Cognition Targets Enterprise Code Legacy as Devin Revenue Hits $73 Million

Devin营收达7300万美元，Cognition瞄准企业遗留代码 该初创公司计划完全聚焦于自主工程任务而非通用聊天智能体，以保持模型中立的独立业务模式。 本次对话探讨了驱动Scott Wu的极强竞争心态，以及Cognition旗下AI软件工程师Devin在企业端的迅速普及；Devin正在为Goldman Sachs和Mercedes-Benz等大型企业带来高达10倍的产出提升。Wu讨论了如何将软件开发从手动编码转变为一种“创造模式”，让用户能够通过模型中立的复合系统部署自主智能体集群来执行长期任务。尽管面临科技巨头的强烈质疑以及在巴西早期部署时的阻碍，Cognition依然坚持保持独立，致力于打造一家跨时代的企业。最后，Wu强调自己的动力源于对技术潜力的第一性原理热情，而非追求财务层面的收购变现。 Scott Wu 将创业与人生决策视作一种算法级的“树搜索”（Tree Search），凭借极度厌恶失败的好胜心，致力于将 Cognition 打造为具有代际影响力的超级企业。 (0:02) Cognition 对 Devin 的终极愿景是作为彻底抽象底层代码的人机界面，推动人类从“生存模式”跃迁至“创造模式”，目前已协助高盛、梅赛德斯等大客户实现十倍的工程交付效率。 (7:30) 从第一性原理来看，AI 智能体正在打破人类对指数级增长的线性认知，能力边界正从单次短任务扩展到能自主独立运行数月乃至数年以解决宏大课题。 (22:57) Devon 在 2024 年初的原型发布虽引发了巨大争议，但团队锚定“自主协作者”定位并通过企业级代码迁移等痛点切入，在 20 个月内实现营收从零跃升至约 5 亿美元。 (30:18) Cognition 采取“瑞士式”的模型中立策略，通过复合模型架构动态调度各大厂商模型以追求最高 ROI，并凭借对复杂工程落地细节的极致专注与巨头拉开差距。 (50:01)

Cognition co-founder Scott Wu argues that autonomous AI software engineers like Devin represent a paradigm shift in human-computer interaction, moving beyond coding assistants to long-horizon, autonomous systems that abstract programming entirely. Drawing on an intensely competitive, algorithmic worldview, Wu asserts that focused startups can outmaneuver tech giants by solving the messy, complex operational realities of enterprise software. Algorithmic Strategy Grounded in Extreme Competitiveness (0:02) Wu frames company building as an algorithmic "tree search," a mental model rooted in an intense, lifelong aversion to losing. His upbringing centered on competitive math, programming contests, and high-level gaming such as Super Smash Bros. Melee and Tetris. This competitive environment was family-wide: his father was a 7-dan Go player whose matches with a university professor directly led to U.S. visa sponsorship from China, while his mother lined their home with academic trophies instead of family photos to reinforce an expectation of excellence. Cognition’s nine-person founding team, largely composed of repeat founders, applies this calculated mindset to company strategy. Accepting that significant losses and public setbacks are inevitable milestones when attempting breakthroughs, Wu views their trajectory through systematic counter-moves rather than emotional reactions. Elevating Software from Syntax to Autonomous Missions (7:30) Wu describes computing history as a continuous ascent up the ladder of abstraction, evolving from manual vacuum tubes and punch cards to assembly, high-level languages, and now natural language. In this framework, Devin is not designed as a mere code-completion tool, but as a primary interface that converts human intent into computer execution. Using a Minecraft analogy, Wu suggests this evolution shifts humanity from "survival mode"—spending baseline effort managing mundane tasks—into "creative mode," where resources are abundant and the core human responsibility is deciding what to create. This abstraction transforms the economics of engineering by enabling single-use software. Historically, writing custom software was only viable if reused millions of times due to expensive labor costs. With AI agents capable of operating over hours, days, or year-long missions, Wu predicts individuals and organizations will delegate complex, bespoke research and multi-step tasks to autonomous agents on demand. Navigating Early Failure to Secure Enterprise Traction (30:18) Devon’s launch in early 2024 sparked intense polarization, with critics calling it a scam and enthusiasts hailing it as revolutionary. At launch, the prototype resolved only 13% of tasks on the SWE-bench benchmark, meaning it failed 87% of the time and broke down on dense corporate codebases. Rather than attempting open-ended system design immediately, Cognition targeted structured, highly repetitive engineering problems such as massive version migrations across tens of thousands of legacy files. To establish product-market fit, the entire founding team relocated temporarily to Brazil to work side-by-side with engineers at Nubank, manually debugging edge cases to refine Devin’s execution loops. This hands-on enterprise focus enabled Cognition to grow from zero to roughly $500 million in revenue within 20 months. Approximately 75% to 80% of current revenue originates from enterprise clients—including Goldman Sachs, Mercedes, and the U.S. government—by compressing typical 12-to-18-month sales cycles to under three months through private cloud deployments and strict airwalls. The Switzerland Model and Start-Up Focus To preserve enterprise trust and maintain efficiency, Cognition positions itself as a model-neutral "Switzerland" in AI. Rather than hardcoding its agent to a specific provider, Devin operates as a compound system that dynamically orchestrates models across Anthropic, OpenAI, Google, open-source alternatives, and internal weights. High-reasoning models tackle initial problem diagnosis and ticket reproduction, while faster, cheaper models handle repetitive boilerplate code. Wu dismisses the assumption that hyperscalers will inevitably eliminate independent AI applications, comparing Cognition to Spotify navigating competition from Apple Music, as well as standalone infrastructure leaders like Datadog and Snowflake. He contends that massive tech conglomerates suffer from corporate inertia, whereas specialized startups win by caring far more about the unglamorous nuances of software engineering—such as ticket integrations, local testing, and legacy codebase maintenance. Cognition rejects acquisition interest from tech incumbents in pursuit of building a generational, independent enterprise. Wu forecasts at least ten subsequent product generations beyond the current pull-request workflow, arguing that as long as human lives can be improved through software, the opportunity space for autonomous engineering remains practically infinite.

Cognition 联合创始人兼首席执行官 Scott Wu 认为，人工智能正在将软件编程的抽象层级推至只需人类表达意图的阶段，推动人类从“生存模式”转向资源与工具无界的“创造模式”。他主张初创企业必须依靠对真实复杂工程细节的极致专注、复合模型路由的中立架构，以及对独立自主的坚持，在科技巨头的围剿中打造新一代平台型业务。 树搜索思维与竞技本能构筑的创业底色 (0:02) Wu 将创业决策比作计算机科学中的“树搜索”（Tree Search），即通过推演行动分支、预测后续事件并规划反制措施来寻找制胜路径。这一思维源自其童年对失败极度不耐受（“Salty”）的竞技经历。他在小学二年级（七八岁）参加初中数学竞赛未获奖后便确立了争胜心态，并在与年长兄长竞争中沉浸于数学竞赛、编程以及《任天堂明星大乱斗》、俄罗斯方块、扑克等高水平游戏。 家庭环境进一步强化了这种专注。其父亲是围棋业余 7 段选手（水平相当于国际象棋 2300 至 2400 ELO），当年正因在中国读研期间与教授下围棋而获得赴美签证担保；其母亲则极度强调卓越与胜利，甚至家中墙面完全不挂家庭照片，取而代之的是摆满兄弟两人的数学和编程奖杯。这种对获胜的执念演化为 Wu “输掉的痛苦远甚于获胜的喜悦”的创业心理基调。 软件抽象演进与创造模式的远期愿景 (7:30) Wu 指出，编程史就是一部抽象层级不断提高的历史，从早期真空管、穿孔卡片、汇编语言，演进到高级语言与现今的 AI 代理。传统软件经济学仅在代码被调用成千上万次或数百万次时才能平衡高昂的人工开发成本；而 AI 代理的出现使得“单次使用软件”（Single-Use Software）在经济上变得可行，例如仅为处理特定表单或定制数据分析而自动生成并执行临时脚本。 吴借用《我的世界》（Minecraft）的隐喻提出，人类历史此前始终处于防范风险、获取资源的“生存模式”，而 Devin 这类自主软件工程师的终极目标是将人类转入“创造模式”——资源与实现能力被无限放大，人类只需明确自身想创造什么。目前 AI 的自主执行能力已从数年前的十到二十秒扩展至数小时，他认为从第一性原理推导，这一窗口有望在未来拓展至数天、数周乃至类似宝丽来创始人 Edwin Land 式的数年期科研使命。 从争议性原型到企业级落地的工程迭代 (30:18) Cognition 由九人创始团队（多数拥有过往独立创业经历）建立，在 2023 年押注多步骤循环自主系统，并在 20 个月内实现约 5 亿美元的营收规模。Devin 在 2024 年 3 月发布演示原型时引发了两极化反响：其在 SWE-bench 基准测试中取得 13% 的成功率（竞品普遍为 3% 至 4%），但这也意味着该系统在早期有 87% 的概率遭遇失败，且在面对庞大且复杂的企业真实代码库时频繁崩溃。 团队随即将产品契合点（PMF）转向高度重复且繁重的工程任务（如跨 50,000 个文件的 Java 版本升级与代码迁移）。在拓展首个重要大客户——巴西数字银行 Nubank 时，Cognition 团队全员飞往巴西与客户工程师并肩排查边界情况与工作流，以此打磨 Devin 的自动化编排闭环。目前，企业业务贡献了该公司 75% 至 80% 的营收（涵盖高盛、梅赛德斯、CVS 和沃尔玛等），并依托私有云部署与严苛数据隔离墙，将传统银行 12 至 18 个月的采购部署周期压缩至 3 个月以内。 瑞士中立架构与垂直专注的防御壁垒 在与底层基础大模型实验室的竞合关系上，Wu 将 Cognition 定位为 AI 领域的“瑞士”（中立者）。Devin 采用复合模型系统（Compound Model System），并不绑定于单一模型供应商，而是在处理工单时根据任务类型动态调用 OpenAI、Anthropic、Google、开源或其专有模型：由深度推理模型解决复杂的架构分析，低成本快速模型执行样板代码与重复任务。 针对微软拥有 GitHub Copilot 及生态的质疑，Wu 援引 Spotify 首席执行官 Daniel Ek 抵御苹果和 YouTube 竞争的先例指出，初创公司的核心壁垒在于对软件工程真实复杂度的绝对专注。科技巨头往往视算法为整洁的沙盒问题，而 Cognition 则深入应对遗留代码库、企业工单对接、本地构建测试等杂乱细节；同时，中立地位使其激励机制与客户的投入产出比（ROI）高度一致，而非追求最大化底层算力或 Token 消耗。 Wu 强调 Cognition 拒绝了多次收购提约，坚持保持独立，并预测 Devin 现有的 PR 审查交互绝非终态，未来软件开发还将经历至少十代产品形态的跃迁。不过，对话也表明，将 AI 代理的连续可靠工作时间从现有的数小时提升至数周乃至整年，仍高度依赖底层推理突破与对企业杂乱现实工程环境的持续解决，其自主性承诺在更大规模的生产实践中仍处于快速演变阶段。

[2:02] Scott Wu: it's all I do honestly. [7:30] Scott Wu: losing feels way worse than winning feels good. Scott Wu: in AI terms 5 years is is like a century Scott Wu: we're going to have forgotten that we ever lived without them. The single most powerful pattern I have noticed is that successful people find value in unexpected places. And they do this by thinking about business from first principles instead of formulas. Daniel Ek: the truth is we're just going to care way more about music than they are. David Senra: Why don't we call it Scott? Scott Wu: Let's be real here okay, agents did not work generally, [57:14] David Senra: those people aren't founders. Founders are rationally optimistic. [57:14] David Senra: They believe even there's no evidence that they should succeed, that they will succeed. Richard Branson: all business is an idea that makes somebody else's life better.

