Sam Altman Scales OpenAI Compute Despite Capital Hurdles

顶住资金压力，Sam Altman 大力扩张 OpenAI 算力规模 OpenAI 将其算力基础设施扩展为史上耗资最高的项目，原因在于非共识性的 AI 押注需要大规模的上游投资，而非遵循传统的初创公司模式。 OpenAI 首席执行官 Sam Altman 探讨了 OpenAI 的创立与发展历程，详细讲述了在缺乏客户反馈的情况下经历数年混乱探索，最终验证缩放定律（scaling laws）和大语言模型的过程。他深入剖析了算力扩张为何正迅速成为人类历史上耗资最大的基础设施工程，迫使 OpenAI 保持高度自律的平台战略，果断砍掉颇有前景的边缘项目以保障算力和人才。Altman 还论及 AI 安全与经济颠覆，指出制度惯性将起到稳定的缓冲作用，同时警告 AI 必须赋能人类自主权，而非将权力集中在技术官僚手中。回顾个人职业生涯，他结合 Paul Graham 和 Peter Thiel 等导师传授的运营经验，指出创业者通过剖析罕见的成功所学到的经验，远多于分析失败教训。 Sam Altman 预计，OpenAI 在芯片、晶圆厂、数据中心和电网等领域的算力扩张，正迅速演变为人类历史上资本密集度最高的基础设施项目。 (10:36) OpenAI 严格执行“双支柱”平台战略——即单一消费者界面与开发者 API，并主动放弃了 Atlas 浏览器和 Sora 等颇具前景的内部项目，以集中算力与工程资源。 (51:30) Altman 借鉴 FAA 的航空安全报告机制来推进 AI 对齐，他认为生存与社会层面的风险无法在象牙塔中通过纯理论解决，必须依赖持续且透明的迭代部署。 (36:11) Altman 预计 AI 最大的宏观经济影响将是推动小型企业创立出现空前激增，因为它从根本上降低了传统创办公司所需的技术门槛、资金与阻力。 (43:14)

Sam Altman argues that artificial intelligence must be pursued as an asymmetric, power-law research bet developed through real-world iterative deployment rather than cloistered caution. Drawing on his venture investing background, he contends that societal inertia, extreme compute constraints, and the necessity of human agency will govern AI's economic integration far more than speculative panic. Economic Inertia and Habitual Friction (2:49) Altman reflects that following the release of GPT-4 in 2023, he overestimated the speed of economic disruption across software. He observes that both consumers and institutions possess immense inertia, routinely buying from existing vendors and maintaining established workflows. Rather than seeing this friction as a systemic failure, Altman views it as a stabilizing force that will allow society to navigate a massive technological shift through a smoother, slower transition. He acknowledges that this resistance to change extends to his own personal habits. Despite developing tools like Codex, Altman notes that he still sorts his inbox manually, hunts for the least painful emails to answer, and copies text between applications. He attributes this to a psychological bias that equates manual computer operation with real work, illustrating that human behavioral patterns evolve far more slowly than model capabilities. Applying Venture Power Laws to Frontier Research (11:56) Altman views leading an AI research lab as the operational equivalent of startup investing, as both environments are governed by extreme power-law distributions where a single non-consensus breakthrough eclipses all other efforts. When OpenAI began in late 2015, academic and industry consensus dismissed artificial general intelligence and deep learning as dead ends—mirroring warnings Altman received at Stanford in 2005. Success in this paradigm, Altman argues, depends on hiring non-standard, independent researchers rather than conventional consensus-followers. Because OpenAI spent its first four and a half years doing research without a shipping product, the organization had to invent internal proxies for market feedback. To gauge whether hypotheses were actually succeeding in the absence of customers, the team introduced mechanisms like an internal competitive leaderboard during its reinforcement-learning work on the video game Dota 2, substituting clear research benchmarks for traditional consumer signals. Iterative Deployment and AI Safety Risks (34:10) Rejecting the deterministic claims of AI 'doomers' who argue that model alignment is an impossible engineering problem, Altman champions an iterative deployment philosophy derived from Y Combinator. By exposing frontier models to public use, observing real-world failure modes, and refining safety systems at scale, OpenAI aims to de-risk the technology incrementally. Altman compares this framework to the Federal Aviation Administration's accident-reporting model, where transparent, clear-eyed investigations of failures transformed aviation into an exceptionally safe system. At the same time, Altman warns of two diverging systemic risks for AI: a loss of control where society surrenders decision-making to an unsteerable system, and monopolistic centralization where an elite hoards access under the pretext of safety. He categorizes both outcomes as misanthropic, arguing that the technology should instead be democratized into the hands of the public to foster entrepreneurship and protect individual agency. Platform Discipline Under Compute Scarcity (51:30) Scaling compute infrastructure has become what Altman calls one of the most expensive infrastructure undertakings in history, demanding vast capital across custom chip design, fabrication logistics, and novel power systems. Faced with severe limits on compute, capital, and engineering talent, OpenAI has repeatedly killed promising initiatives—including the Sora video tool and the Atlas web browser—to channel focus toward core upstream model training and infrastructure. OpenAI's platform strategy centers on a minimalist architecture: a single direct interface (ChatGPT, which incorporated Codex) alongside a programmable developer API. Rather than building specialized vertical products that compete with its users, Altman aims to provide a general-purpose intelligence layer that powers millions of independent enterprises, while admitting that the AI industry must drastically improve public education to demystify the technology. While Altman envisions artificial intelligence driving a historic boom in small business creation and scientific discovery, significant hurdles remain unresolved: managing unprecedented capital and energy constraints, countering deep-seated public skepticism, and overcoming the persistent human habits that slow real-world adoption.

[5:45] Sam Altman: mostly a product failure. [10:36] Sam Altman: probably already or at least rapidly becoming the most expensive infrastructure project in history. [11:56] Sam Altman: Your best investment will outperform all of your other investments put together. Your second best investment will outperform everything else put together after that. [16:04] David Senra: The single most powerful pattern I have noticed is that successful people find value in unexpected places, and they do this by thinking about business from first principles instead of formulas. [18:20] Sam Altman: You can try all of these things. There's all these directions. The one thing we know doesn't work is deep learning. We tried that for a long time. You know, it's the kind of the most guaranteed way to have a bad career. [19:38] Sam Altman: Having done both, I think I can say it's like much much much harder to run a company than being an investor. [30:21] Sam Altman: We're going to give the world a cure to all disease and we're going to make stuff really cheap in exchange for people giving up their autonomy and impact over the future and power. [36:11] Sam Altman: the way you do things is to put things out into the world, get feedback from real customers, see where they break, see where they don't break. That is the way you make a good product. That is also the way you make a safe product. [39:53] David Senra: Everybody uses AI. Everybody hates AI. What the hell's going on there? [39:53] Sam Altman: Dear peasants, we will bequeath upon you these gifts of a cure for cancer and material wealth and some things and, you know, great entertainment. And you stop complaining and we'll make all the decisions about the future and just trust us, you know, we'll be benevolent dictators. [64:03] Sam Altman: When I look at the lessons of where I have failed at something, I learned something generic about grit and determination and something not to do. But most things don't work... there's a lot of reasons why things don't work, and I think it's harder to put together the correct causation. And when I've had something really work—when I understand what parts of Y Combinator really worked, or what parts of OpenAI really worked—trying to apply those lessons going forward has been much more helpful to me than trying to apply the anti-lessons of what didn't work. [73:15] Sam Altman: Well, what do we do now?

[0:02] I just brought up Toby Luque and the [0:04] fact that [0:05] I recorded with him previously. Why'd [0:06] you say that you think he's one of the [0:08] most interesting CEOs right now? [0:10] >> One of the things that [0:13] struck me the most about Toby is in the [0:15] very early days of AI and then at every [0:17] moment along the curve of its [0:19] development, he has been the most [0:21] forward-leaning CEO. He's in there like [0:24] writing the software himself. He is [0:26] experimenting with it. He like sends us [0:28] extremely detailed feedback on the [0:30] product offering, on the capabilities of [0:32] the models. He was before anybody else [0:34] was saying this, he was like, "We are [0:36] not an NPC company and thus we are going [0:38] to adopt agents. Otherwise, you know, [0:40] we're totally screwed. We're going to [0:41] build it ourselves." Every time I talk [0:43] to him, he is at the edge of what [0:45] anyone, CEO or not, is doing. He builds [0:47] himself. He understands it. He has like [0:49] a great deep feel and he is always [0:52] six [0:54] eight months ahead of like any other [0:55] CEO. [0:56] >> Do you remember when he wrote it and [0:57] it's probably like a year and a half [0:58] ago, maybe 2024, he wrote that letter [1:00] saying that like the first thing you [1:02] have to do is see if AI can solve your [1:04] problem. And then even back then it was [1:06] maybe I don't know, 18 months ago, [1:07] three, four months ago, people went [1:08] crazy. They thought it was ridiculous. [1:10] >> It was ridiculous. This is my point. [1:11] Like he's just consistently been ahead. [1:13] He has been correct. He's leaned in. [1:14] He's very no So there's like [1:17] no hype. There's nothing other than [1:18] like, "Here's what it can really do [1:19] right now. Here's what I think it'll be [1:21] able to do soon. Here's where I'm going [1:22] to push the company here." And just like [1:24] extremely deep understanding of where [1:27] it's at. [1:28] >> I never even thought of that. How much [1:30] of a benefit it has to be for somebody [1:31] in your position where you have somebody [1:33] like that giving you intense and very [1:35] direct and clear product feedback. [1:37] >> A lot of people send product feedback. [1:40] He is the only person at the [1:41] intersection of like CEO of a large [1:43] company and extremely accurate, [1:46] detailed, on the cutting edge feedback. [1:48] >> Yeah, he told me, I don't know if it was [1:49] on the episode. They were if it was in [1:51] the episode or if it was after, but he [1:53] said he was very [1:55] adamant. He's like, "We're going to look [1:56] back in 2026." And I I want actually [1:58] your opinion on this. I didn't even [1:59] think to talk to you about this. We're [2:00] going to look back on 2026 as the year [2:02] that every business was up for grabs. He [2:04] said, "So, somebody is going to build [2:06] the AI native version of Shopify." And [2:09] he said, "And it's going to be me." And [2:11] so, at night, he is literally trying to [2:13] rebuild if if you started from scratch, [2:15] what would you do with the current [2:16] technology? [2:17] >> That was the other thing I was going to [2:18] say about him is he does it himself. [2:20] Like he is using these tools himself. [2:22] He's writing software himself. He's [2:23] trying the models himself. He is trying [2:25] to like reimagine his work flows [2:28] himself. Most CEOs, when you get to that [2:30] level, have like teams of people that [2:34] are managing teams of people that are [2:35] trying to implement the thing and [2:36] they're trying to like make you happy [2:37] and they're trying to like, you know, [2:39] smooth the rough edges. I think it's [2:40] very hard to get the feel if you are not [2:43] actually doing the thing and he he does [2:45] it so hands-on all night long as far as [2:47] I can tell. [2:49] I'm not sure if 2026 will be the year [2:50] that every business feels up for grabs. [2:52] I might [2:54] disagree with him a little bit there, [2:55] uh, but [2:57] but I I I I get the spirit of that and I [2:59] do understand that it feels like that's [3:00] happening. [3:01] >> Do you think that's even possible? Like [3:03] whether it's 2026 or [3:05] 2046. [3:07] >> I mean, obviously not literally every [3:08] business. I think there are some things [3:10] that are very anti-AI. Like the better [3:13] AI gets, the more some businesses [3:18] that have nothing to do with AI, I think [3:19] will be harder to compete with because [3:20] we'll really want these like [3:23] authentic, non-technological experiences [3:26] or we'll care more about [3:29] sports teams or whatever. So, no, not [3:31] not everything, but I think there will [3:32] be many software businesses that are [3:33] very up for grabs. [3:34] >> Would you disagree that on the timeline [3:36] then? [3:36] >> Yeah, I disagree on the timeline. I [3:38] think it's going to take a little bit [3:38] longer. [3:39] >> Okay, can you say more about that? [3:41] >> I love startups. Like I think startups [3:43] are the coolest [3:45] thing in the economy and uh, I've spent [3:47] my career trying to like really [3:48] understand startups. [3:50] And I thought [3:52] when [3:54] we got to GPT-4, which was back in 20 [3:57] 23, I think, [3:59] uh, that very quickly after that there [4:01] was going to be much more disruption in [4:05] software business being up for grabs [4:07] right right away than turned out to be. [4:10] And the thing that [4:12] I think I was wrong about a few things, [4:14] uh, but one of them [4:16] in terms of the speed, one of them is [4:18] the economy just has so much inertia. [4:20] People keep doing the same things [4:21] they're doing. They keep buying from the [4:22] same, [4:24] uh, you know, company. They keep sort of [4:25] wanting to use their tools in the same [4:27] way. I think that's actually a positive [4:29] in many ways and it's going to make this [4:30] big transition in front of us go [4:32] smoother and slower. I'm grateful for [4:34] it, but I think it means we've all been [4:36] too ambitious on timelines even with [4:39] this incredible technology. I think AI [4:40] is one of the most incredible [4:41] technologies humanity's ever invented. [4:44] Society and the economy will adapt more [4:46] slowly. [4:47] >> Yeah, it's funny. We were talking before [4:48] we started recording that there's all [4:50] these parallels to the history. [4:51] Obviously, I read history for a living. [4:53] When