a16z Partner Anish Acharya Urges Founders to Build AI Product Loops

A16z的Acharya：AI循环正在重构企业，VC不再青睐小额押注 自动化智能体循环将执行标准的企业工作流直至触及局部性能上限，而人类直觉与创始人不受限的雄心将成为初创企业的决定性护城河。 Anish Acharya认为，AI是一种能动性放大器，它将技能与意愿解耦，而非将劳动者推向被自动化边缘化的底层阶层，但经济扩散与物理制约将决定其在现实中的落地节奏。他解释称，现代企业正在重组为自动化的多智能体执行循环，而人类的直觉与品味对于打破局部最优、指引全新战略方向而言仍然必不可少。在消费科技领域，Acharya指出最大的市场空白不在于生产力工具，而在于解决人类的情感与社交需求，概括为“/loop，让我更快乐”。他敦促创始人树立宏大抱负，采取高端定价模式，并作为“模型侍酒师”在前沿与开源权重架构中培养深厚的实操直觉。 企业对AI的采用受制于物理与流程瓶颈而非单纯的智能水平，这更接近电力长达40年的普及扩散曲线，而非瞬间爆发的递归式飞跃。 (2:36) 企业运营正演变为贯穿工程、销售和战略的层叠式多智能体循环，但自动化循环往往会停滞在局部极值，需要依靠人类直觉去寻找下一座高峰。 (11:25) 消费级AI的核心前沿是情感连接与娱乐而非生产力，其背后的逻辑在于，渴望快乐度过时间的人群远多于想要节省时间的人群。 (32:03) 过去三年中风险投资的评估标准发生了颠覆：投资者如今普遍拒绝温和、渐进式的切入策略，转而青睐能够运用庞大资产负债表进行无上限宏大探索的创始人。 (64:30)

Andreessen Horowitz partner Anish Acharya rejects Silicon Valley fears of an AI-induced "permanent underclass," arguing instead that AI unbundles skill from desire and amplifies human ambition. He contends that while company operations are evolving into cascading autonomous loops, human intuition remains indispensable for moving beyond local optimization plateaus and designing products that fulfill fundamental human needs. Why the "Permanent Underclass" Myth Fails (2:29) Acharya dismisses anxiety about falling irrevocably behind in the AI era as a collective "dark fantasy" that ignores market realities. Unlike the mobile era, which centralized power into winner-take-all networks, today's AI stack features roughly 20 relevant players at every layer, including multiple competing coding agents like Claude Code, Codex, Lovable, and Replit. Furthermore, real-world adoption is buffered by slow economic diffusion and the fact that many industries—such as physical logistics at FedEx or Domino's Pizza—are not bottlenecked by raw intelligence. Inside knowledge-work companies, executive incentives also favor growth over mass replacement. Acharya argues that a CEO like Google's Sundar Pichai wants to build a $40 trillion company rather than a slightly more efficient $4 trillion one. Anecdotally, a Google executive noted that teams are using AI to condense two-year roadmaps into three months rather than pursuing layoffs. Outside of tech giants, operational businesses are also upskilling workers; Mexican used-car marketplace Kavak created a six-week "Jedi Academy" where employees, including vehicle mechanics, learn to build production-ready AI agents. Cascading Business Loops and Human Intuition (11:26) Company architecture is progressing from single prompts to tool-equipped agents, and finally to autonomous business loops. In software engineering, a loop can already reproduce a bug, generate a fix, run code review, and deploy low-risk updates to production within five minutes. Similarly, growth loops can automatically generate test variants, monitor telemetry for statistical significance, and merge winners. At Kavak, per-customer agents handle interactions autonomously and call a human operator for coaching only when stuck. Despite this automation, Acharya emphasizes that AI loops are merely optimization engines that climb to a local maximum before plateauing. Because models cannot perform out-of-distribution thinking to invent new paradigms, human intuition is required to identify non-obvious opportunities and place the loop at the base of the next hill. In product management, this shifts the job from saying "no" and debating spreadsheet priorities to saying "yes" to every idea, simulating user reactions with tools like Simile, and letting ideas compete on merit. Model Economics and the "Sommelier" Mindset (21:41) Organizations must also choose between expensive frontier models and cost-effective open-weight models. Acharya uses Pareto efficiency to distinguish between bounded-upside tasks—like bookkeeping or legal compliance, where open-weight models offer rational cost control—and unbounded-upside domains like drug discovery, strategic engineering, or complex sales, where paying a massive premium for one extra IQ point in a frontier model is justified. Even routine customer support can hide unbounded upside if a bug report contains a breadcrumb that could reshape the organization. Because models are not fungible commodities, Acharya advises builders to become "model sommeliers" by testing every release on an ongoing, low-stakes personal project—a "chassis"—such as a personal DJ app. He notes that models have distinct personalities: Qwen3.8 acts as an open, creative storyteller capable of working autonomously for four to five hours to direct five-minute *Star Wars* documentaries using MiniMax and ElevenLabs, whereas GLM-5.2 and GLM-5.3 behave like precise, "neurotic PhDs" suited for detailed product work. Consumer AI's "/loop, make me happier" Opportunity (31:20) While tech discourse fixates on enterprise productivity, Acharya argues that most consumers prefer to spend time rather than save it. He frames the biggest consumer AI opportunity as "/loop, make me happier"—using background loops to help people feel connected, improve their health, or become better friends. He compares the current consumer landscape to the iPhone in 2010: falling inference costs from open-weight models and new relational interfaces are clearing early bottlenecks, giving startups an edge over risk-averse incumbents in emotionally nuanced categories like companionship. Across the consumer landscape, Acharya highlights three emerging pillars: coding agents (like Claude Code, Codex, and Wabi) used by everyday people for tasks like video editing; personal agents (such as Grok, ChatGPT Work with full-duplex voice and cloud credential caching, and Instinct); and creative or companionship apps like Suno. On the macroeconomic level, he believes AI will win public sentiment if it brings deflationary abundance to education and healthcare, where administrative overhead consumes roughly 45% of expenditures. Uncapped Ambition, Moats, and "Birkin Bag" Pricing (54:30) For founders, the rules of company building, defensibility, and pricing have inverted. Citing Decagon's Jesse, Acharya notes that moats are discovered through shipping rather than designed upfront; Cursor, once dismissed as a thin wrapper, built durability through its UX harness and proprietary reasoning traces before training custom Composer models. Because incumbents now guard their networks against platform growth hacks, distribution relies on organic word-of-mouth driven by remarkable product quality. Consequently, investors now penalize small, safe ideas in favor of massive scope. Acharya also challenges the dogma that consumer software must be free, noting that early adopters already pay $200 to $300 per month for multiple AI subscriptions. He urges founders to imagine a "software Birkin bag"—asking what a product would need to deliver to justify $1,000 or $10,000 a month—and to embrace uncapped ambition because human desires always expand faster than technology's ability to fulfill them. To adapt to the AI era, Acharya advises product builders to view building as an activity rather than an outcome—"the new reading." By adopting a one-week rule to ship small, personal prototypes (such as an automated family keepsake built with Codex), anyone can demystify the tools, overcome career anxiety, and develop practical intuition for where the technology is heading.

[5:26] Anish Acharya: The line of reasoning for that case is always: everything up until now, then something happens that no one can quite articulate, and then fast takeoff. [5:26] Anish Acharya: If you had a data center of PhDs working at FedEx or Domino's Pizza, are they going to be exponentially dominating supply chain and pizzas? Like, I don't think so. [8:23] Anish Acharya: It took 40 years for us to get from the inception of electricity to reorganizing factories. And that means like... burning the buildings down and starting from scratch versus taking what was previously coal and simply swapping it with electricity. So, I do think that like the most ambitious companies are rethinking everything around the models. [20:19] Anish Acharya: If we assume these things are infinitely intelligent and astonishingly cheap, how would we reorganize the company? [31:20] Anish Acharya: In many ways we spent 40 years building a technology that enables better spreadsheets. We built this like technology that extends our intellect, but nothing to extend our soul. [42:48] Anish Acharya: The kind of aura that Anthropic got from having a model that was too dangerous to release was extraordinary. And maybe they had a GPU shortage. Now, maybe the capabilities were more advanced than they actually wanted. You know, maybe they actually wanted to keep that proprietary model internal to extend their own lead. So, I think there's a lot of sort of confounding factors that would cause you to pull back a little bit. [46:29] Anish Acharya: The entire trend of human existence has been that our desires grow faster than our ability to fulfill them. If you look at the things that are expectations today, they were unimaginable luxuries 500 years ago, even 50 years ago for things like therapy, right? Or things like antibiotics 100 years ago... didn't matter how rich you were, you simply didn't have access to it. So, I think we're underestimating human ambition, human desire. People are going to be mad that they don't have a vacation home in Mars in 20 years. Like really mad... They'll be seeing it on Insta and be like, 'Come on, babe. We got to like work harder and make this happen.' [46:29] Anish Acharya: You know what's better than me doing it? It's Claude doing it. [54:30] Jesse (Decagon): Moats are most often discovered, not designed. [59:26] Chris Dixon: come for the tool, stay for the network. [61:24] Anish Acharya: Nobody has a growth problem these days, they have a product problem. [61:24] Anish Acharya: If we imagined our product cost a thousand dollars a month, ten thousand dollars a month, like what if our product was a software Birkin bag? What would it have to do to justify that? Okay, let's figure out how we build that. [64:30] Marc Andreessen: We were going to the moon or we were going to leave a moon-size crater in the ground. And there was no other option. [75:31] Anish Acharya: Don't discover things through painful experience that somebody can just tell you.

[0:00] There's a lot of fear and worry about [0:02] the future with AI. I want to talk about [0:04] this idea that if you fall behind, [0:06] you're going to become part of this [0:08] >> permanent underclass. [music] [0:09] >> It's a funny dark fantasy that we seem [0:10] to have as Silicon Valley collectively. [0:12] Like things have never been better by [0:15] almost every measure. [0:16] >> This is a technology that really [0:17] amplifies our agency. It kind of [0:19] unbundles skill from desire. Not only [0:22] can we dramatically [music] drive [0:22] productivity, we can dramatically drive [0:24] ambition. [0:25] >> Can you get too ambitious? Is there like [0:26] a limit? [0:27] >> In the old days, 3 years ago, we would [0:29] see a company and if what they were [0:30] trying to do was too ambitious, we [0:32] would, [music] you know, not engage. [0:33] Today, we're almost seeing the opposite [0:35] problem. An idea that's too small is not [0:37] something that we want to engage with. [0:39] >> You have this interesting take that [0:40] company building more and more is going [0:42] to become this kind of series of [0:43] creating loops. [0:44] >> We're going to see this sort of [0:45] cascading set of everything from a loop [0:47] per person [music] to loops that can run [0:50] large parts of the company. With that [0:51] said, I think humans are a critical [0:54] ingredient. The loop will help you climb [0:56] to the local maxima, but then it [0:57] plateaus. You need human intuition. You [0:59] need somebody to actually help you land [1:01] at the base of the next hill. [1:02] >> You have this take that the big [1:04] opportunity is this idea of loop make me [1:06] happier. [1:07] >> We believe that people want to be more [1:09] productive, but they don't. I think more [1:11] people want to spend time than save [1:13] time. So, I think that the opportunity [1:15] for this technology is the basics of [1:16] consumer need. How do we feel more [1:18] [music] connected, more loved? How do we [1:20] make progress? How do we have fun? I [1:22] don't think it's a model or a capability [1:24] challenge. It's just a product design [1:26] challenge. [1:29] >> Today, my guest is Anish Acharya. Anish [1:31] is general partner at A16Z where he [1:33] focuses on consumer investing. He is one [1:36] of the most insightful, [1:37] thought-provoking, mind-expanding, [1:39] in-the-weeds product [music] investors [1:41] I've met. He's been at A16Z for over 7 [1:44] years now and unlike a lot of VCs and [1:47] why I loved having Anish on the podcast [1:48] is that he is a long-time product [1:50] builder and founder. He founded a [1:52] company called Social Deck which he sold [1:53] to Google [1:54] >> [music] [1:54] >> and then ended up leading a number of [1:56] efforts within Google, then he started a [1:58] new company called Snowball, [music] [1:59] which he then again sold, this time to [2:01] Credit Karma, where he moved to VP [2:03] [music] of product and then GM of the [2:05] broader consumer product and the entire [2:07] credit card business. This conversation [2:09] [music] will get your mind buzzing. [2:11] Before we get into it, don't forget to [2:13] check out Lenny's product pass.com for a [2:15] free year of the hottest and most [2:17] beautifully crafted AI products in the [2:19] world, [music] available exclusively to [2:21] Lenny's newsletter subscribers. With [2:23] that, I bring you Anish Acharya. [2:28] >> [music] [2:29] >> Anish, thank you so much for being here [2:31] and welcome to the podcast. [2:33] >> Thank you, Lenny. I'm so excited to be [2:34] here, thrilled. [2:36] >> I want to start with a very light topic. [2:38] I want to talk about this meme of the [2:40] permanent underclass. [2:43] It's kind of this joke that people have [2:45] joke about, this idea that if you kind [2:48] of fall behind and aren't just like on [2:50] top of all the latest AI tools, aren't [2:52] becoming the most productive person [2:53] ever, [2:54] you're going to become part of this [2:56] permanent underclass and fall behind and [2:58] have a really hard time. [2:59] And some people are joke about it. I [3:01] think a lot of people take this really [3:02] seriously and stresses a lot of people [3:04] out. [3:05] How real of a concern do you think this [3:07] really is? How seriously do you think [3:08] people should take this? [3:09] >> Not very seriously and it's a funny dark [3:11] fantasy that we seem to have as, you [3:13] know, Silicon Valley collectively. Like [3:14] things have never been better really by [3:17] almost every measure by how sort of [3:20] distributed all the opportunities are, [3:22] by the kind of technology we have access [3:23] to, to the types of ambition we're [3:25] allowed to have, and to the number of [3:27] companies that are sort of independently [3:29] working on things that are winning. And [3:31] yet, there's this sort of discussion of [3:33] permanent underclass, being outside of [3:35] the light cone, I've heard it. And it's [3:37] not just something for deep insiders or [3:39] outsiders. It