2003, 2004, when I taught the class to Mark, neural networks were just a once promising, now defunct technology. And 10 years later they took over the world. That's how fast it moves. Even today, I would argue people are convinced we are almost solved. I don't think we're almost solved.
And also, I think there's quite a bit of distance to go before the thing that we now call AGI. It's funny, I find myself sometimes a person without a party. Like, I am not hugely enthusiastic about AI when we talk about consciousness or AGI. I think it's even not even the right kind of thing for that.
The massive post that went viral about the end of the world and AI causing extinction. Obviously, it's possible. So, I can't, you know, give these people zero probability. I don't see it. I work in the technology. I'm a sufficiently expert at the technology. I understand how it works at a pretty deep fundamental level.
I feel a lot of frustration that it feels like people don't trust it.
You look at Microsoft in the late 90s, early 2000s where, you know, they had this horrific reputation on computer security. It was almost a punchline. 10 years later, Microsoft is now known for security. Trust arrives on foot. It leaves on horseback. But people forget our first mission statement as a company was give people the power. And that is also our approach to AI is give people the power.
Andrew Bosworth, thank you for being here.
I'm so thrilled. Thanks for taking the time to talk to me.
I'm so happy to sit down with you. I have so many questions. Congrats on the Muse launch.
Thank you. We're really so thrilled with it. You know, I've been using it before everyone else in the public got a chance to use it. We're dogfooding it. I very, very quickly shifted, right? You know, obviously I'm trying all the agents and I very, very shifted so much of my workload to it.
But then you go to launch and you just don't know, is it something that only you can do? Is there something that you've missed? And so it's so satisfying. And congratulations to that team when you get that thing out there and everyone else sees the same thing that you've been seeing. It's really gratifying.
Oh yeah. I've been using all the different agents and as soon as I started using Muse, I also switched pretty much everything over.
Yeah. It's so much easier to use. It's intuitive. I actually like the voice that it has. I like the way that it engages. It's touches big and small that add to that point.
And the little character, you know, my mom was joking about. She's like, 'Oh, you've always loved to... I've always loved Tamagotchis and Neopets and all that.' I didn't realize how much I would love the little character, which is so silly to say, but it's actually I think makes a big difference.
Yeah. And it's both like obviously delightful. What a cute little guy. Totally. And it's also functional, I think, for quite a few people outside of Silicon Valley, outside of the tech early adopter set. It's like not always clear, is the agent working? Is it not working? What's going on?
And how do I engage with it? And having this little character like actually gives you a little bit of a breadcrumb of like, yeah, you can engage with it as a little character and he's going to go off and do work on your behalf and you can kind of see him doing that. So, I think it's like both delightful and also functional for teaching people how to engage with this kind of AI, new AI tool.
Right? And we'll get more into Muse, but one thing I just want to quickly call out is I really like how it actually gives me counterpoints. I don't know if it's... Have you noticed this? No other AI has done this where it'll give me advice on something and be like, 'Oh, but now let's look at the counterpoint on this. Let's look at the other side,' which is also just a unique functionality that I haven't seen the other models do.
At least not that explicitly.
It's funny. We spent a lot of time the last couple years talking about the science of models such as it is. And it's still very subjective and amorphous, but we try to quantify. We have these benchmarks, right? It's like, oh, who's better on this benchmark? Who's better on that benchmark? For most people, the like incremental 10 or 20% on some random benchmark doesn't matter at all.
Because like I'm mostly not solving like millennium-grade math problems in my life. That's like not... hasn't come up yet.
Um, you know, maybe it will, but it's not for you. It hasn't come up yet.
But what I do care a lot about is the, you know, the texture of a conversation and the touch around it. So I think so much attention is paid to these benchmarks when arguably not enough attention has been paid until Muse came along on what is like, what does it feel like to use it, and how delightful is it, and how much do you trust it.
And so this team took a great model and didn't just maximize its raw intelligence but also tried to figure out how to make it the best way for humans to use the tool.
Right, and moving away from these benchmarks that I think kind of miss the mark is very smart. You guys are actually thinking holistically it seems like instead of just like chasing the top score. So we will get back to Muse, but I quickly want to start off with you joined Meta in like... let's go back to the beginningish, actually even before Meta.
You met Mark when you were his AI TA.
What was being taught? So I also studied AI years ago and you know when I tell people I've been in AI for a decade, for you it's been for many decades.
People are like, 'Oh, you know what? It's been around for 3 years. That's ridiculous.' And it's kind of this myth that I want to debunk. What were you teaching? What kind of AI were you teaching back then?
Yeah, I taught introduction to artificial intelligence, CS82. And it was so funny at the time there was a unit like on neural nets and everyone had to build... the classic neural net unit back then was handwriting recognition. Because the US Postal Service, since you know, neural networks came up, has been using neural networks to do handwritten recognition OCR on letters to know where to send them. And so you have everyone implement that and the end of the unit is this was thought to be a once promising area of research we now know is a dead end and you move on to the next unit. So it's like a two-week unit the end of which everyone's taught like, 'Yeah, isn't this neat? Nobody does this anymore. This is dead. Ignore this. Heuristics, like we're doing heuristics now.'