you were just talking, I wasn't [4:54] even thinking about OpenAI and AI and [4:57] Sam Altman. I was thinking of like [4:58] reading this biography of Larry Ellison [5:00] in like the '80s. He was just like, [5:01] "Guys, this isn't a software problem [5:03] problem. It's a people problem. We have [5:04] to convince them. Like we can install [5:06] software. They They're not using it. We [5:08] have to change their behavior. The [5:09] technology is there. It's like we have [5:11] to now adapt humans so they actually [5:13] start using the technology." [5:14] >> My own example of this was, uh, [5:16] after Netflix came out and started [5:18] shipping DVDs [5:20] even before they started streaming, it [5:22] was amazing to me that people still went [5:24] to Blockbuster. [5:25] It was incredible to me. Like I would [5:27] just watch this cuz I kind of I drove by [5:28] a Blockbuster on my way to and from [5:30] school. And it was amazing to me that [5:32] people still did it. [5:34] And [5:35] you know, like that is an example that [5:37] has stuck in my head of like force of [5:39] habit in the way people do things. And [5:41] changing behavior is just much harder [5:43] than like the tech nerds realize. So, if [5:45] we go back to this uproar of Toby [5:48] writing that, you know, open letter or [5:50] the letter to the people inside of his [5:52] company, [5:53] you're adopting this faster than anybody [5:55] else because you're partially inventing [5:56] them, right? So, are is there something [5:58] where you're actually shocked at your [5:59] own behavior or like, I know there's a [6:01] better way to do this. I'm even creating [6:03] the product that could be better and yet [6:04] I still can't get over this like habit, [6:06] this force of habit? [6:07] >> 100%. [6:07] >> Okay. [6:08] >> I I love it. I don't think anyone's ever [6:09] asked me this before. I've been waiting [6:11] for this question. [6:13] The thing to me that feels most [6:15] psychologically inconsistent [6:18] about myself [6:19] is that I have [6:22] for 20 years been using computers the [6:24] same way. [6:26] I now have [6:28] a magic thing called Codex. [6:31] So do you, so does everybody. That means [6:33] I should completely be using my computer [6:36] in a different way. I should not be [6:37] clicking around [6:39] uh, you know, pasting from one messaging [6:41] app to another. I should not be [6:42] scrolling mindlessly through my emails [6:44] and trying to figure out which one is [6:45] like least painful for me to open and [6:47] respond when I don't want to be dealing [6:48] with it. I should not be like keeping a [6:51] to-do list and doing sort of this like [6:53] these wrote computer tasks in the same [6:55] way that I have for so long. And yet, [6:57] there's like something in my mind that [6:59] is encoded that like doing this kind of [7:01] stuff is what it means to work and what [7:02] it means to be productive. And [7:06] if you asked me, I would never say I [7:08] like doing it that way. I would in fact [7:10] I would say the opposite and I and I [7:11] think I would mean it. But like by [7:13] revealed preference, I have a better way [7:15] to do it now. [7:17] I can do it faster. I can be using Codex [7:19] for more of just like my day-to-day like [7:22] got to get through this stack of emails, [7:23] got to do this stuff on my to-do list, [7:25] got to, you know, deal with all these [7:26] things. [7:27] And I still do it that way and it makes [7:29] no sense other than I must like secretly [7:31] like like it or feel good about it. [7:33] >> What do you think is going to has to [7:34] change for you to actually adopt your [7:36] own product in a more deep way? [7:39] >> I don't really know. I I mean, it's [7:41] happening gradually and this might be [7:43] the right answer, which is these things [7:45] have to happen gradually and totally [7:46] changing someone's like ingrained habits [7:48] and [7:49] workflows is difficult. Uh [7:52] I think there are better products we can [7:54] build with this technology that will [7:56] make it more seamless to do that, but [7:58] right now it feels like [8:01] we're all kind of straddling these two [8:03] worlds of, you know, we still have a [8:04] computer we can use the old way and we [8:06] have Codex that can use our computer in [8:07] this amazing new way and we're like not [8:10] sure which to use when for what. And I [8:13] think this is mostly a product [8:15] failure. The phase that we're in now [8:16] reminds me of [8:18] like smartphones before the iPhone. I [8:20] was like an early adopter. I had like a [8:22] Palm Trio in, you know, 2003 or 4 or [8:24] whatever. [8:25] >> a Sidekick? [8:26] >> I never had a Sidekick. I thought they [8:27] were super cool. I wanted one. I never [8:29] had one. [8:29] And a lot of the technology was there. [8:31] Like it was missing multi-touch, but [8:33] mostly it was missing like [8:35] the product ideas that made the iPhone [8:37] the iPhone. [8:38] And I feel like we are now in a world [8:40] where we have all of the technological [8:43] pieces, but we have not had the iPhone [8:45] moment of like completely changing how [8:47] someone interfaces with technology. [8:49] >> We were talking about Toby's like Toby's [8:50] out here building these himself, right? [8:53] Uh we have a mutual friend in Josh [8:54] Kushner. He says he's like there's a [8:56] there's a big comparison to be made [8:57] between the way that Steve Jobs thought [9:00] and the way he ran his company to way [9:01] that he thinks he that you do. He wasn't [9:03] the one obviously writing the code, he [9:05] wasn't building the hardware, but he [9:06] he's like I am patient zero. I am making [9:08] products that I myself want to use and [9:11] essentially like everything that we saw [9:12] with Apple was just basically what he [9:14] wanted. There's this great story in one [9:15] of the books um [9:16] where the [9:18] they he they were supposed to have a [9:19] meeting on uh I think one of the MacBook [9:22] like the new MacBook laptops and the [9:24] team prepares like all this huge [9:26] presentation for Steve and they're like [9:28] really nervous cuz of his commanding [9:30] presence and he walks in and they think [9:32] it's going to be like an hour meeting. [9:33] He walks in and he shows them the laptop [9:36] and he's like on [9:38] off. You press the button and it comes [9:39] on. Press off like immediately. And then [9:41] he then he tries to open up the the [9:43] MacBook. There's like a delay. He goes, [9:45] "Make this." Meaning the MacBook like [9:47] that. [9:48] And then walks out the room. And that's [9:49] the whole meeting. Like that there's a [9:51] lot of examples in uh the history of [9:53] Apple like that. Uh how do you approach [9:54] it? Like how do you improve the product? [9:56] Like are you just doing it through your [9:57] own needs? Like how do you think about [9:59] this? [9:59] >> Most of my effort right now is on [10:02] research and compute. I would love to be [10:04] able to spend more time on product. We [10:05] have great people thinking about the [10:07] product here, but the most important [10:09] thing that we can do is to create smart [10:12] models and to be able to run them [10:14] efficiently and abundantly for a lot of [10:16] people. [10:17] If we can get that right, I believe that [10:19] everything else will follow. [10:20] Philosophically, I'm very inclined to [10:22] say, you know, try to find the like [10:24] the high-leverage difficult problem that [10:27] will continue the exponential. And for [10:29] us, this is like models and compute. I [10:32] also just think those are like [10:34] problems that naturally suit me. [10:36] >> Why do they naturally suit you? [10:38] >> To scale compute in the way that we're [10:39] doing this requires like [10:43] it's a complex supply chain. There's [10:45] like a lot of interesting partnerships [10:47] uh to figure out which I like doing. [10:49] There's like interesting financial [10:51] challenges of how you're going to [10:53] uh [10:54] finance what is [10:55] probably already or at least rapidly [10:57] becoming the most expensive [10:59] infrastructure project in history. Um [11:02] the technology questions that go into [11:03] building out compute at this scale from, [11:05] you know, designing your own chip to the [11:06] supply chain of [11:08] fabs and people that make racks to to [11:10] sort of the [11:11] the power systems for these things. I've [11:13] always been interested in energy. Uh all [11:14] come together. So, there are a lot of [11:15] problems that are interesting across [11:18] technology, business, policy, supply [11:20] chain, logistics all together around [11:22] building compute at this kind of scale. [11:24] So, I used to be a startup investor. And [11:26] the thing in my career that I have found [11:27] closest to startup investing is managing [11:30] a research program. [11:32] There are all these ways in which [11:33] they're really different, too. Like, you [11:35] know, the average researcher and the [11:36] average founder have [11:38] I think, on the surface, look different [11:40] for obvious reasons. But, there's like a [11:41] lot of similarities about how you find [11:43] the non-consensus bets, how you decide [11:46] where to have conviction, how you [11:47] understand what exponential growth looks [11:48] like, how you manage like outlier [11:51] talent, and how you how you identify [11:52] even more. This is the research building [11:54] that we're in. It's where I sit. [11:56] >> Great. Say more about why the parallels [11:58] between what you learned at Surbiton [11:59] testing with [12:01] doing research. [12:02] >> One big one is the power law. [12:04] Some people talk about this all the time [12:05] in investing, which is you have to kind [12:08] of like reprogram your brain cuz we [12:09] don't seem naturally well suited to [12:11] think this way. [12:12] Um [12:13] where, you know, your best investment [12:16] will outperform all of your other [12:18] investments put together. Your second [12:19] best investment will outperform [12:20] everything else put together after that. [12:22] And AI research, at least, is like that [12:25] as well. When we started, people thought [12:27] it was totally unlikely or almost [12:29] impossible that AGI was possible. [12:31] >> What year is this? [12:33] >> 2015. End of 2015. I mean, we just got, [12:35] you know, hammered [12:37] in the [12:39] by like all of the intellectual giants [12:41] of the field for saying that we were [12:42] going after AGI. And then when we [12:44] started really focusing on large [12:46] language models, we got hammered again [12:47] saying this is completely ridiculous. [12:50] And we understood [12:53] or I understood, at least, from my kind [12:54] of like startup background, and I think [12:56] other people understood in other ways, [12:58] this point of high [13:01] risk bets are okay as long as you take [13:03] the ones where if they work, it's super [13:05] valuable. And research research looks [13:07] this way. The kind of people that make [13:09] great researchers are sort of [13:11] non-consensus, [13:13] fresh approach, high energy, sort of [13:17] non-standard is the word that keeps [13:18] coming to mind, uh people. [13:20] >> You got to say more about non-standard. [13:22] Can you be more specific? Are they [13:23] spiky? [13:24] >> You don't want to fund a founder who has [13:30] a very slightly different take [13:34] on the same idea as the last thousand [13:35] people you talked to has tried to [13:37] convince you and maybe convince [13:38] themselves that somehow they're [13:40] completely different and doing something [13:41] totally new but it's mostly like trying [13:44] to [13:45] you know, fit in with the herd and be on [13:48] the same track as everybody else and do [13:49] what they're supposed to do which is [13:51] start a startup and you know, they've [13:53] heard Peter Thiel say enough times that [13:55] you know, there's like something that [13:56] you're supposed to [13:58] be doing different that they kind of try [14:00] to emulate that but they don't really [14:01] mean it. Like it's it's very [14:04] clear to me when you have someone who [14:05] just thinks [14:07] differently than most other people and [14:09] has [clears throat] is is willing to [14:10] stand by convictions that are [14:13] very unpopular may well be wrong but if [14:15] right like at least they're going to be [14:17] really right versus someone who is like [14:19] a [14:20] you know, thin veneer on the same idea [14:23] that everybody else has. In 20 If in [14:25] late 2015 we were starting Open AI, [14:27] there were very few AGI efforts in the [14:28] world. There was DeepMind [14:31] one or two others that I can think of. [14:32] It was like a very non-consensus thing [14:34] to do. [14:35] In that same year [14:37] there were probably I'll just pick on it [14:39] just cuz it [14:41] it came to mind but there are other [14:42] categories too. There were probably many [14:44] many thousands of founders starting [14:45] photo sharing apps. [14:47] You know, that was probably like not as [14:48] good of a thing to do. Today [14:50] a lot of people want to start [14:52] AI labs. There are some handful of [14:55] people, you know, two, three, whatever, [14:57] doing something completely new [14:59] um that actually wasn't possible until [15:01] AI got this good but [15:04] uh [15:04] doesn't seem like a good idea yet. And [15:07] like that is the thing that [15:09] as a startup founder I always wanted to [15:10] fund and the thing that mostly worked [15:12] for me. There's a similar thing for [15:14] researchers. [15:15] There were a lot of researchers that [15:17] would chase whatever the last thing was [15:19] that worked and there were a number of [15:20] researchers that had high conviction [15:23] towards a new idea. [15:25] And uh [15:27] I think [15:28] we were and are the best research lab [15:31] for those people. [15:33] >> I remember talking to Demis about this a [15:34] few months ago and he thought, you know, [15:36] it's in terms of like chasing after the [15:37] way that you guys are. It's like, you [15:39] know, there's the big three players, the [15:40] the money, the capital required. Like [15:42] you're not There's not going to be like [15:43] a fourth bigger player. But he's like [15:44] there's like this 10 I forgot what the [15:45] number was. I'll just make it up. Say [15:47] 10% chance