Mark Zuckerberg: Future of AI at Meta, Facebook, Instagram, and WhatsApp | Lex Fridman Podcast #383

The following is a conversation with Mark Zuckerberg, his second time on this podcast. He's the CEO of Meta, that owns Facebook, Instagram, and WhatsApp, all services used by billions of people to connect with each other. We talk about his vision for the future of Meta and the future of AI in our human world. This is the Lex Friedman podcast. And now, dear friends, here's Mark Sackerberg. So you competed in your first jiu jitsu tournament, and me as a fellow jiu jitsu practitioner competitor, I think that's really inspiring given all the things you have going on. So I gotta ask, what was that experience like?

Oh, it was fun. I know. Yeah. I mean, well, look, I'm, I'm a pretty competitive person.

Yeah.

Doing sports that basically require your full attention, I think is really important to my, like, mental health and the way I just stay focused at doing everything I'm doing. So I decided to get into martial arts, and it's awesome. I got a ton of my friends into it. We all train together. We have a mini academy in my garage. And I guess one of my friends was like, hey, we should go do a tournament. I was like, okay, yeah, let's do it. I'm not gonna shy away from a challenge like that. So, yeah. But it was awesome. It was just a lot of fun.

You weren't scared? There was no fear.

I don't know. I was pretty sure that I'd do okay.

I like the confidence. Well, so for people who don't know Jiu jitsu is a martial art where you're trying to break your opponent's limbs or choke them to sleep, and do so with grace and elegance and efficiency and all that kind of stuff. It's a kind of art form, I think, that you can do for your whole life. And it's basically a game, a sport of human chess you can think of. There's a lot of strategy, there's a lot of sort of interesting human dynamics of using leverage and all that kind of stuff. And it's kind of incredible what you could do. You can, you could do things like a small opponent could defeat a much larger opponent. And you get to understand, like, the way the mechanics of the human body works because of that. But you certainly can't be distracted.

No, it's a hundred percent focus sport to, to compete. I, I, you know, I needed to get around the fact that I didn't want it to be like this, this big thing. So I basically just, I, I rolled up with a hat and sunglasses, and I was wearing A Covid mask. And I registered under my first and middle name, so Mark Elliott. And. And it wasn't until I actually like pulled all that stuff off right before I got on the mat that I think people knew as me. So it was, it was pretty low key.

But you're still a public figure.

Yeah, I mean, I didn't want to lose. Right.

The thing you're partially afraid of is not just the losing, but being almost like embarrassed. It's so raw, the sport in that. Like it's just you and another human being. There's a primal aspect there.

Oh, yeah, it's great for, for a.

Lot of people it can be terrifying, especially the first time you're doing the competing and it wasn't for you. I see the look of excitement in your face.

Yeah, no fear. I just think part of learning is failing. So I mean, the main thing, like people who train jiu jitsu, it's like you need to not have pride because, I mean, all the stuff that you were talking about before about getting choked or getting a joint lock, you only get into a bad situation if you're not willing to tap once you. You've already lost. Right. And. But obviously when you're getting started with something, you're not going to be an expert at it immediately, so you just need to be willing to go with that. But I think this is like, I don't know, I mean, maybe I've just been embarrassed enough times in my life. I do think that there's a thing where, like, as people grow up, maybe they don't want to be embarrassed or anything. They've built their adult identity and they kind of have a sense of who they, they are and what they want to project. And I don't know, I think maybe to some degree your ability to keep doing interesting things is your willingness to be embarrassed again and go back to step one and start as a beginner and get your ass kicked and look stupid doing things. I think so many of the things that we're doing, whether it's, whether it's this, I mean, this is just like a kind of a physical part of my life. But at running the company, it's like we just take on new adventures and all the big things that we're doing, I think of as like 10 plus year missions that we're on, where often early on people doubt that we're going to be able to do it. And the initial work seems kind of silly and our whole ethos is we don't want to wait until something is Perfect to put it out there. We want to get it out quickly and get feedback on it. So I don't know. I mean, there's probably just something about how I approach things in there. But I just kind of think that the moment that you decide that you're going to be too embarrassed to try something new, then you're not going to learn anything anymore.

But like I mentioned, that fear, that anxiety could be there. It could creep up every once in a while. Do you feel that in especially stressful moments, sort of outside of the jiu jitsu mat, just in work. Stressful moments. Big decision days, Big decision moments. How do you deal with that fear? How do you deal with that anxiety?

The thing that stresses me out the most is always the people. Challenges. I kind of think that strategy questions. I tend to have enough conviction around the values of what we're trying to do and what I think matters and what I want our company to stand for. That those don't really keep me up at night that much. I mean, I kind of, you know, it's not that I get everything right, of course I don't. Right. I mean, we make a lot of mistakes, but I at least have a pretty strong sense of where I want us to go on that. The thing in running a company for almost 20 years now, one of the things that's been pretty clear is when you have a team that's cohesive, you can get almost anything done and you can run through super hard challenges, you can make hard decisions and push really hard to do the best work even and kind of optimize something super well. But when there's that tension, I mean, that's when things get really tough. And when I talk to other friends who run other companies and things like that, I think one of the things that I actually spend a disproportionate amount of time on in running this company is, is just fostering a pretty tight core group of people who are running the company with me. And that to me is kind of the thing that both makes it fun having friends and people you've worked with for a while and new people and new perspectives, but a pretty tight group who you can go work on some of these crazy things with. But to me, that's also the most stressful thing, is when there's tension that weighs on me. I think maybe not surprising. I mean, we're like a very people focused company. And it's the people is the part of it that weighs on me the most to make sure that we get right. But yeah, that I'd say across everything that we do is probably the big thing.

So when there's tension in that inner circle of close folks, when you trust those folks to help you make difficult decisions about Facebook, WhatsApp, Instagram, the future of the company and the metaverse with the AI, how do you build that close knit group of folks to make those difficult decisions? Is there people that you have to have critical voices, very different perspectives on focusing on the past versus the future, all that kind of stuff?

Yeah, I mean, I think for one thing it's just spending a lot of time with whatever the group is that you want to be that core group grappling with all of the biggest challenges. And that requires a fair amount of openness. And you know, I mean a lot of how I run the company is, you know, it's like every Monday morning we get our. It's about the top 30 people together. And, and we, and this is a group that just worked together for a long period of time. And I mean people rotate in, I mean new people join, people leave the company, people go to other roles in the company. So it's not the same group over time, but then we spend a lot of times, a couple of hours, a lot of the time it can be somewhat unstructured. I'll come with maybe a few topics that are top of mind for me, but I'll ask other people to bring things and people raise questions whether it's okay, there's an issue happening and some country with some policy issue, there's like a new technology that's developing here, we're having an issue with this partner. There's a design Trade off and WhatsApp between two things that end up being values that we care about deeply. And we need to kind of decide where we want to be on that. And I just think over time when by working through a lot of issues with people and doing it openly, people develop an intuition for each other and a bond and camaraderie and, and to me, developing that is, is like a lot of the fun part of running a company or doing anything. Right. I think it's like having, having people who are kind of along on the journey that you're, that you feel like you're doing it with. Nothing is ever just one person doing it.

Other people that disagree, often within that group.

It's a fairly combative group. Okay.

So combat is part of it. So this is making decisions on design, engineering, policy, everything.

Everything. Everything. Yeah.

