Let's bring him in. What's up, guys? Hey, how you doing? There he is. What's going on? Good to see you guys. It's great to have you. Good to see you. I heard you've been doing space repetition to practice for interviews.
I've been doing it. It says your name is Dwarkesh Patel. You have a podcast and you interviewed somebody named Mark who owns a website called facebook.com. Can you tell me about that?
Yeah, in fact, I made an entire Anki deck just for you guys, for Printing.
Great. How's the book launch been going? How's the press tour? How are you doing today?
It's been going good. I have a book out called The Scaling Era, and it compiles the interviews I've been doing about AI with people like Mark and Demis and Dario, the heads of the AI labs, but also researchers and engineers and philosophers and economists. It's been really interesting because AI is one of those topics where there are so many fractal questions you could ask — what is its impact going to be, how are we going to train it, how do you even think about a super intelligence? So I've been dealing with a lot of different kinds of questions, which has made it interesting.
It's great. How contentious was the book process? Did you know you wanted to go with Stripe? I imagine you could have had your pick of the litter in terms of legacy publishers that promised you all sorts of things, but you happen to own your own distribution, so maybe it was just about picking the right sort of underlying partner for it.
It honestly was never a matter of picking a publisher. The main question was whether I should do a book in the first place. Some folks at Stripe Press reached out, and if I was going to do a book, it would be with them because, as you know, their reputation precedes them. So then it was just like deciding do I want to do the book or not, and in retrospect it was the right call. I'm really delighted with how it turned out.
I want to talk about acceleration. Are you feeling the acceleration? Mathematically, we are not accelerating GDP yet, although technically we are today. I think GDP ticked up just a little bit, which means we're technically accelerating. It must be your podcast.
Hopefully. But energy use is not accelerating. Even some of the benchmarks are kind of saturating. We're not seeing acceleration curves, we're seeing solid growth, but at the same time it feels like we're on the precipice of acceleration. But what does it mean to feel the acceleration for you?
I think it's a really good question because we have these models which we think are smart and, as you say, we haven't seen them even automate the things which, like when we're having a conversation, we'd be like, oh, the call center workers they should be really worried — and they still have got their jobs. So what's going on? People have been talking about needing to make the models cheaper. When DeepSeek came out they were like, oh, it's going to be Jevons paradox and we'll be using them way more now that they're cheaper. I think the real bottleneck is just we've got to make them smarter. They're already so cheap — it's like two cents per million tokens or something ridiculous. The real bottleneck for me using them more is not their price but them being more useful, being able to take over more parts of the economy.
Do you think that intelligence is all we need? Andrej Karpathy was talking about the importance of agency. I've talked to other people about maybe it's all — like what makes humans effective? It's not just intelligence, it's also agency, it's also coordination, friendliness, networking. Tyler Cowen talked about like, do we even need to map the different parts of the skill tree that humans have? Like charisma, wisdom — like underrated. The AIs are getting more intelligent but they're already maxed out on wisdom, right? But do we need to think about a different taxonomy here, or do we just need to max out intelligence and everything else will come?
No, I think you're absolutely right. You need a lot more skills. There has been this trend in AI where whenever there's a big breakthrough, we think we've automated a large part of what intelligence is, and in fact in retrospect it's clear that it was only a beginning. The big example here is when Deep Blue came out and beat Kasparov at chess. People thought that this was like a big breakthrough in intelligence in general because we thought that what chess required was general intelligence, and you might have heard this concept of AI-complete problems, where if you solve this problem then you've solved intelligence. So people said that about self-driving, the Turing test was supposed to be AI-complete. We've gone through all of these subcomponents of intelligence, and afterwards we realized there's actually still more left to it.
The thing that's sort of underrated is not even agency per se, although that's a part of it. I think the thing that's underrated is we humans have this global hive mind where the reason we can make iPhones and buildings and whatever is not just intelligence and also not just agency — it's the fact that there's so much specialization, there's so much capital deepening, people are just doing things, trying different ideas. AIs need to be smarter in order to do that, they need to have more agency in order to do that, but once they can, you just — if you have millions of AIs running around trying different things, that's when we get the real acceleration, and you'll feel it in your blood.
When you say millions of AIs — you've said that you think there will be billions of AIs running around. What does that actually mean? How can we quantify — are we just talking like ChatGPT instances? Are we talking about individual — because you can inference multiple threads on a single chip. How are you thinking about it? When you say it sounds more concrete — when you say there will be billions of AIs — but is that really just like there will be a hive mind that is equivalent to billions of people? Are we talking about like one, or maybe each model is like one entity but then there's sub-threads? How do you think about that concept of billions of AIs?
