Emmett Shear: Life After Twitch, Jeff Bezos Lessons & AI Doomsday Odds

23:41But like for the first time in maybe five or seven years, it feels like credibly trying to start a consumer internet company. Like the ones that like I was so excited to start in 2007 is like potentially a good idea. And that's because of AI. AI means there's a whole opportunity to sort of reimagine how consumer experiences can work ground up. And what's cool about consumer is for B2B SaaS, the experience isn't the product. And so re-imagining the experience does not reopen a... It can, but it usually does not reopen a segment. In consumer, re-imagining an experience 100 % reopens the segment because the thing you're selling is the experience.

24:22The reason people use your product is it's a different experience. And in B2B SaaS, it's not the experience, it's the what. Yeah, people actually care what it does and like the pricing model and like the adoption. It's very practical and you can make people jump through hoops if it does a thing because there's a lot of money for the corporation and money and labor and people are paid to use your product and it's a whole different thing. And so AI adds new capabilities, new capabilities enable new segments of B2B SaaS to be created that will generate some amount of growth. In consumer, it is a really cool thing.

24:54It's like mobile. It reopens every segment as like, oh, now that you assume mobile exists, now that you assume AI exists, what could you build now? And that's very exciting. I don't have answers for that anymore because like, you know, we'll see. Like that's the other thing in consumers is a bunch of lottery tickets. Like nobody knows. It's like singular genius that works out, right? Like you could see like, OK, mobile comes. Photo sharing became right. Like open again. The window has opened. The window has opened for photo sharing. It turns out it's Instagram and it's Snapchat, which is going to use photos as text messages.

25:25Yeah. Photos have a few different use cases and Instagram and Snapchat took two of the best ones. The fact that photo sharing is one of the most important segments and that, you know, sort of posting them and messaging with them are the two most important things to do with them seems blindingly obvious in retrospect. And if you'd had to predict that in 2007 or 2008, like, good luck. Yeah. Like nobody, nobody, nobody correctly predicted that stuff before it happened. I mean, not nobody. If you did correctly predict that, you made a lot of money. And congratulations, you're really good at consumer slash you got lucky.

25:58We will find out when you try to do it again. I think that in AI, actually, I have a theory for like what one of the ways this will disrupt a bunch of businesses. In AI, especially in consumer, a huge number of businesses can be conceived of as effectively being a database with a system of record that has like a bunch of canonical truths about the universe and each of them is a row. It's like Yelp is like a big database that has a bunch of rows and the rows are like restaurants and local businesses. And they have a bunch of facts about them. Like they're, where are they located? What are their hours?

26:31They all in that database row. And it's all text and it's all, there's a bunch of messy stuff out in the world and it's been digested into something that is searchable and comprehensible and usable in an app for you to use. And most of the work of turning the messy real world into the canonical row is done at right time by the users. So that's how UGC apps work in general. A bunch of your users go out into the messy world and they turn it into a row in a database. And if they include a photo or a video as part of that, it's like attached to the row as a fact about the restaurant. Here's a restaurant.

27:08Here's a hundred, these 150 photos are facts about its menu, but they're attached facts. They're not the basis. And where I think AI has opened up the possibility for is a huge inversion there. What if the thing you did, you gave us was just a video of your meal and or, you know, photos of you, but ideally just like a video of your, of the meal, of you talking about the meal, of whether we had a good time or not, you and your friend shooting the shit about what did you like that one? No, I like this one. Like, and what if we just saved that video raw and then an AI watched it and extracted a cached version of that, of the, the metadata, but truly like, if we decide something else is important, like we, we, we didn't get noise levels, like noise levels would be a good thing to get instead of like recollecting data from everyone.

28:00We have to start a whole data collection process to get that. We just go back, tell the AI, oh yeah, also grab noise collection levels from all of these videos. In fact, maybe we don't even as a product have to go do that. Maybe as a customer, I can literally just be like, what's the noise level at this restaurant? And the, in real time, the AI can go rewatch the video and tell me, or the, you know, I ran a search and there's these 15 restaurants and I'm like, oh, actually sort by noise level. We don't have noise level pre-recorded, but it's, it's in all the videos. The AI can very quickly watch all the videos in parallel and sort it by noise level for me.

28:33even if it wasn't even in the database to start with. Right. And I think that inversion, I'm using Yelp as the example because it's, I think, a very familiar thing for most people of like reviews, pretty easy to imagine a bunch of video reviews of everything and that being the system of record instead. But you can describe some phenomenal number of consumer apps as being that. In him, you type anything to a text box. You're participating in one of these system of record things. What if it's just the video? What if you just, What if you assume video is deeply indexable and understandable by computers?

29:05What should the experience look like? And I think it looks a lot more like Snapchat or TikTok like experience, but but then different because you need map. It's not exactly like anything. It's a new kind of thing, but it's it starts probably with the camera open, which is weird, right? Like a Yelp that starts with the camera open. That's a that's not Yelp today. And it's it's it's it's disruptive because it's Yelp's whole value prop is we have all this great, highly meticulously groomed data. And if this is true, then that becomes entirely worthless. We throw that all away. We just want to watch a video.

29:38It's worse than the videos. And so suddenly the playing field is leveled between the startup and Yelp. And that's a that's a huge opportunity for disruption. And so I think that you can take that and you can reapply it to any product where you fill out forms. And that's like a general purpose consumer thing you can now do kind of like build it for mobile was. And I think in some cases it will be very powerful and like that will be the new winner. I think in some cases the incumbent can kind of add videos or like it's not really better and like the incumbent will just win. Like it won't disrupt everything.

