Dwarkesh Patel Wants People to Learn Things

Like, how long can an AI do a task? Is it like 7 seconds and it loses its train of thought? Is it 14 seconds? And it seems to be doubling every 7 months. So before, they can maybe write you a little function. Now they can write you a whole package of code, a whole file of code. Now they can maybe write you many different lines of code and keep engaging with you over a session as you say, no, I want to add this feature or something. So if that keeps doubling, then you have this really exponential process where You have this discontinuous impact. Going from like 30 minutes to an hour is a big deal. Going from an hour to 2 hours is a big deal. And going from like a day to 2 days is like an incredibly big deal, right? Like the amount of increased scope you can have as a result will increase exponentially. So I expect these kinds of discontinuities. I am still confused about why these models have no traction whatsoever on things which are, which are like short horizon tasks. Like it doesn't take you that long to, let's say, edit a video or book a flight or start playing a new video game, which you've never seen in your trading data. And there they have very little traction whatsoever, or organize the logistics for like a happy hour or something, right? So the big question, it seems like free training is exhausted. The reasoning paradigm, the big question is how far that gets you. And so we had O1, O2, I guess we skipped O2 and now we went to O3. The next question is like, okay, look, they've conclusively shown that this gets you a ton of traction on math and code. What are the other kinds of RL environments you can spin up where you get this, you know, thing where you're like, now you're doing tasks that are like a human's day-long worth of input? I think in 3 months we will know if, um, do we have models which can reliably do computer use? Can you set up an RL environment for that? If not, then I think the next, like, we're not going to have some huge economic impact from AI that we haven't seen so far in the next 18 months. Because fundamentally, like, there's not that much in the economy that's just like chat-based. But if you do have that, then I think you can start seeing tens of billions of dollars of revenue per AI company and other kinds of big impacts.

Hello. Hello. It is Ashlee Vance back here with y'all for another episode of the Core Memory Podcast. In this episode, the headline is that the podcaster becomes the podcastee. I sat down for about an hour with Dwarkesh Patel. Has a massive YouTube show and podcast where he gets into deep— he goes deep on science, history, you name it, tech, everything. And, and he's a great, great chat. We went to Dwarkesh's studio in San Francisco, so I really turned the tables. On him properly. Um, and there's more headline news, which is that he has a book coming out. It's a, it's an oral history of artificial intelligence technology. Obviously we get into that at length in the show. We get into his background, everything you could ever want, um, from a podcast on Darkest. And per usual, do us a favor, give us a review, give us a like wherever you get your podcast. I promise I'll stop harassing you about this once we have asserted our position as one of the giant podcasts of the universe. But until then, we need your help. Um, anyway, come check this out. Come hang out with us in Tworkesh's studio. Here we go. Tworkesh, how are you?

I'm good, Ashlee. Good to see you.

The tables have turned. You are the interviewer, but I'm in your, I'm in your lair. I'm on the home turf.

We should have actually turned the tables because this is usually the side side I sat on, and I actually should have gone on the other side.

I was wondering, I was wondering which seat I was in, but it's better if you can see what's going on. Um, I'm so excited to talk to you. There's a whole bunch of things I wanted to do. Um, one that we'll start with in a sec is, is you have a book coming out, um, which is fantastic. I feel like you and I've known each other a bit, but I'm, you know, there's not a ton about you on the internet biographically, I guess, because Partly you're so young and, and, but you've been on this rocket ship. And so I wanted to, I thought it was a good chance for people to get to know you a little bit. And then, and then obviously I've got some AI and other questions.

Yes.

Excellent. Um, well, as an author, I wanted to plug your book first while everybody's listening for sure, just because, um, these things are hard to do. And, um, and so it's called The Scaling Era: An Oral History of AI. And it comes out on the—

March 25th.

March 25th. And it's, it's, is it digital at first?

And then it's digital at first. And then July we'll have the print version. Okay. And it's, um, over these last few years I've been doing these interviews with people in AI, the CEOs of the big labs, researchers, people doing everything from low-level research to, you know, the hyper— like Zuckerbergs and Satyas Nadellas. And, um, people, there's like all kinds of different insights that feed into understanding of what's gonna happen in AI from just basically like how much does Satya want to invest too, um, to like, what do primatologists think about what happened between chimps to humans? Yeah. And so there's insights from anthropology, there's insights from computer science and everything in between, uh, that I've gotten a chance to talk to people about over the last few years. And with Stripe Press, I got a chance to put all the best parts from all those conversations into one book, um, added with commentary and footnotes and other insights and useful charts. And even for me, seeing all of that together was actually just super helpful because you forget about how all these different ideas connect together. And so just having this sort of elevated, um, collection of the different interviews, uh, yeah, I'm super pleased with how it turned out.

No, it's awesome. I got a sneak peek. Um, I was, I only got the sneak peek a couple of days ago, so I have read a lot of it, but, and I, and you know, it's an oral history. So I had, I'd heard a lot of the podcast and then it's fat as a, someone who's done some oral histories. It's hard to like stitch these together. So I have questions about that, but just so people know, it's, it's, it's amazing. Is this, you know, you had this. Brilliant. I give you all the credit in the world for seizing on this moment, this fascinating moment in time, in technology, quite likely in human history. And, you know, you call it about the scaling era. I mean, but you, you really capture this, this birth of this moment through, through many of the most important figures who were doing it. So it's, it's, it's really cool to see.

Thank you. Thank you. I mean, it's one of those things where you, I don't think afterwards you can get the sense of how people were thinking about it. At the time, especially if AI is as big a deal as I expect it to be. Maybe you expect it to be. We'll talk about what you expect it to be. Like, how were people thinking about it in this crucial period where we're getting this early progress, but it wasn't clear what the next step was? And yeah, anyways, it's super cool to see, and I'm excited for people to see it as well.

Did you— okay, so my favorite oral history ever is West of Eden by Jill Stein. I don't know if you've ever—

no, so she— no, never mind.

This is West of Eden. She does, uh, she— it's like an oral history of the birth and rise of Los Angeles told through, um, kind of like a movie family, uh, uh, rancher, you know, as, as LA is changing from this, this kind of pastoral unknown place to this booming city. Um, and it's fascinating. She's, she's brilliant. The, it's an art form though, to, to shape these. I mean, you had, you had, I'm sure it was probably like millions of words that then you have to, you have to figure out how you're going to arrange it. Yeah. Just as a writer, I was curious if you— about what the process was like, like how you found it. What if you had— I don't know if you discovered anything magical about putting together an oral history.

I'd be curious. Yeah, let's see. The— and first of all, I should really give credit to my co-author Gavin and our editor Rebecca for helping us put together these different segments. I think what was just really cool is so the first chapter is about scaling, right? And you can see you can talk to Dario and you can ask him like, why does scaling work, right? This— what is it about the universe that if you throw bigger blobs of compute at training with more data, you just get something emergent on the other side? And then you can ask somebody who's like, um, who studied how the brain works and what circuits in the brain map onto, um, uh, how attention works. And then he has a perspective on why scaling works. And then you talk to somebody who's been studying like how primates— what exactly changed between primate brains to human brains? And he has a perspective on what, why scaling works. And you just do that from chapter after chapter, from economics to, um, to interpretability to whatever. So just seeing all these snippets next to each other was a really fun experience. And compiling that together was, um, yeah, it was a fun time.

Yeah, no, it's, uh, it's a fascinating read. All right. So we've done the plug.

Yeah.

