Transcript: ‘Dwarkesh Patel’s Quest to Learn Everything’

The transcript of AI & I with Dwarkesh Patel is below for paying subscribers.

Timestamps

1. Introduction: 00:01:44
2. How Dwarkesh uses LLMs to remember everything: 00:05:37
3. Dwarkesh’s taste in books and how he uses AI to learn from them: 00:11:50
4. Why it’s important to be an early adopter of technology: 00:17:58
5. How Dwarkesh uses Claude to understand complex concepts: 00:20:44
6. Dwarkesh on how you can compound your intelligence: 00:26:36
7. Why Dwarkesh is on a quest to know everything: 00:28:21
8. Dan and Dwarkesh prep for an upcoming interview: 00:39:19
9. How Dwarkesh uses AI for post-production of his podcast: 1:04:14
10. Rapid fire on AI’s biggest questions—AGI and P(doom): 1:08:51

Transcript

Dan Shipper (00:01:48)

Dwarkesh, welcome to the show. 

Dwarkesh Patel (00:01:49)

Thanks for having me, Dan. 

Dan Shipper (00:01:50)

I'm so excited to have you. For people who don't know you—I assume everyone knows you. But for people who don't, you do the best, honestly, the smartest interviews in AI that I've found. You have really incredible guests like Mark Zuckerberg, Demis Hassabis, Patrick Collison. You've created the go-to show for smart people to learn about AI, but you also branch out into lots of other things like geopolitics and history and stuff like that. It's really great. And you're just one of the people that inspires me to make smart content. So I appreciate you coming on the show.

Dwarkesh Patel (00:02:25)

Yeah, that's very kind of you to say. I mean, I've always been sort of trying to have the conversations that I would like to have— If I was getting dinner with one of these professors or CEOs, what would I want to ask them? And I'm glad other people enjoy them as well.

Dan Shipper (00:02:38)

Yeah, it comes through. And I think it's really fun to get to turn the tables on you a little bit. You've done some interviews, but mostly you're interviewing other people. And I think it's probably on people's minds how you use AI in your work and in your life. And so that's what we're going to talk about today. So maybe just start by giving us a little bit of an overview. How is AI integrated into your work and in your life right now?

Dwarkesh Patel (00:03:06)

Yeah. So, it's actually changed a lot. I remember a year ago—I think it was after GPT-4. Somebody asked me, do you use AI to help you with your research or prep? And I was like, not at all. It's completely useless and mid. It gives you these banal— You ask it, what should I ask? So, it's a professor—and it'll give you these banal— Where did you grow up? What's your book about? Whatever. So initially it was terrible. I think recently, the models have gotten to just the point where, with [GPT-] 4o, or especially with the new Claude models, they're intelligent and interrogative and can consider the context which you provide to them. And so they're still not that good at it. What should I ask this person? Because obviously that's why I have a job, right, so that I can come for the questions.

But for the research itself, for me, at least I try to ingest everything they've ever written—all the rebuttals to their ideas, all the other considerations, and there's often a lot involved. Especially given there's many different fields I try to go deep into. The last interview I just did was with Dylan Patel who writes SemiAnalysis. It's a publication about semiconductors and AI hardware and so on. So there's a bunch you have to learn and I can go through my workflow, but it's incredibly useful to be able to have this thing where I'm like, what's going on here? Can you help me explain this? And then I guess one bigger thing I've been thinking about is ever since I interviewed Andy Matuschak. If your audience is familiar, he's the guy who talks a lot about how spaced repetition and other tools can enhance our ability to learn and how the normal mode of learning, you're actually not picking up that much. If you pick up a random book and start reading, you're not getting that much out of it. 

And I really have found that to be the case, to the extent that, if I'm just casually reading a book, I think I'm basically wasting time or entertaining myself, and I have come up with a couple of different workflows and tools that help me really interrogate and make sure I've reinforced what I'm reading about or learning. And a tool like a language model is very helpful because it gives you the content in another context. It can quiz you if you want. So it's super helpful with that kind of stuff.

Dan Shipper (00:05:28)

That's really cool. I want to start back to front with the stuff you're using it to read. Because I think all that reading is one of the inputs to the interviews, and then we'll get into the interviews. And I'm really excited for both. So let's start with using AI to read and to learn.

Dwarkesh Patel (00:05:43)

Yeah, I can show you my screen, if that's right? So as I was talking about, one of the main things I think is important is, if I'm studying a topic over the course of a few weeks, especially if it's a difficult topic, it's new to me, it's incredibly important that I'm not just casually reading. Because if you're just casually reading, every day you're rereading the same key terms, the same concepts in your start over from scratch. 

So one of the things I like to do, for example, I was recently interviewing Dylan, right? So if I go to his publication, SemiAnalysis, there’s just a ton of lingo and things you have to understand. So, the new one was pretty interesting. It's talking about why nobody has built a huge training cluster yet. And then the first thing I do is just, what are the key ideas and concepts I really need to understand? So I made myself a HuggingFace space. You honestly don't need to do anything like this. It's pretty simple to have Claude build you a HuggingFace space, or if you prefer, what it literally does is like apply this prompt to everything I paste in. So you can just copy-paste that prompt into Claude yourself, but basically, I copy-pasted some of the things in Andy Matuschak's post about how to write good prompts and I just asked Claude to make those prompts for me—spaced repetition prompts. 

So when I do this, hopefully in a few seconds, we'll get something back. Initially, this will just give me some ideas of, what are the key ideas here I need to understand? So, super useful, right? I can zoom in a little bit. So, it's more helpful. So, for the audience who's listening, it's given me a bunch of question-answer pairs that consolidate the key things I need to understand about this post. We can go through the specifics here. I'm sure that the actual specifics of AI hardware will bore people, but a lot of the things where it's like, okay, if you don't get this, you've totally missed the boat here. And so you can start with something like this. 

I added it to my spaced repetition app. Or, I can just look through this and I'm getting a sense of like, oh, okay. Here's what it would take to train a GPT-4-level model on a 100,000 H100 cluster. What are the three main types of parallelism you need to use to train on a big cluster or whatever? And this is for a technical post on other kinds of posts. There might be different kinds of cards that come up in history. It might be a different kind of thing for philosophy. It might be a different kind of thing. So this gives me a lay of the land. I'm sorry—you were going to say something?

Dan Shipper (00:08:15)

No, I love this. This is super interesting. I feel like I can go in a bunch of different directions, but where I want to start is how are you reading and when are you reading? So, are you using this specifically for reading that you're doing for the show or are you just doing this for any reading that you're doing that you feel like is serious and you really want to learn?