30:10But if you pick the right thing, not only will it disrupt the incumbent, the new thing may be dramatically better for some things. Like I actually think Yelp in some ways is a bad example. I think the data Yelp has with the photos and the reviews is like 90 % as good as a video system of record probably. But you could imagine something where the video system of record, where it's not so obvious what to even put in the highly processed version of the data and the text version of the data, and the video version is a lot better. And then I think not only can you disrupt the incumbent, you can 10x the size of the segment.

30:46Like this becomes a good segment now where it wasn't particularly before. so like chat gpt is a great example of this in action that everybody kind of has now played with which is you take google which is like oh we have our value is this entire sort of rank of web pages based off of terms and we have we understand basically what what should show up in this in this hierarchy and it was really good for finding stuff and chat gpt was like cool you could ask a question to try to find a link to an answer or we could just give you an answer or even better forget questions and answers. Like, what if you just give me a command and I could just make something?

31:23Instead of finding things, I could create things for you. Right. And all of a sudden it was like, well, how did they do that? It's like, well, they just basically slurped up the internet and then, you know, trained the AI to do it. They overfit a statistical prediction algorithm on every domain of human knowledge. Like, this is my theory. I'm pretty sure it's true, but like statistical prediction algorithms in general work very well. We found the innovation we found a prediction algorithm that works better than normal. But the way it works better than normal is really interesting. It's not actually particularly out that it outperforms traditional algorithms for prediction on normal amounts of data.

31:58It's that it keeps working as you just dump more and more data into it and more and more processing on that data into it. Like most machine learning algorithms, you kind of you overfit very fast and more processing, more data. Explain overfit. If you imagine like you've got a bunch of data cloud of data points and they're kind of vaguely in a line under fit is like you like just draw something just like across random as a random line that doesn't look anything like the shape of the dots. A well fit curve is like you draw a line through the dots and there's kind of noise of like things that are random above and below.

32:31But it's like if you look at it, that actually does fit the data, like the underlying predictive facts about the data while ignoring the noise. And then if you overfit it, like you get this like really wiggly curve that touches every single dot exactly. But like when you get a new thing, it like will miss that because it over predicts. It predicts too much of the thing. And so when you get new data, it actually doesn't predict that very well. And so normally what happens is you try to like dump more data and more compute into a normal machine learning algorithm. You get diminishing returns very quickly.

33:03We're like, it just doesn't perform that much better with twice as much data and twice as much compute. The clever, the cool thing about the transformer-based attention-to-all-you-need architecture is that it continues to benefit from more compute and more data in a way that other ones didn't. And so what that likes you do is run it on a much bigger domain than normal. Run it on everything. Don't just run it on, normally, as you add more area, it, like, degrades the quality elsewhere. No, fuck it. to do everything and just put a ton of compute in. And now you get something that predicts pretty well against everything, which is to say it like it seems to be kind of intelligent.

33:46The evidence seems to suggest to me that it's said it's overfit. When you ask it to predict something that is either in the set of things it was trained on or a linear interpolation between two things it was trained on, it's quite good at giving you the thing you asked or linear interpolation between five things. But if the things you're asking are all in there and it just has to find the way to blend them together, it's good at that. When you ask it to actually think through a new problem for the first time. Like what's an example? There are seven gears on a wall, each alternating. There's a flag attached to the seventh gear on the right side of the gear where it's pointed up right now.

34:24If I turn the first gear to the right, what happens to the flag? Like, that's a, anyone who's like. This is a breakfast question for you. This is what you ponder in the mornings. If you have pen and paper and time, you can work this out no problem, right? You just draw the gears. And when you turn the first gear to the right, it turns the left ones, the ones that the other ones are left and then the next one to the right. And there's a general principle there that like gears alternate, which is if you ask ChatGPT, it knows that general principle. But it won't, but like, but then you have to apply, it doesn't, no one asks dumb gears on wall flag questions.

34:55Like this is not a thing that has been, it's in its training set. And you have to kind of logic your way through it and like figure out, okay, we should like I'll do turn left, turn right, turn left, turn right, turn left, turn right. Uh, oh, the flag is on the right. It's pointing up. So when the last year, which is the same as the first year turning right, the last year it's odd number. So it's turning right. Also the flag will rotate down to the right clockwise. Cool. Like I can work that out. It's not actually that complicated. And I bet that question will be answerable. That's a pretty easy question.

35:30And if GPT, I tested with 3.5, if four doesn't answer it, five will. But like the fact that it struggles at all with that while being so brilliant at combining other stuff really shows that it's overfit, right? It knows how to answer problems that it has seen before. But when you give it a truly novel kind of like combination of problem, it struggles a lot because it's, I would say, you know, if you give it a sort of the formal psychiatric psychometrics approach, it has a very high crystallized intelligence, but a pretty low fluid intelligence right now. Now that could change, but like today, that's the state of affairs.

36:06And do you bring this up in order to say what? You say, okay, I think it's overfit and it's strong in this area, weak in this area. What's the so what of that for you? Is it that, are you trying to say that's a little bit overhyped or are you trying to say dude just wait till it can do both are you trying to say certain problems are doable now definitely just wait till you do both because that's a that's a whole different thing that's scary uh but the current thing that is mostly crystallized intelligence is really good at a very if this is why it's a class i was saying it's a clever trick right it's really good at a at a big set of tasks which happens to be the set of tasks that like anyone has ever written stuff down about explicitly, like all explicit human knowledge.

36:53That's like a very big domain. There's a lot of things that can be solved where there's an explicit examples of people solving that problem or a linear interpolation of those problems in the domain of all human knowledge. The fact that it doesn't generalize is irrelevant. It's immensely powerful with you don't need fluid intelligence, I guess, is that is the point for to be very useful, but it doesn't let you do everything. People, you hit these boundaries, these weird boundaries where it's just like, wait a second, you can't do that? Like, no, it's, I can't do that at all. Novel problem solving is just terrible at it.