[0:02] I want to know why you describe yourself [0:04] as salty. What does that mean? I've just [0:06] always been this way. As a kid, I I just [0:09] hated losing. Like my first competitive [0:12] memory ever is like when I was in second [0:14] grade. I went to this seventh grade math [0:16] competition. It was like a middle school [0:18] competition that was held at the like [0:20] local university or whatever for middle [0:21] schoolers. But you were seven. [0:23] >> Yeah. Yeah. I was like seven or eight [0:24] years old. I was competing in the like [0:26] middle school math and like I did the [0:27] like math test and whatever and then [0:30] they were calling out the names of the [0:31] like here's who got third place here's [0:33] got second and I was kind of like [0:35] waiting for my name to get called and [0:36] then I was none of them and I just [0:38] remember being so pissed about that. [0:40] >> Yeah. I can't really give you a rational [0:42] explanation for for why it is. I [0:45] >> It doesn't have to be rational but like [0:47] how much of your brain is dedicated to [0:48] competition? [0:49] >> I mean it's all I do honestly. I don't [0:52] know. I think the like [0:54] >> Oh, what do you mean it's all you do? [0:55] >> Well, I think strategy game I I don't [0:57] know. It's the the way even building a [0:58] company it feels the same. It's just [1:00] like you're calculating the moves. [1:01] You're thinking about okay, if you do [1:02] this and then this happens and then you [1:04] do that and here are the different moves [1:06] and you're like calculating out what [1:08] comes out to success. You know, it's [1:10] like a it's like a tree search, you [1:11] know, where you're exploring the [1:12] different options in the decision tree [1:14] and you're trying to figure out how to [1:15] lead to victory. Like that's like the [1:16] only thing I do in my life. And [1:17] >> so this is basically just you don't have [1:19] memories when you weren't like this [1:20] basically. [1:21] >> I think that's right. Yeah. Yeah. I was [1:23] uh I was a little brother growing up and [1:25] so my older brother was four or five [1:26] years older than me and naturally we'd [1:28] play video games and similarly I would [1:30] just always be super salty there as [1:31] well. I don't know. It's just like yeah [1:33] it's just always like that. So like I [1:35] spent some time with Demis from Dind and [1:38] what was interesting is I I I draw a lot [1:40] of um like similarities between you two [1:42] cuz I also spent some time with you [1:44] >> and I was like well they're both really [1:46] smart, they're both articulate, they [1:47] have like a friendly UI, right? But then [1:51] underneath that is like this like [1:52] ruthlessly competitive drive and Demis I [1:55] think he said this publicly but I think [1:57] he said like half his brain is dedicated [2:00] to competition and a lot of that comes [2:02] from his early days in chess. [2:04] >> Yeah. [2:04] >> What were you competing in when you were [2:05] younger besides math competitions? [2:06] >> Yeah. Well basically everything. So [2:08] obviously the main thing was math and [2:09] programming competitions and so ever [2:10] since I was really young that was like [2:12] that was like my like life you know my [2:14] whole goal was to become like world [2:16] champion of competitive programming. I [2:17] would do that all the time as a kid. I [2:19] would do, you know, the the really great [2:21] thing about these competitions, too, is [2:23] is, [2:24] you know, you you compete in your school [2:26] competition and if you do well enough in [2:28] that, then you make it qualify for the [2:29] the local or like the city um [2:32] competition and then if you do well in [2:34] that, then you get to like the regional [2:35] competition and the state competition [2:36] and then you get to go to the national [2:37] thing, the international thing, right? [2:39] And so it's it's like a very nice setup [2:42] where sooner or later you get to kind of [2:45] meet people who are like you. Basically, [2:47] when I was a kid, doing these [2:48] competitions, going for that, it was [2:50] like all I really cared about. Those [2:52] people that I met through these national [2:54] international competitions were honestly [2:56] more like they were my childhood friends [2:58] more than like the people around me in [3:00] Baton Rouge, Louisiana were. And it was [3:01] like we would hang out online, we would [3:03] talk about math, we talk about problems, [3:04] but all the other things too. I mean, I [3:06] played basically all the different [3:07] competitive games. So, like I I played a [3:09] lot of Super Smash Brothers. I used to [3:10] go to tournaments for Super Smash [3:12] Brothers. It was a lot of fun. Played [3:13] Melee. And then I played like Tetris. I [3:16] played a lot of poker. I played some [3:17] chess. I was okay at chess. I was not [3:19] good. I played some Go. My dad was a [3:21] competitive Go player. My parents came [3:23] to the US in some sense because of Go, [3:26] which was kind of a funny coincidence [3:28] because my dad um was in grad school in [3:31] China and he had a professor who like [3:33] really liked him. The reason he liked [3:34] him is cuz my dad was like a really good [3:36] Go player. Like he was like a seven dawn [3:38] at Go, which if you were to call it in [3:40] chess would be like I don't know 2,300 [3:42] or 2,400 rating equivalent or something [3:44] like that. he would play with this [3:45] professor, you know, on the weekends and [3:47] stuff and they like, you know, my dad [3:48] would generally win and they would like [3:50] talk about the games and stuff. And then [3:51] that professor ended up moving to the US [3:54] to come and teach. Um, and at the time, [3:57] uh, you know, this was super early on [3:59] in, you know, immigration from China to [4:01] the US. And so it was not a very like it [4:03] wasn't really a path that people knew [4:04] that you could take. The professor wrote [4:06] my dad and said, "Hey, like I came. It's [4:08] great. Like there's so much more [4:09] opportunity. It's so much better. Like [4:10] you should obviously come as well. like [4:12] I'll help you with your like visa [4:14] application. I'll help you like apply to [4:15] colleges here and everything. And so my [4:17] dad applied to grad school uh in the US [4:19] and that's kind of how we ended up here [4:21] in the first place. Uh I was I was born [4:22] after we moved to the US obviously. [4:24] >> So your dad was competitive in go. What [4:25] was your mom competitive in though? Cuz [4:27] I think I read that you said that she [4:28] might have been the most competitive [4:29] person in your family. [4:30] >> Yeah. No, she was always uh she she was [4:33] definitely the most salty I would say [4:34] for sure. I mean she would um [4:36] >> What does salty mean? [4:39] Salty just means that you take offense [4:41] to the idea of losing. [4:43] >> Okay, I love that. [4:44] >> Yeah, she would always be, "Oh, no, no, [4:46] I'm better at this or I'm, you know, I [4:49] can beat you at this, you know, and and [4:50] I don't think she I mean, she played [4:51] ping pong a bunch growing up. Actually, [4:53] she played on her like school ping- pong [4:55] team. Obviously, she studied some amount [4:56] of math and so on. Um, but but it was [4:59] just it was more her personality than [5:01] than any one thing that she really put [5:03] all of her competitive energy into. I [5:05] spent a lot of time obviously reading [5:06] the biographies of history's greatest [5:08] entrepreneurs. [5:08] >> Yeah. [5:09] >> Always fascinated by like there's [5:10] usually two different kind of archetypes [5:12] for the parents. One you have like the [5:14] Larry Ellison and Elon Musk uh their [5:16] dads would literally tell them you know [5:17] you're worthless. There's stories in [5:19] Elon's biographies where his dad just [5:20] gets in his [ __ ] face and yells at [5:22] him for hours. Larry's adopted uh father [5:24] would just tell him you're never going [5:25] to amount to anything. And so that they [5:27] had this like inner fire to to to [5:30] disprove uh you know saying that [5:32] basically no [ __ ] you dad you're wrong [5:34] about this right [5:35] >> and then you have like the SA lers who [5:37] you know their uncle or even their [5:38] father is just like you're really [5:40] special you have a lot of talents if you [5:42] put a lot of effort into this you can do [5:44] whatever you want your mom falls it more [5:46] into like the Estee Lauder like category [5:48] where she would tell you that like hey [5:49] these people are doing amazing things [5:50] you could do even better than them [5:52] correct [5:53] >> I think they would have been happy [5:54] enough if I just got like a more [5:56] traditional cushy job and did all of [5:58] that. Like I don't know that they [5:59] specifically steered me towards [6:01] entrepreneurship and being a founder. [6:03] But um but no, they were always [6:05] >> but your mom gave you self-conidence. [6:06] >> Yeah, I think she she always told me [6:10] that I was the best [6:12] and she was always extremely proud of [6:14] that, you know, it's it's the I had [6:16] these like when we would go to these [6:18] math competition. [6:18] >> Hold on. She so you said she always told [6:21] you that you're the best. Did she say [6:22] that before there was evidence? [6:25] >> Yeah, I think so. I think that's right. [6:26] I I think even when I was like tiny, she [6:30] would tell me that I was extremely [6:31] talented, extreme, you know, she was [6:33] always a huge source of support me and [6:35] she always obviously believed in [6:36] whatever I wanted to do, you know, cuz [6:37] you would compete in math competition, [6:38] you would get these trophies and stuff [6:40] and like we didn't have like, you know, [6:43] growing up like we didn't have like [6:44] pictures of our parents like on the [6:46] walls, you know, we just had old math [6:47] competition trophies. It was like like [6:50] my mom was very intentional about like [6:51] no no the thing that we value in this [6:53] household [6:54] >> other people's trophies are the ones you [6:56] want. [6:56] >> Ours ones that me and my brother want of [6:58] course [6:59] >> we're going to put up pictures of other [7:00] people's trophies and you better damn [7:01] sure replace them with your own. [7:04] you know, so like you know when when I [7:06] was pretty and my brother as well, you [7:08] know, we both really like these [7:09] competitions and so like you know like [7:12] accumulate these trophies and she would [7:14] always like every time we got one she [7:15] would hang it up and we're like put it [7:17] up on the on the on like the mantel [7:18] piece and everything and and no it's [7:20] like it was always um I think what she [7:22] really valued. I think I think education [7:24] was really important to her and I think [7:26] being the best was was very important to [7:28] her as well. [7:30] >> I want to tell you about the presenting [7:31] sponsor of this podcast, RAMP. I have [7:33] been reading a lot about SpaceX lately. [7:35] SpaceX is one of the most valuable [7:36] private businesses in the world. And one [7:38] of the main themes in the history of [7:39] SpaceX is constantly attacking and [7:42] questioning your cost. Ramp helps many [7:44] of the most innovative businesses in the [7:46] world do exactly that. The median [7:48] company running on RAMP cuts their [7:50] expenses by 5%. And one thing SpaceX has [7:53] demonstrated is that a religious [7:54] dedication to controlling costs can help [7:57] actually increase revenue because you [7:59] can pursue opportunities you couldn't [8:01] otherwise. And we see that in the ramp [8:03] data too. The median company running on [8:05] ramp also grows their revenue by 16%. So [8:09] when you're running your business on [8:10] ramp and your competitors are not, you [8:12] have a massive competitive advantage [8:14] that compounds over time. Ramp is the [8:17] only platform designed to make your [8:18] finance team faster and happier. Many of [8:21] the top founders and CEOs I know run [8:23] their business on RAMP. I run my [8:24] business on RAMP and you should too. Go [8:26] to ramp.com to learn how they can help [8:28] your business save time, save money, and [8:31] grow revenue. That is ramp.com. [8:35] What matters to you more? Like what is [8:38] uh is it the the pain of losing is like [8:41] losing is worse than the love like the [8:43] thrill of winning [8:45] >> cuz that's how you started it. You're [8:47] just like I could not stand losing. Let [8:49] me back up. I was reading another Larry [8:51] Ellison biography. He says this. He's [8:52] just like, "Listen, I'm addicted to [8:54] winning, but I fear losing more than I [8:56] love winning." And I actually talked to [8:58] Michael Dell about it. And he's like, [8:59] "Yeah, it's it's just the fear of losing [9:01] is the the pain of losing is way worse [9:02] than the the the good feeling of [9:04] winning." [9:05] >> So, I think that's definitely true in [9:06] terms of how it feels in order to get [9:08] anywhere. You got to lose a lot. If [9:10] anything, almost like a lot of the, you [9:12] know, I mean, a lot of the guys you're [9:13] talking about, if anything, like the way [9:15] they got to where they are is by by [9:17] losing a lot and having their share of [9:19] wins along with that, but like you just [9:20] have to put yourself out there and do a [9:21] lot. So, it's kind of an interesting [9:22] thing to your point of like I definitely [9:24] feel the same way that yeah, like losing [9:26] feels way worse than winning feels good, [9:28] but not by enough that it makes me want [9:31] to stop trying if that makes sense. [9:33] >> Yeah. No, I I think it's impossible. I I [9:35] know