feels like there's a real [3:40] fear kind of from, you know, researchers [3:43] at Foundation Model Labs all the way [3:46] through to the Silicon Valley layman. I [3:48] mean, here's like a couple of points [3:50] that I think are really important. So, [3:51] first, I think the last era of tech was [3:54] a lot more centralized. If you look at [3:55] network effects, that's sort of the gold [3:57] standard. You worked on a network [3:58] effects product. Um that's the gold [4:01] standard of businesses from the mobile [4:03] era. And those things led to dramatic [4:05] centralization, right? Of course, all of [4:07] them are definitionally sort of end of [4:09] one networks. If you look at what's [4:11] happening now, it's like every part of [4:12] the stack, there's not even two relevant [4:15] players, there's like 20. You know, [4:16] you've got labs, you've got open weight, [4:18] you've got different variations within [4:20] both. If you look at coding agents, we [4:22] were talking about it, like your mental [4:24] model for 2 years ago should have been [4:26] would have been, I think it should be [4:27] winner take all. And yet, Claude Code, [4:30] Codex, Lovable, Replit, Wobby, like [4:33] they're all sort of working. So, it's [4:35] really really encouraging to see that. [4:37] You know, the second thing, you've heard [4:38] all the kind of economic data, [4:39] everything from radiologists um who are [4:42] supposed to be cooked uh every year for [4:43] I think about 20 years now. And of [4:45] course, job postings are higher than [4:47] they've ever been, as well as [4:48] programmers, you know. [4:49] So, I don't know that the empirical data [4:51] bears it out. I I think the final thing [4:53] is that there's this sort of discussion [4:56] about RSI, and I know RSI is like [4:58] recursive self-improvement is a fun term [5:00] to throw around. But, if you ask the [5:02] most sophisticated individuals at the [5:04] labs, it's not actually RSI that's [5:06] occurring, which could lead to some sort [5:08] of runaway winner cuz they were an [5:09] epsilon ahead of the others. It's [5:12] autocatalytic effects, which just means [5:14] you're using the new technology to [5:16] improve your process, but it's not truly [5:17] recursive. So, I think everything from [5:20] the most empirical to the most technical [5:21] viewpoints in the other direction, and [5:23] yet we can't seem to let go of this [5:25] fantasy. [5:26] >> Something I've been thinking about [5:27] recently is seeing all these like even [5:29] seeing these crazy um stories about [5:31] OpenAI's models hacking hugging face. [5:34] >> Yes. [5:34] >> All these stories, to me feels like [5:37] there's always been this question of are [5:38] we on the fast takeoff or the slow [5:39] takeoff scenario? And it feels very much [5:42] so that we are on the slow takeoff [5:43] scenario, because every one of these [5:45] milestones is like, "Holy it [5:47] hacked at We had no idea it was doing [5:49] this, But, but like we're catching it, [5:50] we're watching it, we're observing it, [5:52] we're iterating, evolving. There's [5:54] always this fear, okay, but tomorrow [5:55] it's going to take off. [5:56] What I'm hearing from you is that's [5:57] probably not the case, which I think is [6:00] the source of a lot of people's fears, [6:01] this idea that all of a sudden it's [6:02] going to become super super intelligent [6:03] and then we're in big trouble. [6:05] >> That's right. Like the line of reasoning [6:06] for that case is always everything up [6:09] until now, then something happens that [6:11] no one can quite articulate and then [6:13] fast take off. [6:14] So, I don't believe that that's going to [6:16] happen. I do think that model progress [6:18] is happening faster than ever before. [6:20] But if you look at something like [6:22] economic diffusion, you know, I grew up [6:24] in a small town. I went back home last [6:25] summer, like people's lives haven't [6:27] changed that much. So, if nothing else, [6:30] the sort of slow rate of economic [6:31] diffusion will catch it. I think the [6:33] other thing that's under discussed, [6:34] Lenny, is, you know, how many problems [6:36] are truly intelligence bound? Like if [6:38] you had a, you know, a data center of [6:40] PhDs working at FedEx or Domino's Pizza, [6:43] are they going to be like exponentially [6:45] dominating supply chain and pizzas? Like [6:47] I don't think so. So, I think we might [6:49] be overestimating how many problems are [6:51] intelligence bound versus bound by other [6:53] things. [6:54] >> This episode is brought to you by our [6:55] season's presenting sponsor, WorkOS. [6:58] What do OpenAI, Anthropic, Cursor, [7:01] Replit, Sierra, Clay, and hundreds of [7:03] other winning companies all have in [7:05] common? They are all powered by WorkOS. [7:08] If you're building a product for the [7:09] enterprise, you've felt the pain of [7:11] integrating single sign-on, SCIM, RBAC, [7:14] audit logs, and other features required [7:16] by large companies. [music] WorkOS turns [7:18] those deal blockers into drop-in APIs [7:21] with a modern developer platform built [7:23] specifically for B2B SaaS. Literally [7:25] every startup that I'm an investor in [7:27] that [music] starts to expand upmarket [7:29] ends up working with WorkOS. And that's [7:31] because they are the best. Whether you [7:33] are seed-stage startup trying to land [7:35] your first enterprise customer or a [7:37] unicorn expanding globally, WorkOS is [7:39] the fastest path [music] to becoming [7:41] enterprise ready and unblocking growth. [7:43] It's essentially Stripe for enterprise [7:45] features. [7:46] >> [music] [7:46] >> Visit workos.com to get started or just [7:48] hit up their slack where they have [7:50] actual engineers waiting to answer your [7:52] questions. [music] [7:53] WorkOS allows you to build faster with [7:55] delightful APIs, comprehensive docs, and [7:57] a smooth developer experience. Go to [7:59] workos.com to make your app enterprise [8:02] ready today. [8:03] >> What are you seeing inside of companies [8:06] in terms of um is there a more of a [8:08] divide happening? And do you think there [8:10] will be more of a divide between the [8:11] people that are becoming really good and [8:13] embracing versus like I don't have time [8:15] for this. I hate all this stuff. My [8:17] job's already so stressful. What are you [8:19] seeing happening and where do you think [8:20] things will go inside of companies in [8:22] terms of maybe a divide? [8:23] >> I mean, I have so many thoughts on this. [8:25] I don't think we give the average [8:26] employee enough credit. I think we have [8:28] this abstraction of a white-collar [8:30] employee. You know, the white-collar [8:31] manager, the abstraction is like some [8:33] Dilbert-esque manager who's just [8:35] shuffling paper all day long. You know, [8:37] it is abstraction of consumers that [8:39] they're sort of these low-agency NBCs. [8:41] Of course, that would never apply to us [8:42] or our friends. [8:43] You know, we have this abstraction that [8:45] everybody else's job is super [8:47] automatable by AI, but of course ours is [8:48] not. So, I think when you actually get [8:50] into the details, a lot of people are [8:53] actually excited to you know, better [8:55] themselves, get more leverage, and you [8:57] see this with of course sophisticated [8:59] companies like Google, but even a [9:00] company like Kavak where they sell, you [9:03] know, used cars in Mexico, they've got [9:05] this concept of a Jedi Academy where [9:07] they're teaching everybody at the [9:08] company including the mechanics how to [9:10] use the new tools and technologies and [9:12] kind of at the end of the 6-week course, [9:13] they ship a cutting-edge in-production [9:16] agent. So, I actually think that more [9:17] people are embracing the technology than [9:19] we sort of like to discuss. I think a [9:21] big change is going to be using AI um [9:25] versus reorganizing your entire company [9:27] around AI. And a great example of this [9:29] is if you look at the diffusion of [9:30] electricity as a technology, you know, [9:32] it took 40 years for us to get from the [9:34] inception of electricity to reorganizing [9:37] factories. And that means like re- like [9:39] burning the buildings down and starting [9:41] from scratch versus taking what was [9:43] previously coal and simply swapping it [9:45] with electricity. So, I do think that [9:47] like the most ambitious companies are [9:49] rethinking everything around the models. [9:51] And those that are a little less [9:53] ambitious or perhaps a little earlier [9:55] are thinking more about how do we give [9:57] people and existing orgs existing job [9:59] functions access to the technology. [10:01] >> So, a kind of a theme I'm hearing so far [10:03] is we can be a little less stressed [10:06] about where things are going and and the [10:08] future of your job, your careers. [10:10] >> I think so, man. I mean I I think if you [10:12] even just think of the kind of [10:13] incentives for the CEO and executives, [10:15] you know, Sundar running Google, he [10:17] doesn't want to run a more efficient $4 [10:18] trillion company. He wants to build a [10:20] $40 trillion company. So, anytime you [10:23] have an economically productive unit, [10:25] it's rational to kind of especially if [10:27] it gets more productive to maintain that [10:29] sort of presence in your organization. [10:31] And then I don't know what you hear, but [10:32] anecdotally I talked to a good friend [10:34] who's an executive at Google and I said, [10:36] "Hey, have you laid anyone off?" And he [10:37] said, "No, we didn't. What we instead do [10:39] is now rip through our road map. So, 2 [10:41] years of road map happens in 3 months. [10:44] And where actually our hardest problem [10:45] is knowing what to add to the road map, [10:47] which by the way is like every PM's [10:49] fantasy. You know, how much emotion has [10:51] been drained on prioritization [10:53] conversations between you and I?" [10:55] So, yeah, I actually think that people [10:57] shouldn't be as stressed and I think [10:59] they should feel really empowered. And [11:00] and by the way, the best way to do it is [11:02] just to ship stuff. I mean, Claire is my [11:04] muse. Uh she is so awesome cuz she's [11:07] always shipping. Yes, Claire Bow. She's [11:09] shipping. She's trying things. She's not [11:12] afraid to be a little embarrassed by it. [11:14] And if you just see her whole kind of [11:16] affect, she feels like the best version [11:18] of herself that she's ever been. And I [11:20] think we all have an opportunity to be [11:21] that. [11:22] >> Mhm. [11:23] I love that. I want to be Claire Bow [11:24] when I grow up. [11:25] >> Totally. [11:26] >> Uh kind of along those lines, you have [11:27] this interesting take that company [11:29] building more and more is going to [11:31] become this kind of series of creating [11:33] loops and creating series of loops. Talk [11:35] about that. [11:36] >> Yeah, yeah. Well, I think the broad [11:38] concept and you know, it's I [11:40] the loops concept kind of gets teased a [11:42] little on X cuz at sometimes it feels [11:44] like maybe we're big braining it. I know [11:46] there was a a big meme around graphs, [11:48] like the next stage of loops is graphs. [11:50] Um but let me make the kind of steel man [11:51] for it, which is [11:52] you know, we had prompts and then we [11:55] invented agents. Agents, of course, are [11:57] just models in a loop with tools and [11:59] memory and skill files. And then we had [12:02] sort of loops, which are, you know, sets [12:04] of agents that are doing tasks. If you [12:06] look at a lot of work in coding, coding [12:08] such a great domain because you have the [12:10] best models and you have the most sort [12:12] of technically apt customer. Plus you [12:14] have these established loops. Like bug [12:16] fix or sort of bug report comes in, [12:19] uh repro gets generated, bug fix gets [12:21] created, bug fix gets reviewed. If high [12:24] risk, then human should confirm that [12:26] it's okay to ship to prod. And if low [12:28] risk, it just gets shipped. And maybe [12:30] you even email the customer and say, [12:31] "Hey, we fixed the bug you reported." [12:33] That happens in 5 minutes. There are [12:35] many loops like that in engineering, [12:36] everything from, you know, bug fixes to [12:38] customer feedback to sales demos um to [12:42] new feature development. Um [12:44] So, coding sets this up well itself up [12:46] well. So, you have a coding loop and the [12:48] kind of change it makes is to the code [12:50] base. My question is, what are the [12:52] business loops? Right? So, if you're the [12:54] GM of a business, you're looking across [12:56] many job functions and you've got loops [12:58] running in coding and marketing and [13:00] sales and support and legal. The output [13:02] of all of those loops is something that [13:04] is itself a loop that you should be able [13:06] to optimize for. And I think the strong [13:08] form of this is that it sends a message [13:10] to the CEO saying, "Hey, we need to [13:12] actually make a change to one of the [13:14] physical aspects of the business or to [13:16] our business model or to our strategy." [13:18] So, I think we're going to see this sort [13:19] of cascading set of everything from a [13:21] loop per person, loop per job function, [13:24] loop across entire business units to, [13:27] you know, loops that can run large parts [13:29] of the company. [13:30] With that said, I think humans are a [13:32] critical ingredient. I just don't think [13:34] that most work in the organization can [13:36] be done fully autonomously. When you [13:38] think of what a human will do in this [13:40] like AI native company, sales, support, [13:43] strategy, and exceptions. Right? And all [13:46] those things are super critical. If [13:47] we've seen one thing, Lenny, it's that [13:48] the ability for models to do new [13:50] thinking out of distribution thinking is [13:52] still really limited. And I don't [13:54] actually take the point that some of the [13:55] new thinking in math is actually [13:57] representative of new thinking in [13:58] domains like business. So, you're still [14:00] going to need a person to say, "Hey, [14:02] here's the thing I think we should [14:03] make." and have them be right about it. [14:05] >> Let me just kind of make sure this point [14:07] is really clear cuz it's so interesting. [14:09] What you're saying here is engineering [14:11] and building more and more is becoming [14:12] this loop of input, feedback, [14:15] or support ticket, whatever, input of [14:18] just like what to build, and then AI [14:20] more and more is taking that, deciding [14:22] here's a PR, here is this ready, and [14:24] then shipping [14:25] And you're saying that you expect that [14:27] to spread [14:28] to like say go to market, legal, uh [14:31] growth, support. So, maybe describe what [14:33] that loop looks like or may look like [14:36] for within a company. [14:38] >> I mean, a great example is a growth [14:39] team. You worked on the growth team at [14:40] Airbnb, right? [14:41] >> Yeah. Yeah, supply growth. Yeah, that's [14:42] right. [14:42] >> Yeah. Awesome. Right. So, you remember [14:44] those like war I don't know how you ran [14:46] your team, but I'm guessing it was [14:47] something like you got everyone [14:48] together, you built a list of possible [14:49] experiments, you prioritized them, you [14:51] built, then you shipped them, you [14:53] measured them. [14:53] >> Brainstormings. Yeah. A lot of [14:54] brainstormings. A lot of spreadsheets. [14:56] >> So, the the the loop version of that [14:58] should be that every variant gets [14:59] generated, every variant gets measured. [15:02] Once you get to stat sig with a high [15:04] enough P value, you can merge and ship [15:06] that variant. You then have a long-term [15:08] holdout, and you start working on the [15:10] next experiment. You know, and then [15:12] you're going to hit some local maxima. [15:13] And I think this is really important, [15:15] you know, the loop will help you climb [15:16] to the local maxima, but then it [15:18] plateaus. And you need some sort of out [15:20] of distribution thinking, you need human [15:22] intuition, you need somebody to actually [15:24] help you land at the base of the next [15:26] hill. [15:26] >> Yeah, you have this chart. I don't know, [15:27] maybe we'll show it over later as we [15:29] talk about this, which is such an [15:31] interesting way of thinking about it. [15:33] This idea that agents will help you [15:35] hill climb and reach some new plateau [15:38] and then you need a human there to un- [15:40] think about a bigger idea, kind of [15:42] unlock it and then it keeps going and [15:44] going and there's kind of this like [15:45] agent to human kind of back and forth. [15:47] >> Yes. Yeah, and you know what really [15:49] illustrates that? If you've ever tried [15:50] to have an agent come up with a business [15:52] idea for you? [15:53] Like, you know, "Hey Claude, make me a [15:55] million dollars, make no mistakes." [15:56] Like, why doesn't that work, you know? [15:59] And it's because you sort of need to set [16:01] it in the right direction and and [16:02] nothing in the technology has shown us [16:04] that that is not needed. [16:06] >> Yeah, it's interesting what like I've [16:07] been hearing more and more. I just saw a [16:09] tweet that I think at OpenAI, the [16:11] go-to-market team now is using Codex [16:13] more more of the go-to-market team is [16:15] using Codex more often than even the [16:17] engineering team. [16:19] >> Yes. [16:20] And and think of how happy that makes [16:21] them. Like, what does the go-to-market [16:22] team want to do? I mean, this is a [16:24] caricature, but I'm going to stand [16:25] behind it, which is they want to hit the [16:27] gym, [16:28] they want to go to steak dinners, and [16:30] they want to like, you know, [16:32] like raise the trophy up for being [16:34] salesperson of the quarter of the year. [16:36] Um so, I actually think that this is a [16:38] distillation of their job into the thing [16:40] that they're the best in the world at, [16:41] that they're the most interested in, um [16:44] and all the administration that goes [16:45] around doing that core work is now [16:47] handled for them. So, like that's where [16:49] go-to-market is going and it's going to [16:50] be awesome. [16:51] >> This reminds me of a PM friend who has [16:53] this had this really funny take that as [16:56] a PM, [16:57] you're constantly having to say no to [16:59] all these ideas that are coming at you [17:00] and you have like, "I'll put it on the [17:01] road map. We'll prioritize it." He's [17:03] like, "Okay, I'm going to flip this. I'm [17:04] going to say yes to everything. I'm [17:05] going to build everything and then [17:07] simulate every of every idea with like [17:10] Simily or all these products that are [17:12] launching where you can simulate how a [17:13] user will react. And then that'll tell [17:16] you should this be should we build this? [17:19] What a hilarious way to rethink PM. [17:21] >> And you know what's so beautiful about [17:22] that? Actually, there's two things. I'll [17:24] tell you maybe the one that's less [17:25] obvious to me, which is I feel like [17:26] every PM at every company feels like [17:29] they're a true zero-to-one thinker, but [17:31] they're held back by the kind of, you [17:33] know, the heavy hand of management and [17:35] executives and founders and engineering [17:38] capacity. [17:39] And in a world where every story gets [17:41] told, every product story gets told, [17:43] every feature gets tried, I think a lot [17:45] of PMs are going to realize they're [17:46] actually not that good at zero-to-one [17:47] and it's much more fulfilling to work on [17:49] someone else's good idea than your own [17:51] bad idea. [17:52] So, I think even things like that are [17:54] going to lead to a lot more [17:56] organizational health than we've had in [17:57] the past. You know, not to mention the [17:59] fact that the idea that ends up winning [18:01] doesn't have to be the one that's came [18:03] up with by the person who can sell it [18:05] best to an executive. It just gets tried [18:07] and the best idea wins. [18:09] >> So, coming back to this loops idea, the [18:11] the way I'm thinking about it is how can [18:13] every function start to think of their [18:15] function as setting up an agent to be [18:17] able to just go from input to [18:20] >> Yeah. [18:20] >> some kind of impact. And what this makes [18:23] me think about is something actually [18:24] Clear tweeted recently. This point that [18:27] people always used to joke that like [18:28] soft skills are the least valuable and [18:30] engineering skills the most important [18:32] and valuable because they're so concrete [18:34] and it turns out that's what AI is the [18:35] best at. The things that are verifiable [18:37] and and, you know, you know what success [18:40] looks like. And [18:41] so that the question I think about now [18:43] is just like which skills can you not [18:45] just turn into a loop because the output [18:47] is so hard to verify. [18:50] >> Yeah, I think that's right and I think [18:51] that those are going to be the [18:52] rate-limiting factors cuz you can only [18:53] do one steak dinner a night. I guess you [18:55] could do a steak lunch, but, you know, [18:57] to some extent there's going to be these [18:59] rate-limiting factors in every system. [19:02] Um I think a useful way to think about [19:03] it is anytime the model's making a [19:05] mistake or doing something you wouldn't [19:07] do, [19:08] what do you know that it doesn't know? [19:10] Um and there's a really interesting [19:11] example I heard from Ale at Kavak. He's [19:14] probably the most sophisticated thinker [19:16] on this stuff that I get to hang out [19:18] with, where he said anytime they were [19:19] agent they have an agent per customer, [19:21] they sell used cars online, and when the [19:23] agent gets stuck it actually calls a [19:24] human. [19:26] And the human will coach the agent [19:27] through. Now, the magic of that is not [19:29] only does it unblock the agent, but of [19:30] course the agent then captures all the [19:32] traces and learns from it. Like that's a [19:35] really interesting mental model. It's [19:36] either a knowledge gap or a data gap [19:38] that you have to give the agent, and the [19:40] next time it shouldn't have to call you. [19:42] >> I love that example. Basically, the [19:44] takeaway here is [19:46] you need to start thinking about every [19:47] function as an agent loop, and the job [19:50] is to [19:52] figure out where it gets blocked, where [19:53] it goes wrong, and give it more context, [19:55] more insight, more direction, basically. [19:58] >> That's right. That's right. And then [19:59] hopefully a lot of your day-to-day work, [20:01] you know, when you come up with a new [20:02] idea that works, the kind of [20:04] implications of that idea, if it's a [20:06] product, you know, there's marketing [20:07] work, there's sales work, there's [20:08] product marketing, there's [20:09] communication, there's legal, all of [20:11] that should be largely handled for you. [20:14] And your idea, you know, your job is to [20:15] go take a hike and dream the dreams and [20:18] come up with the next hill to climb. [20:19] >> And this begs the question a bit of just [20:21] what will separate the companies that [20:23] win in this world where AI is kind of [20:25] doing a lot of this? [20:27] I imagine part of the answer is the [20:28] human in this local maxima coming in [20:30] with the better idea. Is there anything [20:32] else that you think becomes like a [20:33] differentiator in this world where AI is [20:36] doing so much of the work that humans [20:38] are currently doing? [20:39] >> I mean, I One thing that's under [20:40] discussed is that um [20:43] it's unclear that the sort of [20:44] competitive equilibria that exists in a [20:47] lot of industries will really change. [20:49] You know, so let's say Pizza Hut and [20:51] Domino's and Papa John's and Round Table [20:54] all get a data center of PhDs, [20:57] and you know, [20:58] they're all going to either adopt it or [21:00] have a CEO change and adopt it. So, [21:02] it'll be a rocky period and there'll be [21:04] some, you know, relative shuffling, but [21:05] I think that they're all going to [21:06] embrace new technology cuz most [21:08] companies do. I don't know that one of [21:10] them is going to have 99% of the market. [21:12] So, I think kind of what's under [21:13] discussed is that yes, in the near term, [21:15] I think there'll be winners and losers [21:17] based on adoption of the technology and [21:19] kind of how ambitiously you adopt it, [21:21] but I do think there's a lot of [21:22] industries that will sort of maintain [21:24] their current competitive dynamics [21:25] because they're not intelligence bound. [21:28] Um I think the most useful question to [21:30] ask yourself as a founder CEO is just, [21:33] "Hey, if we assume these things are [21:35] infinitely um intelligent and [21:37] astonishingly cheap, how would we [21:39] reorganize the company?" Cuz that's [21:41] where we're going. [21:42] >> Kind of along those lines, you have this [21:43] interesting take about this kind of [21:44] split that might be coming within [21:46] companies like between these uh [21:48] generalists and specialists and how AI [21:50] plays into that. Talk about that. [21:52] >> Yeah, well, I think that there's so many [21:53] interesting things that this touches on. [21:55] You know, one is the question of open [21:57] weight versus frontier. So, I I'm sure [21:59] Are you familiar with kind of Pareto [22:01] efficiency? I'm sure you are. [22:02] >> Uh please explain. [22:03] >> Yeah, so Pareto efficiency is just, you [22:05] know, um for price performance, for [22:07] example, what is the kind of the [22:09] efficient frontier is what's considered [22:11] the optimal trade-off of, you know, a [22:14] unit of performance for a unit of price. [22:16] And, you know, are you sort of if you're [22:18] along that curve, you're always paying [22:20] the rational amount for the performance. [22:23] And what's interesting is that frontier [22:25] models are actually irrationally priced [22:28] in that, you know, first of all, Mythos [22:30] is infinite dollars a token, you can't [22:32] use it. [22:33] Um but if you look at even something [22:34] like Fable 5, you know, for one IQ, [22:37] conceptually one IQ of extra [22:39] intelligence, you're paying 100X more [22:41] than Opus 48. So, it's not rational, but [22:44] I think there's a lot of jobs in which [22:47] you have unbounded upside like drug [22:49] discovery, you know? And if that one IQ [22:51] point lets you discover the next, you [22:53] know, statin, it's a trillion-dollar [22:55] outcome. So, it's sort of rational to [22:57] pay for the highest intelligence in [22:59] these types of jobs and industries. Now, [23:01] on the other hand, um I'm going to pick [23:03] on, you know, maybe legal. [23:06] There's perhaps only so much upside to [23:07] be had in legal or finance. And for [23:09] those jobs, you actually do want to be [23:11] very Pareto efficient and probably pay [23:13] for, you know, good performance at a [23:15] good price, not infinitely priced [23:18] infinite potential upside performance. [23:21] So, I think what we're going to see is a [23:22] split between job functions that de- [23:25] demand kind of mid-IQ um intelligence. [23:29] And those will often be open-weight, [23:31] sort of biased with reinforcement [23:33] learning, you know, things that make the [23:35] models even cheaper and more performant [23:37] for a narrow job. Along with, you know, [23:39] incredibly um quote-unquote expensive, [23:42] but performant frontier tokens for [23:44] sales, support, research, engineering. [23:47] So, I think you're going to end up [23:48] having both architectures. And, you [23:50] know, we can touch on this in a little [23:51] bit, but I do think there are these sort [23:52] of comparative advantages amongst model [23:55] families. Um and then also amongst, of [23:57] course, individual models that we're [23:58] seeing more and more of. It's not going [24:00] to be one or the other. [24:01] >> Such an interesting insight. So, just to [24:03] make sure I understand what you're [24:04] describing here, you're thinking there's [24:05] going to be this split between kind of [24:07] within an org of [24:09] function and model, where specific [24:12] functions that have a lot more upside [24:13] and in- potential leverage go for the [24:16] frontier models. And you you describe [24:18] these kind of product, sales, [24:20] engineering, research roles. And then [24:22] there's like And then for other [24:24] functions, you don't need you don't need [24:25] Mythos. You don't need Astra, I think. [24:27] Is that Is that the least one coming [24:28] out? [24:29] >> Yeah. [24:30] >> And so, it's both you're saying that AI [24:32] doesn't need to be the frontier model [24:33] and the people don't have to be the [24:34] smartest people in the world to do [24:36] really well in that world. [24:37] >> Yeah, and I don't I don't want to be [24:38] diminutive. Like, there's extraordinary [24:40] people in those job functions. I just [24:41] think that they're bounded upside [24:43] problems that they work on. You know, [24:44] you can only kind of uh close the books [24:48] correctly. You know, you can't close the [24:49] books 100x better. [24:51] So, yeah, that that's exactly what I'm [24:53] saying. [24:53] >> I wonder if it's connected back to that [24:54] discussion we had earlier about it's [24:56] verifiable. If it's a lot more [24:58] verifiable, you don't need the frontier [25:00] model versus I don't know, the potential [25:03] upside. [25:04] >> I'm not sure. I mean, I think for a [25:07] verifiability is a good question is I [25:08] think for a like a drug development [25:10] company, you have infinite upside. It is [25:13] verifiable, but closing the books is [25:15] also a verifiable problem, which has [25:17] limited upside. I think the question is [25:19] really how how much upside is there and [25:22] how hard is it to calculate it? Like [25:23] here's a nuance. Customer support, a [25:25] customer may call in