that there's just some monk [15:49] researcher that's going to approach in a [15:50] way that just this angle we've never [15:52] even considered. [15:53] >> Totally. I don't know how to put a [15:54] number on it, but there is some chance [15:56] >> think you I'm making the number up, but [15:57] it was like a small percentage. [15:58] >> There is some chance of that, for sure. [16:00] And I love that. Like I think that's why [16:03] stuff stays exciting. [16:04] >> I want to tell you about the presenting [16:05] sponsor of this podcast, Ramp. I have [16:07] been reading a lot about SpaceX lately. [16:10] SpaceX is one of the most valuable [16:11] businesses in the world, and one of the [16:13] main themes in the history of SpaceX is [16:15] constantly attacking and questioning [16:17] your cost. Ramp helps many of the most [16:19] innovative businesses in the world do [16:21] exactly that. The median company running [16:23] on Ramp cuts their expenses by 5%, and [16:26] one thing SpaceX has demonstrated is [16:28] that a religious dedication to [16:30] controlling costs can help actually [16:32] increase revenue because you can pursue [16:34] opportunities you couldn't otherwise. [16:36] And we see that in the Ramp data, too. [16:39] The median company running on Ramp also [16:41] grows their revenue by 16%. [16:44] So when you're running your business on [16:45] Ramp and your competitors are not, you [16:47] have a massive competitive advantage [16:49] that compounds over time. Ramp is the [16:51] only platform designed to make your [16:53] finance team faster and happier. Many of [16:55] the top founders and CEOs I know run [16:57] their business on Ramp. I run my [16:59] business on Ramp, and you should, too. [17:01] Go to ramp.com to learn how it can help [17:03] your business save time, save money, and [17:05] grow revenue. That is ramp.com. [17:09] I found one of my all-time favorite [17:10] quotes when I was reading the book Zero [17:11] to One. The quote says, "The single most [17:13] powerful pattern I have noticed is that [17:15] successful people find value [music] in [17:17] unexpected places and they do this by [17:20] thinking about business from first [17:22] principles instead of formulas. That is [17:24] exactly what AppLovin [music] has done [17:26] with their advertising platform. [17:27] AppLovin connects you with over a [17:29] billion potential new customers [music] [17:31] inside mobile games. AppLovin allows you [17:34] to capture undivided attention. AppLovin [17:37] ads are full-screen video ads that are [17:39] watched for an average of 35 seconds. [17:41] That is retention that blows other ad [17:43] platforms out of the water and you can [17:46] launch on AppLovin in minutes. You set [17:48] the goal and AppLovin achieves it. [17:51] There's no complex setup, no expertise [17:53] needed and AppLovin scales quickly. They [17:56] can put your ads in front of over a [17:58] billion potential customers. Other [18:01] businesses have seen immediate results, [18:03] have scaled [music] to hundreds of [18:05] thousands of dollars of spend per day [18:07] and increased their revenue by millions. [18:09] So, you want to get started quickly [18:11] before all of your competitors are on [18:13] AppLovin and you can do that by going to [18:15] applovin.com. That's applovin.com. [18:20] Okay, so what is confusing to me? You [18:21] went from founder to investor to back to [18:25] founder. But, why in 2015, like what got [18:28] you interested in artificial [18:30] intelligence to begin with that you're [18:31] saying, "Hey, this is such a [18:32] non-consensus thing. People think I'm [18:34] crazy. I'm going to do it [18:35] anyways." [18:37] >> Well, I had been interested in AI my [18:38] whole life. I was like a very nerdy kid. [18:41] I like, you know, I was like the kind of [18:42] kid that spent [18:44] Friday nights like playing on my [18:45] computer and watching sci-fi reading [18:47] sci-fi. And I always thought that AI [18:51] would be um [18:53] the most amazing kind of craziest thing. [18:55] I never thought I'd actually get to work [18:57] on it, but I always I always like loved [18:58] it. Uh I even came to college sort of to [19:01] study it. I worked in the AI lab [19:03] the summer between my freshman and [19:05] sophomore year. [19:06] And nothing was working. In fact, like [19:08] very memorably, a professor told me, [19:10] "You can try all of these things. [19:11] There's all these directions. The one [19:13] thing we know doesn't work is deep [19:14] learning. [19:15] We tried that for a long time. You know, [19:17] it's the it's the kind of the most [19:18] guaranteed way to have a bad career. [19:20] >> [snorts] [19:20] >> And you know, I was like an [19:21] impressionable freshman in college [19:23] whatever. I I like I assumed that was [19:24] true. [19:25] So, you know, pursue these other things. [19:28] It was clear to me at the time. [19:30] This is now like kind of 2005. [19:33] That AI was not working. [19:36] And [19:38] I happen to like accidentally get into [19:41] startups but then very much fell in love [19:42] with it and so I wouldn't even call it a [19:44] career detour cuz it was super helpful [19:47] uh [19:48] looking back and becoming like a startup [19:50] investor was was great. The sort of [19:51] normal career path in Silicon Valley is [19:54] you're like a [19:55] or not the a a a common path is that [19:58] you're a founder and then you like sort [19:59] of semi-retire and you become an [20:00] investor. [20:01] We don't want that. [20:03] >> I heard you say that you're going to [20:04] work on this for the rest of your [20:05] career. [20:05] >> Yeah. [20:06] >> I'm hopefully we're starting we're going [20:08] to come on the show multiple times. I'm [20:09] going to hold you to this. We don't need [20:10] more founders that retire and invest. [20:12] >> I don't want that. I don't want that. I [20:13] don't want We have too many investors. [20:15] >> Thank you. [20:15] >> I want to talk to you about this. But [20:17] what I was going to say is I the the [20:18] fact that I got to go in the other [20:19] direction. I was like an investor first. [20:22] And then I got into companies pretty [20:23] unusual. [20:24] >> Very unusual. [20:25] >> And I'm super grateful for it uh because [20:27] the [20:28] like you get this unbelievable [20:31] set of learnings and pattern matching if [20:33] you really study and watch companies as [20:35] an investor that has been super helpful [20:38] to me. [20:39] Running OpenAI. But it's [20:41] it's like it's the opposite of normal [20:43] direction. So, it's just like a very [20:44] rare thing and I strongly recommend it. [20:46] >> Why is it helpful? [20:47] >> If you're running a company, you have [20:50] faced some number of similar decisions [20:52] in your past like some number of like [20:53] crux decisions and you've seen what [20:55] works and what doesn't. And if you have [20:57] to like [20:59] you know, make a high-stakes strategy [21:00] shift or like fire an executive in a [21:03] really messy way. You have like whatever [21:05] your own limited previous experience was [21:07] over the last [21:08] five or 10 years that you've been doing [21:10] it. [21:10] But as an investor, [21:12] you kind of watch all the crux moments. [21:14] So, you don't get the kind of like [21:17] operating practice [21:19] that you do just day in and day out [21:20] running a company, but you've seen a lot [21:23] of the like big [21:25] crux moments, [21:26] a lot. Kind of all day long you see [21:28] those. So, just the the the wealth of [21:31] the data set that I have that was [21:32] awesome. [21:32] >> So, that plays in your head when you [21:34] have a decision to make? [21:35] >> Yeah. [21:36] I'm like, oh, this is what happened when [21:37] this company had a similar thing or I [21:38] saw this founder make this mistake or [21:40] this founder got it really right. [21:41] >> We were talking about that you studied [21:43] the Industrial Revolution. We were [21:44] talking about some great biographies [21:45] that we both read earlier. [21:47] Uh my friend Daniel X says this about me [21:48] cuz like I think the benefit of me doing [21:50] this project my other podcast called [21:51] Founders for 10 years is like [21:54] so, he's like you're like an LLM trained [21:56] on history of great entrepreneurs, but [21:58] with the temperature turned up cuz [21:59] you're crazy. [22:00] >> [laughter] [22:01] >> Cuz I'm like super passionate about it [22:02] in a weird way to be like obsessed with [22:04] dead entrepreneurs, uh [22:06] but it is helpful. It's like, oh, like [22:07] I'll be talking to a founder and they'll [22:09] talk about something they dealt with. [22:10] I'm like, oh, well, like Carnegie did [22:11] this and Rockefeller did this and like [22:12] you might want to try this. Like [22:14] >> And do you find that you have like one [22:16] big insight all that or it's just like [22:17] for any given scenario you have like [22:19] what all these people did and how it [22:20] comes together? [22:21] >> I think it's depending on the [22:22] personality of the founder, right? And [22:24] so like when I was reading I've read [22:25] your blog for years. I think like you're [22:26] a great writer. You're very succinct. I [22:28] like like how the brevity is just really [22:30] appealing to me and I love numbered [22:32] lists. It's weird that we both write in [22:33] the same way and I feel like I'm reading [22:36] your blog and I'm like this is the exact [22:37] conclusion that I would come to based on [22:39] all the reading. And like there's just a [22:41] handful of principles that could be [22:42] applied, but it's like it really depends [22:43] on like who the founder is and what they [22:45] want to do. This is what I'm trying to [22:46] say is like let's go back to what we [22:48] were talking about. I still like [22:50] like you're you're making a big jump cuz [22:52] you know, you're one of the from what I [22:53] hear one of the best investors of all [22:54] time in Silicon Valley. You could just [22:56] be rich and not really have to work cuz [22:58] investors are kind of lazy. Uh I'm just [23:00] kidding by the way. Uh kind of not. [23:02] >> pretty I Having done both, I think I can [23:04] say it's like much much much harder to [23:07] run a company than being an investor. [23:08] >> Exactly. And that's what you people [23:09] should be doing, in my opinion. [23:10] >> with that. [23:10] >> So, [23:11] then you're like, "Fuck that. I'm going [23:12] to not take the easy route. I'm going to [23:14] do the hardest thing ever, the thing [23:15] that people think is impossible, thing [23:17] you're going to be made fun of." [23:18] >> Yeah. [23:19] >> I still need to understand, okay, so [23:20] you're into it kid, like why would it [23:22] appeal to a kid? You were living in St. [23:24] Louis at the time. [23:24] >> Yeah, I was living in St. Louis. [23:25] >> Why would AI appeal to you back then? [23:27] >> Well, I think like it appealed to every [23:30] kind of computer nerd. Like I don't [23:31] think it's that unusual about me. It [23:33] just felt impossible. [23:34] I think most people would say like of [23:35] course that'd be the coolest thing ever, [23:37] but it's totally impossible. [23:38] I think the like the weird thing about [23:39] me was like, "Okay, let's try." [23:41] But I think everybody thought it [23:42] Everybody would think it's awesome and [23:44] something to something to go for. [23:45] >> So, wait, that was a personality trait [23:47] of yours as a kid that people told you [23:50] that you couldn't do something, like [23:51] your initial response was resistance? [23:55] >> Not resistance, but like [23:57] are you sure? Like why not? Let's try. [23:59] Let's see what happens. Maybe I can. [24:01] Maybe we can. I was like a very [24:02] optimistic kid. And also like the more [24:04] something seemed impossible, like the [24:06] more intrigued I was. The idea that we [24:08] could invent a technology that would let [24:11] us do everything else, that would just [24:12] empower people in this way that no other [24:14] single technology [24:16] could, that always seemed like innately [24:18] incredibly appealing to me. It's like I [24:20] want that thing. I want to be able to do [24:21] everything else. Like I think another [24:23] thing about that I kind of was like a [24:24] personality trait as long as I remember [24:26] is like, you know, [24:28] it's like it is interesting to like [24:30] really give people a lot more power, a [24:32] lot more ability. In some sense, this is [24:34] the whole arc of technology and I was [24:35] for sure always a technology nerd. [24:37] But like AI is the strongest version of [24:39] that I can imagine. [24:40] >> What did you think that it would enable [24:42] back then? Like when you were a kid, [24:44] you're like, "This seems like a cool [24:45] technology. I want to do X. I can't do X [24:49] unless AI [24:51] is invented?" [24:52] >> It's always hard to remember like how [24:53] much of this is the stuff that I [24:54] actually thought at the time versus like [24:56] what we're I was to build right Yeah, [24:58] like how much my current work has like [25:00] colored my memories of it for sure. [25:04] As a kid I was like very into robots. [25:06] You know, we had like a robots club in [25:08] my school [25:10] and [25:12] the robots at the time were like [25:13] laughably I even remember at summer camp [25:15] we had this like little turtle that you [25:16] could control with a computer on the [25:18] floor or on the table and thought that [25:19] was just the coolest thing. There's [25:20] something about like physical stuff [25:23] moving controlled by a computer [25:26] that I always thought was amazing. Now [25:28] I am extremely interested in what AI can [25:32] do to advance scientific discovery. [25:35] In my memories as an adult I think I [25:37] thought that was cool as a kid too but [25:38] it feels just implausible and I assume [25:40] that's an example of where like the [25:42] memories have gotten more colored but [25:43] now the fact that [25:45] we can have AI go discover [25:48] new physics and cure diseases and what [25:50] it's already doing for math like I think [25:51] this is I think this will be one of the [25:52] most important areas even more important [25:55] than automation of other tasks that AI [25:58] can do just to help us understand more [26:00] things. We were talking