OpenAI Chairman Bret Taylor Says AI Agents Will End Manual Coding by 2026

OpenAI董事长Bret Taylor称AI智能体将在2026年前终结人工编码 随着自主智能体推动企业软件行业从按席位订阅模式转向按结果定价模式，硅谷软件工程师将在数月内停止手动编写代码。 在这次涉及广泛的交流中，科技界资深人士Bret Taylor探讨了自主AI智能体的快速演进及其对客户服务和软件工程带来的变革性影响。他强调其公司Sierra如何利用基于结果的定价模式，将软件价值与实际业务成果挂钩，而非拘泥于席位数或Token等指标。Taylor详细阐述了组织为何必须围绕端到端数字化流程重构、打破传统的孤岛式组织架构，以释放真正的AI生产力。此外，他还分享了在Twitter和OpenAI备受瞩目的董事会危机期间管理公司治理的独到内幕反思，同时预测人工编码很快将被自动化系统设计所取代。 开源项目“OpenClaw”通过在本地Markdown文件中持续记录并压缩笔记，展示了一种实用的“记忆碎片（Memento）”风格AI记忆机制，与主流大语言模型对话白板式的特性形成了鲜明对比。 (1:24) 代码仓库非常适合AI，因为上下文保存在纯文本文件而非二进制格式中，并通过编译器、单元测试和版本控制提供了结构化的反馈回路。 (2:04) Harness工程（围绕智能体构建由规范、文档和规则组成的结构化支撑环境）正变得比复杂的多智能体架构或Anthropic的模型上下文协议（MCP）更为有效。 (5:55) 软件工程的最终产出可能会从编写“定制化、手工作坊式”的代码，转变为撰写如PRD和系统设计等持久性文档产物，而瞬态的代码生成则交由AI处理。 (5:55) 未来理想的SaaS界面将从面向人类优化的可视化仪表板，演进为充当自主系统操作指南的“智能体支撑环境（agent harnesses）”。 (11:55)

Bret Taylor argues that generative AI is transforming enterprise software from passive databases into autonomous, process-driven agents, rendering traditional seat-based licensing and manual software coding obsolete. He contends that long-term value will shift toward durable documentation artifacts, end-to-end organizational processes, and outcome-based pricing models that align vendor revenue directly with resolved business tasks. From Hand-Crafted Code to Documentation Harnesses (5:55) Taylor observes that rapid advancements in coding agents are fundamentally altering software development, moving engineers away from handcrafted, "artisanal" code toward system design, validation tests, and context harnesses. Because code generation is becoming cheap and disposable, the primary durable intellectual asset becomes the original intent—such as product requirement documents, architectural decisions, and integration guidelines. In this paradigm, software engineering increasingly resembles writing comprehensive documentation that agents consume to generate transient code. For agent memory and execution environments, Taylor favors flat directories of plain-text markdown files and standard Unix utilities over complex protocols like the Model Context Protocol (MCP) or rigid multi-agent orchestrations. He argues that isolated sub-agents often strip the overarching context required for cohesive decisions. Looking forward, Taylor forecasts that by 2026, most software companies in Silicon Valley will no longer write code by hand, relying almost entirely on agentic generation. The Transition to Outcome-Based Enterprise Pricing (60:50) As AI agents perform substantive labor, the historical defensibility of traditional enterprise "systems of record"—such as ERP and CRM platforms—is eroding. Because language models can extract and harmonize information across disparate, legacy databases without requiring unified schemas, value is migrating away from passive database custody toward the active agents orchestrating complex workflows. To reflect this shift, Taylor advocates for outcome-based pricing over conventional seat licenses or token usage tiers. At Sierra, customer support engagements are billed only when an issue is resolved without human intervention, while escalated cases are free; sales agents are compensated via commissions. Taylor likens this model to digital advertising's evolution from impressions to cost-per-click, noting that charging for business outcomes forces software providers to maintain accountability across the entire operational implementation rather than offloading execution risk onto the customer. Agent Economics and Real-World Case Resolution (24:27) Deploying conversational AI across voice and text dramatically changes the baseline unit economics of customer interaction, collapsing support costs from $10–$20 per human phone call down to $0.10–$0.20, and eventually to pennies. Across Sierra's enterprise clients—which include Cigna, SoFi, and SiriusXM—autonomous agents routinely resolve 70% to 90% of incoming inquiries without human intervention, driving Sierra to approximately $165 million in annual recurring revenue within nine quarters. This level of automation triggers several secondary effects. While eliminating repetitive tier-one inquiries raises the average handle time for remaining human-staffed escalations, it also increases staff job satisfaction by leaving more stimulating, complex problem-solving. Furthermore, some enterprises experience Jevons paradox: when customer interactions transition from rigid, decision-tree chatbots to responsive LLM agents, total interaction volumes increase by two to three times as customers actively seek out engagement rather than avoiding support channels. Reorganizing Companies Around End-to-End Processes (73:50) Taylor maintains that the atomic unit of AI-era productivity is an end-to-end business process rather than individual headcount. Corporate inefficiencies, such as vendor onboarding cycles that span weeks across legal, procurement, and IT, persist because organizations are structured into departmental silos without single-threaded accountability. Redesigning workflows around agentic execution allows these multi-department tasks to compress from weeks into hours. This structural evolution favors "hyper-generalists"—cross-disciplinary product engineers who combine technical proficiency, product sense, and design taste. Taylor cautions, however, that while lean engineering teams can outpace bloated hierarchies, extreme operational austerity carries risks: a firm that scales back staffing too aggressively risks losing market share to competitors who deploy slightly larger teams to ship faster. Addressing his governance role on OpenAI's reconstituted board, Taylor highlights how managing a nonprofit entity bound by a fiduciary duty to humanity's benefit fundamentally alters strategic decision-making compared to traditional profit-maximizing boards. While rapid commoditization means specific AI capabilities will continuously be absorbed into foundation models, Taylor argues that enterprise defensibility resides in navigating messy integrations, compliance requirements, and customer workflows. A critical unresolved question remains whether incumbent SaaS giants can transition their business models to outcome-based agent architectures before agile startups establish market dominance.

Sierra 创始人兼 OpenAI 董事会主席 Bret Taylor 认为，AI 正在从底层重塑软件工程、企业软件商业模式和组织协作架构，未来技术竞争的重心将从模型参数转向业务流程封装，软件的价值核心也将从被动的“记录系统”转移至直接交付业务成果的自主智能体。 软件工程重构：从手写代码转向线束工程与文档驱动 (1:24) Taylor 指出，编程智能体在过去数个月内经历了跃迁，彻底颠覆了传统的软件开发认知。当前开源智能体如 OpenClaw 展示了文本记忆的强大效果：系统通过向 Markdown 文件按时序记录笔记和简单整理，便能利用成熟的 Unix 工具（如 grep）实现有效检索，其表现甚至常优于复杂的向量数据库或多智能体层级架构。Taylor 对诸如 MCP 等复杂的外部工具协议及层级式多智能体架构表达了审慎态度，认为后者过度割裂上下文，反而削弱了全局决策能力。 在此背景下，“线束工程”（Harness Engineering）正成为核心实践，即为智能体构建包含产品架构、业务规则与执行约束的结构化环境。Taylor 坦言，告别手工雕琢代码的过程在情感上颇为艰难，但这正如编译器取代人工汇编语言一样不可逆转。他幽默地提出，软件工程未来的最大讽刺在于，开发人员的最终产出可能不再是会被重构或删除的代码本身，而是工程师过去最厌恶撰写的规格说明书（PRD）与架构文档。 智能体经济学：成本坍塌、杰文斯悖论与成果定价 (20:15) 在企业应用层面，Sierra 协助 Cigna、SoFi、SiriusXM 和 Rocket Mortgage 等企业部署面向客户交互的智能体。Sierra 的业务在短时间内快速扩张，其年度经常性收入（ARR）在 7 个季度达到 1 亿美元，8 个季度达到 1.5 亿美元，并在随后的季度首月达到约 1.65 亿美元。在实际运营中，典型客户的智能体全自动解决率介于 70% 至 90% 之间（例如 Ramp 达到 90%），并将 SoFi 的净推荐值（NPS）提升了 33 个点。 技术应用大幅降低了单次交互成本，将人工客服单次通话成本从 10 至 20 美元压降至 10 至 20 美分，未来预计降至 1 至 2 美分。成本骤降引发了杰文斯悖论：部分零售客户由于体验提升，总体交互量反而增加了两到三倍。对此，Sierra 推行“成果导向定价”（Outcomes-Based Pricing）而非席位制或 Token 消耗量计费：若智能体在无人工介入下独立解决工单，客户支付固定费用；若转交人工则不收费。Taylor 将这一商业模式变革类比为在线广告从展示付费（CPM）向点击付费（CPC）的跨越，迫使软件供应商与客户实现利益绑定并承担实施责任。 企业软件护城河演变与传统记录系统的重压 (53:31) 针对公开市场对 SaaS 行业普遍下调 20% 至 30% 估值的现象，Taylor 分析认为，传统企业软件巨头（如 SAP、Salesforce、ServiceNow、Workday）的核心护城河建立在其作为部门“记录系统”（Systems of Record）的数据库垄断地位上。以往所有应用交互都必须依赖其中心数据库，但由于大语言模型具备解析非结构化数据和跨系统调度能力，系统间的数据孤岛优势正在削弱。当执行具体流程（如线索生成、合同审核）的智能体能直接产生业务增益时，价值重心正向动作与流程层转移。 然而，Taylor 强调所谓周末即可“凭感觉编程”（vibe coding）出企业级应用的观点脱离了现实。软件公司的深层护城河不仅是前端表单，更在于底层的合规要求、安全性、欺诈防护、复杂的企业集成以及庞大的渠道分销与社会证明。保守的企业采购者通常希望成为成熟技术的验证者而非首个试验品，这使得既有企业的规模和分销网络依然具备防御力，尽管它们正面临商业模式自我颠覆的严峻竞速。 组织重组：从职能部门转向端到端流程与超能全才 (73:50) Taylor 主张，企业应停止将 AI 简单视作个体的替代品，而应将“业务流程”视为 AI 生产力的基本原子单元。传统组织按职能划分部门（法务、财务、IT），导致一个简单的供应商入驻流程耗时长达 17 天；若打破部门墙、由 AI 贯穿端到端流程，该周期可压缩至 17 小时。这要求企业重新设计组织形态，摆脱单纯向全体员工发放 Copilot 的初级做法。 伴随工程实现门槛的降低，拥有深厚客户同理心和设计审美、兼具工程理解的“高自主权全才”（Hyper-Generalist）正在重新崛起。结合 Elon Musk 在接管 Twitter 后将核心工程团队规模削减 80% 至 85% 仍能维持运转的案例，Taylor 指出，团队规模与工程产出历来不是线性关系。更扁平的组织和高度赋权的通用型人才将借助 AI“外骨骼”爆发出极大的生产力，但企业也须防范过度精简导致落后于速度更快的竞争对手。 OpenAI 治理反思与 2026 年技术跃迁预测 (95:22) 回顾在危机后接任 OpenAI 董事会主席的经历，Taylor 说明自己是在 2023 年 11 月风波平息后受邀担任调解人并协助重组董事会。他强调 OpenAI 的特殊结构在于其受非营利董事会管辖，董事的受托责任并非针对股东利益最大化，而是确保通用人工智能（AGI）造福全人类，这种根本使命彻底改变了董事会的议事逻辑与权衡维度。 针对 2026 年的技术演进，Taylor 提出了三项明确预测：首先，AI 将促成一项能够引发大众广泛共鸣的重大科学突破（类似于“阿波罗登月”时刻），从而将公众视线从生存风险与就业恐慌引导至科技改善人类境况的积极面上；其次，能够长时间自主运行的智能体将在消费级与企业级场景中实现大规模普及；最后，在技术应用最前沿的硅谷，绝大多数公司到 2026 年将不再依靠手工编写底层代码，软件构建将全面由模型生成驱动。 Taylor 的核心论点展现了应用层 AI 正在逼近的拐点：底层模型的通用化正在迅速吞噬早期的临时工程方案，逼迫软件企业放弃对专有代码的执念，转而在深度垂直的业务流程、成果计费模式及分销合规上构建真实壁垒。未解决的挑战在于，传统企业软件巨头能否在维持现有高利润年金业务的同时完成向智能体架构的转型，以及在手工编码走向消亡的过渡期，企业如何化解深层组织重构所带来的摩擦与不确定性。

[5:55] Bret Taylor: It would be the greatest irony if software engineering agents made all of us just write documentation the whole time [21:29] Bret Taylor: Our whole view is that if we were in 1994 and you were doing Cheeky Pint about this internet phenomenon... we would be talking about, like look, this is your digital front door... on the information superhighway. [35:11] Bret Taylor: it's like a multiplayer game that we're all in right now. [43:30] Bret Taylor: If you imagine a reasoning system is right 90% of the time... and you have a supervisor that's right 90% of the time... if you chain them together, you get 99% effectiveness. [53:31] Bret Taylor: The way I think about it is agents are to some degree a system of record of a process of generating a lead or auditing your financials or reviewing a contract... Will those encoded, well-optimized processes start to have more value than the databases? [60:50] Bret Taylor: I think usage-based is like charging for storage or something. Outcomes-based is what business outcome is this agent designed to produce and did it produce it effectively? [65:03] Bret Taylor: If you talk to any company who's ever implemented an ERP system, it's like a multi-year process. It's invading Russia. You don't even remember why you're doing it midway through. You've gone through two CFOs and three CIOs by the time it's done. [83:47] Bret Taylor: Should I get an AI agent to do something? Am I wasting the next eight hours of my life? [90:24] Bret Taylor: no one really cares about enterprise software [95:24] Bret Taylor: to ensure that artificial intelligence general intelligence benefits humanity. [98:24] Bret Taylor: most companies in Silicon Valley won't write code by hand.

