Meta CTO Andrew Bosworth: Our Path To Frontier AI, Renting Models, Consumer AI's Struggles

Meta Chief Technology Officer Andrew Bosworth joins us to talk about the company's AI efforts and why it's building its own new AI glasses. That's coming up right after this. Welcome to Big Technology Podcast, a show for coolheaded and nuance conversation of the tech world and beyond. We have a great show for you today. We're joined today by Meta Chief Technology Officer Andrew Bosworth, who's going to talk to us all about the company's AI efforts, its new AI glasses, the company's culture, and some big thoughts at the end. Bos, great to see you. Welcome back to the show.

Um, we were just talking before we started rolling about what a crazy moment it is in the tech world. We haven't seen progress like this as far as I can remember. The core part of it is the AI model. The AI model underpins everything. Without a working AI model or a leading AI model, it's tough to build. The theory for a long time was that to build a great AI model, you needed a ton of compute and great researchers to work on the algorithm. Meta has a ton of compute and a team of the best researchers to work on the algorithm, but the leading AI model hasn't materialized yet. So, can you talk a little bit about what you've learned there and whether that core assumption about what it takes to make great AI models is wrong?

Which is your latest model?

I just want to talk about the science for a moment. The thing I brought up in the beginning was the idea that you could brute force your way to a competitive model. The answer I'm hearing from you is not anymore, because there are new techniques like mixture of experts and reasoning that require refinement of the base pre-train to build top-tier models. That's what Meta is working through right now.

You said a couple interesting things. First, it's the product that matters, and it's important to have your own model for self-reliance. Let's talk about that. I'm sure you saw what Apple did with Google to distill Gemini. Early reports say Siri is working pretty well. Have you considered doing a similar deal with Google while building your own in parallel for self-reliance, to advance your products as fast as you can?

Spend that much money. It was like a billion dollars to Google.

I think we are working on having incredible models, and I'm confident in the team. My point is that it's not enough. Whether it's enough for Apple to just rent that model, I don't know if they have a broader vision for how it integrates with people's lives. Okay, so you wouldn't rent the model?

From where?

That you can go back to if you need to. It creates a backstop on how much rent somebody can charge you.

So there's this strategic construct of having a leading state-of-the-art model, and that's super important, but it's not like when you have that, you suddenly win.

I'm going to ask you again next time we talk.

Okay, so the way it manifests could be in this personal assistant that knows your context, goes out and gets things done for you. It could happen via a chat interface on a phone or computer, or through glasses like Meta is making. From a product standpoint, don't all AI products converge on this personal assistant use case? OpenAI is trying to create a super app, same with Anthropic, Meta, and Apple. How do you differentiate, and do you agree everything converges on this central assistant use case?

If you've ever had a personal assistant and hired a new one, there's a ramp-up period. If you have this personal assistant embedded in your life, it creates a connection that requires a lot of value to replace. Why do you think consumer AI has been so slow to take off? There have been attempts like character AIs and replicas. OpenAI pivoted from a money standpoint, but they have consumer applications like nutrition and health. You'd think consumer AI would be appealing for entertainment, companionship, and getting things done, but it's been slow.

The work of bringing it to everybody is hard work. It's not just the technology; you have to make the user interface workable and easy to use. People are living their lives successfully without this tool. You're asking them to change their habits dramatically. They mostly don't like it. You have to lead with value. What are the specific things that will make your life better?

You need great models to do it, but great models are not enough. Right. Where do you stand on AI companions? There's a belief that you build functionality and people will come. The other side is building an AI avatar that people feel friends with. Personality matters a lot. I will say that one thing we've learned is that humans care about the way natural language appeals to us. Personality matters for these models. Having said that, there will be a big distribution. Some people want an embodied AI with a personality and a face. Some want 20 different agents for different parts of their lives. I'm not one of those people. I just want my AI to be extremely reliable and trustworthy. I'm fine with it being an amorphous entity. I don't want to deal with 20; I want one that does everything. It's very early. You'll see a big range of how people want to engage. The market will deliver that.

You know, there is a future where these AI companions become the new social media. Social media is where you see what's going on with friends. It can be all-encompassing and fulfilling. Time spent is an important metric, but how you feel after is also important. Time well spent. Maybe that gets replaced by people spending time with AI. Ultimately, it's like how do we...

You engage with something on your computer? Maybe that gets replaced with people spending time with some AI entity that cares a lot about them.