earlier about [26:01] this this book The Beginning of Infinity [26:03] and [26:04] rereading that book [26:06] from today's vantage point I'm like man, [26:08] AI's really going to help us do this do [26:10] this important thing of understanding [26:12] everything or as much as we can. [26:15] I was definitely interested in the sort [26:17] of like Star Trek version of like [26:20] huge prosperity and [26:22] abundance and and what AI could do to [26:23] drive that. Maybe the memory of being [26:25] interested in science is more real. I [26:26] also loved science and and just this [26:28] idea that we could like [26:30] because we were smart we could figure [26:32] out how to understand the world and [26:34] make predictions and do things that we [26:36] couldn't without this deep [26:36] understanding. I don't know, that seems [26:38] like an really awesome. [26:40] >> It's interesting how consistent over [26:41] time what humans want from AI cuz like [26:44] something you're describing is very [26:45] similar. I just reread the biography of [26:46] Claude Shannon for the second time and I [26:48] had forgotten cuz I hadn't read the book [26:50] in maybe 5 years and I forgot that him [26:52] and Alan Turing used to meet every day [26:54] for coffee when they were both at Bell [26:55] Labs. This is like 1940s. Yeah. And they [26:57] would just talk about AI. And they were [26:58] both obsessed with their they thought it [27:00] was inevitable back then. And they [27:01] thought it was going to happen like 15 [27:02] years from there. So like 1955 that [27:04] we're going to have computers, which [27:06] didn't exist, right? They had the analog [27:09] versions. [27:10] Uh that are that are going to be smarter [27:12] than humans. And that anybody that [27:13] thought that wasn't going to occur, they [27:15] thought it was absolutely ridiculous. [27:16] And they're like, "Well, what would you [27:16] want the computer to do?" He's like, [27:18] "Solve math problems, write poetry, like [27:21] solve uh cure diseases." Like you hear [27:23] this over and over again. I have read [27:25] a bunch of things that those guys wrote [27:27] at the time. [27:28] And I'm so sad they are not here to see [27:30] it. [27:31] Cuz they were so right about everything. [27:33] We're finally the moment where AI is [27:35] solving novel math problems. [27:37] It is discovering other stuff. It is, [27:40] you know, you can argue about how good [27:41] or not. I would say not very good, but [27:42] it is writing poetry. [27:43] It gets here like to what these guys I [27:46] think they would have said, "All right, [27:47] you've done it. Like this is it. We've [27:48] got it." Um and that would be so cool. [27:51] Yeah, this is the weird thing where [27:52] everybody's just like, "Oh, it'll never [27:53] do X." Like I talked to [27:55] uh people in the music industry. It's [27:56] like, "It's never going to make great [27:57] music." And and then they're like, [27:59] "Well, do you think it's going to like [28:01] make a podcast?" I was like, "Of course [28:02] it's going to." It's going to do [28:03] everything that we can do at least to [28:06] I would say better than like even right [28:08] now like better than what we can do. [28:09] It's a very bizarre [28:11] thing where it's like it will never [28:13] surpass what is happening at this [28:15] current point that I happen to be alive. [28:17] There's a deep human psychological flaw [28:18] there. But here is a I think a more [28:20] interesting question. Let's say it does [28:22] make a great podcast. [28:24] You know, two AIs are having a more [28:25] interesting conversation than you and I [28:27] are. [28:28] Do you think people will care? Or will [28:30] they want the one with the real people [28:31] because we're all obsessed with people [28:32] and the fact that it's not real people [28:34] Yeah, for this is like more interesting. [28:36] It's like, "Oh, these two people who [28:37] that I may be predisposed to like or [28:38] dislike are having a conversation that's [28:39] interesting to me." I think for like [28:41] strictly like reference, like maybe my [28:43] other podcast where I'm just saying, [28:44] "Hey, this is an interesting ideas I [28:45] read in this book." That could maybe get [28:47] disrupted over the cases, but [28:49] especially for people that were born [28:51] before this happened, maybe it's [28:52] different for your son, you know? [28:54] >> Maybe. [28:54] >> But for me, it's just like I think [28:55] humans are going to always be drawn to [28:57] humans. [28:57] >> I really deeply believe that. I think [28:59] there's like a lot of other jobs that [29:01] could face significant transition, but [29:03] stuff that's about people [29:05] stuff that's about people's connection [29:06] to connection to people and people [29:07] liking other people, that stuff feels [29:09] like it gets uh more valuable in a [29:11] post-AI world, not less. [29:12] >> I may be actually the wrong person to [29:14] talk about this [29:15] >> because I kind of deeply desire, even [29:17] though my entire work is digital, you [29:18] know, broadcast all over the world, [29:20] >> it's just like I deeply desire like more [29:23] like more of an analog life. I like [29:25] reading physical books. Like when I was [29:26] talking to Kelly, I was like, "I don't [29:28] want to get on Zoom. Call me. Or we'll [29:29] talk in person." Like I like physical [29:31] I don't like [29:32] >> like that, too. I don't read e-books. [29:33] >> Yeah. I don't like Zoom meetings. [29:35] I like to be like with people in the [29:37] real world. [29:38] >> I definitely think there's a subset of [29:39] weirdos, and there's probably a lot of [29:40] them that live in this city, that, you [29:42] know, don't like humans and only want to [29:44] communicate with computers, but it's [29:45] just like I think that's a tiny [29:46] percentage of humanity. [29:47] >> I think it's a tiny percentage of [29:48] humanity. I did This is why I think like [29:50] the world is on the whole not going to [29:52] be that different even with [29:53] superintelligence. Like people are still [29:55] going to be very fundamentally wired to [29:58] care about other people, to want to be [29:59] around other people, [30:01] um to interact with other people. And, [30:03] you know, there will be some people who [30:05] just like get obsessed with the models [30:06] and just [30:08] think humans are [30:10] in the way or you know, [30:13] danger to be contended with or whatever. [30:15] >> And for most of people, it'll be the [30:16] whole point. [30:16] >> I think it's very important that when we [30:17] do find people like that, to then be [30:19] called out and make sure they don't [30:20] acquire power. [30:21] >> I certainly agree with that. Maybe the [30:23] two big risks that I'm most worried [30:24] about with AI, which are a little bit in [30:25] tension, one is like a [30:28] a a loss of control, [30:30] where, you know, AI somehow just becomes [30:33] too powerful in a way that we can't [30:34] guarantee the control we want. And the [30:36] other is power gets too centralized, [30:38] where you have, you know, one company or [30:41] model or person with too much power. And [30:44] in both of these, the fundamental thing [30:46] is like [30:48] I think it's a very anti-human position [30:50] for either of these things to happen. [30:52] The right approach is to say like we [30:54] want people deeply in control of the [30:56] future. We want people deeply empowered. [30:58] People are the whole point of this all. [30:59] Like that this is we are not going to [31:01] sit here and, you know, gradually hand [31:03] over control to an AI model because we [31:05] don't trust or like people and you know, [31:07] uh [31:08] it's like a very misanthropic thing to [31:10] say we're going to just like put all of [31:11] our trust in this model and let it have [31:13] all the power and decision-making over [31:14] the world. But I think there are some [31:16] people in the world who [31:18] think that's the right outcome. There's [31:20] like another version of this which is [31:23] because we don't trust people, we have [31:25] to [31:28] limit who gets access to this technology [31:29] and how they can use it and all of these [31:31] terrible things could happen. And and, [31:34] you know, out of fear of those, we are [31:36] going to [31:38] concentrate power in the hands of a few [31:40] companies and they're going to, you [31:42] know, we're not going to let other [31:43] people use this. But we'll give them [31:44] some benefits. Like my my caricature of [31:46] this is [31:47] I think there are some people in the AI [31:50] field who who effectively say, "We're [31:52] going to give the world a cure to all [31:54] disease and we're going to make stuff [31:55] really cheap in exchange for [31:58] people giving up their [32:01] autonomy and impact [32:03] over the future and power and also in [32:06] the name of safety and also like just [32:09] absolutely rampant inequality. Like [32:10] there will be people that have access to [32:12] huge amounts of wealth and power and [32:14] other people just get a pretty good [32:15] everything." And this is like a terrible [32:17] sales pitch. This is a very anti-human [32:19] sales pitch [32:20] that somehow [32:23] people feel willing to make. Why do you [32:24] think they feel willing to make that? I [32:26] think it's like fear and power. I I [32:28] think you can [32:29] when people talk about the risks of AI, [32:32] I think there are a lot of [32:35] people who are so nervous about the [32:37] magnitude of those risks and get so [32:39] taken by that and feel the need to [32:41] protect the world from that um that [32:44] they're like, you know, we should trade [32:46] off a lot of liberty for safety here [32:48] because this is unlike other risks we've [32:50] seen. [32:51] But then I think that also ends up [32:53] kind of like um [32:56] a way to justify a lot of power seeking [32:58] behavior. [32:58] >> Everything when I read like when I was [33:00] saying the Claude Shannon and what [33:01] Claude Shannon was saying or Alan [33:02] Turing, at least in the books that I've [33:03] read, it's more of like an optimistic [33:05] like we're going to invent things that [33:07] make our lives better and can do things [33:09] for us. [33:09] >> You are totally right that if you go [33:12] back to the [33:13] Claude Shannon Alan Turing era, [33:15] uh they talked about how wonderful AGI [33:17] would be and all the things that it [33:20] would do. [33:20] Uh and when we started, we really had a [33:23] lot of pressure from the doomers. [33:26] Now, [33:27] the part of the doomers that I agree [33:28] with is this is a powerful technology [33:30] and we should err on the side of side of [33:32] safety and we should [33:35] act with caution at each level of [33:37] technology. The part of the doomers that [33:39] I don't agree with is that it's an [33:40] unsolvable problem. [33:42] If you go back to the beginning of [33:43] OpenAI, um [33:44] I think there would have been two widely [33:46] held [33:47] opinions. Number one, [33:51] not at all and certainly not in 10 years [33:53] were we going to build something that [33:54] was very AGI like. [33:56] And then conditioned on if we did, [33:59] we certainly were not going to be able [34:00] to make it safe. You know, if you had an [34:01] AI that was smarter in many ways than [34:05] a lot of the smartest people, [34:07] most of the smartest people, then, you [34:09] know, the doomers would say surely at [34:10] that point the world would have been [34:11] destroyed. The alignment thing would [34:12] have failed and they were just these [34:14] very confidently held positions about [34:16] what would have happened. [34:18] A decade on, uh we have built something [34:20] that I think most people would say at at [34:21] the time would have said is very AGI [34:23] like. And [34:25] you know, a lot of good things have [34:27] happened and [34:28] the kind of crazy bad predictions of the [34:30] world ending have not happened. So, I [34:32] think that should update people's [34:35] predictions about the future. There are [34:37] still higher stakes challenges in front [34:40] of us to solve, [34:41] but our approach, this is another thing [34:43] I learned from startups, [34:44] of [34:46] the way you do things is to put things [34:49] out into the world, get feedback from [34:51] real customers, see where they break, [34:52] see where they don't break. That is the [34:54] way you make a good product. That is [34:55] also the way you make a safe product. [34:57] And we have made way more progress on AI [35:00] safety than [35:03] I think most people thought we would [35:04] when we started. [35:05] >> Why? Because so many people there's [35:06] billion people using your products on a [35:09] weekly basis? [35:10] >> And each time we get a new level of [35:11] model, we put it out in the world and we [35:13] see what, you know, what works, what [35:15] doesn't work, where people need us to [35:17] relax the guardrails cuz they have good [35:18] things they want to use it for, where we [35:20] have alignment failures, where where we [35:21] have safety systems failures. ChatGPT's [35:23] only been out like less than 4 years. [35:25] Billion people use it. And [35:27] sensitive important stuff. [35:29] And the fact that we can deliver [35:31] something that is broadly considered [35:32] safe, of course there's issues with it. [35:34] Like in that short of a time frame with [35:36] such a powerful technology, I think [35:37] there is no way we could have done that [35:39] in our three tower. And, you know, this [35:42] is how I believe you build [35:44] good [35:46] safe robust useful technology and [35:48] products. And I think it's a great a [35:51] great learning of Y Combinator. [35:53] And it would have seemed to most of the [35:55] AI safety people, you know, totally [35:57] impossible to get to this stage and [35:59] still have the level of [36:01] safety guarantees we have now. [36:04] I do think it gets harder from here, [36:06] but I don't think you're going to solve [36:08] it by disconnecting yourself from [36:09] reality. [36:11] >> Why does it get harder from here? [36:13] Cuz we're as about the smartest people [36:14] in the world are as about as smart as [36:16] the smartest models in the world and [36:17] that's going to flip right now? [36:18] >> Yeah, direction I think that's right. I [36:20] think that the [36:23] the models are just so incredibly [36:24] capable [36:25] and improving on such a steep trajectory [36:28] is that the unknown unknowns [36:32] maybe they don't get harder [36:35] relatively, but from an absolute [36:36] perspective they they seem harder. [36:38] And I think we'll have to make a bunch [36:39] of difficult decisions about [36:42] you know, when we [36:44] delay development, when we sort of [36:47] say, "Okay, you know what? Let's have [36:49] contact with [36:51] reality now or let's let's wait longer [36:53] to really study this more." So, I was on [36:54] a plane recently [36:55] and [36:57] and something that stuck in my mind is [36:59] that the [37:00] you know, the FAA FAA has helped make [37:02] flying incredibly safe. Flying on the [37:05] surface seems like this [37:07] extremely dangerous thing. [37:09] And you probably get on an airplane [37:10] without giving it much thought. And this [37:13] was certainly [37:14] you know, airplanes are not that old in [37:16] the long tradition of human history and [37:17] this was certainly not the case at the [37:18] beginning of airplanes. They have [37:21] extremely robust accident reporting, [37:24] extremely clear-eyed. [37:26] You're never They never try to like, you [37:28] know, hand-wave over something. They [37:30] want to extract as much information as [37:32] possible. [37:33] And in some sense, [37:36] I think with any new technology, [37:38] an approach like that [37:42] works very well and is often under [37:45] appreciated. So, when we started [37:47] deploying our models, [37:49] when we said, "We're going to put [37:50] ChatGPT out in the world." We know the [37:52] model's imperfect. Um we know it [37:54] hallucinates. We know it can do other [37:55] things, but we also know that [37:59] the world's got to experience this [38:01] technology and we've got to learn how to [38:02] make it safe and we've got to put the [38:03] power in people's hands. We cannot just [38:05] use this to impose our worldview. We [38:08] cannot use this to go sit in a lab and [38:10] try to think through [38:11] all the impacts, which won't work anyway [38:13] because society and the models are going [38:15] to co-evolve. Like, we have to all do [38:17] this together as this [38:19] as this joint product. And then [38:23] we'll do very good accident recording. [38:24] We will study when something goes wrong. [38:26] We will put out a very clear postmortem. [38:28] We will learn as much as we can. We will [38:30] not only improve our own technology and [38:31] products, we'll try to share those [38:32] learnings with other people building AI. [38:35] I think that's worked surprisingly well [38:36] so far, and that was like [38:38] good examples from history of [38:38] technology, good examples from startups. [38:41] >> Deel is how the best founders turn the [38:43] world into their talent pool. I've been [38:45] studying how history's greatest founders [38:46] operate for a decade, and one thing they [38:48] all have in common is they understand [38:50] that recruiting and hiring the very best [38:52] talent is your most important priority. [38:54] A players recognize other A players, [38:56] which is why top companies like [music] [38:57] Ramp, Shopify, Eleven Labs, Uber, and [39:00] DoorDash all use Deel. Many of the top [39:02] founders I know have personally invested [39:04] in Deel after using their product. And [39:06] [music] what they discovered is that [39:07] Deel is the best company in the world at [39:09] building infrastructure for global [39:10] hiring. Deel will help your business [39:12] hire, pay, and manage any worker [music] [39:14] anywhere in the world, so you can retain [39:16] the best talent anywhere, and spend the [39:18] rest of your time focusing on what you [39:20] do best, delivering value to your [39:22] customers. [music] The founder of Eleven [39:23] Labs has a great description of the [39:25] value Deel can give your company. He [39:27] said, "We built Eleven Labs to break [39:29] [music] down language and communication [39:30] barriers. With Deel enabling us to hire [39:33] and support [music] exceptional talent [39:35] anywhere, we can accelerate our [39:37] innovation and bring more voices, [39:38] stories, and ideas to [music] every [39:40] corner of the world." Deel is trusted by [39:43] over 40,000 businesses. Learn how they [39:45] can help your business today by going to [39:47] deel.com/centra. [39:49] That is deel.com/centra. [39:53] >> One thing that has to be disorienting [39:54] for you, cuz I it doesn't make any sense [39:56] to me, like, you know, essentially what [39:58] I focus on is just entrepreneurship and [39:59] entrepreneurship, right? So, all the [40:01] podcasts I make are for the benefit of [40:02] entrepreneurs. I'm glad other people [40:04] listen, but it's like heavily focused on [40:05] just trying to serve [40:07] try to find useful information, whether [40:08] it's in a biography of a dead [40:09] entrepreneur or talking to somebody like [40:10] you, that other entrepreneurs can [40:11] benefit from this conversation, right? [40:13] Or any of the podcasts that I make. So, [40:15] in that, there's not like a lot of [40:17] Actually, we love AI. [40:18] But like what I I I I'm trying to figure [40:21] out like if you can help me reconcile [40:22] this like [40:23] Everybody uses AI. Everybody hates AI. [40:26] What the hell's going on there? [40:28] >> Well, people are always afraid of like [40:31] rapid socioeconomic change. Talked about [40:34] the Industrial Revolution earlier. I [40:36] love I love reading about previous [40:37] technological revolutions, too. And like [40:41] people did not have universally warm and [40:43] fuzzy feelings to the change that was [40:45] happening during the throughout the [40:46] Industrial Revolution. [40:48] I think it's probably a good feature of [40:49] human society [40:51] that we have some built-in inertia. We [40:52] have some skepticism of rapid change. I [40:55] think that probably helps [40:58] society in times of turmoil or in times [41:00] of you know, localized craziness or [41:03] whatever. [41:04] Um so [41:05] some of it is probably good. [41:07] And I think a feature of [41:10] human biology, which I believe in never [41:11] trying to fight too hard. [41:13] I also think a lot of people building AI [41:15] uh [41:16] you know, have been off saying there's a [41:18] 25% chance we're going to destroy the [41:19] world and yet we're going to race ahead [41:21] to do it because otherwise those bad [41:22] guys will do it first or [41:24] you know, it's like [41:25] man, this thing is going to be really [41:26] terrible and there's going to be 50% of [41:28] the jobs are going to go away in the [41:29] next year and we hope you all are okay, [41:30] but seems really scary. Like we have not [41:33] as a field done a very good job of [41:35] explaining to people [41:37] what the benefits are and how the [41:38] downsides can be mitigated and we [41:40] certainly have not done a good job even [41:42] if people have had answers like, you [41:44] know, saying well, there's going to be [41:45] universal basic income or work will be [41:46] optional or whatever. [41:48] There's been very little [41:50] discussion from people about [41:53] how and why it's important that people [41:55] have more power and first personal [41:58] freedom in the world, not less. [42:00] And [42:01] that matters a lot to most people. The [42:03] ability of people to influence their own [42:04] future and collectively to design kind [42:06] of where society is going to go and the [42:08] autonomy that comes with that, very [42:10] important. And I don't think a lot of [42:12] people in the AI field [42:15] they feel it for themselves, but they [42:16] don't spend much time thinking about it, [42:18] reflecting on it, acknowledging how [42:19] important that is to other people. And [42:22] so to go back to that like [42:24] characterization of the sales pitch, [42:25] characterization of this [42:26] characterization of the sales pitch, I [42:28] don't know if that's a word, from [42:29] earlier. It is. Um [42:31] the like, "Dear peasants, we will [42:33] bequeath upon you these gifts of a cure [42:35] for cancer and material wealth and some [42:37] things and you know, great [42:38] entertainment. And you stop complaining [42:41] and we'll make all the decisions about [42:43] the future and just trust us, you know, [42:44] we'll be benevolent dictators." Not [42:46] good, not good. As a lover of [42:48] entrepreneurism and sort of like a [42:49] student of what has [42:51] made [42:53] this incredible economic miracle of [42:56] recent centuries work. [43:00] Really empowering people [43:02] to go do new stuff and to push on the [43:05] things they believe in and to have the [43:07] freedom to create companies and invent [43:10] technology and sort of pursue ideas and [43:12] like [43:14] the system around that that makes that [43:15] happen. There is like nothing I believe [43:17] in more strongly. And I think even if [43:19] people don't see themselves as an [43:20] entrepreneur [43:22] ever, if they they may never want to [43:23] start a big company, they do understand [43:25] how important that is. [43:27] And then when you hear people kind of [43:29] implicitly or explicitly saying there's [43:31] going to be less of that with AI cuz a [43:32] smaller number of people are going to [43:34] have the power, but they're going to [43:35] make great decisions and keep everybody [43:36] safe. I think that's very scary to them. [43:39] I also think that [43:41] even if [43:42] maybe most people don't want to start [43:44] really big companies [43:46] a lot of people want to start smaller [43:47] companies and that has been hard. That [43:50] has been something that has required a [43:51] fair amount of privilege and luck and [43:54] resources to be able to do. And we are [43:56] about to see the greatest boom in people [43:59] starting smaller businesses that we have [44:01] ever seen. I think AI is empowering [44:04] that. Now, for some reason the field, [44:07] including us, not talked about that [44:08] enough even though we see all these [44:10] signs of it and it's it's great and we [44:11] have not built enough products to [44:13] accelerate that but I think we're going [44:15] to see a lot more of that. [44:17] >> It's funny we talked about Toby Lütke [44:18] earlier. At the end of the conversation [44:20] and I think that's in the episode he [44:21] mentioned something I didn't think about [44:22] it. He's like oh yeah you and me are in [44:24] the same business. He's like we're both [44:26] trying to create more entrepreneurs. [44:27] He's building infrastructure for [44:28] entrepreneurs I'm building educational [44:29] and inspirational podcast for them. And [44:31] the crazy thing what you just said is [44:33] like not only I think out of any new [44:36] industry it sounds like that as industry [44:38] the AI industry which [44:40] it's doing the worst job I've probably [44:41] ever seen. And I think part of it is [44:43] just the ability you guys have to get [44:45] out there and talk about the stuff that [44:46] you're seeing and educate. There's a [44:47] great book called the Intel Trinity and [44:49] it tells the story of Intel. And it's [44:51] called the Trinity because the three [44:52] main players are Bob Noyce, Andy Grove, [44:54] and Gordon Moore. And when they there's [44:56] a great story in the book that I never [44:57] forgot. [44:58] They go from inventing I think the [44:59] integrated circuits to the [45:00] microprocessor and they realized that [45:01] that technology was so important and it [45:03] would scare the their potential [45:05] customers. So they went out they stopped [45:07] those three people stopped doing what [45:08] they're doing and went out and started [45:09] educating [45:10] >> Yeah. [45:10] >> potential customers, investors, the [45:12] entire country. And they said at one [45:14] time they were putting on more classes [45:16] than like the local community college [45:19] had in their entire course catalog. [45:20] That's how much they made it a top [45:21] priority like we're going to get out and [45:22] educate about this new technology. It's [45:24] just like why is anyone in AI doing [45:26] that? [45:27] >> I mean no excuse we should be doing [45:28] more. I think we've like tried versions [45:30] of this we haven't gotten any quite [45:31] right. We're doing it right now. [45:32] >> We're like this is the point of one of [45:34] the points of reason I wanted to talk to [45:35] you about because like I use AI all the [45:36] time I think it's fascinating but you [45:38] have such like this you have a view that [45:40] makes mine look like like the the view [45:43] of an ant. Like you have there's so much [45:45] stuff in your head that I want to like [45:47] get out and be like hey you have all [45:48] this context. It's like and you're [45:50] inventing this incredible technology I [45:51] would love to know how other people are [45:52] using it. Actually [45:54] before you even get to other people I [45:55] heard you say something that was [45:56] interesting I think ties to what you [45:57] just said about kind of like we're [45:58] inventing technology the technology is [46:00] increasing rapidly but the adoption [46:02] should be slow and deliberate. And I [46:04] heard you on another podcast saying, [46:05] "Hey, I'm even considering like [46:07] how much should I let AI see every [46:10] single thing that's on my computer?" You [46:11] want to talk about that? [46:13] >> With the latest generation of models, I [46:15] don't want to say they they feel smart [46:16] enough cuz I think we should always [46:17] aspire [46:18] for them to get smarter, but they are [46:20] pretty smart. And I feel more limited at [46:22] this point by the amount of useful [46:24] context the AI has on me. [46:26] Like I want the AI to know [46:28] as much as it can to help me. I want it [46:30] to be doing things I can't or don't want [46:32] to do on my own. Like [snorts] I [46:35] I'm not going to read every [46:38] post on our internal Slack. I'm not [46:40] going to go read every story a customer [46:42] has to tell about where ChatGPT works [46:44] for them or failed them. I can't. And [46:46] then there's like other stuff of like I [46:48] just you know, [46:50] I probably could read more research [46:52] papers than I do, but like oh, it's like [46:55] takes