Dwarkesh Patel (00:08:33)

Both. So just this weekend I was reading. I forgot the author's name, but it's a book called Medieval Technology and Social Change. And it's about how different things that were developed through the last 1,500 years—technology like the stirrups, how they affected society. And it's entertaining. You can read it. And then one of the things is, okay, did I really understand what's going on here with the relationship he's trying to elucidate? So afterwards, in fact, I have some Claude chats where I was just going through while I was reading it. Let’s see if this recollects, I'm curious. I don't know how interesting it will be to your audience for me to go through point by point.

Dan Shipper (00:09:21)

Do it. I want to know. I'm on the edge of my seat because I have this book. It's sitting on the desk in front of me. And so I want to know what you got out of it.

Dwarkesh Patel (00:09:21)

Okay. So, first, I would just ask it to make some spaced repetition prompts for me. First of all, I read the chapter. I'm not sure I got it. So, just explain to me the chapter about how he says that stirrups created feudalism. What exactly was the connection here? So it's much more condensed here's what's going on here. Basically, if you understand this, it's a useful scaffold so that when you're reading the rest of the chapter, you understand where the pieces fit together. Then I added some—

Dan Shipper (00:09:58)

Have you tried— One of the things that I've tried with this is, because sometimes it doesn't know, especially for a book like that, where it's not that popular. Have you tried— One of the things I do is create a little Claude project and then upload the text if I can find it. Have you tried that?

Dwarkesh Patel (00:10:14)

In fact, let me just— Claude.ai, projects. So, I literally just think— I'm a host of a podcast where I try to ask good questions. My upcoming guest is a geneticist, and I just upload— I get the EPUB of the file. I convert the EPUB to a text using an online converter. I uploaded it to Project Knowledge. Then I've only just started prepping for this guest, but I'll just have a bunch of chats where I'm like, how does he explain what groups made up modern Europeans? It has all the context in there. That ends up being incredibly useful like you were saying.

Dan Shipper (00:10:51)

Yeah, that's so cool. I love that. I love that feature. Okay, wait, let's go back to stirrups and this chat you're having with this book.

Dwarkesh Patel (00:11:12)

Yeah. So it explains that the reason stirrups should create feudalism is because you needed a lot of land, basically to support the kinds of people who become heavy cavalry—the knights need a lot of land in order to have the income to have armor and lances and other kinds of equipment and to train themselves. But a knight is only possible if you have a stirrup against which you can brace yourself as you're attacking with a sword. Because otherwise you're just a Mongol who's shooting bows and arrows. But then there's a bunch of stuff that's confusing here. Why is it so expensive to have to be a knight that you need to completely confiscate church lands in order to subsidize this knight lifestyle? And then on these kinds of questions, the author is dead. But I'm just murky about it. I don't know what's going on. So, I can do just these kinds of things the book didn't even talk about, right? But I can always just continue the conversation with Claude and have it explain what's going on. And so this is just a recreational reading that Claude ends up being super helpful with.

Dan Shipper (00:12:13)

I think that's really interesting. What do you think about books like this as a person who likes history a lot—books that sort of single out a specific thing like the stirrup, and then are like, well, you can trace all this stuff to that one thing where it makes so much sense. But then there are things like, I don't know, Guns, Germs and Steel where like Jared Diamond had that whole thesis about—I can't remember the exact thing, but it's like people in warmer climates, or I can't remember the exact things. But it turned out to be totally wrong. How do you feel about things like that?

Dwarkesh Patel (00:12:45)

Yeah. So my opinion on these kinds of books: There's the sort of concise answer is, yeah, there's ones that do it poorly, but just don't read the ones that do it poorly or something. There is a failure mode for public intellectuals where they initially start off with a discipline and they do some exemplary work there. And then they write an initial broad book that's about how this idea explains a lot of the world and it does incredibly well. And now they're in public intellectual mode. And now the next book has to be, here's my theory of everything. And it's just not that satisfying. So I do worry about those kinds of things, but presumably the reason—I don't know, I'm not into reading 500-page books about things like how this chair physically worked. What's the point of that, right? I do want to understand the implications and maybe they're wrong, but what else are we trying to do here, right? Maybe you just intrinsically care about how the stirrup physically works there? I will point out a couple of examples. 

So there's a lot of interesting topics where you really can't get at the heart of the matter without just considering the whole story. And in fact, a couple of biographies especially stand out in this way, where if you look at Caro's biography of LBJ or Kotkin’s biography of Stalin, it's basically a history of the 20th century or, in the case of Kotkin, even before the 20th century, I think the Caro books on LBJ start off with the Comanche raids on frontier settlers in the mid-19th century or something. And it goes through rural life in Texas—why electrification was such a big deal, a whole bunch of other things, right? Now it's basically a history of the 20th century, but it has a very specific point of view or a specific locus, a character that's moving the story along. And I find those to be incredibly helpful in getting a full picture of what's going on in an era. 

There's a couple other books where they really aren't trying to write a theory of everything. I don't think Caro’s trying to write about the history of the 20th century, but they just can't help themselves. They feel like you really cannot understand the very specific topic I care about unless I tell you everything about everything like, Kotkin's story biography of Stalin starts with Bismarck's career as a military general and how that changed the way that different powers thought about colonialism and the need to modernize. And that's where it starts, right? And it's a biography of Stalin. So yeah, I love those kinds of books.

Dan Shipper (00:15:17)

I think there's a very deep point about the universe being interconnected there, but there's also a really interesting point for people who want to make stuff—write or make podcasts or whatever—because there's this deep fear that everyone has about being pigeonholed. And it's like, well, if I pick this really specific topic, I won't be able to bring all of myself to it. I won't be able to be multifaceted. And it's like, no, no, no. If you just pick one guy, Lyndon Johnson, and really get deep into him, you have to explain everything else about the world in order to explain him. And I love that. And, as a creator myself, that's the thing that I think about when I'm like, oh, maybe I'm getting too narrow here. It's like, no, no, the narrow is actually good. You can find the entire universe in the narrow.

Dwarkesh Patel (00:16:00)

Yes. I couldn't have said it better.

Dan Shipper (00:16:02)

So basically what I'm seeing right now is you're using Claude when you're reading books that you care about learning from, and you're using it a little bit to like prepare your mind for what you're about to read, which I think is a particularly good for like difficult books or for thinking through a particular argument before you go through it, you're asking questions. So it's a reading companion, you're getting more out of the books you read from that. But then you kind of take what you've read and throw it into this card generator.

