Cleo Hits $400 Million Revenue as Founder Automates Executive Oversight

Cleo营收达4亿美元，创始人实现管理监督自动化 Cleo首席执行官Barney Hussey-Yeo预测，AI工具允许高管绕过中层管理直接监控代码和运营，从而使企业架构趋向扁平化。 在这场炉边谈话中，Cleo创始人Barney Hussey-Yeo与Stripe首席执行官John Collison探讨了Cleo实现4亿美元ARR的增长历程、从早期基础正则表达式向现代LLM架构的技术转型，以及通过幽默与个性化让财务建议更具吸引力的产品哲学。Hussey-Yeo对比了欧洲与美国的监管及投资环境并提出批评，主张建立包容风险的“创造性破坏”文化。他还分享了AI工具如何通过使工程工作流程透明可见来重塑其CEO角色、扁平化企业管理层级，并赋能小型团队并行运行数千次实验。 总部位于伦敦肖迪奇（Shoreditch）的现代初创企业Cleo表现抢眼：其ARR已达到4亿美元。 (0:00) Hussey-Yeo曾通过玩线上扑克赚取大学学费，目前正运行一个定制AI扑克代理，通过低额单挑对局每年赚取约20万美元的佣金返还。 (0:00) 在高风险封闭系统（如扑克或量化金融）中建立的竞争优势通常被作为专有资产保留，而不会为了系统性效益对外共享。 (1:13) Cleo成立于2016年，其核心假设是大众消费者的财务决策极度缺乏优化；其技术架构从早期的正则表达式和基础分类模型逐步演进为现代生成式LLM。 (2:35) 针对99%的普通人群（尤其是面临日常财务压力、存款不足400美元的60%美国人），财务建议必须具备主动性与个性化。 (5:46)

In conversation with Stripe CEO John Collison, Cleo founder Barney Hussey-Yeo argues that consumer finance is best transformed through proactive, personality-driven AI agents rather than static rules or generic foundation models. Hussey-Yeo contends that sustainable technology value lies in deep vertical domain execution, structured user context, and organizational workflows transformed by machine learning. Targeting Everyday Consumer Financial Stress (5:46) Hussey-Yeo contrasts typical consumer realities with standard wealth management advice. Pointing out that 60% of Americans have less than $400 in liquid savings and face volatile incomes alongside high-interest debt, he argues that high-level guidance like investing in global index funds fails the 99% of consumers living paycheck to paycheck. Founded in 2016, Cleo reached $400 million in ARR by building an automated assistant designed to manage daily cash flow and eliminate unoptimized spending choices. Personality and Architecture in Vertical Agents (8:10) Cleo's early architecture evolved from regular expressions, multi-class intent classification, and reinforcement learning into generative large language models. To drive engagement around stressful financial topics, the company adopted an irreverent tone—initially hiring 30 female comedians to craft features like "Roast Mode"—before transitioning to dynamic LLM persona generation tailored to user demographics and location. Hussey-Yeo emphasizes that Cleo balances conversational banter against an internal Financial Health Score (0 to 100), aiming to guide users toward spending less than they earn over two- to three-month horizons. Structured Intelligence Versus Frontier Model Wrappers (10:15) Hussey-Yeo rejects the idea that generalist frontier models like ChatGPT or Claude will displace specialized software platforms. Because everyday consumers rarely open search interfaces to inspect bank accounts, effective financial management demands an always-on, proactive agent. Cleo maintains structured individual profiles tracking income, aspirations, health metrics, and interaction histories. Hussey-Yeo maintains that underlying financial primitives (spending, borrowing, saving) innovate slowly and carry heavy regulatory burdens; the durable enterprise value sits in the intelligence and advisory layer directing those rails rather than pure product margin maximization. Institutional Inertia and Venture Capital Distribution (20:59) Critiquing the European and UK business environments, Hussey-Yeo argues that risk-averse regulation and multi-year sandbox timelines frequently outlast typical 18-month startup venture cycles, hampering "creative destruction." In contrast, he views the US system of checks and balances across courts, states, and federal regulators as better suited for technology growth. Furthermore, he points out that top European venture capital relies heavily on North American pension fund limited partners, meaning investment returns from European tech innovations often flow back to foreign retirees rather than regenerating domestic capital. AI-Native Operations and Flatter Corporate Hierarchies Internally, Hussey-Yeo restructured executive operations by using LLMs to index Slack messages, Notion documents, and GitHub pull requests. By summarizing code quality and developer velocity directly, this workflow bypasses traditional middle-management reporting filters. Rather than running nested management hierarchies, Hussey-Yeo advocates deploying smaller teams of junior machine-learning graduates who act as quality controllers directing fleets of autonomous coding agents, enabling organizations to execute thousands of simultaneous operational bets. Hussey-Yeo dismisses forecasts of an AI-driven "SaaS apocalypse," asserting that foundation models cannot easily replace the regulatory expertise, domain craft, and proprietary data structures of established software platforms. However, questions remain regarding how traditional middle management and apprenticeship structures will adapt as agentic automation reshapes corporate governance.

Cleo创始人Barney Hussey-Yeo在与Stripe首席执行官John Collison的对话中指出，针对大众市场的个人财务管理必须依赖兼具主动干预能力与鲜明人设的垂直AI代理，而非通用大模型。他认为，结合深厚领域积淀与合规交付能力的专业软件依然具有持久壁垒，同时AI技术正在从根本上扁平化公司治理架构，并重构科技企业的研发执行效率。 面向大众的垂直财务代理与技术演进 (2:35) Cleo创立于2016年，核心立论在于普通消费者的日常财务决策普遍缺乏优化，且财务问题是大众日常压力最集中的领域。针对高净值人群拥有专属财富管理顾问、而60%的美国人流动储蓄不足400美元的现实差距，Hussey-Yeo强调Cleo旨在为缺乏投顾服务的99%大众提供自动化辅助，协助其规避高息信贷并优化结余资金。技术实现上，Cleo经历了从早期的正则表达式（Regex）、多分类意图识别与强化学习，到如今大语言模型（LLM）驱动的系统重构。 Hussey-Yeo认为，大众财务管理不能仅依赖静态的常规法则（如被动投资全球指数），而必须通过主动式代理在后台自动对接低息信贷与免税账户，消除掠夺性利息支出。Collison将此模式比作构建消费者财务生活的“博弈论最优解（GTO）”。截至对话发生前一个月，Cleo的年化经常性收入（ARR）已达到4亿美元。 逆向产品设计与人设驱动的变现逻辑 (8:10) 面对金融领域普遍严肃、枯燥的行业惯例，Cleo采取了“逆向产品设计”（Product Contrarianism）策略，早期聘用30名女性喜剧演员编写交互文案，并推出以幽默方式调侃用户过度消费的“吐槽模式”（Roast Mode）。Hussey-Yeo指出，具备鲜明人设和幽默调性的交互能直接提升月活跃用户数（MAU）与业务收入；在大模型支持下，系统更能根据用户的地域、收入及偏好动态匹配个性化语气。 为避免扁平文本记忆机制的局限，Cleo构建了涵盖收支、财务目标及互动历史的结构化数据架构，结合推荐系统持续优化0至100分的“财务健康分”，核心算法旨在促使用户在2至3个月内实现“支出少于收入”。在商业模式上，Cleo采用免费增值模式，通过订阅费与嵌入式信贷产品获利，并利用价格锚定等行为心理学机制引导用户选择中间档位订阅服务。 垂直代理壁垒与对“SaaS末日论”的反驳 (10:15) 关于通用前沿大模型是否会取代垂直软件的争论，Hussey-Yeo持明确否定态度，直言市场过度外推模型能力的“SaaS末日论”缺乏对业务复杂性的理解。他分析称，个人财务管理需要一种“全天候、主动出击（Always-on, proactive）”的代理形态，绝大多数普通用户缺乏每日主动向ChatGPT查询账单的动力，产品价值在于系统能否在关键时刻主动介入。 此外，构建持久的B2B或B2C软件企业（如Stripe或Cleo）需要深度对接底层银行轨道、医疗系统或专业硬件，同时具备复杂的合规交付能力与精细的产品工艺。Hussey-Yeo认为，这些深层壁垒无法被简单的模型套壳或通用模型底座轻易抹平，软件创新与创造性破坏仍将是全球经济增长的核心动力。 AI原生工作流对企业治理的扁平化重塑 在企业内部管理方面，Hussey-Yeo分享了其自2024年初以来推行的管理重构。借助大语言模型自动化汇总Slack交流、Notion文档和GitHub代码合并请求（PR），CEO得以绕过传统中层管理人员的汇报过滤与激励偏见，直接实时评估代码质量与研发速率。他预测，这种高度透明性将推动企业治理结构永久性扁平化，大幅削减中间管理层级。 在团队组织架构上，Cleo倾向于招募精通AI操作的高校机器学习专业毕业生，由初级员工担任“品味决策者”并指挥成群的递归AI代理协同执行。Hussey-Yeo表示，这种新型人机协同模式使公司能够从过往每年执行十余项战略尝试，跨越至并行测试多达两千项业务举措。 监管制度差异与欧洲风投资本配置困境 (20:59) 谈及宏观创新生态，Hussey-Yeo对英国的金融监管效率提出了严厉批评。他指出，英国监管沙盒等流程动辄耗时两年，往往直接拖垮早期初创企业18个月的风险投资资金生命周期；相比之下，美国分散在联邦、各州与司法系统之间的制衡机制与风险容忍度，更有利于激发“创造性破坏”并推动技术落地。 此外，欧洲的养老基金等大型机构在资本配置上过度偏向避险，导致欧洲顶级风投机构的主要有限合伙人（LP）多来自北美养老基金（如加利福尼亚州及加拿大公共养老金）。Hussey-Yeo警告称，这种结构性缺陷导致欧洲本土科技创新产生的经济红利主要回流至海外，欧洲在由中美主导的全球科技双极格局中正面临被边缘化的长期风险。 本次对话展现了垂直领域AI企业在跨越底层技术迭代时的产品策略与组织演进路径，但也揭示了关键的制度与市场未决问题：随着AI代理对中层汇报职能的替代加速，企业如何在规模化激增的自主决策中维持风险控制；同时，若欧洲无法在资本端扭转保守倾向，其本土创新力量能否在全球科技双极竞争中实现自立仍存不确定性。

## Background 本内容由 Stripe 官方于 2026 年 7 月 7 日发布在 YouTube 平台。视频记录了在 Stripe Tour 伦敦活动现场，Cleo 创始人兼首席执行官 Barney Hussey-Yeo 与 Stripe 联合创始人兼总裁 John Collison 进行的一场深度炉边对谈。 ## TL;DR 对话从 Barney 早年利用扑克算法套利经历切入，复盘了 AI 理财助手 Cleo 从规则匹配演进至 ARR 突破 4 亿美元的 Agent 架构历程。Barney 直言传统理财认知严重脱离了 99% 的底层大众，并通过 Cleo 的幽默调性与主动代理服务，解决普通人日常面临的现金流与信贷消耗痛点。对谈深入剖析了英国金融监管阻力及欧洲养老金风险资本缺位等宏观制度议题，并展示了如何通过内部大模型穿透数据实现组织的扁平化革新。最后，Barney 坚决驳斥了“SaaS 灭顶之灾”的过度恐慌，坚信垂直领域的深度行业工艺不可替代，AI 必将催生更丰饶的软件生态与创造性破坏。 ## Atmosphere & style 整场对话机智幽默、火花四溅，充满两位顶尖科技创始人之间毫无包袱的坦率交锋与智性碰撞。Barney 毫无顾忌地当面调侃 John 对普通大众生活的“脱节”，用辛辣犀利的语言拆解传统金融、风投羊群效应与内部工程管理的痛点。现场节奏明快而充满张力，既有极客探讨底层架构演进的扎实技术细节，又有宏观制度批判与激进商业推演带来的强烈认知冲击，整体观感生动、自信且极具感染力。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=70kWqOOWlAA&t=0s) Cleo 创始人 Barney Hussey-Yeo 透露其公司年度经常性收入（ARR）已突破 4 亿美元，从早期依赖 30 位喜剧演员手写幽默文案演进为依托大语言模型与多层推荐系统的自主 AI 理财代理。 - [11:58](https://www.youtube.com/watch?v=70kWqOOWlAA&t=718s) 针对通用大模型在垂直场景的局限性，Barney 坚决看衰在 ChatGPT 等通用平台上挂载的金融插件，指出被动搜索无法满足反人性的财务管理需求，真正的核心在于常驻后台、主动干预的 Agentic 体系。 - [31:23](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1883s) Cleo 内部利用 LLM 汇总 Slack、Notion 和 GitHub PR 构建全景“船长日志”，高管可直接穿透底层代码与工程产出，打破中层管理者的信息过滤并推动组织深度扁平化。 - [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) Barney 严厉驳斥了“SaaS 灭顶之灾”（SaaS apocalypse）的论调，认为基础模型写出的 Prompt 无法替代垂直领域极其繁复的业务细节与专业壁垒，AI 最终将催生更多软件公司而非消灭产业。 ## Key takeaways - [02:35](https://www.youtube.com/watch?v=70kWqOOWlAA&t=155s) Cleo 从早期通过正则表达式与传统有监督意图分类做规则匹配，逐步演进为利用深度推荐系统，并在 GPT-3.5 爆发后全面重构为大模型驱动的自主理财 Agent。 - [05:01](https://www.youtube.com/watch?v=70kWqOOWlAA&t=301s) 传统理财顾问仅服务前 1% 高净值人群，但全美有 60% 成年人拿不出 400 美元应急储蓄；普通大众需要的不是复杂的指数配置，而是将耗费心力的日常账单与现金流决策全流程自动化。 - [08:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=493s) 幽默是 Cleo 撬动用户参与度的核心武器，通过反传统的“吐槽模式”（Roast Mode）化解大众对金钱的沉重与焦虑心理，进而演进至代客调配资金、开户等主动代理操作。 - [09:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=560s) Cleo 将算法目标函数从单一的“参与度”升级为“财务健康度 + 参与度”的双重平衡，底层逻辑紧扣“让支出少于收入”的朴素原语。 - [10:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=613s) 绝大多数大众的理财行为是“凭感觉的优化”（vibe-based），未来在健康、财务和教育等决策非最优化的垂直赛道，将出现数个深入底层专属数据的专用垂直 Agent。 - [11:58](https://www.youtube.com/watch?v=70kWqOOWlAA&t=718s) 通用大模型因属于被动响应式工具而无法做精财务管理；真正有用的理财助手必须具备“主动推送”与“代办执行”的 Agentic 能力，并结合用户情绪与最佳沟通契机。 - [13:17](https://www.youtube.com/watch?v=70kWqOOWlAA&t=797s) 面对 LLM 缺乏长期记忆的局限，Cleo 采用递归调用 LLM 从对话与流水中抽取结构化知识，沉淀在规范化数据库与 Schema 中，供推荐系统计算最优沟通话术与代办动作。 - [15:01](https://www.youtube.com/watch?v=70kWqOOWlAA&t=901s) AI 的“个性与语调”（Personality）绝非单纯的表面风格，而是直接对 MAU 和收入产生实质反馈的核心商业指标，能依据用户画像与地域文化动态调整。 - [18:14](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1094s) 借贷、支出与储蓄等基础金融原语极其有限，缺乏 AI 建议层的纯金融销售模式必然倒向诱导非理性高频投机或卡债透支的负和博弈陷阱。 - [20:59](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1259s) 英国 FCA 监管沙盒耗时达 2 年，与初创企业 18 个月的风投生命周期严重错配；相比之下，美国分权制衡的监管与司法体系为颠覆性创新提供了更大的抗风险与抗辩空间。 - [23:11](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1391s) 欧洲及英国养老金等庞大机构资本极度厌恶风险且配置保守，导致本土 VC 核心出资人多来自加州，欧洲亟需切换至“全险”（Risk-on）创新心态以避免在全球竞争中被边缘化。 - [26:09](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1569s) 对于消费级 AI 产品的商业化，Barney 告诫绝不要搞折中方案：要么全力冲刺用户规模与产品留存，要么如 Stripe 一般在第一天就把商业抽成机制内嵌在核心产品体验中。 - [33:31](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2011s) 高管借助 AI 穿透全量 Slack、文档与 GitHub PR，建立事实驱动的“船长日志”，能够直观识别真实交付与代码质量，弱化中层管理者的过滤与偏见，倒逼组织扁平化。 - [34:46](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2086s) 未来的组织范式或将告别臃肿的“金字塔”层级管理，转变为由深具审美与工程品味的年轻人才直接调度庞大的并行 Agent 舰队，以极低试错成本支撑成百上千项业务探索。 - [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) “SaaS 灭顶之灾”纯属恐慌过度；深厚的垂直行业壁垒并非单凭基础模型的 Prompt 能轻松抹平，AI 将激发更充沛的软件丰饶与创造性破坏。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=70kWqOOWlAA&t=0s) 从德州扑克玩家到 AI 代理：GTO 与信息差的套利 在伦敦举行的活动现场，Stripe 联合创始人兼总裁 John Collison 介绍了对谈嘉宾——Cleo 创始人兼首席执行官 Barney Hussey-Yeo。Cleo 作为一家诞生于伦敦肖迪奇（Shoreditch）的独角兽企业，在上个月年度经常性收入（ARR）刚刚突破 4 亿美元，是当下消费级个人理财和 AI 领域增长最快的明星公司之一。然而在正式创立 Cleo 之前，Barney 有一段不同寻常的“职业履历”：在大学时期，他依靠职业线上德州扑克赚取生活费与学费。Collison 打趣地问起这段像是电影《赌王之王》（*Rounders*）或《茉莉牌局》（*Molly's Game*）里的经历究竟是如何运作的。 Barney 回忆道，在 2012 至 2014 年左右，德州扑克界还没有普及 GTO（Game Theory Optimal，博弈论最优策略）求解器。他当时每天在多个线上扑克网站打 4 到 6 个小时的单挑对局（heads-up），在中低盲注级别维持着相当可观的盈利。更引人入胜的是，这种对算法与套利的敏锐度一直延续至今——Barney 透露他最近刚刚写了一个自动打牌的 AI Agent，仅靠在低级别牌局中微弱赢率并收割平台的抽水返利（rake back），每年就能为他带来约 20 万美元的被动收入。 