a lot of mental energy and you [46:56] know. [46:57] But I would love to have an AI agent [47:00] that is constantly trying to be helpful [47:02] to me [47:03] and that can [47:05] look at and understand more context [47:09] than I can or that I have time for or [47:11] energy for to do on my own [47:13] and [47:14] can help bring that context to bear and [47:18] give me good advice when I have to make [47:19] a decision. [47:20] So, [47:22] I think we focused [47:23] correctly so much on model intelligence [47:26] that on the product side we have not yet [47:28] thought enough about what it means [47:30] to give a model more context than any [47:32] person could have and help advise that [47:34] person on on their big decisions. My [47:36] sense is we are [47:39] just on the precipice of being able to [47:41] see [47:47] a very different way of working with AI [47:49] on a on a [47:51] on an axis where people just can't get [47:53] this good. There are plenty of like very [47:54] smart people, but there is no one that [47:57] can read like [47:59] you know, tens of thousands of pages of [48:00] context in some small number of seconds. [48:03] And really like use that accurately. And [48:05] this is something that AI can do that [48:07] just is [48:08] going to be very new and an incredible [48:10] supplement. You've read all these [48:12] biographies. There are probably times [48:14] where you vaguely remember something [48:16] that if you could remember a specific [48:17] anecdote from one of them, it would like [48:19] really help an entrepreneur in one [48:20] moment for that particular entrepreneur [48:22] right when you're talking to them. [48:23] But maybe you forgot it or maybe you [48:25] don't remember it exactly right. [48:26] >> I built my own AI tools. I use it [48:28] internally. So, you know what it is? [48:30] It's only trained on since 2018 I've [48:32] kept every single note and highlight [48:33] from every single book that I've ever [48:34] used in this database. And I would [48:37] search it for that. And then when your [48:39] the work that you guys do came out, then [48:40] I added So, I have it trained on that [48:42] then every note, every highlight, and [48:44] then all the transcripts from my [48:45] episodes of Founders. I use this thing [48:47] every single day. [48:48] >> That's awesome. [48:48] >> To to make every single episode. So, I'm [48:50] working on Claude Shannon, right? The [48:51] Claude Shannon episode came out I don't [48:52] know, two or three weeks ago whenever it [48:53] was. And I'm asking questions about all [48:55] I was like, "Hey, what did Bob Noyce say [48:56] about this? Or what did Rockefeller do [48:57] about this?" And I I I made the episode. [49:00] I read the book. I took the note. I [49:01] don't remember it cuz it was like 7 [49:02] years ago. It's incredible. This is what [49:04] I mean. I was like, "AI's [49:05] awesome." [49:06] >> That is so cool. Let's get into like how [49:08] you think about running the company, [49:09] right? So, you're spending your time you [49:10] said your your main focus is getting [49:12] more compute and then research, right? [49:15] Okay. [49:16] So, you want the models to be the best [49:18] in the world. But how do you think about [49:19] like do you have to build your own [49:20] products? And you built Codex, right? [49:23] What I don't even know the the [49:25] product lines. Like where's all the [49:26] revenue coming from? Like [49:28] >> Actually, I be more of a platform [49:29] company than a product company. [49:31] >> Okay. [49:31] >> we will build products, of course. [49:33] >> Well, how many products you have [49:34] Let's back up. How many products you [49:35] have now? [49:36] >> We just merged ChatGPT and Codex [49:38] together. So, we used to have like [49:40] ChatGPT, [49:41] Codex, and the API. [49:43] >> Okay. [49:44] >> You know, and Codex sort of [49:47] unfortunately named, but it was not just [49:49] coding. It could kind of do any kind of [49:51] work, which confused people. [49:52] >> I'm confused by that. [49:53] >> Yes. [49:53] >> Okay. [49:54] >> As were many other people. What I think [49:56] most people want is the sort of like [49:58] single interface to [50:01] their own personal or their company's [50:02] AGI that can kind of help them with [50:04] whatever they need. And then the ability [50:06] with an API to build anything they want [50:07] on top of it. And [50:10] that is the platform that we should [50:11] offer to the world. We're going to sell [50:13] great AI at every at every point on the [50:16] cost curve cost performance curve, we [50:18] will be the best. You know, you want a [50:20] really high-end AI to discover science, [50:21] that's great. You want really like [50:23] inexpensive AI to do like, you know, a [50:26] massive amount of volume of work that [50:28] maybe doesn't require genius level [50:29] intelligence. We got you covered there, [50:31] too. [50:33] And [50:34] you know, thinking about this is a sort [50:36] of people talk about different ways in [50:37] the utility and new commodity, whatever [50:38] you want to call it. Like people want to [50:40] use a lot of AI and they want it at a [50:42] low cost and they want it to be fast and [50:44] to work well and have their context and [50:45] be smooth. We we got you. And then [50:47] there's like a single product which is I [50:51] need to ask AI something. Eventually, [50:53] maybe it's the AI should proactively [50:55] offer me things, but you will have this [50:57] interface, which started as a chatbot [50:59] and now also has coding agents and I [51:01] think at some point will feel like a [51:02] more persistent agent [51:04] to [51:05] this [51:06] AI that is running on whatever you need [51:08] it to run on. But that's it. I don't [51:10] think we should go build every product [51:12] category. I don't think we should like [51:13] go try to compete with all our [51:15] customers. I don't think we should try [51:16] to like subsume the entire economy. I [51:18] think we should offer this platform and [51:21] try to have 100 million new businesses [51:23] and 8 billion people use it in all kinds [51:25] of new ways. [51:26] So, [51:27] one kind of direct interface to the [51:28] product, one API for people to use it [51:30] however they want. Those eventually come [51:32] more and more together, too. And then [51:34] it's all about what people do with it, [51:36] build on top of it, whatever else. [51:38] >> What mistakes did you make to have to [51:39] learn that? I feel like um you've had to [51:42] kill some good ideas to and sacrifice [51:44] and going after the great with your full [51:46] like intensity and focus. [51:47] >> Yeah, I I I think killing the good [51:50] ideas, like sacrificing the good ideas [51:52] to go after the great [51:53] ideas is is kind of the hardest lesson [51:56] for any entrepreneur or business to [51:57] learn. It sucks to kill good ideas. And [52:00] no matter how much you think you're [52:01] going to do it, you you people everyone [52:04] like kind of maybe just by the nature of [52:06] who chooses to be an entrepreneur seems [52:08] to do terrible at this. I'm terrible at [52:10] this. I know I'm bad at this. [52:11] But last year for example, we killed [52:13] Sora. [52:15] Which was [52:16] you know, good product and fun and cool, [52:19] but used a lot of compute and not as [52:20] important as Codex where we put the [52:22] compute. Um [52:24] we killed our web browser called Atlas, [52:27] uh which again, I think it was a great [52:29] product. I think it was the best web [52:30] browser, but not as important for us to [52:31] focus on it. Somewhere else we could put [52:33] that talent. Um in a world of limited [52:35] compute, limited people, limited [52:37] resources, [52:38] we thought really hard and we said, you [52:40] know what? The [52:41] the general intelligence for knowledge [52:44] work and eventually for science, most [52:45] important thing we can do. [52:47] Anything that goes into making that [52:49] upstream of generating that [52:50] intelligence, [52:52] building our own chip, building our own [52:54] data centers, um you know, writing good [52:57] infrastructure software, certainly [52:58] training models, obviously. That's all [53:00] really important. But then let's just [53:01] offer this AI as as a service and get [53:05] people to [53:08] use it for intellectual pursuit, for [53:10] work, for scientific discovery, to be [53:13] more productive in their personal life. [53:15] And let's have the kind of [53:17] flexible general platform and not do a [53:19] lot of other things. [53:20] >> Anybody come engage in complicated work [53:22] and you're going to be the, you know, [53:23] top of the sort top of the list of [53:25] anybody alive right now, needs somebody [53:27] to help organize their thoughts. All [53:28] right? It's extremely beneficial. You [53:30] see this in every single biography. You [53:31] see this in histories. Like you need [53:32] somebody to talk to. There's there's [53:34] actually a funny story of how extreme [53:35] this can be. Uh Charlie Munger has a [53:37] saying called the orangutan theory. [53:38] You've ever heard of this? [53:39] >> No. [53:39] >> Where he said uh relatively smart human [53:41] could go in [53:42] uh sit down with the orangutan tell him [53:44] all his problems tell him everything [53:45] that's on his mind and then you know the [53:47] orangutan obviously knows nothing about [53:48] nothing else the human leaves and the [53:50] human is better off just the idea of [53:52] just being forced to put your thoughts [53:53] into you know into into some kind of [53:56] structure. Now obviously with a very [53:58] intelligent partner. Munger played this [54:00] role for Buffett. Buffett is one of the [54:01] most intelligent people ever lived [54:02] greatest investor of all time still [54:03] needed to organize his stuff to somebody [54:04] else. [54:05] You are going through I can't think [54:07] you're you're you have almost like a [54:08] singular lived experience if you are [54:10] somebody as young as you are. So I'm [54:12] curious like who plays this role in your [54:14] life? Like who do you go to that can [54:17] even yeah remotely empathize with what [54:19] the hell you're dealing with on a [54:20] day-to-day basis? [54:22] >> Kind of three categories here. [54:24] One [54:26] a lot of the researchers that have been [54:27] here forever [54:29] we've kind of like all been through it [54:30] together and we've developed this set of [54:31] like shared [54:33] language intuition standards whatever [54:34] you want to call it and [54:37] that I have not been able to replicate [54:39] with anybody outside of the company. [54:41] Um when it comes to like the shape of [54:43] what's happening and what might happen [54:45] next and where the technology is likely [54:47] to go. [54:48] In in terms of questions of just like [54:50] you know business and the world for a [54:53] long time in my career Paul Graham and [54:55] Peter Thiel have been [54:57] two of the people that I have learned [54:58] the most from [55:00] about lots of different faces of my [55:01] career and are still the two people that [55:03] I go to if I really have like [55:06] a very non-obvious problem that I'm [55:08] stuck on and they're [55:11] I have not found anyone else [55:14] after a lot of looking that has the same [55:17] kind of ability to to just to like think [55:19] in a super non-linear way. Like you know [55:22] if what LLMs do are [55:25] predicting [55:26] what word comes next those are two of [55:28] the people that I can predict the least [55:30] what word is going to come next. [55:32] And that is a super valuable skill. You [55:33] go with like a oh man I feel really [55:35] stuck and I've kind of thought through [55:36] all these options and someone that can [55:38] tell you like, eh "Eh, I think none of [55:41] those options are good. Here's this [55:43] thing that now seems totally obvious and [55:45] correct that you didn't think of." Just [55:47] a completely different view that you [55:49] haven't heard anywhere else. [55:50] >> Is this like more of like a prompt for [55:52] your own thinking as a as opposed to [55:53] like explicit advice do X, for example? [55:57] >> Yeah, it's often like here is a specific [55:58] thing. [55:59] >> Really? [55:59] >> Yeah. [56:00] >> So, what would be an example that you [56:01] could share from like Peter? Peter's [56:03] very fascinating to me. [56:04] >> He is very fascinating. [56:05] >> And I it's it's actually who I thought [56:07] of. The reason I thought of this [56:08] question just now is cuz you're like we [56:09] had to kill these good ideas for the [56:11] great. We're cutting Atlas. We're [56:12] compute constrained. We have to focus, [56:14] focus, focus. It's like that's something [56:16] that is very obvious when you listen to [56:18] him talk about the importance of focus [56:20] and like if you have something that's [56:21] working, making it work better and going [56:22] down this line, like taking an hour away [56:25] from that to like explore something else [56:26] is too expensive. You should just go [56:29] deeper on what's already working. Like [56:31] there's a lot of value at the extremes. [56:34] >> After we launched [56:35] ChatGPT, [56:37] it was this weird thing cuz people [56:38] didn't really know what to use it for. [56:40] And [56:42] it was growing super fast, but it felt [56:45] like [56:46] very unstable or kind of almost like a [56:48] low-value growth. Like people were using [56:50] it to [56:52] just cuz they were like interested in [56:53] talking to it and what they could do. [56:55] And so, there were a lot of people in [56:56] the company who were like, "Uh, this is [56:58] you know, we got to figure out something [56:59] else. This is not sustainable value." [57:00] And we were talking to him about this [57:02] like list of [57:03] five or six other things that we could [57:05] focus on. [57:06] Instead, this is like maybe two months [57:08] after ChatGPT launched, something like [57:09] that. [57:11] And [57:12] he was like, [57:15] "It's an obvious mistake to do anything [57:16] but this. Besides the fact that it's [57:18] growing, which is rare and great." And [57:19] it was not growing this fast then and it [57:21] did start after. He's like, "The power [57:23] of this is the power of the Google text [57:25] box. It's like a text box you can type [57:26] anything into [57:28] and