Dwarkesh Patel (00:16:36)

Yeah. And so that mostly it's just chatting with Claude. And so, let me see if I can find a better example. So, I mean, a lot of topics I just find I've had a vague sense of what's happening, but I don't really get it. And it's super helpful to chat with Claude to make sure I'm on the right track. Dylan has a couple of posts about why packing is a technology super necessary for these advanced chips. I'm not trying to make this podcast all about AI hardware. It just happens to be the last podcast I did. So that's what you're getting. But it's confusing, it's five series posts about how advanced packing works and what the technical specifications are. And I'm like, wait, step back. Why is this necessary? What's going on? All kinds of other questions about when there's questions about how I'm worried about where I might get too deep in the weeds when I'm just explaining. Yeah, basically, I'm just like, how do I think about the broader context of what's happening here? Because I really can't ask good questions unless I have a good mental model of what's going on, what they're talking about. I really get where all this fits together.

Dan Shipper (00:17:56)

That makes sense. And so, Claude is kind of the first thing you flip to when you want to know that you are using it on mobile or using it on desktop?

Dwarkesh Patel (00:18:05)

Desktop.

Dan Shipper (00:18:06)

Okay. Interesting. So you're doing most of your reading and research stuff on the desktop.

Dwarkesh Patel (00:18:10)

Yeah, that's right.

Dan Shipper (00:18:11)

Hmm. And what do you think about Claude being really great right now and— I assume your ChatGPT usage is lower than it used to be?

Dwarkesh Patel (00:18:20)

Yeah. I think these things will keep getting better over time and I think we're just getting in the practice of using these tools. I'll talk a little bit about how these tools relate to my personal life. The post-production process initially was kind of useless, but I did spend a few weekends trying to write a few prompts and create a workflow at the time. It was basically useless. Now it's actually ended up being useful and I can use the same Jupyter notebooks or whatever to get things done. So it is worth investing, even if they don't work perfectly now, to get them part of your workflow so that as they keep getting better, you're getting the returns from that.

Dan Shipper (00:19:00)

Yeah, that makes sense. So, I want to just go back to the Anki card generator, the spaced repetition card generator. So as part of this, once you've done all of that clearing the ground conceptually for yourself to kind of understand the basics of what a guest is talking about or an idea that you're interested in, then you're kind of you're adding to your flashcards, I guess so that you retain the information past even when you talk to that guest. Is that right?

Dwarkesh Patel (00:19:30)

Yes, that's right. I mean, I think the larger mission of the podcast is to consult. Why does the podcast get better over time? And it's because basically I'm getting smarter or learning more things. I'm reducing my ignorance around a bunch of topics. And so if I don't do that I think about all the episodes I did before I interviewed Andy and started using spaced repetition and I just really regret it because I talked to all of these world experts in a ton of different domains. And to be honest, in many cases, I didn't take that much away. I vaguely remember some things. And now that I use it, I can walk you through the kinds of cards I make in these very separate tools I use, but it's totally a game changer in terms of what I can retain. In fact, I think it's not even about making sure I remember what I discussed in a previous episode or what I learned previously. It's more about future learning because I'm sure you've heard the saying about a learning compound because you can use what you've learned in the past to learn future things because they all interconnect. Well, you can't do that if you basically forgot most things you've learned in the past. So, yeah, my learning has a future of other things has become much faster because I have cashed all these different concepts and figures and facts. And so I understand how everything fits together much more. It's not even about the past. It's really about future learning.

Dan Shipper (00:21:00)

I don't know what you use for spaced repetitions. Can we see your deck?

Dwarkesh Patel (00:21:02)

I will point out by the way, as a side note, one use case of Claude that ended up actually being pretty useful. Sometimes you read obscure for a lot of light— I was reading Nick Land’s selected writings about AI and his acceleration and I was like, what's going on? Genuinely, what is his argument? Basically, why does he think that the AI takeover and whatever thing it creates in the aftermath will be good? Because he's a smart guy. I'm assuming he has an interesting argument. So I pulled the PDF of his selected writings. I just asked Claude okay, so why does he think it's a good thing that AI takes over humans? It offers a summary—initially this isn't necessarily that helpful because I kind of did read this in the essay, but what's helpful is that when you go through and I'm like, I respond, I don't get it. What does he think is wrong with human society that you have to erase it? And then he gives an explanation. I'm like, I still don't get it. What exactly are you talking about here? And then here's what I do with the podcast, right? I have the guest on and I ask them, what do you mean here? I disagree. Here's a contradiction—whatever. And going through their writings with Claude and, have I actually found a sort of blind spot in their thinking? Or is this just me being confused by their ideas? It's super helpful.

Dan Shipper (00:22:19)

That is really interesting. It's like you can get down to a deeper level before you talk to them so that you can start there with them as opposed to starting at the surface, which is really cool. I use that too for difficult books, not necessarily for interviewing the author of those books, but, for example, I interviewed Reid Hoffman, I don't know, a month or two ago. And I wanted to talk to him about the kind of intersection between philosophy and AI. 

And he almost became a philosophy professor at Oxford, and was really deep into Wittgenstein. So I read a bunch of Wittgenstein which I hadn't read in a while, and I just used Claude for it, and it was so much better because I haven't taken a Wittgenstein class—or maybe I took one in college a long time ago, but I've read him a lot and there are always those points in those kinds of books where you're like, I think I know what they're saying, but, I'd probably have to go to a graduate school and get a master's in this to really know. And Claude actually makes me be like, oh, I don't need that anymore. Any book I want to read, I basically know. And it just helped me so much in that interview because I could just ask, read really deep Wittgenstein-related questions and he could answer them.

Dwarkesh Patel (00:23:36)

Yep. I think that's totally legitimate. I think some people would be like, oh, you need to read it in the original blah, blah, blah. I think if you care about the ideas and you think the ideas are timeless and not the ideas are not about the specific kind of pros that the original author used, but just generally what is the essence and the gist of what's happening here. If you care about the ideas and I think this is totally valid, right? I don't just agree with the people who are like, no, you need to read the specific syllables that Wittgenstein used.

Dan Shipper (00:24:05)

Yeah. I mean, I'm also just saying, I have the book open and then I just take one of his statements and just throw it in there and then it's like, here's what it means or whatever, which I think is really great. Okay, so you're going to show us the spaced repetition card. So what app is this?

Dwarkesh Patel (00:24:20)

This is Mochi. It's like Anki, but this is the one I use.

Dan Shipper (00:24:25)

Why?

Dwarkesh Patel (00:24:26)

Actually, I don't have any cards today cause I just went through them this morning, but let me give you a sense of what kinds of things I have. So I have if you go through history, recently, I don't know if you can see my screen, how easy this is. Maybe I'll zoom in a little more. I've been planning on interviewing David Reich, who is a geneticist who explores human origins. And these are especially cases where just reading the book, I'm like, I would have totally forgotten. He names all these different ancestral groups and how they combine and in what eras when did the Yamnaya people come through Europe? When did the Anatolian hunter gatherers wash over Eurasia—all these things that were just like you read it in one ear, it goes out the other one, unless you make cards for it. And so I made a ton of cards about this kind of stuff. So there's examples of that here. It's especially useful for hardware and technical things. So here I feel like if I don't make cards, I'm just constantly relearning the same things cause I didn't learn the lingo in the right way. 