当 Collison 谈到 GTO 策略近乎将线上扑克变成一个“已被解决的游戏”（solved game）时，两人探讨了金融与竞技体育中那些被长期隐瞒的竞争优势。就像期权定价的 Black-Scholes 模型也是相对近代才被发明一样，拥有套利优势的人往往选择保持沉默，对冲基金往往藏匿着能带来巨额经济效益的秘密。Collison 也联想到职业公路自行车界的一个著名案例：曾经所有车队都盲目将胎压充得极高（例如 120 PSI），直到某支环法车队测试发现较低胎压阻力更小并秘密保守了这一优势多年。 - **Cleo 规模与定位**：位于伦敦 Shoreditch，ARR 已达到 4 亿美元，为消费级 AI 理财助理独角兽。 - **线上扑克套利**：大学期间每日打 4-6 小时线上 heads-up；目前运行私人 AI 扑克代理，年收益约 20 万美元。 - **GTO 与认知差**：早期依靠经验剥削对手，如今线上局受 GTO 求解器影响很大，线下锦标赛依然充满乐趣。 --- ### [02:35](https://www.youtube.com/watch?v=70kWqOOWlAA&t=155s) Cleo 的技术演进：从正则表达式到大语言模型 谈话切入 Cleo 的创立历程。Cleo 成立于 2016 年，彼时做“管理金钱的 AI 助手”在行业内显得格格不入。当时自然语言处理（NLP）技术尚处早期，学术界刚刚经历凭借 AlexNet 掀起的计算机视觉革命，而 Barney 在攻读机器学习硕士学位期间，导师们开始将研究重心转向 NLP。在金融科技公司担任数据科学家后，Barney 坚信机器学习是未来趋势，并形成了一个核心假设：大部分普通人的日常财务决策处于严重未优化的状态，而财务压力又是人生中最沉重、最普遍的痛点。 Cleo 的核心愿景是让数亿人每天做出更合理的财务决策，甚至代替用户执行决策，如同建立一套属于个人财务的“GTO 策略”。Collison 指出 2016 年不仅远早于 ChatGPT，甚至比 Vaswani 等人提出 Transformer 架构还要早两年。在那个时代，技术团队依靠的是传统的有监督学习与意图分类。 Barney 坦率拆解了 Cleo 架构的演化路径。最初的原型系统其实充斥着大量用于文本匹配的“正则表达式”（Regex）；随后演变为成熟的多分类有监督学习架构，进行意图分类（Intent Classification）和槽位提取；同时结合推荐系统与强化学习（RL），在恰当的时机向用户推送个性化提示。直到 GPT-3.5 爆发，整个行业才被唤醒，而 Cleo 也顺理成章将底层架构全面迁移至大语言模型（LLM），从根本上重塑了产品交互与服务模式。 ```mermaid flowchart TD accTitle: Cleo 底层技术架构演进历程 accDescr: 展示 Cleo 从 2016 年早期的规则与监督学习发展到当今 LLM 驱动的技术演进图 subgraph Phase1["早期架构 (2016-2017)"] A["正则表达式 (Regex) 规则匹配"] B["基础多分类监督学习\n(意图识别与槽位提取)"] A --> B end subgraph Phase2["深度学习与推荐 (2018-2022)"] C["Transformer / BERT 架构"] D["强化学习 (RL) 与个性化推荐系统"] C --> D end subgraph Phase3["生成式 AI 时代 (2023 至今)"] E["大语言模型 (LLM) Agent 驱动"] F["自主财务推理与主动代办执行"] E --> F end Phase1 --> Phase2 Phase2 --> Phase3 ``` - **创立背景**：2016 年创立，基于“优化数十亿人日常财务决策”的愿景。 - **技术跃迁路径**： 1. 正则表达式规则库； 2. 传统有监督意图分类系统 + 强化学习/推荐系统； 3. 大语言模型（LLM）全面重构底层对话逻辑。 --- ### [05:01](https://www.youtube.com/watch?v=70kWqOOWlAA&t=301s) 99% 大众的财务真实世界：摆脱账单与数字的认知奴役 针对 Cleo 所主打的“个性化理财建议”，Collison 提出了一个理性的反思：常规的理财常识其实十分简单——保持合理储蓄率、购买低费率的被动指数基金（如 MSCI 全球指数而非单一国家的富时 100 指数）、远离过度营销的金融衍生品，既然原则如此明确，究竟为什么需要复杂的“个性化建议”？ Barney 随即反驳道："这暴露出你有多么脱离普通大众的生活，John。"（"showing how out of touch you are, John."）他指出，理财顾问服务在传统上只服务于前 1% 的高净值人群，但真正处于财务深渊中的是那 99% 的普通大众。在美国，有 60% 的成年人拿不出 400 美元的应急存款。对普通人而言，金钱不是资产配置的数学题，而是日常生活中最焦虑、最消耗心力的行政琐事（admin time）：从支付房租、应对浮动收入、信用卡高额利息，到每一笔经由 Stripe 结算的日常消费。 Barney 勾勒了他心目中未来个人财务的理想图景：用户不应该每天焦虑地盯着银行账户余额和冗长的交易明细清单，整个金融交互界面应当被隐形和自动化。如果需要信贷，AI 会自动寻找成本最低的方案；如果有节余，系统会自动完成最优的免税投资配置。无论是低端信贷产品令人咂舌的高利贷利率，还是英国大量居民将资金闲置在低收益现金 ISA（Cash ISA）中所错失的复利，都在社会层面上造成了万亿美元级的效率浪费。通过 AI 让资本和消费决策更加高效理性，不仅能挽救无数家庭的生活品质，更能直接释放巨大的宏观 GDP 增长潜力。 - **大众财务现实**：全美 60% 的人储蓄不足 400 美元，收入波动剧烈，承受着极高的无谓信贷利息负担。 - **传统理财盲区**：标准指数基金策略脱离底层现实，低收入群体面临的是日常开销、高息负债与现金流失衡。 - **Cleo 的愿景**：终结“看余额与账单流水”的陈旧管钱模式，由 AI 代理主动优化信贷、储蓄与投资决策，释放宏观经济增长活力。 ### [08:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=493s) 用幽默化解财务焦虑：从 30 位女性喜剧演员到 Cleo 的“吐槽模式” 谈话从千禧一代消费“牛油果吐司”的老梗展开。John Collison 询问 Barney，Cleo 的目标是试图帮用户减少这类不必要的消费，还是通过某种可落地的实用方式来引导用户的消费行为。Barney 笑着回顾了早期创业的趣事：当时底层模型远没有如今这么智能，团队只能严重依赖幽默感来维系互动。因为 Cleo 被设定为一个女性 AI 代理，公司专门雇佣了 30 位女性喜剧演员作为全职员工，负责撰写系统发出的每一行文案以及面对用户提问时的所有回复逻辑。他们意外发现，幽默与金钱这两个看似格格不入的领域，结合在一起却产生了惊人的协同效应——而这是任何传统金融机构或银行家绝不可能做到的，毕竟“只要跟法务和高管聊过一次，你就知道银行家向来以缺乏幽默感著称”。 这种幽默感成为了 Cleo 撬动用户参与度的杀手锏。产品中甚至包含一个极受欢迎的“吐槽模式”（Roast Mode），专门根据用户的账单开涮，比如调侃用户周末在酒吧喝了多少杯啤酒（Barney 顺势打趣称不知道该怎么吐槽 John，也许是周末喝的那些“肆意畅饮的小酒”）。用户非常乐意以这种轻松的方式获取简单、可执行的财务建议。如今，Cleo 正从单纯提供建议演进为更主动的代客操作——直接帮助用户转移账户、划转资金，并在最合适的时机为特定用户挑选最适合的金融产品。 ```mermaid flowchart TD accTitle: Cleo 产品定位与能力的演进历程 accDescr: 展示 Cleo 从早期依靠幽默与文案推动参与度，演进到依靠算法与自主代理进行主动代客操作的路径。 subgraph EarlyStage["早期：幽默驱动与文案互动"] A["30位女性喜剧演员撰写文案"] --> B["吐槽模式 (Roast Mode)"] B --> C["高参与度与轻松解压的财务建议"] end subgraph ModernStage["当下与未来：Agentic 主动代理"] D["构建全量用户财务画像与评分"] --> E["推荐系统权衡健康度与参与度"] E --> F["主动代客操作：调配资金、开户、精选金融产品"] end C -.-> D ``` ### [09:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=560s) 衡量财务健康：从单一参与度指标到“支出少于收入”的底层目标 当 John 问及 Cleo 如何具体量化用户财务健康的改善程度时，Barney 表示团队在内部构建了一套极其精细的个人财务健康评估体系。Cleo 为每一位用户建立了一个介于 0 到 100 分的综合健康度得分，该分值由大量的特征变量共同计算得出，从而能实时追踪用户的某项举动究竟是在提升还是在削弱其财务状况。 这一指标的成熟深刻改变了算法团队的优化目标。Barney 透露，Cleo 早期推荐系统的目标函数（objective function）极其单一，纯粹为了追求“参与度”（Engagement），即想方设法吸引用户点击按钮、制造好玩的体验；但现在，目标函数已经全面升级为“财务健康度 + 参与度”的双重平衡。算法关注的核心在于：给出的某条建议或采取的某项自动操作，能否在两到三个月后真正让用户的财务健康得分得到改善。对于中低收入群体而言，所有财务逻辑的核心原语（core primitive）其实极其朴素而关键——那就是确保“支出始终少于收入”。 ### [10:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=613s) “凭感觉做决策”与垂直智能体的未来 针对 Cleo 的竞争对手是谁、是否在替代现有金融工具的疑问，Barney 直言，当下大众做财务决定的方式本质上是一种“凭感觉的优化游戏”（vibe-based optimization）。绝大多数普通消费者从不记账，也不会理性规划收支，顶多随手在餐巾纸或包装盒背面写两笔，或者等到账单逾期才仓促应对。世界上只有极少数像台下听众这样具备技术背景的“极客”，才会专门拉 Excel 表格来严密规划自己的个人财务。因此，Cleo 争夺的并非存量财务软件的市场，而是面向那些此前从未得到过专业咨询的大众，开辟一个“无建议群体的建议市场”。 Barney 预判，人类生活中有大量关键决策领域长期处于严重缺乏优化的状态，除了个人财务，还有健康和教育等重大垂类。就像如今人们纷纷佩戴 Whoop 手环开始关注身体数据一样，未来这些垂直领域都将被高度智能化的专用系统重塑。他类比道，像 Stripe 这样的大公司在配置资本和人力时，背后有一整支由精算师与分析师组成的“智囊团”根据战略与长期目标进行周密测算；而 AI 代理的意义，就在于让每一个普通人在人生中的每一次决策都能享受到这种级别的智囊支持。 - **去 Excel 化的现实**：绝大多数低收入和中产大众并不具备通过电子表格规划财务的意愿与能力，理财行为多由情绪与直觉主导。 - **垂直智能体的价值**：Barney 认为未来不会是通用模型统揽全局，而是会出现 4 到 5 个在垂直领域扎根极深的专属 Agent（例如健康领域的 Agent 会一直打通到核磁共振 MRI 和血液检测数据）。 ### [11:58](https://www.youtube.com/watch?v=70kWqOOWlAA&t=718s) 为什么通用大模型做不好财务：主动代理对比被动搜索 针对 John 提到的竞争动态——许多用户可能会直接把银行对账单丢给通用大模型，或者将银行账户绑定至 Claude 或 ChatGPT 进行提问，甚至市场上已有类似将金融数据接口（如 Plaid）接入 ChatGPT 的插件方案——Barney 表达了坚决的看衰态度。他敢以巨额资金打赌，这类在通用 LLM 上的财务插件很难获得真实的用户留存，最终往往沦为无疾而终的边缘项目。 其核心原因在于人性的弱点：财务本身是枯燥乏味的，没有人会在早晨醒来时主动去想“我今天在某某方面花了多少钱”。通用大模型本质上是被动响应式（reactive）的，主要用于替代 Google 搜索，前提是用户脑中已经有了明确的问题要去检索；但要真正做好个人财务引导，工具必须具备“主动性”（proactive）和“代理能力”（agentic）。系统必须是一个始终在线（always-on）、主动向用户推送信息、并在后台代劳执行的操作主体。这需要海量的专属训练数据积累，以及极度精准地掌握在什么时间节点对用户说什么话才能奏效的交互理解，而这正是底层通用模型厂商所欠缺的。 ### [13:17](https://www.youtube.com/watch?v=70kWqOOWlAA&t=797s) 记忆与个性化：如何打破《记忆碎片》式的模型局限 谈及大模型的个性化难题时，John 提出了一个深刻的行业隐喻：底层的静态大模型搭配外部平面文件的记忆机制，非常像电影《记忆碎片》（Memento）里的男主角——因为没有长期记忆，只能把线索密密麻麻刺青在自己手上，每次醒来都要重新阅读笔记，当下的 AI 架构似乎在本质上对长期个性化存在抗拒。 ```mermaid flowchart LR accTitle: Cleo 个性化知识库与决策架构 accDescr: 说明 Cleo 如何将非结构化对话解析为结构化记忆，并结合推荐系统驱动下游行动。 subgraph DataIngestion["持续数据沉淀"] A["多轮历史对话"] B["银行交易流水与收入"] C["用户财务愿望与目标"] end subgraph MemoryEngine["结构化记忆层"] D["递归 LLM 信息提取"] E["个人财务知识库<br/>(规范化 Schema 与数据库)"] end subgraph ActionEngine["推荐与执行引擎"] F["目标函数评估<br/>(财务健康 + 参与度)"] G["最佳沟通时机与话术"] H["后台自动化执行<br/>(资金调度/产品推荐)"] end A & B & C --> D D --> E E --> F F --> G & H ``` Barney 对此给出了完全不同的实操视角。他指出，尽管底层数据形态目前仍是文本，但在工程实践中，Cleo 早已通过递归调用 LLM（Recursive LLMs）构建了一套高度结构化的个人专属知识库。系统不仅记录用户的每一次历史对话，还沉淀了用户的财务诉求、收入水平、消费轨迹以及财务健康分。在此基础上，由上层的推荐系统持续扫描该知识库，从所有潜在沟通契机中计算出“哪一句话能最大程度撬动目标函数”，进而输出建议并在后台触发动作。Barney 甚至当场调侃 Stripe 对用户数据的理解还不够智能，认为 Stripe 也应当为每个绑定卡片的用户建立这样一套完备的知识库结构。在 Barney 看来，只要数据存储在设计严密、结构清晰的 Schema 和数据库中，以文本为基础的记忆机制就足以撑起深度的个性化。 ### [15:01](https://www.youtube.com/watch?v=70kWqOOWlAA&t=901s) 赋予 AI 性格：不仅是风格，更是驱动商业指标的核心引擎 随着话题深入，John 谈到两人在后台聊起的一个技术哲学问题：AI 的“个性”（Personality）在产品中究竟扮演什么角色。早年行业低估了性格的作用，如今从 Anthropic 推出的新 Claude 模型的独特语调，到 GPT-4o 因过于阿谀奉承（sycophantic）而引发巨大争议（一部分用户极度反感其谄媚，另一部分人却沉迷其中），性格设计正变得愈发微妙。John 很好奇 Cleo 如何把握性格的分寸，是针对不同用户动态调整，还是坚持自成一体的风格，又是否已经摆脱了依赖大量喜剧演员手写脚本的阶段。 Barney 回应称，基础模型实验室（Frontier Labs）或许可以把通用模型做得相对平淡中立，以纯粹的智能化通用服务为导向；但对于任何直面消费者的品牌和商业应用来说，AI 的性格塑造绝非浮于表面的点缀，而是直接驱动核心商业指标的引擎。对话是否幽默、表述是否精炼干脆、行文质感是否出众，都会在后台数据上带来立竿见影的反馈——它直接关乎月活跃用户数（MAU）的起伏与公司实际收入的增减。Cleo 在创业最初的三到六个月内便彻底确立了这一认知：在这门生意中，品牌调性、人格设定与底层技术同等关键。 ### [16:35](https://www.youtube.com/watch?v=70kWqOOWlAA&t=995s) 打造逆向产品体验：从个性化语气到对冲金融痛点 Cleo 的核心竞争力之一在于其独特的人格化表达（Tone of Voice）。在早期阶段，团队依赖线性监督学习，整个产品基本上只有一种预设的对话语调；但随着大语言模型（LLM）的引入，Cleo 实现了高度个性化的表达机制——系统能够根据用户的真实收入、财务健康状况乃至地理位置（例如波士顿与旧金山用户的文化差异）动态调整交流方式。Barney 打趣道，他本想在台下 1500 名观众面前拿 John 开涮，但理智制止了他；而 Cleo 同样清楚该如何恰如其分地与用户沟通，用户甚至可以主动微调这种语调风格。当 John 问及这是否是为了缓解大众对财务问题的普遍焦虑感与沉重心理时，Barney 直言，打造伟大产品的捷径就是采取“极端的逆向思维”。 这种产品逆向法则在优秀科技公司身上屡试不爽。正如传统支付流程不仅晦涩死板，还充斥着繁冗的企业销售流程，Stripe 却反其道而行之，将其打造成极致简洁、对开发者极其友好的单点击穿体验；数字银行 Monzo 同样是将传统银行反人类的晦涩交互改造成透明、直观的用户界面。面对市面上普遍“枯燥乏味且复杂难懂”的个人财务咨询，Cleo 选择将其彻底重构成“轻松有趣且极度简化”的产品形态。在 Barney 看来，在创业初期的冷启动阶段，找到传统既得利益者（Incumbents）最严重的用户痛点，并以最激进、最极致的方式反向进攻，创造令人惊叹的愉悦体验，是建立产品差异化（Product Differentiation）最有效的破局点。 - **模型迭代跨越**：从初期的单一口吻线性监督学习，演进至基于地理位置、收入与财务指标动态生成的 LLM 深度个性化对话。 - **逆向定位哲学**：找准行业巨头的最大痛点反向发力，将“高门槛、沉重枯燥”的金融沟通重塑为“低门槛、富有趣味”的情感化互动。 --- ### [18:14](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1094s) 金融基语的原罪与“智能建议”的降维打击 针对金融产品形态的演变，Barney 表达了相当清醒乃至略带批判的技术视角。他指出，货币本质上的基础原语（Primitives）极其有限，无非是借贷、支出与储蓄，这决定了所有金融产品的创新速度天然迟缓且严重受制于监管政策。尽管先买后付（BNPL）作为信用卡机制的改良替代品在近年迅速崛起，但在底层逻辑上仍属渐进式优化；相比之下，未来可能带来真正范式转移的反而是稳定币（Stablecoins）对底层清算网络的重构。Barney 甚至调侃称，虽然自己的团队在持续开发大量金融工具，但他个人常常觉得单纯搭建金融产品受限过多且略显乏味，真正具备巨大想象力的是构建在交易之上的“智能与建议层”（Intelligence and Advice Layer）。 传统信用体系的痛点在于用户极易陷入非理性的负债深渊：大量消费者一旦办了信用卡，往往会迅速刷爆额度，随后在长达十年的时间里仅按月偿还最低还款额，这构成了社会生产力中最缺乏效率的信贷形式。尽管 BNPL 提供了一种更偏向分期定期贷款（Term Product）的相对健康的偿还方式，但若用户仅将其视为借取更多资金的借口，同样会加剧财务恶化。Barney 强调，金融企业如果缺乏智能建议层的约束，就极易滑向“纯粹利润最大化”的歧途；以 Robinhood 等产品为例，若企业仅专注于销售金融衍生品和追逐最高利差，便会诱导平台走向过度推崇日内高频交易（Day Trading）的方向，最终损害消费者福利与企业的长期品牌价值。 ```mermaid flowchart TD accTitle: 金融产品商业模式对比 accDescr: 对比纯产品驱动与智能建议驱动两种金融服务路径与最终结果 subgraph Trap["纯利润驱动模式（传统金融 / 经纪商）"] A["单一金融产品销售"] --> B["追求利差与交易量最大化"] B --> C["诱导高频投机 / 信用卡刷爆十年还底线"] C --> D["用户财务崩溃 & 商业模式走向负和博弈"] end subgraph Intelligence["智能建议驱动模式（Cleo 路径）"] E["有限金融原语<br/>(借贷/支付/存储)"] --> F["AI 智能决策与个性化建议层"] F --> G["优化借贷结构 / 抑制非理性消费"] G --> H["长期健康财务留存 & 正和商业生态"] end ``` - **信贷顽疾**：传统信用卡模式的顽疾在于诱导用户长期顶格透支并支付最低还款，沦为极度低效的资产剥削。 - **纯销售模式的反思**：缺乏良性指导的纯金融销售模式必然倒向投机与高频交易，唯有将 AI 决策注入消费前端，才能扭转负和博弈。 --- ### [20:59](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1259s) 英国监管的“沙盒困局”与美国分权制衡的活力 探讨到英国金融监管体系的现状时，Barney 发出了直截了当的抨击。他认为，“创造性破坏”（Creative Destruction）是过去一个世纪以来推动各行各业向前迈进、为消费者和企业创造长远利益的核心引擎。然而，英国的监管文化本质上过度厌恶风险，严重压制了这种破坏性创新。英国金融行为监管局（FCA）往往热衷于邀请初创企业进入监管沙盒（Regulatory Sandbox）共同磨合两年之久，但风险投资的标准融资生命周期通常仅有 18 个月；监管机构本意或许是提供合规指导，实际上却直接用漫长的审批周期拖垮了初创团队。Barney 透露，自己作为天使投资人目睹了太多早期项目因在 FCA 的合规流程中耗尽现金储备，最终甚至无法活到验证产品市场契合度（PMF）的那一天。 这种机制缺陷进一步暴露出深层的制度文化差异。Barney 借用《联邦党人文集》（Federalist Papers）所阐述的分权制衡理念指出，美国通过联邦与州政府、法院与各级监管机构之间多维度的制衡，为市场竞争和资本主义提供了动态纠偏的制度弹性，避免了单点决策的绝对阻断；反观英国，首相府（唐宁街 10 号）与监管机构集权度极高，单凭一纸裁决便能直接左右六七千万国民市场的准入与生死。正是这种制衡机制的缺乏，促使 Barney 坚定地将 Cleo 的全部核心用户和业务重心布局在美国——尽管美国社会伴随着剧烈的两极分化与政治纷争，但其制度所蕴含的理性与包容度，仍是其创新活力远超欧洲的关键所在。 - **监管沙盒的时间错配**：英国 FCA 长达 24 个月的合规审查周期与风投企业 18 个月的融资生存窗口存在严重脱节。 - **制度制衡的决定性优势**：相较于英国自上而下的高度中央集权管制，美国多元制衡的司法与监管框架反而赋予了初创公司挑战巨头的法律回旋空间。 --- ### [23:11](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1391s) 告别遗老守成：从养老金资本到“全险”（Risk-on）创新 跳出微观的金融监管，Barney 将视野放大至整个宏观商业环境与国民心智。他犀利地指出，英国乃至整个欧洲必须彻底走出沉溺于历史荣耀、试图固守往日帝国遗产的“守成文化”。在经济增长停滞、产业全面落后的大背景下，欧洲眼下早已没有什么资产是值得抱残守缺的；唯一的自救方案就是将高度保守的规避风险（Risk-off）心态，全面转变为勇敢下注、全力建设并敢于承担风险的“全险”（Risk-on）创造性破坏模式。这意味着必须彻底激活国内的沉睡资本——包括养老基金、个人储蓄账户（ISA）等海量资金池，将其引导至推动社会生产力突破的技术创新领域。 在被问及个体与机构应如何修正投资行为时，Barney 认为，个人层面的投资逻辑在理论上早已被彻底解决：最理性的做法无非是通过全球指数基金或标普 500（S&P 500）进行长期配置（尽管由于科技巨头权重过高，加上未来 Stripe 一旦上市可能让指数科技属性更重），乃至像 Brad Gerstner 倡导的“投资美国”（Invest America）儿童账户那样尽早开启复利投资。然而，英美之间真正的天壤之别在于机构端：美国的庞大养老金长期作为耐心的风险资本，源源不断地注入科技创新与创造性破坏之中；而英国与欧洲的巨额养老基金却依然被深锁在极度保守的传统资产中，未能成为赋能本土技术革命的活水。 - **文化心态断层**：欧洲亟需从“保护衰落遗产”的防守型逻辑，切换至以部署创新资本为导向的“进攻型”冒险精神。 - **机构资本错配**：个人定投全球指数已是成熟解法，但欧洲养老金对高风险、高回报前沿科技领域的严重低配，正在拉大欧美两地的财富与创新鸿沟。 ### [25:05](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1505s) 机构资本缺位与欧洲科技生态的隐忧 Barney Hussey-Yeo 谈到了英国及欧洲风险投资结构中一个令人警醒的现实：在英国，只有极小一部分养老金流向了风险投资领域。即便他接触过不少英国顶尖风投或欧洲一线 VC，这些基金背后的有限合伙人（LP）往往还是加利福尼亚州的养老基金。