I have to ask just back to jiu jitsu for A little bit. What's your favorite submission now that you've been doing it? What's, how do you like to submit your opponent? Mark Zuckerberg.

I'm in.

Well, first of all, I. Do you prefer no Gi or Gi Jiu Jitsu? So Gi is this outfit you wear that is maybe mimics clothing so you can choke?

Well, it's like a kimono. It's like the traditional martial arts.

Or pajamas. Pajamas that you could choke people with. Yes.

Well, it's got the lapels. Yes. Yeah. So I like Jiu Jitsu. I also really like mma and so I think no Gi more closely approximates mma and I think my style is, is maybe a little closer to an MMA style. So, like a lot of Jiu Jitsu players are fine being on their back, right? And obviously having a good guard is, is a critical part of, of of Jiu Jitsu. But, but in mma, you don't want to be on your back, right? Because even if you have control, you're just taking punches while you're on your back, so that's no good.

Do you like being on top?

My, my style is I'm probably more pressure and, and yeah, and I'd probably rather be the top player, but I'm also smaller. I'm not a heavyweight guy. So from that perspective, I think especially because if I'm doing a competition, I'll compete with people who are my size, but a lot of my friends are bigger than me. So backtake's probably pretty important because that's where you have the most leverage advantage, where people, you know, their arms, your arms are very weak behind you. Right. So, so being able to get to the back and, and, and take that, pretty, pretty important. But I don't know, I feel like the right strategy is to not be too committed to any single submission. But that said, I don't like hurting people, so. So I always think that chokes are, are a somewhat more humane way to go than, than joint locks. Yeah.

And it's more about control. It's less dynamic. So you're basically like a Habib Nurmagomedov type of fighter.

So. So let's go. Yeah. Back take to a rear naked choke? I think it's like clean way to go.

Straightforward answer right there. What advice would you give to, to people looking to start learning Jiu jitsu, given how busy you are, given where you are in life, that you're able to do this, you're able to train, you're able to compete and get to learn something from this? Interesting art.

I just think you have to be willing to. To just get beaten up a lot. Yeah, I mean, it's. But. But I mean, over time, I think that there's. There's a flow to all these things and there's, you know, one of the, One of my, my experiences that I think kind of transcends, you know, running a company and the different, different activities that I like doing are. I really believe that, like, if you're going to accomplish whatever, anything, a lot of it is just being willing to push through and having the grit and determination to push through difficult situations. I think for a lot of people, that ends up being sort of a difference maker between the people who kind of get the most done and not. I mean, there's all these questions about how many days people want to work and things like that. I think almost all the people who start successful companies or things like that are working extremely hard. But I think one of the things that you learn both by doing this over time or very acutely with things like jiu jitsu or surfing is you can't push through everything. And you learn this stuff very acutely doing sports compared to running a company. Because running a company, the cycle times are so long. It's like you start a project and then it's like months later, or if you're building hardware, it could be years later before you're actually getting feedback and able to make the next set of decisions for the next version of the thing that you're doing. Whereas one of the things that I just think is mentally so nice about these very high turnaround, conditioning, sports, things like that is you get feedback very quickly. It's like, okay, I don't counter something correctly. You get punched in the face. So not in jiu jitsu, you don't get punched in jiu jitsu, but in mma, there are all these analogies between all these things that I think actually hold that are important life lessons, right? It's like, okay, you're surfing a wave. It's like sometimes you can't go in the other direction on it, right? It's like there are limits to kind of what. It's like foil. You can pump the foil and push pretty hard and a bunch of directions, but yeah, at some level the momentum against you is strong enough. That's not going to work. And I do think that that's sort of a humbling but also an important lesson for I think people who are running things or building things. It's like, yeah, a lot of the game is just Being able to kind of push and work through complicated things, but you also need to kind of have enough of an understanding of, like, which things you just can't push through and where. Where the finesse is more important.

Yeah.

What are your Jiu Jitsu life lessons?

Well, I think you did it. You made it sound so simple, and we're so eloquent that it's easy to miss. But basically, being okay and accepting the wisdom and the joy in the getting your ass kicked in the full range of what that means, I think that's a big gift of the being humbled. Somehow being humbled, especially physically, opens your mind to the full process of learning what it means to learn, which is being willing to suck at something. I think Jiu Jitsu is just very repetitively, efficiently humbles you over and over and over and over to where you can carry that lessons to places where you don't get humbled as much. Whether it's research or running a company or building stuff, the cycle is longer in Jiu Jitsu. You can just get humbled in this period of an hour, over and over and over and over. Especially when you're a beginner, you'll have a little person kick your ass repeatedly, definitively. There's no argument.

Oh, yeah.

And then you literally tap, because if you don't tap, you're going to die. So this is an agreement. You could have killed me just now, but we're friends, so we're going to agree that you're not going to. And that kind of humbling process, it just does something to your psyche, to your ego, that puts it in its proper context, to realize that, you know, everything in this life is like a journey from sucking through a hard process of improving or rigorously, day after day after day after day. Like, any kind of success requires hard work. Yeah. Jiu Jitsu, more than a lot of sports, I would say, because I've done a lot of them, really teaches you that. And you made it sound so simple, like, I'm okay. You know, it's. It's okay. It's part of the process. You just get humble, get yourself.

I've just. I've just failed and been embarrassed so many times in my life that, like, you know, I'm. I'm. It's a core competence at this point.

It's a core competence. Well, yes, and there's a deep truth to that. Being able to. And you said it in the very beginning, which is, that's the thing that stops us. Especially as you get older, especially as you develop expertise in certain areas, the not being willing to be a beginner in a new area. Yeah, because that's where the growth happens is being willing to be a beginner, being willing to be embarrassed saying something stupid, doing something stupid. A lot of us that get good at one thing, you want to show that off and it sucks being a beginner, but it's, it's where growth happens.

Yeah.

Well, speaking of which, let me ask you about AI. It seems like this year for the entirety of the human civilization is an interesting year for the development of artificial intelligence. A lot of interesting stuff is happening. So META is a big part of that. Meta has developed llama, which is a 65 billion parameter model. There's a lot of interesting questions they can ask here, one of which has to do with open source. But first, can you tell the story of developing of this model and making the complicated decision of how to release it?

Yeah, sure. I think you're right. First of all, that in the last year there have been a bunch of, bunch of advances on scaling up these large transformer models. So there's the language equivalent of it with large language models. There's sort of the image generation equivalent with these large diffusion models. There's a lot of fundamental research that's gone into this and META has taken the approach of being quite open and academic in our development of AI. Part of this is we want to have the best people in the world researching this and a lot of the best people want to know that they're going to be able to share their work. So that's part of the deal that we have, is that we can get, if you're one of the top AI researchers in the world, you can come here, you can get access to kind of industry scale infrastructure. And part of our ethos is that we want to share what's invented broadly. We do that with a lot of the different AI tools that we create. And LLAMA is the language model that our research team made and we did a limited open source release for it, which was intended for researchers to be able to use it. But responsibility and getting safety right on these is very important. So we didn't think that for the first one there were a bunch of questions around whether we should be releasing this commercially. So we kind of punted that for V1 of llama and just released it from Research. Now obviously by releasing it for research, it's out there, but companies know that they're not supposed to kind of put it into commercial releases. And we're working on the follow up models for this and thinking through how, how exactly this should work for follow on now that we've had time to work on a lot more of the safety and the pieces around that. But overall I just kind of think that it would be good if there were a lot of different folks who had the ability to build state of the art technology here and not just a small number of big companies to train one of these AI models. The state of the art models just takes hundreds of millions of dollars of infrastructure. So there are not that many organizations in the world that can do that at the biggest scale today. And now it gets more efficient every day. So. So I do think that that will be available to more folks over time, but I just think there's all this innovation out there that people can create and I just think that we'll also learn a lot by seeing what the whole community of students and hackers and startups and different folks build with this. And that's kind of been how we've approached this. And it's also how we've done a lot of our infrastructure and we took our whole data center design and our server design and we built this open compute project where we just made that public. And part of the theory was like, all right, if we make it so that more people can use the server design, then that'll enable more innovation. It'll also make the server design more efficient and that'll make our business more efficient too. So that's worked. And we've just done this with a lot of our infrastructure.