and report a bug. [25:28] That bug may actually be the first [25:29] breadcrumb to a thing that changes our [25:32] entire organization. And if the CEO is [25:34] on that call or the smartest person, [25:35] they could follow the trail. [25:37] Um but if we actually have this [25:38] quote-unquote mid IQ generalist then [25:40] not, that actually makes the case for as [25:42] a basket of problems always use frontier [25:44] intelligence. Um and I can of course [25:46] make the case for the mid IQ [25:48] basket as well, which is simply that [25:50] we've crossed an intelligence threshold [25:52] for almost every economically useful [25:53] problem. And anything beyond that [25:56] threshold is is simply waste. [25:57] >> Yeah, and like I think I think people [25:59] point people forget is that the models [26:01] that are not the frontier models, they [26:03] were like that was what the frontier was [26:05] I don't know six months ago and we were [26:06] so impressed and we loved it. It was [26:07] like, "Holy it can do all this." [26:09] And now just because there's something [26:10] better, we don't give those models as [26:12] much credit. [26:12] >> You're right cuz it's sort of like this [26:14] thing is so crazy, you know, this is um [26:17] this is AGI and then a day later it's [26:20] like the old thing that you throw in the [26:21] dustbin. [26:22] >> So kind of along those lines, uh I hear [26:24] people jokingly call you a model [26:26] sommelier. [26:28] Tell us with your sommelier credentials, [26:30] what's kind of like the current state of [26:32] the model arts? What what are each What [26:35] are What is each model great at? What's [26:36] each What are models terrible at? [26:38] >> Yes, well, the you know, as you know, [26:40] the secret of every sommelier is 10,000 [26:41] hours or maybe 10,000 bottles. So I [26:44] think that the secret of being a model [26:45] sommelier is just using them all. [26:47] >> Drinking them all. [26:48] >> Um I push myself really hard to ship [26:49] something with every new model that [26:51] comes out. And I think you learn so [26:53] much. You know, I think for people who [26:54] believe the models are commodities or [26:57] totally fungible, you just haven't [26:59] actually used the models. And you know, [27:01] for example, in the last few weeks, I've [27:03] been obsessed with Qwen 38B. [27:05] Um Qwen is an awesome model. It's very [27:08] good at long horizon tasks, and it's [27:11] actually really creative. It's a great [27:13] storyteller. So, I've been using it to [27:15] create these uh impossible [27:17] documentaries, um mostly of, you know, [27:19] planets in the Star Wars universe. [27:21] I did uh Tatooine, which turned out [27:23] great. I did Bespin last night. And it [27:25] can just work for 4 or 5 hours. It tells [27:27] a great story. It generates all the [27:29] video using the MiniMax model via file. [27:32] Generates audio via 11. It actually like [27:35] directs the movie, the 5-minute movie, [27:37] in that it like cuts all of the clips [27:39] in. It overlays it. It watches it. It's [27:41] just extraordinary, you know? And that's [27:43] just It's got a totally different shape [27:45] than GLM 52. GLM 53 just came out, which [27:49] I used for a bunch of product work. It [27:50] doesn't have a vision component, and [27:52] it's sort of like this neurotic PhD you [27:54] put in the corner. And they both have [27:56] their role, right? This is a little bit [27:58] of the kind of these tension of models. [28:00] It's not that one is ahead of another, [28:01] one is more intelligent. It's rather one [28:04] is sort of has a mind that's shaped in [28:06] one direction, perhaps creativity and [28:08] openness for Qwen, and others that are [28:11] shaped in other directions, like, you [28:13] know, neuroticism and precision, like [28:15] GLM 53. Um [28:17] So, that I just make something with [28:18] every model, and that's how I build my [28:21] intuition. [28:21] >> How important is this habit, do you [28:23] think, for people? Because I hear a lot [28:25] that it's really important to be using [28:28] these models. [28:29] And kind of a secondary question is, how [28:31] do you come up with what to do with [28:32] these models? Cuz a lot of people want [28:34] to try these things. They're like, [28:35] "Okay, what do I do?" [28:37] >> I know. [28:37] >> Any suggestions? [28:38] >> Well, I think you Almost all of us have [28:40] got, you know, the most insufferable [28:41] thing for for a long time was your app [28:44] idea friend. You know, every time you [28:46] went to have a beer, they're like, "Let [28:47] me tell you my app idea." You're like, [28:48] "Ugh, here we go again." Like, we have [28:50] got to be the the app idea guys uh now. [28:53] And all the silly ideas, actually, [28:55] especially the silly ideas, cuz those [28:57] are the ones that often have the most [28:58] alpha, are the ones that we should all [29:00] be building. So, I probably got, you [29:02] know, two dozen apps that I build. I've [29:04] got one or two large apps that I iterate [29:06] on. And I think that if you don't have a [29:08] chassis on which to like with which to [29:10] use the models, it's just really hard to [29:12] come up with an idea from scratch every [29:14] time. [29:15] So, I I'd say like work on something. [29:17] It's actually better if it's not [29:18] important with a capital I. And then [29:20] keep finding new ways to invest in it [29:22] and add to it with a model as a kind of [29:24] tool rather than as a goal. Does that [29:26] make sense? [29:27] >> Yeah. Like maybe even zooming out. I've [29:29] been thinking more and more one of the [29:31] most important habits to build right now [29:32] is to whenever you're about to do [29:35] something, ask yourself, "How can AI do [29:37] this for me?" [29:38] >> Totally. [29:38] >> And and visualizing this like input to [29:39] response, you know, like the whole idea [29:41] of there's a space between input and [29:43] response. And meditation helps you [29:45] uh think more deeply before you respond. [29:48] And I feel like the trick now is insert [29:50] in that moment, "How can AI help me with [29:52] this?" [29:53] >> Yeah, no, I think that's that's I mean [29:54] I'm also a long-time meditator. We [29:57] should talk about a lot of you like, but [29:59] yeah, I think that's right. I think in [30:00] our day-to-day knowledge work, sometimes [30:02] it's less obvious to me. I guess my mind [30:04] is I've always been a consumer product [30:05] person. I love products. So, for me it's [30:07] easier to think of a new feature to add [30:09] to my DJ streaming app or um you know, [30:12] my sort of Google Reader for X that I [30:14] use than it is to kind of find a part of [30:16] my life to automate. But you're right, [30:18] there's some fun examples. You know, I [30:19] posted about one a few weeks ago where I [30:21] had my laptop um transcribe everything [30:24] that was happening in the kitchen. [30:25] And then it would um [30:27] award or detract screen time from my [30:30] son's iPad depending on whether he was [30:31] being good or bad. Um So, it was like a [30:34] fun little social experiment. It also [30:36] had a fun outcome, which is that he [30:37] recorded a video of himself saying, "I [30:39] love you, Dad." over and over and put it [30:40] next to the mic [30:42] >> to to hack the metrics. [30:43] >> To hack the metrics. So, yes. Uh so, [30:46] yes, the kind of you know, we [30:48] anything becomes a measure, it's no [30:50] longer useful. Um but, it was just a [30:51] cool little social experiment. I think [30:53] those things are super fun. [30:54] >> That is hilarious. And one of So, one of [30:56] the measures was how often he said I [30:58] love you. [30:59] And I was going to give him more screen [31:00] time. [31:00] >> he saying things that are positive and [31:02] prosocial or or or negative and [31:04] antisocial? And it you know, saying I [31:06] love you is very prosocial. [31:07] >> That's so funny. I had a friend who [31:09] built this little device device that [31:10] measured how often he and his kids laugh [31:13] throughout the day. [31:14] >> Oh, that's nice. [31:15] >> Yeah. He hacked like one of his [31:16] Limitless pendants to do that. And then [31:18] they just look at that metric every day. [31:20] >> Isn't that so And this is what I mean, [31:22] you know, man? Like, I think we spent a [31:23] lot of the time on the show so far [31:24] really like intriguing talking about [31:27] productivity, job loss, kind of all [31:29] these like heady important topics with a [31:31] capital I. But, what you just described, [31:33] how often you laugh, like that's not a [31:35] startup. That's probably not an [31:37] economically consequential idea, but it [31:40] is for like the quality of our life. And [31:42] I think we tend to think that happiness [31:44] is fixed, but what if it's not? You [31:46] know, if you and I were doing jobs 100 [31:47] years ago, like what would our jobs be? [31:49] I promise they'd be less cool than they [31:51] are right now. And And maybe 100 years [31:52] from now, they'll be that much cooler. [31:55] So, I think that the thing that gets [31:56] often missed, and this is why I'm a fan [31:58] of Claire and others, is just like how [31:59] can this thing add texture to our human [32:02] lives even if they don't have economic [32:03] consequences? [32:05] >> Yeah. Along these lines, I I saw [32:07] somewhere you have this really [32:08] interesting take that like the big [32:10] opportunity, maybe just in consumer but [32:12] broadly, is this idea of loop make me [32:15] happier. [32:16] >> Yeah. [32:17] >> Talk about that. [32:18] >> Oh, yeah. I mean, I think that we [32:20] believe that people want to be more [32:22] productive, but they don't. I think more [32:24] people want to spend time than save [32:26] time. There's a reason the biggest [32:28] products in the world are kind of [32:29] entertainment and social. So, we get at [32:32] the heart of how do we sort of deliver [32:34] the value to the consumer? I think for [32:36] most consumers, you know, I sometimes I [32:38] I tease and I say it's like the [32:39] Instagram AI user versus the X AI user. [32:42] The X AI user is like, you know, fearful [32:44] of being outside of the permanent [32:45] underclass, is really opinionated on GLM [32:48] 53 versus Kimmy K3, like they're just so [32:51] pilled. And then the Instagram person is [32:53] like, "Oh, this is like a better Google [32:55] search, kind of. It's cool, you know, I [32:56] don't get what all the hype is about." [32:58] So, I think that a lot of the And it's [33:00] really a product design failure. We have [33:02] the capabilities to radically transform [33:04] people's lives. I mean, Lenny, in many [33:06] ways we spent 40 years building a [33:07] technology that enables better [33:09] spreadsheets, right? We built this like [33:10] technology that extends our intellect, [33:12] but nothing to extend our soul. And I [33:15] think that we have a little bit of a [33:16] spiritual hunger, especially as a lot of [33:18] these cultural institutions have gone [33:19] away that fulfilled that, especially in [33:21] the rest of America, you know, where you [33:23] don't have as many hot yoga classes and [33:25] Pilates and fasting and Friendsgiving. [33:28] So, I think that the opportunity for [33:29] this technology is like, "Hey, that the [33:31] basics of consumer need. How do we feel [33:33] more connected, more loved? How do we [33:36] make progress? How do we have fun?" Like [33:38] the things that we all aspire to, the [33:40] basics. How do we apply these tech this [33:42] technology to those areas? That's what I [33:44] want to see more of. And And again, I [33:45] don't think it's a model or a capability [33:47] challenge, it's just a product design [33:49] challenge. [33:50] >> I love this so much. We talked about [33:51] this idea of loop, like grow my [33:53] business, loop find me more sales, loop [33:56] close support tickets. But the way [33:58] you're describing here, I think you the [33:59] way I had it had my notes here, just [34:01] like loop improve my health or loop make [34:04] me a better friend. And there's no [34:05] reason AI can't just think deeply and [34:08] hard about all those things and figure [34:09] out a way to actually do this. [34:10] >> Yeah, and also, you know, look, there [34:12] there's a we have to dial this in, but I [34:14] think there's a way that AI can sort of [34:16] challenge you, can push you, can be [34:18] disagreeable. This is also why I think [34:19] startups are advantaged over incumbents. [34:22] You know, the idea there's a thousand [34:23] Google committees who would, you know, [34:26] roll in their graves at the idea that [34:27] they're going to release a model that's [34:29] disagreeable or or God forbid it should [34:30] be like sexually suggestive. Or but [34:32] guess what? Those are all parts of human [34:34] existence. So, I think exploring the [34:36] kind of uncomfortable parts of our [34:38] social existence are things that [34:40] startups are uniquely set up to do. [34:42] >> And I know you spent a lot of time [34:44] investing in consumer companies. What [34:45] I'm hearing here is this is a big [34:47] opportunity for consumer businesses to [34:49] basically build a [34:50] app product that is exactly this loop [34:52] around improve my connections with my [34:55] friends and family. [34:56] >> I think that's right. And you know, I [34:56] think the thing that's held back [34:57] consumer so far a little bit and it [34:59] maybe held back is just strong cuz if [35:00] you if we kind of put this into iPhone [35:02] terms, we're in iPhone 2010, right? What [35:05] is iPhone 2010? I think it's pre-Airbnb, [35:08] pre-WhatsApp, pre-Uber, pre-all the [35:10] important kind of consumer companies. Um [35:13] so, it is early days, but what's held us [35:15] back is I think three things. One is [35:17] that the models have been expensive. So, [35:19] if you want to do a kind of free to use [35:21] product, it's that hasn't been easy. [35:24] The second is that we've kind of had an [35:26] interface problem. Like chat makes sense [35:28] if you're the highest agency person in [35:30] the world, which is Elon and Sam. But [35:32] for the average consumer, like their [35:34] ideal interface is TikTok. [35:36] So, we need to find something between [35:38] chat and uh TikTok. And then the fact [35:41] that the technology has been so much [35:43] more focused on productivity than things [35:45] like, you know, human connection and [35:47] entertainment. I think all those things [35:49] are kind of up for grabs. Like the open [35:50] weight models mean things are way [35:52] cheaper. I think that we're starting to [35:54] have conversations about things like [35:56] loop make me happier. And I think that [35:58] founders like Eugenia uh who should have [36:00] been on the show, she's tremendous, are [36:02] thinking really ambitiously about user [36:04] interfaces. Brian Chesky from Airbnb had [36:05] he started a foundation lab focused on [36:08] next gen user interfaces. So, I think [36:10] all those problems will [36:11] get solved or they're at least in a [36:13] better position