Honestly, I don't think anybody knows. I think it'll just depend on how the tech tree shapes out. I have heard these wild ideas in some of these interviews where one person at Jane Street mentioned this idea of 'the blob.' The blob is — right now it's really hard for an institution, organization, or company — it's really hard for the person at the center to have that much awareness of what's happening in the company, to control it to any great extent. Xi Jinping has the same 10 to 15 FLOPS in his brain as any other Chinese person or any other person in general. And in the future you can imagine that the thing at the center just has way more compute, and it's not clear whether — you can think about it as more copies of Xi Jinping or something — but you could just have this huge blob that's constantly learning more things. It's writing every single press release the company releases, it's reading every single pull request, it's answering every single customer response. I don't know if that at all helped answer the question. It's just a very weird question of how this will actually play out.
Will we need to recreate the concept of an individual brain and then copy-paste it a billion times, or do we just need one really big brain? It's unclear to me.
Jord, what do you got? How did you feel on Wednesday — the sort of Ghibli moment? A lot of people were saying, oh, I can't even go on X, it's just all slop. My takeaway was, this is not slop, this is beautiful. This is actually the most beautiful the timeline has ever looked. The most powerful thing about that moment — it was the first time that the entire world could get consistently perfect outputs without any sort of prompt engineering, basically just one-shotting these outputs in a very scaled way. How did you kind of react to some people saying this is bad or I'm going to log off forever?
How many times have those people said that? I think that's a very zero-sum — people were saying things like, oh, you're eating up the fossil fuel, which is like our affection for Ghibli somehow by making these things. I think it's just a very zero-sum view of the world where there can be a limited amount of beauty, there can be a limited amount of joy. I just don't think that's true. Can I be honest? The thing I was really feeling when all these Ghibli images were coming out — I became more convinced of our glorious transhumanist future. Look, you're getting a glimpse just from these early images of how cool and beautiful the things AI makes or helps us make will be. Just imagine this scaled up like 100x, 1,000x, integrated into all our senses, maybe even into our minds, integrated into the way we relate with the people we care about and so forth. The future could be really beautiful.
I agree.
I'm wearing VR goggles and it's making everyone beautiful.
It's really rose-colored glasses, right?
How do you feel — so Manas AI is apparently doing a roadshow in the United States right now, sort of raising for American VCs potentially. I saw some people pushing back on the timeline saying, bad luck to any American VC that does that. The criticism would be we're in this sort of AI Cold War, American venture capital dollars shouldn't be funding companies that are potentially competing with U.S. AI labs or application layer companies. What's your broad take on this sort of cross-border investment in AI?
Yeah. I was in China a few months earlier and it was really striking to me how dismayed the venture capital system there felt and the tech CEOs generally, because after the 2021 crackdowns people really pulled back. And it sounds like after the DeepSeek moment that changed at least in AI, because now the state — the city funds and whatever — are more willing to push in.
And how should people react to this?
I'm of two minds. One, I do believe that there could be an intelligence explosion and you really want to be ahead of that, and you don't want to help them get ahead on that. So I think the export controls and whatever are wise. As for just Manas, it seems like the middle ground here. I wouldn't want to just generally try to harm China by tariffing batteries or cars or something. This is an application of AI and it's complementary to American AI foundation labs because they're using the Claude model, right? So honestly, I don't have a strong take. What do you guys think?
I'm exactly in the same position as you. I don't have a — if American — I did find it a little bit weird that some American venture capitalists were just cheering on DeepSeek, just like blanket statement, open source is good, this is good, when it felt like the way the launch was rolled out, announced, was done in a way to potentially harm American financial markets. But my take has been that just on a pure investment finance level, investing in a Chinese company can be difficult to get your money back at a certain point because the money just gets kind of stranded there, and then depending on who's in charge of America at the time, it could be very onerous to bring that money back.
But wait, I want to stay on China real quick. Fantastic piece — I don't even know if you call it a video essay — but one of the things that really stuck out to me as a creator like yourself was the lack of a Chinese Joe Rogan, basically. Have you thought about that more? Have you unpacked that more? You would almost expect that even if all the crazy censorship is true, why is there no power law winner, there's a Joe Rogan who just spouts propaganda constantly — what is going on that's driving the lack of these long-tail, country-famous people?
I feel like you guys might actually have a good perspective on this, because somebody might have said why doesn't tech have their Rogan equivalent, and you guys started your podcast network. And I think somebody could have said before you guys started your podcast network, why doesn't something like this exist already?
