Mark Zuckerberg — Llama 3, $10B models, Caesar Augustus, & 1 GW datacenters

Mark Zuckerberg on: - Llama 3 - open sourcing towards AGI - custom silicon, synthetic data, & energy constraints on scaling - Caesar Augustus, intelligence explosion, bioweapons, $10b models, & much more Enjoy! Watch on YouTube https://youtu.be/bc6uFV9CJGg. Listen on Apple Podcasts https://podcasts.apple.com/us/podcast/mark-zuckerberg-llama-3-open-sourcing-%2410b-models-caeser/id1516093381?i=1000652877239, Spotify https://open.spotify.com/episode/6Lbsk4HtQZfkJ4dZjh7E7k?si=GOqj7hUdSaWSgi7ULWXjMA, or any other podcast platform. Human edited transcript with helpful links here https://www.dwarkeshpatel.com/i/143705335/transcript. Timestamps (00:00:00) - Llama 3 (00:08:32) - Coding on path to AGI (00:25:24) - Energy bottlenecks (00:33:20) - Is AI the most important technology ever? (00:37:21) - Dangers of open source (00:53:57) - Caesar Augustus and metaverse (01:04:53) - Open sourcing the $10b model & custom silicon (01:15:19) - Zuck as CEO of Google+ Sponsors If you’re interested in advertising on the podcast, fill out this form https://airtable.com/appxGOvFLDLP5dlzv/pagFVrbHRohW6F2bZ/form. * This episode is brought to you by Stripe https://stripe.com/, financial infrastructure for the internet. Millions of companies from Anthropic to Amazon use Stripe to accept payments, automate financial processes and grow their revenue. Learn more at stripe.com http://stripe.com. * V7 Go https://www.v7labs.com/go?utm_campaign=Dwarkesh%20Podcast%20Newsletter&utm_source=Dwarkesh-Podcast&utm_medium=Newsletter&utm_term=Paid-Email is a tool to automate multimodal tasks using GenAI, reliably and at scale. Use code DWARKESH20 for 20% off on the pro plan. Learn more here https://www.v7labs.com/go?utm_campaign=Dwarkesh%20Podcast%20Newsletter&utm_source=Dwarkesh-Podcast&utm_medium=Newsletter&utm_term=Paid-Email. * CommandBar https://commandbar.com is an AI user assistant that any software product can embed to non-annoyingly assist, support, and unleash their users. Used by forward-thinking CX, product, growth, and marketing teams. Learn more at commandbar.com http://commandbar.com. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit www.dwarkesh.com https://www.dwarkesh.com?utm_medium=podcast&utm_campaign=CTA_1

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Dwarkesh Patel?

  Mark, welcome to the podcast.

• 0:01Speaker 01

  Hey, thanks for having me. Big fan of your podcast.

• 0:03Dwarkesh Patel?

  Oh, thank you. That's very nice of you to say. Okay, so let's start by talking about the releases that will go out when this interview goes out. Tell me about the models. Tell me about Meta AI. What's new? What's exciting about them?

• 0:15Speaker 01

  Yeah, sure. So, you know, I think the the main thing that most people in the world are going to see is the new version of Meta AI. Right? So it's, and you know, the most important thing about what we're doing is the upgrade to the model. We're rolling out Lama 3. We're doing it both as open source for the the dev community, and it is now going to be powering Meta AI. So, you know, there's a lot that I'm sure we'll go into around Lama 3, but I think the bottom line on this is that with Lama 3, we now think that Meta AI is the most intelligent AI assistant that people can use that's freely available. We're also integrating Google and Bing for real time knowledge. We're going to make it a lot more prominent across our apps. So, you know, basically, you know, at the top of WhatsApp and Instagram and Facebook and Messenger,

• 0:59Speaker 01

  you'll just be able to, you know, use the search box right there to ask it any question. And there's a bunch of new creation features that we that we added that I think are pretty cool, that I think people enjoy. And I think animations is a good one. You can basically just take any image and animate it. But I think one that that people are going to find pretty wild is it now generates high quality images so quickly. I don't know if you've gotten a chance to play with this, that it actually generates it as you're typing and updates it in real time. So you're like typing your query and it's and it's kind of like honing in on. And, you know, it's like, okay, here, you know, show me a picture of a cow. Okay, in a field with mountains in the background. It's just like everything's popular. Eating macadamia nuts, drinking beer and like it just.

• 1:44Speaker 01

  And it's just like it's updating the image in real time. It's pretty wild. I think people are going to enjoy that. So, yeah, so that I think is that's what most people are going to see in the world, right? We're rolling that out, you know, not everywhere, but we're starting in in a handful of countries, and we'll do more over the coming weeks and months. So that's that, I think, is going to be a pretty big deal. And I'm really excited to get that in people's hands. It's a big step forward for Meta AI. But I think, you know, if you want to get under the hood a bit, the Llama 3 stuff is obviously the most technically interesting. So, you know, we're basically, for the first version, we're training three versions. You know, an 8 billion and a 70 billion, which we're releasing today,

• 2:26Speaker 01

  and a 405 billion dense model, which is still training. So we're not releasing that today. But, you know, the 8 and 70, I mean, I'm pretty excited about how they turned out. I mean, it's, you know, they're, they're leading for their scale. You know, it's, I mean, we'll, we'll release a blog post with all the benchmarks so people can check it out themselves. And obviously, it's open source, so people get a chance to play with it. We have a roadmap of new releases coming that are going to bring multimodality, more multilinguality, bigger context windows to those as well. And then, you know, hopefully sometime later in the year, we'll, we'll get to roll out the 4 or 5, which I think is, is, you know,

• 3:11Speaker 01

  in training, it's still training. But for, for where it is right now in training, it is already at around 85 MMLU. And, and just, we, we expect that it's going to have leading benchmarks on a bunch of, on a bunch of the benchmarks. So I'm, I'm pretty excited about all of that. I mean, the 70 billion is, is, is great, too. I mean, we're releasing that today. It's around 82 MMLU and has leading scores on math and reasoning. So I mean, it's, I think just getting this in people's hands is going to be pretty wild.

• 3:43Dwarkesh Patel?

  Oh, interesting. Yeah, that's the first I'm hearing is a benchmark. That's super impressive.

• 3:45Speaker 01

  Yeah, the 8 billion is, the 8 billion is, is nearly as, as powerful as the biggest version of Llama 2 that we released. So it's like the smallest Llama 3 is basically as powerful as the, the biggest Llama 2.

• 3:59Dwarkesh Patel?

  Hmm. Okay, so before we dig into these models, I actually want to go back in time. 2022 is, I'm assuming, when you started acquiring these H100s. Or you can tell me when.

• 4:10Speaker 01

  Yeah.

• 4:11Dwarkesh Patel?

  Where you're like, stock price is getting hammered. People are like, what's happening with all this CapEx? People aren't buying the metaverse. And presumably, you're spending that CapEx to get these H100s. How, back then, how did you know to get the H100s? How did you know we'll need the GPUs?

• 4:24Speaker 01

  I think it was because we were working on reels. So, you know, we got into this situation where, um, you know, we always want to have enough capacity to build something that we can't quite see that we're on the horizon yet. And we got into this position with reels where we needed more GPUs to train the models, right? It was this big evolution for our services where instead of just ranking content from people who you follow or your friends and whatever pages you follow.

• 4:57Speaker 01

  We made this big push to basically start recommending what we call unconnected content. Basically content from people or pages that you're not following. So now kind of the corpus of kind of content candidates that we could potentially show you expanded from, you know, on the order of thousands to on the order of hundreds of millions. So completely different infrastructure. And we started working on doing that and we were constrained on, um, on basically the infrastructure that we had to catch up to what TikTok was doing as quickly as we would have wanted to. So I basically looked at that and I was like, hey, we have to make sure that we're never in this situation again.

• 5:38Speaker 01

  So let's order enough GPUs to do what we need to do on reels and ranking content and feed. But let's also, let's double that, right? Cause again, like our normal principle is there's going to be something on the horizon that we can't see yet. Did you know it would be AI? Well, we thought it would be, we thought it was going to be something that had to do with training large models, right? I mean, but at the time I thought it was probably going to be more something that I had to do with content. But I don't know. I mean, it's, it's almost just the pattern matching and running the company is there's always another thing, right? So I'm not even sure I had, at that time I was so deep and just, you know, trying to get, you know, the recommendations working for reels and other content. Cause I mean, that's just such a big unlock for Instagram and Facebook to now being able to show people

• 6:22Speaker 01

  content that's interesting to them that are from people that they're not even following. But, um,

• 6:28Speaker 01

  Yeah, that ended up being a very good decision in retrospect. Yeah, yeah. Okay, and it came from being behind. So then it wasn't like I was, you know, it wasn't like, oh, I was so far ahead. Actually, most of the times, I think, where we kind of make some decision that ends up seeming good is because we messed something up before and just didn't want to repeat the mistake.