First, it's not just about learning the terminology. It's about understanding the underlying concepts. Let me give you a good example of that. So maybe I'll sit back and I'll explain, like I go through these cards in the morning. Maybe you can see what it kind of looks like. If I do the Cram cards thing, I can go through them and right now, I'm like, I remember this, right? I remember it's the first one that came up randomly, but it's multi-query to not have to use huge Kv values and then sharing Kv values between layers and using local attention. And that's the answer. Now it seems sort of trivial right now because it's just three things, but I would have totally forgotten about this if I hadn't made a card for this as soon as I read the blog post. And then it's just like I've wasted my time in the future if I'm learning about these technologies in a different context. I just don't have the connection to what was happening here to connect it to, right? If I go to a different category, if I go Cram cards—this is the white thing. I would have totally forgotten about it if I hadn't made these cards. I'm just a big fan right now. I've sort of become a spaced repetition fanboy these days.

Dan Shipper (00:26:55)

How do you think about the usefulness of spaced repetition in a world where any of these questions is possibly pretty much answerable with Claude or with one search?

Dwarkesh Patel (00:27:05)

Yes. So, I think it's about not necessarily remembering this information, but when a future thing comes in, you understand the conceptual— In fact, let me give you a good example of this, right? So I remember sometimes I actually make cards about facts that I don't even understand at the moment, but in the future, as I learn more about the field, as the sort of territory becomes more clear, the things I said in the card make more sense to me. So if I was reading some of Collin Burns papers and so I made this card about why Collin Burns thinks that alignment is a tractable problem or understanding what the model thinks is a tractable problem.

And at the time I wrote things down about features in a linear space. space. What does that mean? Or we can sort of see features in other sorts of categories. And at the time I have no idea what this means, so I'm just going to write it down because I read the blog post and there was no point in reading the blog post if I'm not going to make the card. Later on, as I learned more about how the residual stream model of how attention works works—what that is and so forth— this card made much more sense to me in the future, but I would have just totally just memory-holed, or not even memory-holed, I would have totally forgotten this content, which required future understanding if I hadn't made a card of it. And then when I see the card again in the future, I'm like, oh, this is what Collin Burns meant. Now that I understand how attention works, this is what it means.

Dan Shipper (00:28:35)

This is really interesting to me. So, I want to get into some of the ways that you use AI for interview prep because I think we've mostly covered the reading stuff, but before we do that, I just want to understand what is driving all of this? It feels like you are just consuming massive amounts of information and turning that into knowledge in your head. You have this sort of overdrive of curiosity, which I actually resonate with a lot. I'm surrounded right now by books. And I'm just sort of curious, for you, what do you think that's about?

Dwarkesh Patel (00:29:15)

I think I really just want to know everything, right? I don't know how to express it. There's a beautiful passage in a Will Durant book as he's turning 90, where he's writing a memoir basically of his main ideas called Fallen Leaves. And there's a passage on philosophy where he says as you get older, maybe with all the philosophy and history I've done, I've reached some plateau of higher understanding and clearer insight, or at least I've understood that such a thing as possible. And something that just resonates with me. I don't know. I just find that idea really appealing. I'm nowhere close to it, but I just hope in the years to come, that'll just be a thing that— I also really admire people I've had on the podcast who do have these self-consistent and really deeply interrogated world models. I've interviewed these guests and some of them are people— A couple of names come to mind, people like Carl Schulman or Tyler Cowen or Byrne Hobart. It feels like they've really read everything and everything you know is a subset of what they know. And I just find them to be super compelling as thinkers.

Of course, there's many things that can still be wrong about it. I'm not one of these people who buys there's a thing where you just know everything and you can never be wrong. You always have blind spots, but their ability to—which you can see when you talk to them to connect anything you ask them about, and they're like a Claude 6 in the sense of, you start talking about why a fraction of finance is a percent of GDP. I remember asking Tyler this, and he has a right off the cuff, just a super interesting answer that connects a bunch of different disciplines. You ask Carl about how fast AI hardware could grow and he’s just done this sort of Fermi estimates on how fast algae bloom and how much solar power they consume and how fast TSMC is making. I find that sort of compression of the input they've ingested over their lives and they can not only do they know that stuff, but they can really connect it in a really interesting and compelling novel way. I find it super compelling.

Dan Shipper (00:31:34)

And in terms of developing your own worldview do you have that anywhere where you're creating some sort of living document or is it just all in your head? All the stuff that you're learning, obviously you have the cards, but that feels more like dots in space rather than the ways that they all connect and how you think about everything altogether as a system.

Dwarkesh Patel (00:32:00)

I think I've been trying to do more of this recently. And now that I've sort of built up an underlying vocabulary or understanding because of the podcast, it makes sense to do more of this. Something I've been doing recently— Let me pull this up. I've only just started. Hopefully there'll be more by the time people are looking at this. But I've started writing riffs on different books or things I read. And if I go to it's basically on my website and so I can read a book and I have questions or I connect with other things I've read. I remember, for example, when I was on Steven Pinker's The Language Instinct. He was writing the book before the FOXP2 gene, that can help explain human language, was found. And so he has all these observations that are then later explained by the FOXP2 gene. And so I can just sort of connect with what you're talking about, I can do that by riffing on other people's ideas. I actually am curious. Do you have suggestions on what I should be doing? Maybe I should be writing more blog posts or what do you suggest I should do?

Dan Shipper (00:33:04)

Well, let's see. That's a good question. So well, before we get there, one of the things that this reminds me of, I think Claude is so good for reading old science books because it can tell you what's outdated and what's not. I do that all the time and I love that. I love that little thing. But yeah I think basically developing a worldview is— You have to just try, you know? And you try over and over and over again. And I do think blog posts are really good for that, especially for me. I have to write every week and so I'm forced to take a view on something. And generally, if you're intellectually honest, you want one post to somewhat agree with the last post and your audience will call you out if you're just disagreeing with yourself all the time. So, you're kind of developing a worldview that way.

But for me, right now, I'm actually— My big thing this quarter is, I just have these ideas that are simmering that are sort of the relationship between language models and some deep philosophical questions that we've been talking about since Plato, which is the appearance-reality distinction. And how do we know what's true and what's knowledge and all that kind of stuff. 

I think there's a lot of overlap there and it requires it's going to be a 10,000-word post or something like that. And so what I'm doing is I just have a Claude project with— I have all these little notes and riffs and stuff. And I'm just going into Claude and being like, hey, what's the thread here? What's going on? Can you help me figure out there's something in me that I have all these little ideas for, but I can't quite put it into an argument that all makes sense. And I think just honestly, sitting with that for like a couple of months, I will know what's in there, but there's something in there. And yeah, it's cool.