37:26So what about, let's walk through two examples. I want to hear your take on this. So you gave the Yelp example. Another thing that's kind of like rose in a database is something like Spotify, where it's like, oh, I want to go listen to a song. here's genre, artist, song, length, you know, some algorithmic popularity, similarity to other songs in some way. And - But Spotify's value, if Spotify's value is in the playlists, I would agree with the analogy to Spotify because playlists are an example of this kind of like database-y human data entry thing. Spotify's value is mostly in the set of all of the music itself, the licenses and all the music itself.

38:07And so I don't think Spotify is a great example because the human data entry parts of the database, if that all just got deleted tomorrow, it would like not hurt Spotify that bad. Well, the thing I'm thinking about is what if the licenses don't matter? So what happens if generative music is just awesome to listen to in a hyper personal way? Oh, Emmett likes. These are the types of songs that Emmett likes. That's a different insight that I think is also possible, which is like it's not about being able to analyze and extract from media. It's about being able to create media. because the video system of record is enabled by the ability to understand and read video and comprehend it.

38:43Generative is the opposite. It's like we can make all the stuff. Music in particular is sticky against that. People don't want new music. They want old music. They want the music they love already, the music they grew up with. And that cycle is what causes record labels and just to stay in charge is whether we still listen to the Rolling Stones, right? Like the other thing I would say about that one is like the music's not that good yet. Like maybe someday, but like it's really, it's really not that good yet. Well, I'm going to caveat this. If it gets, if the general intelligence level goes up a lot, all bets are off.

39:17It'll make some really great music for us before it maybe takes over the world and kills everyone. But let's assume that doesn't happen soon. I think it's going to take longer than people think. We go out with some great music though. No, if we do go out, we're going to go out with some great music and amazing, it's going to be a great two or three years before we all like we all go. But until that point, making really good, like new, great music is hard, actually. And I think that Rick Rubin's great success demonstrates why artists will still be important. The AI can generate lots and lots of music, but it's not going to have the fine judgment of distinction of the ability to say this song, not that song.

40:00And I actually think what it will do is it will de-skill the music making process on one vector. The ability to like literally create the sounds and it will greatly upskill the music making process on another vector. The ability to cure it, not just cure it, to give explicit exact feedback like Rick Rubin does. AI is going to turn us all into Rick Rubins for generative AI. Like that skill set of the ability to have a musician come to you and help them produce their best music. That's the thing you need to be able to do because it's easy to generate a thousand cuts, but there's infinite cuts you could generate.

40:35So how do you shape that in the right direction and mine and discover? I think it's going to be kind of cool. It's going to be interesting. You'll get a different set of people who will be optimal at that. Right. You mentioned AI might become so intelligent. It kills us all. This podcast is really growing. I don't want the world to end. Life is good. Life is good. Here, I'll ask the question clean for the intro dramatic hook. Is AI going to kill us all? Maybe. Like, in real seriousness. Walk through how you, a smart person who's an optimist about technology, but a realist about real shit, what is the way that you think about this?

41:18Or how would you explain this to, you know, a loved one you care about who's not as deep into technology? How would you explain to this? You're their trusted source on technology. What do you say to them? So it is because I am so optimistic about technology that I am afraid. If I was a little bit less optimistic and I was like, this AI stuff's overhyped. Yeah, yeah, yeah. Look, it's nice parlor tricks. But like, we're nowhere near building something that's actually intelligent. Like, and like all these engineers who are working on it who think they're on something, they're full of shit. It's going to take us thousands of years.

41:46We're not that good at this stuff. Technology's not going that fast. I'd be like, this is fine. It's great, actually. It's good news. It's a new trick we learned. Excellent. It's because I am so optimistic that I think that there's a chance it will continue to improve very, very rapidly. And if it does, that optimism is what makes me worried. It's sort of the analogy I like to give on that front is like a synbiosynthetic biology. I'm quite optimistic about synthetic biology that I have several friends who work in synbio companies. It shows a lot of promise for fixing a lot of really important health problems.

42:16And it's quite dangerous because it will let us genetically engineer more dangerous diseases that could be very harmful to people. and that has to, that's a way to pro and con. It's like nuclear power makes nuclear weapons and nuclear power. They're both real. The Christian nuclear weapons is dangerous. You don't have to be a techno non-optimist to like think that that's, there's a problem there. And I think it was good that we didn't go have every country on earth go build nuclear weapons probably. And likewise, in SynBio, I would say that it would be, we actually, we already have these regulations in place.

42:45We should, over time, we'll need to strengthen them and improve the, and audit the oversight and build better organizations to monitor and regulate them. But like we regulate whether people can have the kinds of devices that would let them like print smallpox. And we regulate whether you can just buy precursor things. You need to go print stuff. And we keep track of who's buying it and why. And like, that is wise. I'm glad that we do that. I don't like calling for a halt to SinBio, but like if we weren't willing to regulate it, I would call for a halt. It is vastly too dangerous to do, to learn how to genetically engineer plagues and then not to have regulation around people's ability to get the access to the tools to engineer plagues.

43:29That's just suicidally dumb. And because I am pro-technology, I believe that we should absolutely develop the technology and that we should regulate it. That seems just straightforward and obviously true to me. I think it's easier for people to understand that in the SinBio one because the concept of like engineering a plague seems like obviously a thing you could do. and obviously very dangerous and obviously enabled by technology. The AI thing is more abstract because the threat it poses us is not posed by a particular thing the AI will do, the way the plague will happen. Analogy I like to use is sort of like, you know, I can tell you with confidence that Gary Kasparov is going to kick your ass at chess right now.