• 6:47Dwarkesh Patel?

  This is a total detour, but I actually want to ask about this while we're on this. We'll get back to AI in a second. So you didn't sell for one billion, but presumably there's some amount you would have sold for, right? Did you write down in your head, like, I think the actual valuation of Facebook at the time is this, and they're not actually getting the valuation right? If they're already $5 trillion, of course, you would have sold. So like, how did you think about that choice?

• 7:09Speaker 01

  Yeah, I don't know. I mean, look, I think some of these things are just personal.

• 7:14Speaker 01

  I don't know at the time that I was sophisticated enough to do that analysis. But I had all these people around me who were making all these arguments for how like a billion dollars was, you know, it's like, here's the revenue that we need to make, and here's how big we need to be, and like, it's clearly so many years in the future. Like, it was, it was very far ahead of where we were at the time. And I don't know, I didn't, I didn't really have the financial sophistication to really even engage with that kind of debate. I just, I think I sort of deep down believed in what we were doing. And I did some analysis.

• 7:50Speaker 01

  I was like, okay, well, what would I go do if I wasn't doing this? It's like, well, I really like building things, and I like helping people communicate, and I like understanding what's going on with people and the dynamics between people. So I think if I sold this company, I'd just go build another company like this. And I kind of like the one I have. So, so I mean, you know, what's, why, right? But I don't know. I think a lot of the biggest bets that people make are often just based on conviction and values.

• 8:26Speaker 01

  It's actually usually very hard to do the analyses, trying to connect the dots forward.

• 8:30Dwarkesh Patel?

  Yeah. So you've had Facebook AI research for a long time. Now it's become seemingly central to your company. At what point did making AGI, or whatever, however you consider that mission, at what point is that like, this is a key priority of what Meta is doing?

• 8:49Speaker 01

  Yeah, I mean, it's been a big deal for a while. So we started FAIR about 10 years ago. And the idea was that along the way to general intelligence or AI, like full AI, whatever you want to call it, there are going to be all these different innovations and that's going to just improve everything that we do. So we didn't conceive it as a product. It was more kind of a research group. And over the last 10 years, it has created a lot of different things that have basically improved all of our products and advanced the field and allowed other people in the field to create things that have improved our products too.

• 9:28Speaker 01

  So I think that's been great. But there's obviously a big change in the last few years when ChatGPT comes out, the diffusion models around image creation come out and like, I mean, this is some pretty wild stuff, right? That I think is like pretty clearly going to affect how people interact with like every app that's out there. So I, at that point, we started a second group, the Gen AI group, with the goal of basically bringing that stuff into our products. So building leading foundation models that would, that would sort of power all these different products. And initially, when we started doing that, the theory at first was, hey,

• 10:12Speaker 01

  a lot of the stuff that we're doing is, is pretty social, right? So it's helping people interact with creators, helping people interact with businesses to, you know, so the businesses can sell things or do customer support or, you know, basic assistant functionality for, whether it's for our apps or the smart glasses or VR, like all these different things. So initially, it wasn't completely clear that you were going to need kind of full AGI to be able to support those use cases. But then through working on them, I think it's actually become clear that you do, right? They're in all these subtle ways. So for example, you know, for Lama 2, when we were working on it, we didn't prioritize coding.

• 10:53Speaker 01

  And the reason why we didn't prioritize coding is because people aren't going to ask Meta AI a lot of coding questions in WhatsApp. Now they will. Well, I don't know. I'm not sure that WhatsApp is like the UI that people are going to be doing a lot of coding questions. So we're like, all right, look, in terms of the things that, you know, or Facebook or Instagram or, you know, those, those different services. Maybe, maybe the website or meta.ai that we're launching, I think. But. But the thing that was sort of, I think, has been a somewhat surprising result over the last 18 months is that it turns out that coding is important for a lot of domains, not just coding, right? So even if people aren't asking coding questions to the models, training the models

• 11:33Speaker 01

  on coding helps them just be more rigorous and answer the question and kind of help reason across a lot of different types of domains. Okay, so that's one example where it's like, all right, so for Lama 3, we're like really focused on training it with a lot of coding because it's like, all right, that's going to make it better on all these things, even if people aren't answering, aren't asking primarily coding questions. Reasoning, I think, is another example. It's like, okay, yeah, maybe you want to chat with a creator or you're a business and you're trying to interact with a customer. That interaction is not just like, okay, the person sends you a message and you just reply, right? It's a, it's like a multi-step interaction where you're trying to think through, how do I accomplish the person's goals? And a lot of times when a customer comes, they don't necessarily know exactly what

• 12:17Speaker 01

  they're looking for or how to ask their questions. So it's not really the job of the AI to just respond to the question. It's like you need to kind of think about it more holistically. It's really, it becomes a reasoning problem, right? So if someone else solves reasoning or makes good advances on reasoning and we're sitting here with a basic chatbot, then like our product is lame compared to what other people are building. So it's like, so, okay, so at the end of the day, we've got, we, you know, we basically realized we've got to solve general intelligence and we just kind of upped the ante and the investment to make sure that we could do that.

• 12:48Dwarkesh Patel?

  So the version of Lama that, that's going to solve all these use cases for users, is that the version that will be powerful enough to like replace a programmer you might have in this building?

• 13:02Speaker 01

  I mean, I just think that all this stuff is going to be progressive over time.

• 13:05Dwarkesh Patel?

  But end case, Lama 10?

• 13:10Speaker 01

  I mean, I think that there's a lot baked into that question. I'm not sure that we're replacing people as much as giving people tools to do more stuff.

• 13:18Dwarkesh Patel?

  Is a programmer in this building 10x more productive after a lot of?

• 13:20Speaker 01

  I would hope more, but, but no, I mean, look, I'm not, I don't believe that there's like a single threshold of intelligence for, for humanity, because I mean, people have different skills. And at some point, I think that AI is going to be, is probably going to surpass people at most of those things, depending on how powerful the models are. But, but I think it's progressive and I don't think AGI is one thing. I think it's, you're basically adding different capabilities. So multimodality is, is kind of a key one that we're focused on now, initially with photos and images and text, but eventually with videos. And then because we're so focused on the metaverse, kind of 3D type stuff is important. One modality that I'm pretty focused on, that I haven't seen as many other people

• 14:04Speaker 01

  in the industry focus on this is sort of like emotional understanding. Like, I mean, so much of, of the human brain is just dedicated to understanding people and, and kind of like understanding your expressions and emotions. And I think that that's like its own whole modality, right? That, I mean, you could say, okay, maybe it's just video or image, but it's like clearly a very specialized version of those too. So there's all these different capabilities that I think you want to basically train the models to focus on, as well as getting a lot better at reasoning, getting a lot better at memory, which I think is, is kind of its own whole thing. It's, I mean, I don't think we're going to be, you know, primarily shoving context or, or kind of things into a query context window in the future to, to ask more complicated questions.

• 14:49Speaker 01

  I think that there will be kind of different stores of memory or different custom models that, that are maybe more personalized to people. But I don't know, I think that these are all just different capabilities. And then obviously making them big and small, we care about both because, you know, we want to, you know, if you're running something like Meta AI, then we have the ability to, that's pretty server-based, but we also want it running on smart glasses. And, you know, there's not a lot of space in smart glasses. So you want to have something that's very efficient for that.

• 15:16Dwarkesh Patel?

  What is the use case that if you're doing tens of billions of dollars worth of inference, or even eventually hundreds of billions of dollars worth of inference, using intelligence in an industrial scale, what is the use case? Is it simulations? Is it the AIs that will be in the metaverse? What will we be using the data centers for?

• 15:35Speaker 01

  I mean, our bet is that it's gonna, this is basically gonna change all of the products, right? So I think that there's gonna be a kind of meta AI general assistant product. And I think that that will shift from something that feels more like a chatbot, where it's like you just ask a question and it kind of formulates an answer, to things where you're increasingly giving it more complicated tasks and then it goes away and does them.

• 15:58Dwarkesh Patel?

  Hmm.

• 15:58Speaker 01

  So I think that that's gonna take a lot of inference. It's gonna take a lot of compute in other ways, too.

• 16:04Speaker 01

  Then I think that there's a big part of what we're gonna do that is like interacting with other agents for other people. So whether it's businesses or creators. I guess a big part of my theory on this is that there's not just gonna be like one singular AI that you interact with. Because I think, you know, every business is gonna like want an AI that represents their interests. They're not gonna like wanna primarily interact with you through an AI that is gonna sell their competitors' customers. So, uh, sorry, their competitors' products. Um, so, um,

• 16:38Speaker 01

  so yeah, so I think creators is gonna be a big one. I mean, there are about 200 million creators on our platforms. They all basically have the pattern where they want to engage their community, but they're limited by hours in the day and their community generally wants to engage them, but they don't have, they're limited by hours in the day. So if you could create something where an AI could basically, where that creator can basically own the AI and train it in the way that they want,

• 17:06Speaker 01

  and can engage their community, I think that that's gonna be super powerful, too. So, um, so I think that there's gonna be a ton of engagement across all these things.

