ElevenLabs and Legora Scale to $750M ARR Disrupted by Voice, Legal AI

ElevenLabs与Legora ARR达7.5亿美元，语音与法律AI掀起颠覆 AI原生平台正通过垂直化的窄领域模型实现专业工作流自动化，迅速淘汰传统的产品管理模式和律所按小时计费模式。 本期节目探讨了AI给语音和法律行业带来的巨大颠覆。ElevenLabs联合创始人兼CEO Mati Staniszewski详细介绍了公司ARR飞速增至6亿美元的历程、其不设立产品经理的独特产品主导型文化，以及声音克隆、安全协议与名人合作的不断演变。随后，来自Legora的Max Junestrand分享了该平台如何通过自动化复杂法律工作流实现爆发式增长，挑战大型律所传统的按小时计费模式，并借助专业的窄领域AI模型来应对高安全性企业环境的需求。 ElevenLabs 实现了极其迅猛的营收增长，在推出拟真人语音模型后约 35 个月内，其 ARR 从 1 亿美元跃升至 6 亿美元；公司目前拥有 600 名员工，并不设任何产品经理（PM），而是将工程师直接嵌入非技术部门以加速迭代并维护团队文化。 (0:00) AI 语音交互正引发用户行为与心理的深层改变：企业开始采用呼入型 AI SDR，而用户也展现出更偏好直接打断、快速交互的习惯；在处理债务追收等敏感私人事务时，相较于人类客服，客户与 AI 语音代理交流时展现出更少的情感阻抗与更高的坦诚度。 (8:04) 为应对声音克隆与深度伪造风险，ElevenLabs 确立了包含溯源水印、多层审核与公共验证工具的三重安全体系，并已通过创作者市场向声优支付超过 2200 万美元，同时积极拓展与 Matthew McConaughey 等明星合作及语音医疗修复等应用。 (16:03) 法律科技公司 Legora 连续 7 个季度保持 50% 的环比复合增长，营收从 100 万美元快速攀升至 1.5 亿美元；其通过部署类似 Palantir 模式的“法律工程师”重构律所工作流，直接颠覆按小时计费模式，并专注于构建针对特定工作流的狭义模型。 (31:42) 面对全球高达 1 万亿美元但软件占比仅 4% 的法律服务市场，Legora 强调合规与安全性是核心壁垒；即便处理国家机密与军工企业合同，公司也坚决拒绝本地化或私有 VPC 部署，以确保产品的极速迭代与执行能力。 (50:04)

Executives from ElevenLabs and Legora argue that specialized AI is rapidly transforming voice communication and corporate legal services by combining lean, engineering-driven operations with vertical workflow automation. While frontier foundation labs provide core language models, both companies contend that lasting enterprise defensibility relies on domain-specific data, specialized orchestration, and strict security compliance rather than generalized intelligence. Hypergrowth and Engineering-Led Organizational Models (0:00) ElevenLabs co-founder and CEO Mati Staniszewski reports that the company accelerated from $100 million in ARR in 20 months to $600 million currently, while experiencing zero attrition within its core research and engineering team. To manage this scale across 600 employees, ElevenLabs operates without product managers, organizing instead into cross-functional pods of 5 to 10 members dedicated to industry verticals like healthcare, finance, and telecom. Technical engineers are also embedded directly inside non-technical units—such as legal and talent acquisition—to build custom internal automations and inspect AI-generated code for security vulnerabilities. Legora CEO Max Junestrand reports a parallel trajectory in legal tech, sustaining a 50% quarter-over-quarter growth rate across seven consecutive quarters to scale from $1 million to $150 million in revenue. Legora similarly prioritizes technical deployment over traditional corporate overhead by utilizing "legal engineers"—forward-deployed attorneys inspired by Palantir's deployment framework—who sit with senior law partners to transition traditional legal tasks into automated, agentic workflows. Voice Interaction Psychology and Identity Protections (14:34) The widespread deployment of voice agents is shifting end-user behavior across enterprise sales and financial operations. Staniszewski observes that when dealing with sensitive matters like missed payments or debt collection at fintechs like Klarna and Revolut, customers are noticeably more transparent and candid with an AI agent than with a human representative, as the synthetic interface removes perceived social judgment. In enterprise sales, inbound voice SDRs are deployed directly after form submissions to capture richer user intent, while users broadly adopt more direct, interruptive conversational habits with automated systems. Because synthetic audio introduces identity risks, ElevenLabs established a three-pillar safety architecture consisting of cryptographic watermarking, multi-level text and voice input moderation, and public deepfake detection tools. Commercially, the platform manages digital likeness through a creator marketplace that has paid out over $22 million to voice actors, legacy licensing agreements such as Disney's interactive Darth Vader integration with James Earl Jones in Epic Games' Fortnite, and therapeutic initiatives restoring voices for individuals affected by ALS or throat cancer. Dismantling the Legal Billable Hour Model (32:33) The $1 trillion global legal services sector is experiencing structural disruption, as software currently captures only $40 billion (4%) of total industry expenditure. Junestrand notes that seed-stage startups increasingly bypass law firms entirely, leveraging AI to manage compliance, capitalization tables, and contract drafting through their first $1 million in revenue. This shift directly pressures traditional law firm cross-subsidization models, which historically generated high margins by billing junior associate labor at approximately $800 per hour while senior partners commanded between $1,800 and $4,000 per hour. Enterprises are responding by moving routine transactional diligence in-house; Legora completed four internal corporate acquisitions, with the fastest transaction closing in 12 days from letter of intent. As major firms like Kirkland & Ellis—generating roughly $10 billion annually with partner profits between $5 million and $10 million—confront these efficiencies, junior attorneys are shifting from manual document review to supervising autonomous agents, pushing the broader industry toward fixed-fee and success-based pricing models. Defensibility, Specialization, and Frontier Model Competition (25:05) Both ElevenLabs and Legora reject the premise that foundation model providers like OpenAI and Anthropic will commoditize domain-specific applications. ElevenLabs operates as a model-agnostic orchestration layer, enabling clients to swap underlying LLMs while maintaining defensibility through proprietary model architecture and a dedicated network of over 1,000 contractors labeling audio assets. Staniszewski maintains that specialized audio engineering and vertical workflows matter significantly more than sheer computing scale. Similarly, Junestrand dismisses building generalized legal foundation models as inefficient, focusing instead on narrow, latency-optimized models designed for specific high-volume workflows like multi-document tabular reviews. In enterprise sales, Legora relies on rigorous security protocols rather than infrastructure flexibility to establish its moat: despite handling sensitive contracts for weapons manufacturers and government entities, the company strictly refuses client-hosted Virtual Private Cloud (VPC) and on-premises deployments to preserve execution speed and product release cadence. While both leadership teams forecast that domain-specific agents will achieve human parity in conversational voice and complex legal reasoning in the near term, their long-term defensibility remains tied to institutional security clearances, heavy API token expenditures, and the continuous requirement to out-execute the frontier foundation labs powering their backends.

