Pod Insights Summation with Auren Hoffman Dwarkesh Patel - AI Timelines, Productivity & Global Stagnation About this episode Dwarkesh Patel is the host of the Dwarkesh Podcast, where he interviews leading scientists, founders, and economists about AI, science, and progress. In this episode of World of DaaS, Dwarkesh and Auren discuss: Why experts overestimate AI capabilities The continual learning bottleneck in current models Japan's economic crash and global productivity stagnation Spaced repetition and effective learning strategies Looking for more tech, data and venture capital intel? Head...
Transcript 11,739 words · 3 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:02 Auren Hoffman? Hello, fellow data nerds. Welcome to World of DAS. I'm your host, Oren Hoffman, CEO of Incubate and GP of Flex Capital. Discover more episodes, get weekly data as a service news, original content, articles on data and more at worldofdas.com. That's world of DAS, D-A-A-S, dot com. Hello, fellow data nerds. My guest today is Dwarkesh Patel. Dwarkesh is the host of the Dwarkesh Podcast where he interviews leading scientists, founders, economists about AI, science and progress. Before launching his podcast full-time, he started computer science at UT Austin. Dwarkesh, welcome to World of DAS.
0:39 Speaker 01 Thank you for having me.
0:40 Auren Hoffman? I'm very excited. I'm a huge listener of your podcast, of your writings. I'm a massive, massive fan. This is like a fanboy coming out here. You've really interviewed just like tons and tons of super, super smart people. What do you think people are most wrong about?
0:58 Speaker 01 Depends by subject because there's guests on a wider range of topics, but on AI, I'd say that some of my guests have timelines that are too short. They expect the capabilities of AI to be much more valuable, much sooner than I expect. This is not all my guests, but if I did identify the thing I'm most likely to disagree with, some of my most recent guests, that would be it.
1:21 Auren Hoffman? You're not only an interviewer about AI, but you're a user of it, right? You're using these tools. It's funny, like every time I'm using them, I don't know if you're running the same thing, like I run into some sort of limitation where I think it can do X, but it can't actually do that yet. Now that X keeps moving, so every few months, but for some reason, I'm constantly thinking it can do X now when it can't. I don't know if you're running into similar or something similar.
1:47 Speaker 01 A hundred percent. Very similar issues. That's honestly made me update my views around the actual value of AI models today. I think sometimes people have this attitude, and I'm sure you run into this all the time, where they think that the reason, you know, you can look at the revenue of OpenAI and it's on the order of 10 billion ARR, and that's like impressive, obviously, but it's less than Kohl's or McDonald's. Right? So certainly not AGI. And then the justification people have for this is, look, the management at Fortune 500 is too stodgy. They're not adopting these AI workflows, but what you really have is a baby AGI. And it just because people are too lazy to integrate these systems into work environments that they're not being utilized to their maximum potential.
2:30 Speaker 01 I think this is a very minor part of the reason that AI is not as widely deployed as you might naively expect based on how bullish people are about the AI models of today. I think the more important reason is that it's genuinely hard to get human-like labor out of these models. And I've experienced this myself by trying to build these little LLM workflows for my podcast. I'm sure you have these kinds of things set up as well, right? Like getting into rewrite transcripts and find clips and so forth. And I've just been thinking about, well, why is it the case that I still have to hire a human for these tasks? Why can't I delegate these language in, language out tasks to language models? And fundamentally, I think the problem is the lack of on-the-job training. These AI models do have an impressive list of capabilities right out of the box.
3:13 Speaker 01 Any sort of language task you have like this, they'll do a five out of 10 job at it, which is higher than what a default human might do. But the problem is they don't get better over time. At the end of a session, their whole memory gets purged. Unlike an employee, right? Like a human employee will, they might not even be useful until like six months in. But you're training them, they're building up all this context, they're learning from their failures. They're picking up these smaller efficiencies and improvements as they practice a task. So this lack of on-the-job training, I think is a huge bottleneck to unlocking the full value of these models. And I don't see an easy way to add that in right now.
3:45 Auren Hoffman? The puck is moving though, like it is moving, but it feels like my expectations are moving just as fast as the puck. And I don't know if I'm just a weird guy, or that's like what everyone's feeling when they try to use these tools.
3:57 Speaker 01 I think it's fair to point out that our expectations keep shifting. And this is famously, you know, the goalpost we always keep shifting on AI where every year it can do a thing that we said it couldn't do before. But now it's, oh, well, we're still not there because of this X remaining thing. That's a fair rebuttal to say, look, you're shifting the goalposts. But another thing to notice is, look, if you keep hitting these benchmarks, if you do keep hitting, we pass a Turing test, right? Nobody knows. We have models which ace math exams that are made for the smartest grad students in these challenging environments. And if we keep passing through these benchmarks, while at the same time not unlocking the economic value that is implied by AGI, which at the minimum should be able to do white collar work, right?
4:39 Speaker 01 Which is worth tens of trillions of dollars every year in terms of wages paid to white collar workers. If we're not unlocking that while at the same time moving through these benchmarks, then it does make sense to shift the goalpost in the sense of asking, well, what is the remaining bottleneck that we didn't previously identify? Which explains their lack of usefulness.
4:56 Auren Hoffman? What's your thought? Like what's going on or what's happening here?
4:59 Speaker 01 A big part of it is this continual learning, this like lack of on the job training, which I think is how humans mainly contribute value. It's only part of the value humans bring in work is their raw intellect. The rest is this ability to build up context and interrogate their failures and pick up small efficiencies and improvements as they learn on the job. The other big thing is, of course, these models are just chatbots right now. And it's just very hard to be useful as a simple, like you ask it a question and it gives out an answer. That's not how most work happens. And so sooner or later, we're going to see models which can actually interact with the computer in the full range of ways that a normal white collar worker can interact with the computer,
5:43 Speaker 01 including emails, browser, whatever other applications. And of course, it can't do all white collar work until it can at least has a modality to access computer use. Now, there's all these interesting questions about then, how do you train models who are capable of using the computer in this way? Because you're not just doing pre-training on text the way these models are trained right now. But I think that will be a huge unlock as well.
6:05 Auren Hoffman? You have so much insight into just the elite conversation that's happening amongst people. What do you think is like overrepresented and what do you think is underrepresented in this elite conversation?
