spreadsheet of all the little capabilities and nuances of the product that we think really matter to the end user. And maybe each of them matters like 2% to the user, but in aggregate, that's the obsession that will ultimately result in these models that feel magical.
Can you guys describe for me just what goes into training an RL model? Or sorry, a leading-edge frontier model? I'm just curious in that I think a lot of people listening to this have a rough idea where there's there's a pre-training and a post-training step. So pre-training is you kind of ingest this huge amount of data. You look at a lot of pictures and these are cats and these are dogs. And after a while you start to develop a sense for what is catness and what is dogness and what is fur and everything like that. And so that's kind of the pre-training step where you just hoover up data and maybe you can talk about the data acquisition and just kind of what goes into that. And then there's this RL stage, which is the newish advance over kind of the past few years of AI. And again, it sounds like that is where a lot of the specific nuance of this model has come in, where you guys have chosen better domains, better hills to climb for the model. But just at a high level, describe how many people, how long it takes, what the steps are, how you think about those in—
How much time do you have?
This happens to be What we spend all of our time on.
Exactly. And so summarizing—
Yeah, I think at a very high level, these are the stages. It's pre-training and post-training. Ironically, training is the superset of all of it. But that's—
so how many people are on this whole effort, like when you develop these models?
So this, I think you're getting into what is interesting about MSL and the approach that we've taken, which is, I mean, From the very beginning, I think when I first reached out to you, I think we were just looking at this because we were like, just like, wow, okay, how far we've come. And I was just like, I pulled up the screenshot of like when I first reached out to Alex. My first thing was like, hey, do you have time to catch up and talk? And then it's like, all right, we talked for—
44 minutes or whatever. And then after that call, 2 days later, Alex sends me this long thing, which is like, well, here's what I think I would do if I were in your position. And the very first thing is that you don't need a lot of people, and talent density is the thing that matters. So I'd say, like, from the beginning of the plan, this was, like, the core of your strategy, right? It's basically, like, you actually don't need a massive number of people, but what you do need is for, like, each of the people to be as good as possible, which kind of informed this kind of crazy recruiting push and, like, very intense filtering that we had to take, you know, some of the best people from inside Meta and a very intense process to get people from outside Meta. But I actually think like a lot of what has made MSL special and why a lot of people have wanted to join is because I think it's in some ways it's just like more fun to work on a talent and smaller effort.
It feels like the startup.
Yes. That's operating. In some ways it has all the benefits of a startup with the stability and resources of a larger company. And this was like this environment that I think this was like point one in your kind of plan for what, um, for what Alex was like, this is what we need to go do. So I don't know the answer of like how big is a pre-training team, you know, it's, it doesn't need to be that big. But I think if you look at, you know, some of the other labs, um, it's probably quite big in some other places. So, um, I think part of the reason why, uh, these models are doing well is because you You have people who we're betting on and they have a lot of resources because we're taking our compute. We're not peanut buttering it across a lot of competing efforts. We're basically just investing in the key people, a lot of compute per researcher, and then they can do their work.
Was Wonder the one big training effort? That was where most of the training was?
It was the next pre-train. It was climbing up this ladder. So, I mean, we have this tiering system and then kind of Spark. I mean, internally, like, Avocado was the code name for the Muse Spark line, and then Watermelon was the code name for this one. And we're, like, out of fruits now.
Yeah, so we're going to have to switch to a, to a new convention. But, um, I think the only fruit that is technically larger is the durian, and you don't want to go to the durian.
Yeah, yeah, yeah. No, you want to stay far away.
A durian is bigger than a watermelon?
I think it's definitely heavier.
Spikier, yeah.
But no, yeah, I think as Mark mentioned, like, you know, each of the— I mean, each of the stages I'll describe, like, it's tens of people working on it at Meta versus, you know, hundreds at—
So Meta is a huge company, but this is like a really tiny effort inside of Meta.
Oh yeah, I know. The research team is less than 200 people.