本期节目探讨了生成式人工智能在语音和法律领域的商业化突破，ElevenLabs 与 Legora 的创始团队指出，通过垂直领域工程、深层合规壁垒与工作流重塑，AI 正在从根本上颠覆传统的按小时计费服务与人机交互模式。 垂直深耕与超速增长的商业验证 (0:00) ElevenLabs 联合创始人兼 CEO Mati Staniszewski 透露，自 2023 年初发布文本转语音模型以来，公司年经常性收入（ARR）实现爆发式增长：从零到 1 亿美元耗时约 20 个月，随后在 10 个月内达到 2 亿美元，5 个月内增至 3 亿美元，目前已攀升至 6 亿美元，团队规模维持在 600 人。在法律领域，Legora 联合创始人 Max Junestrand 表示，公司连续 7 个季度保持了 50% 的环比复合增长，通过直销模式在短时间内将营收从 100 万美元拉升至 1.5 亿美元，甚至超越了知名客服 AI 公司 Sierra 的增长轨迹。 两家企业的高速增长反映出软件正在快速渗透传统高客单价的服务业。目前全球法律服务市场年支出高达 1 万亿美元，而法律科技软件的支出仅为 400 亿美元，渗透率仅为 4%。初创团队目前已经开始绕过外部律师，直接借助 AI 处理合同审查与合规事务，这表明传统重人力的专业服务正在加速被高效的软件工具所吸纳。 组织重构与跨模型编排架构 (3:45) 在工程组织与架构设计上，两家企业均摒弃了传统的软件开发与项目管理模式。ElevenLabs 内部完全不设产品经理（PM），而是组建由 5 到 10 人构成的垂直行业小组，由跨领域工程师兼顾产品设计与编码；同时，公司在法务、招聘与市场等非技术部门长期驻扎专门的工程师，用于构建自动化系统并审查 AI 生成代码的安全风险。面对与 OpenAI、Anthropic 等前沿大模型厂商的直接竞争，ElevenLabs 将自身定位为对底层模型中立的编排与交互层，允许客户自由切换各厂商的模型。 面对庞大的训练需求，Staniszewski 强调语音领域的核心壁垒在于特定架构和高质量领域数据，ElevenLabs 专门雇佣了超过 1000 名外包人员进行音频数据的清洗与标注。Legora 同样坚持垂类路线，认为斥巨资研发通用法律模型是资源浪费，未来的核心竞争力在于针对特定批量任务（例如数百份文档与提示词相乘的表格审查）训练高效率、低延迟的轻量化专用模型，从而大幅降低 API 调用成本。 声音资产的身份确权与商业拓展 (16:03) 语音技术的逼真度正在逼近甚至达到人类水平，ElevenLabs 正在构建多语种与数字形象的商业生态。平台与马修·麦康纳等知名人士合作，将其标志性的语音语调无缝翻译为西班牙语、意大利语和葡萄牙语，同时还通过创作者集市为配音演员提供授权变现渠道，累计分成已超过 2200 万美元。在企业端，迪士尼通过其技术方案永久授权并数字化延续了詹姆斯·伯尔·琼斯为达斯·维达配音的声音资产，并与 Epic Games 合作将互动对话置入《堡垒之夜》；在人道救助方面，该技术帮助喉癌患者及国会议员 Jennifer Wexton 实现了声音重构。 伴随深伪技术与声音盗用事件的增加，声音作为数字身份资产的安全保障变得至关重要。ElevenLabs 制定了三项防护机制：一是对生成的音频实施溯源与全流程水印标记；二是在输入端对文本和语音进行多层级审查，主动拦截欺诈与商业滥用；三是推出公共验证检测工具，协助公众与机构识别由其自身或开源模型生成的合成语音，并对未经授权的他人声音克隆实施严格的身份认证拦截。 法律服务的解构与合规信任壁垒 (35:15) 传统的律所合伙制依赖“合伙人低估、初级律师高估”的交叉补贴模式，资深合伙人每小时收费最高可达 4000 美元，而初级律师在传统业务中则通过大量耗时的尽职调查赚取每小时数百美元的计费。AI 智能体正在逐步拆解这种计费模式，促使大量法务工作回流企业内部。为协助顶尖律所转型，Legora 设立了“法律工程师”岗位，深入律所指导资深合伙人完成工作流再造，将初级律师的职能从虚拟数据室的关键词检索转变为调度与管理 AI 智能体。 在市场推广层面，技术优势并非唯一的决胜因素，极高的安全准入门槛构成了法律科技的真正护城河。法律行业的核心资产在于责任与信任，Legora 承载着包括国家秘密、政府机构以及武器制造商合同在内的超高敏感级数据。为了兼顾极高安全标准并保持敏捷迭代，Legora 坚持采取纯云端部署策略，严格拒绝私有化（On-prem）部署或客户定制 VPC 环境，以防止过度依赖特定基础设施而拖累产品 roadmap。 尽管语音与法律领域的 AI 智能体展现出从任务辅助走向端到端自主执行的势头，并在特定场景突破了传统图灵测试的交互边界，但全行业在跨法域绝对数据采集的完整性、模型版权边界以及对底层前沿模型高昂 API 支出的依赖等问题上，仍面临长期的商业与技术博弈。

## Background 本期节目由 All-In Podcast 于 2026 年 7 月 27 日发布在 YouTube 平台。根据视频介绍，节目深度聚焦生成式 AI 对语音与法律两大万亿级传统行业的颠覆性重构，核心探讨了 ElevenLabs 冲刺 6 亿美元 ARR 的爆发历程、无 PM 组织架构与声音 IP 防伪体系，以及 Legora 颠覆传统律所按工时计费模式、以专有垂类模型高速吞噬人工法律市场的商业实践。 ## TL;DR ElevenLabs 联合创始人兼 CEO Mati Staniszewski 详述了公司在短时间内突破 6 亿美元 ARR 的增长轨迹，解析了全员嵌入式工程师与摒弃产品经理（PM）的高效敏捷机制，并展示了声音克隆、名人 IP 变现与防范深度伪造的三层安全防护网。随后登场的 Legora 联合创始人兼 CEO Max Junestrand 则剖析了其连续 7 个季度以 50% 环比复合增速突破 1.5 亿美元 ARR 的实战经验，揭示软件如何蚕食每年 1 万亿美元的传统法律服务市场。对话进一步指出了律所过度压榨初级律师按小时计费（Billable Hour）的激励扭曲，并论证了依托 100% 判例物理级全量数据底座与特定窄场景模型构建垂直护城河的技术路线。 ## Atmosphere & style 整场访谈节奏明快且极具张力，主持人 Jason Calacanis 延续了其辛辣幽默的控场风格，穿插着踩踏板意识流 Prompt 输入的独门工作流分享与惟妙惟肖的特朗普口音模仿，不断抛出尖锐问题逼问创业者在面临 OpenAI 和 Anthropic 巨头夹击下的真实生存策略。两位受访创始人则表现得极其沉着务实与自信，从拆解上亿美元爆发式营收的硬核商业推演，到提及利用声音重建技术帮助渐冻症患者在婚礼上重新发声的人文触动，展现出极强的执行定力。现场弥漫着见证软件彻底重塑百年传统巨头（如 LexisNexis 与超级律所）的兴奋感，同时在面对诉讼合规、代码安全与数据窃取等现实壁垒时又保持着严苛冷静的敬畏感。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=0s) ElevenLabs 创下惊人商业化增速，年化经常性收入（ARR）突破 6 亿美元，团队在 600 人规模下未设任何产品经理（PM），并保持前 10 名核心研发人员零离职率。 - [31:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1904s) 法律 AI 初创公司 Legora 连续 7 个季度实现 50% 环比复合增长，通过纯直销模式将 ARR 从 100 万美元推升至 1.5 亿美元，打破企业级软件增长纪录。 - [31:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1904s) 万亿美元规模的全球法律服务市场中人工服务占比高达 96%，软件仅占 4%，随着生成式 AI 的端到端推理落地，软件正迅速吞食传统律所按小时计费（Billable hour）的高昂利润。 - [46:51](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2811s) 在垂直模型路线抉择上，Legora 明确反对烧钱自研通用法律大模型，主张借助前沿基础模型并针对表格审阅等窄场景微调专用小模型，以此降低延迟与 API 成本。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=0s) ElevenLabs 的 ARR 扩张打破传统 SaaS 曲线：首个 1 亿用时 20 个月，2 亿用时 10 个月，3 亿用时 5 个月，当前已突破 6 亿美元。 - [06:17](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=377s) ElevenLabs 采用独特的敏捷架构，全公司不设任何产品经理（PM），依靠 AI 工具赋能工程师和设计师成为复合型全栈人才，并在法务、HR 等业务团队嵌入工程师充当安全与自动化守门人。 - [14:21](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=861s) 企业客服与催收场景发生心理质变：面对冰冷温和的 AI 语音 Agent，借款人卸下羞耻感与防御心，表现出比面对真人客服更高的坦诚度和还款意愿。 - [15:34](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=934s) 面对声音克隆引发的侵权与 Deepfake 风险，ElevenLabs 构筑了包含加密溯源、双模态实时内容审核与全网通用检测雷达的三层防御体系。 - [19:31](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1171s) 声音资产催生创作者经济新模式：配音演员仅需录制一小时样本即可获得长期版税，ElevenLabs 已累计向社区创作者支付超 2200 万美元分成。 - [25:08](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1508s) 面对 OpenAI 与 Anthropic 等巨头吞并语音交互的潜在威胁，ElevenLabs 依靠模型架构重构、超千人专业数据标注团队以及深耕细分行业的专有工作流建立起护城河。 - [31:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1904s) 全球法律服务市场结构严重失衡：每年 1 万亿美元的法律支出中 96% 流向昂贵的人工服务，软件采购仅占 400 亿美元，AI 自动化存在巨大的替代红利。 - [33:08](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1988s) 传统大律所按小时计费的商业模式建立在对初级律师过度收费的基础之上，而企业正在借助内部 AI 工具自主完成尽调与合同审查，倒逼律所向固定收费模式转型。 - [36:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2204s) Legora 借鉴 Palantir 机制推出前向部署法律工程师（Legal Engineer），深入顶级律所协助资深合伙人重构定价策略并落地 AI 工作流。 - [42:31](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2551s) 法律诉讼领域不遵循 80/20 幂律法则，关键判例检索必须做到 100% 覆盖，因此构建合规数据库依赖极重度的实体档案扫描、双盲录入与校对工程。 - [46:51](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2811s) 大模型厂商推出的浅层垂直解决方案不仅未对专业法律 AI 构成威胁，反而起到了教育市场、将受挫客户引流至深度专业系统的作用。 - [50:01](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=3001s) 合规与准入是法律科技的绝对护城河；为保持极高的产品交付速度与敏捷性，Legora 坚决拒绝为客户提供私有化部署（On-prem）或 VPC 定制。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=0s) 营收狂飙：从技术突破到 6 亿美元 ARR 的爆发之路 访谈一开始，主持人便直奔主题，调侃 ElevenLabs 联合创始人兼 CEO 马蒂·斯塔尼斯泽夫斯基（Mati Staniszewski）最近手气正旺、势头如日中天（"You're on a bit of a heater, huh?"）。面对当下的生成式语音赛道，马蒂由衷感慨道：“现在正是躬身创业、构建产品的黄金时代。”尽管各大科技巨头纷纷入局，行业竞争异常白热化，但 ElevenLabs 的商业化变现速度却彻底打破了硅谷传统的 SaaS 增长曲线。 谈及具体的营收爬坡历程，马蒂回顾了公司的创业节奏。ElevenLabs 成立于 2022 年，创业的第一整年全部倾注于底层底层研究与产品研发，力求攻克此前困扰语音合成领域多年的“机械音”瓶颈，打造出业内首个听起来真正具有充沛情感和人性化表达的文本转语音（TTS）模型。2023 年初，该产品正式对外发布，随后便迎来了令人瞠目结舌的几何级增长：团队历时约 20 个月突破了首个 1 亿美元 ARR（年度经常性收入）大关；随后仅仅用了 10 个月便翻倍迈过 2 亿美元；接着只用了 5 个月迅速窜升至 3 亿美元并以此业绩收官上一年；而截至本次对谈，ElevenLabs 的年化经常性收入已经飙升到了惊人的 6 亿美元。 - **公司创立时间**：2022 年（首年专注于底层研究，打磨真正具备人类质感的 TTS 模型）。 - **产品上线节点**：2023 年初正式推向市场。 - **ARR 增长轨迹**： - 首个 1 亿美元 ARR：用时约 20 个月 - 突破 2 亿美元 ARR：用时约 10 个月 - 突破 3 亿美元 ARR：仅用时 5 个月（上一年年底收官） - 当前最新 ARR：突破 6 亿美元 ```mermaid flowchart LR accTitle: ElevenLabs ARR Growth Timeline accDescr: Timeline showing the accelerated ARR milestones of ElevenLabs from launch to 600M dollars. subgraph Launch["2022 - 2023"] A["2022: 专注底层研发<br/>突破拟真 TTS 模型"] --> B["2023 年初<br/>产品正式上线"] end subgraph Hypergrowth["ARR 爆发节点"] B --> C["20 个月<br/>达到 1 亿美元 ARR"] C --> D["10 个月<br/>达到 2 亿美元 ARR"] D --> E["5 个月<br/>达到 3 亿美元 ARR"] E --> F["当前阶段<br/>突破 6 亿美元 ARR"] end ``` --- ### [01:31](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=91s) 组织扩张与核心技术护城河：600 人规模下的文化坚守 面对如此凶猛的营收爆发与暴涨的估值，各路风险投资人几乎是“字面意义上直接找上门来排队送钱”。