Dwarkesh Patel (00:34:52)

Do you make a Claude project to be, here's some of the things I'm thinking about. How do they connect? Or, how do you keep track of those things over those months?

Dan Shipper (00:34:59)

Yeah, exactly. I'll just show it to you. Let’s just pull it up. So, okay. So if I go into Claude, I have a couple of different projects. One project is “Seeing Like a Language Model,” which is the title of this big post, whatever it is. Now there is Zen in the Art of Motorcycle Maintenance. So this is a book that I'm reading as prep for this piece that I'm writing. I've read it a bunch of times before, but now I'm doing a little bit of a deeper read. And so I have the whole book uploaded and then I can ask questions. Then I have another one that I love called “My Psychology,” which has a bunch of journal entries goals I've set for myself over the years. And then also things I've observed about my psychology or things I'm working on, like little aspects of myself that I'd like to grow or change. And so when I'm making decisions or thinking something through, I just go in there and it can reference all that stuff. So it knows who I am, which is really cool. So in seeing a language model, let me see if I can pull it up in the projects directory. So, basically I have this one note in Apple Notes, which every time I have a little thing come into my head, I just put it in there. 

Let me see if I can find it for you. I just throw it in here. And this is huge and messy. And it's different quotes from different books and like just different ideas that come to me off the top of my head as I'm walking around. And I think that there's a thread here and all of this stuff. they're all— I can see how they're all related, bu, I can't quite pull it out. And so what I've been doing is I just throw it all in here. We have all the quotes and all the ideas and fragments. I have a little bit of a draft, like an intro, and then I have a chapter of a book by Richard Rorty that I think is really good called Pragmatism as Anti-Authoritarianism that kind of sparked this whole thing. I read a chapter of that book and then I was suddenly rereading like a bunch of Plato and Aristotle and like I was just down this huge rabbit hole. 

And so what I did, for example, I put all that stuff in here and I was like, hey, I have a bunch of notes and some fragments of text for a long 10,000-ish word piece. I want to write something called, “Seeing Like a Language Model,” but I need to understand what I actually think and make a bit of an outline before I get started in order to do that. I need to understand the patterns of what I've been thinking and writing down. And can you suggest some ways that you can help me do this? I want to get from where I am to an outline. You have access to some fragments, notes, and early unfinished intro. And it just has a bunch of ideas like thematic analysis or argument mapping or chronological development. 

And I'm just sort of going down the rabbit hole with it, where I asked it to do concept clustering. So one of the concepts that I'm playing with is the philosophical divide Plato versus Aristotle, which I think is not quite right. It's actually Plato versus the sophists, but it's close to the evolution of Western thought. How does Plato ladder up into the rest of Western thought and into science and just the way the Western mind works, and then how do language models sort of differ from that paradigm? So that’s the basic thing that I'm trying for because I do have— The reason I asked this question is because selfishly, I feel like I haven't done the big idea thing as much as I really want to because I am writing every week. I am sort of reacting to stuff. And so I want to be a little bit more thoughtful and this is my attempt to put all of it together into something that makes sense.

Dwarkesh Patel (00:38:48)

Yeah, as you're going through this, really, actually makes me want to write more because now that you're talking about it, now that you ask the question, I'm like, yeah, I should be sort of consolidating the things I'm learning in a more comprehensive way and in a way that's also more useful and accessible to other people as well, right? I spend weeks learning about some random—not random, but the things I care about. I'm about to prepare for Daniel Yergin, the guy who wrote The Prize—it's a history of oil. Or, a geneticist, AI researcher, or whatever. To the extent that I'm getting something out of these research processes, I should consolidate it in a way that's not evident in the podcast itself.

Dan Shipper (00:39:30)

Yeah, I selfishly want you to do that because I'm curious what you think.

Let's move on. I want to talk about how you use AI doing the interview prep. So let's move into that. And then we can also maybe even like prep for an interview together.

Dwarkesh Patel (00:40:07)

Okay, let's do it. Honestly, the interview prep requires a lot of work, but fundamentally what's happening is not that complicated. I can just show you a document I might have made in the past. I'll share my screen. So honestly, it literally is just like I come up with a bunch of questions and I sort of group them together and relevant categories. Or if I go to— Okay, if I was interviewing Dylan Patel— I'm sorry, this is not the right one. Just a bunch of different questions basically. It's not complicated, but the process of coming up with them is very research-intensive. So we can go through if, I guess— I've only barely started preparing for them. We can go through the process of preparing for them.

Dan Shipper (00:40:50)

Yeah. I just want to stop at those questions, again, selfishly, because I think it's really interesting. You have these long lists of questions that are organized by theme. Are you going down the list or are you sort of jumping around?

Dwarkesh Patel (00:41:00)

Yeah. So it's really interesting because I come up with these lists of questions, but it really never ends up being, I asked question one and I asked question two and I asked question three. I start off with an interesting question and, if you listen to the interviews, hopefully it comes off more as a conversation because I spend so much time preparing that I have these questions basically memorized. And so the next one that is appropriate to their response, if they say something about memorization in LLMs, I'll have a question prepared about that or related to that. And I'll just ask it next because that's what fits in together. And so I'll have a list and this is what I'll send them if they ask for it. But really just sort of me off the cuff, here's a question. I remember that was relevant to this in the actual interview.

Dan Shipper (00:41:49)

That makes sense. So like the point of the doc is it's almost like writing the doc is, is the prep itself. And you don't even necessarily need it in the interview. Maybe you have it just in case, but yeah, that makes a lot of sense.

Dwarkesh Patel (00:42:05)

And then, yeah, we can even go through. Let's see. I'm doing a couple of interviews in the future—David Reich and Daniel Yergin. So the first former is a geneticist with human origins. The second wrote The Prize, which is the famous book about the history of oil. Which one sounds more interesting to you? We can do that whichever one.

Dan Shipper (00:42:23)

I want to do the famous geneticists.

Dwarkesh Patel (00:42:30)

So, let's go to Claude. In fact, I do have his book uploaded as a project, so we can just use that.

Dan Shipper (00:42:30)

That's great. And so basically what we're going to do is we're going to watch you and I'll do it with you. We're going to prep for an interview with this guy—what's his name again?

Dwarkesh Patel (00:42:35)

David Reich.

Dan Shipper (00:42:46)

David Reich. Okay, cool. Can we get like a little bit of background on David Reich? Maybe we can even ask Claude because I'm obviously a newbie to David Reich's work.