• 17:15Speaker 01

  But these are just the consumer use cases. I mean, I think when you think about stuff like, I mean, you know, I run like our foundation, right, a Chan Zuckerberg Initiative with my wife. And, you know, we're doing a bunch of stuff on science and, um, and there's obviously a lot of AI work that where I, that I think is gonna advance science and healthcare and all these things, too. So I think that it's like, there's a, this is, I think, gonna end up affecting basically every area of the products and, and, and the, and the, uh, the economy.

• 17:41Dwarkesh Patel?

  The thing you mentioned about an AI that can just go out and do something for you that's multi-step. Is that a bigger model? Is that you'll make, like Lama 4 will still, there'll be a version that's still 70B, but will just be, you'll just train it on the right data and that will be super powerful? How do, like, what does the progression look like? Is it scaling? Is it just same size, but different banks like you were talking about?

• 18:05Speaker 01

  I don't know that we know the answer to that. So I think one thing that is seems to be a pattern is that you have the llama, sorry, the llama model, and then you build some kind of other application-specific code around it, right? So some of it is the fine tuning for the use case, but some of it is just like logic for, okay, how like how Meta AI should integrate, should work with tools like Google or Bing to bring in real-time knowledge. I mean, that's not part of the base llama model. That's like part of it. Okay, so for Lama 2, we had some of that, and it was a little more kind of hand engineered.

• 18:45Speaker 01

  And then part of our goal for Lama 3 was to bring more of that into the model itself. And but for Lama 3, as we start getting into more of these agent-like behaviors, I think some of that is going to be more hand engineered. And then I think our goal for Lama 4 will be to bring more of that into the model. So I think at each point, like at each step along the way, you kind of have a sense of what's going to be possible on the horizon. You start messing with it and hacking around it. And then I think that that helps you hone your intuition for what you want to try to train into the next version of the model itself. Interesting. Which makes it more general, because obviously anything that you're hand coding is, you know, you can unlock some use cases, but it's just inherently brittle and non-general. Hmm.

• 19:29Dwarkesh Patel?

  Hey, everybody. Real quick, I want to tell you about a tool that I wish more applications used. So obviously, you've noticed every single company is trying to add an AI chatbot to their website. But as a user, I usually find them really annoying because they give these long, generic, often useless answers. Command Bar is a user assistant that you can just embed into your website or application. And it feels like you're talking to a friendly human support agent who is browsing with you and for you. And it's much more personalized than a regular chatbot. It can actually look up users' history and respond differently based on that. It can use APIs to perform actions. It can even proactively nudge.

• 20:12Dwarkesh Patel?

  Users to explore new features. One thing that I think is really cool is that instead of just outputting text, command bar can kind of just say, here, let me show you and start browsing alongside the user. Anyways, there are a bunch of great products already. You can learn more about them at commandbar.com. Thanks to them for sponsoring this episode and now back to Mark. When you say into the model itself, you train it on the thing that you want in the model itself. But what do you mean by into the model itself?

• 20:42Speaker 01

  Well, I mean, I think like the example that I gave for Lama two, where, you know, it's we, we really, I mean, for Lama two, the tool use was very, very specific. Whereas Lama three has the ability to has much better tool use, right? So, so we don't have to like hand code all this stuff to have it use Google to, to go do a search. It just kind of can do that. So, and similarly for coding and, and kind of running code and just a bunch of stuff like that. And, um, but I think once you kind of get that capability, then you get a peek of, okay, well, what can we start doing next? Okay. Well, I don't necessarily want to wait until Lama four is around to start building those capabilities. So let's start hacking around it.

• 21:27Speaker 01

  And so you do, you do a bunch of hand coding and that makes the, um, the products better for the interim. But then that also helps show the way of what we want to try to build into the next version of the model.

• 21:37Dwarkesh Patel?

  What is the community fine tune of Lama three you're most excited by? Maybe not the one that will be most useful to you, but just you will just enjoy playing it with the most. They're like fine tuned on antiquity and you'll just be like talking to Virgil or something. What are you excited about?

• 21:50Speaker 01

  I don't know. Um, I mean, I think the nature of the stuff is it's like, you get surprised, right? So I think like any, any specific thing that I sort of thought would be valuable, we'd probably be building, right? So, um, but.

• 22:09Speaker 01

  I think you'll get distilled versions. I think you'll get kind of smaller versions. I mean, I mean, one thing that I think is eight billion, I don't think is quite small enough for for a bunch of use cases, right? I think like over time, I'd love to get, you know, a billion parameter model or a two billion parameter model, or even like a, I don't know, maybe like a 500 million parameter model and see what you can do with that. Because I mean, as they start getting, if with eight billion parameters were basically nearly as powerful as the largest Llama 2 model, then with a billion parameters, you should be able to do something that's interesting, right? And faster, good for classification or a lot of kind of like basic things that people do before kind of

• 22:51Speaker 01

  understanding the intent of a of a user query and feeding it to the most powerful model to to kind of hone what the what the prompt should be. So I don't know. I think that's one thing that maybe the community can help fill in. But I mean, we'll also, we're also thinking about getting around to distilling some of these ourselves. But right now, the GPUs are pre-pegged training the 405. So.

• 23:12Dwarkesh Patel?

  What? Okay, so you have all these GPUs. I think you said 350,000 by the end of the year.

• 23:18Speaker 01

  That's the whole fleet. I mean, I was, we built two, I think it's like 22, 24,000 clusters that are kind of the single clusters that we have for training the big models. I mean, obviously, across a lot of the stuff that we do, a lot of our stuff goes towards training, like Reels models and like Facebook news feed and Instagram feed. And then inference is a huge thing for us because we serve a ton of people, right? So our ratio of inference compute required to to training is probably much higher than most other companies that are doing this stuff, just because the sheer volume of the community that we're serving.

• 23:55Dwarkesh Patel?

  Yeah. Yeah, that was really interesting. And the material they shared with me before that you trained it on more data than is compute optimal just for training, because the inference is such a big deal for you guys and also for the community, that it makes sense to just have this thing and have trillions of tokens in there.

• 24:09Speaker 01

  Yeah, yeah. Although in one of the interesting things about it that we saw even with the 70 billion is we we thought it would get more saturated at. You know, it's like we trained it on around 15 trillion tokens.

• 24:21Dwarkesh Patel?

  Yeah.

• 24:21Speaker 01

  We, I guess our prediction going in was that it was gonna ask them to it more. But even by the end, it was still learning, right? It's like, we, we probably could have fed it more tokens and it would have gotten somewhat better. But I mean, at some point, you know, you're running a company, you need to do these meta reasoning questions of like, all right, how do I want to spend our GPUs on like, training this 70 billion model further? Do we want to kind of get on with it so we can start testing hypotheses for Lama 4? So we kind of needed to, to make, um, to make that call. And I think we got it, we, I think we got to a reasonable balance for, for this version of the 70 billion.

• 24:59Speaker 01

  There'll be others in the future where, you know, the 70 billion multimodal one that'll come, um, over the next period. But, um, but yeah, I mean, it's, that was, that was fascinating that you could just, that it's, the architectures at this point can just take so much data.

• 25:12Dwarkesh Patel?

  Yeah, that's really interesting. So what does this imply about future models? I, you mentioned that the Lama 3.8B is better than the Lama 2.70B.

• 25:20Speaker 01

  No, no, it's nearly as good.

• 25:21Dwarkesh Patel?

  Okay.

• 25:21Speaker 01

  I don't overstep.

• 25:22Dwarkesh Patel?

  But does that mean like the Lama?

• 25:23Speaker 01

  In the same order of magnitude.

• 25:24Dwarkesh Patel?

  Does, does that mean like the Lama 4.70B will be as good as the Lama 3.4 or 5B? Like, what is the future of this?

• 25:30Speaker 01

  This is one of the great questions, right? That I think no one knows, um, is, is basically, you know, it's, it's one of the trickiest things in the world to plan around is when you have an exponential curve, how long does it keep going for?

• 25:43Dwarkesh Patel?

  Yeah.

• 25:43Speaker 01

  And, um, I think it's likely enough that it will keep going, that it is worth investing the, um, tens or, you know, a hundred billion plus in building the infrastructure to, um, assume that if that kind of keeps going, you're going to get some really amazing things that are just going to make amazing products.

• 26:03Dwarkesh Patel?

  Hmm.

• 26:03Speaker 01

  But I don't think anyone in the industry can really tell you that it will continue scaling at that rate for sure, right? In general, you know, in history, you hit bottlenecks at certain points. And now there's so much energy on this that maybe those bottlenecks get knocked over pretty quickly, but, um, but I don't know. I think that's, that's an interesting question.