I love the book. I feel like it's important. Um, and well, you know, I was so curious. You had, I don't think I was following you from the very earliest days, but I heard about the Lunar Society and things early on. Everyone's like, oh, there's this, there's this this really smart kid, he's doing a podcast, he's interviewing interesting things. When I got to know you a little bit, and I think it comes through on the podcast a bit, I don't know, you have this old soul vibe, you know? I don't know if it's just the beard or, are you still 23 or 24 now?

24, yeah.

24, okay, so you know.

Time flies by.

Yeah, but both you and I would say even Leopold, who's in the book and who you talk to a lot. I don't know, you guys exude this, kind of wise old soul in a young person's body. I don't even know if you agree with that. I was curious if you've always, if you do agree with that, if you've always like felt like that a little bit.

Yeah. Yeah. I mean, the beard gives me an extra 10 years or something.

I saw these videos of you doing debates in high school. Oh yeah. You got this like chiseled chin. I don't know, man. I think you should, you should go back to that at some point.

That's right. Yeah. I should do like some sort of fundraiser. If we raise $1 million, I'll shave off my beard for something. Thing. Have I always been old?

Well, you know, you have a certain— you've got like a— not to like blow smoke or anything, but you've got like a gravitas. You got to come at— obviously you do a lot of history podcasts, but you kind of come at it from this, um, yeah, I would say this like wise— this like unfair amount of wisdom for a 24-year-old sometimes to have.

Yeah, I mean, I, I will say that I have the good opportunity to ask the questions, and I can just talk to people who have spent a lifetime studying the subject. So I'm under no illusions of having some deep understanding of every single topic I've interviewed people about. But the good thing, and I'm so glad that I got into this career, I mean, I studied computer science and so this was the thing I was doing on the side, but I'm really glad that the way I got into it was with podcasting because it's— I mean, talk about AI, right? These models, you make them better first by imitating somebody else and then you do RL on top of them. Similarly, I feel like I got a similar experience where I interview people. I— you don't really need to understand the whole topic in order to interview somebody. You can just be curious about it. You can read the book, you can have questions about it. It's a great way to give yourself an excuse to read more, to build up your domain knowledge and so forth. And honestly, it was in retrospect such good training to be able to do that. And then, yeah, I guess I, I'm 24, but also I've been doing this for 5 years and so So over time, as I've interviewed different people, you just pick up a lot of different perspectives. You pick up a lot of different contexts. And the other thing is I like really believe in the hive mind of like, once you get into this community, and as I'm sure this has been your experience as well, you keep meeting all these people who teach you more things. And honestly, what I learn offline as a result of producing good podcasts, which leads to me meeting more interesting people who teach me more things, which leads to better podcasts, that is the main feedback loop. The show growing on YouTube or whatever, that is not the main feedback loop. The exponential growth happens on the back end.

Yeah, I mean, it's like favorite part about my job is just you're constantly learning from very fascinating people and being immersed in it.

Um, but, you know, by the way, just as a side note, we were— I don't know if this is breaking Chatham House rules or something. Um, if it is, we can cut it out. But we were at a conference together and people were doing like 3-word intros, and then yours was, "Journalists allowed here." But I feel like that's been part of your edge, right? That you just like know the top people in technology who can teach you what's happening.

It's a weird thing because there's always this debate about like access journalism and things like that. And like, everyone's always like, oh, you're optimistic about tech, which I think I am in my heart. I, you know, but I, I don't know, at least in our documentaries and our TV show and in my writing, you know, I try to tell, I try to like tell both sides or if there are two sides, there aren't always on some things. But I don't know. I just try to do enough reporting where people can make up their own mind. I think like if I have an advantage or something, it's just, I, I'm just, I'm not a super judgmental person. I'm interested in interesting people living interesting lives. And like, you can decide if you think they're, it's like if I'm doing a documentary on Brian Johnson, it's like, you can make up your own mind about this guy, but this is what it looks like. You know? So, but you remember that you do have like a, your memory is, is, is, you seem to have like an encyclopedic type memory. I was curious, you know, like when you do the history podcast, In particular, I was always impressed that you, you're conversant on quite, um, in-depth with these people and, and on the subject matter. I, and I know you do tons of research going into these podcasts. I, so I was wondering if this comes from you being extremely well-read and widely read, or if, if it's these pockets as you're heading into an interview that you deep dive on. I know before this podcast, I was texting you about a couple of my favorite people, like Emma Goldman and and Mayakovsky, and you were like, oh, I, you know, I don't know, I don't know those two particular people. Um, so yeah, I was just kind of curious, growing up, were you reading, were you reading tons of fiction, fiction and nonfiction? Were you reading tons of history, or is this something you've come to a bit later as you're doing these podcasts?

I think mostly it's come through the podcast. And look, I mean, I started the podcast when I was 19, uh, and so Over time, I started the podcast. I was mostly interested in economics and history, and my first guest was Bryan Caplan, and after that, Kyler Cowen. And so that was the main thing I was interested in for a while. And AI almost came to— not begrudgingly, but quite late in the career of the podcast because I just thought, look, this is important, so I should cover it. Now, what I've realized is AI is one of those fields where in order to think about it, you need all these different lenses. If you want to think about what will— what will the world look like in 2040, it's not just enough to say, I expect the benchmarks in AI to saturate in this particular way. You know, I debate with— you had a debate with The Economist about like, will there be explosive growth? Will there be 30% growth? And then Tyler Cowen has all these different arguments that we had at this debate at the Progress Studies Conference that I published on my podcast. But that's like an economics debate, right? That's a— that's not even just an economics. That's a history debate. Like, what has happened? What exactly happened in the Industrial Revolution? How do we best understand that? And so the thing about AI has been that, like, having done these different kinds of interviews where you talk to different kinds of people, you can always ask them about AI and you can always learn something about them from— ultimately, we're trying to predict what this future society will look like, and there's no field that is not relevant to it. So when I interview, like, David Reich, who's a geneticist of ancient DNA, I can ask him about, like, you know, what exactly happened with the Yamnaya? Why were they able to take over so fast? And what does that teach us? Um, but yeah, I mean, that's why it's been fun to talk about AI.

Well, that's a great example because that particular episode was fascinating. But so that's something where you've read, you've done your research for that interview, but you were not— because your knowledge on some of that was impressive. I knew, but you were not, you were not doing like super deep on biology in high school and college. That, that's something really where you gathered a lot of that knowledge for the particular show.

That's right. Yeah.

Yeah. Okay.

Okay. Just, just in time compilation. Okay.

Um, I was curious. Yeah.

I, yeah, maybe I, I really think that it's one, at least a level. Look, I understand I'm not like going in the lab and myself doing research, right? Like, I know there's different levels to this, but for the level of knowledge where you can explain the main arguments, you can explain the evidence for it, you can understand how to evaluate that evidence. I think people would just like underestimate themselves. Like, this is something just a normal person can do and is a more interesting way to engage with intellectual content. So yeah, for that one, I mean, I don't know, as far as my process goes, I have gotten in the habit ever since I interviewed Andy Matuszak of making spaced repetition cards for everything. I have other things in my process, like, fundamentally not that complicated. I have a big Google Doc where I write up my notes and questions, as I'm sure you do as well.

Could you a bit— I mean, I've seen you write about your process. I actually think it's fascinating. I, because I've been doing this a long time and I oscillate between doing tons of research and then sometimes it's actually it's beneficial to go in a little bit blind sometimes. It's funny. I mean, there's, there's like different ways at this depending on who you're talking to.

