I'm Ina Fried, chief technology correspondent at Axios, author of our daily AI Plus newsletter, which if you don't get or any of our newsletters, you can go to axios.com/newsletters. It's free. I'll be in your inbox every day for better or worse. I'm very excited about our final conversation this morning. I've known Andrew Bosworth, the CTO of Meta for a long time and always has a very smart take not just on where we are, but where things are going. So, please help me welcome Andrew Bosworth.
So, you guys have always said the metaverse was a long-term bet, but a bunch of cost cuts last week suggesting, you know, some recalibration was needed. Why? Why the recalibration? What caused it? What does it free up?
Yeah. Well, I want to talk about three different layers of this. The first one is our construct of the metaverse. And this is on us. We've never communicated as well as we wanted to. You know the old saying, the future is here. It's just not evenly distributed. We've always thought when you're looking at your phone, and we've all had this experience, especially if you're at dinner with a teenager, someone's looking at their phone, they are not with you. Like physically, they are there, but mentally they are transported someplace completely different with other people. So, we wanted to bring a rich version of that to life and we still plan on doing that. There's two things. It's like any investment. You're going to look at how you do over the course of years and you're going to reinvest in some areas and trim your losses in others. For us, we're seeing tremendous growth of our metaverse on mobile. You know, Horizon is this thing that started on VR headsets, but obviously there's much more users today on mobile phones. We've been pivoting over the last year to focus on the mobile market, and it's going really well. And so you kind of want to double down on that. Having to build everything twice, build it once for a mobile phone and once again for VR is a tremendous tax in the team. You'd rather just build grow a giant audience and then work from both places from a position of strength. The second piece is on the VR side, we really had this integrated concept of where Horizon and the VR vision were the same thing. So when you put the headset on, you're immediately in this kind of co-present accessible space. That is a real challenging piece of work to land from a standpoint of there's lots of people who put this headset on for lots of different reasons. You want to support all those different use cases. And the lack of focus comes at an expense of user experience and a great expense in terms of development cost. So what we're doing now is that we have two much more focused bets coming out on the other side. VR, we're going to let VR be what it is, what it does great. We're going to have focus a lot more on the third party content library, the ecosystem that's developed there. And on Horizon, we're focusing a lot more on mobile. To do this, of course, it's, you know, tragic. Anytime your plans change and there's a human cost, we found a bunch of roles that we just didn't need anymore. So, we did end up downsizing the effort on the metaverse specifically. Though on net, Reality Labs isn't downsizing. We're taking basically taking all of those and taking the investment on wearables, which is growing so rapidly for us.
Yeah. And that's where I was going to go next.
I'm wearing a pair of the Meta glasses. I use them to capture my softball games. I was playing ice hockey over the weekend.
That's been a, you know, bigger than you thought and bigger than you thought for a couple years now. What did Meta learn? Did you learn something from VR that made that right or was it just the right, you know, you hit a price point? I mean my what I always suspected was what you managed to do in the glasses that you didn't do in the metaverse but frankly few people really do in hardware is say what can we deliver today in the form factor that people want versus what is the form factor we have to deliver in order to have all the features we want.
Yeah there's probably a deeper lesson that I need to introspect on more there which is with VR we've always wanted it to be all things to all people which is always, you know, anyone who's developed a product knows that's a challenging place to start from. And I think we're getting more focused now with the glasses. We never burdened ourselves that way. We were content to build a small loyal audience for people for whom the having the audio and the camera was enough. We got a little lucky there. We built a great product which I'm convinced just those merits you and I would have used it. But then AI hit right at the same time. And you know, people don't realize this. When the Ray-Ban Metas came out and they launched with an AI assistant on the glasses that could access the camera, we didn't plan that assistant when we built the product. We landed that in the last four months before launch. It was never theorized. Tremendous tailwind to have arguably the best interface for AI, contextual AI in the real world, already kind of built on the pipeline when you can bring the software to bear on it. So, we got a little bit lucky there and that's been the big difference maker for us. Now, when you look at where we are with Ray-Ban Meta 2, the Oakley Vanguards, the Oakley Holbrooks, these are all products that are the consumers are coming because they like the AI features and they're staying because oh, you've built a great product here that does a bunch of other things really well like photos and I can take my phone calls and I can listen to my podcasts and I can listen to my music.
So, I kind of come at it the opposite way. I've been using it since Gen one, but you know Meta has an incredibly sweeping policy around AI and data. Basically, the policy says Meta can use that information in almost any way imaginable. And so I actually don't use Meta AI at all and I don't use a feature that would allow Meta AI because essentially the photo that I take, if I ask a question about it, Meta can use that to train its systems. It can advertise me in any way possible. ChatGPT is obviously moving towards ads, but they laid some ground rules of we're not going to use this this way. Do you think given where we are and the broadening adoption, you need to narrow that privacy policy so that someone like me can trust that I can share a photo without it becoming training fodder for your AI?
Yeah, we've got a this is a great market test happening right now in terms of what people are you there's there's some value being created and of course, you know, the AI understanding the types of questions you're asking the specific context around those questions means that in the future the AI is going to be better at answering those types of questions. So you really run the risk of actually robbing future consumers of value by not allowing that kind of training. I do think over time the technology behind AI is going to evolve a lot to allow much more personalized model development, model training, segmentation. Today that's not really possible in the way that models are built and in and inference is run. More and more is being done with context though and more and more can be done on device. Better and better models are becoming available on device. So over time I think there's three different ways that we see this playing out. One of which the way that we do it today is there's a bunch of anything that's happening on device is kind of totally fine. Like that's a device that you bought, it's happening between your phone and your glasses and nobody needs to be worried about it. That's going to be pretty safe and secure. If it's going to leave the device and go to the server, then you get a warning from us. It says, 'Hey, you've got to opt into this because it's going to the server.' Once it goes to the server, that's some risk. That's some risk that you have. Then there's a third tier which is like, okay, when it's on the server, is it staying there and like you guys are having it and it's training on it's part of the data or is it ephemeral? It's you're using it for now, but then it's going away. I think a lot of customers we find for example with yeah I'm okay with you sending this image to the server to run inference on it and then we we remove it. So we run we have versions of all of these different we have on device running we have temporary inference running and then you have permanent training sets running we don't currently have a single policy that unites all this across Meta AI as you use it on your phone versus Meta AI as it's used in the wearables and as you point out different labs are taking very different approaches so there's like a market test running right now to see what consumer preferences are and how that plays out.
Because you know today when you know say you have a strange looking sore on your leg and you're like, 'Hey Meta, what's this?' I don't necessarily want Meta to know that about me forever. And I'm curious, you know, do you think there are categories of information that you Meta will eventually say, you know, we won't use this information? Because again, it's not just for training algorithms. That's one concern, but also you all are great at this. You've built an incredible business. Facebook and Instagram know me super well. They were trying to sell me Diet Coke Crocs yesterday because they know I like Diet Coke. I'm actually fine with that.
No, they were Crocs still. So, you got a little work on the algorithm.
That's a balance. Yeah, I understand.