6:17 Speaker 01 I think people are just really stuck on reasoning or coding or exams which test knowledge. There's famously this exam, this benchmark released last year called Humanities Last Exam. And it was supposed to, well, what it sounds like, it was supposed to be this final benchmark. And why was it supposed to be the final benchmark? Well, because it asked the hardest PhD questions. And of course, if we can answer PhD questions, then you must be AGI, right? Because that's what the smartest humans can do. But that ignores that most work is just not constituted of answering tough chemistry questions or answering math problems. Most work is just mundane, day in, day out. The things we take
6:58 Speaker 01 for granted because it comes so easily to us, you know, just like executing on a project over the course of a month and backtracking when you're, notice you're going on a dead end. But which is not available to these models at all because of the fact that their context cannot accommodate a month long task because of the fact that they won't learn in the way that humans learn on the job because of the fact that they can't use computers.
7:20 Auren Hoffman? It's interesting because like, if I think of like my math, math all throughout college, I think the last math class that was useful to me in my daily life today was probably 10th grade. I think after that, I don't think I've really used that kind of math, even though it's probably important and maybe it's intellectually interesting. Like it wasn't useful in my own life.
7:42 Speaker 01 My day job is a podcaster. So the level of math I'm using is even smaller than that.
7:46 Auren Hoffman? Yeah, exactly. Like, maybe you want to get some probability and statistics and a few other types of things that are out there. One of the things that you're really good at is identifying original thinking and then bring that out on your podcasts. If you had to think through like the commonalities of these original thinkers, what is that? Like, are there any?
8:09 Speaker 01 Yeah, I think one thing that I really appreciate and has led to some of my best interviews, the people I have learned the most from is deep empirical grounding.
8:20 Auren Hoffman? What does that mean? What does empirical mean there?
8:23 Speaker 01 I think sometimes people just have this very like pie in the sky. Here's my grand ideas for how society works or something, you know, like here's the philosopher from the 18th century I read and how it explains everything. And more and more, I'm just like, I don't think that that kind of stuff is that valuable. I think it's more about like, read the main economics papers and like, look at the big trends. Or in the case of AI, measure the trends out in terms of how much, instead of having some grand take about like, what is the nature of intelligence? I think that has its place as well, but also noticing what are the trends in terms of how much computer scaling every year versus how much can be explained algorithmic progress? And what is the sort of uplift that software developers say they're experiencing? Even in terms of grand historical narratives, rather than having some take about like,
9:07 Speaker 01 some Rousseauian take or some Hobbesian take about is mankind screwed? Or what is a primordial nature of man? Just like looking at the population size over time and what has been the growth rates? Like going back to 10,000 B.C. and trying to learn from that pattern, rather than some abstract first principles thought.
9:28 Auren Hoffman? Okay, so it's really just like looking at the numbers and then maybe updating your beliefs based on how those are changing and stuff, or?
9:36 Speaker 01 Yeah, I think otherwise it's so easy to just, especially the kind of line of work I do, where there's no ultimate check, right? You can say whatever and your theory might be correct, your theory might be wrong. And so it's really easy to get yourself in these loops of grand narratives. I found it very useful when somebody just checks that with, no, here's the empirical evidence in other fields. One common failure mode when you get too much in the empirics is just closing yourself off to evidence in adjacent fields, which I think is also a bad idea. Like you're just like, if you give me the numbers on AI, otherwise I don't want to hear that argument in the context of AI discussions. There's also a way to be grounded in when discussing other adjacent fields. So one of my favorite, Carl Scholman made this point to me, he's an AI forecaster and
10:19 Speaker 01 This interesting argument about, look, how do we explain why scaling up AI models makes them smarter? And he noticed there's been this research of primatology by this researcher, Susanna Herculean Huzel, which shows that as you increase the brain size or the brain weight of primates between different primate species, from lemurs or something to humans, their neuron count increases linearly, whereas with other kinds of species, like rodents, that doesn't happen. Okay, well, that is like a very interesting empirical result, which like tells us something about maybe what happened with humans is that we finally landed on a scalable brain architecture and then we were in a evolutionary niche,
11:01 Speaker 01 which like rewarded increases in intelligence. Anyways, that is like a sort of like totally different field that you're learning about AI from. You still have some real thing that you're talking about rather than here's what this 19th century philosopher had to say about will and intelligence.
11:17 Auren Hoffman? If you think of like these big thinkers that are out there, there's this one kind of critique that it's hard to be a big thinker these days because you have to like specialize in a very, very narrow PhD. But then you kind of like meet all these like random people out there, like whether it's like Peter Thiel, Mark Andreessen, Tyler Cowen, like, and they seem to be really big thinkers. They seem to be thinking like across 50 different fields. And I would put yourself as one of those people as well, who you just have like real diverse interests, real knowledge across many, many, many things. Are we losing the big thinker? Are we losing the person who can think across things in today's world? Or are we not?
11:57 Speaker 01 I think all the people you mentioned are super smart. Fans of all of them. But I think those are great examples because I think those people are especially good in domains where they're Which they have like genuine understanding of and that understanding often comes with real tangible. I don't want us to just say numbers because numbers is not what I'm referring to. For example, I think Tyler is like, is he a big thinker? I mean, obviously he's a big thinker, but it's because he's a small thinker in a million different things. In any given subject, he will be able to go toe to toe in terms of, look, if you think this is why Japan's economy cratered, well, how do you explain this other pattern? How do you explain this result from this other paper? Whereas, I don't know, I feel like
12:41 Speaker 01 some of these other people in that list will just have this like, I read this book that was written in the 40s and this explains how the modern world works. And I think that's basically useless. And I don't think it really does. And if you like really get into the nitty gritty, just like this, it's just a bunch of words, right?
12:56 Auren Hoffman? I thought when I first kind of into you, I kind of got into you because you were like talking about AI. But you really do have these diverse interests. And one of the things I love the most about some of your podcasts were like, you had this like whole series on with Sarah Payne on Japan and you brought like Ken Rogoff on Tokyo, Japan, like, why is Japan so interesting to you?