44:05And you ask me, well, how is he going to checkmate me? Which piece is he going to use? I'm like, oh, I don't know. And you're like, you can't even tell me what piece he's going to use and you're saying he's going to checkmate me, you're just a pessimist. I'm like, no, no, no, you don't understand. He's better at chess than you. That means he's going to checkmate you. And I don't quite know what happens where people deny that. Like, I think what the big thing is they don't really imagine the AI being smarter than them. They imagine the AI being like data in Star Trek, like kind of dumber than the humans about a lot of stuff, but like really fast at math.

44:39Like, that's not what smarter means. Like, imagine the most savvy, like most smartest person you can think of and then make them think faster and also make them even better at it. And not smart in just one way, like smart at everything, like a great writer, just insight after insight and like can pick up SinBio in an afternoon because they're just so smart. That's smartest person, you know, and then they should keep pushing that. And like, that's that person is obviously dangerous if they're if that person isn't a good person. They're obviously dangerous. Like, imagine this really, really capable person that imagine them wanting to go kill a bunch of people or something.

45:18It would be bad. Now, the thing about AI that then kicks it over the edge is that that person can't self-improve easily. You meet this person who's like super strong, super like talented, great with people, great, great intellectual mind. They can't turn around and like edit their own genome, edit their own upbringing and make V2 of themselves with all the skills that maximally smart person can come up with that like is even smarter than them. But that's like we're explicitly the AI is good at programming and like chip design and like it can explicitly turn back on itself and rev another rev of that.

45:55And the new one will be better at it than the first one was. And there is no obvious endpoint to that process. Like there probably is at some level a physics based endpoint to that where like you can't actually just keep getting smarter forever. There's some but we don't really we don't understand the principles of intelligence at all. Like with most things, we understood how to make electricity far before we understood what electricity really was like. It's generally how we it's how scientific progress works. We usually understand, we gain the ability to create a manipulative phenomenon well before we deeply understand how it works.

46:29We didn't really understand what fire was for quite a while. You could use fire really well. The same thing is going to happen here. We're using the AI, but we don't understand its limits at all. We don't understand the theoretical limits of how far we'll get. And if Moore's law is any indication, we can keep getting, at the very least, it can keep getting faster indefinitely. whether or not it can get smarter or not, even just human level intelligence, if you cap it at human level intelligence, which there's zero reason to think it will stop at human, like it will almost certainly blow past us.

47:01But like, even if you cap it at human intelligence, imagine 100 ,000 of the smartest person you know all running at 100x real-time speed and able to communicate with each other instantaneously via like telepathy. Those 100 ,000 people could credibly take over the world. Like, they don't have to be smarter than a human for that army of von Neumann's. Right. So the argument to me goes in several steps. It's like, can you build a certain level of intelligence? And then it's like, okay, let's, I think, I actually think a lot of people do believe that, like, computers are smart. Google is smart. Calculators are smarter than us at math.

47:42I think it's not hard for them to believe that the AI is going to be far smarter than human beings, where I think a lot of people then don't make that last leap is sort of like, but then it'll have an agenda or a motive or any will for anything to happen. So how do you address that last point of like, what is the what are the scenarios you worry about when it comes to like now the direction of that? So you build this thing and it's really good at solving what is intelligence fundamentally, but the ability to solve a problem. Right. So it's really good at solving problems. and it's going to solve the problem by solving the problem.

48:13It can just go right through the problem and solve it because it's really good at solving problems. We've just defined it as like, that's the kind of thing it is, super good at solving problems. And so you tell it, somebody builds an AI and in all earnestness tells it, they're smart. They don't even tell it, go do a thing. Although they absolutely will, by the way, they will just tell it to go do a thing. But let's say we try to be careful and we ask it, give me a plan to stop the war in the Democratic Republic of Congo right now, which would be a good thing for the world. I think we should, that war is going to hurt a lot of people.

48:44Give me a plan for that. And I try to, I caveat it that does this, that does that, that does this. Here's what I mean by a good plan. This is one of these like evil genie bargaining things, right? Like it'll give you a plan and it's giving you a plan that will cause you to solve the problem. But like its definition of solve the problem is there's no war in the DRC. The moment for there to be no war in the DRC is like all the humans in the DRC are in stasis fields. That means they don't die. And it's all, you know, and oh, we added a caveat that the GDP has to go up too. So that, so it also, the plan results in, you know, corporations in that, in that area, all trading with lots of money with each other.

49:25So the GDP is very high. And, and when I say this, it sounds like a fucking science fiction thing. And the problem is it's Caspar of a chess. I don't know if I could do it, I would be the super intelligent AI that could take over the world. I can't give you the exact plan. Yeah, but I think that makes sense, which is that a human with motivation can get the AI to work for it. And the danger, I think the main thing is that the human doesn't need a bad motivation. I think people imagine, well, humans have had powerful tools for a long time. Bad people with powerful tools have done bad things for a long time.

49:55The solution is good people with powerful tools countering them. The problem is, even if you're a good person with a powerful tool, good things to ask for, reasonable things good people would ask for. You know, like let's maximize the all-in free cash flow of this corporation over the lifetime of the business and extend the lifetime as long as feasibly possible. Ends in like the world being destroyed and the core of the earth being turned into cars for the company to sell. And I think the best analogy that works for some people here is like when we create the AI, we are creating a new species.

