AI, Learning, and Podcasting | Dwarkesh Patel | Ep. 19

[00:00] First of all, we just didn't realize how much we didn't know about human evolution. Just like the story you learned in high school, all of it is like at least somewhat false about how, when, where, who. What do you mean? Like, did it happen in Africa? Did it? A big chunk of it didn't. We have stuff right up to, you know, a certain amount of history though, right? Okay. Yeah. At least there's something we can hold on to. Dwarakash, I've been really looking forward to this. Thanks for making time for it. Thanks for having me on. So I want to start by talking about your thinking around the state of AI. [00:30] You're a user of it. You have gone really deep with a lot of people who know it on many levels. And you recently wrote this really interesting blog post called Why I Don't Think AGI is Right Around the Corner. And I want to ask you a little bit about that and just this general topic. A lot of my guests so far, probably myself included, have been like a little breathlessly like, you know, this is [00:53] This is here if we just sort of deployed all the AI research that we have or capabilities today, we would have insane GDP growth. I think you have a slightly different take than some of my other guests. So I wanted to start by asking you about how you see the current state of AI. I'm in a similar position as you, where I've also interviewed a lot of people who are breathlessly anticipating what's coming with AI, sometimes in a very optimistic way of the AI researchers. In other cases, they're worried that like... [01:21] The world's going to end in two years. And I think. [01:24] What's changed my mind around how soon we're going to get to these super transformative outcomes is just trying to use these AIs to help me with very simple like script kitty kind of task for my own podcast. And so I have a lot of friends who think, look, if the reason the Fortune 500 isn't using AI all across their stack right now is because the management is too stodgy. They're like just like not being creative enough about how to get O3 into their workflows. And look, I'm like I'm thinking a lot about how to use AI in my podcast post-production setup.

[01:54] tried for 100 hours to get it to be useful for me. And it hasn't been that useful. [01:58] And I think that's because [01:59] It's just genuinely hard to get human-like labor out of these models fundamentally because [02:05] these models can't learn on the job in a way a human can. So if you think about a human employee, probably for the first three, six months, they're not even useful, especially when it comes to knowledge work. The reason they become more useful over time is not mainly their raw intellect, although obviously raw intellect matters, but it's rather their ability to build up context and to learn from their failures in a very, um, [02:27] in a very rich way and to interrogate them. And the models currently, you just get like whatever they can do in a session. You talk to them for 30 minutes and then they totally lose awareness or understanding of how your business works, what your preferences are, etc. And a lot of tasks just require you to like you do a five out of 10 job at something. Then you like talk to your boss, you're like go out to the consumer and then like you learn what didn't go wrong. You like ask [02:57] they just can't do this on-the-job kind of training, which I think, like, [03:00] is what makes humans valuable. Yeah. Like there's a certain set, particularly within like language tasks, and maybe we can get over to coding, which is sort of like a whole different beast, but within language tasks, it seems like there is a limit still to like how good it can be. And so for example, even if you're trying to either pull out the most interesting segment from a podcast or caption it and make clips that are postable, you know, those clips need to be perfect. And like a human still, you know, you're going to trust a person more to do that. Picking what's

[03:30] So offline, we were having a conversation about how to [03:35] how to write a tweet for something for your podcast, right? And then we were discussing, oh, well, once you write it, you might be like, write it for a group chat. Maybe you can add that to, like, you would think this is exactly what LLNs are for, right? It's like, you write the tweet, this is like language in, language out, here's a system prompt. But like, why is it, I assume you aren't getting that much use out of just like having the AI write your tweets for you? And why is that? Not because, by the way, I'm focusing on this not because writing tweets is like the most [04:05] be able to do, right? Why are we not delegating this to them yet? If you posted something, you like notice it doesn't do well, you have this, um, [04:10] You can think about what went wrong. You also have this experience from what do your users want or what do your followers want that these AIs can't pick up easily. Well, it's actually a reasonably high… so think about you posting something on Twitter or Substack or whatever. That bar for you is actually going to be pretty high. There's a lower bar of thing where language models are really useful, like customer support tickets for example. It works really well there because you don't need the language to be perfect. You don't need insane nuance and it's actually okay. [04:40] for a certain class of thing if it's right 97% of the time and if it knows how to say, hey, I'm wrong here. So I do think that what we're describing here still does change [04:51] drive a lot of GDP growth in certain areas, at least, because you don't need the high bar everywhere. Yeah. I mean, I'd be curious to see what the actual numbers are on customer service employment. I think from what I understand, they're not like down that much. Yeah. So it's interesting, even in those areas, you're not seeing these transformative impacts. Yeah. I mean, the venture view against that would be that it's like the beginning of this massively exponential curve. And these things are like, you know, in the first half of the first inning, but all the

[05:21] up. I agree with that, maybe not just in three years, but I agree with that over the course of a decade. But that's just because I think this is such a big bottleneck to getting these models to be valuable that it would have to be like people will want to solve it. Like right now, OpenAI or Anthropics revenues on the order of 10 billion ARR. And that's a lot, obviously. But like McDonald's and Kohl's make more money. And those companies aren't AGI, right? So like if you have real AGI, [05:51] wages. Given that that is sort of the addressable market, and given that this is one of the key bottlenecks to getting there, I just think a ton of effort we've put into this problem. We've had a lot of progress in the past. So I'm not like one of these people who's like, AGI is not coming. Yeah, of course. I'm just saying this definitely needs to be solved before we get there. I guess maybe then to the broader thing about being AGI-pilled in general. You obviously have spent a lot of time talking to researchers. You're very close to that. Do you still feel as confident [06:21] from your guests. - There is one thing really interesting to observe, which is that like, they have correct reasoning. [06:27] It just so happens that reasoning ended up being much easier than something we take for granted. [06:32] which is just that day in, day out, you're going to be picking up information [06:37] in your workplace. But you know, like you go back to Aristotle, [06:41] and [06:42] His big take was like, look, the thing that makes humans special... [06:46] from other animals is that we can reason and the other animals can't and It's sort of funny that like these models just aren't that useful yet. I