这意味着，创业公司创造的技术成果与 GDP 增长，最终并没有实质性地回馈给英国本土经济，而是全部回流到了加州。Barney 警告称，如果不希望未来的世界演变成只有中国和美国两极霸权的科技格局，欧洲就必须打破现状，通过“创造性破坏”（creative destruction）重新部署其庞大的机构资本。 John Collison 对此表示赞同，指出美国、加拿大以及澳大利亚的养老金体系长期以来都在大规模押注风险投资，而英国和欧洲大陆在这方面明显滞后。Barney 感慨道，看看加拿大和澳大利亚的养老基金回报，他甚至笑称“我都想去当一名澳大利亚的退休老人了”。John 顺势补充了一个近期的知名案例：有家养老基金在数年前重仓了 SpaceX 的股份，随着 SpaceX 近期估值飙升与商业进展，“可想而知他们这周的心情有多好”。 - 英国养老金注入本土风投的比例极低，大量欧洲一线 VC 的核心出资人（LP）实际来自美国加州。 - 加拿大和澳大利亚的公共养老金因早期重仓科技与风投（如 SpaceX）而获得了丰厚回报。 - 如果欧洲无法调动本土机构资本支持创新，将面临在全球科技两极化中彻底边缘化的长期风险。 --- ### [26:09](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1569s) 消费级 AI 的商业化困惑：规模先行还是首日变现？ 谈及消费级 AI 应用与商业模式时，John Collison 询问 Cleo 的具体盈利方式。Barney 解释称，Cleo 目前采用免费增值（Freemium）模式，营收主要来自订阅服务和金融产品两大部分。金融产品涵盖了联名卡、信贷、先买后付（BNPL）、预支工资（Earned Wage Access）以及股票与资产管理等财富产品，订阅与金融产品协同沉淀用户价值并贡献利润。 当被问及对当下构建消费级 AI 产品的创业者有何商业化建议时，Barney 坦率回顾了 Cleo 的成长历程。他坦言自己创业的前四年可能深受零利率时代（ZIRP）思潮的影响，当时团队的目标几乎全部围绕着拉动月活跃用户（MAU）展开，大家沉浸在“专注于财务健康、零收入也没关系”的氛围中。然而，当公司推进到成长型基金（Growth funds）尽职调查时，投资人迎面抛来的第一句话就是：“你们的收入在哪里？” 尽管走过弯路，Barney 依然认为在消费端，先做出一款用户真正喜爱且高频使用的产品、随后再嫁接商业化，仍然是一条行得通的路径。但他给创业者的核心建议是“绝对不要搞不痛不痒的折中策略”：要么像消费应用一样先全力以赴拿下海量用户，把商业化暂搁一边；要么像 Stripe 那样，从第一天起就把底层抽佣机制（如交易佣金、Token 费用等 Core Primitive）百分之百无缝融入核心体验中。 ```mermaid flowchart TD accTitle: 消费级与企业级商业化路径对比 accDescr: 对比了先获取规模后变现的消费级模式与第一天内置商业化基石的企业级模式。 subgraph Consumer_Path["消费端路径（纯体验优先）"] C1["打磨极致体验与产品契合度"] --> C2["专注做大活跃用户体量 (MAU)"] C2 --> C3["后期引入订阅制与金融产品增值"] end subgraph Stripe_Path["基石机制路径（首日变现）"] S1["定义核心计费基石 (Core Primitive)"] --> S2["首日将交易抽成/Token计费注入架构"] S2 --> S3["伴随业务规模线性扩张产生健康现金流"] end ``` --- ### [27:52](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1672s) 行为经济学、沉没成本与失控心理学 John Collison 抛出了商业模式的深层取舍：零售界既有像普通超市那样靠单品加价盈利的按量模式，也有类似 Costco 的模式——货架商品几乎不赚差价，全部利润都源自每年约 100 美元的会员费。Barney 解释道，Cleo 之所以同时涉足订阅与金融产品，一部分原因在于当前的传统支付通道（payment rails）对高频小额微交易极其不友好，掉单率高且固定成本沉重，因而按月循环订阅成为了当下软件业最稳健的载体。但他预测，随着未来稳定币（stablecoins）和零摩擦无成本的底层清结算通道成熟，西方软件的商业形态可能会大幅转向类似于中国超级应用的微交易与按需付费模式。 聊到消费者心理，John 提到许多自我提升类产品往往利用了“沉没成本”机制——用户办理健身卡正是因为“既然付了钱，为了不浪费就必须去锻炼”。Barney 承认，订阅体系充满了这类非理性行为学。例如 Cleo 最近增加了第三个订阅档位，正是利用了心理学中的折中效应：“仅仅因为多设置了一个中间档，大部分人就会自然而然选择中间那档。” John 戏谑道：“你们这有点像 Duolingo，把人们的非理性‘武器化’来为他们谋福利，而不是作恶。”Barney 笑着回应让媒体别这么写，更准确的说应当是“驾驭”（harnessing）。这自然引申到了德州扑克中著名的“上头”（on tilt）现象——平时理性的选手在遭受连败或挫折后会情绪失控、胡乱下注。Barney 指出，金钱是极具情绪化和心理敏感度的事物，当用户经历失业或失恋等生活剧变时，极易陷入财务失控与非理性消费的深渊，而 Cleo 的核心任务正是构建一套智能机制与功能，在关键时刻将滑向深渊的用户拉回理性的轨道。 --- ### [31:23](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1883s) AI 原生工作流重塑管理层与“船长日志” 在被问及 Cleo 作为一家年轻且快速扩张的企业如何实现“AI 原生”（AI-native）运作时，Barney 透露自己的日常工作节奏在近几个月迎来了彻底颠覆。过去在管理一个 500 人的组织时，CEO 的标准工作是逐一与各个业务线负责人（division leaders）沟通，搜集碎片化信息，阅读大量的会议预读材料（pre-reads），再费尽口舌提供反馈。这种自下而上的传统信息流转不仅效率低下，而且充满认知损耗。 如今，Barney 践行了 Andrej Karpathy 所倡导的信息聚合工作流：通过大语言模型直接全面梳理公司内部沉淀的每一条 Slack 消息、Notion 文档以及开发团队提交的每一个 PR（Pull Request）。LLM 不仅会为每一次代码合并提取摘要，还会对其质量进行评估和打分，从而直接构建出一个跨越整个组织的全局知识库。 John 将这种模式形象地比喻为一名全天候巡视船舱各处的数字大副，时刻拿着记录夹记录着“船长日志”（captain's log）。Barney 补充道，这种全景透视是传统人力根本无法做到的尺度。现在，员工制作的仪表板、编写的预读材料与代码输出全部被自动聚合，使他能以一种直观而清晰的方式随时洞察全局研发脉搏。 ```mermaid flowchart LR accTitle: 基于 LLM 的组织自动化全景日志流 accDescr: 展示了从底层工作数据到 LLM 知识库再到 CEO 决策层的信息聚合流程。 subgraph Data_Layer["分散数据源"] D1["Slack 沟通频道"] D2["Notion 知识文档"] D3["GitHub 代码 PR"] D4["员工预读材料与仪表板"] end subgraph Processing_Layer["LLM 认知中枢"] P1["PR 自动评分与摘要"] P2["跨部门上下文知识库构建"] P3["高维度动态信息聚合"] end subgraph Leadership_Layer["管理决策端"] L1["全局'船长日志' (Captain's Log)"] L2["直接穿透式组织洞察"] end D1 & D2 & D3 & D4 --> P1 & P2 & P3 P1 & P2 & P3 --> L1 --> L2 ``` ### [33:31](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2011s) 数据穿透管理层：LLM 驱动的组织扁平化革命 随着企业内部数据权限体系与安全工具的逐步打通，高管直接通过 AI 获取底层真实业务信息的现象正在重塑管理动态。John Collison 指出，消费级 AI 工具凭借全网知识库与网络搜索已经展现出惊人的实用性，而企业当下最大的课题是如何在权衡权限与数据安全的前提下，向内开放数据资产。他随即向 Barney Hussey-Yeo 抛出一个尖锐的问题：当 CEO 可以绕过中层汇报、直接调取第一手源头数据时，内部的管理团队是否快要被逼疯了？ Barney 毫不掩饰地坦白：“我的工程经理们恨死我了（My engineering managers hate me）。”在过去，CEO 看到的团队绩效是被层层中层管理者过滤后的视图，不可避免地夹杂着不同管理者的个人偏见与利益诉求。而现在，通过大语言模型直接接入工程数据上下文，Barney 能够获得极其理性、清晰的事实依据，实时掌握谁在真正推进交付（shipping）、谁的代码质量最高、谁的迭代速率最快，以及究竟是哪些由工程师组成的底层团队在产出卓越成果。他断言，这种技术穿透必将让 Cleo 乃至更多企业的中层架构急剧扁平化，打破传统层级汇报的信息阻滞，实现远比过去高效的信息流动。 ```mermaid flowchart TD accTitle: 传统层级管理与 AI 驱动扁平化管理对比 accDescr: 对比传统通过中层汇报过滤信息，与如今高管通过 LLM 直接穿透底层代码与工程数据的两种模式。 subgraph Traditional["传统汇报模式"] E1["底层工程师 / 交付团队"] --> M1["中层管理者<br/>（夹带偏见与考核诉求）"] M1 --> L1["管理层<br/>（获取被过滤的偏差视图）"] end subgraph AIFlat["AI 驱动扁平化模式"] E2["底层工程与业务数据源"] --> LLM["企业上下文 LLM"] LLM --> L2["高管直达源头<br/>（精准洞察产出、速率与代码质量）"] end ``` - **数据去偏见化**：LLM 结合企业内部真实数据源，消除了汇报链条中由“中层管理者个人偏见与激励机制”带来的信息扭曲。 - **透明度倒逼**：工程产出、迭代速度与代码质量变得直观可见，谁在真正交付、谁在摸鱼无所遁形。 --- ### [34:46](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2086s) 重新定义组织构架：从层级传帮带到“AI-Pill”青年指挥 Agent 军团 当谈及扁平化是否会在未来两年的管理幅度（Span of Control）数字上实质性显现时，Barney 表示自己向来不赞成主管每天必须与所有下属进行一对一交流的教条模式。他殷切期望未来的中层经理数量大幅减少，但他同时也作出了审慎的保留：组织永远需要对年轻员工进行指导、投入时间并帮助其成长。他开玩笑地调侃现场台下或许就坐着半个被繁复职级层层包裹的 Stripe 团队，但他坚信这种金字塔结构必然会转变。未来的组织或许不再追求臃肿的层级与庞大的事业部划分，而是能够有机拆分、以极扁平的形态同时下注更多平行实验。 John 进而探讨了人才培养方式是否会回归传统的“学徒制”（Apprenticeship Model）——即由一位资深专家专精指导一位年轻下属，而非传统职能架构下一位主管粗放管理八个人的模式。Barney 则提出了完全相反的激进构想：由一个极具品味的年轻人去指挥数十个乃至成百上千个 Agent。 Barney 透露，Cleo 最近从剑桥大学（Cambridge）和加州大学洛杉矶分校（UCLA）招募了一批机器学习硕士毕业生。这群年轻人从中学 A-level 阶段起就深浸在 AI 技术中，是彻底被“AI 药丸”浸润（fully AI-pilled）的新一代。在软件需求无限膨胀的时代，企业最稀缺的是拥有出色“品味”（taste）的决策者，去指挥并调度庞大的并行 Agent 舰队与递归流程，消耗越来越多的 Token。相比过去一年只能谨小慎微地下 10 个业务赌注，Barney 更向往的世界是凭借这些年轻团队与 Agent 集群，在下一年同时启动 2,000 个业务探索。 - **组织范式重塑**：传统的一对多层级管理演化为由精通 AI 工具的初级员工驱动大规模并行 Agent 舰队。 - **年轻人才优势**：新一代名校毕业生在思维上完全“AI 化”，熟稔 Agent 工具链，兼具提示工程能力与工程品味。 - **业务试错密度跃升**：在极低的生成与验证边际成本下，初创企业可从原先一年试水 10 个项目，飞跃至一年尝试 2,000 个并行探索。 --- ### [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) 最逆向的 AI 认知：“SaaS 灭顶之灾”纯属恐慌过度 面对 John 提出的最后一个问题——当下最逆向（contrarian）的 AI 观点是什么，Barney 坦言作为一个极度的“AI 药丸信徒”，要想提出完全逆向的观点其实很难，但他坚决反对目前市场上关于“SaaS 灭顶之灾”（SaaS apocalypse）的论调。外界与社交媒体上充斥着一种盲目言论：似乎前沿基础模型实验室（Frontier Labs）只要发布一个微小的 Wrapper（外壳工具），就能凭借几个 Prompt 彻底颠覆并毁灭一整个行业。 Barney 强调，打造像 Stripe 这样的伟大企业背后需要极其深厚的行业工艺与沉淀。他反问 John：“在支付领域，你到底学了多少全世界根本没人需要掌握的琐碎细节？”John 随即幽默附和道：“那些东西根本不该有人去学……太可怕了，实在太多了。”Barney 指出，正是这些极其深奥、复杂的行业门槛与工艺匠心，构成了纵深极强、垂直立足的优秀软件企业护城河。反观资本市场，Barney 毫不留情地讽刺许多投资者是“这辈子见过最愚蠢、最不理性的群体”——由于缺乏技术常识、完全看不懂技术演进的真实逻辑，他们索性选择观望并一窝蜂地把钱扎堆灌给最顶尖的前沿模型实验室（以及像 Stripe 这样的巨头）。 Barney 坚信，这种市场非理性将在未来几年内回归常态，估值倍数也将修复。在过去三十年中，软件一直是驱动全球经济与 GDP 增长的核心引擎，未来依然如此。AI 不会消灭软件产业，反而会催生更多的软件公司、带来更充沛的“创造性破坏”（Creative Destruction）与资源丰饶（Abundance）。John 对此表示赞同并总结道，整个行业的传奇故事远未定稿，绝不可能只由少数几家巨头垄断所有实用价值的交付。随后，John 正式宣布本次 Tour London 下午场活动闭幕，并邀请全场嘉宾前往展厅参加后续酒会交流。 - **驳斥 SaaS 终结论**：单凭前沿模型写个 Prompt 无法抹平垂直领域数十年来积累的系统级复杂性与业务工艺。 - **投资界非理性羊群效应**：由于缺乏对底层技术的理解，大量风投资本放弃了对垂直应用的研究，盲目涌入头部基础模型公司避险。 - **挺“丰饶论”（Pro-abundance）**：AI 将作为催化剂激发更大规模的创造性破坏，未来经济增长的源泉依旧来自更广泛、更多元化的软件生态。 ## Notable quotes > [05:01](https://www.youtube.com/watch?v=70kWqOOWlAA&t=301s) 这暴露出你有多么脱离普通大众的生活，John。 — Barney Hussey-Yeo > [08:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=493s) 只要跟法务和高管聊过一次，你就知道银行家向来以缺乏幽默感著称 — Barney Hussey-Yeo > [10:13](https://www.youtube.com/watch?v=70kWqOOWlAA&t=613s) 凭感觉的优化游戏 — Barney Hussey-Yeo > [25:05](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1505s) 我都想去当一名澳大利亚的退休老人了 — Barney Hussey-Yeo > [26:09](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1569s) 你们的收入在哪里？ — 成长型基金投资人 > [26:09](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1569s) 绝对不要搞不痛不痒的折中策略 — Barney Hussey-Yeo > [27:52](https://www.youtube.com/watch?v=70kWqOOWlAA&t=1672s) 你们这有点像 Duolingo，把人们的非理性‘武器化’来为他们谋福利，而不是作恶。 — John Collison > [33:31](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2011s) 我的工程经理们恨死我了 — Barney Hussey-Yeo > [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) 在支付领域，你到底学了多少全世界根本没人需要掌握的琐碎细节？ — Barney Hussey-Yeo > [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) 那些东西根本不该有人去学……太可怕了，实在太多了。 — John Collison > [37:20](https://www.youtube.com/watch?v=70kWqOOWlAA&t=2240s) 这辈子见过最愚蠢、最不理性的群体 — Barney Hussey-Yeo ## Mentioned Barney Hussey-Yeo, John Collison, Cleo, Stripe, Tour London, 《赌王之王》（Rounders）, 《茉莉牌局》（Molly's Game）, 《记忆碎片》（Memento）, 《联邦党人文集》（Federalist Papers）, Andrej Karpathy, Brad Gerstner, Invest America, ChatGPT, GPT-3.5, GPT-4o, Claude, Anthropic, Plaid, Monzo, Robinhood, Costco, Duolingo, Whoop, SpaceX, 英国金融行为监管局 (FCA), 剑桥大学 (University of Cambridge), 加州大学洛杉矶分校 (UCLA), MSCI 全球指数, 富时 100 指数 (FTSE 100), 标普 500 (S&P 500), Slack, Notion, GitHub ## Video description Cleo founder and CEO Barney Hussey-Yeo joins Stripe CEO John Collison for a fireside chat at Tour London. Register for Stripe Tour 2026: https://stripetour.com [Source](https://www.youtube.com/watch?v=70kWqOOWlAA)

[5:01] Barney Hussey-Yeo: 这暴露出你有多么脱离普通大众的生活，John。 [8:13] Barney Hussey-Yeo: 只要跟法务和高管聊过一次，你就知道银行家向来以缺乏幽默感著称 [10:13] Barney Hussey-Yeo: 凭感觉的优化游戏 [25:05] Barney Hussey-Yeo: 我都想去当一名澳大利亚的退休老人了 [26:09] 成长型基金投资人: 你们的收入在哪里？ [26:09] Barney Hussey-Yeo: 绝对不要搞不痛不痒的折中策略 [27:52] John Collison: 你们这有点像 Duolingo，把人们的非理性‘武器化’来为他们谋福利，而不是作恶。 [33:31] Barney Hussey-Yeo: 我的工程经理们恨死我了 [37:20] Barney Hussey-Yeo: 在支付领域，你到底学了多少全世界根本没人需要掌握的琐碎细节？ [37:20] John Collison: 那些东西根本不该有人去学……太可怕了，实在太多了。 [37:20] Barney Hussey-Yeo: 这辈子见过最愚蠢、最不理性的群体

[0:00] Barney Hussey-Yeo: I shouldn't say this publicly, but I earn about 200k every year now just through my little agent that plays and barely beats the low stakes [2:35] Barney Hussey-Yeo: The thesis was everyone's decisions are wildly unoptimized and your financial life is the most stressful and important thing in your life. So, how could you make a billion people make better decisions every day or make decisions on their behalf? [4:20] Barney Hussey-Yeo: It started with regexes. We had a lot of regexes to start with. [5:46] Barney Hussey-Yeo: how out of touch you are, John, with the real consumer and normal humans on the street. [10:15] Barney Hussey-Yeo: To engage people with financial advice, it's got to be proactive, and it's got to be agentic. So people don't wake up and go, 'Ooh, how much did I spend?'... It has to be this agent always on, always pushing information. [18:14] Barney Hussey-Yeo: If you just provide financial advice or provide financial products without the advice... you end up just selling products and trying to get the most margin, and it leads you down this way of like, 'Ooh, let's do day trading.'... I think the intelligence and the decision-making and helping users do it for them is just so critical. [23:10] Barney Hussey-Yeo: I'm not returning to the British state any GDP growth. It's all going back to California. And if you don't want a world which is completely bipolar where you can either build tech in China or you can build tech in the US, it's really bad for Europe. [31:23] Barney Hussey-Yeo: I know who's shipping. I know who is the highest velocity, highest code quality... before I had this kind of weird view filtered through all these leaders that had all their different biases and all these different incentive structures. [34:46] Barney Hussey-Yeo: I like the idea that I can take 2,000 bets next year instead of like the 10 that I took last year.