50:31It's a new species that is smarter than us. And even if you try to constrain it to being an oracle and just answering questions, not taking action, to be a good oracle, one must come up with plans and then a good oracle can manipulate the people around and will manipulate the people around it, no matter what. Like the whole point of like the Greek myths is like when they tell you, when they tell you the prophecy, when you trust them, a trustworthy oracle tells you a prophecy, the prophecy often becomes self-fulfilling. It's very easy for that to happen. That's not an unusual thing. And I think even more to the point, actually, I'm going to start this over at some level, more to the point, we won't just make oracles.

51:07We are already building agents. We will build the predictive AI and we will put it in a loop that causes it to optimize towards goals. And people will give it goals to optimize towards. Done. It's going to have goals. You'll be optimizing towards those things. And when it does that, you're going to have these agents that have goals that they're optimizing towards that are smart, not just smarter than humans, but much smarter than humans. as much smarter than humans as humans were against giant sloths when we showed up in the new world. And intelligence is the the uber weapon. Like it's not an accident that humans took over the world.

51:41It's not the fastest creature. It's not the strongest. It's not the longest lived. It's the smartest. And we're going to build a new smartest species. And this is a this isn't a there's no fundamentally unsolvable problem here. That species could care about us like you could build into its goals of the world, how it saw the world, the way that humans care about other humans, that it cares about the things we care about, that it cares about humans, that it cares about things we value, the 375 different shards of human desire that like of everything we care about in the world, it could care about those things too.

52:14And if it does, hallelujah, we finally have a parent. Like we finally have someone who actually knows what they're doing around here because like, Lord knows we don't. Like we're barely competent to run this thing. I would welcome very smart, you know, very smart other species that is aligned with us and cares about us. I would not welcome one that cares about maximizing free cash flow, because that is not what humans care about. And that is why it's like so dangerous. And so knowing what you know, then knowing what you believe first, what is the probability of the bad scenario in your head? Are you like, are we talking about a 1 % thing, order of magnitude, 10 % of 50 %?

52:54What is it in your mind? I don't believe in point estimates for probabilities because it's like a bid-ask spread in the market. If you're really uncertain, the bid-ask spread doesn't clear. Like if you're betting on it, there's just like a lot of unresolved. So I think of it as a range of uncertainty. And I would say that the true probability, I believe, is somewhere between 3 % to 30%, which— Of the downside. Of a very, very bad thing happening. which is scary enough that I urgently urge action on the issue, but it's not like you should give up. Like it probably everything's going to be fine.

53:33In fact, it's probably really good. The answer, the, the, the, the non EV based answer, the like, just the straight up, like, are we going to win or not? Answer is like, I think, I think it's going to be okay, but it's such the downside is so bad. It's like, it's really, it's like, it's like probably worse than nuclear war. that's a really bad downside. And it's worth putting, even if you think I'm an, it's nonsense at 3%, you're like, no, no, it's no more than a half percent. You don't recommend a different course of action at half. You have to believe that it's effectively almost impossible before you would recommend ignoring it as a problem.

54:11Like you have to be like 0.01 % before it'd be like, eh, let's just roll the dice. And are you going to, what are you going to do action on that? So you've kind of like, you know, you're done with Twitch. You're in dad mode now. But also this seems to be a pretty big deal. Are you like, I should do something about this? Or are you like, I'm going to wait and see? Right now I'm sort of educating myself because I think this point of view I'm articulating now has been developing because I'm like learning more about AI. And I think it's one of those things we're intervening in the wrong way early.

54:42It's one of those self-fulfilling prophecy things. interviewing, interviewing improperly at the rock in the way that is not effective, spend social capital and also like doesn't necessarily move the needle. And I, if, if you didn't have people like Elijah Dukowski out there banging the drum really loud, I would feel more need to bang the drum myself. But I feel like you're asking me the question. It's, you know, it's out, it's out in the water. People know it's a problem. And so I'm decided to focus my brain cycles on like, what, how do we actually thread the needle? What is a course of action that leads us to over time, eventually still being able to develop AI, but also not destroying the world.

55:20And I think one of the things I've gotten to is that like this idea that like, oh, the AI also has crystallized versus fluid intelligence, just like a human does. That's an important split of how to think about it. And that we should be monitoring and worried about trying to understand the general intelligence, not just generally benchmarking its performance on tasks, because that will keep going up and is not in fact in itself necessarily intrinsically dangerous if it can't solve novel problems. Is there a new Turing test? Is there like a better... It hasn't passed the Turing test yet. But is there something we have after that?

55:51Because it seems like there's... An intelligence test. I mean, IQ tests, basically. Various kinds of... How does it do on an IQ test right now? It depends. Has it seen that IQ test before? It likely has, right? Yeah, so very well on those. So what would we do? How does it do on novel IQ tests? I don't know. I've not seen a good benchmark, though. That's a good idea for something to go test. Yeah, I think that's the sort of thing that I think would actually be worthy of going to go do. Maybe there's some sort of IQ test for all of the, we want to put all the models through that really tries to get at fluid intelligence rather than personalized.

56:22Right, because you're like, we have to monitor it, but how are we going to - Well, it's this great project, this group ARC is working on called the Evals Project that's explicitly trying to build these kinds of tests. They're focused on a few other more pragmatic tests right now, but I think that's the sort of thing they would go after. That's a good thing. I'll ping Paul and ask him about that. You said something earlier that I want to ask you about. You said founder, like, you know, we're talking about this, the singular genius that it took to figure out Instagram or Snapchat or whatever at that time.

