Meta’s Path To AI Relevance, According To Meta CTO Andrew Bosworth
Bosworth on what went wrong with Llama 4, Zuckerberg’s AI “founder mode,” and why he thinks the model alone isn't enough to win.
Meta CTO Andrew Bosworth says the big monolithic model era is over, and the battle over AI product has just begun.
In an extended interview on Big Technology Podcast this week, Bosworth claimed that while frontier AI models are valuable, the way they’re implemented will create the most value.
“There’s a strategic construct of having a model, and having it be a truly state-of-the-art one, that’s super important. But having that alone doesn’t mean you win,” Bosworth said. “There are a bunch of pieces you have to connect it to: product, distribution, and the consumer experience. It’s the collection of all four that I think is our advantage relative to competitors, most of whom — whether it’s Apple, Anthropic, OpenAI, Google — only have one of those pieces.”
Meta has struggled to build its own frontier model, and it’s now renting models from some of those competitors, Bosworth said, in service of building “personal superintelligence.” The company also just released a new model today, Muse Spark 1.1, that it says is competitive on leading benchmarks at a lower cost than its rivals.
In our conversation, we cover Meta’s pursuit of the frontier, its latest wearable hardware, and the cultural challenges of meeting this AI moment. You can read the full Q&A below, edited lightly for length and clarity, or listen on Apple Podcasts, Spotify, or your podcast app of choice.
Alex Kantrowitz: Building a great AI model, in theory, takes a ton of compute and great researchers. Meta has both, but the leading model hasn’t shown up yet. What’s going on?
Andrew Bosworth: The only other ingredient I’d add is great data.
There are two stories here. The first is — go back to Llama 1, Llama 2, Llama 3 — we really were at the forefront and advancing things. The Facebook AI Research group goes back a decade or more.
And the real gap, which has been pretty public, was that when we were pulling Llama 4 together, we’d already pulled every stop we had into Llama 3, and unwittingly killed the pipeline.
The way it works is you build a base, you’ve got people pioneering an incremental version of that base, and you’ve got people pathfinding entirely new strategies. Unbeknownst to us at the time — and this speaks to the fact that we weren’t focused enough on it — Llama 3, which was a great model and well received, had pulled forward all the future bets. That meant when it came time for Llama 4, we didn’t have any of the pathfinding the other labs still had going. So now you’re behind on reasoning, behind on mixture of experts, behind on a bunch of the critical technologies that have driven the pace of progress.
That was a pretty public disappointment about a year ago, and it led to Mark shifting from “AI is one of our bets” — which is how we thought of it up to that point, just one of many bets — to “AI is foundational to the entire company.”
This is such a cliché, but I don’t have a better word for it: he went founder mode. He became uniquely focused on getting us all the compute we needed, all the talent we needed, the researchers we signed, who landed about a year ago.
I think Alexandr Wang just hit his one-year Metaversary, and I’ve loved working with him and learned a lot from him already. We’re seeing the fruit of that. If you look at Muse Spark, our latest model, it’s been very well received, and depending on the benchmark, it does really well on the things we care most about, the things we think are unique to our products.
“He goes founder mode. He really did flip into a mode that is like unique and reserved for Mark.”
So you’re absolutely right about where we are in terms of public perception. Model-wise, we’ve built a team I really believe in. We’ve got the compute and the data we need, so I’m confident we’re going to be where we need to be.
The Product Is the Value
I’ll add a second piece I think is strategically important: models are available — you can go rent a model, use Anthropic’s, use OpenAI’s, use Google’s. They’re great models, you can go get them. That’s pretty great. But the real value we’re going to create in the world is the product. The vision we have for personal superintelligence is one we’re uniquely suited to deliver. It’s not just that we have data — we actually have a better chance of understanding you, what you’re trying to do, and who you are in the world than almost anybody else does. So having the model is one piece — you want that strategically, so you’re not dependent on somebody else — but you mostly want to control your own destiny.