[0:01] Bret Taylor is the ultimate Silicon Valley veteran. [0:03] He was one of the creators of Google Maps, invented the Like button, [0:05] was co-CEO of Salesforce. [0:07] He pushed through Elon's acquisition of Twitter when he was on the Twitter board. [0:10] He's now the chairman of the OpenAI board. [0:12] And his day job is founder and CEO of Sierra, which is bringing [0:15] AI to customer service. [0:16] He's one of the smartest people I know on the topic of how AI is [0:19] changing established companies. Cheers. [0:20] Cheers. [0:26] Most important question, have you installed OpenClaw on your work laptop? [0:30] I have not. Or have you played with OpenClaw? [0:33] I have played with OpenClaw. [0:35] I haven't bought a Mac mini. [0:36] You can put these things in virtual sandboxes for less money. [0:41] It's really interesting. I mean, it's very compelling. [0:43] It's probably the first... [0:45] I wouldn't have predicted the first broad... [0:48] I don't know if consumer is exactly accurate, but maybe hobbyist use of AI [0:53] would have been this semi-rogue open-source project that goes [0:58] through three name changes in three days. [1:00] And I love it. [1:02] I love everything about the chaos of it just because [1:06] people in our circles have been talking about AI agents for consumer use and all [1:11] these fancy computer-using agents. [1:14] Instead, you're chatting over WhatsApp with a thing on a Mac mini [1:18] that is mildly unhinged and insecure. [1:22] It's just fascinating. The whole thing is fascinating. [1:24] The thing that seems to me is funny is if you look at the landscape, [1:30] still in 2026, if you open a new Gemini chat or if you open a new ChatGPT chat, [1:36] it's basically a blank slate. There's no memory. [1:38] Then, I mean, Claw, [1:40] people talk about the WhatsApp and Telegram integrations [1:43] and things like that. [1:43] But it feels to me a big part of the value is not only can it do stuff [1:47] proactively, but it has memory. [1:49] But the way it has memory is this super janky... [1:52] It's like the movie Memento. [1:53] It writes things to a markdown file, and it's just writing the things [1:58] to remember, and the compaction is kind of buggy. [2:00] It didn't always write down the exact right things to remember [2:03] and stuff like that. [2:04] But isn't it funny that you can get super polished mainstream consumer apps [2:09] that have no memory at all, or this wildly insecure three-name-changes [2:14] project that almost remembers things by scribbling notes in the margin, [2:19] and that is the state of consumer AI. [2:22] I have a, probably not very thoughtful, but technical theory on this. [2:29] Coding agents to have gone through transformation over the past four months. [2:32] The difference between if we were here in October versus now, [2:36] our conversation about the future of software engineering would [2:38] be materially different. [2:40] How often can you say that about a technology? [2:42] People always, in my circles anyway, [2:46] you look at a coding agent and you extrapolate to other domains. [2:49] You're like, "Could all digital tasks be like this?" The answer is [2:55] obviously yes, over some period of time. [2:58] But it's really interesting because sometimes, I think the hard part [3:02] of engineering is in the details, and code repos have very specific qualities. [3:05] One is all the context is in one place, in files that are largely [3:10] textual, not binary. [3:12] For most broad information tasks, that's not true. [3:17] When you're writing your annual letter, my guess is the sources of information [3:21] were in so many different systems, data warehouses. [3:25] It's not impossible for an agent to use those things, but the idea that you can [3:29] straight-line from coding agents to writing the Stripe annual [3:33] letter, I don't totally buy. [3:35] Then similarly, when the agent's actually performing work on a code base, [3:40] there's feedback, there's compile errors, there's often unit tests, [3:44] there's integration tests. [3:46] There's the history of every change that we made in a really formal format, [3:50] along with code reviews. [3:54] It's almost designed for a robot, and you can self-reflect. [3:57] Maybe we as engineers have always modeled ourselves after robots, and now we can [4:02] actually fully realize that vision. [4:05] What's interesting about it is the idea that it wrote a markdown file for memory, [4:10] I think is maybe more significant than a hack. [4:12] I actually think to some degree— [4:15] Turning your life into code? [4:16] Yeah, it's like you almost want to put everything in a file system [4:21] that looks like source code, not because that's the only way these [4:24] agents can work, but actually it's quite an efficient way to get a mix [4:28] of context and random access memory. [4:31] If you think about a vector database, it's more random access. [4:34] You have to know what to look for. [4:36] But actually, that's not how real memory works. [4:38] There's a mix of it. [4:39] You're loading a markdown file, and as you said, compaction, all these things matter. [4:44] But the messiness of it actually probably produces a more useful agent [4:48] than a lot of the fancier things. [4:50] I use memory in ChatGPT, and I love it. [4:52] But I actually think this idea that there's a directory of just [4:57] everything you've ever done is actually maybe more useful to an AI [5:01] than people think. [5:02] Actually, if you follow over the past couple of months, [5:05] just this emergence of harness engineering, [5:08] where you're building the harness around an agent to do work, [5:13] I wonder if in the short term, it might just be one of those [5:16] idiosyncrasies of history, like mimicking a code base is actually [5:20] the best way to make a general-purpose agent work. [5:23] Maybe over time, we'll get fancier than that, but it's actually a relatively [5:27] efficient harness for an agent. [5:29] Anyway, maybe that's why it's— [5:31] It's very terminal-centric. [5:33] It's backwards compatible with— [5:36] You can use grep. [5:36] You don't need to make some vector database. [5:39] The AIs really know how to use all the Unix tools. [5:44] You got a lot of lift from that. That's exactly right. [5:46] Software engineers were notorious for making tools for ourselves first, [5:49] so then we just almost bend every other domain in digital towards that domain. [5:55] But the reason I brought up things like unit tests, integration tests. [6:00] OpenAI did this blog post—I can't remember the engineer who did the post—but [6:04] on harness engineering. [6:05] One of the more interesting parts of it was documentation. [6:08] Rather than just having a single agent's markdown file, they had a directory [6:13] of essentially the entire product, the architecture, and they're filling this [6:17] out over time, and agent's markdown became pointers to it. [6:22] My hypothesis, having used Codex a lot, [6:26] I wonder if the output of a session where [6:29] you make a change to Stripe's products should be a documentation artifact [6:34] in addition to code, where the documentation artifact is [6:37] actually what the product manager version of John, and the code was the engineer [6:42] version of John, where there's a lot in the code that is more transient. [6:46] You might be fine to delete that. [6:48] What was the intention? [6:49] What was the PRD? [6:50] What was the customer problem? [6:52] Is actually the more durable asset. [6:54] I wrote this on X, and one of the funniest comments, it would be like, [6:57] "It would be the greatest irony if software engineering agents made all of us [7:02] just write documentation the whole time just because notoriously [7:05] every good engineer hates writing documentation, and now that's our job." [7:09] But I don't know, it resonated with me. [7:11] How much are you AI code... [7:14] You're a very prolific engineer in the old-fashioned, [7:17] hand-spun way of writing code. [7:20] How has that changed? Bespoke artisanal code. [7:21] Exactly, yeah. Pour-over. [7:24] That's funny, I would use that. [7:29] I am trying to get to a world where I'm not writing code. [7:32] It's hard, emotionally, if that makes any sense. [7:35] I have a hard time not caring. [7:37] I don't care about the assembly language produced by the compiler. [7:45] Why should you care about the code? [7:46] Why should I care about the code? [7:49] I care about correctness, I care about robustness, [7:51] and I think I know intellectually, I don't need to look at how the compiler [7:55] unrolled this loop to verify its elegance and correctness. [8:01] Yet somehow I feel that way about code. [8:05] I'm not saying the code doesn't matter, but I've been trying to force myself [8:08] to not care because I feel like I won't be a self-actualized software engineer [8:14] in the future if I'm too precious about that artifact which used [8:18] to be so central to me. [8:20] Right now, writing markdown files, maybe that's fine. [8:22] It feels somewhat like a local maximum, and maybe we'll just be like, "Oh, [8:26] of course, markdown is how we work with machines." [8:30] If you think about what a compiler does, there's this interesting mix of [8:35] formality and informality, and if you've used Python versus Rust, [8:39] different ends of the spectrum. [8:40] Now that you're not writing the code, I really wonder what that programming [8:44] system should feel like and look like. [8:47] I don't mind chatting with Codex, that's fine. [8:51] But I also think, as you imagine, all the tests that you care about, [8:56] it showing you demos and mockups. [8:58] I wonder what the future integrated development environment, for lack [9:01] of a better term, will be in that world. [9:04] What I'm trying to do is force myself to not be emotionally attached [9:07] to the code, which is very hard for me because that was my entire life before. [9:11] I was proud of the elegance of the code that I wrote. [9:15] But if I still care about the craftsmanship, what do I want? [9:19] I haven't quite visualized it yet. [9:22] It feels to me like a very interesting time in agentic engineering because you [9:26] were talking about this domain of harness engineering and people having skills [9:30] and MCP and everything like that. [9:31] It's always interesting when not only is the leading [9:39] product in a category changing, we're actually just figuring out [9:42] what the categories are that we need things for MCP or skills or stuff like [9:47] that, and it's all very fast-moving. [9:49] That just feels to me like a very interesting time where clearly a new [9:52] way of engineering is shaking out. [9:55] 2026 is clearly not the final word. Absolutely. [9:58] In fact, I'm growing more skeptical of MCP as a meaningful part of the future. [10:04] It's fine as a protocol, but it's interesting. [10:07] Going back to your joke around OpenClaw just writing a big markdown file. [10:12] I think it works better than a bunch of MCP servers. [10:16] But going back to the point of every AI agent knows how to use grep and knows how [10:20] to use all these things, I feel like this view of a multi-agent [10:24] world was you have all these agents that do tasks that are fraud detection, [10:30] another one over here for personalization. [10:33] Then you make a super agent, it does all these things, [10:36] and it looks really good on a whiteboard like most elegant-looking but completely [10:42] nonsensical architectures. [10:44] Then you realize if you just imagine you anthropomorphize the Stripe experience [10:49] and you're the checkout concierge. [10:52] What information do you need to have no a priori to actually make [10:56] that a humane experience? [10:58] What ends up happening in these multi-agent systems is you stuff all [11:01] the context in the subagents, and the one on top has no ability [11:04] to actually not sound robotic. [11:07] Then in contrast, you look at something like OpenClaw, [11:10] it's just a bunch of markdown files, and the memory feels right, [11:14] even though it's a little bit kludgey. [11:17] Similarly, if you go back to my arguments about a source control, [11:22] like a repo, it has so much context. [11:24] It's not like you just have the myopic view of the file you're editing. [11:28] It really has some expansiveness. [11:30] My sense is we're making true agents over time. [11:33] The way we think about context and how that context is shared so that the agent [11:39] that's orchestrating it actually understands what's behind all these APIs [11:42] and why and the history will maybe look a little bit more like OpenClaw [11:46] and less like MCP over time. [11:49] I think these agents need a lot more context than what MCP affords. [11:55] One thing we've noticed... [12:00] There's a bit of a what's old is new again phenomenon where with this agentic [12:03] commerce stuff that's happening. [12:04] We actually built the APIs for this 10 years ago as part of... [12:08] Do you remember that move of social shopping that was in for a while? [12:11] Buying on Instagram, buying on Twitter. [12:14] It didn't quite happen for a few different reasons at the time. [12:19] But the concepts are very similar that you want some action at a distance, [12:22] you want to be able to go manipulate stuff off-site. [12:25] Similarly, I think Patrick has wanted for the longest time in Stripe is [12:31] the ability to just SSH into your Stripe account. [12:33] What do you mean? [12:34] It's a very ergonomic way for developers to work is you should just be able to log [12:38] into your Stripe account and you have a command line there, [12:41] and then you can list out all your... [12:44] Or you can tail the payments log. [12:45] He wants tail and pipe— [12:47] Exactly, all these things. [12:49] Of course, now we're building that because it's much more relevant [12:52] in the agentic world. [12:53] But I find somehow, all the agentic stuff is also bringing back... [12:58] I don't know if you have this experience as well. [12:59] It's bringing back a lot of ideas that you might have had before. [13:02] Well, it is because to some degree, the elegance of Unix, [13:04] which has been the basis of why everyone wants SSH and the curl command that was [13:09] famously on the Stripe homepage. [13:11] For people who got it, [13:14] it was remarkable because you could have all these tools that did something well [13:18] that was small and useful, and you could chain them [13:20] together to make something great. [13:23] I wonder in the future, we've talked a lot about this, [13:26] if you look at the canonical software-as-a-service application like [13:29] Stripe's console, and obviously you have what consumers see, but the configuration [13:34] that a Stripe customer will log into, you would have a web app, [13:38] and that's the forms and fields and buttons and graphs, [13:42] and then you have the API. [13:44] It was typically like a REST API or GraphQL API, and you [13:48] could do stuff with it. [13:49] This is how computers talk and this is how humans used it. [13:53] I wonder if the web application of the future will actually be... [13:58] Certainly, you want a web app for the rare human who wants to sign in, [14:02] but will you have an agent harness? [14:04] What I mean by that is something more than the APIs, but just like if you think about [14:09] the harness that you provide in a code base, the skills, [14:12] the documentation, the rules. [14:14] Imagine the person who's the greatest Stripe expert who knows how to extract [14:20] the most value from their Stripe account. That's the harness. [14:24] Not the API. That's the button you click. [14:28] Will that be an endpoint on Stripe. [14:30] com, and so that your agent knows how to just get the