Yeah. Yeah.

When you— okay, so you, you were born in India, right? And about 8 years old you moved to the States. That's right. And your dad's a doctor.

That's right.

Did your, did your mom work?

I didn't even know this was on the internet.

I found it.

I found it. Deep research gets you far, huh? The— sorry, the question was about my mom.

Yeah. Yeah. I just didn't know if— I saw that your dad was a doctor. I didn't see anything.

She's a tax preparer. Okay. And she works during tax season. But yeah, mainly it's my dad who's the breadwinner.

And you get shuttled off into like rural America. I mean, you were in like North Dakota, Maryland. It sounded like your dad was getting hired at either different hospitals or— Yeah, I don't know what kind of doctor.

So he was on an H-1B. And so you have to show that you can't get domestic help at the same place. So that's why, you know, North Dakota, they can say, oh, we're having trouble hiring here. And that's why we're hiring. Or similarly, then we went to West Virginia for 3 years after 3 years in North Dakota, then Maryland, then West Texas, where they've been. And so the selection of like, we can't hire somebody to want to live here, selects for these rural areas, which is the H-1B program. And yeah, so then I just like, I guess it was a bit of a culture shock from India to North Dakota.

Yeah, I was curious. Yeah.

But then we moved around a bunch and You know, I guess in the grand scheme of things, I don't really know because, yeah, I mean, how did it impact me or what was it like? I think it was— I mean, one thing I realized is just like a lot of places are quite similar. Like there's not that much of a difference between North Dakota and West Virginia growing up there. Yeah, I'm not sure. I don't know if I have any takeaways from that period other than the fact that we just moved around a lot.

Yeah, I mean, I moved around a lot.

Oh really?

I mean, yeah, it has like a bit of an influence on you, I think.

Where all did you go?

I mean, I was born in South Africa and then I did Connecticut, Houston, Midland, Texas, which really could not be too far. You were in San Angelo, I think.

That's right.

Yeah. Yeah. Not, I mean, Texas is big, so kind of far, but not too far.

And we had debate tournaments in Midland.

I was, I'm going to ask you about it. And then, then I went to New Jersey for a while, then back then to Houston. So yeah, just a bit. That was like a 10-year span. Yeah.

So, so, and my mom's Australian, so we went And was your dad's work or mom's work?

It was my dad. My dad was— this is the truth. People can judge it as they will. He was a lawyer for Exxon doing kind of land deals. So we got moved around a lot. And he had met my mom in Australia. There's quite a scandalous story how they ended up in South Africa. They were both married before and kind of ran away together. And then, yeah, so I traveled a lot as a kid. And, but that part of Texas is, is, I don't know, is San Angelo kind of like middle? I mean, it's middle. Yeah. Yeah. I mean, it is like West Texas, that, that part is, is its own, own country. It's its own culture. Um, middle of nowhere, tumbleweeds, um, rodeo. Yeah.

Yeah.

So that was all, I mean, it sounds like you already had your culture shocks by the time you got to high school, but that's, that's like its own thing.

Yeah. I mean, I will say the, um, there is an element of sort of universal culture. Where I do feel like high school everywhere is kind of similar. You know, it has more to do with generation than with where you are. And I think young people are kind of similar everywhere and older people are kind of similar in a bunch of places. Maybe urban-rural is a notable divide. But other than that, I think, yeah, because I mean, like Maryland to West Texas was just like not that different. Same with like West Virginia versus North Dakota.

Were you doing Friday Night Lights or you were doing Friday night debate prep then?

More the debate prep. Yeah.

So I saw these videos, man. I mean, I don't know what to make of some of the comments on there, but it seemed like you did well at debate. And I saw people, there was one where it's, it's like he had 30 minutes to prepare for this. That was it. And yeah, you deliver. I don't know if it was 30 minutes to prepare or it was a 30-minute speech. I'm trying to remember.

Yeah.

Extent you delivered this like soliloquy that was, you know, just, just spot on and so concise that I saw the people in the comments marveling. And so were you kind of a debate star.

I was okay at it. The event is extemp where you get 30 minutes to prepare a speech on a topic you get chosen randomly. I was actually— I watched it like a year or two ago again, and I was like thinking I've actually gotten worse at public speaking. Like when I do these podcasts, I'm like stuttering all the time and using filler words. And that speech I wasn't. So in some ways I've degraded from then. But yeah, I would in retrospect, there's like all these influences you kind of forget about and I never think about debate that much, but you know, looking back on it, I'm really grateful to my debate coach, Julie Schneers, for giving us that basic training, free training.

Is that the main thing you did outside of school? Or I guess through school, but like your main—

Yeah, I think so. Yeah.

And then there was YouTube and Netflix and whatever, but just consuming, not, not producing yet.

Oh, yeah. Yeah.

Okay. And then you're a computer science major at University of Texas, and in the middle of it, very near at some point in the COVID part, you start the podcast, but did you— and I know I've seen you talk that you, you thought you might do a startup or something at some point, but, but like in the bigger picture, I mean, whether it's high school or as you're getting into college, did you have a thought of what you were going to— what you were going to do?

I mean, genuinely, like all startup, my— I got hooked up with this roommate my freshman or sophomore year who was super into Paul Graham essays, and he introduced me to them. And then I got super into this. And we would talk about it all the time and we're trying out these ridiculous startup ideas. Like he had one where we would buy your— instead of you throwing out your old furniture, we would buy it off you and then store it somewhere and resell it or something. And we spent like a couple weekends. I was just helping him out. We would U-Haul this furniture around. And then I had one. Actually, I had this— I made this website where I was really into David Deutsch's ideas. And he's into this idea of criticism and refutation as this is how knowledge grows. And he's into this philosopher Karl Popper who laid this out initially. So I made this website. I'm like, okay, this will be a cool thing to do. Like there should be a website where the content has these primitives of like coming up with questions and then refuting them and blah, blah, blah. So let me make a forum that does this. And I said, well, you know, it's influenced by the ideas of Karl Popper, so I'm going to call it popperplay.com. I didn't know what poppers were back then. I know like all these people, I mean, at the time I didn't know anybody, so I would just like cold DM people who are in this community like, hey, like Benzedrine and the— yeah, I mean, I think poppers is like what, you know, like loosens up your, loosens up your sphincter, right? So it's like used for—

there's probably many kinds of poppers.

So anyways, I would invite, I would invite cold email, cold DM people who are in this community like, hey, I made this website, I'm inspired by your ideas. It's called Pofferflake.

Um, pretty god interesting, yeah, inbound.

Yeah, that's right. Anyways, I was just doing stuff like that. And then as a student, I was sort of like, I was like middle of the pack basically. I wasn't like, you know, in computer science itself, I wasn't like killing it or something, but I wasn't terribly falling behind either. But this is, the podcast was another thing I was also doing on the side and that, that's the one that took off.

Yeah. Yeah. And the, the, I'm actually not going to do the podcast origin story because there is stuff on the internet about that. So we're going to— great. We're going to go for new ground. What did your— what did your parents say when you told them you were going to become a podcaster instead of a computer scientist?

I never told them that that was my career plan because it was not my career plan. I told them that I'm taking— because I graduated a little bit early, so I'm like, I'm going to use this as a gap year opportunity to— I'm not sure what I want to do. With this podcast, which is the thing I'm doing on the side, is giving me the opportunity to meet all these cool people, and that will help me figure out what I want to do. And so, and that's also what I genuinely thought of what I was doing. It wasn't until quite recently, like 2 years ago maybe, that I decided this is my career. I don't think they were super psyched about it, but I guess they didn't stop me, which is just to their credit.