I honestly — I don't speak Chinese and this is secondhand stuff I heard in China. I feel reluctant to make grand conclusions about Chinese culture based on like, why don't they have a Joe Rogan. The sense I got was that whether it's young people or just people in general, what they want to consume is often more focused on practical matters. And if you listen to Joe Rogan, it's very much like, let's just shoot the shit about whatever. I get the sense that that is just not that interesting, at least to the people I met.
Yeah, it's less practical stuff. I wonder if there's also an effect where some of the first social networks probably accelerated more quickly to these highly diffuse algorithmic-driven TikTok feeds that allow for smaller micro-celebrities essentially, whereas we in America we've been building up this celebrity culture for so long that we have more of a power law dynamic. I don't have a thesis that's super built out, but it is fascinating.
I was there once on a layover, but Jordy did live there for a while. In 2016, I went to Fudan — which is like, it was a little flex — fun called High Flyer. No big deal. I just studied abroad there, so I was there for a semester, and I worked out of China Accelerator, which is a startup accelerator. A very, very interesting experience that doesn't scale.
It's crazy. I was in Guangzhou even just for a day and the scale of the buildings there — it really is remarkable. You need to see it in person because the pictures kind of compress everything; you don't really understand until you're there. The one thing I found that was weirdly fascinating — I don't know if it was the iPhone camera at that time, but I found it very difficult to get high quality images because it was so polluted. Something about your own eyes would be able to kind of adjust, but I would take a picture and I'm like, that's not what it looks like. I realized over time that certain areas were so heavily polluted that the iPhone camera would just kind of bug out.
I had a question for you. Something from this week that I thought was funny was that we were all kind of holding our breath for the next ChatGPT moment, and then it was also ChatGPT with their image generation product. Do you think we should be holding our breath for another company to sort of experience this sort of moment where it's just full, complete takeover of the mindshare? Mindshare is just so important right now. The benchmarks come out and everybody in our corner of the internet is hyper-fixated on it, but the average consumer — one of the things I thought was fascinating is John and I had a couple Ghibli posts that broke containment and a lot of people were quoting it and saying, all right, tell me what this app is, everybody, what's the joke — they just still didn't know like ChatGPT or anything. I think it would be amazing for the industry if another company could have a moment that big. Do you see something like that happening this year? Even like the agent stuff, flight booking — you could see that there's some application that would go real broad. Is there any next milestone that you're waiting for?
If somebody did get a reliable agent to work, I think that would have a similar break-the-internet quality. Something that you personally could use and you could just log in. I think it'll probably be one of the foundation lab companies. People have been trying to build agents for years and they just haven't worked, and it makes me think that's a fundamental limitation of the current models. So it'll just be the company that is building a future model geared towards computer use and so forth — that is what I expect it to be.
I was thinking the other day — remember when Sam got fired and people were posting on Twitter, oh, there's something called Q*? I don't think that's why he got fired, but it is notable that in retrospect, if you were following the Q* rumors, you actually would have been in a good position to anticipate o1 — that there would be this reasoning breakthrough that's kind of what they were talking about at the time. Similarly, GPT 4.5 and Grok 3 being not that much better — if you were following Twitter six months ago you would have seen, oh, pre-training might plateau but we'll have to go with inference scaling or something. So maybe my update has been that you can sort of know what's going to happen. I remember at the time I was just like, these idiots on Twitter don't know what they're talking about, like a rumor mill. And I'm like, yeah, kind of. You take it with a grain of salt, but they kind of had the big picture. There are no secrets.
There was always this funny dynamic where people were criticizing Sam for launching ChatGPT without telling the board, but then at the same time people were criticizing Sam for being non-technical and not driving the product forward. And I was like, no matter what you think of Sam, those two things cannot be true simultaneously. You have to pick a side here — you can't criticize him for both not innovating and also innovating too fast.
Anyway, I want to talk about a flip side of the P(doom) argument that I've been kicking around. Basically, we've seen these accelerating trends before — nuclear energy, 'too cheap to meter' has been talked about before, and we have hit stagnation. It is possible that something can break in our society and all these different economic and political forces can align to just say, hey, nuclear energy is not going to double every couple years. And it didn't. I'm wondering, what is your P(stagnation)? Just your probability that something happens — maybe people freak out, maybe there's one world government or something — but we actually see AI stall for a significant amount of time, like 50 years. There's no intelligence explosion purely for stagnation reasons. Do you think that's a possibility?