主持人提出尖锐问题：在高估值、资本狂欢与激烈的顶级人才争夺战中，公司该如何平衡扩张节奏与组织文化？马蒂透露，ElevenLabs 目前的员工规模已扩大至 600 人。对于一家以硬核技术立足的 AI 公司而言，既要高速纳新，又要确保面试流程、入职培训的高质量闭环，挑战极大。 ElevenLabs 的核心定位是构建一套全方位的“AI 语音通信平台”（Communication Platform for AI）。在研发端，它覆盖了从语音合成、高精度转录到交互式多轮对话编排的全栈能力；在产品端，则打通了从前端多语言营销内容本地化制作，到智能客户服务，再到运营培训和主动销售的全生命周期旅程。这种端到端的闭环对复合型人才提出了严苛要求。值得骄傲的是，马蒂的联合创始人皮奥特·达布科夫斯基（Piotr Dabkowski）本身就是顶级研究员，在他的号召下，公司最初的核心科研与工程团队至今保持着“零离职率”（zero attrition），这在动辄互相挖角的高科技前沿实验室中堪称奇迹。 - **团队规模**：当前全球全职员工约 600 人。 - **全栈技术版图**：涵盖语音生成（Speech Generation）、语音转录（Transcription）以及交互编排（Interaction Orchestration）。 - **业务覆盖场景**：跨国营销资产多语言本地化、客户服务 Voice Agent、企业运营支持与销售话术培训等。 - **核心团队稳定性**：最初建立的核心研究和工程团队（前 10 名核心人员）保持 0 离职率。 --- ### [03:45](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=225s) 敏捷小组架构与嵌入式工程师：防范 AI 时代的生产风险 回溯到 2022 年创业之初，科技圈的聚光灯几乎全被加密货币（Crypto）与元宇宙（Metaverse）所占据。马蒂坦言，在喧嚣之外起步反而是幸运的，这让团队能够避开噪音，全身心沉浸在真正代表未来的语音交互技术上。随着编程范式从低代码、无代码迈入当下的“Vibe Coding”以及非科班人员利用 AI 编写生产代码的新阶段，软件开发的底层逻辑发生了剧变，代码质量与安全风险也成倍激增。 为了在高速运转中保持极高的产品可靠性，ElevenLabs 确立了独特的敏捷组织结构：公司并不设立庞大冗杂的科层体系，而是拆分成一个个由 5 到 10 人构成的紧密小团队（Pods）。这些敏捷单元围绕电信、金融服务、医疗保健等垂直垂直行业进行针对性优化，将产品、工程与市场（GTM）紧密结合。更有趣的创新在于其独创的**“嵌入式工程师”（Embedded Engineers）机制**：ElevenLabs 不仅在技术部门配备工程师，甚至在招聘、法务、营收运营等传统非技术团队中也直接植入软件工程师。 这些嵌入式工程师承担着双重使命： 1. **自动化落地**：为业务部门量身定制工作流与自动化软件工具，直接将技术转化为内部生产力。 2. **安全防线与质量把关（Security Gatekeeper）**：指导业务人员正确使用大模型编码工具，同时充当代码部署的审计防火墙。马蒂指出，当非技术部门拥有了用 AI 快速搭建工具的能力后，最大的隐患在于缺乏代码审查和长远维护机制——一旦工具搭建者离职，系统往往沦为漏洞百出、甚至造成敏感数据泄露的技术负债。嵌入式工程师的存在，正是为了在赋能全员 AI 化的同时兜住生产环境的安全底线。 --- ### [06:17](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=377s) 彻底干掉产品经理：AI 全栈复合型人才与自研 Voice Agent 当主持人追问目前团队如何做日常管理、是否顺应趋势优化了产品经理（PM）时，甚至开玩笑问“你们是不是把 PM 都开除了？”马蒂给出了一个惊人的回答：ElevenLabs 从创立至今**从来就没有招募过任何一名产品经理（PM）**。 马蒂解释了这一激进组织哲学的逻辑：过去行业追求的理想人才往往需要同时精通编码、用户洞察和 UI/UX 设计，但现实中这种“全能通才”凤毛麟角。ElevenLabs 的做法是招募在某一领域（如代码或设计）拥有顶级专长，同时对另一领域有深刻理解的人。如今生成式 AI 带来了颠覆性拐点——借助 AI 辅助，原本只精通一门的人可以迅速跨界，从其他职能的“外行”直接跃升至“中高级水准”。例如在增长工程（Growth Engineering）团队中，一名工程师借助 AI 就能独立完成实验设计、界面交互、代码上线与数据回流分析的全流程，完全摆脱了跨部门拉扯和沟通瓶颈。 与此同时，ElevenLabs 奉行极其严苛的“吃自己狗粮”（Dogfooding）原则。为了帮助全球客户构建高质量的 Voice Agent，团队在内部业务全线上演实战： - 销售部门直接上线了自研的**入境 AI SDR（销售拓展代表）语音智能体**。 - 潜在客户访问官网时，除了常规的文字表单外，可以直接拨打电话与该 AI 智能体对话。 - 实测表明，用户通过语音沟通能够更自然、更详尽地交代自身业务痛点与技术需求，使得意向线索能够以更快的速度精准派发给对应负责人。 对谈最后，主持人深有感触地指出，呼叫中心与自动语音系统的体验正在发生不可逆的转折。过去用户拨打客服电话，深陷层层按键的“语音牢笼”（Voice Jail），恨不得疯狂狂按“0”键大喊“转人工”；而如今，高保真、低延迟的 AI 语音不仅毫无机械感，而且反应极其精准，用户随时打断也丝毫不会产生社交愧疚感，体验已经全面超越了许多疲惫的人工座席。 ### [09:40](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=580s) 从被动客服到主动式语音交互的代际跃迁 在谈到用户与智能体交互的心理变化时，Jason Calacanis 指出，如今人们已经不再需要和客服进行虚与委蛇的寒暄闲聊（small talk），可以直接打断对话，让交流以极高的效率推进。Mati Staniszewski 证实了这一趋势，并指出消费者正在从过去的排斥转向主动索求：“给我一个 AI 智能体/AI 接线员”。这种变化背后，是 ElevenLabs 过去一年（尤其是近六个月内）在企业级产品上的技术爆发：一方面是企业销售团队的强力推进，另一方面则是产品层面上核心可靠性、多 AI 模型编排能力、专业知识库与企业底层系统集成的完美融合。 ```mermaid flowchart TD accTitle: 语音交互界面的代际演进 accDescr: 描述从传统被动式客服到现代低延迟打断交互，再到未来主动式预测服务的演进流程。 subgraph Phase1["传统阶段：被动客服"] A["用户遭遇故障"] --> B["排队等待人工"] B --> C["冗长客套与低效问答"] end subgraph Phase2["当前突破：低延迟语音编排"] D["实时打断机制"] --> E["上下文记忆唤醒"] E --> F["直接切入核心解决方案"] end subgraph Phase3["下一代交互：主动式界面"] G["后台语音感知分析"] --> H["意图预测与动态界面"] H --> I["问题提出前主动化解"] end Phase1 -. 体验断层.-> Phase2 Phase2 ==> Phase3 ``` Mati 描述了一个正在到来的“消费者黄金时代”：当你打开网站或拨通客服电话时，AI 智能体会完整调用你过去的全部交互背景信息，提供丝滑精准的支持。更重大的范式变革在于人机界面的本质迁移——语音不仅是一种输入方式，更会在后台作为感知层辅助寻找信息，整个系统交互将彻底从“出现问题后寻找帮助”的被动响应（reactive），演变为在用户开口提问前就已预测并解决问题的主动服务（proactive）。 ### [11:15](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=675s) 办公室听写场景与“脚踏板”交互创新 回顾语音转文字（STT）的发展史，Jason 吐槽十年前的语音输入简直是一场灾难。当时律师们戴着头戴式耳机、使用笨重且准确率极差的 Dragon Dictate 软件，“在办公室对着电脑大声讲话会让你看起来像个白痴”，只有关起门来在私密办公室里的人才能勉强使用。而如今，办公场景变成了低声轻语（whisper）：员工们平静地对着电脑小声交代 Prompt，或者按一下智能指环与智能体沟通。Jason 盛赞了支持 ElevenLabs 底层能力的转写工具 Whisper Flow，并现场与观众互动，询问台下有谁在使用外接的物理“脚踏板”（pedal）。在发现台下只有三个人举手后，Jason 笑称他们是“怪咖”（dorks），但预言明年这会成为所有人的标配。 Mati 坦言自己还没用脚踏板，但他非常推崇便携可穿戴设备，例如 Plaude 推出的口袋录音设备 Plaude Pocket。他认为在行业峰会等高信息密度的场景中，只要在合规前提下提前声明录音，AI 就能把那些散落在交谈中原本会随风流逝的宝贵信号完整保留下来，自动结构化生成个性化笔记并安排后续跟进事项（follow-ups）。 Jason 随即给现场观众安利了自己在办公桌下配置三个脚踏板的工作流： - **踩下开启，松开即发**：配合 Whisper Flow，脚踏板踩下去麦克风开启，松开瞬间自动停止录入并发送。 - **释放 LLM 潜能的“意识流 Prompt”**：打字输入常常让人疲惫，导致用户写到一半就放弃补充细节；而脚踏板允许用户进行长达一到两分钟无拘无束的“意识流（stream of consciousness）”语音轰炸。 - **自动消化矛盾与自相矛盾的修正**：大语言模型最擅长的恰恰是从这种充满上下文、包含中途自我修正（例如“我想说这个……不对，我改主意了，换个思路”）的大段口语中精准提炼出高阶意图，得到远优于打字沟通的回答。 ### [14:21](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=861s) 人性在 AI 面前的袒露：催收场景与交互习惯差异 当被问及人对 AI 说话与对真人说话究竟有何根本不同时，Jason 开玩笑引用了 Sergey Brin 的段子——声称对 LLM 发出“人身伤害威胁”是一种非常有效的 Prompt 技巧。Mati 则从严肃商业落地的角度分享了极具冲击力的人性洞察。ElevenLabs 与 Revolut、Klarna 等金融科技巨头以及多家欧洲银行合作，处理账单提醒与未还款催收。 在以往的人工电话场景中，面临债务违约的用户面对真人客服时通常会感到极度羞耻和自尊心受挫，往往选择撒谎掩饰或直接挂断拒绝沟通。然而，面对 AI 语音代理时，人们感受到了前所未有的安全感与无评判环境，往往表现得极其诚实和坦诚，非常乐意主动透露自己真实的财务困境与家庭变故。这种面对同类时存在的巨大情感屏障，在冰冷但温和的 AI 面前瞬间瓦解。此外，用户在对 AI 讲话时风格更为干脆利落（snappy），毫不介意直接打断 AI 的说话以直奔主题，这也反向推动了交互延迟和交互模型的优化。 ### [15:34](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=934s) 声音盗用乱象、名人 IP 与三层防御防护网 随着声音克隆技术的门槛断崖式下跌，名誉侵权与深度伪造（Deepfake）迅速演变为严峻的法律与商业争议。Jason 分享了自己亲身经历的荒唐事件：有粉丝向他赞美他制作的“斗牛犬搞笑视频”（Jason 个人养了三只斗牛犬），他查看后发现，有人从他在《This Week in Startups》播客的上千小时归档音频中提取声音，利用 ElevenLabs 克隆了他的音色，制作了一个让各种狗狗讲笑话的大流量 YouTube 频道。Jason 联系对方后，对方甚至毫无防备地坦白就是用了 ElevenLabs。 随后，Jason 现场惟妙惟肖地模仿起了唐纳德·特朗普（Donald Trump），用极其浓厚的特朗普腔调即兴调侃要强索 11Labs 5% 的股份放入特朗普账户，以此交换参观白宫的机会，并顺带毒舌调侃北欧人热爱社会主义。Jason 借此探讨了美国法律在声音权益上的边界： - **合理使用（Fair Use）与模仿**：单纯的脱口秀恶搞、讽刺模仿（Parody）是受到宪法第一修正案保护的合法行为。 - **商业化侵权（Right of Publicity）**：在没有本人授权的情况下，擅自克隆公众人物声音用于广告、带货或欺诈，属于明确违法。 Jason 提到，后来他的团队试图在 ElevenLabs 上克隆他自己的声音，用来快速修补播客口播广告中读错赞助商优惠码（把“jal20”读错成了“jal25”）的片段时，系统直接弹出了封锁拦截提示：“抱歉，您不能克隆 Jason 的声音”，必须由本人进行特定活体验证流程才能放行。 面对声音作为独特“身份与知识产权（IP）”的双重属性，以及平台签约 Jamie Foxx 等一线名人的商业化进程，Mati 阐述了 ElevenLabs 构建的三层防线防御体系： - **端到端生成溯源（Traceability）**：对平台上合成出的每一帧音频实施加密标记与后台追踪，一旦发生恶意传播，可以立即定位源头账户并采取封禁及法律措施。 - **双模态实时内容审核（Dual-layer Moderation）**：在“输入文本层”与“输出音频层”同时设立实时过滤机制，对具有潜在商业欺诈、电信网络钓鱼或冒充名人的输入内容进行即时拦截与阻断。 - **全网深度伪造检测雷达（Universal AI Detection Platform）**：超越自身平台限制，面向公众与行业推出通用的鉴真检测系统。任何人都可以上传可疑音频样本，系统会迅速辨析其是否由 AI 合成，以应对全网不断扩散的语音克隆模型威胁。 ### [19:31](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1171s) 名人声音授权、多语言情绪表达与创作者生态变现 声音作为人类核心的 IP 和个人辨识度要素，在 AI 的赋能下释放出了前所未有的全新机会。ElevenLabs 联合创始人兼 CEO Mati 介绍，他们与知名奥斯卡影帝马修·麦康纳（Matthew McConaughey）达成合作，构建起了一个跨语言的声音世界。现场主持人和嘉宾随即打趣模仿起麦康纳标志性的台词“Alright, alright, alright”，并调侃起演员拍独立电影可能拿不到多少片酬，但 ElevenLabs 的股票期权可相当诱人。更不可思议的是，借助 ElevenLabs 的底层技术，麦康纳标志性的嗓音和声线不仅能在英语中保持高度逼真，还能完美跨越到西班牙语、意大利语和葡萄牙语等语种中，且说话时细腻的情感起伏与辨识度能够被 100% 完整保留。 这种合作正在打破以往教育和培训内容的局限。例如 MasterClass 过去主要提供单向观看的录制视频课程，而如今借助 ElevenLabs，这些内容直接转变为具备交互能力的生动体验。用户在学做菜时，可以由 AI 驱动的戈登·拉姆齐（Gordon Ramsay）亲自实时教学，如果菜没做好，拉姆齐甚至会用他招牌的暴脾气大喊“生贝柱都没熟（Raw scallops raw）”。