most value from Stripe? [14:35] I imagine you don't care if your merchants log in. [14:39] What you want them to do is drive value for themselves, drive GMV, drive payments. [14:46] I'm really excited about that because an API is great. [14:49] APIs are awesome. [14:51] But a harness is basically like, here's the instruction manual for all [14:54] the Unix commands that power Stripe. That's very interesting. [14:58] Yes, and I think if you look at the shape of a lot of APIs that services have, [15:02] and I think Stripe's API coverage is probably more complete than most, [15:05] but ultimately, it is a way to manipulate some [15:09] of the highest-value business objects in the thing. [15:14] Whereas actually what you want is, one, all of the data to be browsable in some [15:20] agentically accessible or textual way, and then all of the actions to be [15:25] able to be taken by agents. [15:28] It turns out there's a lot of switches in the dashboard that there's no API for, [15:32] and we are all as an industry collectively discovering it. [15:35] It might be easier. [15:36] Imagine being a product manager in the future. [15:38] You just need to add the switch to the dashboard. [15:40] You're like, "Yeah, it looks like a Russian submarine [15:42] to switch this, but who cares?" Agents can handle it. [15:45] As long as the harness describes when to use that switch, [15:49] it's easier than UI design in some ways. [15:52] That's fascinating to me. [15:55] One funny point Dario made is it's not clear... [15:58] There's a race between people getting their stuff accessible via agents and just [16:04] desktop computer use getting better. [16:07] It's actually not clear, will the approach be, [16:10] Stripe builds way more APIs, and that's how your agents manipulate [16:16] the Stripe account, or you just give [16:19] your agent access to Chrome and a login. [16:23] Well, so actually, I'll give a funny story here. [16:24] Sierra, my company— [16:27] Sorry, we'll get to Sierra. No, it's fine. [16:29] But it's a real funny story here. [16:31] Sierra powers a lot of healthcare companies. [16:34] On the healthcare payer side, health insurance. [16:36] Well known for their API quality. [16:38] First of all, they're actually pretty sophisticated engineers [16:41] at these companies. I really enjoy working with them. [16:43] You end up with Cigna, Blue Cross Blue Shield on the healthcare [16:46] payer side of insurance. [16:48] Then you have healthcare providers like Sutter Health who we work with. [16:51] Then you have revenue cycle management. [16:53] R1 and revenue cycle management basically help providers get paid [16:57] by the insurance companies. [16:58] Then you have a lot of other people in the middle. [17:00] Pharmacies, PBMs, and they all call each other. [17:04] So a healthcare provider has to call a payer because a procedure [17:08] happened and they have to get paid. [17:10] We have payers with AI agents that pick up the phone. [17:14] And we have providers that have AI agents that pick up the phone [17:19] and make phone calls. [17:19] We have revenue cycle management companies that work to make outbound calls to do it. [17:23] We've already had— [17:25] Do they switch to the agent language? They don't. [17:27] We've done English over the publicly-switched telephone network. [17:30] You have TCP/IP and English over PSTN. [17:35] It reinforces, I guess, Dario's point, which is you can engineer all these fancy [17:40] protocols, but the rails that are already there already exist. [17:42] English is spoken by all AI agents, and the public switch telephone network [17:47] has been around for 100 years, and it all works, which is fascinating. [17:52] You have all these fancy MCP things, and we're doing English over PSTN. [17:56] On one hand, I think I actually agree with the principle that one [17:59] of the powerful parts about AI, with its ability to do text, do audio, [18:04] and I'll say, I'm not sure how you qualify computer use, but you can call it a form [18:10] of image recognition and manipulation. [18:13] Certainly that's useful because you get to the point where you don't need to fully [18:16] finish the last mile to get value. [18:19] The thing I'd say, though, going back to talk about all the actions, [18:23] and they're not all in the product, all the APIs don't exist. [18:27] These visual interfaces were designed for us. [18:29] Think of... [18:30] Stripe is, I think, famously, it was one of the few enterprise [18:34] software companies with good design for a long time. [18:37] The Stripe dashboard is really elegant, right? [18:39] And most enterprise software, you can't say that about their dashboard. [18:43] I don't think the ideal agent harness will be that elegant because it's [18:49] optimized for something else. [18:51] It's optimized for the context that you need to perform complex multi-step [18:56] procedures on behalf of a person. [18:59] And my guess is it's just very different. [19:02] I think seeing the way you write a harness for a software or a coding agent is just [19:08] so different than the way you do UI design. [19:11] I'm certain that it's great that you can click around a green screen or [19:15] whatever—click around a green screen because it's an oxymoron—but [19:18] type around a green screen or click around a legacy on-premises [19:22] enterprise software system. [19:24] I think these harnesses will be really good. [19:26] I wonder if, is there a world two years from now where [19:31] Stripe's ability to work with the agent that manages commerce for a direct [19:38] consumer e-commerce company, that will be one of the ways you're evaluated. [19:43] In fact, if your... [19:45] For lack of a better word, harness is not compatible with the way [19:48] their agents work, that's actually like you're not compatible with them. [19:53] I'm not sure that's right. [19:55] I think it's great that these things are backwards compatible. [19:58] It's great that our agents have spoken over the telephone already in English. [20:03] But I don't think it's the long-term future because there's so much [20:06] value that you can provide. [20:07] Put another way, the agents using a sophisticated [20:10] application harness can just do a lot more and do a lot more with higher fidelity. [20:15] Yes. [20:15] Well, we should get to Sierra because you see a lot of real-world AI adoption. [20:19] Maybe start by grounding us. What is Sierra? [20:21] The business has scaled very quickly. [20:24] One of the latest metrics that you can share because they keep [20:28] changing from month to month as you guys grow. [20:30] So at Sierra, we help companies make AI agents for their customer experience. [20:34] If you have a big phone line, these AI agents can replace your IVR [20:38] system and just pick up the phone. [20:40] If you have a digital chat system, an AI agent can pick it up. [20:43] You don't need to wait on hold. [20:45] These agents can not only answer questions, but take action on your behalf. [20:49] We work with healthcare companies like Cigna. [20:53] We just did a great case study with SoFi, and I'm really proud that we raised their [20:58] net promoter score by 33 points just because it's just so delightful. [21:03] It's really fun to see all these different brands across a wide range of industries [21:08] get so much value from their agent. [21:10] We're the leader in the space. [21:12] You talked about the metrics. [21:13] We reached $100 million in ARR in seven quarters, $150 in eight quarters. [21:18] We're, I think, around $165 now, one month into our next quarter. [21:22] So growing really rapidly and really proud of the momentum that we have. [21:27] That's super cool. [21:29] What is the typical adoption? [21:31] Are people using it for email chat support because that's the easiest modality? [21:35] Do they adopt it for everything, including phone and stuff like that? [21:39] It's changed a lot over the past two years, but I'll say the median customer, [21:43] and then I'll describe some interesting outliers, which I hope are [21:46] glimpses of the future. [21:48] Most will start with one channel and a few use cases. [21:51] At a lot of healthcare companies, phone remains the dominant channel. [21:55] So say, "Hey, for a few types of phone calls, [21:58] let's have the AI agent take them and see how it does." Do people like it? [22:02] Are people comfortable with it? [22:04] Does it lower our cost? [22:05] Does it raise whatever metrics? [22:07] Usually, it's customer satisfaction. [22:10] Does it work more effectively? [22:12] For example, for a car insurance company, it'll be a first notice of loss. [22:16] I got in a fender bender, and that'll be the typical way you start. [22:20] For a lot of more digitally-native companies, they'll start with chat, [22:25] and similar. [22:28] But almost all of our clients will do both. [22:31] SiriusXM, if you call them on the phone, their AI agent, Harmony, which I love that [22:35] name for SiriusXM, will pick up the phone. [22:38] If you go to their homepage and you see the chat, that's also the same agent. [22:41] The neat part is, I think it's pretty neat because you have [22:44] literally all of your, I'll say, customer experience team or whatever [22:48] you might call it at your company. [22:50] They can spend all their time on one thing, and it actually works over [22:53] WhatsApp, it can work online, it can work in your website, it can work [22:56] in mobile app, it can pick up the phone. [22:58] That's a pretty big change. [22:59] A lot of our clients, when we start working with them, [23:01] they'll have a digital team and a call center team and all these different teams. [23:05] We've gotten to the point because we've digitized the last remaining analog [23:09] channel, which is the telephone, those are all unified. [23:12] When I look at a glimpse of the future, we have a few really ambitious customers [23:17] like Rocket Mortgage, great Detroit company. [23:20] They own Redfin. [23:21] They bought a mortgage services company called Mr. Cooper. [23:25] If you go to Redfin, you can search for a home using an AI agent. [23:28] If you go to Rocket. [23:29] com, you can originate a mortgage with an AI [23:32] agent, and you can service that mortgage. [23:33] It becomes product usage rather than just customer service. [23:36] Really end-to-end. Sales, service. [23:39] I think that's really exciting. [23:41] Our whole view is that if we were in 1994 and you were doing Cheeky Pint [23:45] about this internet phenomenon. [23:48] That would have been fun. I was a bit young, but yeah. [23:49] Yeah. I would have my Nirvana shirt on. [23:55] We would be talking about, like look, this is your digital front door, [23:59] or maybe we wouldn't have the prescience to say that. [24:01] On the information superhighway. On the information superhighway. [24:04] I think the same is true of most companies, AI agent singular. [24:08] There are lots of agents, but the one with your brand at the top [24:12] that your customers interact with is special, and that's the one we're [24:15] trying to power for companies. [24:17] You think this, what customer is built on Sierra, [24:22] your aspiration is that it just becomes sometimes the primary way [24:25] people deal with the company? [24:27] I think a company's AI agent will be the vast majority of their [24:31] digital interactions. [24:32] I think digital has come to include the telephone. [24:35] That's a big shift because we think of that differently. [24:39] That's a huge change just because the bigger shift... [24:44] Customer service, which is one big part of what we do, but not the only thing we [24:48] do, has traditionally been thought of as a cost center because it's really expensive. [24:52] I'm sure you have people answering the phone for your clients, [24:55] and depending on where they're located and how well-trained they have to be, [24:59] like how simple is the case? It can cost $10, $20. [25:02] It can be much less if it's a more simple case. [25:06] You have some customers who pay you millions of dollars, [25:10] and you'll answer the phone anytime they call, and you might have one that has [25:13] not even started monetizing yet. [25:16] You might want to call them, but there's a limit to literally how much [25:19] you can afford to talk to that person and still have a profitable business. [25:24] I always joke, it's probably easier for you and me to call Sundar than to get [25:27] Google customer service on the phone. [25:28] It's very hard to get Google customer service on the phone. [25:31] But it's not because they don't like you. [25:33] It's just if you think about the average revenue per user of Google, [25:37] they literally just can't afford to do it. [25:40] Now, if you take that $10 or $20 phone call and you make it 10 or 20 cents, [25:44] and over from one to two cents. [25:46] All of a sudden, not only can you afford to provide a great customer experience [25:51] to more people, even less profitable customers or [25:55] in lower-margin businesses, which I think is very exciting. [25:58] It's not just doing what you did before but new. [26:01] You can provide better customer service. You really can. [26:03] Then just think about running a subscription business where you care as [26:07] much about customer acquisition, you care a lot about churn because that's [26:09] how your lifetime value equation works. [26:12] You think about, "Okay, if I had a budget of how much I spend [26:15] on service, and now I can do 100 conversations [26:18] more than I could before, can I actually reduce my churn rate? [26:22] Can I improve lifetime value?" Then the interesting thing is then you realize [26:26] that, "Wow, all of my competitors have access to the same technology." And then [26:30] you're saying, "Okay, what are my competitors going to do to [26:33] actually take my customers away from me?" [26:36] Then that's where you start to get things like, [26:39] the ATM machine didn't actually reduce bank branches because some bank had [26:43] the great idea of, I'm going to put different people in this branch. [26:46] It'll generate revenue. [26:47] All of a sudden, it wasn't job displacement, [26:50] but something completely different. [26:52] I think the exciting part in our world is you're taking something that was just so [26:56] expensive that people literally hid their phone numbers so people couldn't call [27:01] them, and you're making it inexpensive and delightful. [27:04] The thing I'm excited about is the second and third-order effects are going to be [27:08] really interesting and very hard to predict, and that's pretty exciting. [27:14] I want to come back to this idea of the agent as the UI because [27:17] I find it really interesting. [27:18] We talked about this a bit in our annual letter in the context of agentic commerce, [27:21] where, again, I think people are trying to pitch too much of the end state [27:25] of fully autonomous, the robots just choosing a few. [27:28] The point we always make is, let's just start with not [27:31] having to fill out the web form. [27:32] No one likes filling out forms on the internet. [27:34] Speak for yourself. I wonder just, will using websites [27:41] have been actually a bit of a... [27:43] It's like the fax machine. [27:44] We used to hum emails over the telephone lines as a way [27:48] of transmitting information. [27:50] I wasn't working for this, but there was an era of voicemail memos. [27:55] Were you ever in the working world for that? [27:56] Some people still do this where they do the voice. [27:59] Companies will blast a voicemail memo to employees at the company. [28:03] That's a way of distributing information. [28:05] All these things are very moment in time. [28:07] Maybe navigating websites and filling out forms was a bit of a moment in