to be solved than they [36:14] were 2 years ago. [36:15] >> I want to come back to this the whole [36:17] space of consumer and AI and things like [36:18] that. But I I want to follow this thread [36:21] a little bit more about just [36:23] the optimism around where things might [36:24] go. There's a lot of fear and worry [36:28] about the future as with AI and just [36:30] generally. Marc Andreessen when he came [36:32] on the pod had this really interesting [36:34] take that AI came just in time to save [36:37] us because population is declining, [36:40] productivity is going down, there's all [36:41] this [36:42] war, climate change and all these things [36:45] and we would be in big trouble if AI [36:47] wasn't here to fill that gap. [36:49] Talk about just kind of your bigger [36:50] picture perspective on [36:52] why you think maybe people are [36:53] underestimating the the positive and [36:56] optimism around AI. [36:58] >> One, I think it it's a potential for it [36:59] to be a sort of emotional spiritual [37:02] interface on which we can kind of get [37:03] leverage and explore aspects of [37:05] ourselves that have been really buried. [37:07] If you look at the kind of effect of the [37:08] Industrial Revolution, it's that there [37:10] are these scale advantages which are [37:12] insurmountable. And as much as obviously [37:14] I'm very pro-capitalism and I love the [37:16] economy that we live in, I think that [37:18] centralization it sort of discourages [37:22] the individual in some ways and it maybe [37:24] detracts from their identity. So, I [37:26] think one of the really magical things [37:27] is this a technology that really [37:29] amplifies our identity, our agency. It [37:31] kind of unbundles skill from desire. For [37:34] example, if you want to make music, you [37:35] can make music now. You don't have to [37:37] know how to play the piano. If you want [37:39] to be a programmer and make software, [37:40] you can make software now. So, it really [37:42] amplifies our individuality. It allows [37:45] us to explore aspects of our lives that [37:47] we were never able to explore before. [37:50] And you know, just in terms of the nuts [37:51] and bolts, like we have been in this [37:53] sort of morass of 2% GDP growth. Like [37:56] who said that we have to be there? Why [37:57] can't we be 10 or 15 or 20%? And this is [38:00] a technology with which not only can we [38:02] dramatically drive productivity, we can [38:04] dramatically drive ambition. Like think [38:06] of the maybe this is a caricature, but [38:08] the 1950s and 1960s, we believed that we [38:10] could do anything, right? We were coming [38:12] out of World War where the entire [38:14] economy, the entire world mobilized in a [38:16] way that we didn't think was possible. [38:19] And one of my theories Lenny is that [38:20] like when the stakes are high, we are [38:22] awesome. When the stakes are low, we are [38:24] at our absolute worst. And in a lot of [38:26] ways, I think the world we lived in 5 [38:28] years ago felt like a low stakes world, [38:30] which is why we kind of collectively had [38:32] a lot of these like side projects as a [38:34] society which weren't necessarily [38:36] productive or making any of us happier. [38:38] And now you've got, you know, it's not [38:40] just Elon doing everything he's doing, [38:41] but I think everybody feels like they're [38:43] climbing the ambition ladder. You ship [38:45] your bad ideas so you can discover your [38:47] good ideas. And if you want to, you [38:49] know, build software, great. If you want [38:51] to build a bridge, you don't have to be [38:52] a civil architect to know how to do that [38:54] anymore. So, I really think that we have [38:56] the makings of a sort of dramatically [38:59] happier, more fulfilled, more productive [39:01] society. And, you know, and yet we're [39:03] here talking about permanent underclass. [39:05] >> I love all of that. That all feels so [39:07] right. It's so hard to really believe [39:09] that. [39:10] But if you actually look at how things [39:12] have gone so far with the rise of AI, [39:15] uh unemployment's down, people are [39:16] making a lot of money. You know, [39:17] obviously a lot of people are [39:18] struggling. There's a lot of downsides. [39:20] Data centers causing problems for [39:22] people, things like that. A lot of a lot [39:23] of issues. But it feels like as a [39:25] economy and as a country, feels like [39:27] things are going well so far. Like they [39:29] just almost like cured some kind of [39:31] cancer the other day. So, yeah. [39:33] >> Yeah, you're right. Moderna just did I [39:35] mean, the fact that Dario can write a [39:37] blog post and say, "What happens when we [39:39] cure every disease?" And then we debate [39:41] it as a serious topic. Like what world [39:43] are we living in here, you know? Um I [39:45] also think that we're collectively [39:48] worried always. Like you know, the the [39:50] kind of these abstractions, the world, [39:53] the average worker, the middle manager, [39:55] the person who lives in the vicinity of [39:57] a data center. These are abstractions. [39:59] But our actual lives feel like they're [40:01] getting more fulfilled, we're more [40:02] capable, we're more empowered. So, I [40:04] think that's also something then the [40:06] revealed preferences tend to show how [40:08] people are experiencing individually. [40:10] And the stated preferences tend to show [40:11] how people are sort of observing it or [40:13] believe it's playing out societally. [40:16] You know, if you ask people if they want [40:17] a data center in their neighborhood, [40:19] most will say no, but if you ask them if [40:21] they use ChatGPT today, most people will [40:23] say yes. So, that's kind of the [40:24] dissonance. [40:26] >> And the the obvious issue is that just [40:27] the PR around AI has not been great. [40:29] There's a lot of fear-mongering. [40:31] Thoughts on on that? What's going on [40:33] there? Do you think that all change? [40:34] >> The The most important thing that we can [40:36] do with AI to change the kind of [40:38] conversation around it is make important [40:40] things cheap. [40:41] There's two things that are [40:42] extraordinarily important in America [40:44] that have only gotten more expensive, [40:45] right? Health care and education. [40:47] Um if you look at health care, 45% is [40:49] administrative. [40:51] So, if you take a lot of that [40:52] administrative burden out, you can [40:53] actually see deflationary health care [40:55] costs. Also, things like curing every [40:57] disease, that sounds awesome. You know, [40:59] GLP-1s, also obviously not an AI thing, [41:02] but I think is a [41:04] reason to be optimistic about um [41:06] deflationary health costs. And then [41:08] education as well. I think education now [41:10] has the strongest form of competition, [41:12] sort of traditional education, that it's [41:14] had in 200 years. [41:16] Um and I think it's going to be very, [41:18] very good to kind of unbundle learning [41:20] from institutions. And and and also, by [41:22] the way, like status from credentials. [41:25] You know, like you don't need a Harvard [41:26] degree, you just need to get [41:29] And that's pretty cool. [41:30] >> This episode is brought to you by [41:31] Mercury, radically different banking now [41:34] with Spend. I've been a Mercury customer [41:36] for so many years now. I switched all my [41:38] business banking to Mercury, and [41:40] honestly, I could not be [music] [41:41] happier. It's what online banking feels [41:43] like when it's built by product people, [41:46] not [music] by bankers. And now with [41:48] Spend, you can give your team individual [41:50] cards, set spending limits per person [41:52] [music] or per team, and have expense [41:54] receipts automatically pulled in from [41:56] Gmail or over [music] text. You can even [41:59] give your AI agents their own cards with [42:00] their own limits and policies. Most [42:03] founders start out the same way. [42:04] >> [music] [42:04] >> One card used by everybody at the [42:06] company. It works until it stops [42:08] working. [music] Someone goes over, a [42:10] receipt disappears, you spend 2 days [42:12] trying to figure out who spent what and [42:14] why. Spend is expense management built [42:17] directly [music] into Mercury. All your [42:19] team's cards, budgets, and [42:20] reimbursements all live in the same [42:22] place as your business banking. No [42:24] chasing, no manual reviews, no [42:26] end-of-month scramble. [42:27] >> [music] [42:27] >> The result is a team that can move fast [42:29] and a founder who's no longer the [42:31] bottleneck. [42:32] Learn more and get signed up at [42:33] mercury.com. [42:35] Mercury is a fintech company, not an [42:37] FDIC-insured bank. Banking services [42:39] provided by Choice Financial Group and [42:40] Column N.A. members FDIC. The IO card is [42:43] issued by Patriot Bank N.A. member FDIC [42:45] pursuant to a license from Mastercard [42:47] International Incorporated. [42:48] >> I saw OpenAI recently slow down their AI [42:51] development. They paused [42:52] >> Yeah. [42:53] >> their RL kind of phase on their latest [42:55] model because of [42:57] what they're seeing. [42:58] So, that's obviously a big concern for [42:59] people, just how fast and smart these [43:01] models get. Any thoughts on just that? [43:03] That's like a big shift now. Instead of [43:05] race ahead to the fastest, best model [43:07] ever, okay, we actually have to slow [43:08] these things down. That feels crazy. [43:10] >> I mean, without commenting on OpenAI [43:12] specifically, I think that um maybe I'm [43:15] a little skeptical on some of these [43:16] things where I think the kind of the [43:17] aura that Anthropic got from having a [43:20] model that was too dangerous to release [43:23] was extraordinary. And maybe they had a [43:25] GPU shortage. Now, maybe it was you [43:28] know, the capabilities were more [43:29] advanced than they actually wanted. You [43:31] know, maybe they actually wanted to keep [43:32] that proprietary model internal to [43:35] extend their own lead. So, I think [43:37] there's a lot of sort of confounding [43:39] factors that would cause you to pull [43:40] back a little bit. Look, I do take the [43:42] points about offensive cyber seriously, [43:46] which is we should harden all of our [43:47] systems before we make them trivial to [43:49] penetrate. Um but I think the sort of [43:52] concept of the model that's too [43:53] dangerous to release, it it kind of [43:55] conflates marketing, um inference [43:58] capacity, [43:59] and then also economic considerations [44:02] like do you want to externalize your [44:04] competitive advantage or use it to make [44:05] yourself better? [44:06] >> Yeah, I always think about that when I [44:08] have folks from Anthropic and and open [44:10] AI on the podcast just like how an [44:11] advantage they have when they have the [44:13] best model. It's crazy, right? Just like [44:15] that's a loop right there is just have [44:17] the best model for longer and they can [44:18] move so much faster. [44:20] >> It's crazy though, you know, to take the [44:22] other side for a moment, like it felt [44:23] like Anthropic was unassailable and now [44:26] open AI has had an amazing 6 months and [44:28] open weights are also ripping. So, [44:30] despite the like the sort of scary [44:32] concept of this like you know supremely [44:34] intelligent model that's totally [44:36] proprietary to one company, [44:38] the so far the kind of industry trends [44:39] haven't played that way at all. [44:41] >> Yeah, like Grok bot just came out of [44:43] nowhere and is now like the most amazing [44:45] AI kind of assistant tool. I'm just [44:47] hooked on it. [44:48] >> Oh man, I mean we should talk about the [44:50] personal agent thing. Like actually the [44:52] three big products that I love here, [44:54] Grok bot totally nailed it. And also the [44:56] model underneath it is awesome, right? [44:58] They came out of you right out of left [44:59] field. Um and that's I know a lot of the [45:01] good work that Chris or team did. I [45:03] think chat GPT work it's kind of buried [45:05] in the UI, but it's really really good. [45:07] It's one of the best products they've [45:08] released. And then there's a a startup [45:10] that's really getting some buzz called [45:12] Instinct, which has made some more [45:14] aggressive and interesting trade-offs, [45:16] but all in the same domain. [45:17] >> That's amazing. I I don't know if it was [45:19] a strategy for some ran someone tweeting [45:21] this vague [45:22] tweet about how awesome it is and [45:24] everyone's like, "What the hell are you [45:24] talking about?" [45:26] That was really effective. Cuz it was [45:28] like, "Where did you get that?" Uh chat [45:29] GPT work the episode that will come out [45:32] right before this is the [45:34] PM in charge of that Atara. [45:36] >> Oh, really? [45:36] >> Yeah. [45:37] >> Oh man, they they did such a good job on [45:39] it. It's really I don't know how much [45:41] you use it, but it's really well done. [45:43] >> What is the what is it better [45:45] than co-work there? Is it that that it [45:47] runs in the cloud? Is that the big [45:48] differentiator? [45:49] >> It runs in the cloud. It does a good job [45:51] of kind of caching browser credentials. [45:53] Though it's not as aggressive as Grok or [45:55] Instinct. And it also the remote [45:58] feature, it's got the full deep duplex [46:00] voice mode. So you can actually just [46:02] call it. It it can see all your threads [46:04] and you can just talk to it and say, [46:05] "Hey, what's happening across all my [46:06] coding agents, across this? Can you [46:07] change that?" So, it's just because the [46:09] full duplex voice is so good, you really [46:12] feel like you're calling your assistant [46:14] who knows everything that's happening in [46:15] your world, whereas with Grok, it's a [46:16] sort of one-way transcription. [46:18] Right? Or with some of the other [46:19] assistants, it's a text. So, the voice [46:22] plus the kind of model of it can see all [46:24] threads is really well done. [46:25] >> Awesome. Okay, I'm going to come back to [46:26] this AI assistant consumer stuff, but [46:29] something I wanted to kind of close the [46:31] thread on. I feel like there's in terms [46:33] of [46:34] jobs and the economy as a result of AI, [46:38] I see it almost as the spectrum of [46:40] there's like the Dario end of the [46:41] spectrum of 50% of knowledge work will [46:44] be disrupted and we're all going to have [46:46] massive unemployment to like the Sachs, [46:48] David Sachs spectrum of like it's going [46:50] to be incredible. [46:51] Jobs are going to be fine. So far, [46:53] everything is pointing in the good [46:55] direction. Clearly, you're closer to the [46:57] Sachs direction. Is there anything else [46:59] just along the lines that you think [47:00] might make people feel better about jobs [47:03] in the future? [47:04] >> I mean, I think that the entire trend of [47:06] human existence has been that our [47:07] desires grow faster than our ability to [47:10] fulfill them. [47:11] Um if you look at the things that