So for people who don't know, you did the limited release, I think in February of this year of Llama and it got quote unquote leaked, meaning like it escaped the limited release aspect. But it was, you know, that something you probably anticipated given that it's just released to research.

We shared it with researchers.

Right. So it's just trying to make sure that there's like a slow release.

Yeah.

But from there I just would love to get your comment on what happened next, which is like this is a very vibrant open source community that just builds stuff on top of it. There's Llama CPP basically stuff that makes it more efficient to run on smaller computers. There's combining with reinforcement, learning with human feedback. So some of the different interesting fine tuning mechanisms. There's then also like fine tuning and GPT three generations, there's a lot of GPT for all alpaca colossal AI. All these kinds of models just kind of spring up like run on top of Wood.

Yeah.

What do you think about that?

No, I think it's been really neat to see. I mean, there's been folks who are getting it to run on local devices, right? So if you're an individual who just wants to experiment with this at home, you probably don't have a large budget to get access to a large amount of cloud compute. So getting it to run on your local laptop is pretty good and pretty relevant. And then there are things like, yeah, Llama CPP re implemented it more efficiently. So you know, now even when we run our own versions of it, we can do it on way less compute and it just way more efficient. Save a lot of money for everyone who uses this. So that, that is, is, is good. I do think it's worth calling out that because this was a relatively early release, Llama isn't quite as on the frontier as for example, the biggest OpenAI models or the biggest Google models. You mentioned that the largest Llama model that we released had 65 billion parameters. And no one knows, I guess outside of OpenAI exactly what the specs are for GPT4. But I think the, you know, my understanding is it's like 10 times bigger. And I think Google's Palm model also, I think has about 10 times as many parameters. Now the Llama models are very efficient, so they perform well for something that's around 65 billion parameters. So for me that was also part of this because there's this whole debate around is it good for everyone in the world to have access to the most frontier AI models. And I think as the AI models start approaching something that's like a superhuman intelligence, that's a bigger question that we'll have to grapple with. But right now, I mean, these are still very basic tools. They're powerful in the sense that a lot of open source software like databases or web servers can enable a lot of pretty important things. But I don't think anyone looks at the current generation of LLAMA and thinks it's anywhere near a super intelligent. So I think that a bunch of those questions around is it good to kind of get out there? I think at this stage surely you want more researchers working on it for all the reasons that open source software has a lot of advantages and we talked about efficiency before, but another one is just open source software tends to be more secure because you have more people looking at it openly and scrutinizing it and finding holes in it, and that makes it more safe. So I think at this point it's more, I think it's generally agreed upon that open source software is generally more secure and safer than things that are kind of developed in a silo where people try to get security through obscurity. So I've met, for the scale of what we're seeing now with AI, I think we're more likely to get to good alignment and good understanding of kind of what needs to do to make this work well by having it be open source. And that's something that I think is quite good to have out there and happening publicly at this point.

Meta released a lot of models as open source, so the massively multilingual speech model, the model that's. I mean, I'll ask you questions about those, but the point is you've open sourced quite a lot. You've been spearheading the open source movement where's that's really positive, inspiring to see from one angle, from the research angle, of course, there's folks who are really terrified about the existential threat of artificial intelligence. And those folks will say that, you know, you have to be careful about the open sourcing step. But where do you see the future of open source here as part of Meta? The tension here is do you want to release the magic sauce? That's one tension. And the other one is do you want to put a powerful tool in the hands of bad actors, even though it probably has a huge amount of positive impact also?

Yeah. I mean, again, I think for the stage that we're at in the development of AI, I don't think anyone looks at the current state of things and thinks that this is super intelligence. And the models that we're talking about, the Llama models here, are generally an order of magnitude smaller than what OpenAI or Google are doing. So I think that at least for the stage that we're at now, the equities balance strongly, in my view, towards doing this more openly. I think if you got something that was closer to superintelligence, then I think you'd have to discuss that more and think through that a lot more. And we haven't made a decision yet as to what we would do if we were in that position, but I think there's a good chance that we're pretty far off from that position. So. I'm certainly not saying that the position that we're taking on this now applies to every single thing that we would ever do. And certainly inside the company, we probably do more open source work than most of the other big tech companies, but we also don't open source everything. A lot of the core kind of app code for WhatsApp or Instagram or something. We're not open sourcing that. It's not a general enough piece of software that would be useful for a lot of people to do different things. You know, whereas the software that we do, whether it's like an open source server design or, or basically, you know, things like memcache. Right. Like a good, you know, it was probably our earliest project that I worked on. It was probably one of the last things that I, that I coded and led directly for the company. But basically this like caching tool for quick data retrieval, these are things that are just broadly useful across anything that you want to build. And I think that some of the language models now have that feel as well as some of the other things that we're building, like the translation tool that you just referenced.

So text to speech and speech to text, you've expanded it from around 100 languages to more than 1100 languages and you can identify more than the model can identify more than 4,000 spoken languages, which is 40 times more than any known previous technology. To me, that's really, really, really exciting in terms of connecting the world, breaking down barriers that language creates.

Yeah, I think being able to translate between all of these different pieces in real time, this has been a kind of common sci fi idea that we'd all have, whether it's, I don't know, an earbud or glasses or something that can help translate in real time between all these different languages. And that's one that I think technology is basically delivering now. So yeah, I think that's pretty, pretty exciting.

You mentioned the next version of Llama. What can you say about the next version of Llama? What can you say about what were you working on in terms of release, in terms of the vision for that?

Well, a lot of what we're doing is taking the first version, which was primarily this research version, and trying to now build a version that has all of the latest state of the art safety precautions built in. And we're using some more data to train it from across our services. But a lot of the work that we're doing internally is really just focused on making sure that this is, you know, as aligned and responsible as possible. And you know, we're building a lot of our own, you know, we're talking about kind of the open source infrastructure, but you know, the, the main thing that we focus on building here, you know, a lot of product experiences to help people connect and express themselves. So, you know, we're going to I've talked about a bunch of this stuff, but. And you'll have an assistant that you can talk to in WhatsApp. You know, I think in the future every creator will have kind of an AI agent that can kind of act on their behalf that their fans can talk to. I want to get to the point where every small business basically has an AI agent that people can talk to to do commerce and customer support and things like that. So there are going to be all these different things. And llama, or the language model underlying this is basically going to be the engine that powers that. The reason to open source it is that, as we did with the first version, is that it basically unlocks a lot of innovation in the ecosystem, will make our products better as well, and also gives us a lot of valuable feedback on security and safety, which is important for making this good. But yeah, I mean, the work that we're doing to advance the infrastructure, it's. It's basically at this point taking it beyond a research project into something which is ready to be kind of core infrastructure, not only for our own products, but, you know, hopefully for, For a lot of other things out there too.