your personality type like there's [9:36] it's just impossible for you not to do [9:38] this. What is like let's get into Devon. [9:40] >> What is like winning with Devon look [9:42] like to you? [9:43] >> So, again, you know, we're hyper [9:44] competitive, but also, you know, the [9:45] other thing about us is like we all had [9:46] kind of started our own companies before [9:48] this. So our founding team is was a [9:49] pretty big founding team is nine people [9:51] and most of us had already founded our [9:53] own companies before we had done [9:54] different things. It was true for a lot [9:55] of the early team and we've always [9:57] thought about this as like this is the [9:59] big one and so like we want to go for [10:00] all you know we want to be a [10:02] generational business like we want to [10:04] build a hyperscaler and we want to go [10:05] and do that and like maybe we'll succeed [10:08] maybe we won't I don't know but like [10:09] that's what we're going after and and I [10:11] think to me what that means in our field [10:14] um in building software is like you know [10:16] people sometimes use the term like a [10:18] coding agent or like you know like AI [10:21] programming or something I I always like [10:23] you know I always hear that I think a [10:25] little bit about like well we're not [10:26] always I mean we're not going to be [10:27] interacting with code for that much [10:28] longer you know our programs might not [10:30] be the right like level of abstraction [10:31] but I think what is always true probably [10:34] is that it will be the human computer [10:38] interface and so what I mean by that is [10:39] like like the way that we think about [10:41] Devon if we're successful is Devon is [10:44] the way that humans can tell their [10:46] computers what to do cuz that was the [10:48] whole point of software engineering [10:49] anyway right is just to be able to work [10:51] with your computer computer and tell it [10:52] what you want it to do I mean I doing [10:54] that for the world is like a massive [10:56] opportunity and and and and that's what [10:57] we're really excited to go after. [10:59] >> Why is that interesting to you? You just [11:01] said we all did different things. We [11:02] came together and we're like, "No, this [11:03] is the big one." [11:04] >> Yeah. I mean, the simple answer is, you [11:06] know, we're a bunch of nerds who we're [11:08] all programmers and built software and [11:10] so on before and now this is the idea of [11:12] teaching AI how to go do that. [11:13] >> Wait, say say more about that what you [11:15] just said. We're all nerds. This is a [11:16] big one because we can actually teach AI [11:18] how to do. We're all nerds who've spent [11:20] the last, you know, I don't know, 20 [11:21] years of our lives just coding and [11:23] making little things for the world. And [11:24] the idea that you could teach AI how to [11:26] go make things uh and and have everybody [11:30] have an AI that can help them make [11:31] things and feel the wonder of that. I [11:33] mean, it's pretty sick. [11:34] >> Is that the idea the single idea in the [11:35] world that animates you the most that [11:36] gets you the most excited? [11:37] >> I think it is incredible. Yeah. I mean [11:40] the Well, I I would go further too and [11:42] just say like at a zoomed out level, [11:43] it's I honestly think that [11:48] Because you know everyone talks about [11:50] AGI everyone talks about well what is [11:51] the future going to be like what what [11:52] are human lives going to be you know [11:55] once AI can do everything and I mean to [11:58] some extent I think the thing that is [12:00] most human obviously is [12:02] self-expression creativity like having [12:05] things that you want to make happen in [12:07] the world and being able to go do those [12:08] things right and so I mean a lot of how [12:10] I think about this is basically [12:12] we want to build the tool that gives [12:14] everybody the power to go and make [12:16] things that they want to make in the [12:18] world. My co-founder has a slime which [12:19] I've always loved. Uh which is, you [12:21] know, we've been spending all this time [12:24] living in survival mode as a species, [12:26] you know, and now we're going to be [12:27] living in creative mode. And I think [12:29] that's right. Like I I think um you [12:31] know, Minecraft survival mode is is [12:33] where you're like um you know, you're [12:35] you're growing food, you're like making [12:37] sure you're safe from the monsters at [12:39] night and whatever. Um and like creative [12:41] mode is just like everything's up to [12:43] you. You know, you have all the [12:44] resources at your disposal. If you want [12:46] something to happen, it'll happen. And [12:47] the only question for you is like what [12:49] you want to make happen. I think it's [12:51] going to be amazing. And I mean, I think [12:52] that is like the the world that we're [12:53] going towards and and like we want to be [12:54] the ones building that. [12:55] >> Okay. So, talk about what Devon does [12:57] today. Yeah. [12:58] >> And then I I want you to flesh out that [13:00] idea of like where you see in the [13:02] future. It's just the the interface [13:03] between humans and computers. [13:04] >> Sure. Yeah. So, Devon is uh today, you [13:07] know, what [13:09] what what what folks know as as an AI [13:11] software engineer. And basically what [13:13] that means is that Devon is a tool that [13:16] that anyone can use uh that will work [13:18] with them end to end on building out [13:20] software. Um and so we work with a lot [13:22] of the big biggest companies across the [13:24] world, you know, we work with Goldman [13:25] Sachs, we work with Mercedes, we work [13:27] with, you know, a lot of areas of the US [13:29] government at this point and so on. And [13:30] we work with their software teams to [13:32] just help them build more and do much [13:33] more. And you know, the thing about it [13:35] is in the last 20 30 years obviously I [13:38] mean software is eating the world is the [13:39] is the famous line. I think it's very [13:40] much true. It's still like it's still [13:42] got a couple order of magnitudes to go [13:44] and in practice what it looks like is [13:46] that teams use Devon to ship 10 times [13:48] faster and to do 10 times more. [13:51] >> Okay. So that's where it's at today. [13:52] >> Yeah. [13:52] >> How do you get to where you're saying [13:54] you might there is no like you're just [13:56] the interface between humans and [13:57] computers. [13:57] >> So now we're going to get into a [13:58] philosophical discussion of what it [14:00] means to be a programmer or a software [14:01] engineer, right? And I I think like you [14:03] know if you go all the way back it's you [14:05] know there there was a time where [14:06] programming was like using the vacuum [14:08] tubes like plugging all those in and [14:10] having it do the machine do the [14:11] arithmetic right like the eniac was um [14:14] you know when the first in some sense [14:15] that the first computer out there [14:17] although it was obviously very different [14:18] or it would have been like you know [14:20] filling out the punch cards and like [14:21] setting up the or writing you know the [14:23] like putting down assembly you know or [14:25] writing in basic or something right so [14:27] so we've gone through a lot of [14:28] generations already is my point and what [14:30] does it look like you know going forward [14:32] when you talk about programming all it [14:34] really comes down to is like how do you [14:35] tell your computer what to do and like [14:38] every single piece of software that you [14:39] use if you're using you know Instagram [14:41] or Tik Tok or YouTube or whatever like [14:44] that's a piece of software that somebody [14:46] or like some you know in these cases [14:48] some pretty big teams of engineers came [14:50] together and thought through all these [14:52] details of okay here's what I want it to [14:53] do here's how I want it to look here's [14:55] what I want this button to do here's how [14:57] I want to architect it every single [14:59] little decision obviously was made by [15:01] somebody but the the computer itself is [15:04] then executing it accordingly to to what [15:06] the wishes of its creators was right and [15:09] I think what we'll start to see is that [15:12] abstraction will continue to climb right [15:14] and like you know we kind of see this [15:15] already like at this point you don't [15:17] need to know what programming languages [15:18] like you don't need you don't need to [15:19] know Python or Java or something like [15:21] that in order to build your own software [15:23] right and you can just say hey here's [15:25] what I want I want to make a cool [15:26] website that does this this and that or [15:28] for example you know in my existing [15:30] product you know, here's here's what we [15:32] have today and I want to change this [15:34] thing or add this new plan or add this [15:36] new feature and just have the agent go [15:38] and do that for you, right? I think [15:40] we're going to continue to go further [15:41] down that axis. One important kind of [15:43] like distinction I'd make, which is, you [15:45] know, I think what we'll see a lot more [15:47] of in the near future, is software [15:49] today, [15:51] the the math only really works out to [15:53] create software if it's going to be used [15:55] at least like a million times or [15:57] something, you know, and I'm giving a [15:58] maybe it's 10,000 times or whatever. And [16:00] my point is like if you want to go build [16:02] a product today, you need a whole team [16:03] of engineers. Engineers are expensive. [16:05] You got to pay salaries. You got to go [16:07] build all this out and you got to go and [16:08] do that, right? You need that software [16:10] to be used enough times or to be to to [16:12] create enough value for that to be worth [16:14] it, right? And you know, something like [16:15] YouTube passes that test because [16:18] obviously so many so many hours have [16:20] been spent on building YouTube, but way [16:22] more hours have been spent on using [16:23] YouTube and and that's that's what's [16:25] made that that work out and made it [16:27] feasible, right? But there's so many [16:29] things out there which only, you know, [16:31] very specific things that that only need [16:32] to be used a few times or even like only [16:34] need to be used once, right? And so like [16:36] all the white collar work that we talk [16:37] about today even is very um all right, [16:41] you know, wake up in the morning. All [16:43] right, I'm going to go look through [16:44] these like 15 LinkedIn profiles. I'm [16:46] going to look for this and that or [16:47] whatever. Or I'm going to go fill out [16:49] these forms or I'm going to do this data [16:52] analysis and put this Excel sheet [16:54] together with this research that I [16:55] found. Right? All of these things are [16:58] things that could be done with software. [17:00] It's just it obviously doesn't make [17:02] sense to to hire a whole team of people [17:04] to go make you that piece of software [17:06] which you're going to use one time and [17:07] never again versus just having the human [17:08] go and do that themselves. Right? I [17:10] think what we're going to get to is [17:12] we're going to get to a point where you [17:13] are just giving your instructions to [17:14] that agent and the agent on the back [17:16] end, you know, you don't even have to [17:18] look at this, but but on the back end [17:19] the agent is going to figure out, okay, [17:20] here's I'm going to write this code [17:21] that's going to go do this. I'm going to [17:23] put a a script that automates this part. [17:25] I'm going to do this. I'm going to do [17:26] this. Um, and that's what's going to [17:28] allow it to actually go and do all these [17:30] things. But but but what you start to [17:32] get to, you know, as we're kind of [17:33] saying, it's like this is really just [17:35] how you control your computer and how [17:36] how you do what you want to go do, [17:38] right? And you wake up in the morning [17:39] and it's like here here's what I want to [17:40] go do. You talk to your agent about it. [17:42] You figure out the task together once it [17:44] has it. It can do the part of of the [17:47] literal like, all right, put the put the [17:49] pen to the paper on writing code, but [17:50] that task is like, you know, you're [17:52] you're the one that's deciding what to [17:54] do. So this ideal future that's in your [17:56] mind, right? How far away do you think [17:59] we are to that? [18:01] >> Yeah. Um I mean we've made a lot of [18:02] steps toward it. I would say we still [18:04] got a ways to go. You can use Devon [18:06] today. You can use all the different [18:08] kind of coding tools today. Um and you [18:10] can do a lot more than you could have [18:11] done, you know, 10 years ago. Um but [18:14] certainly um or even one year ago or 6 [18:16] months ago. Um but but but certainly you [18:18] know it's it's it's not at the level [18:20] that we're talking about of you are [18:21] neural linked into the AI and you can [18:23] tell it exactly what you want it to do [18:25] and what you want to see in the world [18:26] and just have the AI go and do that. [18:28] When do we get there? Um it's it's hard [18:31] to say but I honestly I mean I think [18:32] we'll have solved most of that over the [18:34] next 5 years or so. In AI terms 5 years [18:36] is is like a century you know in in in [18:39] the rest of the world terms. Obviously [18:40] it's like it's kind of crazy to imagine [18:42] that things can change that much in 5 [18:43] years but but I really think it will. So [18:45] this is kind of related to something I [18:46] heard you say