Dwarkesh Patel (00:42:50)

So, he is a geneticist at Harvard and, over the last decade or so, their research into how have human populations across the world been formed? Basically, who are the Europeans? What groups make them up? What ancient migrations and genocides and population replacements made them same with the Indians or Native Americans or Africans? it's completely changed— I mean, they've basically sort of made many academic disciplines irrelevant because they actually have empirical data on, here's actually what historically happened. You guys are completely wrong about what you think of your theories of what happened. If you're familiar with the Vesuvius challenges, you have these burned-up scrolls, but with some advanced techniques, you can get some useful information out of them. I feel it's in a similar vein. 

Obviously they're not the same kind of project, but it's a similar vein of once we develop the advanced mathematics or genetics or whatever to understand what's latent in the genome we've just uncovered a ton of insight about what's been going on in human history basically. And sorry, I'm just getting nerd-sniped and just going on riffs here, but one of the interesting things is you can see when one population replaces another, whether it was just like, oh, we met and we were like now intermingling and trading and whatever, or is it like we're committing genocide against you? 

And you can tell that because if in the case where it's genocide or population replacement, it will be that the male line of the population that is invading will overtake the male line of the existing population, but the female bloodline will remain. So mitochondrial DNA only comes up in the female line and you'll see the female line, because they're getting the new men who are coming in are taking them as wives or something. And then, anyway, so you can just learn a lot about what kind of invasion was it, did they conquer or were they just mingling or something? One of the many things you can see from the DNA.

Dan Shipper (00:44:53)

That's really interesting. Wait. And so this is like basically reexamining DNA evidence of old settlements and he's uncovering new ways of being able to analyze the DNA. What's the new methods that they're using to draw new conclusions from existing evidence?

Dwarkesh Patel (00:45:11)

One of them is just that, right? Seeing how the Y chromosome and the mitochondrial DNA—because you can just learn a lot about population based on how the female vs. male male line is propagated about, what was the social structure like and so forth. Another is you can even tell the level of inequality in a society, because if there's a lot, for example, in India. One of the things that was super surprising is the amount of endogamy, which is to say that a certain caste in a certain village would just not— There wouldn't be any sort of intermixing with another caste in a neighboring village to the extent that's true of nowhere else in the world. And they were able to find this in India, where the amount of social stratification, you can see that in the genetic catalog over the last thousands of years, where for thousands of years, these two neighboring castes haven't mixed with 99 percent or something, which is even from sort of infidelity or rape or something, you you would expect there to be more than what actually ends up being the case. So you can understand modern culture in India based on what has happened over the last few thousand years.

Dan Shipper (00:46:20)

That's really interesting. So, I feel like you're doing such a good job of summarizing his main ideas, but I kind of want you to do the same thing with Claude so we can see how you stack up vs. Claude because obviously you’ve input his book into this project. So it has that as reference material. Can we ask it to just summarize a few of his main ideas?

Dwarkesh Patel (00:46:45)

Yeah, that’s a great idea. “Can you summarize the techniques he used to come up with his new—”

Dan Shipper (00:46:57)

Perfect. And so what you're writing is, can you summarize the main ideas from the book and the techniques he used to come up with new insights? Cool. And one thing that's really cool about this is you've been able to do something like this with ChatGPT for a long time, but ChatGPT's context window isn't that long. And so it chops it up and it's not going to really be able to summarize the entire thing because it has to find the right parts of the parts of the book and the embedding search in it is not very good and all that kind of stuff. And Claude, you can just throw a ton of stuff in the context window and that just makes a big difference. So it looks like we've got some answers. Do you want to read them out?

Dwarkesh Patel (00:47:37)

So it tells us that ancient DNA revolutionized understanding of human prehistory. Okay. And then we've learned that populations today are the result of multiple waves of migration and mixture and then just like a bunch of other genetic stuff. Then it talks about the key techniques about whole genome sequencing and how they've enabled these sorts of new discoveries they've been making. Yeah, but anyway, so just a bunch of interesting things about their research.

Dan Shipper (00:48:10)

Well, now I'm interested in, okay, so the key technique that it's using is whole genome sequencing of ancient DNA samples. So is whole genome sequencing a new thing that you can do on ancient DNA samples? So it's said by improved extraction and sequencing technologies, that's a reason.

Dwarkesh Patel (00:48:32)

That is an interesting question. So we can even ask a lot because I'm not sure. “How exactly do you sequence an ancient or a prehistoric genome?” How does that work, right? Okay, so they grind the bone and they have techniques to get the DNA out of that. Now another thing we can ask is— One thing I'm curious about— Let's see. I don't really remember the chapter on Native Americans. I could ask about what exactly happened with Native Americans. Here's one thing I'm curious about. How would— Okay, so I don't even know if David Reich addresses this himself, but, “How would David Reich's theories help explain why civilization suddenly emerges so rapidly, concurrently in the new and the old world after 10,000 BC, a.k.a. the end of the last Ice Age.” And then maybe I'll just ask Claude why I think it is an interesting question. So, this seems like a really remarkable coincidence given how long humans have been around, you know?

Dan Shipper (00:50:06)

That is interesting.

Dwarkesh Patel (00:50:07)

“Coincidence, given that humans—” It’ll correct my spelling. “...”given that humans have been around for hundreds of thousands of years.”

Dan Shipper (00:50:20)

I didn't realize that we believe that it emerged at the same time in different geographically disparate places. So that's totally new to me. I thought it was just in Mesopotamia.

Dwarkesh Patel (00:50:37)

There's a really good book by Peter [Watson] called The Great Divide, and it's one of the most interesting books I’ve read, just as a side note. It's comparing the emergence of civilization and the world vs. your world. So in the new world, the Caral is like a civilization in 3000 BC, and it's based on fishing and not on conventional agriculture, like Mesopotamia. And he talks about how that changed the evolution of the culture in the new world vs. the old world. But anyways— So major population movements and mixtures, exchange of ideas, so it says maybe there were genetic adaptations during that time. There are no major biological changes. So he's saying that, I guess, human population, yeah— We weren't a different kind of human after the Ice Age. Maybe I'll tell it to be more specific. Your answer doesn't help me explain— “Your answer doesn't help me understand why the end of the last Ice Age led to all of civilization, what changed from before, and then you—this is really helpful to interrogate LLMs in this way, because their initial instinct, maybe because of RLHF has to be sort of summarizing mode and comprehensive and, I don't know, just give me the answer.

So maybe climate stabilization is like increased food sources, population growth. Okay. Interesting. Yeah, so this is super interesting. Claude didn't really give me a sort of full answer, but here's why this is still super useful because now that I know Claude doesn't have a good answer, this makes it all the more interesting for me to ask David Reich, because otherwise if it just like given me the right answer, I'm like, oh, okay, this is a known thing. I'm not going to waste his time with this right now that I know that it's not really clear. Now it's going to be such a fun conversation with David Reich.