It's intentionally very, very tight. I think that the—
It's closer to 100, right?
And then there's a bunch of InfoFolks. Yeah.
It's a lot of infra work. So I think at a very high level, to go back to the, to the, to the what is it question, you know, how does it happen?
How does model happen? That's basically what I'm asking.
Where do models come from? You take a huge amount of data for pre-training, and here you're—
And it's here, like, roughly, it's like all the, you know, public internet and like all the books you can get your hands on and just like as much data as possible.
But also different modalities, right? So if you want it to be good at perception, then you want to make sure it has a lot of images and audio and video and this kind of stuff. Um, you know, if you want it to be good at certain domains, you don't just want to RL that in at the end. You want to have it have exposure to that from the beginning.
Yeah. So, so, and in this stage, you're kind of, you're much less, you care a lot less exactly what the data looks like or how it's formatted or You know, all of everything about it. But it's just lots and lots of data.
When you talk about this being the best personal agent model, was there data you had to get for the pre-train where it's like huge amounts of restaurant reservation, call booking? I don't know what it is, but like—
You just kind of want to dump in large volumes of the modalities that people are going to care about.
I mean, there were a bunch of things. I mean, I don't actually think we talked about this when we were talking about where Muse Wonder excels. But some of the areas that it's like, all right, what do people care about, right? It's, I mean, there are things like shopping that people care about a lot, personal advice that people care about a lot. People have personal legal questions, personal health questions, personal tax questions. Muse Wonder, I mean, is like quite good at those. I don't know if there's any quantification that we want to give, but it's, let's say it's—
It's very good on those personal advice questions.
Yeah, and so those types of things, I mean, you can RL them to some degree, but you're like, just want to put in a large amount of data. Like, you can imagine health is a very— there's a lot of data out there about health. And, um, in, in pre-training, you just wanted to like see all these different things. And, um, so you, you kind of just put these different corpora in.
Yeah. So, so that— so as, um, as Mark was mentioning, like, the, uh, we, uh, one of the highlights of Muse Wonder actually is its ability on domain various domain expertises.
And I think our philosophy here is that, you know, the best personal agent should feel like a team of experts in your corner. Like, it should feel like it is your legal team.
And your finance team, and your medical team, and, and, and, and, and. And so, um, on a bunch of these domains, we, you know, Muse Wonder is actually state of the art on, I think, um, a lot of the finance evals, as well as a lot of the, um, uh, legal evals. the state-of-the-art model. And that, you know, these, these domain areas, that starts in pre-training. Those, you know, we look for a lot of domain data to put into pre-training.
But a lot of the, a lot of the art of pre-training— I mean, pre-training is, um, there's this one Rune tweet where it's like pre-training is full of the researchers who like chew glass and stare out into the distance and into the, into the abyss or whatnot. But, but I think there's a truth to that because it is like, um, It is both the highest stakes of the research areas because you, you know, we invested a huge percentage of our compute into training Muse Wonder, into pre-training Muse Wonder. And it was very high stakes. I think probably both of us at various points in that process, you know—
You went through troughs of sadness.
But yeah, troughs of uncertainties, you know, all the 7 stages of grief, you know, happen while you're pre-training a model. And it is the most sort of like fundamentally scientific in an interesting way, because, you know, you just don't have very many samples. So, like, your ability to infer exactly what will happen scientifically really depends on your ability to sort of like, you know, big brain reason about everything that's happening in this process.
Okay. So in the pre-training stage, everyone who's training frontier models basically is trying to get as much data as possible for the pre-train with maybe a bent towards If you want Muse Wonder to excel at legal and finance and tax questions and things like this, then you want to have as much of that specific corpus as possible in the pre-train.
Yeah, you still want as much intentionality about where ultimately you want your model to be, but again, it just, you know, benefits from a lot more data.
Yeah. And this is sort of the massive data, massive compute stage. Do you hit some saturation in the data you put in the pre-train, or ultimately is it just more data is better?