it does the right thing. And the [57:30] fact that it doesn't match the current [57:32] Silicon Valley wisdom [57:33] of, you know, you got to have like feeds [57:36] and you have to have like a network [57:38] effect and you have to have cuz we had [57:39] none of these things and that's why [57:40] everyone was worried. You know, you have [57:42] to have like a [57:45] you know, way that [57:47] people are going to build up more con- [57:49] this before we had memory. People are [57:50] going to build more contacts. People are [57:51] going to get locked in or people are [57:52] going to have all of these like all that [57:54] stuff. Like people have just been [57:56] chasing the Google business model for 20 [57:57] years and this is the first thing that's [57:58] come up and, you know, clearly the empty [58:00] text box worked for Google so why don't [58:02] you just double down on that? It's [58:04] growing like it's very flexible and, you [58:06] know, it has all of the signs other than [58:08] it doesn't fit the current Silicon [58:09] Valley wisdom. And I was like, "Okay." [58:13] And so we went super hard on ChatGPT uh [58:15] and it was great. [58:17] It's like a simple genius to what he [58:18] just said. [58:19] Sometimes there's like more complexity [58:21] to this but that was an example of very [58:22] important simple genius. [58:23] >> What about some advice that Paul Graham [58:26] or some guidance or a direction he kind [58:28] of pushed you in? [58:30] >> When you just said that, the thing this [58:32] is like a meme for many YC founders uh [58:35] where you would go see him for office [58:36] hours [58:37] and he would say, "You know what you [58:39] should do." And he would like shake his [58:40] hand his finger like this. "You know [58:41] what you should do. You know what you [58:42] should do." [58:43] And sometimes the thing that came after [58:45] that was great. [58:47] Sometimes the thing that came after that [58:48] was terrible. [58:49] But the important thing [58:51] was [58:54] there was a kind of there is a [58:57] creativity and open [59:00] landscape and just uh you know, let's [59:01] try a lot of things. We talked about the [59:03] spirit of iterative deployment. And we [59:06] talked about [59:09] you know, what what how like in the same [59:11] way startups he really I think pushed [59:14] the startup ecosystem into this world of [59:16] you got to like get ship a a a V1 you're [59:19] embarrassed of early and it doesn't [59:21] matter if it's, you know, it could be [59:23] much better. You'll you'll you'll get it [59:24] much much better because of the feedback [59:26] to customers. [59:27] I don't even think I asked him before we [59:29] launched ChatGPT like, "Hey, do you [59:31] think we should launch this thing?" But [59:32] I knew what he would say. I knew it was [59:34] still early. I knew it was still [59:35] embarrassing, and I knew the right thing [59:36] was to like get it out and get it in [59:38] front of people. [59:39] >> So, wait. Wait. Your mental model of [59:40] Paul Graham is so complete, you don't [59:41] even have to call him up. [59:42] >> one case. That's That's the one where [59:44] you would say like there's certainty. [59:46] >> Let me tell you something funny. Right [59:47] before he died, a few months before he [59:48] died, I went to Charlie Munger's house [59:49] and had uh dinner with him, and I was [59:51] like, "How often do you talk to [59:52] Buffett?" He goes, "Never." I go, [59:54] "What?" [59:54] >> He goes, "We talked every day for hours [59:56] and hours. Buffett can just pretend to [59:57] pick up the phone call to call me, and [59:58] he already knows what I'm going to say." [60:00] That's obviously comes after 65 years of [60:02] working closely together, but I thought [60:03] it was hilarious. [60:04] >> That is hilarious. That is really a [60:06] funny story. [60:07] >> No, there's many times that I couldn't [60:09] predict what he's going to say, which is [60:10] That's why I think it's valuable. But in [60:12] terms of the like [60:14] launch when you're embarrassed of a [60:16] product, I know what he's going to say [60:17] there. Like that one uh [60:20] That That That has been like [60:22] I won't say the most valuable piece of [60:25] YC of tactical YC advice, but it's been [60:27] up there. I'm astonished looking back at [60:32] all of my data points of YC founders [60:34] over the years how much the ability the [60:37] the like [60:38] moving fast [60:39] and being iterative correlates with [60:40] success. [60:42] >> Okay, you've mentioned YC way many too [60:44] many times in this conversation. I I [60:46] have to explore this because we talked [60:47] before. It's like, "Listen, I'm not a [60:48] journalist. I'm an enthusiast. I don't [60:50] have a list of questions." And I'm like, [60:52] "I have a world-class founder across [60:53] from me. I want to know what the hell is [60:54] in this person's mind, and I want to [60:56] like extract information out selfishly [60:57] for me." So, like why like I just I'm [61:00] shocked at how much you reference it in [61:02] conversations. Like how impactful [61:05] going through YC and then running Y [61:07] through YC, being, you know, affiliated [61:09] with them has is is it clearly on your [61:11] life. Can you like you expound on this? [61:14] >> There is some band that wasn't that [61:17] successful. [61:18] They didn't sell that many albums, I [61:19] mean, but they like influenced all of [61:21] the musicians that came out of college. [61:23] >> The Band. Literally that. [61:25] >> Rick Rubin told this story. I think it's [61:27] Is it might be The Velvet Underground. [61:28] But, but you know the idea I'm getting [61:30] at, whether it's called The Band or [61:31] whatever. Yeah. [61:32] It's not fair to talk about YC in this [61:35] way cuz YC, measured [61:37] by like traditional metrics and market [61:39] cap created whatever, is one of the hand [61:42] would be one of the handful of most [61:44] valuable tech companies. [61:46] But, the degree to which YC [61:50] totally influenced everything that has [61:53] happened in the last 20 years of [61:56] the tech industry and startups, [61:59] entrepreneurship, whatever, [62:00] I think it's is only like sort of [62:02] understood. [62:04] OpenAI is an example of that. Not just [62:06] from like how we have shipped our [62:07] products in the world, but like the [62:09] philosophy of how we run our research [62:10] lab. I think if you go talk to like many [62:13] of the other [62:15] people running like this generation of [62:18] large tech companies, they would tell [62:20] you so much stories, even if they didn't [62:21] go through YC. [62:22] >> Well, what's happening there? Is it an [62:24] operating system that YC is giving you [62:26] because you hear, you know, do these [62:28] five things or whatever, or is it more [62:29] like a philosophy of building companies? [62:31] This is the confusing part for me as an [62:33] outsider. [62:33] >> I think it's two major things. I mean, [62:36] there there is there is some of the [62:38] operating system of what to do. But, I [62:40] think it was the philosophy of how to [62:41] run companies. [62:42] The idea of [62:44] iterative deployment and technical [62:46] people in charge and sort of being [62:49] willing to bet on [62:51] young people with a lot of energy and [62:52] ambition, but maybe less experience [62:54] throughout all levels of a company. [62:56] And then it was also the change the [62:59] related change to the whole ecosystem [63:01] that happened. In the pre-YC tech [63:03] ecosystem, so if you, you know, ran the [63:06] clock back to 2004 and then [63:09] projected technology forward to 2016, [63:11] um but not anything else about the shape [63:14] of [63:17] startup ecosystem, what it meant be an [63:19] entrepreneur, how capital flowed, all of [63:20] the you know, who got to run companies, [63:21] all of those things. I do not think [63:23] OpenAI would have been impossible. I [63:25] think the changes that YC [63:27] induced in the whole ecosystem, you [63:29] know, more leverage going to founders, [63:32] young technical founders having the [63:34] ability to raise lots of capital, um the [63:36] ability to sort of like work on [63:38] ambitious things without a very proven [63:39] resume. I don't think OpenAI would have [63:41] been possible. [63:43] So, this is kind of like a big change. [63:45] >> Is this all tied to the fact that you [63:47] think there was a benefit in you going [63:49] from founder to investor for a long [63:51] period of time back to founder? [63:53] >> There are all those benefits, too. And I [63:55] wouldn't say I really went from founder [63:56] to investor to founder cuz like the [63:58] first time I was founder didn't really [64:00] work out that well. So, like [64:01] >> You still started a company. You you [64:02] learned some lessons from failure, but I [64:04] think you learn way more from success. [64:05] >> No, you got to hold on. You We're not [64:07] moving on from that. You got to say more [64:08] about that. [64:09] >> There's some [64:10] I can't believe I think it's Anna [64:11] Karenina. Some great Russian novel. I'm [64:13] very embarrassed not to know this. [64:14] Starts with like all unhappy families [64:16] are unhappy in their own way. All happy [64:17] families are alike. [64:19] >> Yeah. [64:19] >> That explains why I jumped into YC. [64:21] Um but I think this is really true. Like [64:22] when I look at the lessons of where I [64:24] have failed at something, I learned [64:26] something generic about grit and [64:28] determination and something not to do. [64:31] But [64:33] there's most things don't work. So, [64:34] there's like a lot of reasons why things [64:35] don't work and there's I think it's like [64:37] harder to [64:39] put together [64:42] the correct causation. [64:44] And [64:46] when I've had something really work, [64:49] when I understand like what parts of Y [64:51] Combinator really worked or what parts [64:52] of OpenAI really worked, [64:54] trying to apply those lessons going [64:55] forward [64:57] has been [64:58] much more helpful to me than trying to [65:00] apply the anti-lessons of what didn't [65:02] work. [65:03] And so, you know, you you should of [65:05] course learn as much as you can from [65:07] every data point. So, learn from the [65:09] failures, learn from the successes. [65:11] But [65:13] in my own experience, [65:14] uh [65:15] I have [65:17] when I have tried to apply those [65:18] lessons, [65:20] the lessons I learned from success were [65:24] very good, and I should have applied [65:26] those more. [65:27] And the lessons I learned from failure [65:28] were either fairly generic, and I kind [65:30] of already knew them, or [65:32] got in the way of something else. And [65:34] And I think this is like [65:36] generally [65:37] true for a lot of people. [65:39] >> Yeah, but isn't it like we already kind [65:40] of know what we should do or what we [65:42] should avoid? But it's like the [65:44] reminder, the constant reminder. So, [65:46] like the the best description of my [65:47] other podcast Founders I've ever heard [65:48] is like it's church for entrepreneurs. [65:50] If you really think about it, I used to [65:51] drop it on Sundays, and I should go back [65:52] to doing that. But it's like really I'm [65:54] just telling the same It's the same [65:55] personality type as if you're throughout [65:56] history. [65:57] >> Yeah. [65:57] >> It's just like now this person happens [65:59] to be building ships, and this person [66:00] built technology, but like and they live [66:01] in different [66:02] >> same It's the same personality. [66:03] >> That's for sure. I I'm kind of obsessed [66:04] with this idea of things that last for a [66:06] long period of time. And like uh you [66:09] know, companies The best companies can [66:10] last a long time, but not as long as [66:12] cities. And uh cities don't last Cities [66:14] and countries don't last as long as like [66:15] religions. You know, like so out of all [66:17] the man-made things, what has lasted [66:18] longer? I would say I can't think of [66:20] anything other than religion might be an [66:21] asset. So, then I started studying. I [66:23] grew up My mom was fundamentalist [66:24] Christian. So, I was forced to go to [66:26] church my entire life. [66:27] And I just started analyzing like what [66:28] do all the main religions in the world [66:30] have in common? It's like oh, we have a [66:31] shared base of knowledge. Usually some [66:33] kind of book, right? We meet with [66:35] like-minded fellow believers at regular [66:38] intervals. [66:39] And it's not like I go to church on [66:40] Sunday. It's like okay, we talked about [66:41] Jesus last week, but let's talk about [66:42] this other guy. It's like no, I go back [66:44] to the same books and same stories over [66:45] and over again. So, I read your blog, [66:48] and you even said this uh something [66:49] about like when YC end ended. It's like [66:52] you're repeating the same thing. You're [66:54] telling it to them all the time, and [66:55] then they leave the church, you know, to [66:57] to for this analogy, and then they stop [67:00] doing the same thing. It's like [67:02] >> 100%. [67:02] >> even the lessons. It's like the constant [67:04] reminder that this This important. [67:05] >> I extremely strongly agree with that. [67:09] But I think it is [67:12] better to be reminded of the thing like [67:14] talk to your users more. [67:16] You know, ship products earlier, get [67:18] more feedback, um hold a higher bar for [67:21] who you recruit and who you hire and [67:24] more quickly suddenly. But it's the [67:26] positives that I think are good. [67:28] >> You have one of the greatest tweets. I [67:29] have I say to my phone, you're like, you [67:31] know, skip the conferences, the [67:32] dinners, everything else. Just like [67:34] essentially make the product or sell and [67:35] sell the product. If you're not making [67:37] it, you're not selling it, like that's [67:38] all you actually have to do. And I think [67:39] it's like there again it goes back to [67:40] like that simple genius. So then this is [67:42] the other part that I find most [67:43] fascinating cuz somebody asked me [67:45] yesterday, they're like, "What's your [67:47] Usually there's some kind of historical [67:48] equivalent for every founder I meet that [67:50] I can like, 'Oh, that guy's kind of like [67:51] Verner