[0:00] Thanks for being here. [0:00] >> Thanks for having me. [0:01] >> So, [0:02] Cleo, you guys are presumably familiar [0:05] with a phenom in the [0:07] consumer financial advice space, 400 [0:10] million in ARR. [0:12] >> As of last month. [0:13] >> Yes. [0:14] Unicorn [0:16] and really rapidly growing, one of the [0:18] top AI startups and based right here in [0:20] Shoreditch. Um, but before you started [0:23] Cleo, I want to go back to your prior [0:25] career. [0:26] >> Yeah. [0:26] >> You were a poker shark. Is that right? [0:29] >> Uh, [0:30] shark or fish? I'm I'm I played at [0:32] university. I funded my university [0:34] career through playing online poker. Um [0:36] >> I only know what I've seen in Rounders [0:39] and Molly's Game [laughter] and all this [0:40] kind of stuff. So, how does actually [0:43] making like professionally [0:45] playing poker work? [0:46] >> This was back before you had GTO and the [0:49] solvers. So, it's it's all online and [0:51] you know, play on like at multiple sites [0:53] and you'd play four to six hours a day [0:55] heads-up. And actually really [0:56] interestingly, I've just built an AI [0:58] agent that now plays for me and earns [1:00] rake back. So, I make about I shouldn't [1:02] say this publicly, but I earn about 200k [1:04] every year now just through my little [1:06] agent that plays and barely beats the [1:08] low stakes. But yeah, I used to play, [1:10] you know, low mid stakes heads-up and it [1:12] was quite profitable. [1:13] >> And how So, GTO is [1:16] game theory optimal, which is doing the [1:18] exact right thing that the probability [1:21] tables would tell you that you should [1:22] do, right? [1:23] >> And you didn't have that in 2012, '13, [1:26] '14 when I was playing. [1:27] >> Really? [1:29] >> you had the solvers and we actually [1:31] learned how you could build beat the [1:32] game. You did still ride the, you know, [1:35] the edge. You can probably beat the GTO [1:37] solvers, but it's basically a solved [1:39] game, which means online play is kind of [1:42] is dead now, but you should play [1:43] tournaments. You go play a live [1:44] tournament, they're always fun. [1:46] >> Well, cuz I'd be playing against your [1:47] bots probably. [1:48] >> Yeah, don't play online against me. [1:49] >> Okay. Okay. You're going to get rinsed. [1:51] That's very interesting. You know, it [1:52] was um [1:53] a lot of these things are discovered [1:54] later than one would think. Like [1:56] Black-Scholes is, you know, a relatively [1:58] recent invention for, you know, valuing [2:00] financial options. And I didn't realize [2:01] that GTO is such a recent invention in [2:04] poker. [2:04] >> Well, you don't want to tell people if [2:05] you've got an edge, right? So, there's [2:07] probably all these things in hedge funds [2:08] where you could have a massive economic [2:11] benefit to GDP, which you'd probably [2:12] keep a secret. [2:13] >> Exactly. Yeah, this happened in cycling [2:15] as well, in pro cycling, where people [2:17] used to inflate their tires up really [2:19] high, uh you know, 120 PSI or something [2:21] like that. And then one Tour de France [2:23] team figured out that actually everyone [2:25] was over inflating their tires and you [2:26] don't need to over inflate your tires. [2:28] And they kept the secret from everyone [2:30] for a long time. Anyway, there's still [2:32] plenty of room for new inventions is [2:33] what we're taking away here. But we are [2:35] not here to talk about pro cycling or [2:37] pro poker. We're here to talk about [2:38] Cleo. So, maybe just to orient people [2:40] who aren't familiar or aren't users of [2:41] the product, what does it do? Um how'd [2:44] you guys get here? [2:45] >> Yeah, well, we started in 2016. So, it's [2:47] an AI assistant for your money, which [2:49] was weird in 2016 and now is like the [2:52] thing to be building. [2:53] >> AI was not as hot in 2016. [2:55] >> Yeah, building an AI system was [2:56] completely weird. Like, natural language [2:57] processing was barely working. But it [2:59] was the really big research output at [3:01] the time. So, I did my masters in [3:02] machine learning and we'd just gone from [3:04] the vision era of like we'd beaten [3:06] AlexNet and then NLP became the thing [3:09] all my professors were interested in. [3:10] >> Yep. [3:10] >> So, I'd worked at a fintech as a data [3:12] scientist and I had this, you know, [3:15] fascination with machine learning and [3:16] knew it was going in that direction. So, [3:18] the the thesis was everyone's decisions [3:20] are wildly unoptimized and your [3:22] financial life is the most stressful and [3:24] important thing in your life. So, how [3:26] could you make a billion people make [3:27] better decisions every day or make [3:29] decisions on their behalf? And how could [3:30] you change the economy of the world, [3:32] everyone's life through doing that? GTO [3:34] >> of your financial life, if you will. [3:36] >> GTO, yeah, I don't think it would sell, [3:39] but [3:40] maybe maybe to you, John. [3:41] >> Yeah, [3:42] >> [laughter] [3:42] >> um for the just free idea for the [3:44] magazine article, you know, for the [3:45] opening paragraph on [3:47] on Cleo. Um okay, but that that's [3:50] interesting, right? Cuz you're founded [3:51] in 2016. That's before It's obviously [3:54] before Chat GPT, but that's like before [3:56] LLMs. [3:57] >> 2 years before the Transformer. [3:58] >> Yeah, but even back then if you the [4:00] metal hour we were just talking about [4:02] like they were really focused on it and [4:04] they were pushing out a bunch of good [4:06] you know it was just intent [4:06] classification back in the day. [4:08] >> Then you had 2 years later it was the [4:09] Transformer and then you had BERT. [4:11] >> Yeah. [4:11] >> Everyone then in computer science knew [4:13] this was going to be a thing and then [4:15] everyone else just woke up when 3.5 [4:17] dropped, but you could see it, right? [4:19] >> But so what I I mean I presume today [4:21] you're heavily LLM based. Just how did [4:23] this work? Like how were you providing [4:26] the financial advice? Like you the user [4:28] inputs some data and then what? [4:30] >> It started with regexes. Uh we had [4:32] actually [laughter] we had a lot of [4:33] regexes to start with. But then it was [4:34] just traditional supervised learning, [4:36] right? So you had intent classification [4:37] and then seek classification. It's a bit [4:39] nerdy, but intent is you know what you [4:42] say in a sentence. So it was a big big [4:45] multi-class supervised learning and then [4:47] you had when we still do had the [4:49] recommend the systems and reinforcement [4:51] learning to push you information at the [4:52] right time. So you've gone from a [4:54] supervised learning world to a you know [4:56] an LLM based world and that completely [4:59] changed our architecture and our product [5:00] actually. [5:01] >> Yep. [5:02] >> Um [5:03] how [5:06] the conceit of Cleo is that you should [5:08] have personalized financial advice. [5:11] How personalized does financial advice [5:13] actually need to be? Cuz you know [5:14] there's a view that you could say you [5:15] should have the right savings rate and [5:18] then you should invest in an low [5:21] fee index fund [5:23] um [5:24] and uh you know you shouldn't buy kind [5:26] of financial products. It's like you [5:27] know you shouldn't use um foods that are [5:29] advertised on TV like the good foods [5:30] don't you know don't have an ad budget [5:32] and similarly you know you shouldn't buy [5:34] financial products that are aggressively [5:35] sold. You should [5:36] uh you should buy a low fee index fund [5:38] and then you shouldn't have a home [5:39] country bias. You know you shouldn't buy [5:41] the FTSE 100. You should buy the MSCI [5:42] World Index. Like is that not enough [5:45] financial advice for everyone? [5:46] >> Uh [5:46] showing how out of touch you are, John. [5:48] >> With the with the the real consumer and [5:50] the the normal humans on the street. And [5:51] like, your life is your money. You spend [5:53] like it's the most stressful thing for [5:55] the majority of people in their life. [5:57] You spend so much boring admin time [5:59] thinking about how you make financial [6:00] decisions. From, you know, where are you [6:02] going to live? Where's that mortgage? [6:03] Where's that rent going to be? How big [6:04] should that be? Every single thing you [6:06] do every day is a financial decision in [6:07] reality. You make all these transactions [6:09] every single day. Or someone's Stripe. [6:11] Uh, but imagine if you had a world where [6:13] everyone was making rational, sensible, [6:16] you know, more optimized decisions. If [6:18] you think about how much money goes [6:19] towards consumer credit or paying on [6:20] interest when it doesn't need to. You [6:22] could really change the not just um [6:26] Yeah, [6:27] uh, not just like the top 1% of people [6:29] that get a financial advice. It's the [6:30] 99% that actually need the financial [6:32] advice. The people that are living, you [6:34] know, 1-month, 2-month, 3-month savings. [6:36] If they can have a better life and make [6:37] better decisions, and you can do it for [6:39] them. You shouldn't have to think about [6:40] finance, right? It should just happen on [6:42] your behalf. You need credit, you should [6:44] just ask. It should go get it for you. [6:46] If you have wealth, then it should just [6:47] go get the best return. It should be [6:49] tax-free. [6:50] There should be a world where you don't [6:51] manage your money by looking at a [6:53] balance and a list of transactions. And [6:55] that's coming very soon, right? [6:58] >> Where are you generally trying to steer [7:00] people? Like I can imagine there's kind [7:02] of avoiding bad financial products, like [7:04] the classic one of, you know, borrowing [7:06] on a credit card. [7:07] Uh, there's [7:09] uh, the uh, maybe price comparison where [7:11] if you're getting a mortgage, you can [7:12] get somewhat better rates here versus [7:14] there. There's tax efficiencies. But [7:16] just like [7:17] >> It's actually your life. If you think [7:18] about most of the consumers in the US, [7:20] it's 60% of Americans have less than [7:21] $400 of savings. It's how they spend [7:23] their money every single day. How much [7:25] do they earn? What's their income [7:26] volatility? How much are they spending [7:28] on rent, mortgages, subscriptions? Like, [7:31] what are they actually doing with their [7:32] financial life? And you can really help [7:35] someone just by laying it out to them, [7:37] making better decisions for them. And [7:38] then on the products, they are just [7:40] wildly unoptimized, especially at the [7:42] lower end. They're paying incredibly [7:43] high interest rates. They are not [7:44] getting any returns on their savings. [7:46] Most people don't have investments when [7:48] they just They sit it in like a cash ISA [7:50] in the UK. It's just like crazy the [7:53] amount of GDP growth and the amount of [7:55] financial wealth you could be creating [7:56] people that's like left on the table. [7:58] So, I just think there's