Well, there's a relationship between if you want the model to be bigger, you need more data. Otherwise, it overfits.
Yes, but my point is, like, is there data where it's like, we can't distract the model with that, we can't clutter up its brain attic with knowing that stuff, or is it just more data is better?
There are some of these effects, I think.
Yeah, because this is part of the science.
Okay, we're getting into the secret sauce, sorry. No, it's not even that, it's just like, I mean, I think that there are all these questions about, like, okay, well, we put a lot of these tokens in, is there any zero-sumness in a bunch of places there isn't, or it's actually positive transfer, but then like depending on how it's set up, like you can have some negative transfer.
And so I don't know, there are all these interesting questions. But in general, one of the bigger challenges of scaling bigger models is it's not just about how many flops and how much compute. You want to be able to scale data with that. I mean, this is like part of what, when the original kind of scaling laws for pre-training were written, I mean, this is sort of what came out of that is is that. So if the internet isn't growing that quickly and you want to scale bigger models that can be more intelligent, I think one of the interesting problems is, all right, where's the data coming from?
So I think that probably gets to the end of what we're going to talk about on that. But we could go to the other—
Excellent CEO energy over here.
But so post-training, so then I think Post-training is kind of— is a very exciting field now, I think, and is an area where I think actually, I think almost on a continual basis, the entire field is learning about the optimal recipes to be able to post-train these models.
And again, for people, like, post-training is where you give the model a task that you expect it to do, grade it on how it does. And so if it's a chess-playing model, it plays chess games, learns the—
Well, there's actually 2 stages. There's the SFT, So there's supervised learning. So that's like a lot of when you take this kind of raw thing and you give it both the question and the answer that you would want it to have, right? If like, if you're having a chat conversation, how would you want it to respond in this case? So when you're training things like personality, the SFT is actually like a very important part for molding it. But then I think most of the advances in the recent years have been in RL.
So I think one mental model that can be helpful is like, And SFT is where you kind of initialize a lot of—
SFT is supervised fine-tuning.
Supervised fine-tuning, yeah. Where you initialize a lot of the behaviors that you want the model to be able to exhibit later on. And then, you know, maybe at the end of SFT, it has those behaviors kind of baked in, but it'll only do it like 10% of the time. And then in reinforcement learning, you really sort of give the model reinforcement in various situations to bring the elicitation of those behaviors dramatically up. And I think that the— yeah, so I think there's a bunch of interesting things that are happening in RL and post-training. I think one is that just the scaling of RL and post-training has been one of the clear axes, just spending more compute and gathering more data. Just doing much more of it before you release a model.
Yes, doing much more of it, especially as the pre-trains scale up, doing much more. It's like anything in scaling laws, like it has to go with the rest of the factors or it doesn't work.
Yeah, in general. I mean, if I were to give, say, like one broad statement about AI, it's just scale everything, you know? But, um, uh, but in particular, you have to scale everything. Like scaling one thing in isolation does not work. You, you technically can, you can scale one dimension. You just will get much more optimal performance if you scale many things altogether.
Well, I mean, so one way to think about like Muse Wonder, right?
It's like, all right, so we did the pre-train. Yep. We've done some RL, right? When you look at it like Muse Spark, right? So version 1 to 1.1, 1.2, 1.3, we didn't change anything about pre-training, right? We just continued doing more RL on top of it. So, um, so you, when you think about the early versions of a model of a new class, they're actually relatively less, like the RL to pre-training ratio is going to be lower for RL than the later versions. That is kind of what the point releases in models are.
Well, I think every company has its own convention. For us, that's what it is meant to be. But for a lot of companies, like the big number is the pre-train and then the— I think it kind of lost that semantic value.
Yeah, I think now you actually can't do that. Okay, but that is true at Meta at least. But we haven't even done it. We kind of decided to not even do it. It's true at Muse Spark at least. It is true at Muse Spark and Muse Wonder. It's true that Muse Spark and Muse Wonder are different pre-trains. And we decided to kick the can down the road on this question for Muse Wonder by not giving it a number yet. But when we have our first dot release, we will have to decide that.