[06:53] They can't do almost anything. [06:55] except for the one thing they can do is reason. Given the fact that these sort of ambiguous capabilities do come online, [07:01] That makes me think that like, continual learning might also be something that like in 10 years, it's also useful to remember deep learning is like not that old. It's like 13, 14 years old, or at least like sorry, Alex Knight is like that old when we started training these models. Yeah, like it's very possible to me that in a decade or two, we find a solution to this problem. It's very interesting to hear you continue using language that's like, you know, these models aren't that useful yet, which [07:23] part of me very much agrees with. And part of me, it's like putting language to an experience that I have when I use them myself. The other side of it would be they are an amazing replacement for search in a lot of cases. They do things like code generation in a very effective way. It seems like doctors are able to use them to handle a lot of scribe wars. So there are these things that are, I would say, clearly working. Do you see it that way too, where you're like, some things are [07:53] I'm putting it more in the context of, [07:55] the real potential for AGI, just like a genuine replacement for human labor, I expect that to cause like, um, [08:04] a 10x increase in the level of growth. And so if it's on the scale of the internet, [08:09] I'm like, oh, wow, this is so much shorter of what AGI could be. So this is clearly not that useful yet. And a sort of more tangible reason to expect this kind of change. But one is just that... [08:21] the amount of labor supply just dramatically increases. So I think people often, especially in tech, focus on how it will make a specific

[08:28] industry more productive, this narrow productivity improvements where he's just like, no, imagine like a trillion people in the world who are each specializing and each discovering new knowledge, or we just get all these gains from comparative advantage as a result. But another is that because these minds are digital, they have advantages, even if they're the same amount of intelligence, specifically advantages in collaboration. [08:52] that humans just can't have because of the way our minds work. One example of this is, okay, suppose this problem is solved where we can actually learn on the job. Now, a human can on the job learn from their work, and then over the course of 20 years, they become a master of their craft, they're incredibly valuable. You're one such person, right? You picked up all this context in the tech industry from running companies, from investing in companies, [09:14] If we get models, [09:16] that have this human-like capability, not only could they learn on a single job, copies of the model are deployed all through the economy. They can amalgamate the learnings from basically doing every single job in the economy at the same time. And at that point, even if you don't have further algorithmic innovations, you would still have something that's functionally becoming a super intelligence. You have this broadly deployed intelligence explosion. They're just one of the many ways in which the fact that they are digital just gives them... Like emerging intelligence like ants or something. Yeah, exactly.

[09:46] impact happens is through a trillion new, you know, [09:50] white collar workers. There's another version of it where it never actually does exactly that, but does this other thing, which is just like a higher level of intelligence than anything we've ever encountered. And it like, [10:03] creates new paths and identifies new ways to do things that humans could still then do. Sometimes the way I think about this is like the 400 IQ AI, even if it can't do all the things that a person can do, it could help or fully identify new drugs, help us get to space more effectively, like all sorts of innovations like that. I think a good way to think about this is maybe China. So, okay, why has China been so successful in not only catching up in science and technology, [10:33] a lot of key domains. And obviously China's full of lots of brilliant people. And so that does lend credence to your argument that like, look, they have to have, I think the intelligence is a big part of it, but I think more fundamentally, just like once you've hit a benchmark of intelligence, the scale is what makes China so successful, right? You have just within manufacturing, there's a hundred million people who have built up, who are working in manufacturing in China, who have built up all this process knowledge, [10:58] in whatever subdomain is relevant to whatever is being built. [11:06] That scale, I think, is like, if China's graduating, I think like tens of millions of STEM graduates every single year. It's not that any one of them is super brilliant. It's just that each of them can specialize in whatever radar technology that UID needs or whatever production technology is needed. I'm thinking about this out loud. But it...

[11:25] It opens up the question of would more impact happen from a trillion more super connected, super collaborative human level intelligence people or from one just like demigod level intelligence who could like figure stuff out and tell us all what to do. Do we have evidence in Silicon Valley history that it's more of the latter? It seems to me that. Well, it depends. Yeah, there's the great man theory thing where it's like there are these special people who, you know, that's how the big bleeps happen. [11:55] is some just outlier person who directs the resources and that one person can pull greatness out. Yes. I haven't seen them up close as you have, but it seems to me, I agree that they've had a huge impact. People like Elon and Steve Jobs, it seems to me their impact has been [12:09] more so a product of just like, you will do this, otherwise I will throw a tantrum, which is good, right? Like you should throw a tantrum, and I'm gonna sleep in the office, [12:19] for years on end and just like get people in the right place at the right time. But it's less so like [12:24] only Elon can come up with how the fins on the SpaceX rocket should be designed. And therefore, his uber intelligence allows him to design across five different hardware verticals. It's probably, it's definitely something that's more to do with the leading of people and the clarity of vision, I would say, than probably like engineering genius or something like that. Yes, yeah. But I mean, this is actually another way to illustrate why digital minds are such an upgrade. Elon has obviously been super successful across so many different areas of technology. [12:54] But of course he's just like one person. To the extent that you think there's something unique about him,