ElevenLabs 不仅服务一线明星，更在早期就建立了创作者声音市场（Voice Marketplace）：普通人与配音演员录制并验证自己的声音上传后，即可授权给他人使用并获取分成。配音演员过去只能按小时拿工时费，如今仅需录制一小时声音资产，就能坐享源源不断的版税收益。 - **创作者累计分成金额**：ElevenLabs 已经累计向社区声音创作者和配音演员支付了超过 2200 万美元的分成收益。 - **灵活定价模型**：声音创作者可以选择平台推荐的默认最优分发价格，也可以自主设定授权单价。 ```mermaid flowchart TD accTitle: ElevenLabs 声音创作者变现流程图 accDescr: 展示配音演员或明星从采集验证声音到产生多语种收益的完整闭环流程 subgraph 音频资产采集 A["声音创作者 / 名人 (如配音演员)"] --> B["录制 1 小时高质量音频样本"] B --> C["ElevenLabs 平台认证与声音克隆"] end subgraph 市场与应用 C --> D["上架 Voice Marketplace"] D --> E["开发者与企业调用 (交互/本地化/游戏)"] end subgraph 商业化收益 E --> F["按需计费与版税分润"] F --> G["累计返还创作者超过 2200 万美元"] end ``` --- ### [21:43](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1303s) 声音重建的人文关怀与标志性 IP 的交互延伸 对于普通人而言，声音是个人身份认同至关重要的载体。Mati 动情地回顾道，ElevenLabs 所做过最重要、最有意义的工作，其实是帮助那些因肌萎缩侧索硬化症（ALS）或喉癌失去发声能力的人重获自己的声音。他们曾与美国国会女议员珍妮弗·韦克斯顿（Jennifer Wexton）深度合作，韦克斯顿因病失声，却依然希望激励世人，最终借助 ElevenLabs 的声音重建技术，在国会完成了历史上首个由 AI 声音发表的国会演说。此外，还有一位女士在步入婚姻殿堂前夕不幸失声，后来在 ElevenLabs 的帮助下重新找回了声线，夫妻二人重新补办婚礼并念出结婚誓词，那一刻全场亲友热泪盈眶，展现了技术带来的巨大情感联结。 在文化 IP 领域，传奇演员詹姆斯·厄尔·琼斯（James Earl Jones，曾为《星球大战》黑武士达斯·维达配音）去世前与迪士尼签署协议，授权永久使用其达斯·维达的经典声线以惠及家人。迪士尼并没有选择传统的配音模仿者，而是联合 ElevenLabs 将其搬到了极具沉浸感的交互场景中。Epic Games 在其王牌游戏《堡垒之夜》（Fortnite）中上线了具备实时交互能力的达斯·维达，玩家达到特定关卡后可以与维达实时对话交流、共同完成任务。此外，ElevenLabs 还与冥想软件 Headspace 展开深度合作，负责其多语言的本地化内容；现场嘉宾也半开玩笑地探讨，未来冥想应用是否会推出个性化交互课程，甚至让 David Sacks 声音的 AI 来带领大家“深呼吸、呼气，放松下来听听他为特朗普辩护”。 --- ### [25:08](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1508s) 巨头环伺下的差异化护城河：架构、数据工程与垂类生态 面对 Anthropic 的 Dario Amodei 以及 OpenAI 的 Sam Altman 等硅谷巨头企图将语音交互业务吞并的现实威胁，主持人尖锐提问：ElevenLabs 在底层调用前沿基础模型的同时，是否在“亲手为自己的灭亡铺路”？Mati 沉着剖析了 ElevenLabs 的应对策略与竞争护城河。首先，ElevenLabs 的平台保持“模型不可知”（Model Agnostic），客户可以在 Anthropic、OpenAI、Google 或开源模型之间自由切换，专注于构建自己的 Agent 编排机制与声音表达，而完全不会被任何单一模型厂商锁定。 在基础语音模型层面上，Mati 强调 ElevenLabs 凭借自研架构与工程积累，在文本转语音（TTS）、语音转文本（STT）、交谈轮换机制（Turn-taking）及音乐生成等各维度，持续打败巨头的基础模型。他透露了保持领先的三大支柱： 1. **模型架构而非纯算力规模**：在语音交互领域，单纯依靠扩大模型参数量（Scaling）无法解决本质问题，必须重新从底层设计契合声音交互的模型架构。 2. **专属标注数据集与千人外包团队**：网络上虽有大量未标注的音频数据，但低质且杂乱。ElevenLabs 组建了一支超过 1000 名专业外包人员的内部标注团队，对音频资产进行极致的情感、停顿和语调精细化标注。 3. **垂直化产品栈与深度工作流**：真正的企业落地极其依赖对垂直行业的理解。金融服务、医疗保健以及电信运营商对于语音工作流的要求天差地别，ElevenLabs 投入全部产品团队深耕这些细分场景，而通用大模型厂商根本无暇顾及。 4. **庞大的生态与集成壁垒**：从开箱即用的声音库、认证体系到预置的 Agent 模板，构成了新进入者无法一蹴而就的飞轮效应。 关于大模型厂商是否存在通过蒸馏与数据泄露侵权反哺自身语音模型的问题，Mati 直言这种行为确实屡见不鲜，ElevenLabs 已经部署了多种防御机制来延缓和遏制数据窃取。在自研模型方面，ElevenLabs 并不打算去和巨头争夺通用知识问答或代码生成等领域，而是把全部研发力量聚焦在“人机交互全链路（Interaction Stack）”的绝对所有权上，在语音与沟通这一赛道上直面竞争。 ### [29:07](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1747s) 图灵测试的跨越与通用人工智能（AGI）的新定义 在探讨模型与算力开销时，主持人指出，像 ElevenLabs 这样资源充足的企业，借助当今丰富且优质的开源语言模型，开发专属于自己的模型是顺理成章的演进方向。更现实的考量在于成本——企业每年往往需要向各大前沿模型（Frontier models）公司支付数以千万美元计的费用。ElevenLabs 联合创始人 Mati 坦言，公司在这方面的支出的确相当可观，但双方始终保持着极佳的合作关系，这最终也体现在他们为客户创造的核心价值上。他感慨道，如今的技术进展几乎令人难以置信：只要与 AI Agent 进行交流，就会明显感觉图灵测试已经被攻克，机器已经展现出媲美人类的智力水准；而 ElevenLabs 希望在年内将同样的突破带入语音领域，让每一次语音交流都如同两个真实人类在对话。 主持人对此深表赞同，认为只要场景合适，AI 实际上已经通过了图灵测试。他强调，三四十年前为定义人工智能或通用人工智能（AGI）所设立的各项基准测试，如今早已被全面超越，业界迫切需要一套全新的衡量标准。在主持人看来，现在的评判标准几乎苛刻到了“它是否能比地球上每个人加起来还要聪明十倍”；一旦达不到这种近乎神迹的标准，人们便轻易断言它“还不够聪明”。事实上，AGI 在某种意义上已经被创造出来，只是尚未在全场景中全面铺开；Mati 也由衷认同，在诸多特定领域，AGI 确实早已悄然成为现实。 - **模型开销与自研考量**：头部垂直 AI 企业每年向上游基础模型供应商支付数千万美元推理与 API 账单，促使其积极考虑结合开源生态构建自主专有模型。 - **图灵测试演进**：文本与逻辑层面的拟人度已基本达标，当前核心战役在于语音交互的自然度与实时拟真感。 - **AGI 认知分歧**：技术界所面对的评估框架已从“模仿人类行为”转向“全人类认知能力的超量级倍增”，导致公众低估了当前技术的实质性跨越。 --- ### [31:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1904s) Legora 的爆炸式增长与法律服务的万亿级蛋糕 随着 Mati 结束演讲，Legora（法律 AI 领跑者）联合创始人兼 CEO Max Junestrand 登台。主持人提到 Legora 的惊人增速，Max 纠正并强调这不是常规增长，而是连续 7 个季度保持了高达 50% 的环比复合增长率（Quarter over Quarter）。就在上周二完成数据核算时，Legora 凭借纯直销业务模式，实现了从 100 万美元跨越至 1.5 亿美元 ARR 的行业纪录，在耗时上甚至比 Sierra 还要快整整一个季度，成为企业级软件历史上增速最快的领跑者之一。 对话随即切入整个法律服务市场饱受诟病的痛点。主持人和 Max 共同指出，普通人和企业主在日常中最深恶痛绝的开销之一就是付给律师的高昂账单。然而，整个法律科技的结构性矛盾极其畸形： ```mermaid flowchart TD accTitle: 法律服务市场结构对比 accDescr: 对比每年 1 万亿美元的手工法律服务支出与仅 400 亿美元的法律软件支出 subgraph Spend["全球每年法律领域支出格局"] A["人工法律服务支出<br/><b>1 万亿美元 (96%)</b><br/>高度碎片化、严重受制于人力供给"] B["法律科技软件支出<br/><b>400 亿美元 (4%)</b><br/>极度渗透不足、存在巨大吞吐替代空间"] end subgraph Shift["软件驱动的市场重塑"] C["AI 自动化工作流与内部化"] D["律所重构定价模式与自研产品"] end A -.->|"软件蚕食服务市场"| Shift B -->|"快速扩张与替代"| Shift ``` 全行业每年在传统法律服务上的开销高达 1 万亿美元，全靠人工和高度碎片化的小时账单维系；而同期法律科技软件的总支出仅为 400 亿美元。这意味着整个市场高达 96% 是纯人力服务，软件占比仅有 4%。Max 认为这种极其失衡的比例是“荒谬透顶（bananas）”的，随着 AI 自动化吞吐能力的爆发，软件必将不可逆转地吞食传统法律服务的千亿级营收。 - **超高成长速度**：连续 7 个季度维持 50% 环比增长，通过直销模式刷新 1M 到 150M ARR 增速纪录。 - **万亿市场结构**：全球法律服务每年规模达 1 万亿美元，科技软件渗透率仅占区区 4%（约 400 亿美元），供给侧面临长期严重短缺。 --- ### [33:08](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1988s) 初创公司的“裸奔”法律现状与资深合伙人的溢价黑洞 主持人分享了自己在初创生态中亲眼见证的魔幻现实：有一家已经做到数百万美元年收入、融过多轮资金且拥有数十名员工的创业公司，竟然自始至终连一个专门的外部公司法律顾问都没有聘请。当主持人惊愕地质问创始人：“你们年入百万，谁来审查核心合同？”对方极其松弛地回复：“ChatGPT 帮我们搞定的，兄弟（Chat GPT, bro）。”不仅合同审查，就连极为敏感的股权结构表（Cap table）设计以及人力资源员工协议，第一批新手创始人全部是靠向 ChatGPT 提问并自动生成的规范文本（如知识产权转让协议 IP assignments）撑过来的。 面对这种看似荒诞却普遍存在的“低配操作”，Max 指出，种子期创业公司的激进自救与华尔街大银行的严谨风控不可同日而语，但它直观揭示了法律服务供给受限的本质。在传统大律所（如柯克兰·埃利斯 Kirkland & Ellis）的计费体系中，整个商业模式建立在“对初级律师（Associates）严重过度收费，而对顶级合伙人（Partners）相对低估”的基础之上： - 律所给初级律师开出每小时 800 美元乃至更高的天价，将其作为人头外包的利润机器； - 资深合伙人的账面单价或许高达每小时 1,800 至 4,000 美元，但面对涉及公司生死存亡的“赌上身家性命的大案（Bet-the-company litigation）”或价值数千万美元的商业漏洞规避，顶尖合伙人仅凭半小时经验给出的决策价值远超账面费用。 - 律所之所以固守这种按小时计费（Billable hour）模式，根本原因在于他们只能通过向初级律师大量灌注简单机械的文档审查工时来攫取巨额暴利。 --- ### [36:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2204s) 利益机制冲突与“法律工程师”的前线变革 企业客户正在彻底觉醒并拒绝为冗长的小时工时买单，纷纷将法律尽调全面收缩至公司内部自主消化。Max 以 Legora 自身的实践举例：公司年内一口气完成了 4 起商业并购，所有尽职调查全部使用自研的 AI 工具在内部高效完成。其间最快的一起并购案，从发出意向书（LOI）到最终资金交割，总共仅耗时 12 天。 主持人一针见血地指出了传统法律交易中的核心激励错配：“创始人的动力是促成交易快速交割，而传统代理律师的驱动力不仅是防范风险以防日后被客户起诉，更是要竭尽所能赚取工时费，这意味着即使他们不明说，其内在动力也是把案子无限期拖延下去。” 面对这种范式转移，传统顶级律所正陷入前所未有的焦虑与恐惧之中。Max 透露，像 Kirkland & Ellis 这样的全球超级律所年创收高达 100 亿美元，拥有四五千名律师，权益合伙人的人均年利润甚至高达 500 万至 1,000 万美元。面对 AI 这一既是“生存威胁”又是“存亡机遇”的分水岭，Legora 效仿 Palantir 的前向部署工程师（Forward Deployed Engineers）机制，设立了独具特色的**前向部署律师 / 法律工程师（Legal Engineer）**岗位： - **转型咨询与流程重塑**：派遣前向部署律师与顶级律所的核心合伙人面对面协作，协助他们重构定价策略（如从按工时计费转向固定交易费、基金募资包干费或诉讼胜诉风险提成费）。 - **工具与认知跨越**：将他们从“前 AI 时代”平稳渡向“后 AI 时代”，正如二三十年前律所从纸质文件仓库与实体打印排版向个人电脑（PC）与数字化文档管理系统转型时所经历的全面辅导。 “我们在 Legora 拥有一种非常独特的角色，称为法律工程师。正如 Palantir 拥有前向部署工程师一样，我们拥有前向部署的专业律师，他们的工作就是坐到柯克兰顶级合伙人的身边，帮助他们完成从前 AI 时代向后 AI 时代的业务蜕变。” ### [39:18](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2358s) AI 重塑初级律师角色与跨境法律服务 过去的法律科技工具通常只带来温和的生产力提升，而如今生成式 AI 能实质性地承担大量具体工作，这正在根本上重塑法学院毕业生进入律师行业的职业路径。针对初级律师岗位是否会就此消亡的疑虑，Max Junestrand 强调岗位依然会存在，但核心任务已彻底改变。律所采用的是合伙人驱动的晋升模式，必须像培养软件工程师那样从初级逐步培养出资深专业人才。但未来的晋升之路不再是“把自己关在物理数据室里，逐页审阅每份文件、挑出错别字，然后去发传真”，甚至也不再只是在虚拟数据室里按 `Ctrl + F` 进行文本检索，而是转变为学习如何编排与调度（orchestrate）执行这些脏活累活的 AI Agent。 与软件工程等通用行业不同，法律业务具有极其严格的属地化壁垒。即使在美国本土，各州之间的法律与执业资格也大相径庭，例如波士顿所在的东北部各州对竞业禁止协议（Non-compete）有很强的执行力，而在加州这类条款则几乎无法执行。