But you didn't get a talking to?

I must have. They wanted me to get a master's or something. They kept insisting. Yeah.

Um, you, the, you cover a lot of different subjects in the podcast. I mean, we've already talked about some of everything from history to biology to AI, you know, it's sort of, um, warfare, um, experts in different domains. Obviously you, you gravitated a big chunk towards AI and I, I know you were using the tools and the software yourself. Um, Is that something that's a bit closer to your deepest interest, or this is just this particular moment and you're in San Francisco and, and seize upon it? Um, yeah, I didn't know where it stack ranked in your, your, um, you've obviously, it strikes me that you've gotten deeper on AI than just about anything else. And so you must be, you must be genuinely just drawn to it.

Yeah. I mean, I believe that there's going to be human-level AIs. I believe that they're going to have advantages over humans. Um, I believe there's going to be a population of billions of them. And maybe more, and then buy the craziest versions of like 20% growth, maybe 200% growth. And so if you really believe that's happening in the next decade or two, then it's hard to think about anything else, or at least not come back to this topic again and again. So fundamentally, it's the impact that's driving it, or, you know, like this is the thing that's happening right now, you know.

And usually you're— I see you give your thoughts on some of these things through your questions. Yeah, but this is like my chance. Okay, we're going to go AI land for a minute because I feel like you're usually asking the other people their thoughts. And so, um, all right. You know, I mean, there's, I was sitting, I think in this very office interviewing Leopold for something.

Oh nice.

I was like, this book I'm working on. And he was basically like, well, you better get it out by the end of this year because otherwise everything is pointless.

Yeah. Yeah.

You might not even, there might not even be books. There might not be people writing books in, in 18 months. Um, yeah. So, so yeah, I'm kind of curious where you are on the, you, you just laid out a bit of it, but Clearly over the last, even just the last few months, it feels like people have started, whether it's folks from OpenAI or Anthropic or broader, saying by the end of this year, something very dramatic happens, at least by next year. Yeah, I mean, I don't know, it's a broad question, but what's your sense on the timeline of what you see? What do you see over the next 18 months, say?

18 months. I expect them to keep becoming, just this morning, there's a really interesting result from Meter that the horizon length over which, like, how long can an AI do a task? Is it like 7 seconds and it loses its train of thought? Is it 14 seconds? And it seems to be doubling every 4, 7 months. So before, they're going to be writing a little function. Now they can write you a whole package of code, a whole file of code. Now they can maybe write you many different lines of code and keep engaging with you over a session as you say, no, I want to add this feature or something. So if that keeps doubling, then you have this really exponential process where you have this discontinuous impact. Going from 30 minutes to an hour is a big deal. Going from an hour to 2 hours is a big deal. And going from a day to 2 days is an incredibly big deal, right? The amount of increased scope you can have as a result will increase exponentially. So I expect these kinds of discontinuities. I am still confused about why these models have no traction whatsoever on things which are short-horizon tasks. It doesn't take you that long to, let's say, edit a video or book a flight or start playing a new video game, which would you've never seen in your trading data. And there they have very little traction whatsoever. Or organize the logistics for like a happy hour or something, right? So, um, the big question— it seems like free training is exhausted. The, um, reasoning paradigm— the big question is how far that gets you. And so we had O1, O2, I guess we skipped O2, and now we went to O3. Um, the next question is like, okay, look, they've conclusively shown that this gets you a ton of traction on math and code. What are the other kinds of RL environments you can spin up where you get this thing where you're like, now you're doing tasks that are like a human's day-long worth of input? And I am going to be— it's even hard to forecast, or it's not clear what the point of forecasting is. I think in 3 months we will know if— do we have models which can reliably do computer use? And can you set up an RL environment for that? If not, then I think the next, like, we're not going to have some huge economic impact from AI that we haven't seen so far in the next 18 months because fundamentally, like, there's not that much in the economy that's just like chat-based. But if you do have that, then I think you can start seeing tens of billions of dollars of revenue per AI company and other kinds of big impacts.

And if we don't have that, you're saying we've hit some kind of wall that needs to be another breakthrough.

That's right.

To get somewhere new.

That's right.

Yeah.

And it might come pretty soon. Like the reasoning thing was like, you know, it's not been that long since GPT-4 that they came up with the reasoning.

It's so funny to me because like all these guys, I know, I mean, inside the labs, they sort of know this stuff, but it is funny that right before reasoning came out, it was like, we're fucked. ChatGPT-5 is not going as well as I hoped. We've hit this wall. Like, see, the data's running out. And then it's like, well, and then everybody shows up with reasoning. And then now the debate begins completely.

Goalpost moving.

Yeah.

Though I think it's super striking how there's basically no secrets where if you remember when the Ilya-Sam board drama happened, you have core memory. I've been around that. Maybe producing some interesting content, right? The, um, they were talking about like the reason Sam got fired was because of Q* and this is what people saw, right? I don't know if that's why he got fired. Probably not. But the fact that like they called Q*, like what they said Q* was, was in fact just like reasoning and it led to O1 a year later. So you're just like paying attention to the rumors on Twitter. It's like And then the pre-training is dead. It was like, it's kind of basically right, right? Like, if you could have predicted that before 4.5 was released just from seeing the rumors on Twitter. Honestly, just thinking about that over the last week, I've been thinking maybe I should just like buy the rumors more, just like, just like default to what you hear on Twitter.

It is a crazy field where like, they're, like you said, there's largely no secrets. Yeah, it's still a relatively small group of people, all of whom know each other. And even if they're Anthropic split from OpenAI, all this stuff. It's like everybody's still— not everybody, but there's still lots of friendships. There's, you know, there's a lot of overlap. It's a, it's a— I'm a nerd about Silicon Valley history and semiconductors were the first thing I got really interested in. And it's probably not apples to apples, but I keep finding myself going back and, and trying to put myself in these shoes where, um, You can feel this new industry is being born. It's quite R&D heavy to some degree. And yet, like, even now, if you are a pretty intelligent person, you could learn everything you need to know about AI to, like, do something pretty incredible, possibly. And, and it's— that's wild to me because in some ways this field's been around for 50, 60 years. In some ways it's only been around for what feels like 6. Yeah. And you could learn the 6th part pretty quick.

And in fact, you might be in a better position to do it if you started now.

Yeah.

Because I mean, I remember when GPT-3.5 came out and it felt then, because at the time I was still thinking I'm gonna do a startup or something. I remember then it feeling like, oh, it's too late. Like, it's, the industry is kicking, it's cooking. Um, you can't do anything. And so, yeah, I mean, I think like we'll feel the same way about AI in a couple of years, especially if the next bottleneck has to be qualitatively different, or the next jump has to be qualitatively different. It's got to be the computer use stuff, it's got to be the other capabilities. Then you can just like, if you have a startup, you start now, which is dedicated to those capabilities. If you look at when Cursor hit product market fit, it was only after 3.5 came about, right? Which is like fundamentally have a model which enables this kind of agentic coding that wasn't possible before. So I would, yeah, I would be thinking about like, I think you can start now and still be aiming for those future capabilities.

And it's bizarre because it's like something that at times feels magical. It feels, like I mentioned, R&D heavy. Like this is cutting-edge science. Then there's these brutish, like very just pragmatic, boring aspects to it to some degree. It's like you look at Grok, it's like, well, if you assemble 12 to 20 extremely smart people, have enough money and GPUs, you can leap to like, you know, just in what, 18 months?