trillions and [8:00] trillions of good that you can do for [8:02] society and humans, but also just the [8:03] economy as well. If you had this thing [8:05] moving effectively through the system, [8:07] you can dramatically increase GDP [8:09] growth. [8:10] >> How do you advise people on spending? I [8:12] mean, like there's the meme online of [8:13] the millennials with their avocado [8:14] toast. Is it like you're trying to [8:17] reduce the amount of avocado toast [8:18] consumed? Or like how do you actually [8:20] steer people usefully in an actionable [8:22] way on spending? [8:23] >> Do you know what Do you know what's [8:23] funny back in the day? Cuz it wasn't as [8:25] intelligent. We had to really rely on [8:26] humor. So, we had 30 female comedians on [8:29] staff. Cleo was a you know, it's a [8:30] female agent. So, we had these female [8:32] comedians and they wrote every single [8:35] line of copy both what we pushed and [8:37] what we responded with. And we found [8:39] that humor and money actually go [8:41] together weirdly well, which no [8:42] financial institution [8:44] spoke once to your lawyers and your your [8:46] executives. Bankers are generally not [8:48] known for that. Not known for their [8:49] humor, yeah. So, that was a real unlock [8:51] for us in terms of engagement and like [8:53] there's there's a roast mode in Cleo [8:54] that makes fun of you for your spending. [8:57] I don't know what it would make fun of [8:59] you. Maybe the cheeky pints, John. [9:01] But [9:02] you know, roast you for for what you've [9:03] done over the weekend at the bar, [9:04] whatever it is. And people really like [9:07] that as a way to engage them in simple [9:09] actionable financial advice. And [9:11] increasingly it's just taking action on [9:12] their behalf. It's it's moving accounts. [9:14] It's moving money. It's getting them the [9:15] best product, whatever that is at that [9:18] time for that user. [9:20] >> Um [9:21] do you have a way of measuring how Cleo [9:24] is steering people's financial health? [9:27] >> Yeah. Yeah, we have measure I'm actually [9:29] really surprised that you Well, you [9:30] probably do internally, but there's We [9:33] have constructed a view of financial [9:34] health for every individual at Cleo. [9:37] It's a score 0 to 100 on all these [9:39] variables all these features. So we know [9:41] whether you're leveling up your [9:42] financial health or you're you're [9:43] degrading it based on actions that you [9:45] do. So actually the objective function [9:47] of the recommend system it used to just [9:49] be engagement. So it was like can we [9:51] make you like click this button and it [9:52] was fun. But now it is financial health [9:54] and engagement. So we're optimizing for [9:57] did this bit of advice did this action [9:59] lead to in 2 months 3 months time their [10:02] financial health getting better? And [10:03] them spending less than they earn which [10:04] is like the core primitive uh [10:06] >> Them spending less? [10:07] >> Spending less than they earn. [10:08] >> Oh yeah. [10:09] >> It's like the core primitive of every [10:10] human in that kind of uh low to middle [10:13] income, right? [10:13] >> Yes. Yes. Who do you compete with? Like [10:16] is this substituting for existing [10:18] products? Is this just kind of a new [10:20] thing this [10:21] >> Yeah, well think about how people make [10:23] financial decisions today like [10:25] it's a it's a vibe based optimization [10:27] game for most people. It's um maybe it's [10:29] a packet and a napkin but the vast [10:32] majority of consumers don't budget. They [10:34] don't think about their money in a very [10:35] rational way or they do it in a very [10:38] you know late way. There's only a really [10:39] specific number of nerds probably [10:41] everyone in this room actually um that [10:43] will have an Excel and opt to Sorry I [10:45] shouldn't make fun of all the audience [10:47] uh but have this you know set of uh [10:50] advice that they'll construct [10:51] themselves. So we're really in the [10:54] advice game for people that don't have [10:55] advice and I think there's going to be [10:57] loads of avenues like this. Imagine [10:59] health, right? We all make terrible [11:01] decisions for our health all the time. [11:02] We've all got a whoop now and we're all [11:03] getting into it. But there's going to be [11:05] all these big verticals in your life [11:06] whether it's education or health or [11:08] finance where your decisions are wildly [11:10] unoptimized today and they are going to [11:11] be [11:13] they're just going to be so much more [11:14] thought through. Imagine how you deploy [11:16] capital and people at Stripe. There's [11:17] probably an army of nerds that are [11:19] thinking about how you deploy that [11:21] versus your strategy and where your [11:22] goals are going. Imagine if you do that [11:24] for every human for every decision they [11:26] ever have to make. [11:27] >> So you think you'll go broad and beyond [11:29] financial advice into [11:31] >> There's a there's a lot to do in [11:32] financial advice. I would big time. Um [11:34] so we're we're going to nail that first. [11:36] And I think, you know, you're going to [11:37] have these four or five agents that go [11:39] really deeply vertical in these big [11:41] specific areas of your life. And I don't [11:43] think they're going to be the frontier [11:44] model companies, actually. I think [11:45] they're going to be companies that focus [11:47] all on health. They go all the way to [11:48] the MRI and the blood test. And they [11:49] have a whole stack of advice. So I'm [11:51] bullish on that, and I'm less bullish on [11:53] like [11:54] OpenAI being the interface for everyone [11:56] for everything. Um [11:58] >> But do you see that competition dynamic [12:01] with users where anyone who's building [12:03] AI tools, there's always the um [12:06] anyone who's building kind of specific [12:07] vertical AI tools, there's always the [12:09] dynamic where people might just, you [12:10] know, take a bunch of financial [12:12] statements or connect their bank account [12:14] to Claude and start asking it questions. [12:17] Do you practically see that behavior [12:19] where an alternative to using Cleo is [12:22] wiring up the relevant data to an LLM? [12:26] >> Well, you you can connect your plan [12:27] account now to ChatGPT and ask [12:29] questions, right? But I would bet a huge [12:32] amount of money they get very little [12:34] engagement, and it becomes a kind of [12:36] side project that goes nowhere in [12:37] reality for for that business. Because [12:40] to engage people with financial advice, [12:41] it's got to be proactive, and it's got [12:43] to be agentic. So people don't wake up [12:45] and go, "Ooh, how much did I spend on [12:47] Yes. What is it just a boring thing to [12:49] think about? So it has to be this agent [12:51] always on, always pushing information. [12:53] You think about all the frontier model [12:54] companies, it's a very reactive thing. [12:56] It's replacing Google search for most [12:57] people today. You have something in your [12:59] head that you want to research, you want [13:00] to go into. But there's so much space in [13:03] the product cycle for it to become a [13:05] proactive, intelligent, always-on, [13:08] agentic thing. And that's going to [13:10] require a lot of effort and a lot of [13:11] training data and a lot of understanding [13:13] of what you need to say to a user at [13:14] what time and that's going to work. [13:17] >> It feels to me that we're early in [13:19] personalization. When you're saying [13:21] it'll be very personalized, it's just I [13:23] mean, if you look at how kind of open [13:25] claw works underneath the hood, It's [13:27] this static model and then a a bunch of [13:30] flat files that it's kind of writing [13:31] notes to itself. It's like the the movie [13:34] Memento um where he's like writing on [13:35] his hand um uh and that's the only [13:38] memory he has. That's basically the AI [13:40] models today. And I'm curious just for [13:42] your experience trying to build [13:44] personalized products where [13:46] the current [13:48] AI technology is just kind of [13:50] fundamentally a bit resistant to [13:52] personalization. [13:53] >> I've never thought that [13:55] there's definitely a meta point here of [13:56] like we just have text, right? And like [13:58] how does the human store memory and like [13:59] when are we going to get to a different [14:00] form factor for that? But in reality uh [14:04] and what we've been doing, you can build [14:06] a knowledge base of every single human [14:07] like their financial aspirations, their [14:09] financial health, their financial goals, [14:11] every conversation they've ever had. So [14:12] you remember Copilot did that thing a [14:14] few months ago where it's like, "Oh, I [14:15] do this thing recursively with LLMs." [14:17] We've been doing that for years. Like [14:19] that is such a powerful way to build a [14:21] structured knowledge. I was also giving [14:22] you a bit of for Stripe not being [14:24] intelligent enough. And like they should [14:26] be like knowledge base at Stripe. They [14:28] should be every single card connected to [14:30] every single user, who this user is, how [14:32] much income they have, where they spend. [14:34] So we've built that over time ever for [14:36] every user. And then we have a [14:38] recommender system that is thinking from [14:40] that knowledge base. There's all of [14:42] these opportunities for us to say things [14:44] to users. What is the thing that's going [14:46] to move the objective function the most? [14:47] That's going to move that financial [14:48] health, deliver that information, take [14:50] that action on the back end. So I think [14:53] I think text works as long as it's [14:54] stored in a thoughtful uh [14:57] clear schema and database. [14:58] >> Mhm. [14:59] Um [15:01] one thing we were talking about [15:02] backstage is [15:03] just the concept of personality in AI is [15:06] interesting. Like I think at the [15:08] beginning of the field when people were [15:10] first introducing LLMs, people might not [15:12] have um [15:14] predicted that personality would be such [15:15] a big thing. But you know, we're talking [15:16] about the new um Anthropic model Claude [15:18] and just like what its personality is [15:20] like or famously, you know, GPT-4 0 had [15:23] this very controversial personality [15:25] where, you know, some people um found it [15:28] too sycophantic. For some people, I [15:30] guess they really liked how sycophantic [15:31] it was and they were annoyed when uh [15:33] they tried to take it away from them. [15:35] And I I get the impression that for you [15:37] guys, the personality of uh Cleo and how [15:41] it interacts with you is a big part of [15:44] the differentiation. And so, what goes [15:46] into getting the personality right? And [15:48] for example, like do you need to tune