面对全球跨法域展业的痛点，Legora 构筑了独特的底层数据架构：一边沉淀律所与企业的先例与内部组织数据，另一边投入极重的人力物力抓取全球所有法域的判例、立法文书与监管更新。 这种全球法律知识图谱在跨国业务拓展中展现出显著优势。例如，加州企业总法律顾问（GC）如果刚拿下首个南非客户，过去必须大费周章寻找熟悉当地法条的境外同行，现在通过 Legora 系统即可根据南非当地法规立即生成一份准确率达 80% 的初步意见书供内部研判。长久以来，全球法律与监管数据从未被如此系统性地结构化，而这正是现代社会中巨大的效率黑洞。 ```mermaid flowchart TD accTitle: 初级律师工作范式与数据护城河转型 accDescr: 对比传统法务工作与现代 Legora AI Agent 驱动的工作流程，以及底层双轨数据支撑 subgraph LegoraDataMoat["Legora 双轨数据底层"] A["律所/企业内部数据<br/>(先例与内部组织资产)"] B["全球法域监管数据<br/>(全球立法、判例与监管更新)"] end subgraph ParadigmShift["初级律师工作范式演进"] C["传统模式<br/>(物理数据室审阅、挑错、发传真、Ctrl+F)"] D["现代模式<br/>(编排调度 AI Agent、制定跨国与跨法域策略)"] end A --> D B --> D C -.->|"AI 驱动重塑"| D ``` --- ### [42:31](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2551s) 传统法律数据库的衰落与全量数据的物理护城河 作为长期垄断案例法与法律法规的老牌巨头，LexisNexis 虽拥有巨大的数据壁垒，但其母公司年收入仅有数十亿美元。主持人 Jason Calacanis 打趣道，Legora 与竞争对手 Harvey 的收入相加已有数亿美元，传统巨头看着后视镜里的 AI 初创团队，就像看到《侏罗纪公园》里的霸王龙狂奔追赶一样惊慌失措。Max 解释道，老牌传统机构极难自我转型为 AI 原生企业，一方面它们无法匹配初创团队拼命的工作节奏与工作时长，招揽不到顶尖技术人才；另一方面其内部官僚政治内耗严重。 在 AI 爆发初期，许多人断定“掌握存量数据的巨头必将胜出”，但当前市场趋势已打破了这一假设。Legora 在德国、法国、西班牙等较小法域直接与本地内容供应商合作建构生态；但美国法律检索市场由于 Thomson Reuters 旗下的 Westlaw 与 LexisNexis 的双寡头垄断而显得尤为奇特。Westlaw 甚至实质性垄断了美国政府案例报告的出版与归属权，导致公众判例并未完全公有化。 在顶尖法律诉讼中，数据并不遵循所谓的“80/20 幂律法则”。顶尖律所（如 Wachtell, Lipton, Rosen & Katz）的诉讼律师在接手法尔或涉及数十亿美元的重大诉讼时，绝不可能使用只有 80% 判例覆盖率的不完整数据库，因为任何遗漏的关键先例都可能导致全盘皆输。因此构建合格法律底座必须做到 100% 覆盖。这意味着必须进行大量枯燥且成本高昂的物理工作：从法院调取档案并支付每页 10 美分的复印费、将实体书籍海运至印度、由两组录入人员以双盲形式进行人工录入或 OCR 比对校对以确保零差错，从而获取判例引用必须精确匹配的“卷页引证”（page citations）。 - 传统法律出版巨头面临严重组织惰性：内部官僚政治繁重、无法吸引顶尖技术人才、无法适应高强度开发节奏。 - 美国法律检索市场呈现 Westlaw 与 LexisNexis 的特殊双寡头垄断，甚至哈佛大学的开放法律项目（Free Law Project / CourtListener）也无法完全替代商业全量库。 - 法律检索的严苛性要求 100% 的全量覆盖，缺乏关键判例引用的数据库对顶级诉讼完全不可用。 --- ### [46:51](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2811s) 前沿 Agent 的端到端推理与专有垂直模型的策略 随着 Claude Opus 4.5 与 4.6 等前沿大模型的推出，AI 在法律领域的表现迎来了代际跨越。旧一代数据库仅能根据关键词检索出案例，后续所有研究仍依赖律师人工完成；而新一代 Agent 能够开始进行深度的端到端案例策略推演，将证人证言、证据链与判例综合归纳。法律工作正在从简单的“AI 辅助增强”彻底转变为“由 AI 实际执行任务、人类律师负责编排调度”，这与当前的 AI 辅助软件编程趋势如出一辙。 针对 Anthropic、OpenAI 等大模型厂商是否会在垂直领域构成竞争的问题，Max 指出底层模型商目前推出的垂直套件（例如 Claude 的 Legal 解决方案）主要是一些 Markdown 技能包配置文件与少量基础集成的组合。这种浅层方案不仅不会蚕食专业市场，反而是极佳的市场教育工具。潜在客户在尝试后很快就会触碰到通用模型的浅层天花板，进而主动寻求 Legora 这种深度的专业系统，成为 Legora 高意向销售管道的重要来源。 在技术架构路线上，Legora 每年在 OpenAI 和 Anthropic 等模型供应商处消耗数百万至数千万美元的 Token，但 Max 明确反对盲目跟风训练通用法律大模型： > “我不相信针对通用智能模型的微调或自研，我认为那纯粹是在浪费时间和金钱。我坚信应当针对特定窄场景构建小模型，以此通过规模化把成本和延迟双双降下来。” 例如在 Legora 的“表格化审阅”（tabular review）核心功能中，处理任务涉及上百份合同与上百条提示词的矩阵遍历，会瞬间产生数万次 API 调用。针对合同特定字段的数据提取训练专用的微调窄模型，可以在大幅压低成本的同时降低响应延迟，这比部分竞争对手试图凭空烧钱训练一个“全能法律智能大模型”要现实有效得多。随后，对话切入至在高度监管、容错率极低的司法场景下，如何防范核心诉讼信息泄露至大模型训练集的数据安全议题。 ### [50:01](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=3001s) 法律 AI 的合规壁垒、云端部署与扩张战略 在法律服务这一对容错率极低的行业中，技术失误往往会导致灾难性的后果。Max Junestrand 强调，在法律科技赛道，团队承担着极高层级的安全责任，“合规就是我们的通行货币”（compliance is our currency）。这也是法律软件极难切入和销售的关键症结所在。当前市场上涌现出了大量的法律 AI 初创公司，但真正能够突围、跨越企业门槛的却寥寥无几。这背后的原因并非在于“造不出产品”——事实上，理解法律工作流中哪些环节能够通过 AI 创造价值其实相对直观且容易；真正的死穴在于如何把产品安全合规地送到客户手中、通过企业严苛的安全审查。 对于 Legora 而言，他们在创立初期就率先攻克了合规准入这一硬仗。而法律行业的特性在于：入场壁垒极高，但“一旦成功打入客户内部，后续的横向扩张就会容易得多”。这种高粘性与拓展优势，也正是驱动 Legora 推进后续并购（M&A）战略的核心动力。如今，Legora 承载着极高密级的核心数据，甚至包括多国政府机构的合作项目，以及大型武器军工制造商（weapons manufacturers）涉及国家机密的敏感商业合同。 针对高度敏感的数据环境，主持人询问 Legora 是否被迫为客户提供本地化私有部署（On-premise），或者是否有将本地部署列入未来的产品路线图。Max 明确给出了否定回答。他坦言 Legora 坚决不做本地化部署，未来也绝不考虑。在他看来，在客户的虚拟私有云（VPC）中进行专属部署极其耗费时间，并且会带来海量的外部环境依赖关系；这不仅会拖垮工程团队的交付节奏，更会严重拖慢整个产品路线图的迭代演进与向前推进的执行速度。对话最后，主持人对 Max 的分享表达了感谢，祝愿 Legora 持续取得成功，并在节目的标志性致辞中为对谈画上句号。 - **合规即核心资产**：法律科技的难点不在于功能逻辑的构建，而在于满足极度苛刻的企业合规审查；绝大多数法律 AI 团队都卡死在客户准入阶段。 - **高壁垒带来的网络与并购效应**：一旦突破顶级律所与企业的信任门槛，客户流失率极低且增购意愿强烈，成为平台型并购扩张的最佳跳板。 - **承载极高密级数据**：平台客户已覆盖政府机构及军工制造巨头，系统内直接处理并托管涉及国家安全级别的机密合同。 - **坚决拒绝 VPC 与本地化部署**：为维持极高速度的迭代与产品交付能力，坚决放弃会导致庞大维护包袱和技术依赖的 VPC/On-prem 方案，全面采用标准云端服务。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=0s) 现在正是躬身创业、构建产品的黄金时代。 — Mati Staniszewski > [09:40](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=580s) 给我一个 AI 智能体/AI 接线员 — Mati Staniszewski > [11:15](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=675s) 在办公室对着电脑大声讲话会让你看起来像个白痴 — Jason Calacanis > [15:34](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=934s) 抱歉，您不能克隆 Jason 的声音 — ElevenLabs 系统提示 > [33:08](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=1988s) ChatGPT 帮我们搞定的，兄弟 — 初创公司创始人 > [36:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2204s) 创始人的动力是促成交易快速交割，而传统代理律师的驱动力不仅是防范风险以防日后被客户起诉，更是要竭尽所能赚取工时费，这意味着即使他们不明说，其内在动力也是把案子无限期拖延下去。 — Jason Calacanis > [36:44](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2204s) 我们在 Legora 拥有一种非常独特的角色，称为法律工程师。正如 Palantir 拥有前向部署工程师一样，我们拥有前向部署的专业律师，他们的工作就是坐到柯克兰顶级合伙人的身边，帮助他们完成从前 AI 时代向后 AI 时代的业务蜕变。 — Max Junestrand > [46:51](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=2811s) 我不相信针对通用智能模型的微调或自研，我认为那纯粹是在浪费时间和金钱。我坚信应当针对特定窄场景构建小模型，以此通过规模化把成本和延迟双双降下来。 — Max Junestrand > [50:01](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=3001s) 合规就是我们的通行货币 — Max Junestrand > [50:01](https://www.youtube.com/watch?v=J0bce9WQJ-g&t=3001s) 一旦成功打入客户内部，后续的横向扩张就会容易得多 — Max Junestrand ## Mentioned Mati Staniszewski, Piotr Dabkowski, Max Junestrand, Jason Calacanis, David Sacks, Sergey Brin, Sam Altman, Dario Amodei, Donald Trump, Matthew McConaughey, Gordon Ramsay, Jamie Foxx, Jennifer Wexton, James Earl Jones, ElevenLabs, Legora, Sierra, Harvey, OpenAI (ChatGPT), Anthropic (Claude Opus), Google, Palantir, LexisNexis, Thomson Reuters (Westlaw), Kirkland & Ellis, Wachtell, Lipton, Rosen & Katz, Epic Games (Fortnite), Disney, MasterClass, Headspace, Revolut, Klarna, Whisper Flow, Plaude Pocket, Dragon Dictate ## Video description (0:00) ElevenLabs' $600M ARR Ramp, 600 Employees & Life Without PMs (15:34) Celebrity Voice Deals, Deepfake Impersonation & Racing OpenAI and Anthropic (31:42) Legora's Hypergrowth, Disrupting Law Firms & the Billable Hour (42:31) LexisNexis Decline, Legal Data Moats & Legora's Narrow AI Models Thanks to our partners for making this possible! Airwallex is a leading global payments and financial platform for modern businesses, offering trusted solutions to manage everything from business accounts, payments, treasury, and spend management to embedded finance. https://airwallex.com/allin Oracle powers AI at every scale—from frontier labs to enterprise production. Your data. Leading models. No lock-in. One platform, architected for AI. Built for business. Visit https://www.oracle.com/ai #OracleAIExperienceLive Follow Mati: https://x.com/mati Follow Max: https://x.com/MaxJunestrand Follow the besties: https://x.com/chamath https://x.com/Jason https://x.com/DavidSacks https://x.com/friedberg Follow on X: https://x.com/theallinpod Follow on Instagram: https://www.instagram.com/theallinpod Follow on TikTok: https://www.tiktok.com/@theallinpod Follow on LinkedIn: https://www.linkedin.com/company/allinpod Intro Music Credit: https://rb.gy/tppkzl https://x.com/yung_spielburg Intro Video Credit: https://x.com/TheZachEffect @RAISESummit #allin #tech #news [Source](https://www.youtube.com/watch?v=J0bce9WQJ-g)

[0:00] Mati Staniszewski: 现在正是躬身创业、构建产品的黄金时代。 [9:40] Mati Staniszewski: 给我一个 AI 智能体/AI 接线员 [11:15] Jason Calacanis: 在办公室对着电脑大声讲话会让你看起来像个白痴 [15:34] ElevenLabs 系统提示: 抱歉，您不能克隆 Jason 的声音 [33:08] 初创公司创始人: ChatGPT 帮我们搞定的，兄弟 [36:44] Jason Calacanis: 创始人的动力是促成交易快速交割，而传统代理律师的驱动力不仅是防范风险以防日后被客户起诉，更是要竭尽所能赚取工时费，这意味着即使他们不明说，其内在动力也是把案子无限期拖延下去。 [36:44] Max Junestrand: 我们在 Legora 拥有一种非常独特的角色，称为法律工程师。正如 Palantir 拥有前向部署工程师一样，我们拥有前向部署的专业律师，他们的工作就是坐到柯克兰顶级合伙人的身边，帮助他们完成从前 AI 时代向后 AI 时代的业务蜕变。 [46:51] Max Junestrand: 我不相信针对通用智能模型的微调或自研，我认为那纯粹是在浪费时间和金钱。我坚信应当针对特定窄场景构建小模型，以此通过规模化把成本和延迟双双降下来。 [50:01] Max Junestrand: 合规就是我们的通行货币 [50:01] Max Junestrand: 一旦成功打入客户内部，后续的横向扩张就会容易得多