where you're saying that [18:47] humans just have a really hard time [18:49] understanding exponential curves. [18:50] >> So true. I mean you see this in in in [18:53] the progress itself. You see this in the [18:55] scaling laws with the data. You see this [18:57] in the revenue curves of the companies [18:58] that are building an AI [19:00] >> including your own. And um and it's it's [19:04] a very, you know, it's like humans [19:05] aren't really wired for this, right? [19:07] Like all of our like in inherent like [19:09] fight orflight response are our kind of [19:11] like our ability to kind of like measure [19:13] things to vastly oversimplify if if [19:15] you're just, you know, fighting out [19:17] there, you know, foraging for food or [19:19] whatever it is like, you know, a good [19:21] hunt will bring you, you know, a couple [19:23] days worth of food or something. But but [19:25] but but but obviously with the kind of [19:27] exponential curves that we deal with, [19:29] you know, the equivalent of a good hunt [19:30] here could be a thousand years worth of [19:32] food and and we don't have that [19:34] intuitive signal in our brains to really [19:35] understand that, right, at a really deep [19:38] like native level. People often [19:39] underestimate, I think, how fast things [19:41] can change and how fast the world can [19:42] change. I mean, my parents even like [19:44] drilled this into me because they grew [19:45] up in, you know, in communist China. Um, [19:47] and they came to the US and you know, [19:50] even that like we're talking about what [19:52] are what were ultimately even like much [19:54] slower scales of progress in some sense [19:57] relative to I think what we're seeing [19:58] today. But even that was like, as you [20:01] can imagine, it was incredibly jarring [20:04] to them. you know, the idea that uh like [20:07] you come to the US and everybody has a [20:09] car and all of these different, you [20:10] know, it's like all every everyone has [20:12] all these household appliances and has [20:14] all, you know, it was a very different [20:16] life when they grew up um in in like, [20:19] you know, the '60s in China and they [20:20] were much much poorer and it was very [20:22] very different, right? And and it's like [20:24] I mean, funnily enough, people got used [20:25] to all these things pretty quickly and [20:27] now we can't live without them. But but [20:28] I think we'll kind of undergo the same [20:30] period with AI where 5 years from now [20:32] it's going to be insane to think about [20:34] all these things that we're going to [20:34] have. 10 years from now we're going to [20:36] have forgotten that we ever lived [20:37] without them. Honestly, [20:39] >> what do you think you understand about [20:40] AI that other people don't? [20:42] >> Um [20:43] >> the reason I asked the question is cuz [20:44] we have some mutual friends. I would [20:45] describe our mutual friends as some of [20:46] the most AGI pill that I know. [20:49] >> Yeah. And I feel every time I have a [20:51] conversation with them, I'm like, "Oh, [20:53] even though I pay attention to this [20:54] stuff, I feel like a toddler compared to [20:56] like somebody like you." And so I'm very [20:58] curious like what do you understand [20:59] about a that that most people don't and [21:01] even people within the tech industry. [21:03] >> No, I you're way too generous. I don't [21:04] know that there's [21:06] >> I don't know that I have that anything [21:07] that interesting or that deep of an [21:09] insight. I mean, I think it's it's [21:10] >> you you say that because you're used to [21:12] it. Here's what I'd say [21:15] is the way that folks typically kind of [21:18] predict the future or think about what [21:20] happens next is they pattern match based [21:22] on what they've seen historically and [21:24] they say, "Okay, well for 100 years it's [21:26] always been like this and it's probably [21:28] safe to assume that it will be." And 99% [21:31] of the time that works great, right? Um [21:32] and in these particular periods where [21:35] things that actually move and and [21:36] they're real things that are different, [21:38] those are the the 1% of times where it [21:40] truly is different. Now you just kind of [21:42] you know rather than any kind of pattern [21:43] matching like what what really matters [21:45] is just thinking about things from first [21:46] principles like AI you know there's [21:49] there's the famous like MER report which [21:51] was saying you know a couple years ago [21:52] AI would would would do about 10 to 20 [21:55] seconds worth of human work without [21:58] interruption and then you'd have to you [21:59] know guide it or direct it or it would [22:02] make a mistake or something like that [22:03] right 10 seconds 20 seconds and that's [22:05] just doubled every you know every couple [22:07] months basically and now we're talking [22:09] about like hours of work So, so [22:11] basically an AI can just take a task [22:14] that would have taken humans hours of [22:15] work to go do. If you go and describe [22:17] that task well enough to the AI, it will [22:19] just go and do the whole thing and come [22:20] back to you with the result and then you [22:21] give it the next thing the next thing, [22:23] right? And if you just ask from a first [22:25] principal's question, well, why can't [22:28] that be days or why can't that be weeks [22:30] or months of work? And then what does [22:32] the world look like if everybody has an [22:34] agent that can just do months of work [22:36] for them at a time? then you get to a [22:38] pretty different conclusion from from [22:40] what we've all seen and what we've all [22:41] lived for the last several years. And I [22:43] think that that [22:45] kind of first principles thinking is [22:49] as different as it sounds and as crazy [22:50] as it sounds. You know, this is one of [22:52] those times where it's actually more [22:54] correct than the the the the simple [22:56] pattern match if that [22:57] >> you I think you've even taken this [22:58] further where you're like, what happens [23:00] when they can work for a year [23:01] unassisted? [23:02] >> Yeah. Yeah. Um and I think I think [23:04] that's true and I think we will get [23:05] there. If Devon could work for a year [23:07] without any human assistance. [23:09] >> Yeah. [23:10] >> What would you have it do right now? [23:13] >> Destroy your competitor. [23:15] >> All sorts of things. I mean, no. I mean, [23:17] I I I still wake up and think about this [23:19] in every different like, you know, every [23:21] little thing that I run, you know. Dumb [23:23] example. Yesterday, I was sending out [23:24] like a bulk email [23:26] and I was like trying to get the like [23:28] email formatter to work and it's like [23:30] kind of painful. you know, some of these [23:31] things are still like kind of hard to [23:33] use or it doesn't support a certain like [23:35] styling of the email that you wanted to [23:37] go do. I think the thing I like had [23:40] pasted something with indents and then [23:42] it was like just couldn't like unindent [23:44] them because the editor like I don't [23:46] know there was some weird things where [23:47] the editor like would not allow you to [23:49] unindent one part but not the other or [23:51] whatever. And I was just thinking like [23:54] it's really crazy that like that I'm [23:56] still doing this basically, right? like [23:58] in as much simplicity or honestly more [24:01] as it would take you to explain this to [24:02] another person like hey here's what I [24:04] want to do I just like I'm just trying [24:05] to make it look like this and then this [24:06] and then that like the rest of that [24:09] execution should just be done for you [24:10] you know and then you get to the point [24:12] where you actually really just to get to [24:13] spend all your time thinking about well [24:14] what do you want to do you know what do [24:16] you want to build what do you want to [24:17] create like what are the things that you [24:19] want to see in the world that aren't [24:20] there already [24:20] >> but I think what makes it interesting [24:22] about what you were saying earlier is [24:23] like the the more you increase the time [24:26] the more interesting it gets to me so [24:27] like there's this guy named named uh [24:28] Edwin Land who I won't shut up about and [24:30] he was the founder of Polaroid. [24:31] >> Yeah. [24:32] >> Steve Jobs hero. A lot of Steve what we [24:33] think of as Steve Jobs ideas literally [24:35] just came from Edund Land down to like [24:37] the chairs and the table he would use [24:38] for his presentations. It's like the [24:39] same thing that Hoodland used in like [24:41] the 70s and Polaroid [24:42] >> and he thought of himself as a scientist [24:45] not as an entrepreneur. Edward Land he [24:47] died with I think the third most [24:48] patents. He's like Thomas Edison some [24:50] other dude and then Edwin Land. And what [24:52] he did is he couldn't figure out he [24:54] invented the the the industry of instant [24:56] photography. Yeah. [24:58] >> Uh before you took a picture and you're [24:59] like, "Hey, how's it look?" Like, "We'll [25:00] find out two weeks from now when we get [25:02] back in Kodak." Like, I had no idea. And [25:03] he now we he took a picture of a [25:05] Polaroid. He's like, "We'll find out in [25:06] 60 seconds when it dries." [25:07] >> But that was black and white forever. [25:09] When I read that part where you're like, [25:10] "Well, we're going to have agents that [25:12] can work on a sister for a year." I [25:14] didn't think of what I would do, which [25:16] is a question I just asked. I thought of [25:17] Edin Land hiring this guy. He's like, "I [25:19] want you to think about how we turn this [25:20] from black and white into color." [25:22] >> And before he could begin, the guy [25:24] worked there and just thought for two [25:25] years. [25:26] >> Sure. I'm like, that would be very [25:27] convenient if I can have an agent attack [25:29] this problem and while I'm working in [25:31] the background cuz I can't figure it [25:32] out. They're just thinking about how to [25:33] attack a single problem for two. In his [25:34] case, this guy that [ __ ] solved [25:36] instant color photography. It just took [25:38] him two years of thinking to do it. So, [25:39] this is like I'm going to push you on [25:41] this a little bit more because it's like [25:42] >> I don't want your year-long agent to [25:44] send bulk email. [25:45] >> Oh, I agree. I just to be clear. So, so [25:47] I I and I I think at some point it's [25:49] it's kind of, you know, you see this, [25:51] right, where it's like um you know, when [25:53] you're talking about seconds, you're [25:54] literally talking about just like a [25:55] specific command, right? When you're [25:57] talking about hours, you're talking [25:58] about giving it a task and having it do [25:59] the task, right? And I agree with you. [26:00] Obviously, bulk email is not, you know, [26:02] for for years, what you're talking about [26:03] is like you're giving the agent a [26:06] mission basically, you know, and it's [26:08] it's like this [26:09] >> that sounds way more fun. [26:10] >> Yeah. Exactly. And this this is like [26:12] what do I care about? And and you know, [26:14] the answer might be look, I I want to [26:16] I'll give you a million different [26:17] examples. You know, the answer might be [26:18] like there there's this like one, you [26:20] know, societal problem which is really [26:22] important to me and I think there's like [26:24] I think it's a solvable problem. I think [26:26] we can all be happy, but like um you [26:28] know, we really need to to to to to [26:30] spread awareness about it. We really [26:32] need to get folks to understand the [26:33] points of it and we we you know, we we [26:34] need to figure out how how everybody [26:36] should you know, work together and [26:38] coordinate on. That's your point. That's [26:39] a problem that an agent can can think [26:41] about, right? or or even you know some [26:42] some of the kind of like sillier things [26:43] too like yeah you know there's this [26:45] video game that I really like but I wish [26:48] you know if I were making the game [26:50] here's exactly how I would think about [26:52] all these things and I feel like there's [26:53] like this really cool idea if you know [26:55] you could combine elements from this one [26:57] game that I really like but then [26:58] incorporate some of the elements from [26:59] this other game and set the agent off on [27:01] that mission of like look we're going to [27:02] go and like make the coolest thing ever [27:06] and the coolest game ever and like we're [27:07] going to incorporate those elements [27:08] we're going to think about how those [27:09] like um you know nicely intertwine and [27:11] work together, right? Or if it's like [27:13] here's this this like piece of just like [27:16] novel science which I'm just like really [27:18] passionate about. You know, materials [27:19] have been created this way for for for [27:21] years and years, but like like here's [27:24] this like avenue of attack which I've [27:26] been wondering about and thinking about [27:27] of um maybe there's there's a different [27:30] like novel construction of materials [27:31] this way. Send your agent to go work on [27:33] that for for years, you know, or months, [27:35] right? and and like have it study that [27:37] and explore all these things and run its [27:39] own experiments and try all these. I [27:41] think all of these are are I think soon [27:43] going to be very possible. And to your [27:45] point, it's very different. [27:45] >> Yeah. I like the the framing of we're [27:47] sending the them on missions. [27:48] >> Yeah. [27:49] >> I would have one that would pick the [27:50] missions that I need to send other [27:52] agents on. [27:52] >> Yeah. Yeah. Yeah. Then you'll have the [27:53] AI which is like the the manager AI of [27:55] the the missions. Yeah. And and I think [27:57] it's like I I I think we will continue, [27:59] you know, it's um my example of this is [28:02] like I always joke about how, you know, [28:04] if you think about our ancestors from, [28:05] you know, hundreds of years ago or [28:07] thousands of years ago, imagine them [28:08] looking at us and what we do and it's [28:10] like, you know, you're just pushing [28:11] buttons, you know, and you're like [28:13] sitting in a room and talking with other [28:15] people and you call that a meeting and [28:17] that's like that those things that's [28:18] work for you guys, you know, and it's [28:20] like what do you mean that's work? You [28:22] know, like I'm in the fields like I'm [28:23] doing this every day, you know? I'm [28:25] going in farming. I'm making sure you [28:27] know making all of our our our like you [28:29] know clothes by hand or taking all all [28:31] taking care of all these things and [28:33] that's like work you know but like how [28:34] how can you guys call you know and my [28:37] point is just I think what we will have [28:40] going forward is going to look that [28:42] different from what we have today. We [28:45] will look at people who as we say just [28:46] like have these really interesting [28:48] curiosities that they want to pursue. [28:49] They have like causes that they're [28:51] passionate about. they have like fun [28:53] ideas or or or like art that they want [28:56] to create and like they're they're [28:57] sending off their agents in pursuit of [28:59] those missions and they will think of [29:01] that as work and we will look at them [29:02] and be like wow it's kind of crazy that [29:04] that's what you get up and think about [29:05] all day. I found one of my all-time [29:08] favorite quotes when I was reading the [29:09] book 0ero to1. The quote says, "The [29:11] single most powerful pattern I have [29:13] noticed is that successful people find [29:15] value in unexpected places. And they do [29:18] this by thinking about business from [29:20] first principles instead of formulas." [29:22] That is exactly what Apploven has done [29:24] with their advertising platform. [29:26] Apploving connects you with over a [29:28] billion potential new customers inside [29:30] mobile games. Apploven allows you to [29:32] capture undivided attention. Apploven [29:35] ads are full screen video ads that are [29:37] watched for an average of 35 seconds. [29:39] That is retention that blows other ad [29:42] platforms out of the water. And you can [29:44] launch on Apploven in minutes. You set [29:47] the goal and AppLovven achieves it. [29:49] There's no complex setup, no expertise [29:52] needed, and Apploven scales quickly. [29:54] They can put your ads in front of over a [29:57] billion potential customers. Other [30:00] businesses have seen immediate results, [30:02] have scaled to hundreds of thousands of [30:04] dollars of spend per day, and increased [30:06] their revenue by millions. So, you want [30:08] to get started quickly before all of [30:10] your competitors are on Apploven. And [30:12] you can do that by going to [30:13] apploven.com. That's apploven.com. [30:18] When you started Cognition, [30:21] did you know that you were going to try [30:22] to make an automated software engineer? [30:24] Was that your first idea? [30:25] >> I'll say it it was always two things. [30:26] One, it was always related to code and [30:29] software, which again probably has [30:30] something to do with us all being [30:32] programming nerds. Um and then two is it [30:35] was always around the idea that these [30:37] would be like real multi-step iterative [30:40] processes. Um which was uh I would say [30:43] was was a real hot take at the time like [30:45] this was 2023. [30:48] >> So this is when I met you guys [30:49] >> and uh I was on a walk with a mutual [30:51] friend of ours and I remember first here [30:53] this is you guys scaled from what like a [30:55] million in revenue to like 500 million [30:57] or something like that in what like 20 [30:58] months or something. [30:59] >> Yeah. Yeah. [31:00] I remember before you had no revenue and [31:01] I heard the idea and it was pitched as [31:03] like essentially like an automated [31:05] software engineer and I was like holy [31:06] [ __ ] [31:07] >> this like a huge you're going after [31:08] labor this is a huge market and then you [31:12] released [31:12] >> Yeah. [31:13] >> I remember the demo video. [31:14] >> Yeah. [31:15] >> And then you got a lot of [ __ ] [31:17] >> Yeah. Yeah. Yeah. [31:19] >> Mix of I think it was the the full polar [31:21] ends of the spectrum, right? So there [31:22] were some people who were like this is [31:25] the coolest thing ever [31:27] and like you know all these and then [31:30] there were some people who were like you [31:31] know this is like the worst you know [31:32] thing that you could but similarly like [31:34] in terms of the capabilities there were [31:36] some people who are like oh my god [31:38] everybody's going to lose all their jobs [31:39] tomorrow which is not what we've ever [31:41] really believed uh or thought of this as [31:44] um two also like dude there's no way [31:47] this is ever going to work and like this [31:49] is totally a scam you know. So what was [31:51] the criticism? Did you did you release [31:53] the product earlier than you wanted to? [31:55] >> Well, look, it it's it's it wasn't even [31:57] a product at the time to be honest. I [31:59] mean, it was more just like a prototype [32:00] or like a demo of what was possible and [32:02] and you know, we had been working with [32:04] it and playing with it for a few months [32:05] at that point. And the kind of just [32:08] wanted to show people some of the [32:09] examples of what it was capable of [32:11] because it was a real like like it was [32:13] pretty it was pretty insane for us to [32:15] see as well. Like I remember the first [32:16] time that it did like a real task like I [32:18] could not sleep that night. Uh a and so [32:21] it's just like showing that um [32:23] >> say more about that. [32:24] >> The first task that Devon did was it [32:26] like set up MongoDB for us and it's like [32:28] you know it's a it's a standard thing. [32:30] It's issues that a lot of people run [32:32] into all the time when they're going and [32:33] getting their kind of like initial um [32:35] you know DB setup or whatever. U but we [32:37] would just like run into errors. you you [32:39] have this whole flow where you like um [32:43] you find an error, you Google error, [32:46] find some message on Stack Overflow or [32:47] whatever or you ask Chat GB and it tells [32:49] you, okay, here's here's one you should [32:51] try. You try that thing, you run into a [32:53] different error, you like paste that in [32:55] and then you know and and like for this [32:57] one, we kind of like at some point we're [32:58] just like, "Okay, De just just try to go [33:01] fix it. Just go run the commands like do [33:02] whatever you need to go do it." And then [33:03] it worked [33:04] >> and you couldn't sleep. It was truly [33:06] like because again it's you know it's [33:08] it's like just seeing the exponential [33:10] curve ahead because this was a very you [33:12] know very specific case. It was the one [33:14] success that we had. It was very much [33:16] like a you know like like definitely a [33:20] way better than average run. But no [33:22] there was this feeling of like why [33:24] shouldn't all software and all products [33:27] be built this way now? Like you can just [33:29] tell it what you want it to do and have [33:30] it go do it. I mean it's kind of funny [33:32] because yeah to your point like we did [33:33] get hate in the beginning. We did not [33:35] feel motivated at all from the hate to [33:38] be like, "Oh yeah, I know. Maybe you [33:39] guys are right. Like maybe we should go [33:41] and like, you know, focus on the more [33:42] kind of like chatbot Q&A style product [33:46] experiences." Like maybe more than we [33:48] should have. Maybe maybe we should have [33:49] done something in the middle, you know? [33:50] I I don't know. But but like I I think [33:52] for us like when we had seen that and [33:54] when we had done these different things [33:55] like all of those like demos that we [33:57] showed you know in in in our launch [33:59] announcement were like actual runs of [34:01] dev that we had done ourselves and like [34:03] run into and been like holy [ __ ] this is [34:04] insane for us. It was always kind of [34:06] like look you can you know we can debate [34:08] when or or or or like what level of [34:10] effectiveness or whatever but like it's [34:13] just it's going to happen and that's how [34:15] we always felt about it. [34:16] >> So explain the process of iteration to [34:18] go from that product you're getting a [34:20] lot of hate. Well, uh, actually, you [34:21] know what? I called Jeremy Stern, who [34:23] wrote this excellent profile of you in [34:26] Colossus, which there's a lot of parts I [34:28] would just laugh my ass off, by the way. [34:30] >> And I was asking, I was like, "Tell me [34:31] the stuff that didn't make uh the [34:33] profile." [34:33] >> Yeah. [34:34] >> And he said there wasn't that much [34:35] because you're kind of like an open book [34:36] and a lot of people like um managed [34:38] their media and you know, you could have [34:40] to talk about certain stuff off the [34:41] record or whatever and you were just [34:42] like saying everything. [34:44] >> But he did say something about like you [34:45] made the point where like when you [34:46] released the first product, what was the [34:48] benchmark? [34:49] >> Sweet Bench. Yeah. It was like 13%. [34:51] >> That was Devon. [34:53] >> And your point was that that's already [34:54] better than [34:55] >> Yeah. At the time the best known was [34:57] like 3 or 4% or something like that, but [34:59] obviously Yeah. 13%. It still means you [35:01] fail, you know, 87% of the time. [35:03] >> Was there a pronounced benefit from [35:04] releasing early like that? [35:06] >> For sure. Yeah. I mean, so so, so for [35:08] us, by the way, you know, if you kind of [35:10] think about this overall AI ecosystem, I [35:12] mean, a simple way to put it, it's like, [35:13] dude, we were late by a lot. You know, [35:15] it's like OpenAI started. [35:17] >> Why though? [35:17] >> Google Deep Mind or Google Brain. I mean [35:19] these are obviously you know more more [35:21] than a decade old already. OpenAI [35:22] started like end of 2015 or even you [35:24] know like the anthropic or or you know [35:26] other that that folks talk about in the [35:28] world like had been around already for [35:29] years like we were getting started in [35:31] early 2024 there's like a year plus out [35:34] from the chat GBT launch a lot of the [35:36] existing players were already there you [35:37] know the same was true in code [35:39] specifically I mean there was you know [35:41] GitHub copilot um which had people had [35:43] already you know engineers had already [35:44] used for years and that was very much [35:46] the like Q&A like the autocomplete style [35:49] experience of of you are working with [35:52] AI. You know, when you start a company, [35:54] you kind of have nothing. Like you you [35:56] have no right to exist is maybe one way [35:58] to put it. It's like an interesting [36:00] truth about the world which is obviously [36:02] startups succeed. You know, startups [36:04] succeed all the time and startup versus [36:06] big company has been played out for, you [36:08] know, for for forever. Um, but like [36:11] there's no reason, you know, in terms of [36:13] resources, in terms of people, in terms [36:15] of brand awareness, like you have none [36:17] of the things that, you know, the big [36:20] guys have. And and so from that [36:22] perspective, you have no right to win, [36:24] you know, over over over what they're [36:27] doing, you know. And then the reason [36:28] that you're sometimes able to anyway [36:31] is [36:33] if you really like um you know, plant [36:36] your flag in the ground and and and put [36:38] a stake into like what you think the [36:39] future is and you run like hell towards [36:41] that and and if you like turn out to be [36:43] right on some of the core things and [36:44] like I think we were we were wrong on a [36:46] lot of things. uh to be clear a lot of [36:49] things we you know so definitely we were [36:51] early which is a very fair criticism I [36:53] think a lot of the details and the [36:54] nuances like we learned and adjusted [36:56] over time but the idea that like you [36:58] would work with AI as a co-worker you [37:00] know rather than as like a tool or a [37:02] chatbot I think was I mean over the last [37:04] couple years has obviously like really [37:06] grown and I think it was like very [37:07] important for us for our brand for [37:10] recruiting for you know customer work [37:12] everything for us to be you know the [37:14] first ones that actually planted that [37:15] flag in the ground and said So when did [37:17] you have this idea where you guys are [37:19] going to take a run at very I would say [37:22] ferociously to like these giant like [37:24] Fortune 500 companies or even the like [37:26] the US Army uses you like where did that [37:28] strategy come from? So, funnily enough, [37:31] so that launch was in March of 2024 and [37:33] to your point, it went very viral, but [37:35] again, we didn't have any customers. We [37:36] didn't have any revenue. We didn't [37:37] really have like we had [37:38] >> I thought your first iteration of the [37:39] business model was like $500 a month or [37:41] something, wasn't it? Am I