Dan Shipper (00:52:53)

Yeah. And you sort of know it's not in his book because you've asked the whole book which is, which is pretty cool. So here's what I kind of want to do. I want to see if it is good enough at picking up patterns and how you ask questions that it can help you come up with questions that you think are actually pretty good for David Reich. And I don't know if it's going to work. Are you down to try?

Dwarkesh Patel (00:53:20)

Yeah. That's a great idea. Let's do it.

Dan Shipper (00:53:21)

Okay. So basically who have you interviewed recently-ish that is sort of in the same vein as David Reich, like the same kind of person?

Dwarkesh Patel (00:53:29)

I think the closest would be Tyler in a sense of a more sort of polymathic, less AI focused. So if I look up Tyler Cowen questions.

Dan Shipper (00:53:40)

What about anybody that specifically deals with genetics or sort of population changes? Do you think Tyler—you covered that in your interview?

Dwarkesh Patel (00:53:50)

Oh, no, not at all. But I just haven't interviewed somebody about that.

Dan Shipper (00:53:52)

Oh, you haven't. Okay. So that's new. So, okay. So, Tyler's good then. I mean, I think Tyler's has a broad enough range of ideas that probably work.

Dwarkesh Patel (00:54:04)

Cool. Okay. Let me pull up the questions. Okay. So here are the questions I asked Tyler when I had him on the podcast. And it was basically all these different economists. So I read Keynes's The General Theory of Employment, Interest and Money. I read Smith. I read all of Hayek's essays, and this was actually a super interesting interview. I was asking about the contradictions between Hayek and King. Anyway, and Mill obviously, so this will be a super— Hopefully it'll get what kind of thing I'm trying to do if I add this to Claude. So we go back to prep and then I add content, upload from the device or sorry, I guess, I just add a text—Tyler Cowen questions.

Dan Shipper (00:55:03)

One thing that I want to do first with this is, can you just ask it given all the questions you asked for Tyler Cowen, just ask it to pull out the patterns and how you like to ask questions and how you like to conduct interviews. Do that first—see how it does.

Dwarkesh Patel (00:55:15)

Okay, so maybe, “I’m the host of a podcast. I want you to find patterns.” Maybe I don't want to prompt it with a self-congratulation. “I want you to find the patterns in my questions. Here's questions from an interview of the economist, Tyler Cowen.”

Dan Shipper (00:55:52)

One thing I like to add to these is that the results you give me should be so detailed that another AI who is impersonating can follow your output to generate questions like this for a new guest.

Dwarkesh Patel (00:56:18)

That's a great prompt. Yeah.

Dan Shipper (00:56:38)

Nice. Okay. I like that.

Dwarkesh Patel (00:56:50)

Yep. So it says you often ask about comparisons between different kinds of thinkers. You pose what if questions. You ask how historical economic theories might apply to current issues. You present counter arguments and alternative viewpoints to test the strength of theories. You draw connections between economics and other fields. Okay, so maybe I'll just add, “I don't always interview economists.”

Dan Shipper (00:57:20)

Maybe you should add a couple more examples.

Dwarkesh Patel (00:57:21)

“Find patterns that are not domain-specific.” So let's go to Demis. Oh, shoot. Okay. And then we can do Dylan. So, it's telling me— I feel like it's over. It's specific to the kinds of interviews I've done because it's asking about timelines, which I asked Demis about. Or technical bottlenecks, resource allocation, industry dynamics. I feel like that's more about what I asked Dylan about with semiconductors. Yeah, I'm not sure how good this was but we can try it.

Dan Shipper (00:58:29)

So, what if we said I'm now going to interview a geneticist. He's a geneticist, right? Can you write a guide for another AI to prepare questions? Basically what I wanted to do is write the guide specifically for interviewing a geneticist based on the patterns in your previous questions. See if it can do that.

Dwarkesh Patel (00:59:25)

Yeah, this is much better. Okay. So it's about adaptability, fundamental concepts, future predictions, comparative analysis. Well, sure, you know what? Let's just try it. So, let's put this in. Let's put this in. “Help me come up with questions—”

Dan Shipper (01:00:06)

I might actually start it a new one and just be like—

Dwarkesh Patel (01:00:10)

Maybe I'll delete this so it's not over biased to that. Yeah, let's keep it. Maybe it's like, maybe you learn some context.

Dan Shipper (01:00:20)

Okay. Oh yeah. It'll have examples. It'll have examples. That's good.

Dwarkesh Patel (01:00:22)

“... examples for David Reich based on these guidelines that another AI-generated from other question lists and then—”

Dan Shipper (01:00:46)

Also, I was going to say, you could also tell to reference Tyler Cowen as an example, but—

Dwarkesh Patel (01:00:04)

Maybe one thing I'll point out that's been sort of a big part of me, I've noticed there's a big hesitation, like writing these prompts and reminding the AI here's who I am. Here's why I care about this. Here's the larger purpose of this project. It just ends up being a sort of big sludge. And that's often what keeps me from using AI tools that I know would increase my productivity because I don't want to retype here's, you're an AI and you're trying to help me come up with questions. I'm interviewing blah, blah, blah. And write this as you would write to another AI. So hopefully in the future we get models that just know all this about you and you don't even need to remind them because they would have just been listening to this call and they didn't know what we're trying to do. But in the meantime, I just think I'm just like, don't be lazy and just do the prompting.

Dan Shipper (01:01:49)

Yeah, yeah. I mean, I think that the ChatGPT memory feature is kind of getting there to some degree. I think like Claude projects, it contains custom instructions from one chat to another. So that's kind of nice. But yeah, I agree. I mean, there's a lot of typing to do, but what do you think of the output from this?

Dwarkesh Patel (01:02:02)

So, let's see. So it's asked about, given the rapid advancements in ancient DNA sequencing, what breakthroughs do you anticipate? How do your methods for analyzing ancient DNA differ from those using contemporary genetic samples? So I, one thing I'm noticing is that they're very generic questions and that's kind of what makes it not that useful for me is like, just come up with questions for me because I tried to ask more specific— It’s a more of like, after having read this passage and considering the research in this area, here are some thoughts I have. The 10,000 year BC thing was a perfect example of this, of like, okay, we know that civilization emerged rapidly, what was going on there, right? It's a very specific question, and I don't think that if you just generally aren't that good at, come up with a specific thing from a large context. It really wants to do a summary level or high level kind of questions, right?

Dan Shipper (01:03:12)

I think you're totally right. I think this is a sort of common failure mode with using AI tools—people end up having— They're like, okay, maybe you can do my entire job for me. And then it's like, no, it's too high level. And it doesn't really work because ultimately these questions are about what's interesting to you. And it doesn't have enough context on you to know what you are specifically going to be interested in just from the patterns and previous guests, especially, if you don't have guests that cover the same sorts of topics. 