Cole. That guy's kind of like [67:52] Rockefeller, Ludwig, or any of these [67:53] people.'" And I was like, "What's your [67:55] historical equivalent for Sam?" I was [67:56] like, "There isn't I can't think of one [67:58] cuz like I don't know him enough well I [67:59] don't understand how he thinks yet." [68:01] What do you think now? Well, we This is [68:02] the first of hopefully eight [68:03] conversations so we'll I'll tell you in [68:05] conversation summary. [68:06] But this is very rare. I just talked to [68:08] Doug Leone and he talked about one dude [68:11] that he hired. It's the founder of [68:12] Nubank and he was a like an [68:14] associate in VC and then leaves and [68:16] founds one of the most successful [68:17] companies. I'm like, [68:18] "I've never heard of that. Doug, have [68:20] you?" And he dedicated his life to this. [68:21] He goes, "No, that's the only one." So [68:23] again, it's very rare. Mostly people go [68:25] from founder, sell their business [68:28] unfortunately, and then investor as [68:30] opposed to run the business till you [68:32] die, which is my preferred method of [68:34] things. [68:34] >> What I'm curious about this is like when [68:36] you just said I learned more from [68:37] successes, right? [68:38] >> Well, it's the successes cuz you were [68:40] exposed to what? 10,000 different [68:42] companies in that decade or decade and a [68:44] half that you were doing this. And you [68:45] saw obviously maybe the half a dozen or [68:48] the dozen or were the best in the world. [68:50] So like are you taking their successes [68:51] as well as like instructive to No, no, [68:54] no, okay. [68:54] >> To- to- totally. Uh [68:56] the [68:57] And I think people do I mean, you're [68:59] incredible student but there's a lot of [69:01] pretty good students of [69:01] entrepreneurialism [69:03] and entrepreneurism and people [69:06] I think often try to go look for those [69:07] lessons of the things that that really [69:09] worked. And as you said, it's kind of [69:11] the same thing over and over again, like [69:13] done in different industries. [69:15] But you have to be reminded of it a lot [69:16] and it's unglamorous. [69:18] >> I had no understanding going into this [69:19] conversation. I think I was slightly [69:21] better understanding going into this. I [69:22] told you before we started this is just [69:24] for my own edification. But like even [69:26] the people that influenced you or it's [69:27] just like Peter Thiel saying, "No, [69:30] dummy." He's obviously going to say he's [69:31] like, "No, dummy, this is working. Why [69:32] are you doing anything else but the [69:34] thing that is working?" So there's got [69:36] to be examples where you're like, "Hey, [69:37] I've given this advice to other founders [69:39] a million times." And then you catch [69:41] yourself, "Oh I'm not even [69:43] applying my own advice at this point in [69:45] time." [69:45] >> Totally. Yeah, I'll give many examples [69:47] of that. I I I I think it's also [69:49] instructive to like [69:51] what what was the new thing? What didn't [69:52] you have the [69:54] advice for? [69:54] >> Mhm. [69:55] >> And the thing that was really different [69:57] about OpenAI than anything that I had [69:59] pattern matching with [70:00] is [70:02] it was four and a half years from when [70:03] we started the company till when we [70:04] launched our first product. [70:06] >> The opposite of YC advice, right? [70:08] >> Yes. [70:08] >> Okay. [70:09] >> Yes. [70:10] And although there were all these ways [70:12] which [70:14] managing a research team was similar to [70:17] selecting and advising founders [70:21] learning what it [70:23] to whatever degree we learned it cuz I [70:24] don't think we did it perfectly. Like [70:26] how you [70:28] manage through [70:30] this part of the world where you don't [70:31] have the external signal from customers [70:34] and you're just trying to like [70:37] you know, do what would normally be the [70:39] catastrophic startup advice of not [70:41] uh [70:43] not shipping a product for four and a [70:44] half years. That was very difficult and [70:46] we tried all of these things about how [70:48] we like [70:49] had a [70:51] how we replaced the signal of do [70:52] customers actually like the product for [70:55] is our research actually working. [70:57] One of the things that worked actually [70:58] is during the [71:00] um [71:01] Dota 2 days when we were trying to use [71:03] RL to beat this video beat people at a [71:05] video game, we put up like a leaderboard [71:07] and people could just see how different [71:09] ideas were performing and what was, you [71:11] know, if that was like objective and [71:12] real and people wanted to like go off [71:13] that. But we had to try all of these [71:15] things to [71:17] uh basically like simulate [71:20] end users. [71:22] And that was a totally interesting new [71:24] problem I had no pattern matching for. [71:26] >> How did you work your way through that? [71:27] What were you thinking? Like how did you [71:28] do this? [71:29] >> We asked a bunch of people who had been [71:30] at [71:32] great research labs of the past. And if [71:34] it had been, you know, OpenAI started at [71:36] sort of a time when everybody in Silicon [71:38] Valley as their vanity project, [71:40] including me, uh wanted to start a [71:42] research lab. [71:44] And there were all these books about the [71:45] heyday of Bell Labs or Xerox PARC that [71:47] were very popular. Everybody was talking [71:48] about this. It was a huge amount of [71:49] discussion. [71:50] In fact, I even see one of the books [71:52] over there about uh about Bell Labs. But [71:55] there was not a ton of people that had [71:56] like in living memory how to actually do [71:58] it. So we talked a lot to Alan Kay. We [72:01] talked to a handful of other people. [72:03] And we got some advice from them about, [72:04] you know, what made a really good [72:05] research research lab and some of it was [72:07] really good. Some of it didn't [72:09] translate as well to the current moment. [72:11] >> Well, you also didn't have this giant [72:13] monopolistic profit [72:15] >> [laughter] [72:16] >> profit printing machine like Bell Labs [72:18] was spun out independently. [72:20] Uh Polaroid did a lot more research when [72:22] they had uh essentially like a monopoly [72:25] >> on its photography. I just read the [72:26] biography of the founder of Honda, [72:28] right? Ho- The guy created the most [72:29] successful uh [72:31] uh motor vehicle of all time. The Honda [72:33] Cub has sold uninterrupted for like 60 [72:34] years, millions of yeah, [72:36] vehicles. And his whole thing and he he [72:38] arrived at the same conclusion Bell Labs [72:39] did, that he thought the research and [72:41] development had to actually be separate. [72:43] It was a it was spun out of the company [72:44] and had separate ownership, just like [72:45] Bell Labs did. [72:47] >> We did not have that. [72:48] >> No, you did not have the [72:50] >> When I think back to those early days, I [72:51] mostly feel like I was trying and [72:53] failing to raise money. That's like my [72:54] dominant memory of the early days of [72:56] Open AI. [72:58] So much effort, so frustrating. I wish [73:01] we had some sort of cash machine like [73:03] that. I remember like one of my clearest [73:05] memories of all of Open AI, so [73:08] announced coming in 2015, but the first [73:09] day was right after New Year's in 2016. [73:12] And 12 of us or 11 of us showed up at [73:14] Greg Brockman's apartment, you know, [73:15] like 9:00, 9:30 on a Monday or Tuesday [73:18] morning, something like that. [73:19] Let's say it's January 4th. And [73:21] everybody's there and it'd been this [73:22] like big effort and everybody walks in [73:24] with a lot of excitement. It feels like [73:25] the first day of school, whatever. [73:28] And then very quickly people like look [73:30] around the room and they're sort of [73:31] like, "Well, what do we do now?" [73:35] Someone says, "Okay, we should get a [73:36] whiteboard." [73:37] Greg, you know, gets someone to go off [73:40] and find a whiteboard. Whiteboard comes. [73:42] Um look around again, you know, what are [73:44] we supposed to do now? And you just feel [73:46] the energy in the room collapse. [73:48] And none of us know what to do. Like [73:51] there's no It was not like building a [73:54] product startup. It was not like let's [73:55] build this product. Let's talk to [73:56] customers. It's like, "Okay, we said we [73:58] want to make AGI." [73:59] Maybe we should write some papers. [74:01] Okay, let's write some papers. Maybe we [74:03] should think about some ideas. Okay, [74:04] let's think about some ideas. [74:06] You know, everybody's got their like [74:07] moments of I have no idea what I'm [74:08] doing. That was one of mine. Like I [74:10] have, you know, we have just launched [74:12] this thing. None of us had any idea what [74:14] we're going to do. [74:15] So we did what we knew how to do and [74:16] eventually we figured out a lot of [74:18] things didn't work. Eventually we [74:20] figured out a [74:23] kind of like rhythm for [74:26] making and then evaluating research [74:27] bets. [74:31] And [74:33] all far from perfect, obviously, uh but [74:36] we did find a gradient that we could [74:38] kind of progress along. And we figured [74:41] out how to get, you know, the resources [74:43] that very smart people needed and how to [74:44] make sure that we were like [74:46] not [74:47] completely getting lost in the [74:48] wilderness. [74:50] And [74:51] over [74:54] some number of years [74:57] mostly chaotic stumbling [74:59] we eventually made most of the big [75:01] discoveries. [75:02] You know, what started as the [75:04] unsupervised sentiment neuron paper [75:07] uh [75:08] turned into GPT-1 and then [75:10] eventually GPT whatever. The scaling [75:13] laws work that gave us the confidence [75:16] not only by the compute but the [75:17] understanding about how to scale up our [75:19] models sort of came together. And [75:21] through and many other things, too. Um [75:24] through this process we learned [75:27] things like that idea of leaderboards [75:31] that worked. We also learned the [75:33] incredible power of external demos [75:37] uh for like an eminent person that the [75:39] researchers really wanted to impress. [75:41] And then we learned a bunch of things [75:42] that like didn't work like fake [75:43] deadlines. [75:44] >> It must be so disorienting to live [75:47] through that experience. You're at 12 [75:48] people in an apartment, don't even have [75:50] a whiteboard, don't know what to do. [75:52] Fast forward a decade, you have a [75:53] billion people using [75:55] >> experience. [75:55] >> Do you keep a journal? [75:57] >> When my kid was born, my first kid, I [76:00] would like, you know, [76:02] get home at the end of the day and [76:04] be rocking him to sleep and just like [76:06] talk to a kid or whatever. So I was just [76:07] like [76:08] need to come up with a talk about. So I [76:09] would like just tell him about [76:11] my day and what we were struggling with [76:12] and kind of like what I was worried [76:13] about and what was happening. It was [76:15] kind of like fun for me to do and I was [76:16] like, this is sort of interesting and [76:18] someday I'll be like interesting for him [76:19] to have this. So I started writing him [76:22] like every Sunday I would like write him [76:23] a letter. I would like talk just to talk [76:25] and then I would like write it down. I [76:26] only ever did like eight of them or [76:28] something. [76:28] >> How many kids do you have? [76:29] >> Two. [76:30] >> Okay. [76:31] Bezos has this great line about building [76:32] Amazon. He's like, we're trying to do [76:34] stuff [76:35] that we can tell our grandkids about [76:38] that we're proud of, right? And those [76:39] things are hard. The fact that you were [76:41] writing to your son [76:43] >> Oh, man. [76:43] >> Keep writing the letters. And if you [76:45] don't do that, this is real quick. Just [76:47] because I've read enough books about [76:48] this. [76:49] Most time, guess what? Founders don't [76:51] write autobiographies when they're 40. [76:52] They write them when they're 70 and [76:53] they're looking back and they wish they [76:54] could do it again and so much has been [76:56] lost to the sands of time. They all [76:58] repeat that. They're like, "I wish I [76:59] journaled." So, even if you don't do it, [77:01] you have enough resources, what I would [77:02] do, write have a book written, even if [77:05] it's for internal purposes only. You [77:07] ever read The Little Kingdom by Michael [77:08] Moritz? [77:09] >> I never read it. I have [77:10] >> Oh, you have to. It's like the first [77:11] 6-year history of Apple written by [77:13] Michael Moritz. [77:13] >> Isn't it crazy that he wrote that book? [77:15] >> He's a phenomenal writer. Like, crazy [77:17] writer. Winds up being one of the best [77:18] venture capitalists of all time, I [77:19] guess. But, like the you just The point [77:21] is that have That book ends Steve hasn't [77:23] even been kicked out of Apple yet. So, [77:24] you get like what actually happened. [77:26] You're going to want this. You might not [77:27] want it now, but you're damn sure going [77:28] to want it when you're 60 or 70. [77:31] >> The thing that was so interesting was [77:32] like the mindset of writing to your kid. [77:33] Like, you really can't hide behind [77:36] anything. Like I like I don't really [77:37] care what my kid's going to think about [77:38] me. So, like this thing happened. Like, [77:41] you know, didn't feel great about it. [77:42] Better do it differently next week. [77:44] Like, it was a very interesting [77:45] extremely interesting mental framework. [77:46] Maybe I'll find some way to do it again. [77:48] No, maybe you're going to do it. Okay. [77:50] Sam, thanks for taking the time. This is [77:51] awesome, man. Thank you. Appreciate it. [77:53] I hope you enjoyed this episode. Please [77:54] remember to subscribe wherever you're [77:56] listening and leave a review. And make [77:58] sure you listen to my other podcast [77:59] Founders. For almost a decade, I've [78:01] obsessively read over 400 biographies of [78:04] history's greatest entrepreneurs [78:05] searching for ideas that you can use in [78:07] your work. Most of the guests you hear [78:09] on this show first found me through [78:11] Founders.