time. [28:12] Is that how you see things playing out? I don't know. [28:14] It's interesting because if you look at the past few iterations of technology, [28:20] you had the PC revolution, then you had the internet and the browser, [28:24] then the smartphone came out and the tablet. [28:28] I was more optimistic about tablets than the way the world turned out. [28:33] I see more tablets on airplanes, but I'm guessing if I walked around Stripe, [28:37] I would see very few tablets out. [28:40] Similarly, there's more smartphones than people, but there's still about [28:43] 2 billion PCs in the world. [28:45] I think it peaked some number of years ago, [28:47] but it hasn't gone down as far as I know, and I haven't tracked this. [28:50] That's interesting. We added to our digital world. [28:54] But I think perhaps the more interesting metric is for you and me, what percentage [29:00] of emails were sent through each device? [29:02] Certainly from 2010 to 2020, most of the world might have transitioned [29:08] from percentage of email on desktop to smartphone significantly. [29:15] It's almost like market share of digital interactions, which I think is a really [29:19] interesting way to think about it. [29:21] Certainly, as you think of Stripe's business, where does commerce originate? [29:24] You saw that move to mobile. [29:25] But it doesn't mean that people… [29:27] It's actually a very big… [29:28] You wouldn't want to eliminate the PC commerce business. [29:32] That would actually be catastrophically bad. [29:36] Then you look at AI agents, and I believe most businesses, it will be [29:40] their primary digital interface. [29:42] It's because it works over WhatsApp and it works over the phone. [29:45] If smart speakers make a comeback, they'll work over smart speakers. [29:49] Which they may now. Maybe smart speakers were just too early. [29:51] Yeah, well, it's like ask for the weather just turns out to be not [29:54] the biggest market in the world. But now— [29:56] Set a timer. [29:58] Set a timer. [29:59] I mean, it's amazing they made I'm not much money off a timer-setting speaker. [30:03] It is very future-proofed because it's fundamentally conversational. [30:09] But maybe it's going from punch cards, mice and keyboards, touch screens, [30:14] now voice and chat and probably 3D-immersive at some point. [30:19] Does it just add and make the other ones less important is probably [30:23] the way I think about it. [30:24] I do wonder if we'll see the end of the smartphone at some point. [30:27] It doesn't seem anywhere close to right now, but it is interesting. [30:32] I think most people don't love how much we're addicted to staring [30:36] at this glowing screen. [30:37] On the other hand, you can't talk to TikTok. [30:41] It's fundamentally visual. [30:43] But I wonder if there's a world where you could actually be really productive [30:47] without such an invasive device on your body. [30:51] If that's the case, can it offer an opportunity to [30:55] unwedge some of the addictive properties of these technologies and get [30:58] a lot of the benefits from it. [31:00] Because at least for me, I think all of us are so connected, [31:03] you end up like, "I'm going to check my email," and then [31:05] you're like, "Where have I been for the past hour?" [31:08] The fact that we actually have technology [31:11] that affords that kind of innovation now, I think that's quite interesting. [31:14] I don't know what the future is, but I'm very excited for it. [31:17] I know that sounds really cheesy, but we've now changed the ingredients [31:22] available, and we have a lot more recipes we can cook, and I think [31:25] that's very exciting. I agree with you. [31:27] I'm excited for not having to look at the screen for as many [31:30] things for a variety of reasons. [31:33] When a customer installs Sierra, [31:38] I know there's a significant customer [31:41] satisfaction component as well as cost, but I'm curious what [31:45] cost difference does it make and maybe relatedly, when [31:49] a customer is fully deployed, what kind of mix do they see between [31:52] queries fully resolved agentically, things that end up having a human who is [31:56] presumably somewhat AI-assisted, but just what does a normal [32:00] equilibrium look like there? [32:01] It turns out most of our clients have pretty different priorities. [32:06] Some are very focused on cost savings, and you can automate [32:09] very high percentages of your cases. [32:12] There's a company called Ramp that's a really impressive tech firm. [32:17] We just had Eric here. Oh, that's great. [32:19] Well, they're automating 90% of their cases. [32:23] They're really sophisticated, though, because they're basically getting [32:27] in front of cases before they escalate. [32:29] But I think it's an example of just a really fantastic company [32:33] implementing really well. [32:35] You can see anywhere between 70-90%, which is really incredible. [32:41] The interesting thing, though, is there's counterintuitive effects to it. [32:46] The cases that do make its way to your customer service team can end [32:49] up more complex, by definition. [32:51] What's called average handle time will actually go up. [32:55] We heard from one of our clients that actually their satisfaction of their [33:00] call center agents went way up, too, because it turns out it's way more [33:03] fulfilling to solve a hard problem. [33:05] Totally, than like "Have you tried plugging it out and plugging [33:06] it back in again?"Exactly. [33:08] The other interesting one, we had one retailer whose [33:13] total volume went up almost as much as they saved from the AI agent. [33:19] Jevons Paradox. It was a form of that. [33:22] If you've used a chatbot from three years ago, they were really annoying. [33:27] Three years ago, if you said, "Do you like chatbots?" Zero people would say yes. [33:32] It's so funny that there was a Silicon Valley wave of hype around chatbots. [33:36] It was even earlier than that. It was like 2018. [33:37] It was like pre-LLMs, pre-transformers. [33:40] They were just multiple choice machines or something. [33:43] It was just the worst products of all time. [33:46] Replacing it with something that was delightful. [33:49] People are like, "I'm going to talk to this thing [33:51] a lot more." They ended up keeping... [33:53] Their cost didn't really go down, but their volume of customer [33:56] conversations went up 2 or 3X. [34:00] The CEO was incredibly happy about it. [34:02] They were like, "We're now actually listening to our [34:05] clients." It sounds funny, but it's a little bit of a choice how much [34:09] you want to drive cost savings with AI versus other metrics. [34:14] Most of our clients are interested in the top-line metrics. [34:18] Given if you could save $10 or save $1 and improve your net promoter [34:25] score and competitive positioning by a meaningful amount, everyone [34:30] in the world would choose the latter. [34:31] That's the interesting thing going on right now because, again, [34:34] going back to, it was in 1994, and we're talking websites on this podcast. [34:40] I think if you were to go to a major bank and say, "If you launch a website, [34:45] you're going to have a competitive advantage against every other bank." [34:48] With the benefit of hindsight, that would have been over-promising, [34:51] the correct thing to say was, "If you don't launch a website…" [34:54] Totally, you have to have a website. [34:56] This technology is broadly available. [35:00] As a consequence, you can't just launch… [35:04] In all parts of AI, not just our business, you can't just launch AI, [35:07] absorb the cost savings, pass it on to shareholders, unless you have a monopoly. [35:12] In most businesses, it's just a consumer surplus. [35:14] Exactly. You're either going to lower prices. [35:17] But I think that's why it's an overused analogy. [35:20] But the ATM bank branch thing is really interesting because [35:24] if every single company in an industry has access to technology, I would say it's an [35:28] imperative, not a competitive advantage. [35:31] The more interesting, I would say, board discussion is when everyone adopts [35:35] the obvious things, customer experience, customer service, [35:39] software engineering, legal. [35:40] Just pick the ones where there's off-the-shelf solutions available now. [35:45] What will the industry look like? [35:46] My guess is you could ask ChatGPT, think. [35:51] My guess is there's some really interesting second-order effects. [35:54] When you have competitive markets, you're going to end up investing, [35:58] lowering prices, whatever it may be. [36:00] And that's the thing I don't think gets talked about enough. [36:02] I actually think that it happens with every technology change. [36:07] You project it through the lens of what you're doing today, [36:09] and you don't take into effect. [36:11] It's like a multiplayer game that we're all in right now. [36:15] That's fascinating to me. [36:17] The change is disruptive, but I think it's going to be like... [36:21] I'm very excited for the next few years as the world absorbs the technology, [36:26] we start getting to some of those second and maybe even third-order effects. [36:30] What's the most impressive AI adoption or AI native behavior you've [36:33] seen from a client? [36:36] Oh, that's a really good question. [36:38] I'll probably say Rocket, where we have a really great relationship. [36:43] Varun is their CEO, Shawn Malhotra is their CTO. [36:48] Two people who really are, I would say, not only just curious about AI, [36:54] but very interested in transforming the home ownership experience with AI. [37:00] I don't know when you got your first mortgage, but it's super... [37:04] It's very intimidating. [37:05] It's not a modern process. [37:07] It's not a modern process. [37:09] They literally call it mortgage folders for a reason. [37:12] It used to be a folder. [37:15] For me, it's an example of a company trying to transform an industry. [37:19] The reason I brought it up in the context of our last question is it's not just [37:23] saying, "How can we take AI to do this?" But it's like, if you were to think about [37:26] the homeowner experience from searching for a home on Redfin all the way through [37:30] servicing it, and you had AI available, what would the ideal experience be like? [37:35] It's just really interesting to see Rocket with their acquisition strategy to [37:40] integrate that experience. That's what I'm excited about. [37:43] I think there's an opportunity for CEOs and industries like that to have a bold [37:49] vision of what the future could be. [37:53] Going back to my point, imperative and not competitive advantage, [37:56] it is a competitive advantage right now. [37:58] If you imagine, I haven't tracked the market share of all [38:02] the big US telcos, but if you look at T-Mobile, Verizon, AT&T, [38:07] and you tracked it over the past 10 years, you end up with surges [38:11] in market share growth. [38:12] The iPhone came out, you ended up with 5G, and you end up with these things. [38:18] But my impression of the industry is you end up with these moments that drive [38:23] market share, and then it ends up at an equilibrium. [38:26] What's interesting about it as the iPhone moment for telecommunications [38:30] companies like SoftBank in Japan. [38:33] This is the moment where perhaps if you have a competitive equilibrium, [38:36] you can absorb this technology, use it, and you'll have this window where you can [38:40] actually shuffle the deck a little bit. [38:42] This is a technology that shakes the competitive equilibrium. [38:44] Yes, exactly right. [38:46] We definitely notice that. [38:47] You talked about how coding is such [38:52] a domain that is suitable to AI because [38:55] all of the context you're working with exists in the repo. [39:01] It is in text. [39:02] It's neatly organized to be executed and read by humans, [39:07] and so there's a good balance there. [39:09] The problem is that customer service [39:13] agents are not of that character. [39:17] How do you actually [39:21] smush everything into a format where [39:26] your AI agent can answer it? [39:30] We spend a lot of time thinking about that. [39:31] To some degree, one of our engineers called it almost like [39:34] we're creating a domain-specific language for specifying customer experience. [39:39] What is the mechanism of specifying it? [39:44] We use this metaphor we call journeys, which is what is a customer journey [39:48] end-to-end, and what does the agent need to be successful in that journey? [39:52] What tools does it need to access? [39:55] What information does it need to access? [39:58] If you think about the capabilities of an agent, [40:01] like skills in a coding agent, you'll add different capabilities over [40:06] time as the customer is talking to you. [40:08] The key thing that's been a breakthrough that is probably not surprising [40:12] to the technologist listening to this, but has been a huge difference between [40:16] those crappy chatbots of four years ago is the reasoning capabilities. [40:24] We had one client who had acquired three companies, and they had [40:28] three identity systems, three CRM systems, three of everything. [40:31] They had this big IT project where they were going to unify all those systems. [40:35] But I was like, "Why don't you just have the AI agent go in all three [40:39] of them and just think?" [40:40] They're like, "Well, what if there's duplicate data? [40:41] What if the data conflicts?" They're like, "That's going to fool—" [40:44] Sensor fusion. [40:45] I was like, "Well, what does a person do?" They're like, "Well, [40:49] they think about it." I was like, "Let's just do that." [40:53] That's the interesting thing about these AI agents is they actually... [40:56] The basic humane, basic reasoning, not superhuman ASI, [41:03] turns out to be the huge breakthrough in customer experience. [41:06] The other interesting thing is the innate knowledge of the LLM. [41:09] You don't want an AI agent to hallucinate, obviously, but Sonos [41:14] is one of our clients. Do you have a Sonos speaker at your house? [41:17] You probably have a Sonos. I have had, yeah. [41:19] If a Sonos speaker ever breaks, it's never the speaker. [41:21] It's always Wi-Fi. That's what I've learned. [41:24] It's always true of me, too. [41:25] There's always some Wi-Fi issue. [41:27] If you wanted to make an AI agent to help you with your Sonos speaker, [41:32] you obviously can give it all the manuals for the speakers, all the technical stuff. [41:36] You can give it the device telemetry, all the stuff you need. [41:39] Do you really need to give it the history of Wi-Fi? [41:41] Well, now it turns out large language models have encountered [41:44] every possible Wi-Fi problem. [41:46] Why does the Sonos AI work so effectively? [41:49] Well, it knows a lot about Wi-Fi in addition to all the Sonos things. [41:52] If you look for any given AI agent, it turns out being trained on all of human [41:59] knowledge is useful as a starting point for a lot of tasks. [42:03] I think that's been the big breakthrough. [42:05] How do you give it all of its knowledge? [42:06] Well, first, we've built, I think, the best platform in the market to do so, [42:10] where you can really narrow the guardrails for regulated conversations, widen [42:13] them for less-regulated conversations. [42:15] But the fact it starts with knowledge of obscure Wi-Fi idiosyncrasies turns out [42:20] to be the greatest breakthrough of all that. [42:21] Have you had the opposite problem where there's a customer whose problem domains [42:25] mostly don't exist in the public internet? [42:27] It's like, "We provide the drill bits used in deep sea oil drilling." It