[12:58] or about the small teams he had assembled, like early SpaceX, early Tesla. If they were digital, you could just replicate them, like early SpaceX team, replicate them 1,000 times, throw them at 1,000 different harder verticals, and see what happens. Because you can't scale that with humans. The other thing I'm thinking about now as we're talking is, could a digital intelligence be like a leader of 10,000 people? And I do think that coordinating large groups [13:28] So now I'm wondering, could you have a digital mind at the top, or does it need to be a human? I think you'd have a much easier time if it was digital. This is another one of the key advantages of being digital. Because right now... [13:39] Elon has the same 10 to the 15 flops in his head. [13:42] that every single other human has. Now this could be negative as well, right? Like Xi Jinping also only has 10 to 15 flaps and it's that every other human has. The ratio of compute or deliberation happening at the top versus [13:53] through the hierarchy is just so lopsided, and that requires a lot of delegation. Now, that's good in a sort of like free society sense. You don't want the president to have that much control over your life. But if you... [14:04] within these specific domains where we want, like, companies have this outer loop where if they fail, they can just go down, unlike a country. So there it might make sense to have more of a, [14:13] Have the company be the product of a single coherent vision. This is the founder mode idea, right? But obviously that's limited by the fact that if a company goes to a certain size, it's just hard for a single person to monitor everything. If it was an AI, and especially if you could like inference scale, mega Elon, the thing that's like running on a huge data center that's dedicated to just him, mega Elon can

[14:35] read every pull request, every comms input/output into the company. He can micromanage every single employee down to the technician at the dealership. Especially, I don't know, Elon, of 10 years ago, it'd just be quite incredible. So I guess what I'm wondering, is that the AI COO under the human who is giving them superpowers, or is that person, is that AGI rather, actually leading? I think in the near term it will be [14:59] like the latter. But over the long term, obviously, like if AI... You think they'll be AI CEOs basically in the long term. Yeah. And so it does seem like the thing they're lacking at least now is like, [15:09] this sort of more taste oriented stuff, right? They can do the research for you. O3 can go do the research for you, but then you're not going to let it make the investment decision for you, right? That just comes down to your own unique insight into the field or your unique taste. And maybe like running a company will work like this for a while where they can just curate a lot of information for you. And then you still have to make the final call. It's just like the Steve Jobs model of like he makes the final call. The designers come to him with a bunch of different [15:39] Maybe in the near term, that's what it looks like. So maybe back on AGI, all the time that you spent with researchers and companies and you really inspecting this and you're very truth seeking, you still do believe that AGI is going to just like wake up like you think it's all going to happen still? Or has anything, as you've learned over the last year or so, changed your mind? And you're like, well, it's a really cool idea, but it's not for sure it's going to happen to you. Yeah, I think the biggest consideration there is the progress for the last 10 years in AI. I mean, more than 10 years, actually, but even going back to the 60s, the progress in

[16:09] has been driven by increases in compute. And especially in the deep learning era, it's been like stupendous increases in compute dedicated to frontier systems. If you just look at public systems, [16:20] announcements of how big the compute runs are. It looks like the trend since I think [16:25] 2012 2016 has been 4x per year so this over four years is 160x the biggest system Train from the one before in terms of the compute used that physically cannot continue to [16:37] past this decade from how much energy is used to like what fraction of advanced ships at TMC you need to procure to even like raw fraction of GDP. So then you would need just need progress from [16:48] other, the progress would just have to come from algorithmic progress or, I mean, it would just have to be that. And because this key input into AI progress would stop after the next five years, there is this dynamic that the yearly probability of AGI is like quite high now. Not in the sense that like it will happen. [17:06] But there's a decent chance every year until 2030. And then it just has sort of craters. Because then you're just like, okay, we'll just have to think hard about what's missing. We can't just throw more compute at the problem. Yep, yep, that makes sense. Do you feel right now like AI is making us smarter? Or is it like increasing brain rot? And do you think like over time that that goes in a particular direction? Did you see the meter uplift study from the other day? Oh, it was super interesting. Meter is this organization that does evals on basically how AI is progressing. [17:36] interesting result the other day. So they had open source

[17:42] developers who are working in repositories that have tens of thousands of stars. They did a randomized control trial with these people. [17:50] where they would issue them a random pull request that was open in these repositories. And they'd work on it in one case, just by themselves, in another case, with the help of Cursor and Cloud 3.7. And then they measured... [18:06] in the case where they're working with the AI, one, how much do you think you were sped up? [18:11] And then two, how much were they actually sped up? And the developers thought that they were 20% more productive as a result of AI. They were actually 19% less productive as a result of AI. It was really interesting to read the threads of the developers who participated. So people who have looked at this, who are even more bullish on AI, are like, [18:30] They think their experimental design was extremely robust. Even senior engineers misread themselves? Especially the senior engineers. Wow. The senior engineers, from what I remember, the biggest decrease is in productivity. So these are people who are experiencing this project. They've been working on it for decades. [18:44] There's a couple explanations. One is just that I think in a lot of domains there is this common failure mode in intellectual work where you default to procrastinating by doing a thing which seems productive but is not moving the ball forward that much. The classic example of this is in college, instead of rereading the textbook, you should just do the practice problems. [19:06] And maybe using these AI tools and then like going on social media for 30 minutes while you're waiting for the completion to complete is another example of this. Why did this happen? So at least for now, to go to your original question, it's like not obvious to me that it's making us smarter. It's interesting. The version for everybody is I think a large number of people are...