Do you think the LLAMA or the language model underlying that version two will be open sourced? Do you have internal debate around that? The pros and cons and so on?

This is, I mean, we were talking about the debates that we have internally and I think, I think the question is how to do it. Right. I mean, it's, I think we, you know, we did the research license for V1, and I think the big thing that we're thinking about is basically like, what's the right way?

So there was a leak that happened. I don't know if you can comment on it. For V1, we released it as a.

Research project for researchers to be able to use, but in doing so, we put it out there. So we were very clear that anyone who uses the code and the weights doesn't have a commercial license to put into products. And we've generally seen people respect that. Right. It's like you don't have any reputable companies that are basically trying to put this into their commercial products. But, but, yeah, but by sharing it with, you know, so many researchers, it's, it's, you know, yeah, it did leave the building.

But what have you learned from that process that you might be able to apply to V2 about how to release it safely, effectively, if you release it?

Yeah, well, I mean, I think a lot of the feedback, like I said, is just around you know, different things around, you know, how do you fine tune models to make them more aligned and safer? And you see all, all the different data recipes that you mentioned. A lot of different projects that are based on this. There's one at Berkeley, it was just all over and people have tried a lot of different things and we've tried a bunch of stuff internally. So we're making progress here, but also we're able to learn from some of the best ideas in the community. And I think it's know we want to just continue, continue pushing that forward. But I don't have any news to announce on, on this if that's, if that's what you're, you're asking. I mean this is a, a thing that we're, we're still, we're still kind of, you know, actively working through the, the, the right way to move forward here.

The details of the secret sauce are still being developed. I see. Can you comment on what do you think of the thing that worked for GPT, which is the reinforcement learning with human feedback. So doing this alignment process, do you find it interesting? And as part of that, let me ask, because I talked to Yann Lecun before talking to you today. He asked me to ask or suggested that I ask, do you think LLM fine tuning will need to be crowdsourced? Wikipedia style? So crowdsourcing. So this kind of idea of how to integrate the human in the fine tuning of these foundation models.

Yeah, I think that's a really interesting idea that I've talked to Jan about a bunch and we were talking about how do you basically train these models to be as safe and aligned and responsible as possible and different groups out there who are doing development test different data recipes in fine tuning. But this idea that you just mentioned is that at the end of the day, instead of having kind of one group fine tune some stuff and another group produce a different fine tuning recipe and then us trying to figure out which one we think works best to produce the most aligned model. I do think that it would be nice if you could get to a point where you had a Wikipedia style collaborative way for a kind of a broader community to fine tune it as well. Now there's a lot of challenges in that, both from an infrastructure and a community management and product perspective about how you do that. So I haven't worked that out yet, but as an idea I think it's quite compelling and I think it goes well with the ethos of open sourcing. The technology is also finding a way to have A kind of community driven. A community driven training of it. But I think there are a lot of questions on this in general, these questions around what's the best way to produce aligned AI models. It's very much a research area and it's one that I think we will need to make as much progress on as the kind of core intelligence capability of the models themselves.

Well, I just did a conversation with Jimmy Wales, the founder of Wikipedia, and to me, Wikipedia is one of the greatest websites ever created and it's kind of a miracle that it works. And I think it has to do with something that you mentioned, which is community. You have a small community of editors that somehow work together well and they handle very controversial topics and they handle it with balance and with grace, despite sort of the attacks that will often.

Happen a lot of the time. I mean, it has issues just like any other human system. But yes, I mean, the balance is. I mean, it's amazing what they've been able to achieve, but it's also not perfect. And I think that that's. There's still a lot of challenges.

Right. It's the more controversial the topic, the more, the more difficult the, the journey towards quote, unquote, truth or knowledge or wisdom that Wikipedia tries to capture. In the same way AI models, we need to be able to generate those same things, truth, knowledge and wisdom. And how do you align those models that they generate something that is closest to truth? There's these concerns about misinformation, all this kind of stuff that nobody can define. And it's something that we together as a human species have to define, like what is truth and how to help AI systems generate that. One of the things language models do really well is generate convincing sounding things. They can be completely wrong. And so how do you align it to be less wrong? And part of that is the training and part of that is the alignment. And however you do the alignment stage, and just like you said, it's a very new and a very open research problem.

Yeah, And I think that there's also a lot of questions about whether the current architecture for LLMs, as you continue scaling it, what happens? A lot of what's been exciting in the last year is that there's clearly a qualitative breakthrough where with some of the GPT models that OpenAI put out and that others have been able to do as well, I think it reached a kind of level of quality where people are like, wow, this is, this feels different. And like it's going to be able to be the foundation for Building a lot of awesome products and experiences and value. But I think that the other realization that people have is, wow, we just made a breakthrough. If there are other breakthroughs quickly, then I think that there's the sense that maybe we're closer to general intelligence. But I think that idea is predicated on the idea that I think people believe that there's still generally a bunch of additional breakthroughs to make and that we just don't know how long it's going to take to get there. And one view that some people have, this doesn't tend to be my view as much is that simply scaling the current LLMs and getting to higher parameter count models by itself will get to something that is closer to general intelligence. But, um, I don't know, I, I tend to think that there's probably more, more fundamental steps that need to be taken along the way there.

But still, the leaps taken with this extra alignment step is quite incredible, quite surprising to, to a lot of folks. And on top of that, when you start to have hundreds of millions of people potentially using a product that integrates that, you can start to see civilization transforming effects before you achieve super, quote, unquote, super intelligence. It could be super transformative without being a super intelligence.

Oh yeah. I mean, I think that there are going to be a lot of amazing products and value that can be created with the current level of technology to some degree. I'm excited to work on a lot of those products over the next few years and I think it would just create a tremendous amount of whiplash if the number of breakthroughs keeps, like, if they keep on being stacked breakthroughs. Because I think to some degree, industry and the world needs some time to kind of build these breakthroughs into the products and experiences that we all use so we can actually benefit from them. But I don't know, I think that there's just a, A, like an awesome amount of stuff to do. I mean, I think about like all of the, I don't know, small businesses or individual entrepreneurs out there who now we're going to be able to get help coding the things that they need to go build things or designing the things that they need or we'll be able to, you know, use these models to be able to do customer support for the people that they're, that they're serving over WhatsApp without having to, you know, I think that that's, that's just going to be, I just think that this is all going to be super exciting. It's going to create better, better experiences for people and just unlock a ton of innovation and value.

So I don't know if you know, but you know, what is it? Over 3 billion people use WhatsApp, Facebook and Instagram. So any kind of AI fueled products that go into that, like we're talking about anything with LLMs, will have a tremendous amount of impact. Do you have ideas and thoughts about possible products that might start being integrated into into these platforms used by so many people?