• 26:24Dwarkesh Patel?

  What does the world look like where there aren't these bottlenecks? You know, suppose like progress just continues, uh, at, at this pace, which seems like plausible.

• 26:33Dwarkesh Patel?

  Like zooming out, forgetting about like Lamas.

• 26:35Speaker 01

  There're gonna be different bottlenecks.

• 26:37Dwarkesh Patel?

  Right? So if not training, then like, oh yeah, go ahead.

• 26:40Speaker 01

  Well, I think at some point. You know, over the last few years, I think there's this issue of GPU production. Yeah. Right. So even companies that had the models, sorry, that had the money to pay for the GPUs,

• 26:55Speaker 01

  couldn't necessarily get as many as they wanted because there was, there were all these supply constraints. Yeah. Now I think that's sort of getting less. So now I think you're seeing a bunch of companies think about, wow, we should just like really invest a lot of money in building out these things. And I think that that will go for, for some period of time.

• 27:15Speaker 01

  I think there's a, there is a capital question of like, okay, at what point does it stop being worth it to put the capital in? But I actually think before we hit that, you're going to run into energy constraints, right? Because I just, I mean, I don't think anyone's built a gigawatt single training cluster yet, right? And, and then you run into these things that just end up being slower in the world. Like getting energy permitted is like a very heavily regulated government function, right? So you're going from, on the one hand, software, which is somewhat regulated. I'd argue that it is more regulated than I think a lot of people in the, in the, in the tech community feel.

• 27:57Speaker 01

  Although it's obviously different. If you're starting a small company, maybe you feel that less. If you're a big company, you know, we just interact with people, but different governments and regulators are, you know, we have kind of lots of rules that we need to kind of follow and make sure we do a good job with around the world.

• 28:13Speaker 01

  But I think that there's no doubt that like energy, and if you're talking about building large new power plants or large build outs and then building transmission lines that cross other private or public land, that is just a heavily regulated thing. So you're talking about many years of lead time. So if we wanted to stand up just some like massive facility to power that, I think that that is.

• 28:40Speaker 01

  That's a very long-term project, right? And so I don't know. I think that that's, I think people will do it. But I don't think that this is like something that can be quite as magical as just like, okay, you get a level of AI and you get a bunch of capital and you put it in. And then like all of a sudden, the models are just gonna kind of like, it just like, I think you do hit different bottlenecks along the way.

• 29:00Dwarkesh Patel?

  Yeah. Is there something, a project, maybe I really did, maybe not, that even in company like Meta doesn't have the resources for? Like if your R and D budget or capex budget was 10x what it is now, then you could pursue it. Like it's in the back of your mind, but Meta today, and maybe you could like, even you can't even issue a stock or bond for it. It's like just 10x bigger than your budget.

• 29:20Speaker 01

  Well, I think energy is one piece.

• 29:21Dwarkesh Patel?

  Yeah.

• 29:21Speaker 01

  Right? I think we would probably build out bigger clusters than we currently can if we could get the energy to do it. So I think that that's.

• 29:34Dwarkesh Patel?

  And that's fundamentally money bottlenecked in the limit. Like if you had a trillion dollars.

• 29:39Speaker 01

  I think it's time.

• 29:40Dwarkesh Patel?

  Yeah.

• 29:40Speaker 01

  Right? Well, if you look at it in terms of, but it depends on how far the exponential curves go, right? Like I think a number of companies are working on, you know, right now, I think, you know, like a lot of data centers are on the order of 50 megawatts or 100 megawatts or like a big one might be 150 megawatts. Okay. So you take a whole data center and you fill it up with just all the stuff that you need to do for training and you build the biggest cluster you can. I think you're, that's kind of, I think a bunch of companies are running at stuff like that. But then when you start getting into building a data center that's like 300 megawatts or 500 megawatts or a gigawatt, I just, I mean, just no one has built a single gigawatt data center yet. So I think it will happen, right?

• 30:24Speaker 01

  I mean, this is only a matter of time, but it's, it's not going to be like next year, right? It's, I think that some of these things will take, I don't know, some, some number of years to, to build out. And then the question is, okay, well, if you. I mean, just to, I guess, put this in perspective. I think a gigawatt. It's like around the size of like a meaningful nuclear power plant only going towards training a model.

• 30:51Dwarkesh Patel?

  Didn't, didn't Amazon do this? There's like, they have a 950 gig, a megawatt.

• 30:55Speaker 01

  Yeah, I'm not exactly sure what you did. You'd have to, what they did, you'd have to ask them.

• 31:00Dwarkesh Patel?

  But it doesn't have to be in the same place, right? If distributed training works, it can be.

• 31:03Speaker 01

  That I think is a big question.

• 31:04Dwarkesh Patel?

  Yeah.

• 31:04Speaker 01

  Right. Is, is basically how that's going to work. And I do think in the future, it seems quite possible that more of what we call training for these big models is actually more along the lines of inference generating synthetic data to then go feed into the model. So I don't know what that ratio is going to be, but I consider the generation of synthetic data to be more inference than training today. But obviously, if you're doing it in order to train a model, it's, it's part of the broader training process. So, um, I don't know. That's an, that's a, an open question is to, to kind of where, what the balance of that and how that plays out.

• 31:42Dwarkesh Patel?

  If that's the case, would that potentially also be the case with Lama 3? And maybe like Lama 4 onwards where you put this out and if somebody has a ton of compute, then using the models that you've put out, you can just keep making these things arbitrarily smarter. Like some Kuwait or UAE or some random country has a ton of compute. And they can just, actually just use Lama 4 to just make something much smarter.

• 32:06Speaker 01

  Um, I, I do think that there are going to be dynamics like that. But I also think that there is a fundamental limitation on, um, on kind of the network architecture, right? Or the, the kind of model architecture, right? So I think like a 70 billion model that kind of we trained with the Lama 3 architecture can get better, right? It can, it can keep going. Like, like I was saying, it's, you know, we felt like if we kept on feeding it more data or rotated the high value tokens through again, then, then, you know, it would, it would continue getting better. But.

• 32:46Speaker 01

  And we've seen a bunch of other people around the world, you know, different companies basically take the Lama 2 70 billion base, like take that model architecture and then build a new model.

• 32:57Speaker 01

  It's still the case that when you make a generational improvement to the kind of Lama 3 70 billion or the Lama 3 405, there's nothing open source anything like that today. Right? Like a, it's, it's not, I think that that's like, it's a big step function and what people are going to be able to build on top of that I don't think can go infinitely from there. I think it can, there can be some optimization in that until you get to the next step function.

• 33:21Dwarkesh Patel?

  Yeah. Okay. So let's zoom out a little bit from specific models and even the many years lead times you would need to get energy approvals and so on. Like big picture, these next couple of decades.

• 33:32Speaker 01

  Sure.

• 33:32Dwarkesh Patel?

  What's happening with AI? Is, does it feel like another technology like metaverse or social? Or does it feel like a fundamentally different thing in the course of human history?

• 33:45Speaker 01

  I think it's going to be pretty fundamental. I think it's going to be more like the creation of computing in the first place. Right? So you'll get all these new apps in the same way that when you got the web or you got mobile phones, you got like people basically rethought all these experiences and a lot of things that weren't possible before now became possible. So I think that will happen, but I think it's a much lower level innovation. It's, it's, it's going to be more like going from people didn't have computers to people have computers is my, my sense.

• 34:24Speaker 01

  But it's also, it's, it's a, I don't know. It's, it's very hard to reason about exactly how this goes. I tend to think that.

• 34:36Speaker 01

  You know, in like the cosmic scale, obviously it'll happen quickly over a, you know, a couple of decades or something. But I do think that there is some set of people who are afraid of like, you know, it really just kind of spins and goes from being like somewhat intelligent to extremely intelligent overnight. And I just think that there's all these physical constraints that make that so that that's unlikely to happen. I just don't, I don't really see that that playing out. So I think you'll have, I think we'll have time to kind of acclimate a bit, but it will really change the way that we work and give people all these creative tools to do different things that they, yeah, I think, I think it's gonna be, it's gonna really enable people to do the things that they want a lot more is my view.

• 35:20Dwarkesh Patel?

  Okay, so maybe not overnight, but is it your view that like on a cosmic scale, if you think like humans evolved and then like AI happened and then they like went out through the galaxy? Or maybe it takes many decades, maybe it takes a century, but like, is that like the grand scheme of what's happening right now in history?

• 35:39Speaker 01

  Sorry, in what sense?

• 35:40Dwarkesh Patel?

  I mean, in the sense that there are other technologies like computers and even like fire, but like the AI happening is as significant as like humans evolving in the first place?