The model itself isn’t the value. I think we’re going to get to a world very soon where consumers don’t care which model they’re using — they don’t care if it’s 4.7 or 4.8, the same way you don’t care whether you’re using Oracle or a SQL database. You just want the functionality to work well. That’s the standard we’re all going to be held to. Today the discussion is about models, which suggests to me that we’re a little under-indexed on the user side of it — on how humans are actually going to benefit. That’s the story we need to tell, in addition to showing the technical work — we need to demonstrate the value to consumers
So the brute-force theory doesn’t hold anymore. To build frontier intelligence, you need more than compute, data, and researchers. You need techniques like mixture of experts and reasoning on top of the base pre-train. That’s what Meta’s working through now?
It’s not just that. The era of the monolithic model kind of died around the Llama 3 launch — the idea that there’s one model, you test how smart it is, and that’s how good it’ll be at everything.
We’re now in a world where these harnesses — Claude Code, Codex, whatever it is — are shopping underneath to lots of different models depending on the task. If you’re using Gemini, it’ll farm image-generation tasks out to Nano Banana.
We’ve moved past the world where one model rules everything.
We’ve moved past the world where one model rules everything. What you want is a very expensive, exquisitely intelligent model that you can distill down in interesting ways, and use only when necessary — because it’s expensive to run — with cheaper, faster, lower-latency models for tasks that don’t need genius-level intellect.
I really believe in scaling laws — you’ll see continued growth as compute scales up, raw model intelligence scales up — but human tasks don’t have infinite intelligence demands. A lot of human tasks can be done with conventional levels of intelligence. So there’s going to be a stratification. It’s not “what’s the one model that rules them all,” it’s “what’s the collection of models that come together to solve these problems with the right balance of performance, price, and value.”
The interview continues below…
How Databricks Scales Modern Identity Governance with Opal Security (sponsor)
Most identity governance platforms create more work than they eliminate. Databricks took a different approach with Opal, using automation and developer-friendly policy controls to manage access at scale without sacrificing visibility or control.
Automated workflows speed up provisioning and reduce manual toil.
Policy-driven governance scales access across complex distributed environments.
Unified visibility shows who has access, why, and for how long.
Read the Databricks case study.
Apple struck a deal with Google to build its product on a version of Gemini, and Siri seems to be working well with it. Would you consider a similar deal, while building your own model in parallel?
We use lots of different models today, and you want to give consumers the best model for them — there’s a price and performance calculus, and a latency calculus. But having your own model gives you the ability to not just control your destiny, but also have much stronger negotiating leverage on the deals you make to get consumers the best available answer.
Apple didn’t spend that much money. It was like a billion dollars to Google?
I don’t know what the experience is going to be yet — I don’t have access to it, so we’ll find out.
For us, at least, we’re talking about personal superintelligence — the ability to bring tremendous, specific capability to bear, not just general intelligence, for the products we build. We’re not seeing this as a value-add bolted onto an existing system — we’re seeing it as an entirely new way people are going to interact with their computers. It goes back to a lot of the work we’ve done in Reality Labs.
We’ve always tried to model ourselves after pioneers like Xerox PARC, the Stanford Research Institute, or Bell Labs — thinking about how we get information from our brains into the machine, hence our work on neural interfaces, and how we get information from the machine back into our brains, hence our work on AR and VR.
AI is potentially the best tool we’ve ever seen for getting information from our brains into the machine, especially if it’s able to observe what’s around us. Those are unique capabilities we’re trying to bring to bear — it’s not just the model, it’s the model’s ability to work with all these novel inputs and create a closed-loop system. We’re working on having incredible models, and I’m confident in the team we’ve assembled. My point is just that the model alone isn’t enough — and I don’t know if it’s enough for Apple to just rent one. I don’t know if they have a broader vision for how it integrates into people’s lives.
So you wouldn’t rent the model?
No, we do rent models —
From where?