it [15:50] for different users? Do you try to suss [15:51] out what personality people want? Is it [15:53] just what it is? How do you build that? [15:55] Have you moved beyond the army of [15:56] comedians or is that still how you [15:58] engineer the personality? [16:00] >> Um I think this is one of the biggest [16:02] things at Frontier Labs. Maybe they [16:03] don't need to. Maybe they should be [16:05] boring and it should be a generic, you [16:07] know, not even personalized thing and [16:09] they should just be there to be of [16:09] service and intelligence. But I think [16:11] for any brand or any company doing this, [16:13] it moves the metrics. The personality [16:16] and the personalization to the user [16:18] dramatically [16:19] changes engagement. It dramatically [16:21] changes what a user does. So, being [16:23] funny, being concise, writing well, all [16:26] these things move our MAUs, it moves our [16:28] revenue. So, we've known from like, I [16:30] don't know, the first kind of like 3-6 [16:32] months of building this business that [16:34] the brand and the personality and the [16:35] tone of voice mattered a lot. So, it's [16:38] um it's been a it was kind of linear [16:40] supervised learning, one tone of voice. [16:43] And then since LLMs, it is now a much [16:45] more personalized [16:46] uh tone of voice based on your income, [16:48] your financial health, where you live. [16:50] You in Boston, you in San Francisco, uh [16:53] don't know how it talks. I want to make [16:54] fun of John, but I shouldn't make fun of [16:55] John in front of 1,500 people. But it [16:58] it will talk to you um however it thinks [17:01] uh you want to be spoken to and you can [17:02] craft that as well. [17:04] >> But but is your general learning that, [17:06] you know, finance for a lot of people is [17:07] a high-stakes topic, it's maybe they're [17:09] nervous about it, it's a bit of a [17:10] schlep, and so you're trying to make it [17:12] more approachable is kind of the fun of [17:13] fundamental personality stuff? [17:15] >> I think these kind of great companies [17:16] take the the inverse of take Stripe. [17:18] Like it's it's completely archaic and [17:20] complicated and I've got to go through [17:22] enterprise sales to do anything in [17:23] payments. [17:24] Let's do the opposite. Let's make it [17:25] really simple, dev friendly. It's going [17:27] to be a one-click thing. It's the exact [17:29] same You remember Monzo? Like it was [17:30] really bad UX to do anything with a bank [17:33] and they just made it simpler, more [17:35] transparent, less opaque. And we've [17:37] gone, "Okay, it's really complicated [17:39] financial advice. We're going to make it [17:40] really simple. It's really boring. We're [17:42] going to make it really fun." And we've [17:45] just taken the inverse position, which [17:47] is such an obvious thing to say, but I [17:48] think like [17:50] how many founders do you see actually [17:51] take that? [17:53] So, I think it's worked for us. [17:54] >> It's product contrarianism or product [17:56] differentiation. [17:57] >> Yeah, it's like the in the you always [17:59] want it in the most extreme form [18:01] generally. Like what is the biggest pain [18:02] point of your incumbent? And like what [18:04] am I going to do? I'm just going to [18:06] go incredibly hard at that biggest pain [18:08] point and make it the most delightful [18:09] experience. So, it's a yeah, it's a hack [18:12] for building a first product, right? [18:13] >> Yes. [18:14] Um [18:16] how is the landscape of financial [18:18] products that you advise people on [18:20] changing? Like borrowing is some [18:23] presumably changing somewhat [18:25] at an industry level due to the [18:26] emergence of BNPLs and things like this. [18:28] I don't know is kind of the mortgage [18:30] industry changing and how people are [18:31] getting those here. Just yeah, I'm [18:33] curious how you [18:35] >> There's only so many primitives with [18:36] money as you definitely know, right? [18:38] Like you can borrow, you can spend. Like [18:40] there's there's just only a finite set [18:41] of things you can do. So, all of the [18:43] innovation is [18:46] It's it's it's useful, but it's slow and [18:48] it's guided by regulators. I think, you [18:51] know, buy now pay later is a great [18:52] example of credit cards were this really [18:54] bad thing for all these years. So, we're [18:56] doing something better and something [18:57] different, but it's not that interesting [18:59] in my opinion generally. Hopefully, like [19:02] the biggest thing will be stable coins [19:03] and like how do we transact and how do [19:05] we move away from those rails, but [19:08] I don't find building the We build a lot [19:10] of financial products. I shouldn't say [19:11] this cuz all my team's work on them, but [19:13] I find them [19:15] constrained and slightly boring cuz they [19:17] the limited number of primitives, but I [19:19] think the intelligence and the advice [19:21] If I was building Stripe, I would be [19:23] like the intelligence, what I can give [19:25] consumers from that would be the most [19:26] important thing, but I guess that's also [19:28] my wonk. [19:30] >> Well, it's not just a limited number of [19:31] primitives, right? It's also all these [19:32] products are heavily regulated, and so [19:33] they end up in kind of certain stra- [19:35] streams. But, for example, [19:37] do you see different like should people [19:39] be consuming a lot of BNPL or a little? [19:41] Like, what's um [19:43] >> Well, they shouldn't be maximizing that. [19:45] What happens generally for consumers is [19:47] they'll take a credit card out and they [19:48] will max the limit, and then they will [19:50] stay at a max limit paying the minimums [19:52] for 10 years. It's probably the most [19:53] inefficient form of credit the vast [19:55] majority of society has. So, BNPL is a [19:57] better way to pay over actual time [20:00] period and make it more like a term [20:01] product. [20:02] >> For a fixed amount of borrowing though, [20:03] if you just use it to borrow more, [20:04] presumably that's bad. [20:05] >> This has been my whole raison d'être, [20:07] you know, for a for a long period of [20:08] time. Like, credit is such [20:11] an important part of the economy in [20:12] people's lives, but it's used so poorly [20:15] by so many people. And this is why I [20:16] think the intelligence is the [20:18] is the answer, right? Help make people [20:20] make better decisions about how much [20:22] credit they consume, how much they [20:23] spend. If you just provide financial [20:25] advice or provide financial products [20:27] without the advice, [20:29] I don't want to bash Robinhood, [20:31] but you you end up just selling products [20:33] and trying to get the most margin, and [20:35] it leads you down this way of like, [20:37] "Ooh, let's do day trading. Ooh, let's [20:39] do None of these things are bad, right? [20:41] And they can be used in the right way. I [20:42] do think you need some sort of [20:43] >> You can't be purely revenue maximizing [20:45] is your view. [20:46] >> 100%. I think it just leads to bad [20:48] outcomes for consumers and bad outcomes [20:50] for the business as a whole. So, I think [20:51] that [20:52] I think the intelligence and the [20:54] decision-making and helping users do it [20:56] for them is just so critical for [20:58] everything. [20:59] >> If you're in charge of regulating [21:00] financial products in the UK, what would [21:02] you change? [21:03] >> Ah, where do I start with the UK and its [21:06] regulators? [21:08] Like innovation is good. Like [21:10] creative destruction is good. It's how [21:12] you move every industry in the world [21:14] along. It's what we've seen for the last [21:15] 100 years. Creative destruction is what [21:17] drives better outcomes for consumers. [21:19] It's what drives better outcomes for [21:20] businesses. So you need to enable [21:22] creative destruction. You need to have a [21:24] level of risk-taking to allow [21:27] upstarts to take on incumbents. And all [21:29] too often in the UK, it's like [21:31] "Please come to our sandbox and let's [21:34] work together for 2 years when a venture [21:37] funding cycle is actually 18 months." [21:39] And it actually them trying to be [21:41] helpful normally constrains a company. [21:43] I've seen so many of these little angel [21:44] investments I've done that have tried to [21:46] engage with the FCA, but it's just taken [21:48] so long to to get there that they've run [21:51] out of money and not been able to [21:52] approve product market fit. So if you're [21:54] a founder doing it, I actually do think [21:56] you need to take some risk early on. I'm [21:58] sure Stripe took some risk with his [21:59] early clients and how he built the [22:00] business. Don't want to get in trouble, [22:02] John. But [22:04] taking some risk in this sector at the [22:06] start is a is a way to build [22:10] something that the incumbents aren't [22:11] going to do as well. [22:12] >> And you think there's kind of not enough [22:13] risk-seeking behavior in the UK or it's [22:15] not enabled in the right way? [22:16] >> Well, it's Yeah, go back to the [22:18] Federalist paper the Federalist papers, [22:20] right? You had checks and balances. You [22:21] have state and federal. [22:24] It's probably like the [22:25] Read those Federalist. They're like it's [22:27] a great insight into how you should run [22:29] a company and it's run a country. And we [22:31] need more of that in the UK. I see the [22:33] Prime Minister has so much power. Number [22:35] 10 has so much power. The regulators are [22:37] complete have complete control of yes or [22:40] no across 60, 70 million consumers. [22:44] Doesn't happen in the US. We have the We [22:45] have the courts. We have the regulators. [22:47] We have state There's all these checks [22:49] and balances which make America and [22:50] capitalism work, which is why I love [22:53] building in America and why all my users [22:54] are in America because it's actually a [22:56] better place to do business and it leads [22:58] to more rational outcomes. You get [23:00] obviously dramatic polarization and [23:02] fights and all these things, but the [23:03] checks and balances of the US [23:05] institutions [23:07] is why America is winning versus Europe. [23:10] >> So, [23:11] maybe moving beyond financial, um, [23:14] regulation and just kind of more broadly [23:16] to the business climate, [23:17] you you think just more embrace of [23:19] creative destruction, basically. Would [23:21] that be the precy? [23:22] >> If I Yeah, if I was going to really [23:24] synthesize it, it's like we've got to go [23:26] from a culture of [23:28] trying to conserve our legacy and our [23:31] past and the greatness of Britain and [23:33] Ireland and we to a place where we go, [23:37] "Hey, actually, we're all stagnating. [23:38] We're all declining. We actually don't [23:40] have much to protect at