• 35:48Speaker 01

  I think that's tricky. I think people like to, you know, the history of humanity, I think, has been people basically, you know, thinking that certain aspects of humanity are like really unique in different ways. And then coming to grips with the fact that that's not true, but humanity is actually still super special. Right? So it's, it's like, we thought that the earth was the center of the universe. And it's like, it's not, but like, it's like, humans are still pretty awesome, right? And pretty unique. I think that another bias that people tend to have is thinking that intelligence is somehow

• 36:36Speaker 01

  Kind of fundamentally connected to life. And it's not actually clear that it is, right? I think like, like, people think that, um, I mean, I don't know that we have a clear enough definition of consciousness or, um, or, or, or life to kind of fully interrogate this. But I think there's all this science fiction about, okay, you create intelligence and now it like starts taking on all these human like behaviors and, and things like that. But I actually think that the current incarnation of all this stuff at least kind of feels like it's going in a direction where intelligence can be pretty separated from consciousness and agency and things like that, that, um, I think just makes it a super valuable tool.

• 37:21Speaker 01

  So I don't know. I mean, obviously it's, it's, um, it's very difficult to predict what direction this stuff goes in over time, which is why I, I don't think anyone should be dogmatic about, you know, how they plan to develop it or what they plan to do. I think you want to kind of look at like each release, you know? It's like, we're obviously very pro open source.

• 37:38Dwarkesh Patel?

  Yeah.

• 37:38Speaker 01

  But I haven't committed that we're going to like, release every single thing that we do. But it's basically, we, I like, I'm just generally very inclined to thinking that open sourcing it is going to be good for the community and also good for us, right? Cause we'll, we'll benefit from, from the innovations.

• 37:54Speaker 01

  But if at some point, like there's some qualitative change in what the, the thing is capable of, and we feel like it's just not responsible to open source it, then we won't. But, um, so I don't know. It's, it's, it's all, it's all very difficult to predict.

• 38:07Dwarkesh Patel?

  Yeah. What is a kind of qualitative change? Like a specific thing? You're training Lama Phi, Lama Four, and you've seen this and like, you know what? I'm not sure about open sourcing it.

• 38:20Speaker 01

  I think that that, it's a little hard to answer that in the abstract because there are negative behaviors that any product can exhibit that as long as you can mitigate it, it's like, it's okay. Right? So, um, I mean, there's bad things about social media that we work to mitigate. Right? There's bad things about Lama Two that we spend a lot of time trying to make sure that it's not like, you know, helping people commit violent acts or things like that. Right? I mean, that doesn't mean that it's like a, a kind of autonomous or intelligent agent. It just means that it's learned a lot about the world and it can answer a set of questions that, um, we think it would be unhelpful for it to answer. So.

• 39:01Speaker 01

  I um, I don't know. I think the question isn't really what behaviors would it show? It's what things would we not be able to mitigate after it shows that? And um, and I don't know. I, I, I think that there's so many ways in which something can be good or bad, that it's hard to actually enumerate them all upfront. If you even look at like, what we've had to deal with in, in um, you know, social media and like the different types of harms, we've basically gotten to, it's like, there's like 18 or 19 categories of, of harmful things that, that people do. And we've basically built AI systems to try to go identify what those things are that people are doing and try to make sure that that doesn't happen on our network as much as possible. So um, yeah, I think you, you can, over time, I think you'll be able to

• 39:45Speaker 01

  break down um, this into more of a taxonomy too. And I, I think this is a thing that we spend time researching too, because we want to make sure that we understand that.

• 39:54Dwarkesh Patel?

  So one of the things I asked Mark is what industrial scale use of Llama would look like? You see this in previous technological revolutions where at first they're thinking in a very small scale way about what's enabled. And I think that's what chatbots might be for other Llama. And I think the large scale use case might look something like what V7 Go is. And by the way, it's made by V7 Labs who's sponsoring this episode. So it's like a spreadsheet. You put in raw information like documents, images, whatever, and they become rows and the columns are populated by an LLM of your choice. And in fact, I used it to prepare for Mark. So I fed in a bunch of blog posts and papers from Meta's AI research. And as you can see, if you're on YouTube, it summarizes and extracts exactly the information

• 40:39Dwarkesh Patel?

  I want as columns. And obviously mine is a small use case. But you can imagine, for example, a company like FedEx has to process half a million documents a day. Obviously a chatbot can't do that. A spreadsheet can, because this is just like a fire hose of intelligence in there, right? Anyways, you can learn more about them at v7labs.com/go or the link in the description. Back to Mark. Yeah, like it seems to me it would be a good idea. I would be disappointed in a future where AI systems aren't broadly deployed and everybody doesn't have access to them.

• 41:09Speaker 01

  Yeah.

• 41:10Dwarkesh Patel?

  At the same time, I want to better understand the mitigations. Because if the mitigation is the fine tuning, well, the whole thing about open weights is that you can then

• 41:21Dwarkesh Patel?

  remove the fine tuning, which is often superficial on top of these capabilities. Like if it's like talking on Slack with a biology researcher. And again, I think like models are very far from this. They're right now, they're like Google search. But it's like I can show them my Petri dish and they can next lane like here's why your small pox sample didn't grow. Here's what to change.

• 41:40Dwarkesh Patel?

  How do you mitigate that? Because somebody can just like fine tune that in there, right?

• 41:44Speaker 01

  Yeah, I mean, that's true. I think a lot of people will basically use the off the shelf model. And some people who have basically bad faith are going to try to strip out all the bad stuff. So I do think that that's an issue. The

• 42:01Speaker 01

  the flip side of this is that, and this is one of the reasons why I'm kind of philosophically so pro open source. Is I do think that a concentration of AI in the future has the potential to be as dangerous as kind of it being widespread. So I think a lot of people are, they think about the questions of, okay, well, if we can do this stuff, is it bad for it to be out wild? Like just in kind of widely available.

• 42:31Speaker 01

  I think another version of this is like, okay, well, it's probably also pretty bad for one institution to have an AI that is way more powerful than everyone else's AI. Right? So if you look at like, like, I guess one security analogy that I think of is,

• 42:50Speaker 01

  you know, it doesn't take AI to basically, okay, there's security holes in so many different things. And if you could travel back in time a year or two years, right? It's like, that's not AI. It's like you just, let's say you just have like one year, two years more knowledge of the security holes. Pretty much hack into like any system, right? So it's not that far fetched to believe that a very intelligent AI would probably be able to identify some holes. And basically be like a human who could potentially go back in time a year or two and compromise all these systems. Okay, so how have we dealt with that as a society? Well, One big part is open source software that makes it so that when improvements are made to

• 43:31Speaker 01

  the software, it doesn't just kind of get stuck in one company's products, but it can kind of be broadly deployed to a lot of different systems, whether it's banks or hospitals or government stuff. And like, just everyone can kind of, like, as the software gets hardened, which happens because more people can see it and more people can bang on it. And there are standards on how this stuff works. The world can kind of get upgraded together pretty quickly. And I kind of think that a world where AI is very widely deployed in a way where it's gotten hardened progressively over time, and is one where all the different systems will be in check in a way that seems like it is fundamentally more healthy to

• 44:13Speaker 01

  me than one where this is more concentrated. So there are risks on all sides, but I think that that's one risk that I think people, I don't hear them talking about quite as much. I think like, there's sort of the risk of like, okay, well, what if the AI system does something bad? I am more like, you know, I stay up at night more worrying, well, what if like, some actor that, whatever, it's like, from wherever you sit, there's going to be some actor who you don't trust. If they're the ones who have like the super strong AI, whether it's some like, other government that we, that that is sort of like an opponent of our country, or some company that you don't trust, or whatever

• 44:51Speaker 01

  it is. Like, I think that that's potentially a much bigger risk.

• 44:59Dwarkesh Patel?

  As in, they could like overthrow our government because they have a weapon that like nobody else has?

• 45:06Speaker 01

  Cause a lot of mayhem, right? It's, I think it's like, I, I, I mean, I think the intuition is that this stuff ends up being pretty kind of important and, and valuable for both kind of economic and, and kind of security and other things. And I don't know, I just think, yeah, if, if like, if someone who you, you don't trust or is an adversary of you, gets something that is more powerful, then, then I think that that could be an issue. And I think the probably the best way to mitigate that is to have good open source AI that

• 45:37Speaker 01

  That basically becomes the standard and in a lot of ways can become the leader. And in that way, it just ensures that it's a much more even and balanced playing field.

• 45:49Dwarkesh Patel?

  Yeah, that seems plausible to me. And if that works out, that would be the future I prefer. I guess I want to understand mechanistically how if somebody was going to cause mayhem with AI systems, how the fact that there are other open source systems in the world prevents that? Like the specific example of like somebody coming with a bioweapon. Is it just that we'll do a bunch of R and D in the rest of the world to like figure out vaccines really fast? Like what's happening?

• 46:13Speaker 01

  If you take the computer, the security one that I was talking about, I think someone with a weaker AI trying to hack into a system that is protected by a stronger AI will succeed less. Right? So I think that that's, I mean, that's like in terms of software.

• 46:28Dwarkesh Patel?