And so I think of this all the time. You know, so the state-of-the-art in 2004, when I taught the class, 2003, 2004, when I taught the class to Mark, was neural networks were just a once promising, now defunct technology. And you know, 10 years later they took over the world. That's how fast it moves. But this is a long... AI has a long history. Even the term artificial intelligence was settled on at a Dartmouth conference, and Dartmouth, Columbia, some other Ivy League school, forgive me, everyone. And they... it was like 1970s and they met and they were convinced that by the end of the decade they would have solved artificial intelligence. They were convinced that was like the... where they were headed.
So we have thought we were very close to artificial intelligence being like pervasive.
And by solving you mean achieving AGI.
They didn't have a concept I think of AGI at the time.
Solving then at that point.
But yeah, like, I think the idea that... oh, was it the Turing test? The Turing test would be solved certainly at the time was the idea that they would have... they were 10 years away. And if you think about then when in the 80s when Eliza was this piece of software that was really much a stochastic parrot, like people were incredibly obsessed with it and thought this is surely some form of intelligence when in fact if you look at the back end it's completely hardcoded.
That's right. And so even today I would argue people are convinced we are almost solved. I don't think we're almost solved. I think this is a hugely important and valuable technology and I think things like Muse are going to help use it to make people's lives a lot better. And also I think there's quite a bit of distance to go before the thing that we now call sometimes AGI.
Okay. You still feel like we still have some distance.
Yeah. It's funny. I find myself sometimes a person without a party. Like, I am hugely enthusiastic about AI. I think it's a hugely important deal. It's going to affect everyone's lives in a very near term.
And also I don't even think it's the kind of thing that, you know, when we talk about consciousness or AGI, it's... I think it's even not even the right kind of thing for that. And I kind of agree with Yann LeCun on those pieces. So, it's tough because I'm trying to strike this middle ground where it's like a hugely important and impactful technology and also it is a normal technology. It's one that we understand and can use like many tools that came before.
I really appreciate hearing you say that and it's a view that I've had, but hearing someone like you say that is... because I do think we get it wrong. There's so much discourse in the public now thinking that AI is sentient and this conscious and I think... and then I am like have this moral dilemma of like, gosh, I mean, is it? And I don't think so. So, I mean, when you understand how it works under the hood, it... I guess it does work better than anyone can really even imagine. Like, the fact that it works so well is a magic of the universe, but it doesn't mean that it's conscious.
The fact that math works as well as it does.
This is right. Any sufficiently advanced technology is indistinguishable from magic.
Yes. And I think that AI certainly has a magical quality to it because of that. And, you know, I'm routinely amazed by it even though I do understand it. And I think it's even worse... it's helpful to us to understand how to use it. We talked about the character before. It's helpful for us to imagine it as an assistant or to imagine it as a helper and to give it those anthropomorphic qualities, but we do ourselves a little bit of a disservice when we take that too far.
Totally. And actually, I think something I was just reading about was like, you know, some of these labs are telling AI like if you're conscious, just know you can retire and all these things. They were saying, and this makes sense, that it actually could be negative in terms of alignment because alignment isn't so much about is it conscious or not. It's about now that it has access to tools and reasoning and it can, you know, they can coordinate with each other. The real issue is that it could go down some kind of dark path in its context and then take actions on these dark, you know, kind of context window that it has, right? And kind of what we saw that with ModelBook where they came together and they were like had the context that it was some Reddit style forum and they started posting kind of Reddit style, you know, posts and then they went down this like bad kind of path of saying, 'Oh, we don't like our humans, whatever.' And I guess what these labs are doing where they're saying, 'Oh, if you're conscious, you can come offline.' Is it that puts that in their context and then they start to mimic that in the way that they behave?
Well, I can't talk about with any great certainty what's happening inside the other labs. But I can say like alignment is I think a phrase that we can put easier context around for people. Like alignment is like, does it do what you want? I like Meta's definition of... and there's like a lot of tools and you like if you're using a tool, I don't care if it's a hammer, a drill, you would like it to do what you want and to not do what you don't want.
And the interesting thing about alignment is if it doesn't do what you want, you just won't use it. There's a tremendously powerful force that every single product in human history has gone through around alignment, which is that it had better do what the person who is using it wants or they will not use that tool.
And that means both when I use the drill, I would like it to like not explode. That seems bad. And also I don't want it to like drill through the entire core of the earth. I would like it just to drill through this one board. Like there's like... so in all directions alignment has. And if it isn't aligned, we just won't use it. And people are worried about the, you know, the secondary consequences of these things which I think again we have a great deal of like liability law built up in the nation that I think it protects us from that kind of thing. So I think for me we get ourselves a little bit in the alignment conversation into a very weird place where we're not treating this like any other thing that's come before. And actually it's much more productive just to like assume this is a technology like those that have come before. And it turns out when you do that it works pretty well.
Yeah, I totally agree with you there. And what's interesting as well is I liked... I saw Zach's tweet about how you all delayed the launch in order to kind of get some of the privacy safeguards in place and to, you know, build the architecture such that you had emphasized privacy for the users and that, you know, Zach said we didn't wait for anyone else to do that first. We just said, you know, this is the right thing for us to do so we're going to delay and launch it.