Yeah, I think there's three main categories of things that we're working on. The first that I think is probably the most interesting is there's this notion of you're going to have an assistant or an agent who you can talk to. And I think probably the biggest thing that's different about my view of how this plays out from what I see with OpenAI and Google and others, is everyone else is building the one singular AI. It's like, okay, you talk to ChatGPT or you talk to Bard, or you talk to Bing. And my view is that, that there are going to be a lot of different AIs that people are going to want to engage with. Just like you want to use, you know, a number of different apps for different things and you have relationships with different people in your life who fill different emotional roles for you and I. So I think that they're going to be people have a reason that I think you don't just want a singular AI. And that I think is probably the biggest distinction in terms of how I think about this and a bunch of these things. I think you'll want an assistant. I mentioned a couple of these before. I think every creator who you interact with will ultimately want some kind of AI that can proxy them and be something that their fans can interact with or that allows them to interact with their fans. And this is like the common creator promise. Everyone's trying to build a community and engage with people and they want tools to be able to amplify themselves more and be able to do that. But you only have 24 hours in a day. So I think having the ability to basically bottle up your personality or give your fans information about when you're performing a concert or something like that, that I think is going to be something that's super valuable. But it's not just that. Again, it's not this idea that I think people are going to want just one singular AI. I think you're going to want to interact with a lot of different entities. And then I think there's the business version of this too, which we've touched on a couple of times, which is I think every business in the world is going to want basically an AI that it's like you have your page on Instagram or Facebook or WhatsApp or whatever and you want to point people to an AI that people can interact with. But you want to know that that AI is only going to sell your products. You don't want it recommending your competitors stuff. So it's not like there can be just one singular AI that can answer all the questions for a person because that AI might not actually be aligned with you as a business to really just do the best job providing support for your product. So I think that there's going to be a clear need in the market and in people's lives for there to be a bunch of these.

Part of that is figuring out the research, the technology that enables the personalization that you're talking about. So not one centralized, godlike LLM, but one, just a huge diversity of them that's fine tuned to particular needs, particular styles, particular businesses, particular brands, all that kind of stuff.

And also enabling, just enabling people to create them really easily for the, you know, for, to, for your own business or if you're a creator, to be able to help you engage with your fans. And I think that's so yeah, I think that there's a clear kind of interesting product direction here that I think is fairly unique from what any of the other big companies are taking. It also aligns well with this sort of open source approach because again we sort of believe in this more community oriented, more democratic approach to building out the products and technology around this. We don't think that there's going to be the one true thing. We think that there should be kind of a lot of development. So that part of things I think is going to be really interesting and we could, we could go. Price spent a lot of time talking about that and the kind of implications of, of that approach being different from what others are taking. But there's a bunch of other simpler things that I think we're also going to do. Just going back to your question around how this finds its way into like what do we build? There are going to be a lot of simpler things around, okay, you, you post photos on Instagram and Facebook and you know, and WhatsApp and messenger. And like you want the photos to look as good as possible. So like having an AI that you can just like take a photo and then just tell it like, okay, I want to edit this thing or describe this. It's like, I think we're going to have tools that are just way better than what we've historically had on this. And that's more on the image and media generation side than the large language model side. But it all kind of plays off of advances in the same space. There are a lot of tools that I think are just going to get built into every one of our products. I think every single thing that we do is going to basically get evolved in this direction. It's like, in the future, if you're advertising on our services, do you need to make your own kind of ad creative? No. You'll just tell us, okay, I'm a dog walker and I am willing to walk people's dogs and help me find the right people and create the ad unit that will perform the best and give an objective to the system. And it just kind of like connects you with the right people.

Well, that's a super powerful idea of generating the language. Almost like rigorous a B testing for you that works to find the best customer for your thing. I mean, to me, advertisement, when done well, just finds a good match between a human being and a thing that will make that human being happy.

Yeah, totally.

And do that as efficiently as possible.

When it's done well, people actually like it. I think that there's a lot of examples where it's not done well and it's annoying, and I think that that's what kind of gives it a bad rap. But yeah. And a lot of this stuff is possible today. I mean, obviously a B testing stuff is built into a lot of these frameworks. The thing that's new is having technology that can generate the ideas for you about what to a B test. So I think that's exciting. So this will just be across like everything that we're doing, right? All the metaverse stuff that we're doing. It's like you want to create worlds in the future. You'll just describe them and it'll create the code for you.

So natural language becomes the interface we use for all the ways we interact with the computer, with the digital, more of them.

Yeah, yeah, totally.

Yeah. Which is what everyone can do using natural language. And with translation, you can do it in any kind of language. Um, I. I mean, for the personalization is really, really, really interesting.

Yeah.

It unlocks so many possible things. I mean, I for one, look forward to creating a copy of myself.

I know we talked about this last.

Time, but this has. Since the last time, this becomes now we're closer, much closer. Like, I could Literally just having interacted with some of these language models, I can see the absurd situation where I'll have a large or a Lex language model and I'll have to have a conversation with him about like, hey listen, like you're just getting out of line and having a conversation where you fine tune that thing to be a little bit more respectful or something like this. I mean that's, that's going to be the. That seems like an amazing product for businesses, for humans. Just not, not just the assistant that's facing the individual, but the assistant that represents the individual to the public. Both, both directions. There's basically a layer that is the AI system through which you interact with the outside world, with the outside world that has humans in it. That's really interesting. And you that have social networks that connect billions of people. It seems like a heck of a large scale place to test some of this stuff out.

Yeah. I mean I think part of the reason why creators will want to do this is because they already have the communities on our services. Yeah. And a lot of the interface for this stuff today are chat type interfaces and, and between WhatsApp and Messenger. I think that those are just great ways to interact with people.

So some of this is philosophy, but do you see a near term future where you have some of the people you're friends with are AI systems on these social networks, on Facebook, on Instagram, even on WhatsApp having conversations where some heterogeneous, some as humans, some as AI?

I think we'll get to that if only just empirically looking at. Microsoft released this thing called Xiaoice several years ago in China and it was a pre LLM chatbot technology that was a lot simpler than what's possible today. And I think it was like tens of millions of people were using this and just really became quite attached and built relationships with it. And I think that there are services today like Replika, where people are doing things like that. So I think that there's certainly needs for companionship that people have older people, I think most people probably don't have as many friends as they would like to have. Right. If you look at, there's some interesting demographic studies around that like the average person has. The number of close friends that they have is fewer today than it was 15 years ago. And I mean that gets to like this is like the core thing that I think about in terms of building services that help connect people. So I think you'll get tools that help people connect with each other are going to be the primary thing that we want to do. So you can imagine AI assistants that just do a better job of reminding you when it's your friend's birthday and how you could celebrate them. It's like right now we have the little box in the corner of the website that tells you whose birthday it is and stuff like that. But at some level, you don't just want to send everyone a note that says the same note saying Happy Birthday with an emoji. So having something that's more of a social assistant in that sense and that can update you on what's going on in their life and how you can reach out to them effectively help you be a better friend. I think that that's something that's super powerful too. But yeah, beyond that, there are all these different flavors of kind of personal AIs that I think could exist. So I think an assistant is sort of the kind of simplest one to wrap your head around. But I think a mentor or a life coach, someone who can give you advice, who's maybe like a bit of a cheerleader who can help pick you up through all the challenges that inevitably we all go through on a daily basis. And that there's probably know some role for something like that. And then, you know, all the way you can, you can probably just go through a lot of the different type of kind of functional relationships that people have in their life. And you know, I would bet that there will be companies out there that take a crack at a lot of these things. So I don't know, I think it's part of the interesting innovation that's going to exist is that there's certainly a lot like education tutors, right? It's like, I mean, I just look at, you know, my kids learning to code and you know, they love it. But you know, it's like they, they get stuck on a question and they have to wait till like I can help answer it, right? Or someone else who they know can help answer the question. In the future they'll just, there will be like a coding assistant that they have that is like designed to, you know, be perfect for teaching a five and a seven year old how to code. And they'll just be able to ask questions all the time and you know, be extremely patient. It's never going to get annoyed at them. Right. I think that like there are all these different kind of relationships or functional relationships that we have in our lives that, that are really interesting. And I think one of the big questions is like, okay, is this all going to just get bucketed into, you know, one singular AI. I just, I just don't, I don't think so.