  How do you know everything in the world is like that? Like what if bioweapons aren't like that?

• 46:32Speaker 01

  No, I mean, I don't know that everything in the world is like that.

• 46:38Speaker 01

  I think that that's, I guess, one of the, bioweapons are one of the areas where I think the people who are most worried about this stuff are focused. And I think that that's, I think that makes a lot of sense to think about that.

• 46:52Speaker 01

  The, and I think that there are certain mitigations. You can try to not train certain knowledge into the model, right? There's different things, but yeah, I mean, it's some level. If you get a sufficiently bad actor and you don't have other AI that can sort of balance them and understand what's going on and what the threats are, then that could be a risk. So I think that that's one of the things that we need to watch out for.

• 47:21Dwarkesh Patel?

  Is there something you could see in the deployment of these systems where you observe like you're training Lama for and it's like lied to because you thought you weren't noticing or something? And you're like, whoa, what's going on here? Not that this is probably not likely with Lama for that system, but is there something you can imagine like that? Where you'd like be really concerned about deceptiveness and if like billions of copies of things are out in the wild?

• 47:48Speaker 01

  Yeah, I mean, I think that that's not necessarily, I mean, right now, it's we see a lot of hallucinations. Yeah. Right. So I think it's more more that. I think it's it's an interesting question how you would tell the difference between a hallucination and deception. But yeah, I mean, I look, I mean, I think there's a lot of risks and things to to think about. The um. The flip side of all this is that there are also a lot of. I try to in in in running our company, at least. Balance what I think of is these longer term theoretical risks. With what I actually think are quite real risks that exist today.

• 48:27Speaker 01

  So like, when you when you talk about deception, the form of that that I worry about most is people using this to generate misinformation. And then like pump that through, whether it's our networks or others. So the way that we've basically. Combated a lot of this type of harmful content is by building AI systems that are smarter than the adversarial ones. And I guess this is part of this kind of informs part of my theory on this, right? Is if you look at like the different types of harm that people do. Or try to do through through social networks. There are ones that are not very adversarial. So for example, like.

• 49:06Speaker 01

  Hate speech, I would say, is not super adversarial in the sense that like, people aren't getting. Better at being racist, right? They're just like, it's you just like, OK, if you you kind of, that's one where I think the AIs are generally just getting. Way more sophisticated faster than people are at those issues. So we have and we have issues both ways. It's like, people do bad things that whether they're trying to incite violence or something. But we also have a lot of false positives, right? So where we where we basically censor stuff that we shouldn't. And I think understandably make a lot of people annoyed. So I think having an AI that just gets increasingly precise on that. That's gonna be good over time. But let me give you another example, which is like nation states trying to interfere in elections.

• 49:51Speaker 01

  That's an example where they're absolutely, they have cutting edge technology and absolutely get better each year. So we block some technique, they learn what we did, they come at us with a different technique, right? It's not like a person trying to, you know, I don't say, say mean things, right? It's like, it's, it's they're, they're basically, they have a goal, they're sophisticated, they have a lot of technology. In those cases, I, I still think the ability to kind of have our AI systems grow and in sophistication at a faster rate than theirs have. It's an arms race, but I think we're at least currently winning that arms race. So I don't know, I think that that's, but this is like a lot of the stuff that I, that I spend time thinking about is like, okay.

• 50:34Speaker 01

  Yes, it is possible that whether it's Lama four or Lama five or Lama six, yeah. We need to think about like what behaviors we're, we're observing and it's not just us. And part of the reason why you make this open source is that there are a lot of other people who study this too. So, yeah, we want to see what other people are observing, what we're observing, what we can mitigate. And then we'll make our, our assessment on whether we can make it open source. But I, I think for the foreseeable future, I'm, I'm optimistic we will be able to. And in the near term, I don't want to take our eye off the ball of what are actual bad things that people are trying to use the models for today. Even if they're not existential, but they're like, they're like pretty bad kind of day to day harms that we are familiar with and running our services.

• 51:21Speaker 01

  That's actually a lot of what we have to, I think, spend our time on as well.

• 51:24Dwarkesh Patel?

  Yeah, yeah. Actually, I, I, I found the synthetic data thing really curious. I'm, I'm actually interested in why you don't think, like current models, it makes sense why there might be an asymptote with just doing the synthetic data again and again. If they get smarter and you use this kind of techniques you talk about in the paper or the blog post that's coming out on the day this will be released where it, it goes to the thought chain that is the most correct. Why you, why this wouldn't like lead to a loop that, of course, it wouldn't be overnight, but over many months or years of training potentially with a smarter model. It gets smarter, makes better output, gets smarter and so forth.

• 52:01Speaker 01

  Well, I think it could within the parameter of whatever the model architecture is. It's just that like at some level.

• 52:10Speaker 01

  I don't know. I think like today's 8 billion parameter models, I just don't think you're going to be able to get to be as good as the state of the art, multi hundred billion parameter models that are incorporating new research into the architecture itself.

• 52:29Dwarkesh Patel?

  But those will be open source as well, right?

• 52:31Speaker 01

  Well, yeah, but I think that that's if, I mean, subject to all the

• 52:35Dwarkesh Patel?

  Yeah.

• 52:36Speaker 01

  questions that we just talked about. But yes, I mean, we would, we would hope that that'll be the case. But I, but I think that at each point, I don't know, it's like, when you're building software, there's like a ton of stuff that you can do with software. But then at some level, you're constrained by the chips that it's running on, right? So there are always going to be different physical constraints. And it's like, how big are the models is going to be constrained by how much energy you can get and and use for inference. So

• 53:10Speaker 01

  I guess I'm simultaneously very optimistic that this stuff will continue to improve quickly. And also a little more measured than I think some people are about. Kind of it's, I, I just don't think the runaway case is like a particularly likely one.

• 53:31Dwarkesh Patel?

  Hmm. I think it makes sense to keep your options open. Like, there's so much we don't know. There's a case in which like, it's really important to keep the balance of power. So when nobody becomes like a totalitarian dictator, there's a case in which like, you don't want to open source the architecture because like China's catch, it can use it to catch up to America's AIs. And like, there is an intelligence explosion and they like win that. Yeah, a lot of things seem impossible. Just like keeping your options open, considering all of them seems reasonable.

• 53:57Speaker 01

  Yeah.

• 53:58Dwarkesh Patel?

  Let's talk about some other things.

• 54:00Speaker 01

  Okay.

• 54:01Dwarkesh Patel?

  Metaverse, what time period in human history would you be most interested in going into? A 100,000 BCE to now. You just want to see what it was like.

• 54:10Speaker 01

  Through the past?

• 54:11Dwarkesh Patel?

  Huh?

• 54:11Speaker 01

  It has to be the past?

• 54:12Dwarkesh Patel?

  Oh, yeah.

• 54:12Speaker 01

  It has to be the past.

• 54:18Speaker 01

  I don't know. I mean, I have the periods of time that I'm interested. I mean, I'm really interested in American history and classical history. And I'm really interested in the history of science, too. So I actually think seeing and trying to understand more

• 54:35Speaker 01

  about how some of the big advances came about. I mean, all we have are like somewhat limited writings about some of that stuff. I'm not sure the metaverse is going to let you do that because I mean, it's, you know, we can't, it's going to be hard to

• 54:48Speaker 01

  to kind of go back in time for things that we don't have records of. But

• 54:54Speaker 01

  I'm actually not sure that going back in time is going to be that, that, that important of a thing for them. I mean, I think it's going to be cool for like history classes and stuff, but

• 55:02Speaker 01

  that's probably not the use case that I'm most excited about for the, for the metaverse overall. I mean, it's, I think the main thing is just the ability to feel present with people, no matter where you are. I think that that's going to be killer. I mean, there's,

• 55:16Speaker 01

  I mean, in the AI conversation that we, that we're having, and it's you know, so much of it is about physical constraints that kind of underlie all of this, right? And you want to move. I think one lesson of technology is you want to move things from the physical constraint realm into software as much as possible because software is so much easier to build and and evolve and like you can democratize it more because like not everyone is going to have a data center, but like a lot of people can, can kind of write code and take open source code and modify it.

• 55:49Speaker 01

  The metaverse version of this is I think enabling realistic digital presence is going to be just an absolutely huge difference for, for making it so that people don't feel like they have to physically be together for as many things. Now, I mean, I think that there are going to be things that are better about being physically together.

• 56:12Speaker 01

  So it's not, I mean, these things aren't binary. It's not going to be like, okay, now it's, you don't need to do that anymore. But

• 56:19Speaker 01

  But overall, I mean, I think that this it's just gonna be really powerful for for socializing, for feeling connected with people, for working, for, I don't know, parts of industry, for medicine, for like, like so many things.

• 56:36Dwarkesh Patel?