I think if you had five other interviews with other geneticists, it might do a little bit of a better job. And I think usually the solve for that is backing up and thinking about what are the different micro tasks that I do as part of getting to those questions that this could be helpful for. So, instead of just doing the whole thing all at once, it's backing up to the particular context that I'm interested in currently, what are the micro tasks that I could replace myself.

Dwarkesh Patel (01:04:09)

Yeah, totally. For example, the things we're doing— Research is pretty helpful because then you can narrow down on, I want to research X but not complete the job.

Dan Shipper (01:04:18)

So fortunately you still have a job. Claude 3.5 Sonnet, good for research, not going to replace you, but you are also using AI not just for reading, not just for research, but also for helping you sort of in the post-production process of putting episodes out there. Do you want to tell us about that?

Dwarkesh Patel (01:05:13)

Yep, totally. So, obviously there's a bunch of things that go into post-production of the episode, which I've been trying to come up with workflows to help me out with. Before I started having a human make transcripts of the podcast and before I had ads to help me subsidize that. What I would do is I would have a text-to-speech or speech-to-text— AssemblyAI is the API we use. Or I have it come up with the transcript, a first-draft transcript. And then I have GPT-4 ,just literally— It says prompt transcript. I came up with a couple of methods to do this. And it just says, so you're you, and then I came up with some guidelines so that like, here's some prompts to make sure that when you rewrite the transcript or removing filler words, you're making the thing more readable. You are cleaning it up. And it actually worked decently well, but it just wasn't good enough that it beat out a human yet. So, then it's just like worth it for me to just have a human do it. but other things we're doing is I'm trying to come up with a workflow where I can just upload the MP3 of the episode and it makes an auto-generated transcript. And from that transcript, we can just have it come up and generate different ideas for titles and clips and highlights. And it's a work in progress right now but, you know, we just do a sort of a few-shot learning in the prompt of, here's what good titles look like. Here's what good clips look like. And this is just a sample of a two-minute random interview. We're trying to figure out ways to get this kind of workload going.

Dan Shipper (01:06:49)

That's really great. And the reason I think it's great is because we actually built these two at Every and we use it all the time. And then we just released it as an app. Have I shown you Spiral? So if you go to spiral.computer. So, basically we just built that into an app where you can create these things called spirals, where spirals are basically few-shot prompts for repetitive creative tasks that you do. So for me, when I do this podcast, I always have to come up with a tweet for it and I put in some few-shot examples. It creates a little style guide for itself. And then you get a form that you can share with your team or just use yourself or share publicly. And then every time you have a new podcast, you just paste it in there. It can pull out tweets. It can do transcripts. It can do highlights and all that kind of stuff in your style and in your voice using the few-shot stuff. It works really well. I use it a ton internally and it's going kind of viral. We're about to pass 3,000 signups for it. It's been like a couple of weeks. So, I would love

Dwarkesh Patel (01:07:53)

Can I set custom prompts? I want to generate a tweet using these guidelines or something like that?

Dan Shipper (01:07:55)

Yes, you can. 

Dwarkesh Patel (01:07:56)

Okay. I would love to use this.

Dan Shipper (01:07:48)

Okay, cool. You should use it. I'll hook you up with an account after this.

Dan Shipper (01:08:32)

Cool. Anyway, we'll hook you up with a license and I would love to see what you think of it. I think it's just a problem that all creatives face. There's all this kind of drudgery of creative work that's not about the core thing that you're doing. And Claude just got to this place where it can do a lot of that. And, yeah, I just needed a form. I just needed a little builder thing. And it sounds like you have the same problem. I don't want to prompt it every time.

Dwarkesh Patel (01:09:06)

Exactly. Yeah. I'm so excited to use this. I'm going to experiment with it later today and it would save me so many hours per production of an episode to have this kind of stuff because it's hard to overestimate how much time these kinds of things take, as I'm sure you know.

Dan Shipper (01:09:03)

Totally. Totally. I do. I love to hear that. Let me know if you have any feedback. So before I let you go, I have to ask you a couple important questions. So, what are your AGI timelines?

Dwarkesh Patel (01:09:40)

If I would give a 25th percentile, 75th percentile sort of bounds. And I would also say, What do we mean by AGI? It's not just that it's productive or can generate trillions of dollars of value, but really it's that you can replace a remote worker. Any remote worker you can just replace with AGI. For that, I'd say 25th percentile, maybe 2029, and then 75th percentile, like 2050.

Dan Shipper (01:10:10)

Okay. Interesting. And then what about p-doom?

Dwarkesh Patel (01:10:12)

So something along the lines of, the thing that's taken over doesn't really have any sentient experience and it just doesn't have culture, doesn't have individuality, doesn't really— It's not just that humans are disempowered, right, because I think humans disempowered chimps and I think we didn't doom the universe. I think it just has to be the paper clipper that— What are the odds of that? Something like that. I don't know. 10 percent or something like that. Things had to go really wrong.

Dan Shipper (01:10:50)

That makes sense. I can't decide whether that's higher than I thought it would be or lower than I thought. I think it's higher than I thought it would be.

Dwarkesh Patel (01:10:52)

I think it is fair to be like, that's crazy. You know, 10 percent odds that everything we care about is going to not exist in the far future. I think that's like the right reaction to that, honestly, because often it's easy to bandy about these numbers as abstractions are, I don't know, what are the odds the Patriots will win or something? And no, we're really talking about technology that will be alive in our lifetimes and might result in something really bad and we're taking that seriously.

Dan Shipper (01:11:28)

So, I really appreciate you coming on for anyone who has been listening to this and does not know you yet and wants to find out more. Where should they find you?

Dwarkesh Patel (01:11:37)

I have a podcast called the Dwarkesh Podcast available on YouTube, Spotify, Apple Podcasts, wherever. And there's also the newsletter, as you saw, hopefully thanks to Dan's prodding, I'll be doing more writing. And that is at dwarkeshpatel.com as for and then you can also follow me on Twitter or something. So that's dwarkesh_SP, or just look up Dwarkesh Patel on Twitter.

Dan Shipper (01:12:00)

Amazing. Thank you so much. This is awesome.

Dwarkesh Patel (01:12:01)

Thanks so much for having me on, Dan. It was super fun to go through that workflow and also get your tips on how to use these tools better. I'm actually pretty excited for people to see how these tools have been useful in my workflow. So, I'm excited for this to be out.

Thanks to Scott Nover for editorial support.

Dan Shipper is the cofounder and CEO of Every, where he writes the Chain of Thought column and hosts the podcast AI & I. You can follow him on X at @danshipper and on LinkedIn, and Every on X at @every and on LinkedIn.