Yeah. But yeah, no, I think the RL stage, part of what has been— part of what's very interesting is first is just producing the environments and producing the tasks. And I think that is one of the areas where there's a lot of intentionality in training these models. It's also one of the areas where you notice very funny things. Like if you— in a lot of cases, less is actually more when it comes to data in RL. And so I think I think the sort of, like, the feel and the science of it, like, is pretty flipped, generally speaking, from pre-training.
Yeah. The other is that I think it can give very magnified kind of instruction to the model, if that makes sense. So I think that, like, a lot of safety cases or alignment cases that we care a lot about are, like, you know, and, you know, where you would see when you do large high-compute RL runs on frontier models, are where you apply a lot of RL or optimization pressure in scenarios where the only way for the model to succeed is to do something misaligned or do something unsafe. And so that's another thing that we had to be very careful about and thoughtful about in training Muse Wonder is ensuring that we curated the RL environments very thoughtfully to ensure that all of them were eliciting and driving towards more aligned behavior from the model.
So the point that you're saying about quality, I think, is an important one because I think a lot of the previous questions back and forth were about the quantity and have the flavor of like, oh, you need to scale the quantity. Yes. But I mean, I do think like a huge part of this is just, is, is the point that you just made of like the even a smaller amount of high quality.
Yeah, I think, I think in RL, like we, you know, as we've seen pretty big jumps in our new Spark line, I think these have been like accompanied with, you know, real meaningful improvements to the underlying recipe. And each of these improvements in the recipe have been like, I think the improvement between Muse Spark 1 to Muse Spark 1.3 feels night and day. I think probably to most people, if you ask them, do these feel like they came from different pretrains? I think they probably would say yes, because I think you can—
Mm-hmm. I think it's happening industry-wide for what it's worth. It's just like we are learning to squeeze so much capability and so much and elicit so much agentic capability from these existing base models. Now, that's not to say, like, we believe scaling is very important. That's why we have to use Wonderland in the first place. Yes. But, you know, this sort of like, these two things are happening at the same time, which is also, by the way, for us, this is why we're seeing so much progress in general. Like, I think it's the fact that we're scaling both pre-training and scaling our recipes in RL kind of like concurrently.
At a nice clip. Yeah. And then there's these other scaling axes too. I mean, like making sure that your agent harness is good and that you can co-optimize these things is probably just another aspect of, or dimension for scaling.
But if I'm understanding you guys correctly, are you saying that alignment is not about coming up with some alignment-specific RL steps and instead about making sure the entire RL process is kind of steering the model—
You have to be thoughtful about it in the entire process of RL. Yeah. So I think this is a holistic effort, I think. Yeah. But also, I think different labs have very different theories. I mean, alignment is a very philosophical concept, right? Of like, who do you want it to be aligned to? I mean, a lot of our theory is like, okay, we care more about it being aligned to, you know, we're trying to give people agents, right? And it needs to, if you ask it to do something, you need to have a high confidence that it's going to do what you ask it to do. Yes. You know, like, you want it to do that with high reliability. You don't want a moralizing lecture for why you're wrong for asking it to do the thing. Obviously, if you're trying to do something that's catastrophic, it needs to gracefully, without pissing you off, like, not do it. But in general, you know, we view alignment as more of a question of instruction and values alignment with the person who it's intending to serve. And I think that that's kind of a different mode of thinking about it. It's very different from, I think, how some of the others do.
But one of the more interesting things I think that is maybe a question related to your original question is also just how long the process of training one of these models takes. Like Muse Wonder, I mean, in general, we've planned Muse Wonder since last, you know, August.
Yeah, so yeah, I remember the offsite when we were— so this is like when they first presented this.