[19:26] having chat GPT sort of like give them guidance in life at this point. You know, I don't mean that in some like huge philosophical way, although maybe in some cases, but even just like day to day people are like, here's my whole setup. This is everything about me. Like, what should I do? And that is a big influence. And so it's sort of, we kind of, to the extent that these models have quietly gripped a lot of people, either through their decisions, through their engineering work, [19:56] But have you found it useful, that kind of stuff? Just like, I mean, it is sort of like, okay, plan out a cute date for me or something. I use it some. I think I use it less than a lot of my friends. I think more I'm commenting that I think broadly a lot of people do use it for that. Even if you or I happen to be in the group that uses it a little bit less, I think a lot do. [20:17] Yeah. You know, and even probably taking personal advice and things like that. Yeah. Which is probably good in a lot of cases. Like, I think it is very smart and people know how to use it and things like that. But it just speaks to even with engineering and then also on, you know, people using it and sort of like search and question asking for personal use. That's got a lot of influence.

[20:47] in synthetic biology. And there, most of my prep time was just dominated. Talking to these models and then just telling them, teach this to me as if you're a Socratic tutor. Don't move on in the explanation until you're satisfied that I have [21:00] completely understood. I guess I'd be curious, yeah, if people in other domains have- Do you feel, actually, on biology in particular, have you spent time with Patrick at ARC by chance? A little bit, yeah. I feel like the bullishness that a lot of people in biology have for AI's ability to do drug discovery and things like that seems very promising to me. I don't know what you've found as you've spent time in biology. One interesting question I have for these people is, in biology, we- [21:23] we can either employ models which think in thought space. So just like humans, they can come up with hypotheses and so forth, or models which think in protein space or DNA space or capsid space. You know, like humans just aren't [21:36] We can't like, you know, I think G sounds really good here. Then let's do T next. And so I'm curious if [21:43] Which one they think is a more promising or more useful complement to the current progress in biology? Is it having better models that can think in, you know, like the alpha-fold type stuff? Or is it just like have O3 come up with hypotheses and just like write them out in English? At least George Sirson to think it was the biospace, like thinking in proteins or DNA or so forth. Because this is, while we have millions of life science PhDs who can like come up with ideas, being able to prune through them in simulation. Yeah, exactly. Like a digital cell kind of thing. [22:13] is the more useful compliment there. - Yeah, I mean, that seems like that would be like a very unequivocally positive output for humans if we can sort of wildly change biology and pharmacology and things like that. - Yeah, I think in the long run, I sort of worry about the fact that

[22:28] Um, [22:30] We like we know ways in which [22:33] things can just go horribly wrong. Like we have the equivalent of nuclear weapons, but in different domains. So mirror life and biology, [22:41] Apparently, if you have life with the opposite chirality, there's just no defense. [22:49] like, plausibly it could render many life forms unviable on Earth. And so George Church was one of these people who wrote this letter saying like, "Look, this thing exists, let's not work on it." But like, I don't know, over the course of 100 years, what's the equilibrium here? In physics, I interviewed this physicist, [23:05] One of the things he's worked on is thinking about this thing called vacuum decay. And the TLDR is basically like, it might be plausible to just literally destroy the universe. Well, what's the idea there? Look, I'm a podcaster, so you're asking a content creator about physics, so I'll do my best. [23:23] the [23:25] In quantum field theory, we're in a sort of what's called a metastable state where if it's like sort of having a huge valley and then a little bit of a hill and then we're in this like little rump here, it's possible to throw such a huge amount of energy. And what would happen is that this like bubble would expand at the speed of light. [23:43] which would just be like total destruction. It sounds like some wild sci-fi thing. No, but it actually takes me to the next thing I wanted to ask you about, which is you have this wide range of guests and interests. So maybe first, outside of AI, what domains are you most interested in right now? Because I feel like I've seen you talk about

[23:59] you know, politics and Russia and math and science and longevity? Like, what are you interested in most right now? I'm interested in what the year 2050 looks like. Obviously, you need to understand AI in order to understand what happens in 2050. But throughout history, [24:14] There's never been a case where... [24:17] There's just been a single technology which explains why, say, the Industrial Revolution happened. Right. It wasn't just that we made better textile machines. You have [24:26] improvements in sector after sector, which are enabled by key innovations in specific sectors. But like, for example, yeah, it's important to learn about what's happening in bio and robotics, etc. So I want to get into those fields. Also, I've just been interested in the fact that we are finally getting to [24:44] a pace of change that we actually have seen before in history, but not for a long time. I most recently interviewed this biographer of Stalin, Stephen Cawkin. And I think Stalin is born in the 1870s. And you just think about his life. [24:59] From the 1870s onwards, railway, planes, airplanes, steamships, radio, telegraph, light bulbs, combustion. I mean, World War I is a crazy example where you start off the war. [25:16] I think there were... the Wright brothers had flown, but there weren't like many... they were like on the order of hundreds of planes in the world. And there were no tanks. Like tanks was not a thing. And World War I ends with... [25:29] Like it's a tank war, it's a plane war. In not that many years. Yes. You go from almost no trucks to tens of thousands of trucks in the course of four years.

[25:38] And planes? Yes. So I think even planes were at an extremely low level before the war. And then the military use during, obviously. Yes. So I mean, the reason Germany thought it was going to [25:47] win is because it had this like real network and there was this plan on how you could do this two front war and knock out both France and Russia. [25:56] at the same time, just by like, you'd be amazing at the railway logistics. And I think the, um, von Moltke or whoever the leader of the German command was said at the end of the war, like we lost because of trucks, right? Like we didn't anticipate that like there was another way to ship enemy combatants to the front. Yeah. We're just going to see like this level of change across so many different sectors. There was, um, somebody posted on Twitter, Bucko posted something that I thought was an interesting point about you, whether or not it's [26:26] But I'm curious how you react to it, which was it seems like you went through this evolution where you were learning a ton about AI. And then it seems like you believed that AGI was coming. And then your interest started expanding out into all these other things like, you know, geopolitics, biology, all these other areas. I'm guessing, you know, which was sort of like saying, you know, the technology is what, you know, creates technology. [26:48] moment for change, but then this backdrop of the world is what really influences it. I'm guessing that the chronology of your interests were a little bit different than the framing, but I'm curious just how you think about it. [27:00] Given what I do, I'm actually quite pessimistic about being able to learn from other fields. I just know people who will read some philosopher in the 19th century and they think, "Oh, this explains how Silicon Valley works," or "This explains AI." And I'm like, "No, I think you just have to read the papers about AI." I just think it's very hard to generalize. You're saying to understand the technology, you have to read the papers. Yeah. I think just people have this idea that I'll come up with my grand theory of history. Yeah, you're like, "It's not philosophy, it's science."