are [47:13] expectations today, they were [47:15] unimaginable luxuries 500 years ago, [47:17] even 50 years ago for things like [47:18] therapy, right? Or things like [47:20] antibiotics 100 years ago, right? I [47:22] mean, didn't matter how rich you were, [47:23] you simply didn't have access to it. So, [47:25] I think we're underestimating um human [47:28] ambition, human desire. You know, people [47:30] are going to be mad that they don't have [47:31] a vacation home in Mars in 20 years. [47:33] Like really mad, like this really mad. [47:35] They'll be seeing it on Insta and be [47:36] like, "Come on, babe. We got to like [47:38] work harder and make this happen." Every [47:40] CEO is going to want to build a much [47:41] bigger company. So, don't they already, [47:43] it feels like we're living a much larger [47:46] form of sort of human existence than we [47:48] could have imagined 100 years ago? [47:50] And there's no reason that trend won't [47:51] continue or accelerate. [47:53] >> This idea of ambition, I was I'm glad [47:54] you even brought that word up again. [47:56] It's something that's coming up a lot on [47:57] this podcast. That not only is AI making [48:00] it easy to be a lot more ambitious, it's [48:03] almost making us have to be more [48:04] ambitious because everybody else can [48:07] just do all the easy stuff now. [48:09] And now what separates us is just how [48:10] big can you go? Okay, I'm going to make [48:12] a personal website. Okay, it's cool. [48:14] It'll be this white simple background. [48:15] No, okay, I'm going to make it this 3D [48:17] game where you have to like walk through [48:18] a world and discover all the things that [48:20] you like everything is just getting more [48:21] epic. Uh thoughts on just this idea of [48:24] ambition becoming a bigger, I don't [48:26] know, scale and habit. [48:27] >> A lot of things get conflated because [48:29] when you and I say ambition on this pod, [48:31] I think people have a very specific idea [48:32] of what that ambition is, you know? [48:34] Ambition to be a founder, ambitious to [48:36] build a beautiful software product. Like [48:38] those are types of ambition, but there [48:40] other types of ambition. Let's talk [48:41] about creative ambition. You know, when [48:43] you're five, nobody says, "Lenny, you're [48:45] good at painting, but you're bad at [48:46] drawing." You know, you just have an [48:48] ambition or a desire to make something [48:50] and you make it. Like that's something [48:52] that's very unique and can now actually [48:54] be encouraged. [48:55] >> You know, think of a very local [48:56] ambition. So, you know, they have the [48:58] NHS in the United Kingdom. I think it's [49:00] a sort of treasured institution that's [49:02] not working well. Maybe the way AI shows [49:04] up in their society is making like the [49:06] NHS as good as the iPhone. Right? That's [49:08] very specific and local to them. Or [49:10] simply the ambition to be more connected [49:12] to our family, to be more present [49:13] parents. So, the ambition doesn't have [49:16] to be sort of ambition in the narrow [49:17] economic sense. It can be really [49:19] anything that we want to do more of and [49:21] who doesn't have that in their bones? [49:23] >> Yeah, that's such like my example is so [49:24] dumb now that I think about it. Like the [49:26] It feels like the thing we have to uh [49:28] unblock in our brain is like, "Okay, AI, [49:30] solve cancer." [49:32] We're not We're not like [49:34] Like that's where we can start to think [49:35] now and it's so unnatural to us, [49:37] especially as product people that always [49:39] have to think about the MVP and the [49:40] constraints. Now we have to think big. [49:42] Okay, what can What's the big What's the [49:44] 10x the 1000x version of this? [49:46] >> Yes. I know, right? In a way we have to [49:47] think small because this thing that [49:49] we've all Our whole lives have been [49:50] built around how precious software and [49:53] intelligence is, and now it's totally [49:54] not precious. Like that probably will be [49:57] a harder change for you and I. [49:58] >> Yeah. So, like all these new little [50:01] habits [50:02] uh I think a lot about on the Claude [50:03] code team, they have a principle. [50:05] Uh you know what's better than me doing [50:07] it? It's Claude doing it. [50:09] And that just created this habit in [50:11] everyone on the team. How do I help [50:12] Claude do this thing for me? And I feel [50:14] like that's a thing we all have to start [50:15] to build in our head. And And Grok Bot's [50:17] really good at that. Like and not an [50:19] investor not no affiliation, but it's [50:21] just like so simple and good at this [50:22] stuff. [50:22] >> I also think that there is so much [50:24] learning that happens through doing. You [50:26] know, it's funny, if you look at the [50:27] number of people that are talking about [50:28] vibe coding versus the number of people [50:30] that are talking about their projects, [50:31] people are a little embarrassed. I'm a [50:33] little embarrassed to talk about a lot [50:34] of my projects cuz they don't seem [50:35] important or substantial enough. But so [50:37] much of it is learning through execution [50:40] or shipping or or being fulfilled [50:41] through execution and shipping. Like I [50:43] don't think we can underestimate that [50:44] either, you know. [50:45] >> I have a guest post coming together from [50:47] uh someone at Google that works on a lot [50:49] of their lab stuff. And I don't want to [50:51] give away the the goods, but just [50:53] broadly her concept is that building is [50:55] now the new reading where you build to [50:58] learn and infuse and experience, and [51:00] it's totally okay for most of it to go [51:02] throwing away because that's still it it [51:05] building your your muscle, you know. [51:07] >> So well said. Yeah, it's uh building as [51:09] an activity rather than an outcome. [51:11] >> Right. [51:13] >> Yeah. [51:13] >> Which I think a lot of people feel bad. [51:15] I shipped all these things, but no one's [51:16] using it or ever use it. And I think [51:18] that the key here is like that's [51:19] actually okay, and that's totally fine. [51:21] >> Oh man, I mean and it happens in every [51:22] other domain, you know? Like I make a DJ [51:24] set and like three people listen to it, [51:26] and I listen to it a hundred times. [51:28] And it's like very fulfilling. It [51:29] doesn't matter. [51:30] >> Mhm. Uh we're going to talk about your [51:31] DJ stuff later. [51:32] >> Okay. Okay. [51:33] >> [laughter] [51:35] >> Okay, let's come back to consumer stuff. [51:37] So, you focus on consumer at A16Z. [51:40] What's kind of like what's happening in [51:41] consumer these days? What's kind of like [51:43] the landscape? What are you excited [51:44] about? [51:45] >> So, I think there's three big areas. [51:46] We're very early as we kind of [51:48] discussed. You know, the labs have done [51:49] a good job, but there aren't as many [51:51] sort of independent mass market consumer [51:53] products. Um coding agents are awesome, [51:55] and I think a total lightning bolt. You [51:57] know, I think it's easy to say most [51:58] people don't want to make code, but the [52:00] the big change in my thinking is that [52:02] coding agents are a way to interact with [52:04] the world generally. And you've seen a [52:06] lot of this on X. You know, people use [52:08] Claude code to edit videos, you know, or [52:10] Codex to you create a game that they [52:13] play with their kid on an airplane ride. [52:16] Like there it's sort of this general [52:18] problem-solving tool that consumers can [52:21] use in very unique ways. Um you know, [52:23] and and probably the best example [52:24] company that we've invested in is Wabi, [52:26] where it's sort of a platform for mini [52:28] apps. People can create them, consume [52:29] them, share them. So, coding agents is [52:31] one big area I think that's important. [52:33] Personal agents, we had this credible [52:35] moment around Open Claude, but guess [52:37] what? Open Claude is a dev thing, you [52:38] know? And Hermes, of course, and [52:41] Multibook. Like all of that is now [52:43] getting distilled into mass market [52:45] consumer and enterprise agents that [52:47] people understand, which is what we [52:49] discussed previously. And then the last [52:51] is I'm going to call it entertainment, [52:52] but that doesn't fully do it justice. Um [52:55] I think that it's a lot of sort of [52:56] creative tools. You know, Suno has done [52:58] such an amazing job. It's companionship [53:01] products. Um all of that that entire [53:03] area is, you know, uh uncomfortable to [53:06] talk about, so I think it's under [53:07] discussed, but there's some huge [53:09] fast-growing products there. So, I think [53:10] those are the three big areas that we're [53:12] watching right now. [53:13] >> That's really interesting. Just thinking [53:14] of it these three buckets. Coding [53:16] agents, [53:17] AI assistants, kind of Open Claude, but [53:19] much simpler and easier and more [53:20] reliable. [53:21] Uh sounds like basically the three [53:23] described there that you're excited [53:24] about are um [53:26] Instinct, uh Grok Bot, and and ChatGPT [53:29] work. [53:30] >> That's right. [53:30] >> And then Wabi, I guess, would fit in [53:32] that first one. It's like a personal [53:34] coding agent that can build whatever you [53:35] want for you. [53:36] >> Yes. Yeah, that's exactly right. Yeah. [53:37] >> And then entertainment is the third [53:38] bucket. Like like AI girlfriends and [53:41] that kind of stuff, right? Yeah, there's [53:42] companies it's like crazy. You look you [53:44] guys put out these market maps and like [53:45] five of them are these different like [53:47] companions. [53:48] >> Though though actually to be accurate is [53:50] more boyfriends and girlfriends [53:51] actually, you know. [53:53] Um the majority of people using [53:54] companion products are women that are in [53:56] their 40s and 50s actually. [53:58] >> Interesting. Mhm. Okay, cool. So So [54:01] those are the three kind of areas you [54:03] think the biggest opportunities will [54:04] come from? [54:06] And it's interesting that they connect [54:07] to this idea of loop. I don't know, make [54:09] me happier. Like Like they Yeah, these [54:11] are kind of These are kind of like the [54:13] jobs to be done for humans. Make me [54:15] happy. Uh make me healthier. Make me [54:17] live longer. That kind of stuff. [54:18] >> Yes. Yes, give me a channel for my [54:21] ambition, a channel for my sort of [54:23] fulfillment, and then a thing to do when [54:25] I'm not doing everything else. [54:27] >> Mhm. That makes sense. [54:29] Okay. [54:30] The other question I have for you along [54:31] these lines is [54:33] there's so many companies [54:35] and so many startups, so many products [54:37] everyone's launching. The speed at which [54:39] companies and products are shipped is [54:41] like thousand x-ing. How do you think [54:43] about durability and moats when you look [54:46] at a startup? Because a lot of founders [54:48] get that question. Everyone's getting [54:49] that question. How am I not a rapper? [54:51] What are What are signs that tell you [54:52] this might be a durable thing? [54:54] >> Well, I think there's two important [54:55] ideas. One is something that um Jesse [54:58] from Decagon said, which I love, and [55:00] that is that moats are most often [55:02] discovered, not designed. [55:04] I think it's really easy, I've done this [55:05] as a founder, to get in your own head [55:07] about like hey, I need a business plan [55:09] that survives scrutiny from MBAs and [55:11] VCs. I've got to have some really [55:13] sophisticated, you know, idea of what my [55:16] moat will be. And for that team, they [55:17] just started shipping and it developed [55:19] over time. Another great example of this [55:21] is Cursor. You know, they were [55:22] criticized a lot for not having a moat, [55:25] but it turned out that initially being a [55:27] high-end PSD AU product was really good. [55:30] And over time they captured all the [55:31] reasoning traces, they trained their own [55:33] models, the composer one two models, and [55:35] you know, so on and so forth. We know [55:37] how that story plays out. So, moats can [55:39] be discovered. They don't have to be [55:40] designed is one. And then, I think the [55:42] other is that we seem to have forgotten [55:44] that the classic moats, none of the [55:46] classic moats are based on how hard it [55:48] is to make the software. [55:49] You know, like we're not building [55:50] self-driving cars. Most of us aren't. [55:52] So, it's network effects, it's scale [55:53] advantages, it's brand effects, [55:56] proprietary sort of data or what was [55:57] historically called a cornered resource. [55:59] Every moat from 5 years ago generally is [56:02] still a good moat. We just need founders [56:05] that have ambition in those directions. [56:06] We need more multiplayer products. We [56:08] need consumer social. We need products [56:10] that get dramatically better the more [56:12] you use them like town. Um so on and so [56:14] forth. [56:15] >> So, still read Hamilton Hamilton Helmer [56:17] and uh all that stuff still applies. [56:20] >> Yes. Love this book. [56:22] >> Yeah, he's been on the podcast. [56:23] >> He has? [56:24] >> Yeah. [56:24] >> Maybe I missed that. Yeah. I feel like [56:26] there's only five real business books in [56:27] the world. Every other one is uh in the [56:29] business of selling business books. And [56:31] they're fake. And his uh his is on my [56:33] list of five. [56:34] >> Are there any other in this list that [56:35] come to mind real quick or anything? [56:37] >> The two that are so obvious are High [56:39] Output Management. Which is like that is [56:42] as good as it ever was. And then, Ben's [56:44] book, you know, Hard Things is it was [56:46] the first emotionally honest book about [56:48] business that was ever written. [56:51] And that's why founders love it. That's [56:52] why I love it. Cuz you read it and [56:53] you're like, "Wow, I'm not the only one [56:55] that's, you know, anxious and feels like [56:57] a failure and can't tell anyone what I'm [56:59] going through." Ben went through it, [57:00] too. [57:01] >> Mhm. Two of the most mentioned books on [57:03] this podcast, turns out. [57:04] >> That's a surprise. [57:05] >> So, so on this moat idea. So, say you're [57:07] a founder and you're just like, you [57:08] know, you're putting a pitch together [57:09] trying to pitch you [57:11] uh or other VCs. What's the best way to [57:14] talk about a moat? Is it like Can you [57:16] just say, "We're going to discover it. [57:17] We're not sure. Nobody really knows [57:18] yet."