There’s no reason not to. When we’re doing development internally, a lot happens on our own models. There’s also development we do on models we use from Google, Anthropic, or OpenAI. Being model-agnostic in a way that’s economically sensible actually hinges on having a competitive model of your own — one you can go back to if you need to. It creates a real backstop on how much rent somebody can charge you. But it’s also worth noting — whether I’m talking about a developer inside the company or a consumer — I don’t want them to worry about the model over time. Today they have to, because it’s all tightly tied together. Over time, they should just have a goal they’re trying to accomplish, and that should be their focus.
So there’s a strategic construct of having a model, and having it be a truly state-of-the-art one — that’s super important — but having that alone doesn’t mean you win. There are a bunch of pieces you have to connect it to: product, distribution, and the consumer experience. It’s the collection of all four that I think is our advantage relative to competitors, most of whom — whether it’s Apple, Anthropic, OpenAI, Google — only have one of those pieces.
Merging With AI
Last time we spoke, you said you wouldn’t merge with AI. But what you’re describing — getting your thoughts into a computer and back — sounds a lot like that. Have you changed your mind?
No. I don’t see this as merging with AI. I still want a very clear separation between things.
It’s really a continuation of a trend, an acceleration of a trend, where the bit rate between us and machines, and back, goes up over time. There are funny versions of this we’ve already been doing. Autocorrect is like a little AI that sits between you and the computer, reducing loss and improving that bit rate. There are all these little tools we use to accelerate the loop. QR codes are one of my favorites. I want to enter a URL, but I definitely don’t want to type one, because the error rate would be too high. So we use QR codes. If you have an AI that’s really able to understand things in human language terms, that’s a potentially profound improvement in our ability to take advantage of the compute we already have, even just on the input side. Combine that with the AI’s ability to synthesize information more effectively on the way back, and you’ve tremendously improved the bit rate.
This goes back to Doug Engelbart. When he left NASA to start the Stanford Research Institute, his idea was that human problems were getting harder at a steeper rate than human capability was improving. He wanted to create a human-computer symbiosis, and he believed the only way to do it was for people to merge with computers in some way. That’s why he led the first-ever video call, the first-ever joint document editing, the mouse — all of it came from wanting to increase the bit rate. I think AI is exactly that kind of thing.
AI Personal Assistant Competition
Don’t all AI products converge on the same personal-assistant use case eventually — OpenAI, Anthropic, Meta, Apple, all building toward the same thing? How do you differentiate?
I think everyone’s doing exciting work, and we’re at the very forefront of it, so it’s hard to say. The business Anthropic is doing, and that OpenAI appears to be increasingly pursuing under Greg Brockman, is an enterprise business — building harnesses, because that’s where the money is. I understand they need money, and it’s an important place to start. It’s attached to the enterprise, which is where the revenue is as a practical matter. There’s a lot of money in one place, so you have a smaller number of sales to make and can raise larger amounts of capital. Iit’s the capital-intensive game they’re playing.
Their major focus is on work use cases, and those are super valuable. We take advantage of them too, for our own professional work. But that’s not our major focus. Our major focus is 100% on how this helps consumers in their lives. I don’t think the AIs become indistinguishable from one another at all. There’s a real question — you framed it yourself — that these are kind of like a personal assistant with access to information about you that you wouldn’t want broadly distributed. It’s a trusted assistant. If you’ve ever hired a new personal assistant, there’s a ramp-up period. If you have one that’s actually embedded in your life and doing well, that creates a real connection — one that requires a lot of value from a competitor to go replace.
Why has consumer AI been so slow to take off, given how appealing it should be — entertainment, companionship, getting things done in your life?
The hype cycle is an evergreen concept our industry keeps falling for. People often misunderstand it: there’s a peak of hype, then the valley of disillusionment, then eventual product-market fit. The point of the hype cycle isn’t that the technology is fake. It’s that the people willing to go through a bunch of hoops to make it work are a relatively small percentage of the population, and the work of bringing it to everybody is genuinely hard. It’s not just solving the hard technology problem — it’s making the user interface workable, making it easy to use. People are living their lives successfully without this tool. You’re asking them to change their habits, to change how they deal with computers, in a pretty dramatic way.