13:15 Speaker 01 I don't know if it's a central interest of mine, but the reason I find the Rogoff interview so interesting, or his claims about Japan so interesting. So Japan famously had in the early 90s, it had a famous crash. It was actually after World War II, obviously Japan got decimated. I think in 1950s or 60s, Japan's per capita GDP was like 20% of Americas. And then by 1989, Japan's GDP per capita, at least in like pure nominal terms, is 125% of Americas. Which is to say that it went from being this destroyed nation to richer than America on a per capita basis. And then it crashes again, and it hasn't
13:59 Speaker 01 recovered for the last three decades. The GDP per capita in Japan is now less than 50% of Americas, which is just like a shocking. And then the question is, why did that happen? Ken Rogoff, who is a former chief economist at the IMF, and he was on my podcast, and he said, this was actually the result of interventions the Japanese government made at the behest of US pressure. Specifically appreciating their currency and deregulating their financial markets faster than they should have. And this just caused a huge financial crisis. His claim is that unwinding financial crises is just a super tricky problem because it's not like near term shock where you
14:39 Speaker 01 can, we had the 2008 recession, eventually recover, supply recovers. But in this case, you just have so much bad debt that needs to be unraveled. Anyways, a big update from me from that example is just that how bad financial crisis can be, that not only are they a short term pain, but they can permanently alter the growth trajectory of a country. Now, many people have responded to that claim by Rogoff by saying, look, he's underestimating the impact of demographics. That Japan is just a much older society. And if you also compare Japan to other countries in East Asia, like South Korea, South Korea is obviously wealthier than North Korea or China on a per capita basis, but it's also not doing as well as America. And so this could have to
15:18 Speaker 01 do with the nature of high end electronics was just not the most competitive industry in the 21st century as it was in the late 20th century. So maybe it's more to do with demographics and what things are exporting. But maybe a bigger question I've been thinking about in Japan is just one example of this, is how much of the world is simply not doing that well right now? We're in America and America's been doing great for many, many decades. And we're just quite used to the stock market. Like, stock market goes up, there's some amount of stability, new businesses are created, there's growth, there's new technology. And as much as we complain, that's been the pattern. Whereas you just like, look at, look at different parts of Europe, you look at parts of East Asia, you look at even China now. It's just remarkable how much of the world really does seem stuck.
16:02 Auren Hoffman? It is interesting because like in 2008, the GDP of Europe is like roughly the same as the US, maybe has more people. So per capita, it's like slightly lower. And then in that timeframe since then, 17 years since then, we just see a wide divergence between the two. And is it because maybe they didn't handle the crisis better? Or I don't even understand why that divergence has happened at such a rate.
16:28 Speaker 01 I have no idea either. I feel it's quite mysterious. They also missed out on the internet wave, which I think has been a huge part of why, between the internet and between energy abundance with fracking, I think has been a big like source of tailwind. I mean, that's quite interesting in itself, that if not for two or three key inputs like this, America would also be in a pretty bad position. Like, imagine if fracking was not a thing. America would be in a pretty bad position, or imagine if big tech didn't end up being a thing. Our pensions would be in trouble. You subtract that from the S and P returns and it's a very different picture.
16:59 Auren Hoffman? That just may come with like, okay, if you over-regulate too much, you might just miss a wave that you don't know about, like the fracking wave, which could have hit Europe but didn't because of, you know, a whole bunch of different reasons there.
17:11 Speaker 01 Yeah, totally. It would be curious to hear from economists about what they think the main causes of Europe slowdown. I mean, one that I hear often is the labor market inflexibility. The fact that it's so hard to hire and fire people really limits the dynamism of those economies. I don't know if you have a theory on what's up with Europe?
17:25 Auren Hoffman? I don't know enough, but my guess is that's not totally it because it's easier to hire and fire, I believe, in Eastern Europe than Western Europe. And so you would just think there'd be more arbitrage. It could be like in the US where it's like harder to have unions in the South than the North. So more jobs move to the South, but it just doesn't seem like it's happening at that rate there.
17:45 Speaker 01 Well, it's also crazy. I think in Italy, this might be wrong, so somebody will let me know, but I think productivity per hour has not increased since the 90s or like some ridiculous amount of time, which is crazy to think about, given the fact that we've had all this technology in the meantime, internet, advances in logistics and energy and whatever. And the fact that like the net impact of that is still zero, presumably because other things are reducing productivity. This is also a pattern in China. So since 2008, if you just look at aggregate, the way you calculate productivity is you just look at output, you look at how much the inputs increased, capital and labor, and then whatever's not accounted by that must be the productivity number. And since 2008, I think China's total factor productivity has also been flat.
18:28 Auren Hoffman? Oh, really? I didn't know that. What's the theory behind it?
18:31 Speaker 01 My explanation is that they did a huge stimulus after 2008, so their reaction to the financial crisis was... This clearly shows that just market-driven growth will lead us to some sort of financial crisis. We need to manage this investment. And so they did a huge investment boom, which has now resulted in these ghost towns, as they're called, or the huge leverage that local governments, building out these huge infrastructure projects that then are now becoming less and less marginally useful. Because it wasn't directed by the market, a lot of this investment, some of it was, it just wasn't as productive. And so the aggregate productivity numbers are actually surprisingly flat.
19:07 Auren Hoffman? Yeah, I wasn't in the workforce 40 years ago, but I do believe that at least today, there are a lot of people, a significant minority of people, at least in the US, that are negative productivity. And I don't know if that has gone up over time, but I believe it has. I believe there is this weird, like, in some ways, like, companies have this bargain with the government and with society, where they have these extra folks there and they keep them there. And there's actually a lot of people who are not just zero, but they're actually making the company worse. It's just like hard to see it in the numbers that there's all these people who are way more productive and there's people who are dragging them down.
19:50 Speaker 01 What is the sort of market explanation for why wouldn't companies just fire? Or like, why wouldn't the companies that decide to fire just outcompete the ones who don't?
19:57 Auren Hoffman? I don't know. I don't totally understand it. But if you just like, even if you are like the most capitalist, you know, which is like the private equity firms and stuff, you'd think they're the most capitalist out there. They just want to win. If you actually interact with firms that are like run by private equity firms, like, they're not run better usually, in my opinion. And it's unclear why, like, they're still bloated. They still have all this bureaucracy. They still do like all these random things that probably don't add value, probably add negative value, all these like weird committees internally and stuff, and all these weird processes that don't probably help things and just kind of like make people move much, much slower. I'm really not sure why. Maybe it's just ingrained in people's head.
20:41 Auren Hoffman? If everyone's doing it, then it doesn't like really matter to outcompete if enough people start.
20:45 Speaker 01 I've never worked a real job, so I have like no sort of firsthand experience of. My intuition is that there might actually just be like some surprising amount of value in these people. One, because I think markets work and I think it would, unless there's some other reason this distortion is created. I think it would be corrected for it, but maybe the other is the sort of Chesterton's fence kind of thing where even though we don't realize why the fence is there or what these extra people are doing. Now, I mean, obviously there are sectors in which this is not the case, like NGOs or something, right? Or government, it'd be very easy to understand why there's not a direct incentive to have efficient headcount. But how many extra workers does McDonald's have or Walmart have? Or a lot of these firms that employ a lot of people.
21:26 Auren Hoffman? I would say more in the white collar than blue collar. I think in blue collar, they're actually, there's a direct correlation to the output there. Whereas in the white collar, I think it's very mysterious.