Yeah.

And so sometimes when I go back to this, semiconductors, I, I picture, um, you had Shockley kind of just birthing the stuff and then Fairchild comes and does the integrated circuit and, and TI does the integrated circuit almost at, at the same time. And then you have this like explosion of, of the Fairchild, like all these, these semiconductor startups and you had this sharing of information. It was weird. You were at the cutting edge. People are chasing, they're in a race. They're, they're trying to win everything, but it's such a hard field. They're engineers and scientists by nature that they're going to the bar and they're like, well, How did you solve this problem? And everybody's sharing. So you're in this collective race. And this is, and then I just wonder at some point things do get where it's like so far along that then you can't just catch up. And so you can't know everything. It's things start to get so specific. You have to be a fantastic physicist or chemist or, you know, and you have this isolated part of the puzzle. Yeah. And it still feels like we're possibly like years from—

that's right.

I don't— at least to me. Yeah.

I mean, if anything, the labs' approaches seem to be coalescing. They literally call their products the exact same thing, right? Everybody has a deep research. And it's interesting the extent to which just knowing a thing is possible is— because there was this big thing, Leopold talked about this, and other people talked about this, we got to protect our secrets. Um, and it seemed like the most relevant thing for DeepSeq was just the knowledge that that a certain, a certain kind of thing is possible. So reasoning is possible, and you can learn a lot just from a capability being deployed. What is the throughput? How— what is the latency? These things you just literally can see from using the API, just give you a ton of information on how something was built. Um, and it's like a 4-minute mile kind of effect where like, you know, it's possible, then you, then you can just like, it just gets demolished.

Yeah.

So, um, at least so far it seems the labs are coalescing, and the big, big question is is by the point at which you have automated engineers, automated researchers, does that give the first lab that's a couple months ahead this huge lead where they get this 10x, 100x edge over their competitors? There's a bunch of reasons why I'm skeptical, which we can talk about, but I think that honestly, that's like, that's the big question the labs are probably asking themselves is just like, how are we putting ourselves in a good position for this sort of intelligence explosion?

Let's do that for a second. So I think your audience probably gets what you just, you were talking about. I feel like the general public does not understand this maybe on some level about this race aspect to this. You know, there is a belief among at least some people in this arena, exactly what you said. It's like, from the outside, it looks like they're all so close and it's like somebody does a breakthrough, then the next person does, the next company does the next one shortly thereafter. It's like, what's the point of all this? But there is this belief that if somebody really, really hits whatever, however you want to define superintelligence, that you in that moment, you're kind of hitting this singularity point where you are getting such exponential gains over whoever's in second place, the second place never catches up. Right. And then when I hear that, I am skeptical because I'm like, I get what you're saying, but like, it's hard for me to imagine that that curve necessarily holds that way. And that— so you're also, you're also a bit, it sounds like, question whether that's how it would play out.

Yeah, the, um, and maybe I'll explain that if I— no, no, you're explaining it well. Maybe the fact that you're skeptical means that I should try to steelman that position. Let me try to steelman. So, um, here's the reason, as you were saying, look, like xAI has 20 people and it was able to make all these breakthroughs and catch up. Well, I think that if they had the, you know, Google DeepMind having 2,000 people is what you need, or 20,000 or whatever. And if that was like, it was just like you needed more super intelligent people, because the whole thing AI is going to give you is you're just going to have so many of them in parallel each of them can try out a different experiment, maybe a different training method, a different optimizer, whatever. And what I hear from my friends in the labs is they're really bottlenecked on compute for experiments. And so xAI had a lot of compute because of Elon's super fast build. Maybe that's why they were able to catch up. And this is what's really bottlenecking researchers, is you can't just, you know, one-shot this in your head. You got to have some compute for experiments. There's like two dynamics that I think the labs probably had to think really hard about. One is that the cost of training these models is declining like 10x a year.. So, um, and there's a dynamic where it's really easy to fast follow because if you know capability is possible, you can copy it and you can do it for much cheaper. You can, um, uh, the other dynamic is that hopefully, potentially, these models are multipliers on your existing researchers, and maybe they can like totally automate the process of researching future algorithmic breakthroughs, maybe optimizers or architectures or whatever it is. The big question there is, is that a thing? If you just like automate it, you just get 1,000x or more of it, right? As you were just mentioning, xAI has these 20 engineers and they were able to catch up pretty fast. Why didn't they need 2,000 engineers or 20,000 engineers or something? Why isn't OpenAI trying to hire every single math PhD at Harvard and trying to turn them into AI researchers if they're just like more of these people is what matters? And so that's why I'm skeptical that you can just have a million parallel agents who are each running, you know, these ML experiments and that gives you this software-only singularity. The thing that I do still think you get this enormous impact from— maybe these AIs are helping you with chip design, and there's a lot of optimization you can do there just in software alone. Um, over the longer term, there's like, maybe you can actually improve the process of like, uh, physically manufacturing the chips, but that looks much more like the world economy is improving and growing. And if you were to ask yourself like, how would you produce the iPhone in the year 1900? It would just be like the whole economy has to— like, all your tech across many different industries has to be upgraded. It's not like you can just like make the iPhone. Um, and maybe we're in a similar position with the kinds of— maybe there's hardware breakthroughs or software breakthroughs that are needed for this God that we're trying to make, right? Where it will happen, but it won't just be because a company is like one month ahead and then three months later they've got God. Um, and I'm more inclined towards this view that it's like the iPhone where the, the the whole country, or sorry, the whole economy's tech stack has to be upgraded, um, rather than just like you make this one change that a million people in parallel can find.

That's what I, that's what I think. I think even if you had that lead and some things were, were rising exponentially for you across all those tools, I think you run into, you run into the physical world at some point and there's, there's still things that have to be solved. There's slowdowns. It's not like you just have this blindingly clear path to, to perfection on this.

Yeah.

This curve.

And yeah, and then, uh, there are some interesting implications from this model. So one is compute is going to be incredibly, incredibly valuable because if you're bottlenecked not by the agents but by the compute that they have to play around with and make, run these little mini ML experiments, then that compute is just like incredibly, incredibly valuable into producing more intelligence, which is itself like worth at least $60 trillion a year because that's how much we pay workers in the world. That's like how much labor commands of output. The other is I think we can use this early position we have where, look, we're going to have AGIs, but the first ones are just going to be ones that we've trained less efficiently. We haven't really optimized them for cost or for, or for speed. They're going to— maybe it's going to cost like $1,000 per token. It's like God spitting out a token at a time, just like slowly enunciating, right? Right? Uh, and it's going to say like, let's say it's like $100 billion or $500 billion to train. Um, the, the worst kind of world is one where you just have these incredibly cheap AIs that somebody can build in their basement that are just like running around and none of the world is ready for it. You know, they can hack into our systems, they can find zero, uh, zero-day exploits in cryptocurrencies, they can make these super weapons or whatever, um, exploit the stock market or something. And so I think we're in a good position if we have early on these inference-weighted AGIs which are super expensive massive and you, people will later on fast follow, but you use these to basically upgrade the tech stack of the entire world so it can kind of keep up. And then over time, I mean, you still have this fast period where it's like, if that is getting 10x cheaper a year, you're still in this crazy whirlwind period, but that's better than just getting the, somebody making ASI in their basement.