[27:30] hard to have this like I think there's I know a couple of people who can do this and I find it really impressive but what I noticed about them they're just like not hand wavy at all so there are people who like for example if you're trying to model how AI will impact economic growth one way is just to like read the sort of like firsthand accounts of people going through it in the 1500s and like oh like let's read the biography of the Medici and so forth and there's other people who's like okay let's look at the growth rates going back 10,000 years and [27:57] What is the long run secular trend? What actually explains what changed? Well, there's the endogenous growth theory where the key change is that population growth and more people come up with more ideas. Okay, well, AI is more people, they'll come up with more ideas, there'll be more specialization. So there's this very different mode of learning from other fields, which is [28:17] empirical and, I mean empirical is maybe the wrong word, but just like very falsifiable and grounded versus I'm just gonna read, I'm gonna go to the library and just like read a bunch of random books. Totally. So how do your interests connect to each other then? Like are you following any particular threads or you know, does one connect to another in a certain way? Like what's, what would you say is driving what creates various interests over time? Honestly, it's just a super, [28:43] uh bespoke whatever I happen to be interested in that week if I'm reading an interesting book how are you spending your time like are you reading a lot like are you is it do you learn mostly through reading through talking to people is are those are there other methods reading reading um [28:56] I think there's a couple of people you learn a lot from talking to. In general, I've just sort of been disappointed about, like, I mean, given the fact that I'm a podcaster. That's interesting because you're talking to, like, some of the smartest people, though. It's also really interesting, especially, so in some domains, people can be super, like, generative outside their domains. I've sort of been disappointed about the fact that, look, you might hope that you could talk to some historian about World War I or about the history of oil or something. And then they'd have insights about, like, how this applies to AI.

[29:26] not from them. When I was interviewing Daniel Juergen, who's the author of The Prize, it's this book about the 200-year history of oil. One thing I thought was really interesting is that [29:35] uh, [29:36] Drake discovers the first oil well in Pennsylvania, I think in the 1850s. [29:41] And then the Model T car, or I think it's like 1905 or something, that like you finally have cars with internal combustion engines that people are using to transport all around the world. And this is a sort of industrial [29:54] case use of oil. Before that most of oil was just wasted. It was only the kerosene component. So all of Rockefeller, all of that history that you sort of think about as a Gilded Age like oil bear and stuff, that's happening when a small fraction of oil is being used just for lighting before the electrical light bulb was invented. And in fact when the light bulb was invented, I remember getting my days wrong on the Model T and the light bulb, but it's like around that area. When the light bulb was invented, [30:22] People are like, oh, [30:23] oil is like standard standard oil is going to go bust because what's the other use case of oil and so i do find it interesting that it took more than 50 years to go from we have discovered limitless energy in the earth [30:34] to here's a way to use billions of gallons of this stuff. And I think it has maybe interesting implications for AI, where look, we have like, [30:43] Basically, it's sort of shocking how cheap AI is. I think that's why I always find it confusing that people are like optimizing on cost. Like, do I want like two cents per million tokens or 0.2 cents per million tokens? So we have this like, [30:53] commodity that we could potentially use at an industrial scale [30:56] and

[30:58] We don't know how to, like, part of it is just technological. We don't know how to, like, get those tokens to be more valuable. And part of it is just like, yeah, what do we do? What's the internal combustion engine equivalent for AI? So you feel like most of your learning comes from what you read rather than... [31:13] you know, through your conversations with people. [31:16] Yeah, I'm very lucky that there's maybe a... [31:19] six to twelve people who I've known for five years, almost all of them I've had on the podcast, but who I'm in just extremely regular touch with. I've genuinely learned a lot of what I know from like this handful, like this group chat. In one sense, it makes me feel like, I know that this is weird that I've like known them for five years. And now they're also like super successful. But like, we were all college students at some point. Yeah, there's a lot to be said for that, you know, that five closest friends. Yeah, it's a big deal. Yeah. How do you feel about this? Do you do you learn more from talking to people? [31:49] I also think that there's different sorts of things that you can be seeking. Seeking the truth versus seeking a good decision are very related, but not exactly the same thing. For example, if you're trying to learn from people about how do you spot great talent and what can you pick up from somebody, you're not going to get to the truth. You're just going to get to techniques and things that have been useful for other people that you try to apply to yourself. So for that kind of thing, I wouldn't know how to read about it anyway.