Do you think? Well, this is actually a question from Reddit with the long term effects of human communication. When people can talk with, in quotes, talk with others through a chatbot that augments their language automatically. Rather than developing social skills by making mistakes and learning, will people just communicate by grunts in a generation? I. Do you think about long term effects at scale, the integration of AI in our social interaction?

Yeah, I mean, I think it's mostly good. That question was sort of framed in a negative way, but I mean, we were talking before about language models helping you communicate with. It was like language translation helping you communicate with people who don't speak your language. At some level, what all this social technology is doing is helping people express themselves better to people in situations where they would otherwise have a hard time doing that. So part of it might be okay because you speak a language that I don't know, that's a pretty basic one that I don't think people are going to look at that and say, it's sad that we have the capacity to do that because I should have just learned your language. That's pretty high bar. But overall, I'd say there are all these impediments and language is an imperfect way for people to express thoughts and ideas. It's one of the best that we have. We have that. We have art, we have code.

But language is also a mapping of the way you think, the way you see the world, who you are. One of the applications I've recently talked to a person who's actually a Jiu Jitsu instructor. He said that when he emails parents about their son and daughter that they can improve their discipline in class and so on, he often finds that he comes off a bit of more of an asshole than he would like. So he uses GPT to translate his original email into a nicer email, a more polite one.

We hear this all the time. A lot of creators on our services tell us that one of the most stressful things is basically negotiating deals with brands and stuff, like the business side of it, because they're like, I mean, they do their thing right? And the creators, they're excellent at what they do and they just want to connect with their community. But then they get really stressed, they go into their DMs and they see some brand wants to do something with them and they don't quite know how to negotiate or how to push back respectfully. So I think building a tool that can actually allow them to do that. Well, is one simple thing that I think is just an interesting thing that we've heard from a bunch of people that they'd be interested in. But going back to the broader idea, I don't know. I mean, Priscilla and I just had our third daughter.

Congratulations.

And it's like one of the saddest things in the world is like seeing your baby cry. Right? But like, it's like, what? Why is that? Right? It's like, well, because babies don't generally have much capacity to tell you what they care about otherwise. Right. It's not actually just babies. Right. It's you know, my five year old daughter cries too because she sometimes has a hard time expressing, you know, what, what matters to her. And, and I was thinking about that and I was like, well, you know, actually a lot of adults get very frustrated too because they can't, they have a hard time expressing things in a way that going back to some of the early themes that maybe is something that, you know, is a mistake or maybe they have pride or something like all these things get in the way. So I don't know. I think that all these different technologies that can help us navigate the social complexity and actually be able to better express our, what we're feeling and thinking. I think that's generally all good. And there are always these concerns like, okay, are people going to have worse memories because you have Google to look things up And I think in general, a generation later, you don't look back and lament that. I think it's just like, wow, we have so much more capacity to do so much more now. And I think that that'll be the case here too.

You can allocate those cognitive capabilities to deeper, more nuanced thought. Yeah, but it's change. So with, with just like with Google Search, the, the addition of language models, large language models, you basically don't have to remember nearly as much. Just like with Stack overflow for programming now that these language models can generate code right there, I mean, I find that I write like maybe 80%, 90% of the code I write is now generated first and then edited. So you don't have to remember how to write specifics of different functions.

Oh, but that's great. And it's also, it's not just the specific coding. I mean, in the context of a large company like this, I think before an engineer can sit down to code, they first need to figure out all of the libraries and dependencies that, you know, tens of thousands of people have written before them. And, you know, one of the things that I'm excited about that we're working on is it's not just, you know, tools that help engineers code. It's tools that can help summarize the whole knowledge base and help people be able to navigate all the internal information. I think that that's in the experiments that I've done with this stuff. I mean, that's on the public stuff. You just, you know, ask, ask one of these models to build you a script that does anything and it basically already understands what the best libraries are to do that thing and pulls them in automatically. It's. I think that's super powerful. That was always the most annoying part of coding, was that you had to spend all this time actually figuring out what the resources were that you were supposed to import before you could actually start building the thing.

Yeah, I mean, there's, of course the flip side of that, I think for the most part is positive. But the flip side is if you outsource that thinking to an AI model, you might miss nuanced mistakes and bugs, you lose the skill to find those bugs. And those bugs might be the code looks very convincingly right, but it's actually wrong in a very subtle way. But that's the trade off that we, that we face as human civilization. When we build more and more powerful tools, when we stand on the shoulders of taller and taller giants, we could do more, but then we forget how to do all the stuff that they did. It's a, it's a weird trade off.

Yeah, I agree. I mean, I think it's, I think it is very valuable in your life to be able to do basic things too.

Do you worry about some of the concerns of bots being present on social networks more and more human? Like bots that are not necessarily trying to do a good thing, or they might be explicitly trying to do a bad thing. Like phishing scams. Yeah, like social engineering, all that kind of stuff. Which has always been a very difficult problem for social networks, but now it's becoming almost a more and more difficult problem.

Well, I think there's a few different parts of this. So one is there are all these harms that we need to basically fight against and prevent. And that's been a lot of our focus over the last five or seven years, is basically ramping up very sophisticated AI systems, not generative AI systems, more kind of classical AI systems to be able to categorize and classify and identify. Okay, this, this post looks like it's promoting terrorism this one is, you know, like exploiting children. This one is, looks like it might be trying to incite violence. This one's an intellectual property violation. So there's, there's like, it's like 18 different categories of, of violating kind of harmful content that we've had to build specific systems to be able to track. And I think it's certainly the case that advances in generative AI will test those. But at least so far it's been the case. And I'm optimistic that it will continue to be the case that we will be able to bring more computing power to bear to have even stronger AIs that can help defend against those things. So we've had to deal with some adversarial issues before, right? I mean, for some things like hate speech, it's like people aren't generally getting a lot more sophisticated, like the average person, let's say if someone's saying some kind of racist thing, they're not necessarily getting more sophisticated at being racist. Right. It's okay, so that the system can just find. But then there's other adversaries who actually are very sophisticated, like nation states doing things. And, you know, we find, whether it's Russia or, you know, just different countries that are basically standing up these networks of bots or inauthentic accounts is what we call them because they're not necessarily bots. Some of them could actually be real people who are kind of masquerading as other, as other people, but they're acting in a coordinated way. And some of that behavior has gotten very sophisticated and it's very adversarial. So they, you know, each iteration, every time we find something and stop them, they kind of evolve their behavior. They don't just pack up their bags and go home and say, okay, we're not going to try. You know, at some point they might decide doing it on meta services is not worth it. They'll go do it on someone else if it's easier to do it in another place. But, but we have a fair amount of experience dealing with even those kind of adversarial attacks where they just keep on getting better and better. And I do think that as long as we can keep on putting more compute power against it, and if we're kind of one of the leaders in developing some of these AI models, I'm quite optimistic that we're going to be able to keep on pushing against the kind of normal categories of harm that you talk about, fraud, scams, spam, IP violations, things like that.

What about Creating narratives and controversy. To me, it's kind of amazing how a small collection of. Yeah, what did you say? Inauthentic accounts. So it could be bots, but.