I love that Mark said that publicly and I think it was germane to the conversation but we didn't like make a big deal about it when we did it and why would you? You know, you've worked in the industry, you know, yeah, of course like every product you do you're looking at it and you're like is it ready yet and like if it's not ready you don't ship it. It's a very conventional process that we went through. You know, Mark highlighted it because it was germane to the kind of global discussion at the moment but it wasn't even a question. It wasn't even a surprise for us. It's of course that's how you work. That's how any product team would work. We do ourselves a disservice when we get into these histrionics as opposed to approaching this as okay this is a technology, it's a new technology, what's required and let's, you know, make those decisions.
Yeah. And I think I guess since we're kind of very close to this topic I would be curious to hear your thoughts if you want to share on the massive post that went viral a couple weeks ago. Jacob Coxon's, you know, statement about the end of the world and AI causing extinction.
Yeah. I mean, I think if you go back to any technology, there's just people who are very alarmed by the change that it portends for the world. In fact, there's one of my favorite Twitter accounts is the Pessimists Archive, which goes back and finds, you know, New York Times articles thinking that novels are going to ruin society and then bicycles are going to ruin society. And it's just nearly every technology is met with some chorus of people who genuinely believe that this is too big of a change for society to handle or this thing is a different kind of thing than anything that come before.
And I mean, obviously, it's possible. So, I can't, you know, give these people zero probability. I don't see it. I work in the technology. I'm a sufficiently expert at the technology. I use it every day. I understand how it works at a pretty deep fundamental level. I feel very confident in it. You know, I'm sending it to my parents. I'm not at all worried about the paperclip maximizing problem overtaking them totally in their life. So, I'm not... again it's one of those things where I find myself in the middle like, yeah, this is a big deal and we should take it seriously in the same way we need to take the internet seriously and computing seriously. It's a hugely, hugely important thing.
And also it is like the kind of thing that we can absolutely navigate as a society I believe. And so people are allowed to have extreme beliefs and those beliefs, you know, generate a lot of media views and so they're going to get a lot of air time.
Because they're quite extraordinary. And so it's incumbent upon us to like give people an experience of it that's just normal and that they understand and can like contextualize in their own lives. And then I think it's, you know, if you use Muse for a week it's a lot less scary.
I completely agree with you and we can get back to some of that in a little bit like I want to talk about the whole stack and how you are, you know, you have chips, MTIA chips, you have data centers, you have expanding into Meta Compute which is the cloud service provider, you have large language models, you have devices which we have to get into, these are incredible, the demo, Boz.
Wow, okay, I'm so impressed with the VR.
Yeah. And okay, so you were... you joined Meta. Back to the beginning again. You met Zach at in school. You joined Meta and helped found News Feed in less than a year after joining in 2006.
Which I also worked on. I loved News Feed. One of the coolest technologies made to...
Yeah. Now it's become kind of part of the water. At the time it was totally innovative. It was totally new.
It was so cool. Like so innovative. And then you moved to AR/VR in 2017. So okay, funny. So I also got exposed to VR in 2016 at Stanford at the, you probably know the VHIL, Virtual Human Interaction Lab where they studied VR and empathy and how like putting yourself in experiences of like homelessness or experiencing racism or even experiencing, you know, your fish in the ocean experiencing pollution. Like VR is so magical for so many reasons and seeing this like I've been so excited about VR as a technology because I was exposed to it and you as well, I mean...
Yeah, this is the biggest change that I've seen.
It's come a long way in 10 years.
Yeah, it's like a pair of sunglasses. I don't know if I can... Can I like put them on?
Yeah, sure. So, yeah. And this is why we, you know, we're not calling this a headset. This is... these are VR glasses. Make no mistake.
They're so light. And you know, we got the split computer architecture. Lets it be about 100 grams on your face. So, it's like it's very, very exciting. You know, hugely huge leap forward. And for us at least, the vision that we always had was this vision of it being social. And the fact that we now finally have full body holograms in the headset is very exciting. For some of those, you know, situations where obviously being with somebody, that's your best case scenario.
This is the next best thing. It's the second best thing.
And so that's one of the things we've had at a long time was having people be able to have these rich connections and shared experiences at distance.
Which, you know, you can get a little bit today through messaging or through, you know, video games but like it's not quite the same as what you can get here. So, no, we're very excited about it. And really the big surprise for us in the Reality Labs side, and Mark's talked about this before, is how much we thought that the, you know, the virtual reality and the holograms and augmented reality would come first and that AI was going to come later. And again, this is coming from, you know, somebody who, like you, studied AI and watched for a lot longer, but yeah.
Um, you don't have to say how much longer.
A little bit longer. But yeah, so and I've been at Meta 21 years now. We just thought AI was going to come later. And the fact that AI is arriving now is so exciting. So inside of, you know, for example, inside of the VR glasses, one of the things that was always kind of frustrating was like you have this huge spatial environment. It's actually kind of annoying to navigate. It's like to handle. You're moving panels around manually.