just turns [42:33] out there's nothing on Reddit about that. Yes, 100%. [42:37] We work with this medical device company, and it's a deep cut of human knowledge. [42:42] You can train it all on that. [42:44] In fact, we do a lot. [42:46] One of the things you want to be really careful about is you have [42:48] a really well-known brand. We work with... [42:50] I want to say a third of our clients have over 10 billion in revenue, [42:56] over half have over a billion in revenue. [42:58] Most of our clients are actually quite well-known. [43:00] One of the challenges when you're offering either sales or service or customer [43:06] experience to a really well-known brand is it's harder to ground it. [43:08] It's actually easier when the internet has never heard of you and you want to make [43:13] a well-grounded agent, it's actually pretty easy because [43:15] there's no temptation from the LLMs to go off-script. [43:19] Actually, I would say— They know what they don't know. [43:21] Ironically, the harder challenge is when it's a very well-known brand, it's like, [43:24] "No, I got this." I'm like, "No, you don't. [43:26] You got to go look it up." That's actually a harder problem. [43:30] How do you force the LLMs? [43:32] Mechanically, how do you force them to not answer off the top of their head, [43:37] but actually look it up? [43:39] We call it a constellation of models. [43:43] Our platform, we call it Agent Studio, you essentially configure the goals [43:47] and guardrails of a process. [43:51] Goals and guardrails, not the sequence of steps because you want [43:54] agency but you want guardrails around it. [43:57] Within that, we'll use reasoning, but we use supervisor models [44:01] to actually inspect that reasoning. [44:03] If you were an AI agent in Sierra and you [44:09] decided to go off-script, "I got this." What would end up happening is [44:14] a supervisor agent would observe your reasoning, [44:17] say, "I think John should have actually looked up the policy here" and send it [44:21] back with notes and say, "Actually, you're not allowed to make that decision. [44:25] Here's the reasons why. [44:27] Go redo that decision." It's a really effective technique. [44:31] The way I think about it, which is a little simplistic, [44:34] but I think basically right. [44:35] If you imagine a reasoning system is right 90% of the time but has some [44:39] either guardrail, malfunction, or hallucination 10% of the time, [44:43] it's obviously better than that. [44:45] Then you have a supervisor that's right 90% of the time. [44:49] If you chain them together, you get 99% effectiveness. [44:52] That methodology of layering reasoning and intelligence has been really effective. [44:58] In general, it makes sense. [44:59] You're basically layering, compute, you're layering reasoning on top of it. [45:03] What's the neat about it, though, is we can abstract [45:06] that complexity from our clients. [45:08] They're expressing the goals and guardrails, and we have all these [45:12] evals and tests and all these other things. [45:15] We can find ways to make it more and more robust over time, but it doesn't require [45:19] you to prompt engineer, write in all caps or whatever the [45:24] hacks that people use to get these things to be conformant. [45:27] You started in '22, '23? [45:31] We launched the company on February 13th, two years ago. [45:36] I guess a little more like our—Oh, '24. Yeah. [45:39] Our two-year birthday was a couple of weeks ago. [45:42] What I was thinking as you were saying that is, did you co-evolve chain [45:47] of thought and RL and some of these things that are now in the models, [45:50] but did you have to build your own janky version of them before [45:54] they were in the models? [45:55] Yes. [45:56] I'll also talk about just the weird part about building a product right now [45:59] in a company right now, because so much of what we write, [46:01] we plan to throw out later. [46:03] It's just a very weird way to build a company. [46:06] Google's chain of thought paper, which preceded o1 and doing reinforcement [46:12] learning on chains of thought, was out roughly when we started the company. [46:15] It was an earlier paper and effectively provided a substantive basis of why asking [46:21] a model to explain its reasoning step-by-step produced more robustness. [46:26] We use chain of thought all the time, and it was a methodology we used. [46:31] Then OpenAI very innovatively came up with the idea [46:36] of we could do reinforcement learning on those chains of thought, [46:39] which is where o1 came from, and then most labs are doing that now. [46:44] We'd throw out things all the time. [46:46] You do it and you're like, "Okay, the model just does this for us now." [46:50] We work with a lot of financial services firms. [46:52] We work with one bank that is a large Hong Kong business, [46:56] and they speak Cantonese. [46:57] Okay, well, we need really good Cantonese voice support. [47:01] It turns out that that's really hard, and there's not an obvious [47:04] model that does that. [47:06] We spend all this time evaluating all these models. [47:09] What certainty would you ascribe to every voice model [47:13] supporting Cantonese well in three years? 100%? [47:16] 99%? Pretty close. [47:17] We did all this work. [47:18] In fact, we, I think, have the best Cantonese support on the market. [47:22] Great for us, and it's a huge selling point. [47:24] It's a technology that will certainly be commoditized in three years. [47:29] A lot of what we think about is going from essentially technology innovation now. [47:36] I think a large part of why we work with the largest companies in the world [47:41] is because our technology works. [47:43] In three years, those same clients will work with us because we [47:46] have the best product. [47:48] If you look at the early marketing for early software-as-a-service companies, [47:53] they'll explain why having multiple tenants in the same database is safe. [47:58] That was a huge part of their marketing. [48:01] Nowadays, if you came and you marketed your product that way, people are like, [48:04] "What are you talking about? [48:05] I don't care what database Stripe uses." I think we're just at this period [48:10] where the technology is so immature. [48:12] It's a very technology-forward conversation just because [48:17] people are figuring it out. [48:19] Just like when Netscape's business was monetized through a web [48:22] server back 100 years ago. [48:25] It will evolve from being a technology-forward conversation [48:28] to a product-forward conversation. [48:30] The interesting part about building an applied AI company is you can't have [48:33] the luxury of waiting for all the models to catch up with your aspirations. [48:37] But you know they will. [48:39] You have to have the best technology and have to be comfortable [48:43] with throwing it out. [48:45] It's a real momentum and pace of innovation game rather than thinking [48:49] of this as precious intellectual property, if that makes any sense. [48:52] It absolutely does. [48:53] But isn't this organizationally hard where, [48:59] if I'm the head of Cantonese language [49:06] at Sierra, my incentive, [49:09] and not disingenuously so, I'll notice all the corner cases where the [49:14] models aren't that good at Cantonese yet. [49:15] Obviously, we saw this in prior tech waves, where the cloud adoption laggards [49:20] were companies that had their own on-prem stuff, and they had a million reasons—half [49:25] real, half fake—as to why cloud did not suit their business purposes. [49:29] But how do you avoid getting stuck in this mode of thinking where, "Oh, well, [49:33] their chain of thought doesn't do what we need," is the classic thing you would [49:37] hear from someone within the organization. [49:39] It's a huge shift. [49:40] Going back to the first thing we were talking about, it's hard for me to not [49:44] care about the elegance of the source code, which I think [49:50] is an impediment to my fully realizing being a software [49:54] engineer in this new world. [49:56] I think teams that start to treat the code that they wrote as precious, [50:02] that has been obviated by a general-purpose AI model, [50:06] will fundamentally fall behind. [50:08] Public markets deem the software industry 20-30% less valuable than [50:15] they did maybe three months back. A day ago. [50:18] Yeah, exactly. Very recently. [50:21] The two sides of the debate are, one, the valuations were based on what [50:28] the businesses will do in in 2030 or 2035, like far in the future, [50:32] and just there's much more uncertainty there, and so this is deserved. [50:35] The counterargument is that it's still not the case that agentic software production [50:42] is really going to build you a Workday. [50:44] Indeed, Anthropic just installed Workday, very famously. [50:48] Where do you net out on, is this a rational response or not? [50:53] I think it's rational, but I think it's a bit overblown at the same time. [50:57] I think it's rational just in the sense that there [51:03] hasn't been more uncertainty in this market ever. [51:06] Unless you have a strong thesis about an individual company, [51:12] my guess is, will these companies be less valuable 10 years from now? [51:17] I think the answer is probably yes. [51:18] Will that be true for every individual company? [51:20] I don't think that's true. [51:22] If you're just thinking about a portfolio of investments, I think it's an indictment [51:27] of the sector more than it is an indictment of an individual company. [51:31] I don't know if the value of these platforms was who could vibe [51:34] code it in a weekend ever. [51:37] Not that we knew what vibe coding was. [51:40] My point is everyone who's ever built a software-as-a-service application has [51:44] had a Hacker News comment of, "I could have coded this [51:46] in a weekend." Every single one. [51:48] Famously, Dropbox, I'm sure you have as well. [51:50] Every single product I've ever made. It just happens. [51:53] It's like a write-up. [51:54] In fact, if no one says it about your product, I'm sorry. [51:56] It's not relevant. It's not relevant. [51:59] Obviously, I see most of those comments were incorrect, [52:02] but if you think about all the work you've done in compliance or the relationships [52:06] you have with large financial services institutions, the work you do on fraud, [52:10] the things under the surface that aren't the forms and fields in the web browser [52:15] are actually incredibly valuable. [52:17] If you think about a large software company, they'll have thousands [52:22] of quota-carrying account executives that represent sales [52:24] capacity, which is basically a channel, and distribution turns out to be [52:28] a very important part of software. [52:31] There's social proof. [52:33] There's the old saying, "No one gets fired for buying IBM," [52:36] which few people say right now, though IBM is actually doing [52:40] really well under Arvind. [52:42] You want to be maybe the first healthcare insurance company to adopt something. [52:45] There's another healthcare insurer who says, "I want to be the fifth. [52:49] I want other people to prove it." There's all these network effects around these [52:54] businesses and scale and moats and Silicon Valley speak around them. [53:00] I think the big risk is where is value in the software industry n years from now? [53:05] One risk is that more people will build than they do now versus buy because [53:12] the marginal cost of writing software goes down. [53:15] I think that'll be true for some software, particularly developer platforms [53:19] and things like that that are already being consumed [53:23] and purchased by other engineers. [53:26] Little libraries or— [53:28] Things that already were part of a build-versus-buy calculus, [53:31] it shifts the balance of power. Absolutely. [53:33] The other part of it is systems of record. [53:37] I think these systems of record have always been the gravitational center [53:42] of their relative solar systems, and it roughly breaks down by department. [53:47] ERP systems are associated with the finance department, and SAP [53:50] and Oracle and Workday have ERP systems. [53:53] You have Adobe in the marketing department, and you had Salesforce [53:57] in the sales department, and you had ServiceNow in the IT department. [54:02] Everything rotated around them. [54:06] Why? [54:07] Well, first, their database was truly the system of records. [54:10] Every application that wanted to interact with the data in that had to... [54:16] You essentially collect taxes from your ecosystem. [54:19] And then similarly allowed each of those systems of record company to essentially [54:24] have revenue expansion opportunities to go to adjacent areas where they're all [54:28] sold to the same buyer and all of that. [54:32] The thing that's really interesting is AI agents are actually [54:34] performing valuable labor. [54:38] Is the database in the system a record? [54:41] Does that continue to be the gravitational center of each of those workflows? [54:45] I'll just take marketing as an example. [54:48] The database of your customers that you use to drive sending out an email [54:54] blast on Black Friday has some value. [54:58] But if you had an AI agent that drove way [55:01] higher, more leads for your sales team [55:05] from that marketing blast, you probably... [55:08] That's worth more to you than the system of record itself. [55:12] Similarly, if you imagine, I'll just take a CRM system, [55:20] and you think about the AI agent that's carving your territories if no one ever [55:24] logs in to actually do it manually, all of those things have a lot of value. [55:29] And a lot more value, relatively speaking, than they did because they're [55:33] actually performing the action. [55:35] The real question to me is, does it upend this? [55:39] I would say something that's been true for 30 years, which is all the values [55:42] in these systems of record. [55:44] The way I think about it is agents are to some degree a system of record [55:48] of a process of generating a lead or auditing your financials or reviewing [55:53] a contract or whatever it might be. [55:55] I don't think we've ever had a piece of software like that. [55:58] Will those encoded, well-optimized processes start [56:03] to have more value than the databases? [56:05] I don't know that's the case. [56:07] For example, if your ERP system is your company's ledger, [56:11] that'll have a lot of value. [56:12] But I wonder for all these others, and my theory is, the closer you get to, [56:17] literally the database is the value, i. e. [56:19] a ledger, the more durable it is. [56:22] The closer you get to be in a system of engagement, the less durable it is. [56:27] That's a very interesting framing. [56:32] It gets back to the point you were making about the company that was looking [56:37] to standardize and not have three different ERPs and stuff like that. [56:40] You're like, "Why? [56:41] Just try not doing that." I think maybe the consumer example of this is I think [56:46] people have probably had the experience of you paste data into an LLM to do [56:49] something with it, and the formatting is all messed up and the tabs and the spaces [56:55] don't come through and everything like that. [56:56] It doesn't matter. [56:57] LLM doesn't care. [56:58] You can just paste it in whatever and it'll work with it. [57:01] This idea that, as you say, if the system of record is important [57:06] because it's your general ledger and it matters to the auditors, that's one thing. [57:10] But if it was a system of record in this all your data in one place [57:15] way and because it was easier to build incremental software atop [57:19] it, maybe