21:39 Speaker 01 Maybe the explanation for this distortion is that everybody knows that, say, 20% of your employees are net negative value. But in white collar work, since there's no way to identify who those 20% are, you'd be worse off by just indiscriminately firing or firing based on the signals you think you have. And so maybe it's a measurement problem. So it's rational for you to keep these people on board because the remaining 80% are worthwhile that you can't just randomly get rid of people.
22:03 Auren Hoffman? If you remember like four years ago, when Meta started doing all these layoffs, they did it a very small percentage at a time. They did many, many layoffs, each one very small in the single digit percentages. And everyone will tell you that's the wrong way to do it. Everyone says you should cut big and then announce to the team, well, never doing a layoff again because it's so bad for morale. And maybe the theory is like, well, they just did, they couldn't cut that big because they had no idea who to cut. And so they had to do these scalpels very small over time to actually make it happen. And okay, we'll remove this layer of middle management, see what happens and then kind of adjust and then move this other layer and see what happens and adjust, et cetera.
22:44 Speaker 01 Yeah, interesting, but more like liposuction instead of amputation.
22:48 Auren Hoffman? Yeah, exactly. But if you lipo enough, then you have the same effect.
22:52 Speaker 01 Right. Instead of Elon's companies, I like Elon saying is you should have cut enough that.
23:00 Speaker 01 You have to hire people back. This is the same thing of like, if you've never missed a plane, then you were arriving to the airport too early. But yeah, potentially just as some sort of like pain threshold that you're not willing to go through.
23:12 Auren Hoffman? One of my friends is an investor has a similar theory about scams, which is like, you need to have a relatively decently high scam rate where people scammed you out of your money. Otherwise, you're not believing enough in the future.
23:26 Speaker 01 Hmm, interesting.
23:27 Auren Hoffman? And so the theory is like, actually investing in Theranos or FTX, like, that's not necessarily a bad thing, as long as you're not doing too big of a percentage of your investments in these things.
23:39 Speaker 01 I think FTX might have cost more money to investors, not in the raw investments they made in FTX, but in the in the risk aversion that that created. Not just like normal risk aversion, but like specifically, I noticed this in AI where the big people who have a lot of money are making these bets in wrapper companies like Cursor and Perplexity and so forth, which I think might be fine bets, but are not to me as compelling as, because I believe that future AIs will be so much more powerful, given these unlocks that are coming down the line with computer use and continual learning. I think that's sort of like the things underlying that basic research or that the unlocking of this key capabilities, rather than building wrappers on top of the existing abilities.
24:22 Speaker 01 I think that building that base layer is more valuable, but people are not willing to invest in it as much as it would make sense to me too, because.
24:28 Auren Hoffman? Sure, but just to double click on that, because there's not that many companies that are building like the real underlying layer that's out there. So wouldn't that just like really confine the number of things you can invest in there in the AI world?
24:44 Speaker 01 It would in terms of just if you need to want to invest in a foundation lab. But I think there's going to be, well, there's the physical compute itself. You know, you could invest in the components that go into building the chips. Nvidia is probably like priced in, but the HBM's, SK Hynix is whatever. Yeah. Then I think there's like the other components to trading, which are not raw compute, which are, I don't know, like people who are building data sets or something, or have novel techniques there. I guess if I had more ideas, I would stop becoming a podcaster.
25:15 Speaker 01 I certainly feel like right now, just pouring money into, because some people, the mentality like, "Look, we want to be grounded and therefore we want to see raw user numbers." And who has a lot of user numbers? Well, just rapper companies, right? Because rappers are who have users. And then they have inflated valuations, which might be merited in the sense that AI is going to be such a big deal that even they will be worth more. But I think it's like relatively less undervalued as compared to the other parts of the stack, which don't have a clear like monthly active user type flashing light.
25:44 Auren Hoffman? World of DAS isn't just a podcast. World of DAS membership is a private invite-only community for founders and executives of important data and AI businesses. It's a place for off-the-record, high-trust conversations focused on the biggest questions data leaders are facing. The community is about peer learning, curated intros, and access to private events. Everyone in the group is building or leading something meaningful in data. If that sounds like you, apply to join at worldofdas.com. That's world of D A A S dot com.
26:26 Auren Hoffman? The problem is with some of these research things that are out there, their valuations may have zero ARR and their valuations are the same as another one that has 100 million ARR and growing, you know, 100% year-over-year. And so it's hard to know where you should put your money. If one was like super cheap, yeah, it's great. And maybe you should just be giving your money to Hinton's Lab or something like that and you donating it somewhere there. Are you an angel investor? How do you think about like that side of things? Like, because you have a very unique view of the world. In some ways, you could be a great investor. When I asked Tyler this, I was like, "No, I don't invest. I have no desire to do it. I invest in people, like giving money away to people and seeing what they can do with
27:11 Auren Hoffman? it, but not to make a financial return for myself."
27:15 Speaker 01 It's also crazy how much Tyler has accelerated so many different ecosystems. But even within AI, which is the one I have visibility on, and which is, I don't think his primary focus at all, there's many people I know who are doing very interesting things, who in some way... Can source themselves back to Tyler and the initial 20K career grant that he gave them. So if he was in a position to receive the financial dividends from all the surplus value he's created, I think he'd be a multi-billionaire.
27:42 Auren Hoffman? I honestly, I think for most people, when they make billions of dollars, then they just spend it on other things like respectability or other, you know, status or other types of things. I feel like Tyler has, he's gotten that billions of dollars of status and respectability back in some sort of way. The whole, the whole ecosystem really appreciates him in a way that maybe wouldn't appreciate him if he was just a billionaire.
28:09 Speaker 01 This reminds me of that quote that Patrick McKenzie has about being a salary man in Japan, where he says, "You might be wondering, why are Japanese salary men willing to work so hard for so little money?" And his answer is, "In other parts of the world, you work hard, so you get paid a lot, and you spend that money to buy status. And Japanese conglomerates have figured out a way to cut out the middleman and just pay you in status." But anyways, it's your original question about, I gotta figure out something to do with the podcast ad money. So I am not actively doing it, but I have friends who often start or are working on interesting things. And maybe like this year, I've started seeing if I can be involved in some way, but just if something falls in my lap, basically.
28:51 Auren Hoffman? You said, "Most people are kind of sleeping on like the collective AI advantages." Like, how do we shift our mental map from these smart tools to the smarter, let's say, ecosystems?
29:03 Speaker 01 I think people are underestimating, when you think about AI, you're just like thinking about the intelligence of an individual copy of a model. And I think this leads people to ignore the unique advantages that AI have, which have nothing to do with the raw intelligence, but the fact that they are digital. A lot of the value humans have also comes from the fact that we can cooperate with each other, we can form firms and institutions and governments and so forth. This ability to cooperate and coordinate will be so much stronger with AIs. One obvious way is, okay, suppose this continual learning thing is solved, right? So suppose AI can actually learn on the job.