  I want to go back to something you said at the beginning of the conversation where you didn't sell the company for a billion dollars and like the metaverse, you knew we were gonna do this even though the the market was hammering you for it. And then I'm actually curious, like, what is the source of that edge? And you said, like, oh, values, I have this intuition, but like, everybody says that, right? Like, what if you had to say something that's specific to you? What is, how would you express what that is? Like, why were you so convinced about the metaverse?

• 57:06Speaker 01

  Well, I think that those are different questions. So what I mean, what what are the things that that kind of power me? I think we've talked about it on the theme. So it's, I mean, I I just really like building things. I specifically like building things around how people communicate and sort of understanding how people express themselves and how people work, right? When I was in college, I was I was said computer science and psychology. I think a lot of other people in the industry studied computer science, right? So it's a it's always been sort of the intersection of those two things for me. But I think it's also sort of

• 57:44Speaker 01

  this, like, really deep drive. I don't know how to explain it, but I just feel like in constitutionally, like, I'm doing something wrong if I'm not building something new. Right? And so I think that there's like, you know, even when we're putting together the business case for, you know, investing like $100 billion in AI or some huge amount in the metaverse. It's like, yeah, I mean, we have plans that I think make it pretty clear that if our stuff works, it'll be a good investment. But like, you can't know for certain from the outset.

• 58:25Speaker 01

  And so there's all these arguments that people have, you know, whether it's like, you know, with advisors or or different folks, it's like, well, how could you like, it's how are you confident enough to do this? And it's like, well, The day I stop trying to build new things, I'm just done. I'm gonna go build new things somewhere else, right? It's like, um, it's like, it is, I'm fundamentally incapable of

• 58:52Speaker 01

  running something or in my own life and like not trying to build new things that I think are are interesting. It's like, that's not even a question for me, right? It's like whether, like whether we're gonna go take a swing at like building the next thing. It's like, it's, I like, I'm, I'm just incapable of not doing that.

• 59:11Speaker 01

  And

• 59:14Speaker 01

  I don't know. I, I'm kind of like this in like all the different aspects of my life, right? It's like, we built this like, our family built this ranch in Kauai and like, I just like, worked like design all these buildings. I'm like, kind of trying to like, we started raising cattle and I'm like, all right, well, I want to make like the best cattle in the world, right? So it's like, how do we, like, how do we architect this so that way we can figure this out and like, and build like all the stuff up that we need to, to try to do that?

• 59:43Speaker 01

  So I don't know, that's me.

• 59:46Speaker 01

  What was the other part of the question?

• 59:48Dwarkesh Patel?

  Look, Meta is just a really amazing tech company, right? They have all these great software engineers and even they work with Stripe to handle payments. And I think that's just a really notable fact that Stripe's ability to engineer these checkout experiences is so good that big companies like Ford, Zoom, Meta, even OpenAI, they work with Stripe to handle payments. Because just think about how many different possibilities you have to handle. If you're in a different country, you'll pay a different way. And if you're buying a certain kind of item, that might affect how you decide to pay. And Stripe is able to test these fine grained optimizations across tens of billions of transactions a day to figure out what will convert people. And obviously, conversion means more revenue for you. And look, I'm not a big company like Meta or anything, but I've been using Stripe since long before they were advertisers.

• 1:00:33Dwarkesh Patel?

  Stripe Atlas was just the easiest way for me to set up an LLC. And they have these payments and invoicing features that make it super convenient for me to get money from advertisers. And obviously, without that, it would have been much harder for me to earn money from the podcast. And so it's been great for me. Go to stripe.com to learn more. Thanks to them for sponsoring the episode. Now back to Mark. I'm not sure, but I'm actually curious about something else, which is, um, so a 19-year-old Mark reads a bunch of like antiquity and classics, uh, high school, college. What important lesson did you learn from it? Not just interesting things you found, but like, there aren't that many tokens you consume by the time you're 19. A bunch of them were about the classics. Clearly that was important in some way.

• 1:01:13Speaker 01

  Tokens.

• 1:01:22Speaker 01

  I don't know, that's a good question. I mean, one of the things that I thought was really fascinating is, um, so when Augustus was first, so he became emperor and, um, and he was trying to establish peace. And the, there was no real conception of peace at the time. Like the people's, people's understanding of peace was, it is the temporary time between when your enemies will inevitably attack you again. So you get like a short rest. And, and he had this view, which is like, look, like, we want to change the economy

• 1:02:02Speaker 01

  from instead of being so mercenary and like, and kind of militaristic to like, actually this positive sum thing. It was like a very novel idea at the time.

• 1:02:18Speaker 01

  I don't know. I think that there's like something that's just really fundamental about that. It's like in terms of the, the bounds on like what people can conceive at the time of like, what are rational ways to work? And, um, I don't know, I'm going back to like, I mean, this applies to both the metaverse and the AI stuff, but like a lot of investors and just different people just can't wrap their head around why we would open source this. And it's like, are you like, like, I don't understand. It's like, open source, that must just be like the temporary time between which you're making things proprietary. Right? And it's, um, but, but I actually think it's like this very profound thing in tech that has actually, it, it creates a lot of winners. Right?

• 1:03:03Speaker 01

  And it's, and, um, so I don't know, I don't want to strain the analogy too much, but, but I do think that there's, um, there's a lot of times, I think, ways where you can.

• 1:03:15Speaker 01

  That are just like models for building things that people can't even like they just like often can't wrap their head around how that would be a valuable thing for people to go do or like a reasonable state of the world that it's I mean it's a I think that there's more reasonable things than people think.

• 1:03:36Dwarkesh Patel?

  That's super fascinating. Could I give you my answer, what I was thinking?

• 1:03:40Speaker 01

  Sure.

• 1:03:40Dwarkesh Patel?

  You might have gotten from it. This is probably totally off, but just how young some of these people are who have very important roles in the empire, like Caesar Augustus is like, by the time he's 19, he's actually incredibly one of the most prominent people in Roman politics. And he's like, leading battles and forming the Second Trireme. I wonder if you're like, the 19 year old is like, I can actually do this because like Caesar Augustus did this.

• 1:04:01Speaker 01

  I think that's an interesting example, both from a lot of history and American history, too. I mean, it's, I mean, one of my favorite quotes is it's this Picasso quote, that all children are artists and the challenge is how do you remain an artist when you grow up? And it's like, basically, I think, because when you're younger, I think it's just easier to have kind of wild ideas and you're not, you know, you have no, there are all these analogies to the innovator's dilemma that exist in your life as well as your company or whatever you've built, right? So, you know, you're kind of earlier on your trajectory, it's easier to pivot and take in new ideas without disrupting other commitments that

• 1:04:43Speaker 01

  you've made to to different things. And so I don't know, I think that's an interesting part of of running a company is like, how do you, how do you kind of stay dynamic?

• 1:04:55Dwarkesh Patel?

  Going back to the investors in open source, the $10 billion model, suppose it's totally safe, you've done these evaluations, and unlike in this case, the evaluators can also fine tune the model, which hopefully will be the case in future models. Would you open source that, the $10 billion model?

• 1:05:11Speaker 01

  Well, I mean, as long as it's helping us, then yeah.

• 1:05:13Dwarkesh Patel?

  But would it like to $10 billion of R and D? And then now it's like open source or anyway?

• 1:05:17Speaker 01

  Well, I think here's, I think, a question, which we'll have to evaluate this as as time goes on, too. But

• 1:05:26Speaker 01

  We have a long history of open sourcing software. We don't tend to open source our product. So it's not like we don't take like the code for Instagram and make it open source. But we take like a lot of the low level infrastructure and we make that open source. The probably the biggest one in our history was Open Compute Project where we took the designs for kind of all of our servers and network switches and data centers and made it open source and ended up being super helpful because, you know, I mean, a lot of people can design servers, but now like the industry standardized on our design, which meant that the supply chains basically all got built out around our design and the volumes went up. So it got cheaper for everyone and saved us billions of dollars. So awesome, right?

• 1:06:09Speaker 01

  Okay, so there's multiple ways where open source, I think, could be helpful for us. One is if people figure out how to run the models more cheaply, well, we're going to be spending tens or like a hundred billion dollars or more over time on all this stuff. So if we can do that 10% more effectively, we're saving billions or tens of billions of dollars. Okay, that's probably worth a lot by itself, especially if there's other competitive models out there. It's not like our thing is like be giving away some kind of crazy advantage. So is your view that the trading will be commodified?

• 1:06:45Speaker 01

  I think there's a bunch of ways that this could play out. That's one. The other is, is that, so commodity kind of implies that it's going to get very cheap because there's lots of options. The other direction that this could go in is qualitative improvements. So you mentioned fine tuning, right? It's like right now, it's pretty limited what you can do with fine tuning major other models out there. And there are some options, but generally not for the biggest models. So I think being able to do that and be able to kind of do different app specific things or use case specific things or build them

• 1:07:24Speaker 01

  into specific tool chains, I think will not only enable kind of more efficient development, it could enable qualitatively different things. Here's one analogy on this is.