It’s not enough to just point at the technology. You have to lead with value: what specifically are we going to do for you that makes your life better? My favorite example is agentic work. Like a lot of people in our industry, I was early — playing with agentic frameworks in December, and since — and I find them very powerful, but they’re not very user-friendly. They’re hard to build, hard to maintain, and they drift over time. I built one for my wife and me on a WhatsApp chat. She never uses it — I use it all the time, she doesn’t. It’s hard to integrate into a workflow; she just asks me to do things, I’m the agent, and then I go to the agent.
My point is: we haven’t made these things easy to use yet. We’ve done a great job on search and research use cases — people understand those. People understand generative AI for content, like making a funny image. There are a few use cases people understand and want to use now, but we haven’t done the work to make AI something people want to integrate into their daily life. It’s not easy enough to use, it doesn’t create enough value, it’s too fussy — that’s the problem to tackle. You need great models to do it, but great models alone aren’t enough.
AI Companions and the Future of Social Media
Where do you stand on AI companions?
We know personality matters a lot. I think Anthropic has learned this too, over the various generations of Claude — we care a lot, as humans, about the way natural language appeals to us or doesn’t, so personality matters for these models. That said, I think you’ll find a big distribution across the population. Some people would absolutely like the AI to be embodied, to have personality, to have a face. There are people in the agentic world who want to create 20 different agents, each with a different personality for different parts of their lives — a trainer, a nutritionist, a doctor’s assistant.
I’m not one of those people — I just want one reliable, trustworthy AI that does everything I need. It doesn’t need a human structure for me to care about it, and I definitely don’t want 20 of them.
I think you’ll see a big range in how people want to engage with this technology and what makes them comfortable. The market will deliver that.
There’s a future where these AI companions become — this is a blunt way to put it — the new social media. Social media is a place you go to see what’s going on with your friends and engage with it — it can be all-encompassing and, in its best case, fulfilling. Time spent is a pretty important metric, though how you feel after you spend that time matters too.
Time well spent.
Time well spent — and maybe that gets replaced by people spending time with some AI entity that cares a lot about them.
I try not to judge the way people choose to engage with technology. My instinct is that for the overwhelming majority of people, the major benefit of AI is going to be increased time for human contact with people they care about, people they love. I talk about this a lot in the context of augmented reality — even just the camera glasses we have. When I’m with my kids, I’m able to both record something and share it with my wife, which is meaningful to us, and also be fully present — I don’t have a phone between me and them, which matters to me. If you’re more effective at work, that’s more time not spent commuting, more time not spent away from your family, from the people you love.
My sense is that the value of authentic human connection only goes up over time for the overwhelming majority of people — it doesn’t go down — and I think we’re seeing that a little in how people are reacting to AI early on. People are worried it’s a replacement for connection. I don’t find it that way myself — I’m an avid user, and mostly I’m spending more time not having to be at my computer, thanks to it, not the opposite. That’s my prediction for how it affects most people’s relationship to media and to the people they love — a premium on authentic connection and authentic human moments. But I’m sure the entire distribution will exist.
And the AI glasses are core to that vision.
That’s right.
Please make the case: Why do we need AR or AI glasses when we could just use our phone?
Phones are great — I love phones, I have two of them, they’re wonderful devices. From the very beginning, the question we asked ourselves with glasses was exactly that: phones are great, what’s something you wish you could access without having to take your phone out of your pocket? We came up with a camera and audio — very simple. AI has been a tremendous tailwind, because it unlocks a much larger swath of potential capability over time than what the phone can do through a Bluetooth connection alone. So it’s much more promising now than it looked two or three years ago. Back then it looked like, at some point you’d have to put a display on this and it would have to become a standalone system with a bunch of accessories. Now it looks like there’s plenty of room in the market for a big range of wearable devices — glasses, certainly, and probably other things too. Some of those devices will just be input and output for your phone — that’s cool, your phone’s great, and if it’s making your life more efficient in terms of input and output, that’s awesome. Some will be more complete. For the Meta Ray-Ban Display glasses, for example, we just launched a vibe-coded platform — anyone who wants to can build whatever app they want for the glasses. Right now you build the app and put it on the glasses, but in the future there’s no reason that couldn’t happen in real time, with you telling the glasses what app you want and having it build that app for you on the fly.