? [57:19] >> Yeah, I think that we would happily take [57:21] a bet on a product that doesn't have a [57:22] quote-unquote moat or durability story [57:24] if it has, you know, a lot of momentum, [57:27] a lot of craft, a lot of, you know, sort [57:29] of growing engagement. So, actually the [57:31] thing I've learned over the years, I [57:32] used to be very worried about people [57:33] stealing my idea, [57:35] but I've learned in that the big ideas [57:37] are always supported by a dozen small [57:39] ideas that are invisible. And even if [57:41] somebody replicates your big idea, they [57:43] they never see the small ideas that make [57:44] the big idea work. So, I actually think [57:46] that there's some there's some just [57:48] something special in the water with [57:50] certain products that make them [57:52] incredibly successful despite [57:54] extraordinary competition. I mean, look [57:55] at Granola. You know, 2 years ago, it [57:58] was really criticized and I don't know [58:00] that they have a super strong durability [58:02] story today and yet it is like beloved [58:04] and dominant. [58:05] So, they you know, you sort of I listen [58:07] to like what the customers are saying [58:08] more than, you know, what the what the [58:10] business books say. [58:11] >> You get a free year of Granolas become a [58:12] subscriber to Lenny's newsletter. I'm [58:14] going to I should do this more often [58:16] when it's part of this product pass. [58:17] Lenny's product pass.com. You get a free [58:19] year of all these amazing products [58:20] including Granola. [58:21] I just saw Ramp put out a report of the [58:23] fastest growing companies according to [58:25] their data and Granola is like number [58:27] two or three. [58:27] >> It's a tremendous product. It's the [58:29] craft is is really high. [58:31] >> Yeah, and then that I think about that a [58:33] lot these days [58:34] because there's so much like there's [58:37] co-work, there's chat GPT work, there's [58:39] cursor, there's grokbot and it's crazy [58:41] how quickly one can switch from one to [58:43] the other. And the underlying model is [58:45] not that different. All it really is [58:48] most of it is the harness and slash UX [58:51] of the product. And so to me that tells [58:53] you there's so much opportunity in the [58:55] actual user experience being a moat or [58:58] at least giving a lot of time to find [59:00] something that is durable. [59:01] >> 100% and I think for the people that are [59:03] at the edge, they pay for all of them [59:05] because they all have their respective [59:06] areas of specialization. [59:08] >> Yeah, I have many $200 a month plans [59:10] right now. [59:10] >> I know. I know, right? [59:12] So, it's a little painful, but yes, me [59:14] too. [59:14] >> And I think Cursor has a $300 a month [59:17] plan now. [59:17] >> So does Grok. Oh yeah, you're right. [59:19] Yeah, that was the Grok one. It's pretty [59:20] with [59:21] Grok heavy. [59:21] >> RIP. Yeah, RIP Cursor. I think they've [59:24] transitioned away from that brand. [59:26] Um, something I've been talking a lot [59:28] about is the is the distribution. [59:31] I call it distribution's new moat, but [59:32] it's like it's always been a moat, but [59:34] it feels like more and more [59:36] that is actually a massive [59:38] advantage because everybody is There's [59:41] like a thousand launch videos a day. [59:42] Everyone's launching launch launch [59:43] launch. And really [59:46] the ability to get your stuff into [59:48] people's atten- into people's feed, get [59:51] them continue to be reminded your [59:53] product exists feels like increasingly [59:55] is powerful and important. Thoughts on [59:57] the rising value of distribution [59:59] existing distribution being a big lever [60:01] for growth and success. [60:03] >> This is so important and you know, I [60:04] asked Chris Dixon about this um, because [60:06] he sort of author- authored the famous [60:08] come for the tool stay for the network. [60:10] I think the issue is that our entire [60:12] generation of founders and CEOs were [60:13] trained on the theory of networks and [60:15] network building. [60:17] And as a result, every network that [60:18] exists today is hyper trained to ensure [60:21] no one else builds a network on their [60:22] network. [60:23] So I actually think that the sort of [60:24] network effect has gone back to this [60:26] grassroots like true word of mouth. When [60:30] somebody is getting a ton of mentions on [60:32] X and on YouTube and on Instagram and [60:35] all these places organically, that is [60:37] probably the best form of the sort of [60:40] third-party network effect that you can [60:42] hope for today. And actually just like [60:44] the Web 2.0 era, unlike the mobile era [60:46] where you had the App Store and you had [60:47] growth hacking and you had all of these [60:49] sort of, you know, little cottage [60:51] industries, we have to kind of build our [60:53] own channels off of that word-of-mouth [60:55] growth. So in a sense, it's a it's a [60:57] purer growth problem but a harder one [60:59] than we've had in a couple of product [61:01] cycles. [61:02] >> And to get word of mouth, you need to [61:04] build something. I always think about [61:05] Seth Godin's line, build something [61:07] remarkable. [61:08] Something that people that is worth [61:10] remarking about. Which is basically, you [61:12] know, build an amazing product that [61:14] people want to talk about. Which is a [61:16] and you know, a very hard to do. And it [61:18] makes sense that because there's so much [61:20] happening, people are just going to pay [61:21] attention to what are my friends using [61:22] saying is worth paying attention to. [61:24] >> Well, I you know, here's the one thing I [61:26] would say that the here's the hopeful [61:27] point, which is I always say that nobody [61:29] has a growth problem these days, they [61:30] have a product problem. [61:32] I mean, the reason for that is like you [61:34] can build such a wildly ambitious [61:35] product in any direction, you know, [61:37] functional or emotional. [61:39] You can charge a lot of money for it. [61:41] So, my challenge is like, "Hey, is it [61:44] that you have a growth problem or is it [61:45] a failure of our collective imagination? [61:48] You know, if we imagined our product [61:49] cost a thousand dollars a month, ten [61:51] thousand dollars a month, like what if [61:52] our product was a software Birkin bag? [61:55] What would it have to do to justify [61:56] that? Okay, let's figure out how we [61:57] build that." [61:59] >> And it comes back to the ambition [62:00] question. [62:01] Yes. And still though, you still need [62:04] some [62:05] advantage to get in front of people, to [62:07] get it out there, at least initially, [62:09] because there's like a thousand things [62:11] launching every day. Imagine that's [62:12] still a big opportunity and and I don't [62:14] know, advantage, which to me make me [62:15] feel like incumbents have a huge [62:17] advantage. They have a their products, [62:19] they can tell you, "Hey, go use Gemini." [62:20] And every time you go search Google. I [62:23] guess Do you think Do you feel like [62:24] that? Do you feel like it's harder for [62:25] startups now because of this [62:26] distribution challenge? [62:28] >> I don't know. I I think it's easier [62:30] because like Gemini, despite all the [62:32] kind of heavy cross-selling Google has [62:33] done, nobody would say they're winning. [62:35] Um [62:37] you know, startups can build in [62:38] directions that incumbents are [62:39] uncomfortable building in, like [62:41] everything we talked around companion, [62:42] but there's many others. [62:44] >> It's true. [62:45] >> You know, prices can be pretty high. [62:47] Like people are open to paying 200 a [62:49] month for their enterprise, they're open [62:50] to signing million-dollar ACV contracts [62:52] without really knowing what they're [62:53] getting. So, I think like the kind of [62:55] the floodgates are open. It feels like [62:57] Christmas 2009, where everybody got [62:59] their iPhone and wanted to download new [63:00] apps. You know, that'll change at some [63:02] point. People will feel like they you [63:04] know, they're done and they're tired and [63:05] they don't want to try any more apps, [63:06] but the windows are open for now. I [63:08] think it's easier for startups. [63:09] >> That's a really interesting insight and [63:11] your line about how it's not a [63:13] distribution or growth problem, it's a [63:15] the problem, is such an important one. [63:16] Because if your product was that good, [63:19] people would talk about it and share it [63:20] and use it. [63:21] >> And it can be. I mean, what are the wild [63:23] social experiments that we're going to [63:24] see [63:26] with this technology kind of [63:27] intermediating them? [63:29] Um [63:29] You know, I think a lot about actually, [63:30] here's a fun example from a few years [63:32] ago. Have you heard of MSCHF? You know [63:33] MSCHF? [63:34] >> Yeah. Yeah. Yeah. [63:35] >> Right, they're awesome, right? They're [63:36] sort of like this creative studio that [63:38] uses technology as their medium. I'm [63:39] very web 2.0 in that way, actually. Many [63:41] of the kind of, you know, the Ev [63:42] Williams, the Kevin Roses, that's who [63:44] they were painters except technology was [63:45] their canvas. [63:47] So, they created this very funny product [63:49] called card versus card where they [63:51] shipped, I think, 100,000 people a debit [63:53] card and then every day they would text [63:55] those people with [63:56] um a location that they had to spend the [63:58] money at and they would put $100 on the [64:00] card and everyone would rush out spend [64:02] the money and one or two would be able [64:04] to spend it and everybody else would get [64:06] declined. And it was just this hilarious [64:08] social experiment that went hyper viral [64:10] and for me it was always inspiring in [64:12] the world of fintech cuz it was like, [64:13] "Wow, why don't we build more products [64:15] like that?" You know, money is [64:16] inherently social and yet all the [64:17] financial products we have are so dry [64:19] and personal and embarrassing. [64:21] I think there's a sort of similar moment [64:22] happening in AI right now where we can [64:24] build these wildly ambitious products [64:26] that touch on many of our social nerves. [64:28] We just have to do it. [64:29] >> So, let me follow that thread. [64:31] You get to see tons of companies both [64:33] pitching you and also companies you're [64:35] working with that you're investing in. [64:38] What are some [64:39] counterintuitive lessons you've learned [64:42] from watching the companies that operate [64:44] well and have succeeded? Uh lessons that [64:47] maybe go against typical wisdom, startup [64:50] wisdom. [64:51] >> I'll tell you the biggest one and you [64:52] know, in the old days, which is 3 years [64:54] ago, um we would see a company and if [64:56] what they were doing, trying to do was [64:57] too ambitious, we would you know, not [65:00] engage. Just too crazy, too complex. [65:03] You know, and implied by that is you [65:05] wouldn't do a $100 million seed because [65:07] just it's too much money for almost any [65:10] problem. It's too much money for any [65:11] person to actually manage. It's too much [65:13] money to build a the sort of talent to [65:16] absorb. Like it doesn't make sense as an [65:18] inception round. I think today we're [65:20] almost seeing the opposite problem [65:22] where, you know, an idea that's too [65:23] small is not something that we want to [65:25] engage with. And you can talk about how [65:27] you put $100 million to work in the seed [65:30] productively. Now, [65:31] I'm not recommending you raise $100 [65:33] billion, but I think that there's this [65:35] sort of no ceiling on ambition is also [65:36] showing up in how we're picking [65:38] companies. And and maybe how they're [65:40] picking us as well. Because when we [65:41] invest, we tell every founder like we're [65:44] here to help you build the strongest [65:46] form of your vision. You know, Mark told [65:48] this to me when him and I were talking [65:50] about coming here. And he said, "Aneesh, [65:52] the way we um sort of told our story [65:54] when we were raising our first fund was, [65:56] you know, we were going to the moon or [65:57] we were going to leave a moon-size [65:58] crater in the ground. [66:00] And there was no other option." That's [66:01] sort of how um how we want to work with [66:04] our founders as well. [66:05] >> I love that point. So, like is can you [66:07] get too ambitious? Is there like a [66:08] limit? Obviously, you look at the you [66:10] know, the team and what they're But kind [66:12] of the key lesson here is be more [66:14] ambitious. Uh it's like the opposite of [66:16] what used to be of like here's our [66:18] here's our wedge. Here's where we're [66:19] going to go. What you're looking for is [66:20] just how big is the idea. [66:22] I mean, look at Atoms. [66:24] >> Yeah. [66:24] >> How crazy is Atoms, you know? I mean, [66:26] what an incredible hero's journey for [66:27] all of us collectively. But also what [66:29] they're trying to do is something that, [66:30] you know, 5 or 7 or 10 years ago would [66:32] have felt insurmountable. [66:35] And now it's like, "Okay, it's [66:36] challenging. Let's see." You know? [66:38] >> Interesting. Is there anything else [66:39] Anything else that has changed or I [66:41] guess you've changed your mind about [66:42] around [66:43] what you think it takes to build a [66:44] successful company these days? [66:46] >> I mean, I think that the the kind of old [66:48] wisdom around consumer products have to [66:50] be free. I'm almost taking the opposite [66:52] take, which is let's think about [66:53] consumer products that are [66:54] extraordinarily expensive. I think every [66:57] um every part of consumer discretionary [67:00] spend is up for grabs right now. [67:02] And I think a really useful because [67:03] price is a measure product market fit, a [67:05] really useful product exercise is what [67:08] is the Birkin bag $10,000 a month, [67:10] $1,000 a month version of our product. [67:12] So, I think expensive consumer software [67:14] is something like new and important that [67:16] we wouldn't have thought about 5 years [67:17] ago. [67:18] >> I love that framing. Cuz it just pushes [67:20] you again to be more ambitious. [67:21] >> Yeah. [67:22] >> You mentioned Mark and you work closely [67:24] with Mark Andreessen and Ben Horowitz. [67:26] >> Yes. [67:27] >> What's What's one thing you've learned [67:28] from each of those guys? [67:32] >> They're such extraordinary leaders, [67:34] founders. I mean, the thing that I [67:36] actually feel so grateful to be a part [67:39] of that they really embody to their [67:40] core, I think is a feeling of [67:43] stewardship for the technology industry [67:46] and for really the country and the sort [67:49] of maybe the western way of living and [67:50] thinking. [67:52] You know, and if you look at sort of a [67:53] Ron Conway, you know, [67:56] or [67:58] Brook Byers, Tom Perkins, there was this [68:01] feeling, I think, of obligation [68:05] to sort of leave it better than you [68:06] found it from an industry perspective. [68:09] And that sort of aspiration goes way [68:12] beyond just being the best investor in [68:13] the world, though we want to do that, [68:15] too. I see them both show up that way [68:17] over and over again, where they want to [68:19] do hard important things that