29:36 Auren Hoffman? It can have memory and et cetera.
29:39 Speaker 01 Yeah. And I'm not sure if like memory really captures what this is. I think we like really don't even have a good mental framework for what it is humans are doing that makes you so much more valuable in six months after you start the job than you are day one. But suppose we get that going in AIs. A human who learns on the job is upskilling in the specific job they're at. Suppose an AI model is widely deployed through the economy and it's learning how to do different jobs across the economy. It can learn from the experience of all of its copies. So a single model is potentially learning how to get better at every single white collar job to begin with. And even if there's no more software progress after this point, right? We're not making like found some new way to train models.
30:20 Speaker 01 That alone is potentially enough for like a what looks like a broadly deployed intelligence explosion. This thing is not functionally becoming a super intelligence. And that's one of the many ways you could have these sort of like collective advantages because of the fact that they're digital that humans don't have. Yeah, I don't know how we make people more aware of it. Or I think it's probably hard for even us to sort of anticipate in advance just because predicting the future is very hard. What the implications of these collective advantages are, but you can even lower bound it to these kinds of really wild outcomes.
30:49 Auren Hoffman? If you think of our listeners of this podcast, RoboDass, so many of them lead data companies or work at data companies. And we haven't yet seen like an explosion of revenues to these data companies from the AI boom. What would you be thinking about if you were working or leading one of these data companies?
31:09 Speaker 01 What kind of data?
31:11 Auren Hoffman? People have so many different kind of random things that they sell data about some widget. I mean, usually data is about four nouns, right? Like either about people, places, products or companies, right? And then they could be crossed for each other or sometimes crossed with time and price. So you could have data about molecules, it would be about a product or data about some person out there or what they're doing or whatever it might be.
31:36 Speaker 01 I don't know that much about this industry, so I don't know what the main use cases for this data are. Or what is the theory of why it would intrinsically become more valuable with AI? My intuition is that it's like, it's not that useful for AI training. For AI training, I think it's just like genuinely getting the raw. It would be much better to have like, here's how you write software better, kind of building environments made for AIs to practice in kinds of stuff. Not like here's a couple hundred million tokens of spreadsheets. Then there's a question of, okay, well, what else is that data useful for, for AIs? And I'm not sure, like, maybe there's an answer, but yeah, what's the theory of why it's supposed to become more valuable with AI?
32:13 Auren Hoffman? I don't know that there is a theory, but I think a lot of people are having, most of these data companies today, they're way more profitable than they were a couple of years ago because just like creating the data is easier than it was before. And so they've been able to get costs out and they maybe haven't yet passed those costs on to their customers. So there's this big profitability delta, but at some point they're all in a very competitive industry. At some point they'll probably start passing on those costs to the customers and then they won't be any more profitable than they were, let's say a few years ago that's out there.
32:49 Speaker 01 Who are their main customers right now?
32:50 Auren Hoffman? I mean, every data company is different. So you just think of like any type of data company you've ever thought of. Okay, data that sells to real estate, data sells to hedge funds, data that sells to marketing, data that sells to academia, data that sells to, so so many different kind of data companies that are out there that sell. Part of the problem is it's hard to use the data. Sometimes you're selling like software companies who then take that data and like put it in a product, make it sing in some sort of way. The number of users has not gone up very much.
33:21 Speaker 01 Unfortunately, I have no insights on, which is tragic, given the title of this podcast, but.
33:26 Auren Hoffman? One of those I really wanted to ask you about, when I think of you, I think of like an incredibly curious person. And in some ways that is the it factor. That's one of the things we really want more. It would be great if we could have 10x more people who are really curious in this world. Is there a way to manufacture that somehow?
33:45 Speaker 01 I think curiosity is often driven by meditating on your own confusion. I think it's very easy to just like keep reading stuff or think you're learning and never pause and ask, wait, do I really understand something? I don't know. I have this experience all the time. I don't know if you have this experience where I thought I understood something and then. I've been reading a book about a topic and I'm a dozen hours in and then somebody asked me the most basic question, like, how does Concept Day work? And I'm like, I have no idea. And I wait, so then like, what have I been learning? And I guess I've become more and more aware that there's these huge holes in my understanding. And I had to become more aware of it because that's how I prepare for my podcast.
34:28 Speaker 01 When I'm interviewing somebody, I'm just trying to notice, wait, what is the stupid question? Or the basic question that I still feel is a big crux? I don't feel fully clarified on. And then once you notice it, you can't stop thinking about it. I just get super excited. I'm just thinking about the upcoming interview and I cannot wait for them to tell me. If they don't give me a good answer, I'll like, keep passioning them because like, I really want to know the answer. How do we make more people curious? Putting people in a position where they can notice their own confusion, the blind spots in their understanding. One way, this is a sort of micro solution. I found it really helpful to have L L M's act as a Socratic tutor, which is to say that like, you give them a problem that says, explain Concept Y to me. Do not move on until you're fully satisfied that I've understood the concept. Ask me test questions, et cetera.
35:14 Speaker 01 Do not move on until I've understood the relevant sub-concept to your satisfaction before explaining more. It's like such a much more effective and faster way to learn. And you like, realize how much.
35:23 Auren Hoffman? That extra sentence of do not move on, quiz me, et cetera. That actually is the key sentence, not the first one in a way, right?
35:31 Speaker 01 Yeah. What's your answer? How do you stay curious?
35:34 Auren Hoffman? I don't know, but I think you're right. But sometimes my learning is not very, you, you listen to a whole book or you read a whole book or you, whatever it is that you have, or you're listening to 16 podcasts about why. And then six months later, you don't retain anything. So maybe you did learn during the moment. I don't know. Maybe I'm actually not learning during the moment from reading or my retention is not high. I have a list of books that I love. I have this list of 42 books that I send to people. Recently, one of my friends asked me about some of them. You know, I read this book 10 years ago and I put it on the list because I thought it was like one of the defining books that someone should read. And now I barely even remember what's in that book.
36:15 Speaker 01 This is another one of those like boring strategies that I recommend to people, which is like, not some grand, this will, this is like, is like a boring thing, which works. It's spaced repetition. It's crazy how well it works. It works so well that I regret all the years I've been doing the podcast without using it. Because I feel like I would have actually learned something from all these interviews that I did before starting it. Which has just like sort of totally slipped through my mind. It's crazy how often I'll make spaced repetition cards for stuff.
36:44 Auren Hoffman? I don't exactly know what you're talking about, so why don't you explain it?