• 1:07:38Speaker 01

  So one thing that I think generally sucks about the mobile ecosystem is that like you have these two gatekeeper companies, Apple and Google, that can tell you what you're allowed to build. And there are lots of times in our history. So there's the economic version of that, which is like, all right, we build something and they're just like, I'm going to take a bunch of your money. But then there's the qualitative version, which is actually what kind of upsets me more, which is there's a bunch of times when we've launched or wanted to launch features, and then Apple's just like, nope, you're not launching that. I was like, that sucks. Right? And so the question is, what is like, are we kind of set up for

• 1:08:17Speaker 01

  a world like that with AI where like, you're going to get a handful of companies that run these closed models that are going to be in control of the APIs and therefore are going to be able to tell you what you can build? Well, for one, I can say for us, it is worth it to go build a model ourselves to make sure that we're not in that position. Right? Like, I don't want any of those other companies telling us what we can build. But from an open source perspective, I think a lot of developers don't want those companies telling them what they can build either. So the question is, what is the ecosystem that gets built out around that? What are interesting new things? How much does that improve our products?

• 1:08:59Speaker 01

  I think that there's a lot of cases where if this ends up being like, you know, like our databases or caching systems or architecture, we'll get valuable contributions from the community that'll make our stuff better. And then our app specific work that we do will still be so differentiated that it won't really matter. Right? It's like, well, we'll be able to do what we do. We'll benefit in all the systems, ours and the communities will be better because it's open source. There is one world where maybe it's not that. I mean, maybe the model just ends up being more of the product itself. In that case, then I think it's a trickier economic calculation about whether you open source that because then you are kind of commoditizing yourself a lot. But I don't, from what I can see so far, it doesn't seem like we're in that zone.

• 1:09:42Dwarkesh Patel?

  Do you expect to earn significant revenue from licensing your model to the cloud providers? So they have to pay you a fee to actually serve the model?

• 1:09:52Speaker 01

  We want to have an arrangement like that, but I don't know how significant it'll be. And we have this, this is basically our license for Lama.

• 1:10:01Dwarkesh Patel?

  Yeah.

• 1:10:02Speaker 01

  You know, in a lot of ways, it's like a very permissive open source license, except that we have a limit for the largest companies using it. And this is why we put that limit in, is we're not trying to prevent them from using it. We just want them to come talk to us because if they're going to just basically take what we built and resell it and make money off of it, then it's like, okay, well, if you're like, you know, Microsoft Azure or Amazon, then yeah, if you're going to be reselling the model, then we should have some revenue share on that. So just come talk to us before you go do that. And that's how that's played out. So for Lama 2, it's, I mean, we basically just have deals with all these major cloud companies and Lama 2 is available as a hosted service on all those clouds. And I assume that as we, as we release bigger and bigger models,

• 1:10:47Speaker 01

  that'll become a bigger thing. It's not the main thing that we're doing, but I just think if others are, if those companies are going to be selling our models, it makes sense that we should, you know, share the upside of that somehow.

• 1:10:55Dwarkesh Patel?

  Yeah. With regards to the other open source dangers, I think you have a genuine legitimate points about the balance of power stuff and potentially like the harms you can get rid of because we have better alignment techniques or something. I wish there was some sort of framework that Meta had, like other labs have this where they say like, if we see this concrete thing, then that's a no go on the open source or like even potentially on deployment, just like writing it down. So like the company is ready for it. People have expectations around it and so forth.

• 1:11:25Speaker 01

  Yeah, no, I think that that's a fair point on the existential risk side. Right now, we focus more on the types of risks that we see today, which are more of these content risks. So, you know, we have lines on, we don't want the model to be basically doing things that are helping people commit violence or fraud or, you know, just harming people in different ways. So.

• 1:11:49Speaker 01

  In practice for today's models, and I would guess the next generation and maybe even the generation after that, I think while it is somewhat more maybe intellectually interesting to talk about the existential risks, I actually think the the real harms that need more energy being mitigated are things that are going to like to have someone take a model and do something to hurt a person with today's parameters of and kind of the types of kind of more mundane harms that we see today, like people kind of committing fraud against each other, things like that. So that I just don't want to shortchange that. I think we have a responsibility to make sure we do a good job on that.

• 1:12:33Dwarkesh Patel?

  Yeah, Meta's a big company. You can handle both.

• 1:12:35Speaker 01

  Yeah.

• 1:12:37Dwarkesh Patel?

  Okay, so as far as the open source goes, I'm actually curious if you think the impact of the open source from PyTorch, React, Open Compute, these things has been bigger for the world than even the social media aspects of Meta? Because I've like talked to people who use these services who think like it's plausible, because a big part of the internet runs on these things.

• 1:12:55Speaker 01

  It's an interesting question. I mean, I think almost half the world uses our

• 1:12:59Dwarkesh Patel?

  Yeah, that's a tricky point.

• 1:13:03Speaker 01

  So I think it's hard to beat that. But no, I think I think open source is it's really powerful as a new way of building things. And yeah, I mean, it's possible. I mean, it's, you know, it may be one of these things where, I don't know, like Bell Labs, right, where they, you know, it's like they were working on the transistor because they wanted to enable long distance calling. And they did. And it ended up being really profitable for them, that they were able to enable long distance calling. And if you ask them five to 10 years out from that, what was the most useful thing that they invented? It's like, okay, well, we enabled long distance calling. And now all these

• 1:13:48Speaker 01

  people are long distance calling. But if you ask 100 years later, maybe it's a different question. So.

• 1:13:55Speaker 01

  I think that that's true of a lot of the things that we're building, right? Reality Labs, some of the AI stuff, some of the open source stuff. I think it's like the specific products evolve and to some degree come and go. But I think like the advances for humanity persist. And that's like a, I don't know, a cool part of what we all get to do.

• 1:14:14Dwarkesh Patel?

  By when will the Llama models be trained on your own custom silicon?

• 1:14:22Speaker 01

  Soon, not, not, not Llama 4. The approach that we took is first, we, we basically built custom silicon that could handle inference for our ranking and recommendation type stuff. So reels, newsfeed, ads. And that was consuming a lot of GPUs. But when we were able to move that to our own silicon, we now were able to use the more expensive Nvidia GPUs only for training. So at some point, we will hopefully have silicon ourselves that we can be using for probably first

• 1:15:03Speaker 01

  training, some of the simpler things, then eventually training these like really large models.

• 1:15:11Speaker 01

  But in the meantime, I'd say the program is going quite well and we're just rolling it out methodically and have a long-term roadmap for it.

• 1:15:19Dwarkesh Patel?

  Final question, this is totally out of left field, but if you were a mate CEO of Google Plus, could you have made it work?

• 1:15:25Speaker 01

  Google Plus? Oof. Well, I don't know. I don't know. That's, that's, that's a very difficult, very difficult counterfactual.

• 1:15:35Dwarkesh Patel?

  Okay, then the real final question will be when Gemini was launched, did you, was there any chance that somebody in the office uttered Carthago de Linda est?

• 1:15:44Speaker 01

  No, I think we're tamer now.

• 1:15:47Dwarkesh Patel?

  Cool, cool. Awesome, Mark.

• 1:15:52Speaker 01

  Yeah, I don't know. It's a good question. I don't know. The problem is there was no CEO of Google Plus. It was just like a division within a company. I think it's like, and you asked before about what are the kind of scariest commodities, but you asked about it in terms of dollars. And I actually think for most companies, it's, of this scale at least, it's focus, right? It's like when you're a startup, maybe you're more constrained on capital. You know, you just are working on one idea and you might not have all the resources. I think you cross some threshold at some point where the nature of what you're doing, you're building multiple things and you're creating more value across them. But you become more constrained on what can you direct and to go well.

• 1:16:35Speaker 01

  And like, there's always the cases where something just random awesome happens in the organization. I don't even know about it and those are, that's great. But like, but I think in general, the organization's capacity is largely limited by what like the CEO and the, and the management team are able to kind of oversee and kind of manage. It's, I think that that's just been a big focus for us. It's like, all right, keep the, as I guess Ben Horowitz says, keep the main thing, the main thing, right? And, and try to kind of stay focused on your key priorities. Yeah.

• 1:17:15Dwarkesh Patel?

  All right. Awesome. That was excellent, Mark. Thanks so much. That was a lot of fun.

• 1:17:17Speaker 01

  Yeah, really fun. Thanks for having me.

• 1:17:20Dwarkesh Patel?

  Absolutely. Hey, everybody. I hope you enjoyed that episode with Mark. As you can see, I'm now doing ads, so if you're interested in advertising on the podcast, go to the link in the description. Otherwise, as you know, the most helpful thing you can do is just share the podcast with people who you think might enjoy it. You know, your friends, group chats, Twitter, I guess threads. Yeah, I hope you enjoyed and I'll see you on the next one.