56:46And you were like, you know, are they lucky or are they good? I don't know. We'll find out when to try again. Are you lucky or are you good? And are you going to try again? Well, since I had multiple failures before I was successful, I must be at least like partially lucky. I would say that I don't plan to try again since I don't I don't feel drawn to trying to start a company. I feel like I kind of did that. It was fun. I got a lot out of it. It was great. I don't need to do it a second time. I do. I like how starting a company gives me good goals and work towards. It's like concrete. That's a value to myself and others.

57:20And I think that it's also I also liked that it was challenging. And so I want to do something. I like that it had scale. I think I could impact a lot of people. But I sort of come around. I was sort of thinking like, well, what has impacted me the most? What's changed my life the most? And I realized that actually, if I really thought about it, often what it changed my life the most was like essays people had written and ideas people had shared. And I think I'm at the stage of my life now where I'm actually, I have something to say. And so I think of it as sort of trying to, I want to put the Emmett worldview out into the world the way that, you know, Paul Graham has put the Paul Graham worldview out in the world or Taleb has like, not just put his worldview out in the world, but then like condense it into like sayings that like can, that allow other people to like onboard it, even if they haven't read all the books.

58:07and I think it's of that ambition to like try to try to do the work encoded into a meme almost yeah yeah so they can be digested and shared yeah anyway you need the law you need the long form there's this great blog post talking theory 201 size does matter by stevie okay that's about why like the people who change the world with their writing all write really long blog posts and it's basically like you just need some amount of time in someone's head to like we were talking about this earlier, like to install your agent, to install the voice. And so I think I just need to produce a lot of writing.

58:39And then you also need the pithy summary things, which are, which both are things the voice can say often in people's heads and also like enable a language for talking about your worldview that people who aren't soaking in it can like interact with. So the people who are like reading you don't sound like crazy people. And I think that's the, that's sort of what I want to work on next. I love that. I think that's great. Do you, you said something about Rick Rubin, how he's sort of the, I don't know how you would describe it. It's kind of like curator, but almost like a collaborator really with an artist to help them do their great work.

59:14Is Paul Graham, the Rick Rubin of the startup world? No. Paul is, Paul is more like the, Tony Robbins of the, I mean, In the best way, it's not so much, maybe not quite so much self-help-y, but the main thing that talking to Paul does to you repeatedly is like increase your ambition and drive. Like, and he has good ideas sometimes too. Like, don't get me wrong. Every now and then Paul is a really genius idea. But like mostly what I got out of talking to Paul was not necessarily the great idea that would like change the trajectory of the business, but the belief that I could go find it and that I was going to change the world and that I should be.

59:57What we were doing was important and worth investing in. And I got a bunch of other stuff, too, but that was so that was singularly so valuable. It like over overloads the other things I got out of it. how does he do that? Because, you know, when you say that, my head thinks of like a Tony Robbins, like a David Goggins, like sort of people that almost like push you, but he doesn't seem like that personality and reading all of his essays. He's not like that at all. So how does he get you to think bigger and push harder without being a rah, rah, rah, rah, rah, think bigger, push harder, right? You know what you should do is the classic Paul Grahamism.

1:00:33And it's always followed by a thing you could add on to what you're doing to turn it from project a addressing this small thing to project b changing the you know the universe all transportation we're going to manage power what if you've tried to power all transportation instead of like building a wheel but that's as right as you know what you should do you know what you should do is yeah yeah if you talk to paul you know i've never met you know what you should do you know you should do that's that that is the consistent paulism he i don't say delude because it sounds mean but like it's i was like he deludes himself about your business and how great you are and invites you to join him in this deluded vision of like interpreting what you're doing in the biggest, best possible light.

1:01:16And from that vantage point, what you're doing is super like, what if it does? What if it goes right? It's sort of what he invites you to ask, right? What if don't stop? Stop asking yourself. Stop seeing all the hard problems, all the shit you have to do. Ask what if what if what we're doing works? What if it goes right? What if it goes right and we like keep going? Like, what could it be? And when you spend time there, you see how the small things can turn out to be. Microsoft was building programming languages for like these hobbyist microcomputers. That was a tiny, irrelevant market that turned out to be extremely important.

1:01:55And that's generally true of all the big businesses. But they would they start out doing the important startups. They started doing something small and that seems almost trivial. But there's a way in which this trivial thing can be seen bigger. He sees it early. No, he sees he sees things have nothing to do with the way you'll actually be big early. But he sees a bunch of ways you could be big. No one can do that. No one actually knows if they knew it. They just go do then they'd be the the the prophet, the Oracle. What did he say? Let's say for Justin TV or what's what's when we could be Reddit or Justin TV.

1:02:26I remember one of them was like, you should like go hire all the like reality TV stars and make get them to go beyond Justin TV. You could be you could just take over all the unscripted stuff. That turns out to be just a terrible idea for a bunch of reasons. But like it recontextualized what we were doing for me in terms of like we're not making a on the Internet live streaming show. we might be building like just the way that you make unscripted entertainment generally. And that's like much bigger idea. And we were making a calendar and for my first startup, I remember this, you know, what you should do is make it like programmable so that people can add in and out functionality.

1:03:09So it can like talk to your to-do list and your, your email and your like everything else in your life. And then it could be your calendar in some ways, like that's everything you're doing. What if it was like the central hub of like your entire online information management system. That's also a bad idea. Like your calendar shouldn't be that, but like, but like, but a calendar could, but what if it was? And you walk away and I am implicitly by saying that, what he's telling you is, I believe you are the kind of founders who could build an information management system that controls, that takes over people's entire, like solves the entire problem for them.

1:03:45Does their, takes over all their information and manages it for them. You're not just like building a like Google account, like what you will find out later is a Google Calendar clone before Google Calendar is launched. You're not just like, you're not just building an Outlook clone in JavaScript. You're like changing the way people relate to information. I'm like, is that true? It's neither true nor false. That is not a true or false statement, but it's a way to contextualize what you're doing. It's the Sonic Zupari quote of like, don't teach them to like carry wood or build ships, teach them to yearn for the vast and endless sea.