36:48 Speaker 01 Basically, you make flashcards and then there's a software. And what the software does is every so often it'll resurface the card and add quiz you on it.
36:57 Auren Hoffman? So you just spend X hours a week just doing this?
37:01 Speaker 01 Yeah, the harder part is writing the cards because you had to pause yourself while you're reading something. In order to ask yourself, okay, what is the thing I really want to remember here? The key relationship or fact or conceptual handle, which in and of itself increases your understanding and retention.
37:16 Auren Hoffman? It's kind of like the people who are highlighting, back in the day, people would like highlight the book or something like that. And I don't even know why they would do it because they would never come back to it. But essentially whatever you're highlighting is almost becomes a card. And then you kind of quiz yourself on it later.
37:30 Speaker 01 Yeah, one thing the software does is it increases this length of the intervals as you get cards right. So once you make a card, it'll show up tomorrow. Then if you get a right, it'll show up a week from now. Then if you get a right a month from now, then so on. And it's crazy how often I'll be making a card while I'm reading something where I'm like, this is so basic, like, of course I'm going to remember this, but I got to write something down. So I'll write this down. And then you come back to it a week later and I'm like, I got this wrong or I almost got this wrong.
37:57 Auren Hoffman? What's an example? Because I would think you'd want to know, like a fact is not as interesting as like a concept, right? And
38:03 Speaker 01 You can do both, but I would not underrate how important it is to remember facts. In order to understand concepts, like a lot of concepts are just, for example, in AI, right? Like, here's a card you can save. How much has the compute for frontier models increased per year? What's the trend been since 2016? I think it's like 4.5X a year. Okay, so that's just like a raw number you stored. Like, isn't that just like a useless memorization exercise? Okay, well, next time we're debating where AI is headed from here, it's very important to know, oh, wait, like, what is the key trend? And you can like recall this.
38:34 Auren Hoffman? Oh, I didn't even know that. It's 4.5X for the last 10 years?
38:38 Speaker 01 Yeah.
38:38 Auren Hoffman? Every year?
38:39 Speaker 01 Yeah, the amount of compute used in the frontier systems. That's right.
38:41 Auren Hoffman? I actually didn't realize it was that high. That's an interesting compounder that I didn't understand until right now.
38:47 Speaker 01 And that has all kinds of interesting implications that we can get into, but like, that's what I mean is that you can now consider those implications if you have that fact cached. I mean, another thing is it's very hard to learn a field if you're not accumulating these like nuggets of understanding to build. You know, before it felt like I'll be trying to understand a difficult topic and every day I'm trying to relearn the exact same like four keywords or something or four key concepts, and I just wouldn't move on. And now I can like just make the card. It feels like I'm consolidating, so I can actually keep stacking my knowledge up over time.
39:20 Auren Hoffman? Though I mentioned, just like the fact that you make the card, the chance that you're going to retain that goes up by 10X or something, just by the fact you made the card.
39:28 Speaker 01 Yeah, yeah.
39:29 Auren Hoffman? You make a habit of doing this, like?
39:31 Speaker 01 To be honest, I'm quite bad about like, some weeks I'll just totally forget, and then some weeks I'll go through like the whole stack again and.
39:37 Auren Hoffman? Because like, sometimes like you're in a conversation, you probably go to a super interesting dinner with a buddy, and they tell you all these things. Are you like taking notes and stuff?
39:47 Speaker 01 I don't, but then I don't remember them. That's why it's tragic. I'm very lucky to have friends who have taught me incredible amounts of things. And I'm now thinking about how much I've lost from those conversations because I haven't been consolidating them in a system like this.
40:00 Auren Hoffman? What is the process? Do you like decide like, okay, once a week I'm gonna do like a three-hour thing, or you're on a train and you're bored, so you just go through it instead of going on, you know, social media or something, or?
40:13 Speaker 01 That's the idea. I mean, I think the thing is you're supposed to do it daily. I don't. I like sometimes build it build up and then do it once in a while. Did you mean it in the sense of like, how do I make the cards or?
40:22 Auren Hoffman? Yeah, yeah, like you have to make it and then you actually have to like remind yourself to do it. And then you have to decide, okay, I'm gonna stop doing this now because I'm getting frustrated or bored with this or whatever. And then, of course, like reminding yourself is also impeding new learning and taking some other time. It's kind of like working out, right? Working out has some value. You're building some muscle and stuff like that, but you just can't do it the whole day.
40:48 Speaker 01 Right. I mean, maybe the big picture thing here is, look, some of these questions, as I was going through them, I'm like, one part of you is like, why is this information important to remember in the first place? But then you decided to read about it, right? So it must contribute to a world model in some way. If not, why read about it? And if it does contribute to a world model to know how does DNA synthesis work? And what are the main constraints to bridging it together? If that's important, no, why wouldn't you want to remember that? And how are you going to remember that if you don't make some sort of a system like this? To the extent that you find something like this monotonous and you're like, okay, why am I, I don't need to remember these facts? Then you should really ask yourself like, why are you reading in the first place? And I think genuinely, the answer might be-
41:28 Auren Hoffman? Is it just for pleasure? Are you reading just like you'd read for fiction or is it?
41:32 Speaker 01 Maybe that's it. I think in many cases, you're just killing time, but then be honest with yourself. And if it's not just to kill time, then if you're not using a system like this, do you really think you're like remembering the things you're reading? For me, it's like my job to then accumulate the things I'm reading, have them cash in my mind so that if a topic is referenced in an interview, I know what the relevant follow-up could be. Or even when I'm preparing for a future guest, I like have a bunch of ideas baked in my mind. Ask yourself what the equivalent thing in your career or life is. If there isn't one such thing, then maybe like you're just like, go watch a movie instead. Like, reading is kind of boring, right?
42:07 Auren Hoffman? I find that everyone is different about how they remember and what they remember. And so they tend to remember, like for me personally, for whatever reason, I'm just incredibly good at understanding dates and dates relative to one another.
42:22 Auren Hoffman? So if you gave me like 12 things that happened in history throughout time, I might not be able to tell you like the date it happened, but I would probably be able to order it on those 12 almost perfectly. So for whatever reason, that's like the thing I can remember really well, but like just other things I'm terrible at. And then maybe I just gravitate more to those things that I'm good at remembering and I kind of go away from the things I'm not. And that might hinder me or might give me some sort of local maxima that I can get because I like the reward of knowing I can remember it or something. If you're good at, I assume you're good at like remembering a certain type of thing. You like the fact that you're good at it. So at least for me, I like the fact I'm good at it. So I just like learn more of those types of things. And I can impress people.