Yeah, so I think all of the stages that we just talked about are like, you know, you really work on them for quite a long time. And, and, um, it's not that we, we didn't start working on Muse Wonder when we finished Muse Spark. It's that we've actually just been planning and working on this for quite a long time. And we really view Muse Spark, if anything, as like I think we mentioned this in the first blog post where we wrote about it. That was a stepping stone to get to Muse Wonder. And that really was how we thought about it.
Even at the time you thought about it.
Yeah, and in terms of how long it takes, I mean, we're also even just talking about the research part. I mean, there's also the compute part where we've been investing in building up, first of all, a huge amount of compute overall. But a lot of, you know, part of our philosophy on this is that you can kind of measure the compute strength of a lab by how big its largest single computer is, supercomputer, for being able to do these large training runs. So, I mean, we built this Prometheus cluster, we call it, in Ohio. Mm-hmm. That's scaling to larger than a gigawatt. And now we're building this Hyperion cluster in Louisiana that's going to scale to 5 gigawatts. And those actually predate MSL. Because like building a whole like data center site from scratch, at least the plans for those, they're coming online now. But that was—
You guys thought it might be the feed, right? That you just have used for a lot of compute generally?
No, no, no. This is kind of predating, even going back to the, to kind of the LLaMA program. I see. I mean, it was probably in 2023, or yeah, either 2023, 2024 was when we basically started making the plans to build those clusters. They're now kind of fully coming online. Well, I guess the Louisiana one will come online over a couple years. Yeah.
But if you want to go really end-to-end on undoing this stuff, then, I mean, obviously the kind of full stack, and I mean, Meta is a full-stack tech company. I mean, we build our own data centers, we build our own chips, we build our own low-level infrastructure. And now this is kind of the next component, which, I mean, arguably is probably the most important component in all of this for the next wave of technology is kind of building really good models and then delivering them to people. But I mean, yeah, if you want to go fully end-to-end, I mean, this is a long process.
I do think this is like a— I mean, maybe it's well appreciated by folks in the tech industry, but it really is, you know, it feels like models are coming out constantly. So maybe you'd have the mental model that like, oh, a company like launches a model and then they figure out how they're going to make the next model and, or something like that. But it's actually these like very long planning processes and these sort of like very long work streams that kind of all come together at the very end and then result in these beautiful models. Yeah.
How do you build good products when the feedback cycle is so long? Like ultimately you start planning the model in August and you don't know until more than a year later if it's good. And, you know, Silicon Valley famously likes to run on metrics, but, like, ultimately, do you just need someone who is the aesthete and says, I'm sorry, I don't like it?
It's nice when you have metrics, but I think the, like, state of being an entrepreneur is, like, how well do you make decisions in an information-poor environment, right? So—
But this is, like, that on steroids, right?
Yeah, I think that this is— and I think Marc gets a huge amount of credit for this, like, because of how long everything takes, like, there were many points along over the past year since I've been here, but also, you know, as you mentioned, like, long before that, where you have to make very large investments or make very big decisions about, you know, how you're going to invest or develop the technology or, you know, what decisions you're going to make long before there's proof points around whether or not even the thing is working. And so I think this is like, in some sense, it's like entrepreneurship on hard mode. It's like the opposite of Lean Startup. Like, Lean Startup is like, yeah, you like see if you get something working and then you double down and then you see if that works, then you double down. Like, you cannot do that in AI. You have to like weave— I mean, we've been investing into these data centers since, you know, for years and years and years. We've had to—
And you'll be radically behind if you wait to see if things are working.
So we have these conversations where it's like, at the end of last year, we're like, all right, we are quite excited about the progress. Yeah, we haven't seen the models yet. Like, it's like, I'm going to, like, you know, we're going to write this check for, like, I don't know, like $100 billion, $200 billion. I'm like, Alex, are you sure this is going to work? It's like, um, yeah, so it was, it was so extreme that, you know, we had, we had like the whole exec team. I think Susan scared you a bit.