miserable? [37:43] >> So, we had that that was actually later [37:45] on actually. That was that was end of [37:46] 2024. [37:47] >> Okay. [37:47] >> But, you know, initially what we started [37:49] was just like, you know, a bunch of [37:50] people came asked us for the product. We [37:52] were like, "Guys, I I don't know if this [37:54] is a product that's ready for prime [37:56] time, but if you really, you know, you [37:57] can try it and like just let us know." [37:59] And so, like, people ran, you know, we [38:02] had to go and scramble to build a system [38:03] of like a okay, let's do like a pilot or [38:05] like a PSC, you know, and like we tried [38:08] our best to like not overpromise and be [38:09] like, "Guys, like really, it's like very [38:11] early, but if you want to try it, you [38:12] know, you can try." And we did all these [38:14] and perhaps unsurprisingly, I mean, they [38:16] were all just like failing [38:18] um which is kind of natural. like it's [38:20] like you could do some pretty cool [38:21] things with it. You could you could do [38:23] some pretty interesting like toy demos [38:24] or projects or whatever, but it was [38:26] certainly not ready to work on like [38:29] actual companies like real codebase and [38:30] so on, right? And so from that point, [38:33] this would have been like April and May [38:35] of 2024. We were make making agents work [38:37] with like GP4, you know, this is like a [38:40] very different era. So they they were [38:41] much more primitive agents. We kind of [38:43] talked about this and got to this [38:46] question of like okay well what do we [38:47] think it actually does look like when [38:48] the agents start to get good enough for [38:49] adoption and I think what we kind of [38:51] came to was well [38:54] different tasks are different right and [38:55] so so like there are some tasks out [38:57] there that are just really mundane and [38:58] really repetitive and you're just doing [39:00] the same tedious thing over and over and [39:01] over again and there are some tasks on [39:03] the other end of the spectrum that are [39:04] like really tough like architecture [39:06] problems or like deep like you know deep [39:08] issues that you have to like really [39:10] understand all the context and have all [39:12] the knowhow how to to know how to fix [39:14] and like people were trying to use Devon [39:17] for all of those and we're failing at it [39:20] understandably and so then the question [39:22] was like okay what are the the the [39:24] natural tasks that are like if there's [39:26] like a first task that is going to have [39:28] PMF you know and real value from these [39:30] kind of agent experiences where it can [39:32] do the whole thing end to end what is [39:33] that going to be we kind of said okay [39:35] well it should be some of these like [39:36] really repetitive tedious ones it's not [39:38] so cut and dry that you can just like [39:40] have an automated script that does the [39:41] exact thing every time. So it does take [39:43] obviously some intelligence and some [39:45] amount of meandering, but it is like [39:47] repetitive enough and scoped enough and [39:49] like on a tight enough feedback loop [39:51] that you can have an agent do it and it [39:54] would be able to kind of go and diagnose [39:55] and like fix that problem, right? And so [39:57] that's kind of what brought us to some [39:59] of these like naturally to to some of [40:01] these like you know initial use cases [40:03] which were things like migrations or [40:04] version upgrades or you know helping [40:06] people upgrade from like Java 7 to Java [40:08] 8 or something like that which were kind [40:10] of like you know as you can imagine [40:11] enterprise had enterprises had these [40:13] massive code bases where they would go [40:15] and do all that you know and it's like a [40:16] 50,000 file codebase where you have to [40:18] go you know it's like the same eight [40:19] things that you need to change in each [40:20] one you have to be a little bit [40:21] thoughtful about how you make the [40:23] trade-offs but like it's a very [40:24] repetitive task right Our first success [40:28] ended up being with a company called New [40:30] Bank, uh, biggest bank in Brazil, you [40:32] know, by market cap at the time. And the [40:34] use case was like one of these big [40:35] migrations. And we had, uh, kind of a [40:38] custom Devon that was like extremely [40:40] extremely optimized um, for doing that. [40:43] And as we grew from there, you know, [40:45] later on we had kind of as things got a [40:48] little bit more mature, you know, we had [40:49] both self-s served business and [40:51] enterprise business and so on. But I [40:52] think from the beginning we had always [40:54] seen this value that building software [40:56] in the real world and managing massive [40:58] massive products that like millions of [41:00] people use every day was pretty [41:02] substantially different from you know [41:04] just building a cute website or a cute [41:06] demo from scratch right and I think we [41:07] really really leaned into that and [41:09] naturally it's you know all the Fortune [41:11] 500 or all the biggest companies in the [41:13] world are software companies in 2026 you [41:16] know even the ones that that folks don't [41:17] necessarily think of that way right and [41:19] like Walmart or CVS or JP Morgan or um [41:22] or or Mercedes or what, you know, [41:24] they're they're all software companies, [41:25] right? They have massive massive teams [41:27] of software engineers. They have tons of [41:28] things that they're building and [41:29] shipping and um and maintaining. And [41:31] that kind of became like a natural thing [41:32] for us to work on because like I think [41:34] we had had learned very early on that we [41:36] want to work on real problems and we [41:38] want to work on things that that that [41:39] matter and that people actually care [41:40] about. [41:40] >> So wait, what percentage of your revenue [41:42] is coming from enterprise then? [41:44] >> Today [41:45] around 75 80%. [41:47] >> What are people using on the self-s [41:48] serve? Like what are examples of [41:50] >> Yeah. So we have a lot of teams who use [41:52] it in self-s serve. I mean that's grown [41:53] actually you know that's grown quite a [41:54] bit as well lately. Um but um you know [41:57] we have startups who who you know we [41:58] have exa um who uses it a ton or open [42:01] router or built you know I I I ran into [42:04] someone in in my apartment in the [42:06] elevator the other day like oh you're [42:08] the dev guy like we use de uh the [42:11] >> are you Devon? [42:13] >> I' I've gotten that as well. Hey are you [42:14] Devon? And I was like you know I'm [42:16] actually not but that's that's that's [42:18] okay. Uh [42:18] >> why don't we call it Scott? [42:21] um you know there there's kind of both [42:22] self- serve and enterprise. With that [42:24] said, it is entirely in both sides. It [42:26] is still like actual engineering teams [42:28] building real output. And so so we don't [42:30] focus at all on like you know individual [42:32] hobbyists who are just like trying to [42:34] make a cool thing or something like [42:35] that. We focus on like real teams who [42:37] are building real products that they [42:38] want people to use and um and getting [42:40] output out of that. [42:41] >> Okay. Can you walk us through I'm very [42:43] curious like I'm a big enterprise. Yeah. [42:45] >> I contact you. Yeah. [42:46] >> Walk us through what happens to being a [42:48] customer. you know, it starts with [42:49] education. Um, and everyone's gone crazy [42:51] over AI and agents and so on. And those [42:54] are the buzzwords of the last six [42:55] months, obviously. So, they want to know [42:57] more about this, but there's still a lot [42:58] of detail and like what does it actually [42:59] look like to deploy them? What? So, so, [43:01] you know, we'll we'll show them what [43:02] this looks like. We'll talk them through [43:03] like how we work with with teams and how [43:06] we partner, you know, how we direct them [43:08] to the right use cases, how we give them [43:09] guidance on like how, you know, to [43:11] maximize their ROI or like what projects [43:13] are or are not feasible with agents. And [43:16] then from there, you know, enterprises [43:18] typically obviously have some very messy [43:20] processes. And so, you know, for most [43:22] software or most, you know, just [43:24] generally like vendors that they'd want [43:25] to work with, it's often, I mean, for, [43:28] as you can imagine, for a massive bank, [43:29] you know, adopting software, giving it [43:31] access to all of their, you know, um, [43:33] the repos getting through security or [43:35] whatever, that's like a usually um, uh, [43:38] for for typical companies is like a 12 [43:40] to 18 month cycle. The thing that we do [43:42] naturally is is is we just work with [43:44] them to figure out how we go and do that [43:45] as fast as as humanly possible. [43:47] >> Did you send employees down to South [43:48] America for New Bank? [43:49] >> Well, so New Bank I mean the first one [43:51] honestly our entire team was the foreign [43:52] deployed team. like we all flew to [43:54] Brazil like I mean it was this like the [43:56] first case you know it's like it's like [43:58] let's be real here okay Asians did not [44:00] work generally okay and so there was it [44:02] was a lot of like how do you make it [44:04] work very specific like we all flew to [44:06] Brazil I was there the whole team was [44:08] there we were sitting there with their [44:10] engineers understanding okay so this is [44:11] what you do in that case this is what [44:12] you do in that case and this is what [44:14] Devon needs to know and Devon needs to [44:15] be able to read these things and like [44:17] going and debugging their like exact [44:19] problems now obviously it's not like [44:21] that at all because you Hold on. We'll [44:23] get to where it is now. But the idea of [44:25] like No, we didn't deploy a a team. We [44:27] deployed the whole company. [44:29] >> Oh, yeah. Yeah. It was I mean it was [44:31] getting the first customer obviously was [44:33] like a real, you know, it's like uh I [44:35] mean I wonder what they thought of that. [44:37] >> But but but like yeah, it was like a [44:39] literally like okay, let's go through [44:40] all these different things that you guys [44:41] think could make sense. Let's go through [44:43] each one. Let's try some of it manually [44:45] ourselves to understand what the task [44:46] looks like. let's see if we can teach [44:48] Devon to do this correctly and build in [44:49] the right kind of like um you know the [44:51] the the right orchestration for Devon to [44:53] be able to do this and like let's just [44:55] like basically building the product was [44:57] like almost like building for for one [44:59] company and yeah it was fun. [45:01] >> So what do you do today? So today it's [45:03] it's obviously much more um uh it's it's [45:06] it's it's much more self-start and you [45:08] know agents are are so much more capable [45:10] obviously and we've we've figured out a [45:12] lot more things with the onboarding [45:13] experience but but a lot of it is you [45:15] know we're saying these cycles typically [45:16] take 12 to 18 months. We try to get [45:18] deployed with folks you know than like 3 [45:19] months and a lot of that requires folks [45:22] to obviously first of all it requires [45:24] them to really appreciate that it's a [45:25] priority. Um, I mean, if you have 25,000 [45:28] software engineers that you're, you [45:29] know, in an org that's that's that's [45:32] running that that that that that [45:34] costs, you know, 10 billion a year or [45:36] something that you think can move three [45:38] times faster. You know, usually it is a [45:40] priority, but then it's like figuring [45:42] out how we get through the security [45:43] reviews. You know, we we have all the, [45:45] you know, we can deploy in their private [45:47] cloud. We have very strict data [45:49] agreements. We have like tight airwalls [45:51] on on on on all these things. Obviously, [45:54] um you know, [45:55] >> do employees physically have to go? Are [45:57] they are they working with like are [45:59] these deals so big that the that they're [46:01] working with a specific company and only [46:03] that company for a period of time or no? [46:05] >> Typically, no. Um we have uh you know, [46:07] we have a full like foreign deployed [46:10] motion, but a lot of what that looks [46:11] like as we're saying is is a little bit [46:13] more like user education and guidance. [46:14] And so we will fly, you know, we we [46:17] still do this where we'll fly out and [46:18] kind of like go and see customers and [46:20] work with them, point them to the right [46:22] use cases, teach them how to make get [46:24] success, help them with their like setup [46:26] and their playbooks, like all these [46:28] things for how to use Devon. But it's [46:30] much more a kind of like look, we're [46:32] here to assist you guys. We're we're [46:33] giving you a lot of this kind of like [46:34] direction and so on, but like you are [46:36] yourself, you know, and your team is the [46:39] one that's using Devon and running all [46:40] that, right? So that's a lot of what [46:42] that looks like. We're set up to be as [46:44] incentive aligned with our customers as [46:46] possible. And so a lot of it is like is [46:48] not just like oh here's this tool like [46:49] hand it over to your engineers and like [46:51] find out what they say about it. It's [46:52] like okay well let's figure out like [46:54] what are the initiatives that you guys [46:55] really care about and like we will point [46:57] you to for those initiatives. Here are [46:59] each of the