that advantage is going away because the agents are fine plucking [57:25] data from 10 different places. [57:26] That's roughly my view. [57:28] But the bull case, I always mix up bear and bull. [57:31] I need to spend more time on Wall Street. [57:33] The bull case, though, is I think all these companies have a right to win. [57:37] They're all big, they still have sales capacity, they have all these advantages, [57:42] but it's a race. [57:44] How fast will smaller companies build differentiated, scaled businesses [57:50] before the incumbents go into this new world? [57:53] But for a wide variety of well-documented reasons, like disrupting their [57:58] own business model, it's harder. [58:00] But I think your ask is, "Is it irrational?" I don't [58:03] think it's irrational. [58:04] I think there's just more uncertainty now than there's ever been. [58:08] I think markets are telling you there's a lot of uncertainty, [58:11] and that's why you see people recede from the whole category, basically. [58:16] I feel like there's also a totally separate thing playing out here, [58:22] where for a long time, certain companies were criticized for not [58:27] taking profitability that seriously. [58:29] And at some level, there's just a return to normal valuation levels on a fully [58:34] loaded, stock-based, comp baked in and everything basis. [58:38] That's independent of the AI thesis, but maybe just some return to, again, [58:44] on a fully loaded GAAP basis, more normal valuations. [58:48] Well, essentially, if you look at a traditional [58:50] software-as-a-service company, the way most people model it is you have [58:55] annual recurring revenue, which is basically an annuity, [58:57] and it should throw off that much cash every year. [59:01] Then you have attrition, which is subtracting from the annuity. [59:04] Then you have net new ARR, which is adding to the annuity. [59:07] Your salespeople sell software to add to the ARR. [59:11] You typically have an account management team or customer success [59:15] team to keep churn down. [59:17] You grow that annuity and you grow your headcount, [59:20] often just a little bit ahead of that annuity because you [59:23] need to grow a new business. [59:27] If that annuity is not an annuity… [59:28] Yeah, that makes sense. [59:30] Then that math really changes. [59:31] It really changes. [59:32] Because the whole idea of software-as-a-service is you can just [59:37] slow down hiring and you become very profitable because the annuity [59:40] starts throwing off cash. [59:41] That's been the thesis of every private equity firm who acquires slow growth [59:47] software-as-a-service companies. [59:49] If you don't assume that that revenue is going to be there two years or three years [59:53] from now, your discounted cash flow analysis looks pretty different. [59:58] I don't think it's actually It's quite so dire in the time frames that people think. [60:02] But again, if you're asking markets are, there's all those great quotes [60:06] about weighing and I don't know. [60:08] Voting/weighing machines. Voting/weighing machines. [60:11] I get it. [60:12] There's probably more safer sectors to invest in, but I don't think it's [60:16] an indictment of individual companies. [60:17] That's my point. I actually think... [60:19] When we first met, I doubt either of us had an extremely [60:24] positive view of the future of Microsoft. [60:26] At the time, it felt a previous generation company. [60:31] Now you look at Azure, their OpenAI relationship, all these [60:34] things, what an impressive turnaround. [60:36] I think any one of these companies could do it. [60:39] I think it's just... [60:41] But it's more of an indictment of the market. [60:43] I have a lot more questions. Would you like another Guinness? [60:46] Sure. [60:50] Brett has been through a few platform shifts, and one thing he's been pretty [60:53] consistent about is being mindful of the external forces that are [60:56] shaping the ecosystem you're in. [60:58] He talks a lot about building with the broader wave of AI agents in mind. [61:02] Stripe Sessions is our way of helping builders see that wave up close. [61:05] What's changing in the internet economy, what's actually working in production, [61:08] and what the next era of software looks like when agents are running [61:11] real commerce workflows. [61:12] It's not the usual conference fluff. [61:14] It's insights into what the fastest moving companies are actually up to. [61:18] If you want to experience the next chapter of the internet economy firsthand, [61:21] join us at Stripe Sessions this April. [61:24] Use the code Cheeky Pint for 50% off a conference pass at sessions. [61:28] stripe. com. [61:35] You talked about business models. Are you guys usage-based? [61:37] How are you innovating on the business model front, or are you? [61:39] We are trying to. [61:41] We do outcomes-based pricing. [61:43] For a customer service context, that means if the AI agent resolves [61:47] the case, no human intervention, there's a pre-negotiated rate for that. [61:53] If we do have to escalate to a person, that's free. [61:55] For sales, it would be a sales commission. [61:57] And wherever possible, there's a way to align our interests [62:00] with our clients', we choose it. [62:02] I'm a huge believer in this. [62:05] I think the analogy of going from impression-based ads to CPC ads is apt. [62:11] I don't think any ad platform thinks like, "Man, think of all the impressions we're [62:15] giving away for free." Because when you charge for something closer to a business [62:20] value, it's actually more valuable. It's more efficient. [62:22] It's a lot more efficient. [62:24] I think the idea... [62:25] If an agent's outcome is measurable, it's a really a compelling way both [62:32] for clients, obviously, because it's aligned with their business. [62:35] But it's also quite disruptive because most, [62:37] I'll say, legacy software companies are not necessarily equipped to do it [62:42] for a variety of reasons I'm happy to go into, but it's just [62:44] a very disruptive model. [62:46] Yeah, there's a few lens you can have on it. [62:48] One is that you get more alignment, like usage-based is more aligned [62:52] than other ways of charging. [62:54] As you say, it's more efficient because you're [62:56] incentivized to drive the right outcomes. [62:58] People also make the analogies to it's almost more [63:03] correct for the labor substitution dynamics that you get, [63:06] or just because you have real inference costs, you kind of have [63:11] to do a usage-based model. [63:13] I mean, do those factor in at all, or it would not be possible, almost, [63:18] to do a fixed-price contract? [63:19] I would actually argue outcomes-based is pretty different than usage-based. [63:23] Just because think of it this way. [63:25] If you have an AI agent that is making sales for Stripe to small businesses, [63:31] and I told you I will sell one-tenth [63:37] the number of new Stripe GMV, [63:40] however you value that, but I'll use one one-hundredth [63:45] of the tokens, you probably wouldn't care. [63:47] You care about the value to your top line of your business. [63:51] I would argue there's not a strong correlation between token [63:54] usage or utilization and value. [63:58] There may be, but there's not always. [64:00] There was that infamous, I think it was called Folklore, [64:03] but it's this website where that Apple engineer used to put just [64:06] all this Apple folklore. Folklore. [64:07] org. Folklore. [64:08] org, yeah, I love it. [64:11] It's like, if you're an engineer, it's a fun site to go to. [64:13] But there was a story about some new bozo manager asking for lines of code every [64:18] day, and one of the engineers wrote a negative number as a way of saying, [64:21] "F you" to the man because he refactored a code base or whatever. [64:25] I think that is the essence of why tokens are not correlated with value. [64:30] They may be, but the idea that they definitely are, I don't [64:33] think stands to reason. [64:35] I think usage-based is like charging for storage or something. [64:40] Outcomes-based is what business outcome is this agent designed to produce [64:44] and did it produce it effectively? [64:46] That is really aligning because it creates this whole vertical alignment. [64:51] As a company, reducing your token utilization [64:55] for the same outcomes is your problem, not your customer's. [64:57] That's a great incentive to drive more efficiencies over time. [65:02] It means that to grow your relationship with the client, you actually have to make [65:06] your product better and not just theoretically better over a steak dinner. [65:10] Like better, better. [65:13] How do you have usage-based... [65:14] Or sorry, outcome-based when you move beyond customer service where there's [65:18] a clear, was this resolved or not, to product usage where people are shopping [65:23] and yeah, they didn't buy a house there, but they mostly don't buy a house on [65:28] most website visits, but it was a successful visit. [65:34] It's the right question, and there's not a great way to do it [65:37] for every type of agent right now. [65:39] You can always fall back to usage-based, which is fine. [65:44] But in that, over time, it's like, wouldn't it be interesting? [65:47] I think AI agents should have memory. [65:49] I think AI agents should drive relationships, not conversations. [65:54] It would be really interesting to say, could we make an AI agent that actually [65:58] drives home ownership over time. [66:00] I think that's actually... [66:02] It's hard, but it's not— [66:03] As an AI agent, you have a territory, kind of. [66:06] I think so. [66:07] I mean, even because it's hard today and [66:11] we're a pragmatic company, I think it's the right thing to ask, [66:16] though, because that's fundamentally the value the software is designed to produce. [66:19] Actually it's a really values-aligning thing. [66:25] It also changes the dynamics of a software company's relationship to its partners, [66:30] to its clients, because if you go back ancient history, [66:34] four years ago, there was a really stark separation [66:37] between software and implementation and usage. [66:41] It was the client's accountability to use the product well. [66:45] It was either your IT team or a systems integrator's responsibility [66:51] to implement the software. [66:52] The job of the software company was just to make it and throw it over the wall. [66:56] Obviously, it's not exactly that, but that was the market we were in. [66:59] Everyone had good intentions, but what's the saying, [67:02] "Success has a thousand fathers, failure is an orphan." [67:04] When the software didn't go well, everyone was blaming everyone else. [67:09] The client was like, "I'm using it just fine. [67:13] It was implemented poorly." The person to implement it was like, "No, [67:16] the platform's broken." The platform people would say... [67:19] It was like everyone's pointing at everyone else. [67:21] What's nice about outcomes-based, whether or not the client sets it up, [67:26] you become more accountable to help them be successful because until they do, [67:30] they can't use it. [67:31] If there is some long last mile of implementation, [67:36] creates a strong incentive for the software company to have skin [67:39] in the game to just help you navigate that last mile. [67:42] I think so many of the problems in the software [67:46] industry are due to that lack of accountability. [67:49] If you talk to any company who's ever implemented an ERP system, [67:52] it's like a multi-year process. It's invading Russia. [67:56] You don't even remember why you're doing it midway through. [67:59] You've gone through two CFOs and three CIOs by the time it's done. [68:03] We're like okay with that. [68:04] That's just the way software works. [68:06] My view is just like AdWords changed the advertising [68:12] industry on the internet, just drove it. [68:15] I think you can even pay for mobile app install now directly [68:18] and truly pay for outcomes. [68:20] I think it's a really positive step forward. [68:22] It's not going to be possible for everything. [68:24] You have to have a pragmatism. [68:26] But I think it's the right way to actually have a partnership. [68:29] You should share in the outcomes. [68:31] You want to wire the company to be thinking in this outcome-based way. [68:34] In your main customer service stuff, you can do that. [68:36] In other ways, you might not be able to yet, but you want people to be [68:38] spring-loaded to be thinking that way. That's right. [68:40] If the whole company is incentivized towards outcomes, it's... [68:45] We're a way better partner to work with because of it. [68:48] We find this at Stripe where, again, we have outcome-based pricing. [68:51] Exactly. [68:52] It's transactional. [68:54] But we find there's a lot of uplift we can get on just getting people more [68:58] revenue and finding ways... [69:00] We're sometimes hammering customers where it's like, "You should be accepting [69:04] local payment methods for internationalization," or "You're [69:07] crazy not to be turning on this feature." But we really feel it because we have [69:10] the same incentive as the customer. [69:11] It's like this will be revenue maximizing for both of us. [69:14] I'm going to ask a very AGI-brained question. [69:17] I just can't resist. I'm glad we're on our second Guinness. [69:19] Exactly, yeah. [69:20] Now we get to it, which is you described building stuff that you know you're going [69:24] to throw away because the model capabilities will get there [69:27] and you're occasionally... [69:29] They are developing capabilities that you developed yourself. [69:33] Isn't Sierra itself "short AGI?" Sorry, I said I couldn't resist. [69:38] No, it's the right question. [69:41] The short answer is I don't know. [69:42] I mean, the fog of war in the software industry is pretty thick right now. [69:46] I really believe in the applied AI market, though. [69:50] I think most companies don't want to buy models or buy software. [69:57] They want to buy solutions to their problem. [70:01] If you just go back to the cloud industry, why doesn't Amazon and Microsoft [70:07] do everything for everyone? [70:09] There's not really a... [70:11] By somewhat similar logic, why should any software-as-a-service [70:15] company exist when you have bigger scale, all this technology? [70:19] In theory, they could just develop all the software. [70:21] Actually, many of them have tried. [70:24] There's actually competitors to Salesforce and almost all the above. [70:28] I think there's so much nuance in how these companies align themselves [70:32] with different departments at these companies, solve their very unique [70:36] problems in very specific ways. [70:38] That is a mix of product, not technology, but product, go-to-market. [70:43] It's an ecosystem around it. [70:45] I think a lot of that still exists because I'm not sure coding the software [70:49] was necessarily the hard part. [70:51] Then similarly, I actually think, especially in enterprise [70:55] software, how you engage with your clients really matters. [71:00] I think it turns out that [71:04] GPT-5 and Claude, whatever version it's [71:07] on right now, or Opus, excuse me, is sold to a different buyer than the CFO [71:13] or the Chief Customer Officer, the Chief Digital Officer. [71:16] It seems small, but it's actually big. [71:19] I think you tend to see software companies orient around individual [71:23] buyers within companies. [71:25] You tend to see consolidation around departments and around buyers. [71:30] It's possible that you can go beyond those lines, but it hasn't [71:33] happened traditionally. [71:35] I think the reason