I think we’re headed toward a zone that’s a little less app-garden-specific. You’ll still have content homes — content continues to be an evergreen, important thing, as it’s been on TV, on social media, everywhere — so there will still be places where the media you want to reach lives, and those look kind of like apps or channels. But there’s a long tail of things — why does my toaster need an app?
I don’t think it needs one.
I don’t want it. I just want to tell my AI agent, get me the toast I want — it’s the same toast I have every day, just get it for me. I don’t want to go through whatever the steps are.
What does your toaster app do?
I honestly refuse to install it. I absolutely won’t do it
You have to stand up for something.
There’s a line.
It’s sometimes so cool that you can have a specific app to control every aspect of a thing, and I respect that — I’m a tech guy, I like the fidgety nature of it — but it’s kind of gotten out of hand. What I really want is to tell an intelligent system, get me the thing I want, and have it do that for me. We see an early form of this with our Spotify partnership — you ask the glasses to play music, and if you have a Spotify account linked, it goes and gets the music you want. That’s what I wanted — I didn’t want to go through a bunch of steps. So the way I’m thinking about this isn’t that phones aren’t great — they are, and they’ll continue to be — but I don’t think the app model is the way the future looks. I think the future is valuable services provided to you, and you getting access to those services the way you want, when you need them, paying the people who provide them, negotiated either in advance or on demand. I really believe in this.
I noticed the Meta AI app now has a Garmin connector. I’d love to just say, “find me a 5K in my area this weekend and sign me up” —
Totally.
— and do it while I’m on a run, instead of spending an hour figuring it out myself.
Agreed completely. And taking it to a higher level — your Meta AI would ideally already know you’re training, that you have a goal you’re trying to reach, and it’s tied into all the pieces that matter — your nutrition, and so on. That’s the direction we want to take this. There are a lot of steps between now and then, but that’s where we’re going.
Orion and the Future of Full AR
Where do the Orion glasses stand? Those are the full AR glasses we talked about last time.
Orion was such an important moment for us. Having had this AR vision for such a long time, it finally gave us a device we could use to start playing with the software. Even though we couldn’t get the price to a point where we felt comfortable launching it as a consumer product, we designed and developed it with consumer intent — it’s quite wearable, quite workable. I have a pair at home, and we use it to test software. We’ve continued to iterate on the software, and we’ve made a lot more progress — not just because AI has gotten better, though that makes a huge difference, but also because having Orion to develop on makes a big difference.
We continue to be focused on the entire spectrum. We’ve hinted that in addition to display glasses and camera glasses, there’s a whole range of glasses that may sit below that in price. I still believe in full AR as a future for the space. We’re going to continue the same approach we’ve taken so far, and the same reason we didn’t launch Orion — it’s not enough that it does all the functionality, it has to look great, be comfortable enough that you want to wear it, and be at a price point where a reasonable person would say, yeah, this is a good value.
How far away is that?
I’m not going to give an exact number — I’ll say I like the progress we’re making.
Measured in years or months?
I’m not going to answer that.
That’s fair — I had to ask.
I appreciate the hustle.
I know some of my reticence — people who’ve been at companies like ours know this — we’re constantly looking at vehicles and asking ourselves, is this the one, is it ready yet? And, man, we’re getting into the zone. It’s pretty exciting.
Inside Meta’s AI Culture
Let’s talk about Meta culture. You’re running this applied AI division —
I run the Agentic Transformation Accelerator.
Right — which has been the subject of some reporting.
One of the groups in that is the applied AI team.
A Wired source called it “literally the gulag” — zero purpose, barely interacting with anyone, just tasks every week. What’s going on there?