43:05 Speaker 01 This is why people read more history than I think makes sense too. Everybody has like a hundred history books and then like a handful of science books or something. Because history does have this quality of like, oh, there's like a bunch. It's like really long. There's a sub-history and everything. I agree. It is this sort of adverse selection. I don't know if it's adverse is the right term, but the selection towards more memorable or more parseable things. Whereas I think like often learning is really hard. It's not an unwinding experience to go through dense text that really has good explanatory frameworks for something in biology or whatever.
43:37 Auren Hoffman? Going back to your history thing is really interesting. So I was just on a long drive and I had two books queued up. And one was like the history of chiefs of staff in the White House. And that was the easiest book to listen to. It was like so easy and it was great stories and it was super easy. And then the other one was like this deep analysis of interest rates. And this was a terrible book to listen to on a drive because it was just like, it was very, very, very dense and very hard to comprehend. And I probably shouldn't have done it on a drive. I should have actually been reading it and spending time and stopping. And there's also just different modalities that people want information. And like a podcast like this, probably a lot of people are doing something else while they're listening to this podcast.
44:21 Auren Hoffman? It might be hard to get super deep into something because it's like, you don't want to like stop the podcast and like write it down and stuff. Whereas uh, there might be other types of modalities where you really will want to go more deep.
44:33 Speaker 01 Yeah, 100%. Even in a book like that, reading it versus like stopping to any time some important explanation is made, make the cards, because you realize that you like, you ask yourself the question, you're like, I actually don't understand. I feel like I'm going to do like the rabbit hole of like, what is a specific question I had or something. But there's been many cases where I'm like, okay, here's the question that's like obvious that I should ask about this text. And wait, like, I actually don't understand what exactly, like, how does he explain why this is true? And then you start talking to the LLMs, the LLMs give you a few, a bunch of ideas. This is another great thing about using AIs is that you can keep asking follow-up questions. So there's this famous effect called Bloom to Sigma where this researcher noticed that
45:13 Speaker 01 the effects of having a one-on-one mentor is two standard deviations better than teaching in a group. But you can with, if you know how to use LLMs well, you can basically create a one-on-one tutor for yourself. You don't have to exhaust your list of questions and everything is a rabbit hole. There's like an immediate way in which you can like keep banging the pinata.
45:36 Auren Hoffman? I do think this is where a wrapper can be helpful. Because you came up with this like smart prompt that I hadn't thought of before. But just like better wrapper with a nice UI that like gamifies it a little bit and like has some sort of exploding stars when I get four in a row, right? It makes me feel good about it and tells me how great I am. Gives me some positive encouragement, stuff kind of like Duolingo. Duolingo essentially really, maybe it's just a wrapper on these things, but it's great. Like people actually really like it. Something like that actually really does have value.
46:10 Speaker 01 I've made these kinds of wrappers for myself and I've like stopped using them because the thing about these AIs is they're so general. And so I wanted to do Socratic tutoring and then sometimes I just want to stop and be like, okay, stop, we're moving too slowly. Let's stop doing the Socratic thing. Now I just genuinely answer my question in a normal way, like tell me this. It's a good type of flexibility. So I've still found using the chatbot itself to be the best way to access the full generality of the models.
46:37 Auren Hoffman? One of the things you're doing is really just working on your craft and becoming a better podcaster. I feel like somehow you just like got out of the gate really good. Whereas like most people like over time get much better. They kind of start on a very low basis. And then if they keep working on it, but what are you doing to just get to the absolute excellent point? Is it just preparation?
46:59 Speaker 01 Well, I mean, I've been doing it for five years, so I don't know if I was like, it was good out of the gate. I mean, from the inside, it feels quite slow. Like five years is a long time. Now, it's an exponential growth process. For anybody hearing about it for the first time, they're likely to have heard about it recently, or the average person who's heard about my podcast has heard about it recently. And then, sorry, your question was?
47:18 Auren Hoffman? Are you doing something to make yourself just a better podcaster? Like preparing, but is there some other thing you're doing to just become a better podcaster?
47:27 Speaker 01 The main thing I do is I try to learn more things. I've been writing more. That helps me sort of consolidate my own takes because it's very hard to react to somebody if you don't have your own mental model that you're pushing back using. But fundamentally, I don't think, other than that, there's any other skills that I've been rehearsing. You think that one of the things that'd be involved in being a podcaster, and I'd be curious to get your takes on this, but like, oh, you gotta be able to talk. I'm like, fine at talking. I don't know. I haven't like worked on it. I stutter a lot. I'm not that coherent. I'm talk too fast, which is like a common piece of feedback I get, but I think it doesn't really matter. I just like focus on the content and that makes up for it somehow. How about you? What are you working on to hone the craft?
48:05 Auren Hoffman? It kind of depends what your listener wants to get out of it. But do you think that most of the popular podcasts are way too long? They're much more about entertainment than they are about giving your listener more knowledge and helping your listener be better. So if you just think of like the knowledge per minute on some of these podcasts, I think I've gone down. And maybe what I could do better is increase that, maybe shorten it somehow. And this is what I'm trying to struggle with is how do you give your users? The goal isn't to make it like social media where you want them to consume as much time with you as possible. In fact, you kind of want the least time for the most bang for your buck.
48:48 Speaker 01 The thing that takes up so much time is not even entertaining. It's questions you feel like you have to ask. What's your book about? Why did you write it? Where'd you grow up? Whatever. You know that you don't care about. Who cares where they grew up, right? Their like book is about fucking sociology or like-
49:04 Auren Hoffman? This is where I go back and forth. I'm like, I haven't had any questions to you about things about where you grew up or all the other types of stuff of your background and stuff like that. But if you and I were having lunch together, I might actually find that very interesting. And that might help me understand something about you, or you may tell me something interesting anecdote about your sibling or your parents or something that could give me a better sense of things. So it's very hard to know. Maybe these podcasts like I never do in my podcast, but maybe that's a failing of this podcast.
49:37 Speaker 01 There are situations in which I find like a sort of person mysterious and I want to just like understand what contributed to their mind. But like, in most cases, I'm just like especially interested. The thing that got me obsessed is just the object level of interest in the thing they're studying or have expertise in. And then as they're talking about their own personal life, I'm just like, well, let's get back to, you know, whatever.
49:58 Auren Hoffman? Yep, that's a good point.
49:59 Speaker 01 Like, look, if you figure out a formula where you can actually, I'm probably bad at asking questions about people in a way that reveals interesting things about them. If you've got a formula to crack that in a way that's genuinely interesting to you, I think it's a ton of sense to ask those questions. The bigger problem I have is with people who ask questions that they aren't interested in. If you aren't interested in their personal life, or you don't think it's relevant, but still asking it because it feels like something the audience might want to hear about. I've seen some podcasters do things like that, which I think, like, if you don't find it interesting, why do you think the audience will find it interesting?