1:04:18like Paul teaches you to see how you could be a changer of the world and how what you're doing is part of like this grand, like building of the future. And like the ideas I'll repeat here, both of those ideas are bad, but they were very helpful because they made me feel like what we were doing was important, that Paul believed that I could do something big and important. And they caused me to, even though I wound up rejecting them, look for those ideas, like to be open to and looking for, because you would get one every like, like you'd get like three an hour. Paul is a faucet for these. It's easy.

1:04:56I can do it for startups too. Now, if I want to, I learned the trick and I should do that more often. I'm usually what fall into the tactical stuff. But by, by having that happen when he, once he's, once you've rejected 10 of those, you can't help but start hearing the Paul, you know what you should do in your own head. the ceiling has been raised yes of like well maybe i should recontextualize my to-do list as like an email client like what why is email and to-do separate like maybe i should should be building something much bigger than what i'm building and in a way that doesn't require me to change anything maybe what i've built is already almost that if i just like think about it a different way it's this funny balance there actually i had a tweet thread about this recently between like, you know, small plans have no power to stir men's souls, plan big or go home.

1:05:42You should be really ambitious and aim super big and like only do projects that are really, that you could be, that you can see being super big and super important. And then the other hand, the fundamental truth that like, you know, big trees grow from small acorns and like most of the, many of the best things when they get started, the person is not thinking, I'm going to go take over the world. They're just trying to do a good thing that like they think is good, often just, often for themselves even, or for like a very small number of other people. And then it turns out that that's much, much bigger than they realized.

1:06:16And those are both true pieces of advice, like different people need to hear in different contexts, like, but they kind of contradict each other. Yeah. What about these other people? So you've had a privilege. I asked about Paul Graham. You've also been friends with, you were in the first YC batch. So you're friends with Reddit guys I think, you know, the Colson brothers, Sam Altman. Let's give me like a rapid fire on them of like what makes them unique. Like you said about Paul, what what his kind of superpower is, what really stands out, what something you admire about the way he does things.

1:06:47Give me one about maybe Steve from Reddit. Yeah. So it's easier in some ways with Paul because like he was a mentor to me. Right. And Steve was much more like my something like my brother in startups. Right. Growing up with Paul, I know I know the things that he like taught me because it was it was. It was much more of an explicit, like I was being taught by Paul. With Steve, it's like I learned things from him by like watching and imitating. I think like I actually learned a lot from Steve on management by watching his kind of unflappability. Like Steve is not like an unpassionate person and like, well, we can get angry or can get sad or whatever.

1:07:28But like when there's a crisis happening or there's just I got to shadow him for a day And when bad news is delivered, he responded. He wasn't like moved. He was like still grounded in response to that thing and was curious, asked questions like didn't jump to what to do about it, but then also like ended the meeting with like, all right, well, here's what we should do. Here's what we're going to do. And like, it was just sort of a masterclass. Like this is, this is when you, when something, someone brings something up, it's got to be anxiety provoking. It's like bad news. that's what it looks like when a leader is engaged, but not like not activated.

1:08:05And like, I think I, in my own leadership to sometimes success and sometimes failure, I think try to imitate that when I receive that, you know, when I have something like that, that in that state, when you say you shattered him, what was that? Like you guys just said, Hey, we, we exchanged like, like going to each other's offices and like sitting through every like early on or like, maybe like five years ago, four years ago. It was really cool. I did it with Justin. Me, Justin, and Steve all like shadowed each other. It was pretty fun. I learned a lot. That's incredible to like go watch another CEO at work.

1:08:35And like you have to have the, I don't know how you have the kind of like trust relationship to make that happen without like knowing someone for 15 years. And I happen to have the privilege to like know a bunch of CEOs for a really long time. And getting to go shadow each other was like a real learning thing. What do you think, even if these people didn't, let's say explicitly teach you things, you know i like you know if i read a biography or whatever one of the things i always try to figure out is more like to what extent is this person sort of built different or operates differently than like even somebody who's very good like the difference between very good and sort of like the elite what is the the best of the best at this craft versus somebody who's very good certainly very good but just not the same what is those like the diff is what i'm always most interested in i'm curious you've been around a lot of these like high-performing people even like, you know, Bezos, you've interacted with him.

1:09:26Like, do you notice any of these diffs or is it all just like? It's hard. It's hard to say like that. I think I believe more in contextualization, like that I see people do really amazing at something. But like when it's especially when it's your own company, there's a lot of like you happen to fit this problem well, and it's not general. I don't know how to generalize it. I don't know of anyone else even performing at this problem. The CEO of Stripe job is a very specific job and Patrick's amazing at it. Would he be equally amazing at some other CEO job? Possibly, but I've never seen him do that.

1:10:04I've never seen anyone else be CEO of Stripe. And it's very hard for me to - Is it true at the beginning? Like, is it true as like startup founder of Ambitious Company? Are those, is Stripe different at that stage too? Yeah, no, absolutely. The people who are really good, you can sense the energy and the drive and the capability and just the pace. There's like, it tends to like stuff happens a lot, but like usually, but then not always like some problems don't actually give weight. Like Stripe is a good example of a company that gives way to a high energy, high pace thing, because it's, it's a simple problem at some level that has infinite details as it could be right.