for it is most business users want actual solutions to their [71:39] problems, and they want a company that serves their unique problems [71:43] in a very specific and bespoke way. [71:45] I actually am extremely bullish on applied AI. [71:50] I actually think we could accelerate. [71:52] I'll make one statement, which is I think if we paused model [71:55] development, we'd still have trillions of dollars of economic value. [71:58] I totally agree. That have yet to be realized. [72:01] I think if we had a mature applied AI market where the CFO could go buy [72:06] that agent to onboard new supply chain vendors that just worked, [72:11] we could actually accelerate that trillions of dollars of economic value. [72:15] I think not only am I somewhat skeptical that there will only be two companies [72:19] in the world, I actually think one of the main things [72:22] impeding adoption of AI is the lack of existence of all those other companies. [72:27] So many of the startups, particularly around here in San Francisco, [72:30] are basically doing relatively rote [72:34] tools around the AI rather than actually [72:37] building agents for business processes that are boring [72:39] but important and valuable. [72:41] I'm really bullish on it. [72:43] I guess you help companies ensure that they can always have access [72:47] to the latest models, which sounds like a minor thing, [72:50] but the leading model is always changing, and so that's not a trivial... [72:53] I agree. I don't know. [72:56] I'm not sure how much of a long-term value it is. [72:57] I think it is. I think your customer— [72:59] Up to this point, the race is led by a matter of months, right? [73:04] Well, every single month, there's a new frontier model, [73:08] and your customer experience doesn't change that frequently, [73:10] so you're absolutely right. [73:12] But I also think there's just a big product, like our clients use [73:15] it to optimize their sales. [73:17] That is a product, not a technology. [73:20] It's very particular to the workflows of people building customer experience [73:23] teams, building sales teams, and that's really what we're focused on. [73:27] I think those departments deserve purpose-built software, and I think [73:32] there will be enduring value there. [73:34] But it's interesting. [73:35] It's the right question to ask. [73:36] I don't think we've ever lived in a world where production of software was easy. [73:41] Software engineering was the most scarce asset in a company, and now it's the most [73:46] plentiful, and I don't think we've ever lived in that world. [73:50] Well, that gets to one of the biggest conundrums in Silicon Valley right now is [73:55] what will the shape of AI productivity be? [74:02] I think there's this strong sense that [74:09] the AI has gotten really good, [74:11] and it should change the composition of companies and it should [74:15] change the hiring plans somewhat. [74:17] You've seen this in some corners. [74:19] Block announced their 45%, [74:22] 50% AI layoff yesterday, and you have [74:27] some companies not growing as quickly. [74:29] At the same time in coding, [74:33] you see a lot of AI benefits. [74:36] You can argue that either way. [74:39] You can say, "Engineers have gotten much more [74:41] productive, therefore, we should hire fewer engineers." Or you [74:44] could say, "Engineers have gotten much more productive. [74:46] The ROI on a single engineer is way higher. [74:49] We now have super engineers that we can hire, therefore, we should hire way more [74:52] of them because there isn't a fixed amount of stuff for Stripe or any other company [74:56] to do." Then the AI productivity story in other roles is just a bit less clear [75:02] because, as we've discussed, AI is uniquely well-suited to coding. [75:08] What do you make of just... [75:10] How does the AI productivity show up? [75:12] I feel like every company in Silicon Valley is trying [75:14] to figure this out right now. [75:16] Well, first, I think I'll go back to my why I believe in applied AI. [75:19] I think the atomic unit of productivity in AI is a process, not a person. [75:25] I don't think AI... [75:27] I don't know if you have an assistant, but if you do, [75:30] he or she might help you prepare for a podcast, might help you prepare [75:36] for a meeting, he or she might also get you a cup of coffee. [75:40] AI will be really good at the first two, but quite poor at the last one. [75:45] No matter of AGI, short of robotics, will get you a cup of coffee. [75:49] I think it's wrong to think about AI as replacing people. [75:54] In addition to being inhumane, it's just nonsensical because AI operates [76:00] in the world of digital technologies. [76:02] I think if you go to an example of even a mundane process in your business, [76:06] like onboarding a new supplier, think about all the departments [76:11] and people involved in that. [76:12] There's a legal department to do a contract. [76:15] There's some finance department procurement to negotiate the relationship. [76:20] You probably have IT that's involved to onboard them into your core systems. [76:25] Then there's usually a business that's sponsoring it. [76:28] Fairly mundane, happens all the time. [76:31] Let's just say you tracked what is the median amount of time it takes [76:35] to onboard a new supplier, and it was 17 days, just for argument's sake. [76:40] I bet you could say as a CEO of a company, I want to use AI to optimize that process [76:45] and make it 17 hours or one day. [76:50] You could go through, and if you had a product manager [76:53] on that and optimize every part of it, I bet you could achieve that. [76:56] But the hard part isn't a person's job. [76:59] It's actually all the systems and people in between it. [77:02] I think part of the reason why I think it's been slow to get the productivity [77:05] enhancement is we ship our org charts as companies naturally. [77:09] That's the natural state. [77:11] There's not usually a person responsible for that process. [77:14] There's the legal team responsible for the contract, there's the procurement team. [77:18] I think actually we will end up reimagining our companies with the benefit [77:23] of AI, will we actually think of our companies as a collection of processes, [77:27] have people responsible them with KPIs who can apply AI. [77:32] I think I bring it up just because that's my theory of the world. [77:35] I might be wrong, I might be right, but I'm not sure companies are set up to [77:40] essentially absorb the benefits of AI efficiently right now, [77:43] and we need to do that to really do so. [77:45] But the bigger point, I think, is that there's the paradox of, well, [77:51] you want more software engineers. [77:53] But on top of that, most of the world isn't just digital technology. [77:58] I think a lot of the people in the AGI community have only ever worked [78:02] at a research lab or a software company. [78:04] You look around, you're like, "Wow, AI is going to do all of this." [78:07] As they walk by the flower shop and get their coffee at the coffee shop, [78:12] and you think about the local flower shop, if you took all the AI in the world [78:17] and gave it to that, you gave it super intelligence, [78:20] how much would it impact the flower shop's operations? [78:23] Maybe a little. [78:24] I mean, I'm sure it would help, don't get me wrong, [78:26] but someone's still clipping the ends of the stems of the flowers, [78:30] arranging the bouquets, and thanking you on your way out the door, [78:34] and congratulating you for your daughter's wedding, or whatever it is. [78:38] I think if you think about what parts of the economy can absorb intelligence [78:43] really efficiently, it's certainly software, and we're seeing that already. [78:46] Finance seems particularly meaningful here because so much of finance today [78:51] is just digital information. [78:52] Everything's in digital systems now, not even just crypto. [78:56] I mean, just everything's in digital ledgers everywhere. [78:59] It still doesn't touch a wet lab. [79:00] It still can't do a clinical trial. [79:03] You still need to get a crate from this country to that country on a ship. [79:09] As a consequence, I think I'm not sure we'll see the productivity [79:14] enhancement we see in software in every sector as quickly. [79:17] Then on top of that, I think companies need to stop just giving [79:22] Copilot to every employee and be like, "We're AI now," and start to think about [79:26] from first principles, what are the parts of your business [79:29] that have a lot of digital workflows? [79:31] Where can AI have a real big impact? [79:33] How do you actually set up your company to actually have someone [79:36] accountable to drive that? [79:37] That feels like a real big change management opportunity that most [79:41] companies haven't done. Just to push on that. [79:43] Software engineering, I think we clearly are seeing a lot of AI productivity gains. [79:49] Software engineers have always loved tools and the latest tools and are [79:52] just headlong diving into it. [79:54] Then you have stuff like you see in the flower shop, and [79:59] stuff that requires really good robotics that we're far away from. [80:02] That will take a while. [80:04] What I'm talking about is like— [80:06] And by the way, I might prefer a flower shop with the florist. [80:12] Totally. Yeah. [80:12] Just to say it. [80:13] I'm not sure it solves a problem I have in my flower shop. [80:18] I might be wrong. I might be unique in that. [80:20] I think a lot of the economy is actually [80:24] white-collar knowledge work, not coding. [80:27] Think of finance departments, legal departments, [80:31] things like that, where you should be able to see a lot of AI uplift [80:37] and a lot of AI productivity improvements. [80:39] It just feels like a current course in speed. [80:42] We're not on track to get those productivity— [80:47] Well, I'm not sure I'm right, but I would argue thinking about it [80:51] by department rather than by process is where it's off. [80:54] We can talk about the processes as well. [80:55] Hear me out, though, in this, because if you said, [80:58] "I want to make the legal department more productive, so I want to make it easier [81:02] to do red lines," and you optimize that. [81:04] But why is the contract there? [81:08] What is it for? [81:11] If you're, for example, onboarding a supply chain vendor and you [81:14] have hundreds of them, you might actually say actually making an [81:19] abstract technology for your legal department to red line contracts more [81:24] efficient is actually a harder, more general problem than [81:28] for your supply chain vendors, because you might actually have very [81:31] rigid rules around your supply chain. [81:33] Let's say you're a CPG company, and you might actually have very specific [81:38] saying, "Look, if you want to work with us, [81:40] here's our core legal terms, here's the axis of independence." If you [81:44] want to make an AI agent to automate that contract, that's actually a much more [81:48] narrow problem domain that doesn't require general purpose redlining technology. [81:54] In fact, if you reduce it, you could say, "Well, [81:57] there are 10% of our suppliers where we let them negotiate their contract, [82:01] but only for this spend." [82:02] Let's have them go through our legal department. [82:04] The rest, let's do it all with AI. [82:06] My point on it is, if you look at it through the lens [82:09] of an end-to-end business process, you can turn science into engineering. [82:13] I think solving legal through AI, that's a science problem. [82:18] This is my point, though, which is I think people are going [82:21] through department by department. [82:23] Similarly, there's not a person accountable for that end-to-end process. [82:27] The more you can narrow the domain that you're solving with AI, [82:31] the more you can build a harness or a scaffolding with existing technology [82:35] to actually fully automate it. [82:37] My hypothesis is most companies just aren't set up that way. [82:40] That's just not how we're organized. [82:42] As a consequence, we're all optimizing our silo. [82:45] We're all just installing Copilot. [82:47] Copilot is great, by the way. [82:49] I didn't mean to insult it. [82:51] But it's not actually like... But it's incremental. [82:54] To be fair, that's the thing we're doing. [82:57] Obviously, good companies did this before [83:03] AI, continuous process improvement, and it [83:05] feels like that is the best thing to do. [83:07] I think what you're saying is there's no such thing as an AI lawyer. [83:10] Instead, there's improving your commercial contracting. [83:13] That is a thing that you can— [83:14] Even more narrowly, pick one domain of commercial contracting and solve that. [83:21] I actually think those are truly solvable. [83:24] I think the companies that really think [83:28] about their business that way, I think they can see the value. [83:31] Again, I'll go back to the immaturity of the applied AI market is probably [83:35] one of the bigger barriers right now. [83:38] My hope is that as the applied AI market matures over the next few years, [83:42] we'll see a step change in productivity. [83:47] There is a canonical way to build a Silicon Valley company. [83:52] You have engineering and product and design, you have this number of ratios [83:55] of engineers to product managers and engineering managers. [84:00] Then you have your go-to-market organization, and you have these pipeline [84:02] coverage ratios, and you have the product marketers, and all this stuff. [84:07] But I find it interesting how similar so many Silicon Valley companies are [84:11] to each other because they've all learned from each other. [84:13] There's a shared recipe and a shared playbook as to how to build a company. [84:18] It gets tweaked, but ultimately, I think it's a pretty good IP. [84:22] Certainly, companies are much better off with it than without. [84:25] How is that canonical template for building a company [84:30] different post-AI than before? [84:32] Yeah, it's a really interesting question. [84:34] One is, I've always believed in the primacy [84:38] of tech leads over engineering managers. [84:41] Both Google and Facebook, where I spent some of my early career both [84:46] did this well, where in a product review, you weren't just talking to a manager. [84:53] You were talking to the tech lead and PM who were product manager [84:57] who were building the product. [85:00] Whereas if you went to companies that produced worse software, [85:03] I'd notice you move up the chain of the command, like the military. [85:07] I think that we will end up with individual [85:14] tech leads who, because of the existence of AI agents, [85:19] will become even more important, where [85:22] if you are a, I'll say, product engineer, [85:26] I'm trying to find the right word for it—we might invent one—who has taste [85:30] but didn't necessarily know CSS, who has infrastructure [85:36] ability, meaning that you understand the basics of distributed systems [85:41] and debugging, and you understand your customer very deeply. [85:45] With the presence of Codex, you can produce amazing results. [85:51] Those people are truly worth a thousand X [85:56] other people because it's relatively easy [85:58] to find someone who's a great infrastructure engineer. [86:01] Not easy, easy, but relatively. [86:03] Finding someone with good taste, that's relatively easy. [86:08] Finding someone who also understands your customers extremely well, the nuances [86:11] of the problem they're solving. [86:13] Those people who can combine that will, I think, end up being able to actually [86:18] produce products like, capital P, valuable products with