Yeah. Yeah. Okay. I'm going to change gears a little bit. No, no, I think I'm going to bring us back here. I'm looking for, uh, your philosophical thoughts on a couple things. So bear with me, 'cause this may feel like an unusual line of questioning. Do you live in a, like a shared house?

No, no, I don't.

Did you ever?

Yeah, for like a year when I moved here.

Was it kind of like AI heavy?

No, I mean, this became a thing where when people read profiles of me, they'd be like, the thing about Dwarkesh is he lives in this group house of AI people or whatever. It's like, it just was the cheapest rent I could find in SF.

I hadn't read that. That's not fully where I was going. I was just curious. Because I run across, I've, I'm not naive to the fact that shared houses have been like a thing, especially in California for decades and decades. And I've, I've written my last book, um, it was about space and it was, rose out of shared houses in a different era. Yeah. Um, but you know, I have, as I've been reporting this current book I'm doing on, on OpenAI and AI, there's shared houses everywhere. It's interesting to me that, that, um, there are some people that tended to live in like EA-heavy houses, rationalist-heavy houses, AI-heavy houses. Like, it just— I find it interesting about this moment in time, this kind of philosophy. But okay, so you were not in one of those. My question is, like, when I hear you doing your interviews, you're able to talk with, like, pretty much everybody in the field. But then do you identify— you know, there is a, there is a sort of, you can argue, an anthropic school of thought on this. There's an OpenAI school of thought on some of this stuff. Um, you know, is there, is there a camp that— I don't want to put you in an awkward position because you still have to talk I get to ask the questions. You know, is there, is there a, is there a camp you identify a little more with who's, who's maybe both their philosophy, whether it's safety or more broadly, who you feel like is approaching this in the best way?

I'm going to say something that might be surprising, but I think like we're in a good position with all the labs in the respect that like, look, I get that people have their concerns about particular CEOs or particular decisions. But compared to the counterfactual worlds, right? I think in a lot of counterfactual worlds in which we're building AGI, potentially the people who are building it, like AI safety is not even a coherent concept. The idea of misalignment is just not a thing people talk about at all. I think that we very well could have ended up in such a world, and we didn't. And all the labs at least pay lip service to this idea, which is important. And the reason I mentioned that is people very cavalierly say things like, oh, we should nationalize the labs or something because that way we'll prioritize safety. There won't be any race dynamics and the government will prioritize misalignment safety. I'm just like, yeah, appreciate what you have. I think you're just like, don't take for granted that we just ended up in this situation where the CEOs of all the labs say they care about misalignment and safety. I'm not saying that this is the— like, I'm not— my P-Doom is, let's say, like 10-20%. So this is not like the overwhelming—

that is where you put your P-Doom?

Okay. But look, I mean, yeah, it's just like, yeah, basically I think we're like in a pretty good position relative to where we might have been in another universe.

I've thought about this a lot. I mean, it's kind of funny. I agree with you on a lot of, a lot of fronts. And I'm not necessarily saying this is my opinion. This is something I've thought about. And you could frame it this way, which is that you've kind of got like Demis gets on this early. Um, sort of, this is not like fully a commentary on him, but you could imagine a world where he's, he's like into the, he's in the AI, he's assembled these researchers, like, you know, I think I'm a decent person. I, I can shepherd this technology into the world quite well, you know, and I'd be a good person to help oversee this. And then OpenAI appears and sort of, um, kind of throws a wrench in the plan and that now there's like a new player in the arena and maybe they actually get there first and now, now it's a race and, and now we kind of a new personality, you know, and that, that personality at the time is like this almost like four-headed monster between Elon and Sam and Greg and Ilya, you know, and it, um, and then as things have evolved, you know, you sort of got the— this is, this is like the, the, you know, 30,000-foot view, but you've got the— you've got Demis, you've got Sam, you've got Elon, maybe Zuck, um, um, you've got Dario, and, and you could make an argument that I think in each one of these persons' souls, they kind of think, The world would turn out best maybe if I'm the one who gets there first, because I got a pretty good handle on how to do this. But you're totally right. There's like a— and like each one of those people, whether it's lip service or not, is like aware of risk and talks about it and like does do some safety things. There is like a total world where you just have someone who just does not care about that at all. And they obtain $1 trillion worth of GPUs that are just like barreling forward. I mean, it is— I think each one of those people comes with a lot of baggage and, and, um, different motivations that are not fully clear, but at least they're like cognizant of, of the risks.

Yeah. And then this is really important because I think people, when they're saying we either have to have a pause or we need to nationalize them or something like that, what is the alternative? The alternative is either you slow down enough that China becomes in the lead. And I haven't heard exactly Lianwen Feng say much about safety. He might care about this, I actually don't know. But how much scope he has in his system to prioritize that over the considerations of the CCP is another question. And I'm like, I'm just like not that optimistic about China. And I think just to back up on safety, what does it mean, right? Like, I think fundamentally it's not like you have to do this galaxy-brain ELIAS or thing. I think a lot of it is just just like, if you pay lip service to it and just like do the basic minimum of, hey, don't let them talk in their own secret language, like try to understand what they're saying, maybe make them talk in human English that you can understand. By the way, this is not taken for granted, right? Like Meta released a paper where they explicitly had it talk in this like neural-ese, or just like, hey, let's set up this like 10% tax, which will be this like monitoring. We're going to use some of the compute to pay attention to what the AIs are saying. We're going to try to at least attempt to understand what's happening in their Minds, whatever. Anyways, so just like, we want them to just like— the lip service level is enough in many cases to avoid a lot of risks. Um, okay, so if it's nationalized or if it's like a big public-private partnership, what you'd worry about is, does the national security community care that much about these kinds of speculative risks? Uh, I doubt it. And then so I'm worried about them having more leverage. The alternative world I'm imagining is where you've got these different labs over time because training costs are getting so much, um, cheaper, and then they, you know, it's like easy to fast follow. Any one of them is not in a position to like take over the world, right? They're kind of getting commodified over time. And that especially because of the— I don't buy the software-only singularity that we just talked about. I think we'll be in a good position there. And then you just have this broad deployment where AIs are a little expensive at first. Over time they get cheaper and much better. The whole tech stack of the world keeps getting upgraded. And we're in a good position with defense because like we're slowly accumulating putting, um, these AIs into our society. And again, if it was nationalized, the thing I'd worry about is that government keeps certain capabilities secret, so you can't adequately prepare if it's not being deployed. Um, I think just like robustly bad. Uh, so anyways, that's the sort of vision I have where just like, yeah, just like let people develop this. Um, I do think there's risk, like I think like 10-20%, but I don't see an alternative where you can just get rid of that.

Yeah, and I, I know people hate— lots of people hate tech, lots of people hate tech billionaires. In the scenario that you were talking about on the government side, I'm still not sure like what magical person you're putting in charge that, like, knows all the right decisions to make, isn't felled by their own human nature. And, and, you know, it's like, I don't know, I feel like we're chasing some idealized person that doesn't exist to, like, oversee this, this whole process.

Yeah. And they often say, like, oh, well, there's— government has, like, lots of checks and balances. And I'm like, are these people who are, like, so confident about the checks and balances that will keep ASI that's nationalized in check and not let the president take over or something, are they just like, so, um, are they just super, uh, relaxed about what Doge is doing because they're like, oh, you know, the checks and balances or whatever? I mean, like, you may like what Doge is doing. I'm just saying the kinds of people who are like, oh, the government can run ASI are not the kind of people who are okay with what Doge is doing, right? They don't believe the checks and balances will stop Doge. So why do you think with ASI the president's going to be totally, um, totally checked and balanced?