Yeah, I mean, we have sort of this funny name for it, but we call it coordinated inauthentic behavior.

Yeah. It's kind of incredible how a small collection of folks can create narratives, create stories, especially if they're viral. So if. Especially if they have an element that can catalyze the virality of the narrative.

Yeah. And I think there, the question is, you have to be very specific about what is bad about it. Right. Because I think a set of people coming together or organically bouncing ideas off of each other and a narrative comes out of that is not necessarily a bad thing by itself if it's kind of authentic and organic. That's a lot of what happens in how culture gets created and how art gets created and a lot of good stuff. So that's why we've kind of focused on this sense of coordinated inauthentic behavior. So it's like if you have a network of whether it's bots, some people masquerading as different accounts, but you have kind of someone pulling the strings behind it and trying to kind of act as if this is a more organic set of behavior, but really it's not. It's just like one coordinated thing. That seems problematic to me. Right. I mean, I don't think people should be able to have coordinated networks and not disclose it as such. But that, again, we've been able to deploy pretty sophisticated AI and counterterrorism groups and things like that to be able to identify a fair number of these coordinated and authentic networks of accounts and take them down. We continue to do that. And I think it's one thing that, if you told me 20 years ago, it's like, all right, starting this website to help people connect at a college and, you know, in the future, you're going to be, you know, part of your organization is going to be a counterterrorism organization with AI to find coordinated and authentic. I would have thought that was pretty wild. But. But, but it's. But no, I think that that's. That's part of where we are. But, but look, I think that these questions that you're pushing on now, this is actually where I'd guess most of the challenge around AI will be for the foreseeable future. I think that there's a lot of debate around things like, is this going to create existential risk to humanity? And I think that Those are very hard things to disprove one way or another. My own intuition is that the point at which we become close to superintelligent superintelligence is. It's just really unclear to me that the current technology is going to get there without another set of significant advances. But that doesn't mean that there's no danger. I think the danger is basically amplifying the kind of known set of harms that people or sets of accounts can do. And we just need to make sure that we really focus on basically doing that as well as possible. So that's definitely a big focus for me.

Well, you can basically use large language models as an assistant of how to cause harm on social networks. You can ask it a question. You know, Meta has very impressive coordinated inauthentic account fighting capabilities. How do I do the coordinating authentic account creation where Meta doesn't detect it? Like literally ask that question and ba. And basically there's this kind of part of it. I mean, that's what OpenAI showed that they're concerned with those questions. Perhaps you can comment on your approach to it, how to do a kind of moderation on the output of those models, that it can't be used to help you coordinate harm in all the full definition of what harm means.

Yeah, and that's a lot of the fine tuning and the alignment training that we do is basically, you know, when we ship AI is across our products, a lot of what we're trying to make sure is that you can't ask it to help you commit a crime. Right. So I think training it to kind of understand that. And it's not that, not like any of these systems are ever going to be 100% perfect, but just making it so that this isn't an easier way to go about doing something bad than the next best alternative. Right. I mean, people still have Google, right. You still have search engines. So the information is out there. And for these, what we see is for nation states or these actors that are trying to pull off these larger coordinated and authentic networks to influence different things. At some point when we just make it very difficult, they do just try to use other services instead. Right. It's just like if you can make it more expensive for them to do it on your service, then kind of people go elsewhere. And I think that's the bar. It's not like, okay, are you ever going to be perfect at finding every adversary who tries to attack you? I mean, you try to get as close to that as possible. But. But I think really Kind of economically we were just trying to do is make it so that it's just inefficient for them to go after that.

But there's also complicated questions of what is and isn't harm, what isn't, isn't misinformation. So this is one of the things that Wikipedia has also tried to face.

Yeah.

I remember asking GPT about whether the virus leaked from a lab or not. And the answer provided was a very nuanced one and a well cited one, almost, dare I say, well thought out, one balanced. I would hate for that nuance to be lost through the process of moderation. Wikipedia does a good job on that particular thing too. But from pressures from governments and institutions, you could see some of that nuance and depth of information, facts and wisdom be lost.

Absolutely.

And that's, that's a scary thing. Some of the magic, some of the edges, the rough edges might be lost to the process of moderation of AI systems. So how do you get that right?

I really agree with what you're pushing on. I mean the, the core, I think the core shape of the problem is, is that there are some harms that I think everyone agrees are bad. Right. So sexual exploitation of children. Right. Like you're not going to get many people who think that that type of thing should be allowed on any service. Right. And that's something that we face and try to push off the, you know, as much as possible today. Terrorism, inciting violence. Right. Like we went through a bunch of these types of harms before. But then I do think that you get to a set of harms where there is more social debate around it. So misinformation I think has been a really tricky one because there are things that are kind of obviously false that are maybe factual but may not be harmful sense like, all right, are you going to censor someone for just being wrong if there's no kind of harm implication of what they're doing? I think that there's a bunch of real kind of issues and challenges there. But then there are other places where it is just take some of the stuff around Covid earlier on in the pandemic where there were real health implications, but there hadn't been time to fully vet a bunch of the scientific assumptions. And unfortunately I think a lot of the establishment on that kind of waffled on a bunch of facts and asked for a bunch of things to be censored that in retrospect ended up being more debatable or true. And that stuff is really tough and really Undermines trust in that. So I do think that the questions around how to manage that are very nuanced. The way that I try to think about it is that it goes. I think it's best to generally boil things down to the harms that people agree on. So when you think about is something misinformation or not, I think often the more salient bit is, is this going to potentially lead to physical harm for someone? And kind of think about it in that sense. And then beyond that, I think people just have different preferences on how they want things to be flagged for them. I think a bunch of people would prefer to kind of have a flag on something that says, hey, a fact checker thinks that this might be false. Or I think Twitter's Community Notes implementation is quite good on this, but again, it's the same type of thing. It's like just kind of discretionarily adding a flag because it makes the user experience better, but it's not trying to take down the information or not. I think that you want to reserve the kind of censorship of content to things that are of known categories that. That people generally agree are bad.

Yeah, but there's so many things, especially with the pandemic. But there's other topics where there's just deep disagreement fueled by politics about what is and isn't harmful. There's even just the degree to which the virus is harmful and the degree to which the vaccines that respond to the virus are harmful. There's just. There's a. Almost like a political divide around that. And so how do you make decisions about that? Where half the country in the United States or some large fraction of the world has very different views from another part of the world. Is there a way to stay out of the moderation of this?

I think we. It's very difficult to just abstain, but I think we should be clear about which of these things are actual safety concerns and which ones are a matter of preference in terms of how people want information flagged. So we did recently introduce something that allows people to have fact checking, not affect the distribution of what shows them their product. So, okay, a bunch of people don't trust who the fact checkers are. All right, well, you can. You can turn that off if you want, but if the. If the. If the content, you know, violates some policy, like it's inciting violence or something like that, it's still not going to be allowed. So I think that you want to honor people's preferences on. On that as much as possible. But look, I mean, this is really difficult stuff. I think the. It's really hard to know where to draw the line on what is fact and what is opinion. Because the nature of science is that Nothing is ever 100% known for certain. You can disprove certain things, but you're constantly testing new hypotheses and scrutinizing frameworks that have been long held. And every once in a while you throw out something that was working for a very long period of time, and it's very difficult. But I think that just because it's very hard and just because they're edge cases doesn't mean that you, you should not try to give people what they're looking for as well.