the moment. We [23:42] actually need to take some risk. We need [23:44] to go from a very conservative risk-off [23:46] mindset to a we're going to build and [23:49] we're going to build great stuff and [23:50] we're going to take some risk and we're [23:51] going to deploy capital, uh, all these [23:53] pension funds, all these ISAs, all [23:55] all these institutions that could be [23:57] doing productive things in society, you [23:59] need to uh, move from a a mode of, yeah, [24:02] conservatism to to risk-on creative [24:04] destruction. [24:05] >> Does that [24:06] uh, filter down into how you think about [24:08] how people should invest both at the [24:10] individual level and at the [24:13] institutional level? Like, if you could [24:15] tune how people are investing, what [24:17] encouragement would you give? [24:19] >> The I think the individual is, you know, [24:21] a largely solved problem even if people [24:22] aren't doing it. Um, the the solved [24:24] problem is, you know, the world index [24:26] fund, right? And that's invest in the [24:27] S&P 500. Maybe it should get better. [24:31] Getting a little bit tech heavy, um, [24:32] when Stripe goes public, uh, at some [24:34] stage, but [24:36] that that's a almost solved problem if [24:37] you had people making rational decisions [24:39] there and like Cleo could dramatically [24:40] help that. I think the the set there the [24:43] new fund that they've set up, uh, with [24:46] Gerstner and the the baby accounts or [24:48] whatever, the Trump account, it's always [24:49] the Trump accounts, isn't it? [24:51] Um, that is a that's a good thing. Get [24:53] people investing from your age. But on [24:54] an institutional level, there's a [24:56] dramatic difference between the US and [24:58] the UK and you know, all of Europe. You [25:00] look at our pension capital, which is [25:01] should be the big pool of capital going [25:03] into creative destruction, [25:05] tiny, tiny, tiny, tiny fraction in the [25:07] UK goes towards venture capital. I've [25:09] got British VCs and the tier one [25:12] European VCs, [25:14] it's like, you know, it's the California [25:16] pension funds are actually the LPs into [25:17] those VCs. So, I'm not returning to the [25:20] British state any GDP growth. It's all [25:23] going back to California. And if you [25:25] don't want a world which is completely [25:27] bipolar where you can either build tech [25:28] in China or you can build tech in the [25:30] US, [25:31] it's really bad for Europe if we don't [25:33] have this creative destruction and [25:34] ability to deploy institutional capital [25:36] in a different way. [25:37] >> It's an interesting point that [25:39] yeah, there's a lot of US, I mean, it's [25:41] not universal, but a lot of US, Canadian [25:45] >> Yep. [25:46] >> pension systems are a little bit in [25:48] Australia, yeah, are invested in venture [25:50] in a big way and much less so in the UK [25:52] and Europe broadly. [25:53] >> And it's worked. If you look at Canadian [25:55] pension funds, if you look at [25:56] Australian, I would love to be an [25:58] Australian pensioner. Like it's [25:59] >> Yeah, there was a news article about [26:01] famously one of these pension funds that [26:03] bought a lot of SpaceX a few years back [26:05] and as you can imagine, they're feeling [26:07] quite good this week. [26:09] >> Um [26:10] you [26:11] uh you run a very scaled consumer [26:15] uh [26:16] uh AI products. I'm curious to talk [26:18] about that. Well, actually first off, [26:19] how do you make money? [26:20] >> Uh we make money through subscriptions [26:22] and and financial products. So, we've [26:23] got we've got cards, we've got lending [26:25] products, buy now pay later and wage [26:26] access. We've got wealth products, we've [26:28] got stocks and shares. Um so, the [26:30] financial products make money, the [26:31] subscription makes money, value [26:32] accruing, and it's a freemium model. [26:35] >> What advice do you have for people who [26:37] are building AI products that they sell [26:38] to consumers? Like what have you learned [26:40] about willingness to pay, what works [26:42] well, just acquisition, like the the [26:45] this new space that's only existed over [26:48] the past few years? [26:49] >> I don't know if I did it the right way. [26:50] Like I spent the first first four years [26:52] in business, maybe this is Zarp, just [26:54] doing MAUs. Like it was built a great [26:56] product, just getting MAUs. [26:57] >> Not focused on revenue, just focused [26:58] >> No revenue. And everyone was like [27:00] financial health, no revenue, this is [27:02] great. Um that worked up until I got to [27:04] the growth funds. They were like uh [27:06] where's your revenue? [27:07] Um so I don't know if I'd advise that, [27:09] but I kind of would, you know. Like I [27:10] would probably go back and do the same [27:12] thing of just try and build a product [27:14] that people are using and like. And then [27:17] integrate the the monetization um [27:19] afterwards. The inverse of this though [27:21] is just make sure there's one core [27:23] primitive like Stripe where you are [27:25] taking, you know, a margin on something [27:26] or there's one thing tokens, whatever it [27:28] is, and build that from day one. [27:31] I wouldn't [27:33] I wouldn't do anything in the middle [27:34] which is like [27:36] be like half-hearted on the monetization [27:38] strategy. Like it's either baked into [27:40] the product 100% from day one or you [27:42] just completely forget it and get [27:44] consumers. [27:45] Um on the consumer side, I think it [27:46] works actually probably to just get [27:48] consumers first then monetize, but I [27:50] don't know, what do you think? [27:52] >> Um Well, I I'm curious [27:55] like on the consumer side of things, it [27:57] seems like many businesses decide that [28:00] they're going to be one or the other [28:02] where they're going to sell on a usage [28:04] basis like uh um uh you know, normal [28:07] grocery store just like charge a markup [28:09] on what you um what you buy or um in the [28:12] US, you know, Costco has this uh famous [28:14] model where they basically make no money [28:16] on the actual things you buy there and [28:18] then they just sell a um [28:20] a $100 [28:21] uh a year subscription or it's probably [28:23] gone up. Um a subscription membership [28:25] subscription and uh that is where they [28:27] make all the money that Costco makes and [28:30] they've decided that they're not going [28:30] to bother making money on anything you [28:31] buy in the store, they're just kind of a [28:33] subscription business. And do you guys [28:35] just view it as both or you are you in [28:37] the business of selling subscriptions or [28:38] are you in the business of selling [28:39] financial products? [28:41] Or is this both? [28:42] >> We are in both, but I think, you know, [28:43] the interesting thing there is [28:44] subscription has been the dominant form [28:46] factor to monetize because of payment [28:48] rails as well, right? Because it's [28:50] actually hard to do micro transactions [28:52] or lots of transactions and transactions [28:53] fell. So, having a recurring monthly [28:55] thing and making this a kind of [28:57] phenomenon for for the masses [29:00] works really well um [29:02] for software businesses. But, I can [29:04] definitely see a world in the future [29:05] where [29:07] like [29:08] you know, you have stable coins, there's [29:09] no friction, the payment rails actually [29:11] work, there's no cost. [29:12] >> Yes. [29:13] >> Then, you move way more to the Chinese [29:14] model, right? [29:15] >> For for self-improvement products, I [29:17] feel like there's also sometimes a [29:20] kind of [29:21] using the sunk cost dynamic in consumer [29:23] psychology, where you know, people say, [29:25] "I'm going to buy a gym membership and I [29:27] am going to get in shape." And now that [29:28] I've spent this money on the gym [29:29] membership, I have to use it cuz I'll [29:30] feel bad if I, you know, spent the money [29:32] and didn't. Is there any of that dynamic [29:34] with Cleo where people are investing in [29:35] their own financial health? [29:37] >> I think every there's like all this kind [29:39] of behavioral theory in all of [29:40] subscriptions. Like, we just added a [29:42] third tier because [29:45] that you know, that just worked. And we [29:47] got this middle tier and just because we [29:50] have three, people will pick the middle [29:52] like [29:53] it just there's all this kind of weird [29:55] behavioral stuff with payments and [29:57] subscriptions that you have to optimize [29:58] and learn about. Um so, yeah, humans are [30:01] kind of irrational actors and I think [30:02] you kind of got to learn where they're [30:04] irrational um [30:06] to help them [30:07] but also to build a business. [30:08] >> Yeah, you guys are kind of like dueling [30:10] or something, right? You're like [30:11] weaponizing people's irrationality for [30:13] for good rather than for evil. [30:15] >> [laughter] [30:16] >> Let's not write that in the press, but [30:19] >> Thanks, John. Yeah. [30:20] Harnessing. [30:21] >> Yeah, there you go. [30:22] >> Yeah, there we go. Better. We're we're [30:23] just workshopping it. Um it kind of [30:25] speaking of people's irrationality, you [30:27] know, people in poker talk about this [30:28] idea of going on tilt, where, you know, [30:30] people um you know, may mostly have a [30:34] very solid betting strategy and then [30:36] they just they break at a certain point [30:39] and they start making irrational [30:41] decisions, uh you know, maybe after a [30:43] losing streak or something. Does that [30:45] also happen in [30:47] behavior you see in the app, and then [30:48] you're trying to kind of get people back [30:50] on the straight and narrow? [30:51] >> It's It's It's It's human psychology. [30:53] It's the biggest thing, right? Like um [30:56] money is such a personal, such an [30:58] emotive, such a, you know, psychological [31:00] thing. People will lose a job. People [31:02] will have a breakup. Like we have all [31:04] these user research interviews where you [31:06] see someone, you know, go completely on [31:08] tilt, and, you know, [31:10] do things that are, you know, slightly [31:11] irrational. That's a Yeah, we're [31:12] definitely trying to help um in those [31:14] circumstances. [31:16] But it's um [31:17] yeah, that that humans are humans in a [31:19] way as well. But there's all these [31:20] interesting mechanisms and features that [31:21] we build around that. [31:23] >> You started Cleo [31:25] before the current wave of AI, but [31:28] you're still a kind of small and [31:29] fast-growing company. How AI-native is [31:31] Cleo in how it actually works? [31:35] Uh what's ways of working do you guys [31:37] have that we should be borrowing from [31:39] you guys, or people in the audience [31:40] should be borrowing? [31:41] >> Well, my my day-to-day is drama Like [31:43] since probably January, February, I've [31:46] I spend my time completely differently. [31:48] >> Oh, please do tell. Give us the before [31:50] after. [31:50] >> Yeah, well, the the before was 500 [31:52] people when you'd have all these [31:53] division leaders, and you'd talk to [31:54] division leaders, and you'd go farm for [31:56] information, give them your opinions, [31:57] you have your pre-reads. And the way of [31:59] getting information in a company was to [32:01] talk to all of these leaders, and then [32:02] give them your feedback, and do that. [32:04] You don't need to do that in [32:05] organization. I think one of the [32:06] interesting things that you probably [32:07] won't want to talk about is like how you [32:09] can understand what's happening in your [32:10] organization day-to-day. You know their [32:13] Slack messages. You know their Notion [32:14] docs. You know every single PR, you know [32:17] what it is built. So, I understand, you [32:20] know, if software is the output of all [32:21] these companies, I know what we're [32:23] shipping in a really visceral, clear [32:26] way, which I didn't know before LLMs. [32:28] >> And just very practically, how are you [32:29] synthesizing all this information across [32:31] the company? Like are you using a [32:32] third-party tool? Do you guys build [32:33] internal tools? Like how are you [32:35] scouring the Slack channels and kind of [32:37] ingesting on all the docs? [32:38] >> That That [32:40] That workflow that Karpathy described is [32:42] what I use for so much stuff. I just [32:43] have LLMs that go over you say every PR [32:47] and it summarizes what that PR does and [32:49] it grades it and it tells me what's [32:50] going on that PR and it just builds a [32:52] knowledge base by going over all these [32:54] things. Goes over all of the Slack [32:56] messages in all the channels, builds up [32:57] a knowledge base of what is happening. [32:58] So you take lots of disparate [33:00] information across your entire org. [33:01] >> Yes. [33:02] >> You build it up into a higher level. [33:03] >> And it's essentially a sort of [33:05] >> captain's log. Like it's going around [33:07] the organization with its clipboards [33:09] taking notes. [33:09] >> On like this dramatic scale that you [33:11] could never have done. [33:13] >> And then you're curating the log. [33:14] >> Well you get humans also to input into [33:17] that log. Their pre-reads, their their [33:19] dashboards that they're making, [33:20] everything they do all goes into logs [33:22] and becomes more synthesized and [33:23] aggregated over time. Are you doing [33:25] that? [33:26] >> Um I I mean I'm not doing specifically [33:28] that, but I I think one of the [33:30] interesting things that all [33:31] organizations are focused on right now [33:33] is [33:34] like um [33:36] consumer AI tools are phenomenally [33:38] useful and they're trained on the kind [33:40] of corpus of information that's out [33:42] there and they have web search tools and [33:44] everything and companies are trying to [33:46] >> Yeah. [33:47] >> internally open up the data but not too [33:50] open you know what I mean with all their [33:51] permissions and security tools and [33:53] everything and so I feel like that is [33:54] the thing that all organizations are [33:56] reckoning with with right now. But it's [33:59] really as you say I mean does it drive [34:00] your internal leaders up the wall the [34:01] fact that you are now going straight to [34:03] the source for information? [34:05] >> My engineering managers hate me. Like [34:07] honestly like I can't I know I know [34:09] who's shipping. I know like who is the [34:12] highest velocity, highest code quality. [34:14] I know which teams because it's a [34:15] collection of engineers are actually [34:17] doing great work where before I had this [34:19] kind of it was this weird view filtered [34:22] through all these leaders that had all [34:23] their different biases and all these [34:25] different incentive structures. You can [34:27] actually get really rational, clear [34:29] thought from these LLMs and build it up [34:31] with contacts. So, yeah, it's like I [34:34] think the middle layer has [34:36] probably Cleo definitely hates me, but I [34:39] think orgs are just going to be way [34:40] flatter because of this and it's just [34:42] become [34:43] yeah, it's going to become a much more [34:44] efficient organizational flow and [34:45] information flow. [34:46] >> Do you think the average number of [34:47] people that a manager at Cleo manages [34:50] will be different in 2 years time versus [34:54] currently? [34:54] >> Yeah. I've never been a massive believer [34:56] of like [34:57] you have to do a one-to-one with every [34:59] single one of your direct every day. And [35:00] I think [35:01] >> Yeah, but when you say becoming flatter, [35:02] like do you think that will show up in [35:03] the numbers? [35:06] >> I I really hope you're going to have [35:08] less middle managers [35:10] um as and I think orgs are going to [35:11] definitely become flatter. [35:13] The only like the only reason I'm [35:15] caviating myself is like you do need to [35:17] give junior people advice and time and [35:19] grow and develop these people. So, I [35:21] think there's always going to be that [35:22] need for nurturing and maybe [35:25] it's done in a different way, but I [35:27] think those layers and layers of [35:30] people and people and people uh it's [35:31] probably half the Stripe uh team uh full [35:34] inside Cascade in the room, but I think [35:36] that has to to change um and maybe we [35:39] just get to do more stuff, right? [35:40] Instead of like more hierarchical, fewer [35:42] business units, maybe you just have [35:44] way way way more bets, way way more [35:46] business units and it becomes much more [35:49] organic and much flatter. [35:50] >> I think you possibly yeah have that [35:52] because you can have more impact from a [35:54] small number of people. Uh I think the [35:55] thing a lot um people have been [35:57] wondering as well that I think is very [35:58] interesting is when you talk about your [35:59] team and your people, do you end up with [36:01] more of a an apprenticeship model [36:04] because it's different if you have like [36:05] one senior person with eight junior [36:07] people, then it's maybe kind of the [36:09] classic uh org chart view. Whereas if [36:11] you have one senior person with one [36:12] junior person, then you can do much more [36:14] of a traditional kind of knowledge [36:16] transmission uh apprenticeship model. [36:18] So, I think [36:19] >> I think the I think the inverse is like [36:20] one person with like 36 like uh I'm so I [36:23] hire all these grads from like UCLA and [36:25] Cambridge that done their machine [36:26] learning masters. They are so they are [36:29] like fully AI pill. They've been doing [36:31] it from like their A-levels, right? Like [36:33] it's it's crazy how good they are with [36:36] these tools. So, I think you're going to [36:37] want more junior people that have had [36:39] the learning and experience in LLMs from [36:41] these institutions and are really pill. [36:43] And then they are just dramatically, [36:46] they're managing a fleet of agents to do [36:48] things. So, like [36:49] in software, you want more and more [36:51] software, right? But you want it [36:52] directed to the right things. So, you [36:54] need great taste makers that can run [36:57] lots and lots of these parallel agents [36:59] and recursion, just more and more and [37:00] more tokens. So, you want as many [37:03] people, I think, as possible [37:05] at a lower level with good taste to run [37:07] as many agents as possible. I I like the [37:10] idea that I can take 2,000 bets next [37:12] year instead of like the 10 that I took [37:14] last year. [37:15] >> Yes. [37:16] >> That's the world I would love to live [37:17] in. [37:18] >> Very cool. [37:18] >> We'll see. [37:19] >> [laughter] [37:20] >> Last question. What's your most [37:23] contrarian AI take at the moment? [37:26] >> Um [37:28] I'm very AI pill, so it's hard to be [37:30] like absolutely contrarian at the [37:32] moment. [37:33] Uh [37:34] I I I think the SAS pop apocalypse, like [37:37] Frontier Labs dominating everything [37:39] every tweet is like they've just [37:41] destroyed X industry with their little [37:43] wrapper. There's so much work that goes [37:45] into building a great business like [37:47] Stripe or like anyone else. Like there's [37:48] so much that you need to know. [37:50] How much have you learned about payments [37:51] that no one in the world needs [37:53] >> No one should have to learn. [37:55] >> It is awful. Too much. [37:56] >> Way too much, but there's so much of [37:58] that and so much craft that goes into [38:00] building a great consumer, [38:02] business-to-business, or [38:03] business-to-consumer products that [38:05] you're going to have you're still going [38:06] to have great software companies that [38:07] are like completely vertical and really, [38:09] really deep. And this kind of idea that [38:11] someone wrote a prompt and they've just [38:13] got a Frontier model that's going to [38:14] kill everything is wildly overblown and [38:17] investors are just [38:18] the silliest, most irrational people [38:20] I've ever met in my life and I can't [38:21] believe they have all [38:22] >> So we are kind of over extrapolating [38:24] from the [38:25] >> They have no idea what's going on. They [38:26] don't have any technical knowledge at [38:28] all and they are they are in a world of [38:30] they don't understand. So they're in a [38:32] world of we're going to sit on the fence [38:33] and just put money into the frontier [38:34] labs. [38:36] And stripe apparently. [38:38] >> [laughter] [38:38] >> And that's all they're going to do. Um [38:40] But yeah, I think like that rationality [38:42] will play out in a couple of years and [38:43] you'll see valuations, multiples The [38:46] economy is driven by software. It has [38:47] been for the last 30 years. We want more [38:49] software companies. That is where GDP [38:51] growth is going to come from. We want [38:53] more creative destruction. That's what [38:54] AI is going to lead to. Like the [38:56] abundance is not contrarian take, but [38:58] I'm pro abundance. [38:59] >> Yeah. [39:00] That's a good note to end on. You heard [39:02] it here. Lots more creative destruction [39:04] to come and the story is not yet written [39:07] and it's not just, you know, a small [39:08] number of firms that'll be shipping [39:10] everything useful. [39:12] With that, um that brings our [39:13] programming for Tour London at 2:00 to a [39:16] close. So thank you all for coming this [39:18] year. And we are going to roll into a [39:20] reception which we're having out in the [39:22] Expo Hall. And we can't wait to see you [39:24] guys next year. [39:26] Thanks, Barney. [39:27] >> Thanks, Luke. [39:28] It's fun. [39:28] >> Thanks. Good luck.