Yep. And of course, the AI glasses are kind of core to that vision.

Hi everyone, Alex Canitz here. I want to tell you about a documentary I've made with Gravity to explore the future of AI agent security. To find out if we're truly ready for autonomous agents, I sat down with MIT professor Ramsh Roskar, former White House CIO Terresa Payton, Michelin's group chief data and AI officer Ambika Roger Gopal, and Sharon Guy, a former executive at Alibaba. They each offer unique insights into this evolving landscape. We conclude with Rory Blundell, CEO of Gravity, to discuss the path forward. With Gravity leading the way, join us on this journey. You can watch the full documentary at the link in the show notes.

And we're back here on Big Technology Podcast with Andrew Bosworth, Bos, the CTO of Meta. Bos, great to see you again. Thank you for taking the time to speak with me. If we go to the wide shot, we can see we're here in New York at a moment where you and your team are releasing three new pairs of Meta designed glasses. It's something we've been debating on the show: is your phone the AI device or is it a wearable? And we've had this moment again going back to Apple where it looks like they're preparing to release a version of Apple intelligence that actually works, that knows your context to a degree and might be able to get things done for you. And then we see the opposite side is the Snapchat specs release, which got a lot of people saying maybe we don't... I mean, those were so bad that people were just... You don't have to comment. I'll say it.

Let's just say my comment reflects what the market did. Evan Spiegel wore them out to some presentation. I think Snap stock went down like 6% immediately. It's just what happened.

Yeah. But I'd love to hear your thoughts. Obviously Meta has invested a lot in this. You believe it's a compelling use case. If I were to say maybe we don't need AR or AI glasses, we can just use our phone, what would you say that makes you feel the other side of that bet?

Interesting.

I don't think it needs one.

I honestly refuse to install it. I refuse to install it.

I refuse. I absolutely won't do it. And so...

Yeah. Listen, there's a line, there's a line that nobody...

And we see an early form of our partnership with Spotify. You ask the glasses to play music, and if you have a Spotify account linked, it goes and gets the music you want, and it's like, 'Yeah, this is great. This is what I wanted.' I didn't want to have to go through a bunch of steps to do this. So for me at least, the way I'm thinking about this is not that phones are great and they're going to continue to be great. I don't think the app-y thing is the way the future's going to look. I think the future is going to be valuable services that are provided to you, and you getting access to those services the way that you want when you need it, and paying money to the people who provide those valuable services, all negotiated either in advance or on demand.

Totally.

Yeah.

The Orion glasses, we talked about those last time. Where do those stand? Those are the full AR, full glasses experience.

How far away is that?

Measured in years or months?

All right, that's fair.

Have to ask.

Okay, cool. Let's talk about metaculture for a moment. You're running this applied AI division, that's right, which has been the subject of some reporting.

Okay. I'm just going to read the quote from Wired. One employee told Wired, 'It's literally the gulag. You have zero purpose in life all of a sudden. You barely interact with anyone. You just have these tasks every week.' Apparently talking about how the employees there have been put on some AI puzzles that they have to try to accomplish that helps train the AI. What's going on there?

Doesn't seem like it, but the fact that they would say that...

Yeah. So we've been spending a lot of time on this internally. It's a hugely important topic for us. You've been covering us a long time, so you know this: this is a company that goes into lockdowns when we have an urgent opportunity ahead of us. We did it with mobile, we did it with video, we did it with stories. Every one of these things pivots the entire company, but there are moments where we're like, 'Wait, if we put exquisite effort on something right now, we think there's a tremendous opportunity for us in the market.' In this case, we saw that when we came out with Muse Spark, and I want to be careful: Muse Spark is a great model and we're really excited about it. Coding had not been a focus for us on the model, but it actually was better out of the box at coding than we had expected it to be. And we found early on through experiments that actually giving it a relatively modest number of trained, expertly guided examples, we could post-train the model and dramatically improve its competitiveness. So when you start to run the numbers on this, you're like, 'Oh, this is an incredible opportunity for us to build a coding model that not only allows us to have independence in how we operate the company, but also something that we think is going to be valuable both inside and if you give users AI that's able to code, that's obviously one of the very powerful tools that's become very common in these AI systems over the last year.' And then also for us to be able to make the model itself more widely available over time. So we basically saw this huge opportunity, such a big opportunity that we pivoted on a dime and brought a lot of people across the company, thousands of people, out into this AI organization to do these expert traces. We absolutely need their expertise. It doesn't work if you do a bad job. It turns out if you use a bad piece of coding to train the model, you do some damage to it. They have to be well done. They have to be expertly guided. Now, we did it very quickly, and as a consequence, it did not have a lot of structure. It did not have great communication around it. I've been on record... Actually, that's not true. I wasn't on record. I was leaked calling it atrocious.