Yeah, yep, yep. No, I agree. Um, okay, so writ large, okay, when I asked that question initially, I actually I, you could tell me if, please tell me if I'm wrong. I, I, I felt like you would be more in Anthropic camp. And I say this because I feel like they have, when I interview people from OpenAI and Anthropic, it's, it's weird. It's like not universally, but it's kind of like a different personality. I always felt like your vibe was, was, was Dario vibe. Um, do you think that's, it's even like a vague, terrible question, but, but, um, I don't know. Are you into I'll ask you a different way. I mean, do you think they are— you said security, maybe we overblow it, alignment and things like that, but do you think they have the best of all approaches so far? Do you think they genuinely take this more— the impression is they take this more seriously than others. Do you believe that?

Yeah, I think it's just I think it's just like unwise to cheer for a particular group or a particular person rather than the kinds of things you care about.

Yeah.

I think like, honestly, there's not that much difference between the labs. Like, they all work on safety research, they all work on capabilities research. I'm not saying it's bad that they're doing either of those things. I'm glad they are, actually. But I'm like, I don't think they're like that fundamentally different.

Even though like we saw this huge schism of safety people leaving OpenAI and then now them trying to kind of reboot that team, that didn't, that didn't strike you as significantly different or a departure for them.

Yeah. And to be clear, I don't know that much about the internals of these different companies, but I'm just going to be, to the extent that I have opinions of what people should do or questions for them, I'll keep asking them and keep, keep, you know, pitching for different things. But I, yeah, I think, for example, Tyler Cowen, who's an economist, he's just like whenever people say, who do you want to win the presidency or something, they'll often just say like, look, politicians are going to disappoint you. That's why I never cheer for any particular politician. I'll, I'll, I'll tell you about my worldview, but like, I don't think you get the right person in there and they're going to do everything right or something.

Yeah. Sorry, you mentioned Tyler. It was actually somebody I wanted to ask you about. I mean, not that he needs help. Lots of people know who he is, but like such an influential, like underestimated, influential person in in the world. It's so interesting because he didn't— he kind of helped.

Yeah, that's right.

That's right. Yeah.

He's— I mean, the thing is, he's so generous with his advice, mentorship, time. He didn't have to come in. When I started my podcast, him, Bryan Kaplan, these early people didn't have to come on. There's— in fact, when I get cold emails from people who are like, can you come on my podcast where it's a totally new one, many times I say no. Right. And then they're obviously way more prominent than that. And they said yes. This. And, um, he, he's done that not just for me, for so many other people, like literally hundreds of people, yeah, given the mentorships or funding with EV. Um, and it's the kind of thing where you don't like learn publicly about every single instance of it, but the cumulative impact and what people are able to do. And there's also the— just like the fact of his example, the fact that somebody can just try to read everything, try to understand a bunch of different topics, add commentary on it, host the podcast. Every time I listen to this conversation with Tyler, which is his podcast, I think about like, oh, this is where I learned it. This is where the inspiration came from. Yeah.

He's a, he's a super interesting guy. I didn't mean to pull us away from AI talk. I actually, I did have a thought coming into this. We could like set like a podcasting record and I should torture you and take this for like 9 hours and we could, we could outdo Lex. I have to say I enjoy your podcast much more than his. The, the, so there's just a couple last questions. If people want to hear more of your thoughts on AI, I think there's lots out there and they should, they should check out the book if they haven't, you know, um, they should check out the book. Um, there's, there's somebody else I wanted to ask you about, which is Sarah Payne, who you, um, this is like fascinating character. I mean, you know, she works at the Naval War College. Yeah. Name, right? I mean, you know, an academic, people who don't usually get this kind of stage. I know you know, when you did the— had her on your platform, um, the Naval War College was like, we haven't had this, this sort of reaction in ages. But man, I mean, it's amazing that you guys seem to have this very unique relationship. It's like kind of crazy. She's so fascinating. Um, this is just like somebody that you just— you happen to like run across a talk she gave, or—

I know, I, I asked on Twitter, who should I interview about World War II? You know, as you were saying, I'm sort of like an old soul, so the kind of hobbies I have when I'm 50 or I'm like reading about World War II.

Yeah, yeah, yeah, that's early. You're 20 years early.

That's right. Yeah. Um, and anyway, so somebody recommended her. I didn't know that much about her. The podcast at the time was kind of small, so I'm just like, whatever, somebody recommends a name. It was Tanya Greer. I'll try them out. So I interviewed her and the response has just been— if you look on sort of like a viewer weighted minute, um, uh, adjusted basis of like Um, I, I host basically the Sarah Payne Podcast where I occasionally talk about other topics like AI, right? Um, so anyway, so she's this military historian at the US Naval War College and she's been to all the archives, all the continents, has a deep understanding of so many different conflicts. Yeah. And then we did an episode together initially, and then now she— we've been doing these lecture series together where she will deeply explain a conflict. Like the Indo-Pakistani War or how Mao took over China or Japan's culture going into World War II and why they lost. And her ability to explain this is like so tremendous. The thing I didn't appreciate before the interviews is why did I find this person who hadn't— she was obviously well regarded in her field but wasn't publicly well known. And now she gets recognized in the street all the time and stuff. And it was like a big profile in Time recently, actually. And I think it was like she's in a generation where, like, genuinely— I mean, she'll never talk about this, but like genuinely misogyny was like a, was a hindrance to her career and like put a glass ceiling on her. And if you look on Amazon and you look up her books, the title is— the author title is S.C. Payne, not Sarah Payne.

She's kind of hiding it.

Yeah, exactly. But anyways, yeah, super grateful that I found her because she She's like, her coming on and agreeing to do the podcast with me has been a huge driver of the channel growing and also helping me understand a bunch of topics better. Um, and yeah, it is a sort of like really funny relationship that, uh, um, I mean, there's so many people like this, not so many actually, only a couple where I've interviewed them and the episode just like takes off and I have no explanation. There's nothing where I can give like this MrBeast type explanation of like why, oh, the thumbnail, the, like the red circles or whatever. Um, um, like I interviewed David Reich. Genetics of Ancient DNA. That's my second most popular episode. It's not Satya, it's not Zuckerberg. Um, and then Daniel Yergin, this like guy who wrote about the history of oil.

Yeah.

Um, yeah. Anyways, it just like really convinced me that like you really don't know what's going to be popular. Just interview people who you personally are interested in. They might be previously totally unknown, but they might take off.

Yeah. I mean, she does this, this, she walks this amazing line where she's so well-informed that she, she has a point of view. But she mutes it a bit. And then she's kind of presenting all available evidence to you in this very confident, like, you know that she knows her stuff, but, you know, she does have a point of view, but it's never coming across as like this, this kind of bias, this slant for one side or the other. It's very like, eases you in. And, and man, I mean, yeah, just a fascinating person. Do you— are you guys— you guys, do you, do you only chat when you're doing those moments, or are you now, are you fast friends and hang out?