50:33 Auren Hoffman? Is there certain people, when I like just meet up with them, have nothing to a podcast, I meet up with them for lunch, where I'm pretty much guaranteed to have like a hundred percent interesting conversation from start to finish?
50:44 Auren Hoffman? And certain people you love where you're really much more in catch-up mode with them. Oh, how are things? How are the kids? What's going on? What's happening? What vacations did you go on? Or something like that. You're still some sort of like real value in those things, too. And you might not be learning as much. It might not be like intellectually interesting, but there's something still just like great for humanity about those things.
51:08 Speaker 01 I sort of feel like I've become less patient over time. I'm like too young to be a grouch, but every time I have a coffee with somebody, I'm just like, I'm never doing this again.
51:17 Auren Hoffman? It's funny, I have a friend who is, I would say, one of the most interesting people in the world. And I think that he designs his life to only have interesting conversations with people. And he has figured it out somehow that he does not suffer uninteresting conversations. And probably for most people, you just have to deal with stuff. And somehow he has just aligned his life. And then he gets more interesting by default because of that.
51:45 Speaker 01 A hundred percent. I think a huge flywheel to the podcast has been that I've met smarter people who have taught me a lot of things. In my personal life, I've sort of way overcorrected from my job. And my new rule in my social life is just nothing can be learned and nothing can be taught. It has to be like exclusively stupid banter and riffing and jokes. 11:00 PM is just a weird time to get somebody's monologue on Chinese EVs. And also just send me the blog post, like you're really bad at explaining stuff.
52:12 Auren Hoffman? In the interest of actually participating, preaching, keeping this short, we have two more questions that we ask all of our guests. One is, what is a conspiracy theory that you believe?
52:20 Speaker 01 One is that I think the war with, I'm not a historian. I don't have like too much at stake in this. I could definitely be wrong. The war with Japan, America's involvement in World War II. I don't think Pearl Harbor was an inside job or something. This is not my level of conspiracy, but more so the preceding negotiations could have just been handled much smarter on both sides. But since the American side is the one that like we had control over, we could have avoided the events that slid up to Pearl Harbor.
52:50 Auren Hoffman? Interesting. I mean, almost certainly, we probably could have, right? Is that a conspiracy theory or?
52:55 Speaker 01 So basically, what happened is Japan invaded China, committed horrible atrocities, obviously. And then we embargoed oil on Japan. The Japanese empire could not survive more than like two years if they ran out of oil. And so they're like, okay, well, we're going to probably lose a war against America.
53:13 Auren Hoffman? This is my main takeaway from the prize, which is one on that 42 list.
53:18 Speaker 01 So we're probably going to lose the war if we go to war with America, but we're like, definitely going to lose the war if we don't because we're going to run out of oil. And I think it was not America's main geopolitical objective to take out Japan. The main geopolitical objective was to take out Germany. And there are multiple like miscommunications where fundamentally the problem was the Japanese government had like a civilian dominated part led by Prince Canoe. And then the military dominated part by Tojo. And Americans would get messages from the Japanese side where we would read them as, oh, they're like, they're double playing us. Sometimes they want to be diplomatic, sometimes they want to be aggressive. Whereas really, it was just the different parts of the Japanese government trying to get an edge in.
53:59 Speaker 01 And we should have given the more accommodating civilian side more room to save face, basically.
54:06 Auren Hoffman? Yeah, I have this theory that basically like the US has lost basically every war battle it's ever been in, in the history of the US. But somehow we still win the wars so often just because we just outlast them. In this case, like Japan, like that was a tough adversary. Our last question we ask all of our guests, what conventional wisdom or advice do you think is generally bad advice?
54:29 Speaker 01 Maybe to tie it into something we were discussing earlier. I don't think reading widely is as good for you as like it's conventionally considered. I think reading specific things with intention and depth has huge benefits, but this is a much slower process where you might spend 15 minutes on a page. And the ratio of times talking to other lungs to reading the main text might be incredibly high. But just like scouring through all this stuff behind me is like, it might do you some good, but I'm not actually convinced it would.
55:01 Auren Hoffman? And partially it's because people are, they want to put points on the board. So they want to read more books or they want to listen to podcasts at 3X speed or go through more things or talk to people about that they read this thing and stuff like that. Or what do you think is driving that?
55:18 Speaker 01 Yeah, it's just very legible to have the sense of I read 50 books this year and less legible of, okay, I got through 20 pages today, but now I like, I can really tell you like how interest rates work. The thing you were discussing earlier about the, you do the thing which feels easier or which like comes easier to you.
55:36 Auren Hoffman? I have this thing where I have a monthly newsletter, I write some things I read, and then I used to list all the books I read the month prior. So first of all, I often realized at the end of it that I would recommend none of the books. If I read five books in the month prior, I'd recommend none of them to people. So that wasn't really like helpful to my audience. And then I also realized like, well, if I wouldn't recommend any, why am I even reading? What's the value of me reading these books? So it's just like, I've tried to reduce the number of books that are, because most of them are just not adding that much value in the end.
56:10 Speaker 01 Yeah, 100%.
56:11 Auren Hoffman? This has been awesome. Thank you, Dorakesh Patel, for joining us at World of DAS. I follow you at Dorakesh underscore SP. I don't even know why you have that handle. I think you need a better handle on X.
56:23 Speaker 01 They won't give me the word cash.
56:25 Auren Hoffman? So you had to do the underscore?
56:27 Speaker 01 Yeah, mistake, but you know, we're locked in now.
56:30 Auren Hoffman? It's too late now. I mean, you have so many followers, so you've got to keep it. I follow you on X. I definitely encourage our listeners to engage with you there. This has been a ton of fun. I really appreciate you coming on World of DAS.
56:40 Speaker 01 I had a lot of fun, too. Thanks so much for having me on, Auren.
56:45 Auren Hoffman? Thanks for listening. If you haven't already done so, please subscribe to World of DAS wherever you consume your podcasts, YouTube, Apple, Spotify, and more. And please help us get discovered by leaving a review and check out worldofdas.com. That's worldofdas, D-A-A-S, dot com. And of course, connect with me on Twitter, at Auren. That's A-U-R-N. Would absolutely love to hear from you.
Previous Morning Consult CEO Michael Ramlet - AI, consumer sentiment and reinventing survey research Next Airtable CEO Howie Liu - Excel vs Access, Founder Mode, and the Future of SaaS Page updated November 1, 2026 · transcript and summary are machine-generated