You said maybe not the worst it's ever been in 20 years here, but it's up there. It's definitely up there.

I didn't say that.

Yeah.

Can we talk about the tracking briefly? I actually, if I was an employee, I don't think I'd be a fan of it, but I actually sort of made the case for why you might be doing it on our show recently. And now that we're sitting next to each other, let's talk about it. So basically, the reports have been that Meta has started to track some keystrokes and the way that employees type, and basically use that as a way to train models. My perspective on this was as model training moves into reinforcement learning, where I think Scale AI, where Alexander Wayne came from, said most of their training is reinforcement learning now, as opposed to pre-training, which we talked about previously. As the technology moves into reinforcement learning, it's very valuable for these models to learn how to accomplish tasks in what's typically called gyms or different areas, simulations of real world activity that they go in and try to accomplish. So am I right in thinking that this program is basically a massively scaled up version of that, where the models watch employees work through their tasks and then learn how to accomplish tasks on their own?

So you are posting a lot of your old blog posts to Substack, and I've been getting them in my email and reading them. There was a very interesting one that I read recently talking about how you were doing some biology research and the doctor said the pain is rehab. You need that pain in order to be able to heal. You write that at some point you have to embrace the pain to make real progress. Given two otherwise equal stories, humans remember the story that evoked stronger emotion. Emotion is how our brain triages memories. Sometimes it has to hurt for your brain to prioritize it.

AI is evolutionary bio.

So this, I love this. Very small aside: my Substack, one of the things I did is I assigned my agent the task of bringing my blog posts over to Substack so at some point I could do both. I didn't realize until very recently that it wasn't any bulleted list. It would just strip out. My agent did not understand bulleted lists. So we have a long ways to go on agents, is my takeaway. Phase one. The pain is the rehab. There was a question we were studying the neurobiology that would occur during withdrawal from drug use, and a student asked, 'Hey, we have all these symptoms, why don't we just give people a pain medicine?' And the professor was like, 'You don't understand. The pain is the medicine.' Experiencing the desire to pursue drugs, drug-seeking behavior, and then having it be immensely painful is the way you reprogram your brain to overcome the drug-seeking behavior. And if you get rid of the pain, then the person is never going to do it. So this is a productive form of pain. By the way, I would argue AI, all these paroxysms happening not just at Meta but at every company, is the pain I'm talking about. That is the pain that there is no way out but through, and you have to figure out the path through it to figure out what works and what doesn't. It's just gritty. We do have lots of other types of pain in our society that have nothing to do with real value being created. This comes up a lot in education, is a good example. I remember being told, I'm sure you were, when I was in school, 'Hey, you can't use a calculator on this test. You will not have a calculator with you as you go about your day in the real world.' I have at least three calculators on my person at all times. Not to mention, I can just ask my glasses math problems. I'm filthy with calculators. It turns out doing a math test without a calculator is a certain kind of pain, not a particularly useful kind. Doing a harder math test that requires critical thinking with a calculator is probably the more valuable way to do that thing. So I do think it's important to align the pain that we're experiencing with the value we're trying to create in the world. I think learning to integrate AI, you could avoid that pain. You just skip it. You don't do it. You and I both know that puts you at real risk. You're going to fall behind people who are able to do AI and want to do the same job as you. You're going to fall behind other companies that have integrated AI either economically or in the products that you offer. There's this Cheryl Sandberg has this great quote, which is that companies don't usually fail by setting tough goals and missing them. They fail by setting easy goals and hitting them all the way down. So I think you could easily avoid the pain today by just being like, 'Yeah, we're just not going to do it. We're just going to let it happen and then we'll figure it out later.' So I think there is productive pain and unproductive pain, and maybe a little bit of judgment to know which one's which.

Bos, it's really always a pleasure to speak with you. Thanks so much for coming on.

All right, everybody. Thanks so much for listening and watching, and we'll see you next time on Big Technology Podcast.