projects that we think Devon [47:00] right now can make you 10 times faster [47:02] on and we'll tell you for the ones that [47:03] that it's not and here's here's what [47:05] workflows you should use instead or [47:07] here's how you should get to value [47:08] instead. Um but it looks more like that [47:09] than like a um than than like a literal [47:12] like you know we are using Devon on your [47:15] on your on your behalf or something. [47:18] >> Deal is how the best founders turn the [47:20] world into their talent pool. I've been [47:22] studying how history's greatest founders [47:23] operate for a decade. And one thing they [47:25] all have in common is they understand [47:27] that recruiting and hiring the very best [47:29] talent is your most important priority. [47:31] A players recognize other A players, [47:33] which is why top companies like Ramp, [47:35] Shopify, 11 Labs, Uber, and Door Dash [47:38] all use deal. Many of the top founders I [47:40] know have personally invested in Deal [47:42] after using their product. And what they [47:44] discovered is that Deal is the best [47:45] company in the world at building [47:46] infrastructure for global hiring. Deal [47:48] will help your business hire, pay, and [47:50] manage any worker anywhere in the world [47:52] so you can retain the best talent [47:54] anywhere and spend the rest of your time [47:56] focusing on what you do best, delivering [47:58] value to your customers. The founder of [48:00] 11 Labs has a great description of the [48:02] value deal can give your company. He [48:04] said, "We built 11 Labs to break down [48:06] language and communication barriers. [48:08] With deal enabling us to hire and [48:10] support exceptional talent anywhere, we [48:13] can accelerate our innovation and bring [48:15] more voices, stories, and ideas to every [48:17] corner of the world. Deal is trusted by [48:20] over 40,000 businesses. Learn how they [48:22] can help your business today by going to [48:24] deal.com/enra." [48:26] that is deal.comra. [48:30] >> Wait, so tell more about how you align [48:31] the incentives between your company and [48:33] then your customers. [48:34] >> Yeah, so a couple things here uh which I [48:36] think are really important one obviously [48:38] is is just like really working with them [48:41] on the the just very clearly defining [48:44] the ROI. Um, which I think is a very [48:46] important I mean it's like everyone's [48:48] talking about this right over the last [48:49] few months like people are going crazy [48:50] on their token budgets and you know one [48:51] engineer can spend so much and you know [48:54] you want to know that that's actually [48:55] doing real value for you and you want to [48:56] be able to identify where you are. [48:58] >> You think all this stuff is a little [48:59] crazy. [49:00] >> I think it is directionally correct but [49:01] I think there's some you know the there [49:04] are definitely some places where people [49:05] have gotten carried away you know people [49:06] talk about oh like yeah like we rank our [49:08] engineers by how many tokens they're [49:09] spending. Well, let's let's try and rank [49:11] people by how much output they're [49:13] actually producing or how much good [49:14] work, you know, is actually getting [49:15] done, you know, but but obviously I [49:16] think that like the math works out in [49:18] terms of like, you know, the GPUs are [49:20] expensive, but like if your engineers [49:22] are actually able to ship three times [49:23] more, you know, then it's very clearly [49:25] worth it. You just want to make sure [49:26] you're doing it the right way, right? [49:27] Obviously, a lot there on just like [49:29] tying to specific outcomes, tying to [49:31] okay, what are the actual tickets that [49:33] are getting done? what are the what are [49:34] what what what are the projects and the [49:35] initiatives and like this project which [49:37] was scoped out for 18 months and was [49:39] going to be handed off to you know an [49:41] outsourced contractor um and and was [49:43] going to cost you 15 million like let's [49:45] just talk about how that you do this all [49:46] internally with your own team and you do [49:48] it for 1 million and you get it all done [49:51] in 3 months you know uh stuff like that [49:53] uh which I think is super important the [49:55] second thing which I'll call out [49:56] especially is just being neutral a lot [49:59] of discussion obviously what do you mean [50:00] by that being being neutral with respect [50:02] to all the labs, the models themselves. [50:05] >> I think you said you like being [50:06] Switzerland. [50:06] >> Yeah, we like being Switzerland. [50:07] Exactly. And so I think it's like an [50:09] important thing of of you know, we are [50:12] just as incentivized as they are to [50:14] figure out how to make their token spend [50:16] efficient, right? And so so Devon is [50:18] purposefully meant to be uh you know, a [50:20] compound model system. And so all of [50:23] these different things, you know, for [50:24] the right task or even the right subtask [50:27] of a task or something like that. [50:28] >> Describe what a compound model system [50:30] is. [50:30] >> Yeah. Yeah. So, so for each different [50:32] part of your use case, I mean, imagine, [50:34] so let's say you tell Devon, hey, um, [50:36] customer just reported this bug. Uh, you [50:38] know, there's a jer ticket. Can you take [50:40] a look at the ticket and like go and [50:42] solve the whole thing, right? What does [50:43] Devon actually go do, right? First of [50:45] all, Devon's probably going and like [50:46] investigating, understanding like what [50:48] does the report say? What's going on? [50:50] Second of all, probably, you know, as [50:51] any engineer should, like it's going to [50:53] go and try to reproduce the bug itself, [50:54] right? So, it it'll say, "Okay, let me [50:55] spin up the product locally. we click [50:57] around, try to follow the same steps [50:59] that they followed and see if I can make [51:01] the bug happen as well. Right? And if [51:03] you did that, then like, okay, now [51:04] you're looking at the logs. You're [51:05] trying to figure out what went wrong. [51:06] You're pinpointing what what are the the [51:08] particular files or what was the flow [51:10] of, you know, what was the command flow [51:11] that led to this. Then you do the [51:13] debugging. Then you go and test it [51:14] again, make sure, you know, all these [51:16] steps, right? Then you put it up for [51:17] review. And it turns out that there are [51:19] different models that are good for [51:20] different parts of these tasks, right? [51:21] And even for across different tasks [51:23] obviously also it's like some tasks are [51:26] these really crazy hard ones where you [51:28] want to actually use max thinking and [51:30] you want to use the the the very best [51:32] models you can get your hands on in the [51:33] world, right? And then many other tasks [51:36] are, you know, boilerplate enough or [51:38] repetitive enough that what you care [51:39] about is just getting it done really [51:41] fast, getting an an immediate answer, [51:44] having the ability to verify that it was [51:46] correct, but then beyond that making [51:48] sure that it was like as cheap and fast [51:49] and efficient as possible, right? And so [51:51] Devon rather than being pegged to one [51:53] model and saying, "Oh, we're only going [51:55] to serve you GPT for this or we're only [51:56] going to serve you for this." Devon can [51:58] use any of the different models it has [52:00] in its arsenal, which include all of [52:01] these models from anthropic, OpenAI, [52:03] Google, etc. Um, but but also our own [52:05] models, right? Or open source models out [52:07] there, and it will, you know, [52:09] dynamically go and choose these models [52:11] for these tasks. [52:12] >> How do you think about this? Like you're [52:13] a customer of them, but also competitor [52:17] with. [52:18] >> Yeah. No, I mean look I I think in in [52:20] practice there's a lot of positive some [52:22] work for us to do together. Um and so [52:24] like you know the way that we think [52:26] about ourselves a couple things. One [52:27] software is the only thing that we care [52:29] about obviously. Um and there's a lot of [52:31] value in focus and in building products [52:33] and building our whole home whole [52:35] motions specifically around that. [52:37] >> Say more about how you think about focus [52:38] and the and the value in it. [52:39] >> You know back to what we were saying [52:41] about startups, right? Like why do [52:42] startups ever win at all? And if you do [52:45] everything, you will lose to Microsoft [52:46] or Google who does everything but also [52:48] has like trillions of dollars more [52:49] resources and 100,000 more people than [52:51] you, right? And and infinitely more [52:52] brand name, right? Um and and like the [52:55] way that you build, you know, like a [52:56] real kind of like um you you know a real [52:59] like lasting business or lasting product [53:02] is by [53:05] really really focusing and narrowing on [53:07] one specific thing, making a very [53:09] concentrated bet, and then obviously, [53:11] you know, your bet has to end up being [53:12] right. But but but like you know from [53:14] there it's like a lot of just like [53:15] really tight execution. Um and so in [53:17] software you know I love um you and I [53:19] have talked about this. I love the quote [53:21] from from Daniel in Spotify, right? [53:23] People are saying so like you know [53:25] there's there's like like YouTube's [53:26] trying to go and do this and Apple has [53:28] Apple Music and like why why should [53:29] there be like a he was like you know I [53:31] can give you all the other reasons but [53:32] the truth is we're just going to care [53:33] way more about music than they are. And [53:35] I think for us it's true. It's like we [53:37] we are just going to care so much more [53:38] about like what does it look like to [53:40] build software end to end at Goldman [53:43] Sachs or at like Mercedes-Benz or [53:46] something like that, you know? And and [53:47] there's a lot of nuance in that and [53:48] there's like a lot of messiness in that, [53:50] right? It's it's not as simple as like, [53:52] oh, here's a sandbox algorithms problem. [53:54] Go and code me the correct, you know, 30 [53:56] line program that solves this problem or [53:58] something. It's like how do you work [54:00] with all the messiness of the real [54:01] world? How do you understand the [54:02] codebase as it exists today? How do you [54:03] like collaborate with all the humans on [54:05] it? How do you plug into their ticketing [54:06] system? How do you give the agent the [54:08] ability to test its own code and run [54:10] everything locally? Right? Like all of [54:11] these are are obviously just super [54:13] hairy, messy problems. Uh and that's [54:15] that's that's what we care a lot about, [54:16] right? And so from that perspective, [54:18] it's it's a very kind of nice like, you [54:20] know, the labs have their own products. [54:21] I'm sure they'll continue to do more, [54:22] but but in practice, like there are a [54:24] lot of nice ways for us to cooperate. [54:25] Obviously like on top of that like being [54:27] the Switzerland means that um you know [54:30] folks can work with us and kind of like [54:31] trust in us that that that we will kind [54:34] of route them to the right models that [54:35] we will optimize the price performance [54:37] for them that we will kind of like [54:38] direct them to the right use cases and [54:40] so on because you know we're not [54:42] incentivized for them to spend more on [54:44] the models either you know we're [54:45] incentivized for them to get value and [54:47] to get output out of it. I love that you [54:49] use the example of the war between [54:51] Spotify and the rest of the like Apple [54:53] Music. For example, Jimmy I who also [54:55] came on the show and now is actually a [54:56] friend of mine. He actually called me [54:57] yesterday about this cuz anything [54:59] anytime anything happens with AI music [55:01] and Spotify, he calls me. But we talked [55:03] about this because he's like, "You don't [55:04] understand." Like he had like 40 years [55:05] of experience in the music industry, had [55:07] all the relationships. He's like, "Apple [55:08] bought me for $3 billion. Spotify at the [55:11] time." He's like, "We're gonna go [55:12] headtohead with them." Spotify only had [55:13] three million paid subscribers at the [55:16] time. Jimmy's gonna fight the war. And [55:17] he's like, I have the one of the biggest [55:19] companies in the world. And then he [55:20] talked about, he's like, that wind up [55:21] being a huge, he thought it was an [55:23] asset. It was a huge liability. Yeah. [55:24] >> Because they're like, we don't give a [55:25] [ __ ] about getting a couple million [55:27] more, tens of millions or more of [55:28] subscribers. We invented the most [55:31] successful consumer product of all time. [55:33] And he said something like they just [55:34] clipped his wings [55:35] >> like he tried to do something like, "Oh, [55:37] this isn't important to us." Where [55:38] Daniel was going to die if he was not [55:40] successful. He just cared about a lot [55:42] more. Dude, even in the two years that [55:44] we've been around, you know, I've heard [55:46] so many different versions of the same [55:48] argument because when we started, as you [55:49] can imagine, this was true for us. It [55:51] was true for everybody in the space [55:52] building the space. The the the number [55:54] one push back that you would always hear [55:55] from investors, from other people, [55:57] people. It was like, but like Microsoft [55:59] already has GitHub Copilot. Like, isn't [56:02] that everybody's just going to use that, [56:03] right? And it's like it was a very [56:04] reasonable thing to say in some sense [56:06] because you know they did have all you [56:08] know they did own all of GitHub and they [56:09] did have the partnership with OpenAI and [56:11] they did have like VS code and so on. I [56:13] think in practice the reality is there's [56:15] so much more innovation and so much more [56:16] to build that there was a lot of like [56:18] positive you know Microsoft's like a [56:19] great partner of ours and we do a lot of [56:21] things together and we build even more [56:22] together right and and I think the [56:23] reality is like people have said this [56:25] forever of like oh like yeah like [56:26] startups versus you know like why should [56:28] you can give all the rational [56:30] >> 10 years ago it was like Google do this [56:32] or Facebook it's like oh why should data [56:34] dog exist or why should snowflake or [56:36] data bricks exist or whatever you know [56:37] the clouds the clouds have all you know [56:39] the clouds care about observability too [56:40] the clouds care and the reality it's [56:42] Like it it's it's in some sense a bit of [56:45] an uncreative way to think about things. [56:47] I think like if if there was like one [56:49] thing that okay here's what we all know [56:51] is going to be the endstate future and [56:53] everybody's just working toward it and [56:54] whoever has the most resources toward it [56:56] wins. Of course, yeah, you know, it's [56:57] like we know who has the most resources [56:59] today, right? But if there are millions [57:01] of problems out there to solve, there's [57:02] lots of different things. The world is [57:04] dynamic. Things change all the time. [57:06] There's lots of new, you know, ideas or [57:08] opportunities or or or ways to to to [57:10] build new products. Um, then the reality [57:12] is like, of course, there's there's [57:15] there's lots of different niches to own. [57:17] There's different things to really bet [57:18] on. There's different focuses to to to [57:20] spend your time on. And I think that [57:22] will continue. Yeah. I mean, I I think [57:23] it's like a for better or for worse [57:25] after the last couple months of news. [57:26] You know, Cognition, I think, has been a [57:28] bit more of the like we've become a bit [57:30] more known as as the, you know, the [57:32] folks betting on independence in some [57:33] sense because obviously there have been [57:35] some high-profile acquisitions. It's [57:36] funny that you said that because I was I [57:38] want I wanted one of the questions I'm [57:40] going to ask you which you probably [57:41] won't answer is like how many different [57:42] acquisition offers have you had? [57:44] >> I will not answer that question. [57:46] >> Dozens more? [57:47] >> Dozens is a lot, dude. I don't know [57:49] about dozens but okay. [57:50] >> The time it's probably only a handful [57:52] that could actually have afford to buy [57:54] you if you would sell. But the amount of [57:56] times they keep coming back. [57:57] >> Oh, I see. Well, yeah. Anyway, [58:01] >> no, this this independence part is [58:02] really interesting to me. We were on the [58:04] phone, me, you, and a mutual friend on [58:06] three-way like I don't know, I think it [58:07] was actually last summer. And um we're [58:10] not going to say the company, uh but I [58:14] was like, "Scott, what do you what do [58:15] you what do you think about uh you know, [58:18] an acquisition with X, not X, the [58:21] platform, X the blank company." You're [58:23] like, [58:24] >> "I don't know how we'd be able to afford [58:25] them." You just assumed I meant you [58:27] buying them. And I just [ __ ] cracked [58:29] up laughing. I was like, "It's [58:30] hilarious." [58:31] >> No. So, so I I I think like um Yeah. [58:36] Yeah. And it's it's like folks like [58:38] there there folks taking acquisitions [58:39] and doing things. There's some big [58:40] high-profile ones. I think those are [58:41] great to be clear and I think it's a [58:43] very reasonable path for exit. As [58:45] mentioned for us, it's like we we've [58:47] we've all come into this like we want to [58:48] build a generational business. That's [58:49] something we're really excited about. [58:50] And and I think people I think today [58:52] sometimes I I've I've seen some more of [58:54] this nihilism where they say, "Oh, like [58:56] yeah, maybe it's just maybe it's too [58:57] late and maybe it's not possible [58:58] anymore." And like [59:00] >> what does that mean? What's not [59:01] possible? like maybe it's not possible [59:02] to build a new independent business [59:03] because everything else, you know, it's [59:05] like [59:05] >> the labs are going to do everything. [59:06] >> All the opportunity is take it, you [59:07] know, and it's like guys, [59:09] >> those people aren't founders. Founders [59:11] are rationally optimistic. [59:12] >> Yeah. [59:12] >> They're just not founders. [59:14] >> Yeah. [59:14] >> They they believe even there's no [59:16] evidence that they should succeed, that [59:18] they will succeed. [59:19] >> So people like that just need to get a [59:21] job. Dude, I'd say this even with Devon, [59:22] like if if you think that going onto the [59:25] Devon web app and like giving the or any [59:27] other coding product out there today, [59:29] you know, and and like giving the [59:30] instructions the way that you do right [59:31] now and then you get the pull request [59:32] out and then that's how you review it [59:34] and that's how you build software. If [59:35] you think that that's going to be the [59:36] way that software is built forever, then [59:38] yeah, then then nothing will change. But [59:39] like I think we have 10 more generations [59:42] of these different product experiences [59:43] to come, right? And like building those [59:45] and doing those is like that is what [59:47] innovation is. Like that's that's that [59:49] that's what's going to happen. The [59:50] advantage I have of reading, you know, [59:51] for the last decade of all like all [59:53] these biographies of entrepreneurs, like [59:54] you're studying somebody's life, but in [59:56] many cases if they you were so [59:57] successful, uh, they wrote a book about [59:59] your life, you usually it's almost like [60:00] you get a minor in like new industry [60:02] creation. Every single time it's just [60:04] like we're too late, it's over. There's [60:05] no more opportunity. And the way I [60:07] thought about this was like the the best [60:08] definition of a business I ever heard [60:09] actually came from Richard Branson where [60:11] he says all business is an idea that [60:12] makes somebody else's life better. And [60:14] if that's true, which I believe it is, [60:15] then that means there's infinite [60:16] opportunity in the future now and in the [60:18] future because there's infinite ways to [60:19] make somebody else's life better and [60:20] that's all business is. So this idea [60:22] that's like we got to the end of history [60:24] is just [ __ ] [60:25] >> Yeah. [60:25] >> Uh there and you know I don't mean to [60:28] push you on this like I am cur [60:30] personally curious though because like I [60:31] know you're already rich. Uh, I don't [60:33] think money is your north star based on [60:36] the conversations we've had in the past, [60:37] but like there's got to be some crazy [60:39] number somebody can throw at you where [60:41] you're just like, "Fuck, I have to take [60:43] this." [60:45] Not really on. I mean, it's like the, [60:49] you know, people have asked me sometimes [60:51] before, they've asked me like, um, okay, [60:53] but like really that would you guys like [60:55] >> this is what I'm doing right now. [60:58] And the way that I say it sometimes um [61:00] is like we would sell if we thought it [61:02] was the most ambitious thing to do. It's [61:03] kind of an oxymoron because obviously [61:05] but but but you know it's like my [61:07] genuine answer in the sense that like [61:08] that's like what we care we care about [61:10] you know it's it's like the I mean it's [61:14] funny you talk about money like I mean I [61:16] I don't even dude I don't have like I [61:18] don't have a car. [61:20] >> You live in an apartment. I just I just [61:21] realized [61:22] >> I have rent apartments. Yeah. It's like [61:23] I don't know. I think I like eating [61:25] sushi. That's fun. As soon as she [61:27] doesn't cost that much, you can do that [61:29] off of an engineer salary as well. [61:31] >> Just so you know, if this is true, like [61:33] then you are the the type of [61:35] entrepreneur that I find the most [61:36] fascinating in the world cuz like when [61:39] startup founders talk come talk to me, [61:41] it's just like I don't give a [ __ ] about [61:42] your startup. I like I asked the same [61:44] question. It's like is this your last [61:45] business? [61:45] >> Yeah. [61:46] >> Right. I asked Kareem from ramp like [61:47] that's before we did this like deep [61:48] partnership [61:49] >> and it's just like [61:50] >> it's like okay you could sell it like [61:53] the Zuck example where they're like why [61:54] didn't you take a billion dollars I [61:56] think you own 25% of the company you [61:57] would have made 250 million you're like [61:58] 22 he's like well what would I do with [62:00] the money I would just start another [62:02] social network I kind of like the one I [62:03] have like I just want to build [ __ ] [62:05] anyways so like what do I do and then [62:07] the other element of this which I think [62:09] is almost tied into you where it's like [62:11] I feel like you're just having a lot of [62:12] fun like even being around here it's [62:15] just like There's your company's weirder [62:18] in the composition of the people because [62:20] it's literally just all nerds [62:22] >> like and I think you like Yeah. Know [62:24] what I mean? But like you know like [62:25] there's a usually a mixture of obviously [62:27] the nerds and some other people and like [62:28] you kind of built this like almost like [62:30] your social network is like [62:32] >> I'm not here if you're not a nerd. Yeah. [62:33] >> Yeah. It's like a physical manifestation [62:34] of your social network but like the way [62:35] I way I would describe this is like [62:37] >> Okay. Cool. Cool. No. So my friends are [62:38] all also [62:41] enduring way. [62:42] >> I love that. Yeah. But what what what [62:43] the the people I'm most interested in it [62:45] is just like there's no price, right? [62:48] Like the example I use is go ask Steve [62:51] Jobs. I'll give you $2 trillion, but you [62:54] can't work on Apple. [62:55] >> Yeah. [62:55] >> What the [ __ ] do you think he would have [62:56] said? [62:57] >> Yeah, I agree. No, [62:58] >> no, he's like, well, I want to work on [62:59] Apple. That's what I want to do. There's [63:01] a funny thing that I wanted to tie [63:03] together which I didn't find the [63:04] opportunity till now [63:05] >> is you were currently the second [63:08] wealthiest entrepreneur from Baton [63:11] Rouge, Louisiana. [63:12] >> I don't know if you remember we've [63:13] talked about this. [63:14] >> We talked about this. Yeah. Yeah. [63:15] >> Yeah. [63:16] >> The first one [63:16] >> Chicken Figure is great business, but [63:18] I'm a customer of that that business as [63:19] well. [63:20] >> The first one is Todd Graves. Yeah. [63:21] >> Owns Raisin Canes. [63:23] >> Yeah. [63:23] >> I I talked to him. He was on the show. [63:25] Yeah. [63:26] >> You tell him you'll give a hundred [63:28] billion dollars for your chicken finger [63:29] empire. [63:30] >> Yeah. He No, he there's no He's turned [63:32] down and I know this for facts. I've [63:33] talked about it. He's turned down crazy [63:35] acquisitions offers. He's like, I don't [63:36] care about it's not the money. [63:38] >> Yeah. No, I mean for us it's very like [63:41] >> like again I this not rational but but [63:44] like the way that I would describe it is [63:46] like [63:48] I feel and I think all of us do like [63:52] that [63:55] it'd be one thing if we tried and we [63:57] gave it our all and we just weren't good [63:58] enough that'd be fine. Like, okay, it [64:00] wouldn't be that, dude. I'd be I'd be [64:02] salty as hell, you know? It wouldn't be [64:04] fine. But, but like, but like it would [64:06] be like it would be an outcome. It would [64:08] be like an outcome that I could live [64:09] with, you know? But if we felt like, [64:12] >> you know, we we could have gone for it [64:14] all. We could have pushed harder and we [64:16] did it. Like that I think is like I just [64:19] I I I don't think we would like live [64:20] with ourselves in that outcome. And [64:22] that's like the if you have me explain [64:24] it's it's kind of circular. I don't [64:26] know. But but it's like why why are we [64:29] so excited to do this? Why do we do, you [64:30] know, spend all it's like we want to [64:32] achieve our potential and and build what [64:34] we were meant to build, you know, and [64:36] and and maybe that's something, maybe [64:37] that's nothing, but like you'd rather [64:38] find out and see. Yeah. [64:40] >> I think this idea of you have one life, [64:43] go. [64:43] >> Yeah. [64:44] >> Is a perfect place to end. Thanks for [64:45] the time, Scott. [64:46] >> Yeah. Thanks for having me. [64:47] >> I hope you enjoyed this episode. Please [64:49] remember to subscribe wherever you're [64:50] listening and leave a review. And make [64:52] sure you listen to my other podcast, [64:53] Founders. For almost a decade, I've [64:55] obsessively read over 400 biographies of [64:58] history's greatest entrepreneurs, [64:59] searching for ideas that you can use in [65:01] your work. Most of the guests you hear [65:03] on this show first found me through [65:05] Founders.