[32:19] what's happening in... [32:21] robotics. I don't know. I've been sort of underwhelmed by how... Yeah, I mean, the schools of thought to me would be either you can try to go [32:31] learn it for real yourself. Or if that hill is too high to climb, which for me, getting into robotics in the white papers way, I would be kidding myself to think I could catch up and then get to the edge of anything on any timescale that mattered. And so for me, there's the other move of, is there a way, if you're going to do it, to shortcut the decision somehow to somebody? And so who can you most trust to give you good information or something? I think we're also in a lucky position where we have [33:01] enough public output that we can sort of reach out to people and they'll say yes. This is a tougher position to be in if you're like 19 and I want to like [33:08] I want to learn about biology. I feel very lucky because a lot of people do great work in many different domains. [33:14] Mine just happens to be public facing by default. So I get like more [33:20] ability to reach out to people than I would have default. - Than someone doing great work in private. - Exactly. Which does create this flywheel where if I do make good content, smart people will be willing to talk to me. That helps me make better content. I think that's actually more relevant. [33:36] to why the podcast grows than just like audience tells people. Some natural flywheel you're saying? Yeah. Before sort of moving on to a new topic, I wanted to ask about that's related to this. You mentioned a bunch of really interesting ideas there around oil and history and Stalin and biology. And I'm curious if there's any other just ideas recently that have just really stuck out to you that you can't stop thinking about that have really gripped you, just because I love hearing about these.

[34:06] geneticist of ancient DNA, David Reich. And what his lab and his research has revealed is that human history, first of all, we just didn't realize how much we didn't know about human evolution, just like the story you learned in high school, all of it is like, [34:18] at least somewhat false about how, when, where, who. What do you mean? Like, did it happen in Africa? [34:24] - Did it? - A big chunk of it didn't. Like there was this, there was a group that went out 400,000 years ago, and then they mixed in back with the group that left out of Africa 70,000 years ago. A lot of the evolution that led to this branch of humanity maybe just didn't even happen in Africa. When did it happen? We were like learning more stuff about it. And then the key thing we're learning is how did it happen? [34:45] And it seems like [34:48] we just see this pattern again and again in history, which is very disturbing, but super recurring, is that some small group will figure something out. And it's not clear from the genetic record what it is, right? Like 70,000 years ago, there's this group of 1 to 10,000 people. [35:05] in the Near East, so like where the Middle East, North Africa are right now, they figure something out. [35:12] And they wipe out... [35:14] every single other species of humans across all of Eurasia. There was half a dozen different species of humans. There were the hobbits, obviously the Neanderthals. The hobbits? I forget what their real biological name is. I think they're called the hobbits. That can't be it. Yeah, yeah. That's good. The Denisovans. [35:35] Um,

[35:36] They're all wiped out by this one group. It starts out with 1-10,000 people, expand all through. 10,000 years ago, Anatolian farmers, also from the modern-day Middle East, they expand out, kill off 90% of the hunter-gatherers in Europe. [35:51] through Asia. This also happens again, by the way, with a group that goes through the land bridge to America. They also keep doing this. There are like multiple waves and like one of the waves killed off the remaining ones. Wow. The only people who survive, by the way, interestingly, are that we have genetic evidence of is this group in the Amazon where because the Amazon is so densely [36:11] uh it's so dense to get through like the genocide wasn't completed and so it was like more of an intermixing then 5 000 years ago the yumnaya [36:20] which is this group of like step nomads, they sweep through all of Eurasia again. Like, and we're talking about like 90% death rates of the domestic population. So, and not just like multiple continents, like, [36:35] All of Europe, the people who built the Stonehenge are killed off by these people. This is insane. And they're pretty sure this is right? Yes. Wow. By the way, the way you learn why it's a genocide or why it was like violent is you look at the fraction of maternal versus paternal DNA that comes from the native population versus the invading population. And what they'll find is that the maternal DNA comes exclusively from the native population. Wow.

[37:05] killed off. Wow. It also explains a bunch of... So, like, India today is a mix of the original Indus Valley civilization from 3,000 to 5,000 years ago, plus the Yamnaya. And so, all of India is just, like, this, like, gradient, basically. Like, north is more... [37:21] of this Yamnaya ancestry, south is this Indo-Indus Valley civilization ancestry. So basically just had like a lot of this wrong? Yes, but you know what else is really interesting here? For hundreds of years, [37:34] anthropologists, archaeologists have been doing this Indiana Jones thing. We're going to read the scrolls and we're going to think hard about how [37:45] what they're trying to say here, and read the literature or whatever. And it was just so useless in comparison to one mathematician who went into this field and he's like, okay, let's just look at the haplotypes. [37:57] and see how they compare. And just like totally redefined our understanding of [38:03] Basically all of history going back millions of years even to things that happened like 500 years ago. All of that is like totally a We can re-understand it all these mysteries about like why did the Mycenaean civilization fall and who exactly were the Mycenaeans in Greece? Just like all around the world. We have stuff right up to you know a certain amount of history though, right? Okay. Yeah At least there's something we can hold on to. But I just think it's very interesting actually. It's like that you [38:29] No, it was just like all this sort of like esoteric reading and understanding. Right, like what was happening in the year 500, like that we could have that super wrong, for example. Yes. Oh, I mean, something really interesting we learned speaking of the year 500 is, uh,

[38:41] The Roman Empire, I think, is like... [38:43] was it 540 or so? Is that there was the, basically the black death. [38:49] kills off like half close to half of big fractions of the Roman Empire and this is [38:57] around when the roman empire falls there were also there were also previous plagues like the antonine plague in [39:02] in the second century or third century. I interviewed somebody about this. I didn't realize the extent to which [39:08] how Rome actually had something as bad as a Black Death and how that contributed to their collapse. Because another bias we have in history is just to think about, oh, we had the four good emperors, and then they did a really good job, but then the next guy fucked it up. No, just like there was a climate optimum during the quote-unquote four good emperors where the breadbasket was really heavy. Yeah, just civilizations happened to fall after a certain amount of time. It's like, no, a thing happened. Yeah, exactly. I often wonder, because over the last few hundred years, like we keep like learning a new thing, you know, [39:35] something was completely wrong, whether it was like, you know, the Earth's round or gravity works this way or whatever else it is. And I'm like, we must still have big fundamental things wrong, you know, like human history and that kind of thing would be like a good example. My five-year-old recently [39:49] which was a funny question, was like, do you believe in Jupiter? And I was like, I do, but that's a great question. And you could ask me about something else in the universe, and I might think we have it super wrong. But we must still have big basics wrong. [40:03] Yeah, I feel like this is especially when I talk to [40:06] Physicists. We're just like very basic questions of like, [40:09] Is the universe infinite or not? We have no idea.