relative autonomy. [86:25] I wonder if it will change our view on generalists broadly. [86:30] I've always identified myself as a generalist just because I've been both [86:34] a software engineer in a suit, basically, and I've gone back [86:38] and forth in that world. [86:40] As companies grow, you tend towards more specialization just because [86:47] the person who's the Jack or Jill of all trades ends up not fitting in. [86:51] There's not really a place for them because, okay, well, [86:55] you're not really the deepest engineer, you're not really the best designer, [86:59] you're not really product manager. [87:28] If you've been at the company for a while, we'll give you an honorary something [87:29] to do, and you have to lead through influence, and da, da, da, da, da. [87:30] Could that person actually endure as one of the most valuable [87:30] people at these companies? [87:30] I don't know whether it's naive optimism or true, but I actually think those people [87:30] who often exist in early-stage startups are often the people who get sidelined, [87:30] but in a way that actually really harms the company. [87:31] I'm hopeful that in a world of AI agents, those generalists who, again, [87:32] I think the most important part is understanding the customer need [87:36] with agency, no pun intended, and empowerment can end up [87:40] more powerful in the Silicon Valley company in the future. [87:43] I've noticed the exact same thing at Stripe. [87:45] The exact same thing, which is high [87:48] agency, really caring about customers, [87:53] just really caring generally. [87:57] High work ethic people who maybe weren't the best engineers [88:03] previously or now, those people are massively ascendant, [88:08] as far as I can tell, because they suddenly got the exoskeleton. [88:14] They always have the ideas as to what we should be doing, and this is the better [88:17] way to serve the customers and everything like that. [88:18] But now they have the way to make all their schemes real. [88:22] I've really noticed that at Stripe. Well, it's interesting. [88:24] You talked about work ethic. [88:27] It's addictive right now because you can do much with the technology. [88:32] Everyone I know who's really used it works harder because it's like, "Wow, [88:38] I could do so much." You think you're about to go to bed. [88:42] You're like, "Should I get an AI agent to do something? [88:44] Am I wasting the next eight hours of my life?" [88:49] That might be a novelty that wears off, but I think it's really exciting. [88:54] I'm hopeful on the product engineering design side, you end up with these hyper [89:00] high-agency people who really deeply care. [89:03] I really like the way you said it, actually. [89:05] It's right. It's not just customer problems. [89:06] It's care, period. Just care can end up more empowered. [89:10] I'm curious what that means for organizational structures. [89:13] I think we have a new job role we need to invent. [89:16] It's like, what role do these people in? Hyper-generalist. [89:18] Yeah, like product managers, but sometimes maybe without a product, [89:24] like minister without a portfolio. [89:25] They're just doing stuff, but now they can do much more. [89:27] It's almost like product designer, product manager, engineer. [89:31] That's why I said product engineer, but that means something different. [89:35] But it's interesting because we've talked about this. [89:37] You end up where the grass is always greener with org structure. [89:40] You go functional organization. [89:41] Okay, we're going to have engineering, product, design. [89:43] Let's go to business units. [89:44] They're like, "Wow, that led to silos and worrying factors." [89:47] "We're one reorg away." Yeah, you sway back again. [89:50] That's welcome to—Just one more reorg. [89:52] Yeah, exactly. [89:53] "I'm a middle manager now." I think that it is interesting if these [89:59] people become extremely important. [90:02] What does it mean to organize around them? [90:04] I think it does feel like something that will end up flatter just because [90:10] of the amount of impact an individual can have. [90:15] That feels really exciting to me, but it feels like a blurry picture right now. [90:20] Enhance. I agree. [90:21] It's very blurry. It's so interesting. [90:24] You were on the Twitter board during the [90:32] super interesting takeover [90:33] battle with Elon Musk. [90:35] What are your reflections on that experience a few years later? [90:39] It was really interesting to be in the public spotlight. [90:44] I hadn't really experienced that in my career before. [90:48] I joke that no one really cares about enterprise software. [90:52] I worked for Salesforce for six and a half years. [90:55] Sorry, what's the joke? [90:56] I worked for Salesforce for six and a half years, and I don't think my mom [90:59] knows what Salesforce does. [91:02] To have something that was not really just a business issue or a technology issue, [91:08] but in the mainstream. [91:11] I realized I didn't love that very much. [91:14] You actually prefer enterprise software. Yeah, exactly. [91:17] I'm like a builder. [91:19] I like to build things and have people use them. [91:22] I know it sounds funny and reductive. [91:24] That's what gives me joy. [91:25] The one thing I realized is the conflict of it all. [91:30] However it turned out, victory, defeat, whatever it was. [91:34] It just didn't... [91:35] It wasn't something that filled my bucket very much. [91:38] What do you make of the fact that in all these head count debates, [91:43] Elon is now running Twitter with 80-85% fewer people. [91:47] I think Nikita Bier tweeted recently that all of eng, product, [91:52] and design at Twitter is 50 people. [91:58] It's really impressive. [91:59] Yeah, it's been a little flaky in pockets or just at times, but mostly the service [92:03] works, and they have shipped new features. [92:05] I think those two statements are undeniable. [92:08] But just what's your takeaway from that? [92:10] I don't know. I haven't followed as much as the… [92:14] I didn't see that tweet as an example. Do you call it tweet still? [92:17] Sorry, I'm retro. I'm old-fashioned. [92:19] I don't know about that, but it is interesting right now because [92:25] obviously a lot of that predated AI. [92:30] But I mean, any person has been an individual [92:33] contributor, engineer, knows that the size of the team does [92:35] not produce linearly greater outcomes. [92:38] Everyone in the world has experienced that. [92:41] I think the idea of can you actually give [92:47] individuals with good taste [92:48] more agency, no pun intended. [92:50] I think it's always been an enduring thing. [92:53] What was the Jeff Bezos—"It's a two pizza box thing." [92:55] But then do large tech companies underrate this phenomenon? [92:59] Do they pay lip service to small empowered teams and two pizza teams, but X shows [93:05] that maybe they should be doing more? [93:07] I think their companies largely act somewhat rationally. [93:12] I can't remember who the CEO was, but it might have been [93:15] the Rippling CEO just talking about there's this idea of being lean and agile, [93:21] and then there's you want to capture market share and grow your [93:24] product and grow your platform. [93:27] At the end of the day, you can be clever but not smart. [93:32] You might be so clever to think, "I'm not going to have anything more than [93:36] two people on these features." If you have a competitor who maybe does something [93:40] a little less elegantly but wins, who cares that you are clever with your [93:45] two pizza box team or two-person team or one AI agent team or whatever it is. [93:50] When someone said, "We're going to have a X billion dollar company with one person," [93:54] I think that might have been right, but it's not— [93:57] Maybe you could have had a $10 billion company if you'd hired a bit more. [94:00] That's right, and I would actually argue the more specific thing is if all [94:03] of a sudden, for some clever reason, you want to prove you can, [94:09] the idea that a competitor might have 10 people and beat you is probably more [94:13] likely than even having a $10 billion company. [94:16] I think at the end of the day, when you're building a business, [94:20] especially one that's in hyper growth, which successful businesses in tech tend [94:24] to be, if you are too clever and austere, and going back to your point about [94:30] Silicon Valley culture is all being the same, there are examples of companies [94:35] that really innovated in culture. [94:37] You wouldn't think of it this way, but HP, a lot of the traditional open [94:42] office floor plan came from them. Facebook. [94:45] Woz worked at HP. Oh, I didn't know that. [94:47] That's interesting. [94:49] Then Google offered free food to their employees, which a lot of people did. [94:54] Then Facebook, a lot of the layouts of offices all [94:58] looked like Facebook for a long time. [95:00] But then you have other companies like, "We're going to innovate in HR, [95:03] and they spend all this time and energy on it." In fact, the smart thing to do [95:07] is just be like, "It's not what we do. [95:09] Let's just do the same old thing as everyone else because [95:11] everything is push button. [95:13] I don't need to worry about it." I do think it's the right question [95:18] to ask for every technology company. [95:22] After being on the Twitter board during the Elon takeover, you were then [95:26] on the OpenAI board when Sam got fired. [95:30] Have you considered that you are the problem? [95:33] You are bringing the drama. [95:34] I came in after the drama there. Oh, right. [95:37] You joined after. Sorry. [95:38] I was brought in as the mediator. I see. [95:41] Post the... [95:42] Okay, your hands are clean. [95:44] Maligning my reputation here. [95:46] I wasn't actually on the other side of it, but I got a phone call, [95:50] was it Saturday or Friday after? [95:54] Basically, my understanding was I was the person that both the existing board [95:59] and Sam agreed upon to help mediate the situation. [96:02] What have you learned on the OpenAI board? [96:05] A lot. [96:06] I mean, certainly the most interesting part is the AI research. [96:12] I've never been affiliated with a true research lab before, [96:15] and that's fascinating to me. [96:18] It is very inspiring. [96:20] I mean, it's very easy to grow, [96:24] not cynical, but you can look at OpenAI, [96:29] Google, Anthropic, and say, "Whose model scores better on this leaderboard?" [96:34] To actually go in and see this company where every single researcher is trying [96:39] to make safe AGI and not come out of those board meetings inspired is impossible. [96:42] It's amazing. [96:44] The other thing is it's the first not-for-profit board I've [96:47] been affiliated with. [96:49] That's really interesting as well just because— [96:52] It's a different thing. [96:53] Well, when I mentioned the fiduciary duty, is you have a duty is you [96:56] have a duty to commission. [96:57] That is really clarifying and interesting as well, because when you're making [97:03] decisions and you realize your sole duty is to ensure that artificial [97:09] general intelligence benefits humanity, that's really different. [97:13] It's really interesting. [97:14] I've never had a fiduciary duty to a mission before. [97:17] That's really interesting to me because I take those duties really seriously [97:20] and reflecting in a board meeting and you're making a decision, [97:23] you think about it very differently through that context. [97:26] Then the other thing was because I was brought in after that crisis, [97:32] there was three people on the board on the other side of that when I [97:35] agreed to temporarily be the chairman. [97:37] It's still there. [97:39] We had to grow the board essentially from scratch. [97:42] That was really interesting, too. [97:44] Just to think about… [97:46] Normally, you add one board member at a time. [97:48] This one was like, "Do you have a bulk rate?" [97:51] "We're going to build a board, put together a team." [97:54] You really think about spend time with the other two board members just [97:57] really thinking about what does the composition an OpenAI board look like? [98:01] How do you represent the not-for-profit part of it? [98:04] How do you represent safety? [98:06] How do you represent the economic impact of AI? [98:10] We're doing lots of infrastructure investments. [98:12] How do we find someone with that specific type of financial expertise? [98:18] That was really rewarding as well, just building a board, not from scratch, [98:22] but effectively from scratch. Last question. [98:24] What are your AI predictions for 2026? [98:27] I think we will have some scientific breakthroughs with AI that positively [98:34] break through into the mainstream press and awareness. [98:38] We've already had some interesting math proofs, [98:42] but I joked with one of my friends, "Until I can understand what the title [98:47] means, I'm not sure it's going to make..." [98:49] You're not excited about the interventional manifold space. [98:52] Yeah, exactly. [98:54] It won't quite be like the Apollo landing, but I remember [99:02] the Kasparov chess match, and certainly things like [99:07] AlphaGo were really meaningful. [99:10] Given the progress in math, I'm hopeful we have [99:13] at least one moment of discovery that is inspiring, because I think [99:19] a lot of the dialog around AI right now is economic opportunities, [99:24] but also what could go wrong. [99:26] I actually think one of the main things that can go right is actually discovery [99:30] in science that actually can improve the human condition. [99:33] I'm really excited for it because I think it will contextualize why so many of us [99:38] are excited about this technology in a way that captures attention. [99:41] As you said, something beyond an end-dimensional manifold blah, blah, blah. [99:46] I feel not confident in that, but it certainly feels like [99:49] the ingredients are there for that. [99:51] I think we'll continue to see mainstream adoption of AI by both [99:56] consumers and companies. [99:58] That doesn't really feel like a prediction, but I think this will be [100:02] really a year of adoption of agents. [100:06] We're certainly seeing that in Sierra's customer base, but I think we're [100:10] going to see it more writ large. [100:12] Then you've already seen ChatGPT growth, really unprecedented levels. [100:17] Things like OpenClaw, you can see that translate over to agents and the more [100:23] long-running autonomous tasks. [100:25] It does feel like by the time we exit this year, can that go from a niche [100:30] community to something more mainstream? [100:31] It feels probable to me. [100:34] Then the other thing is, I think most companies in Silicon Valley [100:41] won't write code by hand. [100:44] That might seem... [100:45] It's funny that it feels obvious right now. [100:47] You're like, "Oh yeah, of course." You're just nodding like, [100:49] "Yeah, why not?" But if I had said that four months ago, [100:53] that would have been a bold prediction. [100:55] But I think that's really interesting just because that's [100:58] such a fundamental state change. [101:00] I say in Silicon Valley because I do think it takes a while for these tools to [101:05] diffuse through society. [101:07] Silicon Valley is insular enough that I think it will here, [101:11] but I'm not sure it will happen through every company in the world yet. [101:15] The year of agents across businesses and just people finally getting their [101:19] Claw-style agents, and then all code written by AI as well. [101:24] It's a good set of predictions. Great. [101:25] Thank you. Thanks for having me.