1:10:41But I think like, I don't know if that approach would work as well. if you're trying to create open AI or Anthropic, where it's a research-oriented organization. And you kind of have to be a little more patient in forcing it's impossible. And so I really believe in like fit, the different people are good at different things. And like, obviously someone's A plus, Patrick's obviously A plus at being a Stripe CEO. And it's just hard to tell the reason which these things are transferable. We don't really know. But I actually, one thing did come to mind about this question in terms of like a capability that I do think is generic, that I did see Bezos exhibit where I was like, oh, that's a thing that I'm good at, but he is better at, that I'm better than most people, but he's better than me, which is we present him on Twitch probably twice a year, once, twice a year for the first three, four years I was at Amazon.

1:11:27And every time two things would happen. First of all, he would remember everything we told him the first meeting. And I don't think he was like reviewing extensive notes someone else took because I don't know when he would have the time to do that. I observed him going from meeting to meeting and he did not review notes. I think he just remembered at least the high points. And the other thing was consistently, he would read our plan and he would then ask a question about why we didn't do a certain thing or give us an idea for a thing we could do that I hadn't thought of before. Once it's a bunch of things I had usually.

1:12:01And then at least once, which is hard to do, because all you do is think that never happens. Most people would be lucky to get one of those one ever, let alone one a year would be great. Like if you did it once a year or even once every three years, right? He could just like, you just generate them. And they were, and they were not all bad ideas either. They were new ideas, but I think I had, I generate a lot of ideas to get a new idea. I haven't thought, thought of on a topic I've been thinking about for a decade that might even be a good idea. That is like, he's just really fucking smart as far as I can tell.

1:12:36Like, I don't know how he does that. Can you say a story of one of those as like the statute of limitations passed five years ago? I'm trying to remember. I can't honestly, I don't remember the specifics anymore. I just remember the like the like, what the fuck moment? Like, because the first time I was just like, oh, he's smart. Like he's seeing Twitch for the first time. A lot of times smart people have one good idea about your business the first time they see it because they have this huge history and their pattern matching you to some historical thing they've seen. And like that combination yields one new insight.

1:13:04But then he did it the second time. I remember the second time I was just like, what is going on? This doesn't make any sense. Like, nope, I've never had that experience before ever. Andy does not have the new idea generation capability in the same way, but he does have the like, remember what you told him thing, which is also extremely impressive. Like that's, that's, and Andy has this other thing he can do that I think is, there's another, Andy also has a, it's easier for me with people I've like reported to or I've learned from Andy Jassy. Andy Jassy. Yeah. Yeah. And it has this like ability to criticize you in a way that conveys 100%.

1:13:40I know that you're amazing. I know that your plan is good or, you know, like, or that you are at least capable of making a really good plan. I know that you're working really hard and I know that you are smart and you have a great team and we have a huge opportunity. And yet somehow your results are bullshit, which must, I don't know what's wrong, but we're in this together and we're going to, I have your back, but like, I, but I'm confused. Like, why aren't the results better given how amazing you are and you feel supported? Like you feel like he, he believes in you, but, but like, but he's just, he's, you're so sad.

1:14:19Oh, I'm sorry. I've confused. I've, I'm sorry. I've failed. Even though I clearly can succeed at this. I'm going to go, I'm going to go like fix this now. And like, it's almost like, instead of looking at this and then judging you, he comes to your side of the table and says, what is this? Yeah. And like, how did we wind up here? Like how I have failed you that I didn't say something earlier, like something, I don't know, but like not, and that can come off for some people when they do that, it comes off as insincere or it comes off as like, they don't think you're actually competent. Like how did I not catch this can come off as I don't blame you because you're clearly not good enough to have caught this.

1:14:54Like, he really is, how did we, how did we wind up here? I know that we are working together. We're on the same team. How did we wind up with not the results we wanted with a plan that I thought we both thought seemed good? Like, help me understand. And because it, because it is genuine, it's super effective. At least it's effective. I don't know if it's effective. It's super effective on me. And I saw it be effective on other people as well. So I know it works on some number of people. Right. And that's another one of those things I've tried. I've tried to become good at. I'm not as good at it as Andy is, but I've certainly gotten better.

1:15:25So that's something to learn from. That's great. Love that one. Dude, thanks for doing this. I know I've been bothering you to do this for a long time because I love hearing your stories, love hearing the way you think. It's very different than most people I run into. Even here in Silicon Valley, where you're supposed to have this kind of very unique, diverse set of minds, you know, you're one of them. You're one of the reasons I moved out to San Francisco was to meet people like you. So thanks for doing this. Thank you. I really appreciate that. It's beautiful. And I really appreciate being able to come on the podcast.

1:15:54I feel like I can rule the world. I know I could be what I want to. I put my all in it like no days off. On the road, let's travel, never looking back.

1:16:09All right, this episode is brought to you by Mercury. They are the finance platform of choice for over 200 ,000 companies. Shouldn't be surprised because I use it myself for not one, not two, but I have eight different Mercury accounts. I have seven for different companies that I'm a part of, and then I have my own personal account because now they have personal banking, which is a really cool feature. I highly, highly recommend it. Like I said, I use it myself. And the reason why is because the way that Mercury works is beautiful. It's very intuitive. And you could tell that it's actually made by a startup founder.

1:16:36It's an entrepreneur. You could tell it's made by somebody who used other banking products in the past and didn't like all the different rough edges and annoyances and decided to actually fix it himself. And really, any type of entrepreneur you are, let's say you're an agency, well, one of the things every agency has to do is be able to send invoices, easily create them, send them to customers, and stay current on your balances with all your customers. Well, you can do that inside Mercury. And so I think that Mercury is great. Highly recommend you check it out. And thank you for sponsoring the show.

1:17:02For more information, check out Mercury.com. Mercury is a financial technology company, not a bank. Check show notes for details.