I'd like to think we're friends. Uh, um, I mean, she lives in, uh, she lives on the East Coast, so we don't, we don't grab drinks or whatever. But, um, the, yeah, I mean, she's also super kind with her, um, with her thoughts. And she is somebody who comes from this old school, uh, uh, like really generous mentality of people— she's like super famous now, and people email her like, here's like sometimes hate email where they're like, here's why I think you're totally wrong about Ukraine. And she'll respond to each of them with this like long message about, um, and then reasoning out why she disagrees and everything to get each person individually. And then they come back like, oh, I didn't really think you'd actually read this and respond and everything. Um, uh, yeah. And, and also she's from this world where she understands fans, like, look, we're in SF and there's this, like, culture. I'm sure you notice this when you go to parties where people just, like, fucking say shit. Like, they have no idea. I mean, I'm like this. I'm like, fucking, I'm a podcaster. I'm just like, oh, you know, I've got these takes. And she's like, look, I spent like 10 years studying this one conflict and I've, like, literally read all the direct sources. I've got the goods. And that's the thing I'm going to talk about. That's the world she comes from.

We're all living in, like, hot take land.

Yeah, yeah.

She has, like, proper take.

Yeah, that's right.

Yeah. Well, I feel like there's some kind kind of buddy movie where the two of you go on a trip somewhere together and it's just this odd pairing. So these were some of the biggest questions I had. I mean, I could talk to you about AI forever. I think I'm gonna use my last couple questions just to, yeah, to flip the script, the table on you one more time. I mean, maybe I have a sense of it from things you've already said, said, but you said where your P-Doom is at. So yeah, that gives us some frame of reference. But, but net-net, you are— you're, you're optimist or pessimist?

I'm an optimist. I mean, I think we're gonna have, um, look, I, I think people say this in a corny way, and they say it in a way that's like maybe reflects their self-interest, but I think it's like basically true that we live in a world where because of scarcity, there's a bunch of things we don't get a chance to do. Like, um, uh, and it's not just like a scarcity of the basic goods, like poverty and, um, uh, like homes and whatever, but just like aesthetics. Like, how much energy can you devote to making things look good if you have any— if you just have 8 billion people? Or, um, what are the kind of peak experiences that you can explore What is the kinds of things your civilization as a whole can do? Can you go to the stars? Can you do all these cool things that you've written entire books about? And I think you'll be able to do more of that with AI. I think like we can explore things like we've never— in the long run, I'm a transhumanist. I think like the long-run way in which humanity stays in control or is a stakeholder in the future is that our intellectual descendants are in some ways integrated with AI or upgraded by AI, and they'll do all kinds of cool shit.

You're pro that, or you think it's inevitable? Inevitable?

Both.

Okay.

Yeah, yeah. And there's people who have this mentality where it's very much this zero-sum game over the future, where even though you've got trillions and trillions of stars, you still have to think from this perspective of like, what percentage are we getting for my value system versus your value system, um, or for the AIs versus for us? And I think it's just like a wrong way to think about things. I think you should have this— there's this enormous overhang where there's literally trillions of times more resources out there, there's trillions of times more technology we could discover and new ideas we could have. And every single person who has a different value system will have some representation in the future. There's enough scope and abundance for that to happen. And that, like, it's hard to know in advance what that will look like, but just this mentality of like, look, that's like, don't fuck it up basically, but otherwise it's going to be a good time.

And do you think one part that this is going to sound, I think, sort of like a stupid question because I'm not even sure I can convey what's in in my head fully, but I do get worried sometimes that the, this current world, this path that we're on, if this does take off, is being shaped by any, this is a too broad of a statement, but I think at its heart it's true. It's, it's, it's a very like mathematically minded group of people. It's a relatively small group of people, almost, almost, let's say 80% of the folks I've been interviewing, you know, they did math prodigies, math contests as a kid. It's not like they all share the same soul or something like that, but wired in a, in a, in a particular way. You know, and this is, this is the, this is what is shaping this intelligence that's being brought upon us. It is not, it's, it's not just a uniform person, but it's similar enough for me for Yeah, I just wonder about this sometimes. I mean, I don't really know how you solve this, but I feel like, I feel like this new universe that's being created is coming through like a narrow, a bit of a narrow scope.

When I look at the history of technology, I, it's just not clear that the values or motivations of the person inventing them really matter that much. It's more about the technology itself. Like, what were Gutenberg's politics? Did he want to kick off this big war between the Protestants and Catholics? I don't know if he did, but that's what it is like. Technology kind of ended up doing. The people who invented the first automobiles, did they want a bunch of teenagers having sex in them and kicking off the baby boomer generation or something in the back seats or something? Probably not. Just like a side effect. And maybe the kinds of people who hang out in SF should be scared about this because it's really hard for them to reliably instill their values, which they're working day and night to build into this machine god. It's going to be hard to get that carried over in the future. But I don't think they're going to like who exactly it is going to have that big a difference.

Yeah. Okay. All right. Last question. I mean, where do you see all this going for you? You're going to— you'll keep pursuing AI, you'll keep pursuing— I love, I love like the thesis of what you do, which is like, let's not dumb things down. Let's be smart and explore, explore different areas. And so, so my question's a little bit on areas of focus, but also You know, is it, is it you keep podcasting, you keep, where do you see this all going for you? Like the next, I don't know, call it 10 years.

Man, 10 years is such a long time.

I will allow 5.

I've been doing this, I guess, 4 or 5 years. So it's just basically what am I doing in the second half of this? I have ambitious targets about how fast I want the podcast to grow. I kind of want it to become the place where if you want to understand, I think the, like, AGI is a thing that will happen. I'm not saying, I don't think the intelligence explosion happens, but the explosive growth, this period where you have billions of extra people, you have all these kinds of craziness that causes, is a real thing. And I want to be the place where when it's happening or before it's happening and you want to understand what's hap— you know, how to make sense of what's going on, what are the different perspectives from all these different fields you can apply to it, this is the place you come to, to understand it. And long run, that's basically what I'm trying to do. That's, I think, going to involve more writing than I'm doing currently. I'll keep doing the podcast, but that's the vision.

Do you enjoy writing? No.

Do you?

I love it. Yeah, I think it's hard. It's hard work at times, but like in my best days, yeah, I love it. You don't, it's not your, your kind of.

I don't enjoy it, but I know that it's crucial to even posting a good podcast is requires you to have takes, right? To ask a question is like has to come from a place of understanding and writing has been super helpful at that. But I, it is sort of a struggle.

Yeah. Yeah. Well, this is no BS, you know, like in a world, it's like, I feel like you go so deep in these areas. I've always tried to kind of like entertain people, but also educate them and not talk down. Um, I, you know, I, I think I, some of the things I do, like making movies for Netflix or HBO or something is like a slightly more mainstream. I don't know. You know, it's just like hitting people with I didn't phrase that the right way, but I think you get the vibe of what I'm saying. And it's like, there's so much pressure to dumb things down to like, just cater to the lowest common denominator. Even, even like someone like MrBeast, whether you love him or hate him. I mean, it's like, you know, there's a, there's this, you know, IV of shock value and entertainment. And that feels at least where a big chunk of the world is going. So it's like, I just really like admire what you're done and done it at such a young age and just that you're like, you found such a big audience. I think it's great for the world, right, to actually get smarter and learn things. And so it's really cool to see, see how you did this and like, I'm glad you're doing it.

Thank you. Thank you. And thanks for interviewing me, Ashlee Vance. I'm— yeah, it was really fun to talk.

And yeah, no, thank you. Thanks for letting me crash in your place. And, and I hope to learn learned from you.

Likewise. Likewise.

Core Memory is hosted by me, Ashlee Vance. Our guest today was Dwarkesh Patel. The show is produced by me and David Nicholson. Our theme song is by James Mercer and Jon Sortland. And Core Memory is edited, as always, by the stupendous This John Sortland.