[40:12] And there's many different kinds of infinities it could be. But just that very, it just seems so consequential. Or the fact that time warps when you go different speeds. I'm like, we just can't possibly have it all exactly right. It's just too crazy, right? Or like dark matter. There's just enough big stuff that I'm like, I bet we don't have it quite right. But that might sound like some PhD physicist is listening and being like, what is this idiot Jack talking about right now? [40:42] A lot of stuff has to be. Yeah. Okay. This kind of flows into the next thing I wanted to ask you about, which is your broad sort of perceptions of the way learning is happening. And, you know, you've obviously been sort of like learning in public. You do a lot of self-directed sort of education and things like that, you know, you, but you're also, you've got a, you know, one foot sort of very tied to people in academia and at the top of research fields and things like that. And I'm curious, sort of, you know, your opinion on sort of this transition that's happening where [41:12] that the stuff should get consumed is self-directed and not part of the big institutions and things. The standards people have for like... [41:18] "Is this thing true? Do I really believe it?" have just really degraded, especially in sort of podcast land to criticize my own tribe. People will just like, people will just fucking say shit. Whatever you say about academia, there is this idea like, okay, does this make sense? Like, have you actually made like a clear argument? Like, do you even have a clear like end statement or is it just like, [41:34] the thing that people are saying. On the other hand, [41:37] Look, I mean, is it like net good for the world? If you read history, you read about the worst things that ever happened. The Cultural Revolution in China, the Great Terror and the Soviet Union.

[41:47] You can complain that Twitter has low average IQ, that like a conversation is very dumbed down, but you just don't need that many IQ points to realize the cultural revolution is bad. You just need some mechanism where you could have gone on and been like, why is Mao having us kill all the sparrows? Like, isn't that actually really bad for, you know, like pest control and just like making fun of this deification? And I think that actually has worked, right? Like woke was a thing for a second. And then I think social media contributed to that being less. [42:17] in front of Trump is like a thing that people do and has worked. And so on that, I just [42:22] Getting rid of the worst accesses is much more important for making history go well than making sure that we can have these gigabrain genius level takes all the time. And I think social media does a reasonably good job of helping us correct the worst accesses. On the sort of truth point, I think part of the issue is that the legacy media corporations, in my view, have lost quite a lot of trust from people, myself included in many cases, where they seem like they've got agendas. [42:52] of for-profit and maximizing eyeballs. And so I'm like, citizen journalism on Twitter, it's not like I trust that completely. But I also don't trust the institutions completely. So I don't know that one's that much better than the other at this point. I disagree with this. My attempt to do this thing has... [43:09] actually given me more respect for the media. [43:13] in a couple of ways. One, I think they genuinely are better at holding power to account.

[43:19] than sort of independent creators. [43:23] like talking to somebody, an extremely powerful politician or business leader, and then asking tough questions is a thing that the media will do. [43:32] And often... [43:34] It won't happen if they could just get to go on the podcast of their choice. And it's like, it's harder than it looks. And they're willing to uphold these standards when they do these interviews. Now, I mean, I agree that it can often be sanctimonious when they do this. Is that to do with the person or the institution? Like as an example, like, you know, Tucker Carlson. [43:53] goes Fox News to Indy, same guy theoretically. Does that change his truthiness? I mean, I think this is another example of... [44:01] his show and many others um this is not coming from a place where i i'm like making an object of political point i'm a libertarian i'm like sort of close enough in embedding space to these people um but the standards of discourse in these new places are just abysmal they'll just um uh if you just like if you had a conversation you could just like save one of a thousand things and you just like pause on one of these like what exactly do you mean here why do you think this and obviously [44:31] They just like genuinely have fact-checkers. [44:34] who will go through content. I know that there's many cases where they failed by their own standards, especially in tech journalism. And I don't like their bias in these kinds of things. [44:44] The standards of an order of magnitude difference. I mean, my view is we actually probably...

[44:49] At first with AI, it might have looked like that was going to be a big problem for the New York Times, and maybe in some ways it is, but I would actually, as time's gone on, [44:58] I think we probably need these institutions more than ever in a certain way, which is that, you know, like, AI also now brings, you know, a whole new layer of FUD to what's true with deep fakes and random content generation by bots everywhere. And so you kind of at some point do go back to needing somebody to, like, really hold the standards of, like, truth as much as possible. So I would think that should make these institutions more necessary. [45:20] Yeah, and I think like we always had to compare AI against the counterfactual. I mean this goes back to the are they making us smarter or dumber? Yes, do they hallucinate? But they are probably I don't know when I talk to an AI this I feel like this goes back Do you remember like 10 years ago people were like oh we can't trust Wikipedia because anybody can edit it? Yeah, I feel like that was always a sigh off Yeah, I think it was like all the reasonably trustworthy always well, there's one version We have this attitude towards AI. We're like oh hallucinate, but there's one version of AI where you're asking it directly there's another version where somebody who has a [45:48] propaganda intentions uses it, you know, to their advantage. Right. [45:51] We don't have a lot of time left, and I wanted to get to the last topic. So up until you, all of my guests have been VCs or founders, and I've been sort of using this as an excuse to kind of publicly learn from them. One of the things I wanted to sort of learn from you is about... [46:06] the way you've done podcasting, because you've done it as well as anybody I've seen. You were like the one person I reached out to when I was getting started for advice. You gave me really good advice, which was to just, you know, basically all centered back to just like be authentic, follow your interests, like don't post stuff you're embarrassed to like all that. And that was basically like the one North Star that I had. But I'm curious because you've.