don't [68:21] directly benefit the firm or at least [68:22] not singularly because they're just sort [68:24] of important. You know, Ben has done a [68:26] lot of that in the direction of the [68:27] industry and and Mark in sort of the [68:29] direction of the country, though of [68:31] course both work on both. [68:33] Another really cool thing is just to see [68:35] how I think Mark and Ben, but maybe the [68:38] firm a little bit, has shaped our [68:40] collective ambition as a founder [68:42] community. I think if you look at 5 or 7 [68:44] or 10 years ago, deep tech was deeply [68:48] unpopular. You know, it wasn't a [68:49] high-status thing to be working on. It [68:50] was very fringe. [68:52] And and now it's become very popular, [68:55] high-status, and mainstream. And I think [68:56] there's a lot of firms that have, of [68:58] course, pulled in that direction, but I [69:00] think that um like Mark has been very [69:02] full-throated in his support of working [69:04] on sort of capital I important work in [69:06] the national interest. [69:07] And all of Silicon Valley has changed as [69:09] a result. [69:10] >> And you guys have had some big wins uh [69:12] in the past couple weeks investing wise, [69:14] too. [69:15] >> Yeah. [69:15] >> Thank you. [69:16] Congrats. [69:17] Uh final question before we get to our [69:20] very exciting lightning round. [69:22] What's your advice to product people who [69:25] are trying to [69:26] think about what they might want to [69:28] shift in how they work, how they think, [69:30] how they operate to be more successful [69:33] in the future in their careers and with [69:34] their companies? [69:35] >> Just make more things. And I know it [69:37] sounds silly. I know everybody says it, [69:39] but just please like come up with a [69:41] project. You don't have to tell anyone [69:43] about it. It can be totally unimportant. [69:45] Um but use it as a chassis to use all [69:48] the new models, ship things, talk about [69:50] them, build your own intuition. I [69:52] promise you're one sort of slightly [69:55] frustrating and then very fulfilling [69:57] week away from being as pilled as [70:00] anyone. So, you just got to use the [70:02] technology. Um and if not now, then [70:04] when, right? This is all of us got in [70:06] the game to build the products we saw in [70:09] our mind's eye. [70:10] Um and now you have a chance to do it. [70:12] So, just please please use the models. [70:14] You know, and and tell me what you [70:15] built. Text me. You know, tag me. Like I [70:18] will reply and respond and engage with [70:20] you and so will everyone else because we [70:21] that the magic of Silicon Valley is that [70:24] it's a very positive sum mindset. You [70:26] know, it's sort of everybody is building [70:27] on each other and vulnerability is [70:30] really rewarded. Um so, I definitely [70:31] would encourage people to use the [70:32] models. [70:33] >> What's like a good heuristic if you were [70:36] doing this enough? Is it like build [70:38] something once a month? Is it sit uh [70:40] like some number of hours per day [70:42] sitting, talking, building? Anything [70:44] that you think might help people be [70:45] like, "Okay, you're doing a good job." [70:46] >> I mean, just just ship something once a [70:48] week. And it doesn't have to be crazy. [70:49] You know, I mean, for example, um when I [70:52] was playing around with Codex, I had it [70:54] build a slide deck for um Mother's Day [70:57] for my wife that pulled from my text [71:00] messages. It looked at my photo gallery. [71:03] Um it set some music to it. Created like [71:05] a 20, you know, slide deck of our [71:08] relationship and pulled some cool old [71:09] texts from when I first asked her out. [71:11] It was a really nice Mother's Day, you [71:13] know? I mean, it wasn't important. It [71:14] wasn't something that I It might come [71:15] back to, but it was shipping something. [71:17] So, it can be that small. [71:18] >> I know that you're a our mutual friend [71:20] Nikil Singh Goel. [71:21] Uh you worked at Credit Karma with him. [71:24] He had a really good way of thinking [71:25] about this. He finds that [71:27] people sh- [71:29] flip on AI and how they feel about it [71:31] once they find some moment of joy [71:33] that it had created for them. And this [71:35] Mother's Day idea is such a good [71:37] example. [71:38] And so, I think that's kind of a tip I [71:39] always think about is just like what's [71:40] something that just would will bring you [71:42] joy if this works. [71:43] >> What's something you can do for someone [71:45] else? You know, maybe that's a good [71:46] starting point as well. I love that. [71:48] >> Uh Anish, before we get to our very [71:50] exciting lightning round, is there [71:51] anything else that you want to share, [71:54] anything else you want to [71:55] double down on before we get into the [71:57] lightning round? [71:58] >> I don't think so. I've loved the [72:00] conversation so far. [72:01] >> Me too. And with that, we reached our [72:03] very exciting lightning round. I've got [72:04] five questions for you. Are you ready? [72:06] >> I am. [72:06] >> Okay. Here we go. Uh what are two or [72:08] three books that you find yourself [72:10] recommending most to other people? [72:12] >> Yeah, so okay. Conquests and Cultures is [72:15] my very favorite book. Um it's Thomas [72:17] Sowell, and it just talks about how um [72:20] conquests have led to culture change in [72:23] different societies around the world. [72:25] Sometimes positive, sometimes negative. [72:27] I think to me, it's just the best [72:28] historic view of sort of culture as the [72:31] biggest driver of outcomes. Um and I've [72:33] experienced a lot of that as, you know, [72:35] somebody who was born in Canada and [72:36] lived here to America, which has a very [72:38] different culture of ambition. There's [72:39] that word again. Um [72:41] So, Conquests and Cultures is great. [72:43] Seven Powers, you mentioned, um is [72:44] actually just an awesome book. It's I [72:46] think it's a very intellectual [72:48] distillation about kind of, you know, [72:50] moats in business theory and compounding [72:52] advantages. I really like it. You know, [72:54] the maybe the third is the one that Mark [72:56] I I actually thought that he was um [72:58] maybe punking me when he sent this to me [73:00] before I started. I said, "Mark, are [73:01] there any books that I should read?" and [73:02] he sent me a couple. And one was a book, [73:04] a textbook, called um Increasing Returns [73:07] to Scale. [73:09] Which I think is Brian Arthur is the [73:10] author. It's It's an awesome book. It's [73:12] It's a you know, it's a slightly dense [73:14] study of why things like software have [73:18] such outlier economic effects, but it [73:20] really helps put things in perspective [73:21] or it helped me in terms of like why [73:23] does our industry work in the way that [73:25] it does? Why does our culture work in [73:26] the way that it does, right? Why are we [73:27] so positive sum? [73:29] I'd love to say that we're better [73:30] people, but I think it may be a sort of [73:31] better system and structure we work [73:33] within. [73:33] >> Favorite recent movie or TV show you [73:35] have really enjoyed? [73:36] >> Oh, man. I watch trashy movies and TV. I [73:39] mean, we we watched House of Dragon. [73:42] Um which is pretty cool. Yeah. [73:44] >> Yeah, I don't know. Yeah, maybe it's [73:45] it's it's not highbrow. It's not [73:47] important with the capital I, but it was [73:48] awesome. We loved that. Um [73:51] and then I uh I did see Odyssey. Um I [73:55] saw it in London. I was there for a [73:56] board meeting in in an IMAX theater [73:59] packed with people drinking pints and [74:01] having fun and amazing. It was cool cuz [74:03] it's just a whole The movie was great, [74:04] but it was just like a theater [74:05] experience, like a weird social [74:07] experience, kind of alone together. [74:09] Um those are two recent ones. [74:11] >> Uh I have I have not been able to get [74:13] tickets to the Odyssey at at an IMAX. [74:16] So, I actually have a Grok bot just [74:17] watching the site constantly and finding [74:20] me good seats. [74:21] >> Perfect. [74:22] >> Perfect. The they've got them dumb [74:23] capture though on the AMC website that [74:25] hasn't been able to get through. [74:27] >> Really? [74:28] >> Uh it's like a tricky one. It's a really [74:29] tricky one. [74:30] >> Is it the like select the fruit? Anyway. [74:32] >> Yeah, it's like you have to click three [74:34] different matching shapes, which like [74:36] come on, you can't do that. [74:39] Hopefully by the time this comes out [74:40] I've seen it. I think [74:42] there's few podcasts that are on like I [74:43] haven't seen it yet. [74:46] Oh, man. Okay. [74:47] uh [74:48] favorite new AI product. I don't know. [74:50] Favorite AI product recently that you've [74:52] that give me joy. [74:54] >> Oh, man. I've I've spent a lot of time [74:55] with the personal agents. I think Grok's [74:57] okay. I'm going to give it to Grok bots [74:59] only cuz it's just so unhinged for it to [75:02] be so ambitious about sort of caching [75:04] credentials and getting work done on [75:06] your behalf. Like, I love it and I [75:08] expect it from a startup. Um but they [75:10] actually are doing things that I think [75:12] no other sort of big company would do. [75:15] It's a really, really well done. It's a [75:16] really thoughtful UI. It's got a really [75:18] powerful foundation model. I think it [75:19] also is like, okay, wait, maybe this is [75:21] not a two-horse race on the sort of [75:22] model side. [75:24] So, I I just think the product is like [75:26] fun and ambitious and uh and is sort of [75:29] taking risks that other products like [75:30] that wouldn't take. [75:31] >> Two more questions. Do you have a [75:32] favorite life motto that you often come [75:35] back to in work or in life? [75:36] >> Oh, man. I've got one. It's [75:39] I I learned this or I sort of, you know, [75:40] this is from my founder days, but it [75:42] also is something that's very true of [75:43] parenting. [75:45] Which is um, you know, don't discover [75:47] things through painful experience that [75:49] some somebody can just tell you. [75:51] Um so, and I unfortunately have had a [75:53] bad habit of sort of discovery versus [75:55] learning from somebody who's just a few [75:56] steps ahead of me and I find my children [75:58] have that habit, too. [75:59] >> What's one example of something that you [76:01] wish someone had told you? [76:02] >> I mean, for my kids, it's don't touch [76:04] the hot stove. Um for my startup, it was [76:07] like [76:08] >> Literally. For my startup, it was don't [76:10] build a a product and a platform at the [76:11] same time. You know, if you're going to [76:13] be a platform company, build one. Our [76:14] first company we tried to build a a sort [76:15] of a social platform for mobile games [76:17] and be a gaming studio. [76:19] And you know, somebody wise told me [76:20] right away, like, look, you can't being [76:22] a studio is so hard, much less being a [76:24] studio and a platform. Pick one. And and [76:26] we did and and it was it took us years [76:28] to figure out we were wrong. [76:30] >> Final question. I asked Ben Horowitz [76:32] what to ask you about and he just said, [76:33] uh ask him about his DJing. He's very [76:36] good. And I went to your I found your DJ [76:38] site uh I don't know if that's what you [76:40] call it on SoundCloud/illscience. [76:42] Uh it's very good. I'm just like [76:44] listening to it while I work. Um [76:47] any tips for somebody that wants to get [76:50] into DJing, any tools you found useful, [76:52] any I don't know, insights that might [76:54] help someone become better at this? [76:56] >> Totally. I mean, I think that this is [76:58] why I love music models so much. So I [77:00] DJing has always I I mean, I love DJing. [77:02] I've been playing for 30 years now since [77:04] '95, actually 31. Wow. Um [77:07] and it's an awesome way to kind of [77:08] express yourself musically if you're not [77:10] a classically trained musician. You [77:11] know, you select the music, you pick the [77:13] records, you mix them together, so it [77:15] requires some technical skill. But now I [77:17] think you can go a step further and just [77:18] make music with the models. And the best [77:20] part is when you can come up with music [77:22] ideas and have the models do the kind of [77:24] strong form of them. So I think music is [77:26] just such a visceral, satisfying way to [77:29] kind of, you know, experience and and, [77:31] you know, provide experiences in the [77:32] world. So whether it's DJing or making [77:34] music, I just suggest everyone do it. [77:37] >> I hadn't thought about how DJing has [77:38] changed now that you have Suno and [77:40] things like 11 Labs and all these things [77:41] where you could just generate music, not [77:43] have to just splice together existing [77:45] music. [77:46] >> I mean, think of the history of it. You [77:47] know, you went from Okay, you know, [77:49] first you could only hear music if you [77:51] were there with the person playing it on [77:52] an instrument. You know, recorded music [77:54] on the phonograph. The big change in [77:56] music actually, from a medium [77:58] perspective, was um the cassette tape. [78:00] Cuz the cassette tape was really the [78:01] first time you could create, right? You [78:04] could actually compose your own {quote} [78:06] album. I think a lot of why music [78:08] struggled in the 2000s is the sort of [78:10] that went away and we went back to [78:11] broadcast, and now that people are [78:13] making music again, I think the music [78:15] industry is going to be bigger than it's [78:16] ever been. [78:17] >> Wow. Man, so much disruption. [78:21] Anish, this was incredible. We covered [78:23] so much ground. Uh final question, how [78:25] can listeners be useful to you? [78:27] >> I mean, show me what you're building. [78:29] Um please don't be despondent. Build [78:30] something and then tag me, um and I'd [78:32] love to see it. [78:34] Um if you're interested in hearing more [78:35] stuff like this, please follow me on X. [78:37] I try to kind of engage and follow back. [78:39] Um and otherwise, [78:41] make sure you check out all the amazing [78:42] folks in Lenny's network. Claire is a [78:45] star. Elena is so, so good. And there's [78:47] just so much compelling content here. [78:49] >> Nikkel, which we talked about. [78:50] >> Nikkel is the best, man. That guy is is [78:53] everything. [78:54] >> And amazing. And Nish, thank you so much [78:57] for being here. [78:58] >> Thank you, Lenny. [78:59] >> Thank everyone. [79:01] Thank you so much for listening. If you [79:03] found this valuable, you can subscribe [79:04] [music] to the show on Apple Podcasts, [79:06] Spotify, or your favorite podcast app. [79:09] Also, please consider giving us a rating [79:11] or leaving a review, [music] as that [79:12] really helps other listeners find the [79:14] podcast. You can find all past episodes [79:17] or learn [music] more about the show at [79:18] Lennyspodcast.com. [79:20] See you in the next episode.