Let me ask about something you've faced in terms of moderation is pressure from different sources, pressure from governments. I want to ask a question, how to withstand that pressure for a world where AI moderation starts becoming a thing too. So what's Meta's approach to resist the pressure from governments and other interest groups in terms of what to moderate and not?

I don't know that there's a one size fits all answer to that. I think we basically have the principles around. We want to allow people to express as much as possible, but we have developed clear categories of things that we think are wrong that we don't want on our services, and we build tools to try to moderate those. So then the question is, okay, what do you do when a government says that they don't want something on the service and we have a bunch of principles around how we deal with that? Because on the one hand, if there's a democratically elected government and people around the world just have different values and different places, then should we, as a California based company, tell them that something that they have decided is unacceptable, actually, that we need to be able to express that? I think that there's a certain amount of hubris in that. But then I think that there are other cases where, you know, it's like a little more autocratic and you know, you have the dictator leader who's just trying to crack down on dissent. And you know, the people in a country are really not aligned with that and it's not necessarily against their culture, but the person who's leading it is just trying to push in a certain direction. These are very complex questions, but I think so it's difficult to have a one size fits all approach to it. But in general we're pretty active in kind of advocating and pushing back on requests to take things down. But honestly, the thing that I think a request to censor things is one thing, and that's obviously bad, but where we draw a much harder line is on requests for access to information, right? Because, you know, if you can. If you get told that you can't say something, I mean, that's bad, right? I mean, that, that, you know, is. Is obviously it violates your sense and freedom of expression at some level. But. But a government getting access to data in a way that seems like it would be unlawful in our country exposes people to real physical harm. And that's something that in general, we take very seriously. And then, so there's. That flows through, like, all of our policies in a lot of ways, right? It's. By the time you're actually, like, litigating with a government or pushing back on them, that's pretty late in the funnel, I'd say a bunch of this stuff starts a lot higher up in the decision of where do we put data centers then? There are a lot of countries where, you know, we may have a lot of people using the service in a place. It might be, you know, good for the service in some ways, good for those people if we could reduce the latency by having a data center nearby them. But, you know, for whatever reason, we just feel like, hey, this government does not have a good track record on basically not trying to get access to people's data. And at the end of the day, I mean, if you put a data center in a country and the government wants to get access to people's data, then they do, at the end of the day, have the option of having people show up with guns and taking it by force. So I think that there's a lot of decisions that go into how you architect the systems years in advance of these actual confrontations that end up being really important.

So you put the protection of people's data as a very, very high priority.

But that, I think is a. There are more harms that I think can be associated with that. And I think that that ends up being a more critical thing to defend against governments than, you know, whereas if another government has a different view of what should be acceptable speech in their country, especially if it's a democratically elected government, and, you know, it's. Then I. I think that there's a certain amount of deference that you should have to that.

So it's. That's speaking more to the direct harm that's possible when you give governments access to data. But if we look at the United States to the more nuanced kind of pressure to censor, not even order to censor, but pressure to censor from political entities, which has kind of received quite a bit of attention in the the United States. Maybe one way to ask that question is if you've seen the Twitter files, what have you learned from the kind of pressure from US Government agencies that was seen in Twitter files? And what do you do with that kind of pressure?

You know, I mean, I've, I've seen it. It's really hard from the outside to know exactly what happened in each of these cases. You know, we've obviously been in a bunch of our own cases where, you know, where agencies or different folks will just say, hey, here's a threat that we're aware of. You should be aware of this too. It's not really pressure as much as it is just flagging something that our security systems should be on alert about. I, I get how some people could think of it as that, but at the end of the day, it's our, it's our call on how to, on, on how to handle that. But I mean, I, I just, you know, in terms of running these services won't have access to as much information about what people think that adversaries might be trying to do as possible.

Boy. So you don't feel like there would be consequences if you know, anybody, the CIA, the FBI, a political party, the Democrats or the Republicans of high powerful political figures write emails you don't feel pressure from.

I guess what I say is there's so much pressure from all sides that I'm not sure that any specific thing that someone says is really adding that much more to the mix. It's. There are obviously a lot of people who think that, that we should be censoring more content. There are a lot of people who think we should be censoring less content. There are, as you say, all kinds of different groups that are involved in these debates. Right? So there's the kind of elected officials and politicians themselves, there's the agencies, but I mean, but there's the media, there's activist groups, there's. This is not a US specific thing. There are groups all over the world and kind of all in every country that bring different values. So it's a very active debate and I understand it. Right? I mean, these kind of questions get to really some of the most important social debates that are being had. So it gets back to the question of truth because for a lot of these things, they haven't yet been hardened into a single truth and society's sort of trying to hash out what, you know, what we think. Right. On, on, on certain issues, maybe in a few hundred years everyone will look back and say, hey, no, it wasn't it obvious that it should have been this? But, you know, no, we're, we're kind of in the, in that meat grinder now and, you know, and working through that. So, so now these are, these are all, are all very complicated. And you know, some people raise concerns in good faith and just say, hey, this is something that I want to flag for you to think about. Certain people, I certainly think, come at things with somewhat of a more kind of punitive or vengeful view of I want you to do this thing and if you don't, then I'm going to try to make your life difficult in a lot of other ways. But I don't know. This is one of the most pressurized debates, I think, in society. So I just think that there are so many people and different forces that are trying to apply pressure from different sides that it's. I, I don't think you can make decisions based on trying to make people happy. I think you just have to do what you think is the right balance and accept that people are going to be upset no matter where you come out on that.

Yeah, I like that pressurized debate. So how's your view of the freedom of speech evolved over the years. And now with AI, where the freedom might apply to them, not just to the humans, but to the, the personalized agents as you've spoken about them.

So, yeah, I mean, I, I've probably gotten a somewhat more nuanced view just because I think that there are, you know, I, I come at this, I'm obviously very pro freedom of expression. Right. I don't think you build a service like this that gives people tools to express themselves unless you think that people expressing themselves at scale is a good thing. Right. So I didn't get into this to like, try to prevent people from, from expressing anything. I, like, want to give people tools so they can express as much as possible. And then I think it's become clear that there are certain categories of things that we've talked about that I think almost everyone accepts are bad and that no one wants and that are illegal. Even in countries like the U.S. where, you know, you have the, the First Amendment that's very protective of enabling speech. It's like you're still not allowed to, you know, do things that are going to immediately incite violence or, you know, violate people's intellectual property or things like that. So there are those, but then there's also a very active core of just active disagreements in society where some people may think that something is true or false. The other side might think it's the opposite or just unsettled. Right. And those are some of the most difficult to kind of handle like we've talked about. But one of the lessons that I feel like I've learned is that a lot of times when you can, the best way to handle this stuff more practically is not in terms of answering the question of should this be allowed, but just like what, what is the best way to deal with someone being a jerk? Is the person basically just having a like, repeat behavior of like causing a lot of, a lot of issues? So looking at it more at that.

Level and its effect on the broader communities health, the community health of. It's tricky though, because like, how do you know there could be people that have a very controversial viewpoint that turns out to have a positive long term effect on the health of the community because it challenges the community.

No, that's true. Absolutely. Yeah. No, I think you, and I think you want to be careful about that. I'm not sure I'm expressing this very, very clearly because I certainly agree with your point there. And my point isn't that we should not have people in our services that are being controversial. That's certainly not what I mean to.