[46:26] had so much success with it, if you can kind of reflect it all about what's made it work, or, you know, like, why has it played out the way that it has so far? It's sort of very hard to say from the inside. [46:39] I feel extremely lucky that my job is I get to sit down this morning and decide what I want to learn about over the next few weeks. I'll interview the person who's the best in the world at that. I get to pepper them with questions for a few hours and then I get to repeat that week after week. I try to ask. [46:52] the questions that I generally want the answers to [46:55] Including the questions I want to answer to after having done two weeks of prep in that field and hopefully having had learned about it over the previous years. And so much content is very much like, give us the intro chapter of your book again. Explain this very basic thing in your field. I think people just really appreciate the feeling of being a fly on the wall. Like one of the reasons it's valuable to be in San Francisco is that you go to dinners or events where you will miss a ton of context. [47:25] You won't know what they're talking about, but it sort of raises the bar and immersion learning works. I try to provide that kind of environment in whatever field I'm trying to learn about. And I think people appreciate like not being talked down to. Having a sense that like the host is actually interested in the questions they're asking. These are like... [47:42] if they were having a private dinner. I think the dynamic to replicate is if you were at a private dinner party, you wouldn't just be deferential at a dinner party. You'd hassle them if you disagree with them about something. But if you have a fun vibe, can you explain this concept for everybody else here? You wouldn't have that dynamic. Somebody that I spoke to recently who worked closely with Steve Jobs, who I'm actually going to have on the podcast at some point soon, said something that I really liked, which was that one of the things that made Steve Jobs special was

[48:12] of just, [48:13] operating day to day, the way you talk to people, the way you give feedback, the way you, you know, ask questions was just really good at those fundamentals. Before we started, I asked you about like, what have you learned about conversations? Because, you know, I see that as like a big sort of fundamental thing that everybody does that you're very practiced at. But you made a point, which is that actually what it is for you is preparation. That's the center piece. And like, that's your fundamental. I still think that's highly applicable to everybody, [48:43] going to interviews as either somebody on the candidate side or the employer. We're all meeting people for [48:50] all sorts of things, but everybody's preparing for stuff. So what is your preparation like? What does that mean for you when you're saying, "I'm preparing really hard for this," and it's the center of your YouTube? In some sense, this is the very obvious stuff. [49:02] It's if you're interviewing a researcher and if you read the key papers, um, [49:07] a couple years ago before, back when I was just like, started getting into AI and I was about to interview Ilya, I'm like, okay, I'm gonna like, [49:13] Program the transformer. This is like how I'll learn about this and then try to talk to any of these researchers as I could if I'm interviewing a scholar and feel that Interviewed um this person who actually wrote a rebuttal to the power broker, which is this book about how Moses changed in New York City that itself I think is a 1500 word book or something a certain page book I read that and I've read like his rebuttal of that book and I read like review articles or whatever like different things about like New York construction history I just do this where I try to do this for all my guests so they but in some sense is like a [49:40] very obvious, just like read the things that could potentially be relevant and then write down questions. Obviously the thing I've changed over the last year is I've started using space repetition. Space repetition is this like tool where you basically write flashcards for yourself and this software.

[49:56] serves them to you every couple of months. Like you have to be preparing for something well ahead of time? No, this is for consolidating knowledge across interviews. So if I do an interview, I'm actually gonna like retain a [50:07] what I've learned, because a lot of the concepts connect, especially with AI. With AI, you're just trying to predict what a future civilization of different kinds of beings will look like. And there's no domain of knowledge that is not relevant to this question. So obviously, technically, AI stuff is relevant. But history, anthropology, even primatology, like what happened between primates and humans, everything is relevant. It'll come up in the interview. And so just having it cached. [50:30] through tools like this is extremely valuable. That's cool. So it's like you're like retaining a curriculum of all your work basically. Yeah. It's the kind of thing where if you're reading a book, [50:38] I think either you shouldn't read, at least if you think you're doing it for learning, or you should have a very intensive practice around, I'm going to make the cards, I'm going to write the... whatever thing is the equivalent to doing practice problems for the domain you're trying to study. [50:54] It's talking to me how often I will make a flashcard for a topic where I'm like, OK, this is like so basic. I don't write this down, but I just have to do something right now. And a week later, I'm like, I was on the verge of forgetting it. And you just think about how many books have you read in your life? Hundreds. Right. How much have you like? [51:11] taken away from them or conversations, whatever other medium. The lack of efficiency here is really striking. And so I've been thinking a lot about how to [51:18] make this a process where over the coming years I can be like, I'm really getting better over time. I'm not just like doing the next thing. It's sort of the spiritual opposite of the idea of an AI that's always listening and remembering for you so you don't have to remember any conversation you ever had. And it's sort of just like some

[51:33] thing on your person that is constantly listening and you can always go back to it, but everything's captured for you. You're like, "It needs to get in my brain so that I can learn the next thing." This is exactly full circle from where we started because people will say in response to the continual learning on the job training stuff that, "Oh, we'll just have an external memory system that it will be just like a document of things the model has learned. ChatGPT already has this." And I think a lot of cognition [51:58] is just memory. Like it has to be on board. It has to be cash the whole time. -That's a great place to end. Thank you so much for doing this. I hope you didn't mind being a guest and keep doing your amazing work. I love to watch it. -It was super fun. Thanks for having me on.