[0:00] Mati Staniszewski: it's the best time to be building [3:45] Mati Staniszewski: We don't have any PMs [19:41] Jason Calacanis: all right, all right, all right [19:41] Gordon Ramsay: raw scallops [25:05] Mati Staniszewski: being in the arena [31:42] Max Junestrand: exponential [32:33] 4% software, 96% service [32:33] fun diligence targets one day [35:15] overcharge for the associates and you actually undercharge for the partners [35:15] drag it out [37:47] orchestrating the agent that will be doing that work [40:27] 80% accurate response [42:31] end-to-end work [47:26] Max Junestrand: total waste of time and money [50:04] responsibility and compliance is our currency

[0:00] You're on a bit of a heater, huh? [0:02] >> It's It's the best time to be building [0:05] >> and revenue has surged, but you face [0:10] really intense competition. Let's go [0:13] right at that to start. [0:16] >> If you were building a global financial [0:17] system from First Principles today, you [0:19] wouldn't build it on 50-year-old legacy [0:21] rails. You'd build Airwalks, one AI [0:24] native platform for global accounts, [0:26] cards, and payments. is designed to make [0:28] the entire world feel like a local [0:29] market. Others are bolting AI onto [0:32] broken infrastructure, but Airwall was [0:34] built for the intelligent era from day [0:36] one. Stop paying the legacy tax and [0:38] start building the future at [0:39] airwalls.com/allin. [0:41] Airwalls built for the future. [0:44] >> 350 million in what 2 or 3 years and I'm [0:48] hearing numbers five or 600 million now. [0:51] Tell us about the revenue ramp of the [0:53] company from the moment you released the [0:55] software to today. the the product's [0:57] been in market for 40 months, 50 months, [0:59] you tell me. [1:00] >> Spot on. We started company 2022. First [1:03] year was all about building the the [1:05] research and the product to really [1:07] kickstart the work. We built the first [1:10] texttospech model that finally could [1:12] sound human. Released it in 2023, [1:14] beginning of 2023. Then it took us [1:17] roughly 20 months to get to the first [1:18] 100 million in ARR. Roughly 10 months to [1:21] get to 200, 5 months to get to 300. Um [1:25] and um and that's how we closed end of [1:29] the last year and now we are at 600. [1:31] >> You're at $600 million in revenue. This [1:34] is just extraordinary. Um how many [1:36] employees now? Because the company's [1:38] obviously hit incredible valuations, [1:42] but you have to fill in that valuation [1:44] and you you're competing at a very high [1:46] level for talent. So, so tell us about [1:49] how many employees you have now and how [1:51] you maintain the culture of the company [1:53] when revenue is ripping, investors are [1:56] throwing money at you, showing up at [1:58] your doorstep. I mean, quite literally. [2:01] Um, but you've got to run the company. [2:03] You've got to build a culture. So, how [2:05] many employees now and how are you [2:06] dealing with these competing [2:10] uh priorities? [2:11] Yeah, that's the that's the key element [2:13] of how you all for us the the element of [2:16] like how we can maintain the culture [2:17] despite the quick growth is is is kind [2:20] of critical and how we optimize both the [2:22] interview cycle how we are bringing [2:24] people on board how we on board them [2:26] with 600 people today um so also very [2:28] quick growth on that people's side and [2:30] as a company we combine research and [2:32] product so we we we are building a [2:35] communication platform for AI on the [2:38] research side this includes everything [2:39] across audio generating speech [2:40] transcribing speech, orchestrating [2:42] speech for interactions on the product. [2:44] This is how we can complete the entirety [2:46] of the customer journey from marketing [2:48] and creating assets and localizing them [2:50] internationally through customer support [2:52] with voice agents to proactive [2:54] enablement of how voice agents can help [2:56] in operations, training and sales. So [2:58] this requires a lot of different talent [3:01] um and and and and a part of that [3:03] revenue growth is actually reflection of [3:04] the functions we've grown over time. So [3:07] from the original team very research [3:09] very engineering heavy from the first 10 [3:12] people we had zero attrition everybody [3:13] is uh uh still at the company from those [3:16] core research and engineering talent [3:18] building together with us so so far been [3:20] able to out compete and I think the [3:21] common [3:23] Fred and and then credit to my [3:24] co-founder who is incredible researcher [3:27] himself we've been able to assemble the [3:29] team that is truly excited about solving [3:32] audio solving interaction and building [3:33] that research and if they if they are [3:36] looking for an opportunity out there and [3:38] looking for a company to join and solve [3:40] that we are we are one of the the the [3:43] leading if not the leading place to do [3:44] that [3:45] >> and you started before [3:47] AI was so impactful at making software [3:51] >> right [3:52] >> so when you were starting four years ago [3:54] 5 years ago and working on this building [3:57] software was limited to low percentage [4:00] uh of the population of planet earth you [4:03] know the number of people could write [4:04] code and now here we are, you know, went [4:07] from vibe, we had a no code moment, then [4:09] vibe coding, and now we actually have [4:12] people building production code who are [4:14] not developers. You have developers [4:16] going 10x and token maxing. [4:19] How has building software changed [4:22] internally? And how do you deal with [4:24] making sure that the code is really high [4:27] quality? because people are paying you [4:29] this money [4:31] >> uh but they're going to demand really [4:32] high quality product since they're [4:34] spending so much money with you. [4:36] >> Yeah, that's it's also true that 2022 [4:38] was still the year where topics of the [4:39] day were crypto and um and metaverse. So [4:42] with the building there was also the [4:44] best time to start because we could [4:45] actually take a take a bit of time to [4:48] focus on what we thought is the future. [4:50] Um but the the way we are structured is [4:53] a lot of small teams especially across [4:55] the product engineering but also in how [4:57] we think about go to market optimized [4:59] for specific industries telco financial [5:01] services healthcare. So every unit is [5:03] very tightly knit together. Um and we do [5:06] that across the company. Uh so it's [5:08] usually five to 10 people teams that [5:10] that that run ahead. And inside of each [5:12] of those teams the decision we took [5:14] which is slightly different than how [5:16] it's usually structured. We embedded [5:18] engineers in in in every place and even [5:20] in the places which aren't engineering. [5:22] So our talent team will have an [5:23] engineer, our legal team will have an [5:25] engineer, our revenue engineering or go [5:28] to market engineering have engineers [5:30] embedded all across and those people [5:32] will have two roles. One is of course [5:34] creating automations and bringing the [5:37] software inside of that team, but second [5:39] is actually helping everybody else do [5:41] what you said, which is make sure that [5:43] people are adopting AI, but also there's [5:46] a security check for everything they [5:47] deploy because ultimately if if you're [5:49] not using a lot of the coding software, [5:51] a lot of the co-working software, then [5:54] you're probably in the wrong spot. If [5:55] you're using too much of it, that is [5:57] also a flag because you maybe are not [5:59] doing that in the right way. Um and of [6:01] course as you start bringing that into [6:04] the sites of the organizations that [6:05] never were exposed they frequently can [6:07] create but not necessarily review [6:10] whether that's actually doing behind the [6:13] scenes all the secure ways or anything. [6:14] So that's an essential role in the in [6:16] the in the company. [6:16] >> Yeah. We the it's fantastic that [6:19] everyone can build software until you [6:22] put it into production and you have a [6:23] leak. [6:24] >> Yeah. uh or that person leaves the [6:26] company and people forget they built [6:28] that software and it's just deprecating [6:30] uh on its own. The other thing that [6:33] seems to have changed is management. [6:35] When you had 10 developers in your pod [6:38] or six, you had a UX designer, you might [6:41] have a pure graphic designer, you'd have [6:43] a product manager, they rolled up and [6:46] then suddenly, you know, we watched over [6:48] the past 3 years. Oh, hey, this is [6:51] pretty good at um summarizing what [6:54] happened on the call. Oh, it is actually [6:56] creating action items and it's telling [6:58] us what to do next. Oh, and it's, you [7:00] know, doing uh all the different stories [7:02] in our conbon board. And now, how do you [7:05] think about product managers and [7:08] management as the CEO and as the [7:10] co-founder? [7:11] >> Yeah, we we So, we [7:12] >> you fired them all, right? [7:13] >> We don't we don't have any PMs, [7:15] >> right? [7:16] >> Did you ever or did you have you never [7:17] did? Okay. [7:18] >> Never did. thought it's it's a little [7:20] bit of what you mentioned was also [7:22] before the true AI impact started which [7:25] was ideal ideal person in that role can [7:28] code can understand the customer can [7:30] understand design of course that's very [7:31] hard to find there's no truly that many [7:34] people that are expert in any all all of [7:36] those fields at the same time so we [7:38] optimize for profiles that are experts [7:39] in at least one of those fields but [7:41] understand at least one other field [7:44] really well to your point what we are [7:45] seeing now there's if you can do a [7:48] little bit of all with AI you can maybe [7:51] step change from being an amateur to [7:53] being a advanced level maybe not an [7:56] expert level so suddenly you are not [7:58] bottlenecked on all the other functions [7:59] to do your work in u in growth [8:02] phenomenal part growth engineering a [8:04] person can design experiment ship an [8:05] experiment it's working and bring it [8:07] back um we also have the privilege where [8:09] we are using a lot of our product [8:11] ourselves so to be able to do that [8:14] ultimately to help everybody else create [8:16] voice agents we ourselves need to create [8:18] voice agents too. Um so we are seeing [8:20] that also in the non-traditional [8:22] functions even in go to market like you [8:24] need to be able to create a version of [8:27] that if we are offering that to the [8:28] customers too and and we do we created [8:30] our inbound AI SDR agent that in [8:34] addition to the form that you fill on [8:36] the website you have an agent that you [8:39] can call [8:40] >> and uh and people are um of course e can [8:44] give all the information in a much [8:45] easier and quicker way but the second [8:47] thing that happen is people also leave a [8:48] lot more information so you can get [8:50] connected to the right problem and right [8:52] person a lot a lot quicker. Um so we are [8:54] seeing that kind of phenomena all the [8:55] time where actually using a lot of [8:58] tooling makes you yourself better in [9:00] your job overall and in 11lapse in our [9:03] specific tooling that we are solving for [9:04] customers. [9:04] >> Yeah, it seems like the use case of [9:08] calling on the phone and talking to a [9:10] computer or previously going through [9:12] voice jail and and it was incredibly [9:15] arduous and painful and annoying. [9:18] made you just say operator and hit the [9:20] zero button like as fast as possible. [9:22] But now it seems to have turned a corner [9:24] where [9:25] >> talking to a human you I almost feel bad [9:27] talking to a human where I'm like I am [9:29] so sorry I'm wasting your time with this [9:31] and the AI is just so much more precise [9:34] and the fidelity is so great that when [9:36] you tell them what you're looking to do [9:38] and you cut them off you don't feel bad [9:40] you don't have to make small talk. Is [9:42] that what you're seeing in your customer [9:43] base in terms of the ability in real [9:47] time to interrupt the agent to interrupt [9:50] the uh you know conversation and just [9:52] move faster has made consumers and [9:56] companies basically embrace the [9:58] technology. [9:58] >> Yeah, it's it's slowly becoming that you [10:00] will be asking for a give me an AI agent [10:02] effective call AI operator. But uh we we [10:05] are seeing a transition. We're suddenly [10:07] and that's you know the the the biggest [10:09] fuel of the recent growth for us is [10:11] enterprises sales team just doing [10:14] incredible work but then finally the [10:16] product combines the reliability that's [10:19] core with [10:21] the orchestration for a lot of the AI [10:23] models but also the knowledge and the [10:26] integrations to provide you the right [10:27] experience. And yeah, I think it it was [10:29] a step change in the last 12 months and [10:31] especially in the last six of of how [10:33] good that experience became where it's [10:35] like this golden era of consu for for [10:36] the consumers out there customers and [10:38] and customers is coming where you're [10:40] going to actually open a website um um [10:44] call an agent and have um the agent have [10:47] information from your past interactions [10:48] and deliver that help. And I think we'll [10:50] see this kind of interesting phenomena [10:51] combining your previous question and [10:52] this where now of course you are [10:54] reaching frequently when you have a [10:56] problem and you're asking for help but [10:58] ultimately a the whole interface will [11:00] change and morph depending on how you [11:02] are operating with that inter interface [11:03] with voice being helping you in the [11:06] background find that information will [11:08] shift from reactive to proactive to help [11:09] you get uh that help before you [11:11] potentially ask for it. Um and we are [11:14] seeing those examples those examples too [11:15] >> seemed to me that speech to text um had [11:19] a major blocker again in fidelity 10 [11:22] years ago lawyers would put on dragon [11:24] dictate if you remember that terrible [11:26] software they get a headset [11:28] >> and it seemed like the the big blocker [11:31] was you felt like an idiot talking to a [11:34] computer in an office, right? And so [11:37] people who did it quietly in their [11:39] office, you know, they they kind of got [11:40] away with it. Um, but now we see [11:42] something very different. The whisper in [11:45] the office. People very quietly talking [11:48] to their computer, giving it a prompt, [11:50] you know, and talking to their agents. [11:52] And now there's a ring out. You can [11:53] press it. And uh I use a really cool [11:56] product called Whisper Flow. I don't [11:58] know if they use 11 Labs on the back [11:59] end. I [12:00] >> they they use they use and and a few [12:03] others uh as well. And they they are [12:05] doing phenomenal work too. [12:06] >> Whisper flow is just a tremendous [12:07] product. And then I got a pedal. Does [12:10] anybody here use a pedal on their [12:12] computer? Raise your hand if you're a [12:14] com There's one dork, two dorks. [12:18] Any others? Raise it high. Oh, she's [12:21] half dork. Okay, so there's about three [12:23] and a half dorks here. Next year, this [12:25] is going to be Do you have a pedal? [12:27] >> I don't. I I have I [12:28] >> Have you considered a pedal? [12:30] >> I I should consider a pedal. I love the [12:32] devices that you can wear and the [12:34] >> I have the Plaude. It's incredible. [12:36] >> Plaude Pocket. Phenomenal. Like so good. [12:38] And especially in events like this, I [12:40] feel [12:41] >> if you pre pre preempted that you're [12:44] recording, of course, [12:45] >> but how incredible would it be that all [12:47] the signal on the conversations that [12:49] otherwise disappear, you maybe tap tap [12:50] few notes here and there to try to get [12:52] signal afterwards if you can just have [12:54] that automatically fill your specific [12:56] notes and make sure you do your [12:57] follow-ups. Phenomenal. [12:58] >> All right, so let me make the case for [13:00] the pedal. Okay, I have a I have three [13:04] pedals under the desk and I I think I'm [13:07] trying to figure out what the company [13:08] is, but with Whisper Flow, you press [13:11] down, it turns on, and uh you talk and [13:14] then you let it go. And one of the [13:17] annoying parts of working with an LLM is [13:20] typing. And you're kind of like [13:21] exhausted when you're giving it the [13:23] prompt. So, you stop prompting. But if [13:26] you're a professional artist [13:28] like me and a talker, this is like [13:30] incredible because when I press the [13:32] pedal down, I just give a stream of [13:34] consciousness now. And it turns out what [13:36] these LLMs actually do really well with [13:39] is taking a massive stream of [13:42] consciousness where you just keep [13:44] talking and talking and talking. So I'll [13:45] give it a one to twominut prompt. Then I [13:49] let go and it has changed everything. [13:53] Everything. It's it's you know like the [13:56] whole experience is changing so much a [13:58] similar version of what we see happen is [14:02] you know how you have a you want to say [14:03] a thought and then you're like okay I [14:05] actually want to change and say [14:06] something else [14:07] >> now you have those two context combined [14:09] and the experience you get as an answer [14:11] is so much better uh so we already see [14:14] that as an experience but even the the [14:16] previous example of like people are [14:17] adjusting how they speak to AI versus [14:19] how they speak to human people are [14:21] asking [14:21] >> how so yeah what what how should you [14:23] speak to the LLM. We saw Sergey Brin say [14:26] threaten it with bodily harm. It's a [14:28] very effective technique if you haven't [14:30] tried it, but what are the things that [14:32] are different when you're talking to the [14:33] LLM? [14:34] >> The the the a specific emotional [14:37] example, we we work with a lot of um [14:39] financial services companies uh uh uh [14:42] Revolute, Clara, Pogbank. And some of [14:45] the frequent case not in all of them is [14:47] of course how you remind people about [14:48] payment or that you collect the de that [14:52] from the people that aren't answering [14:53] and frequently people would naturally [14:57] feel ashamed of telling the real [14:59] situation with AI people are much more [15:01] open to share what actually happened [15:04] give the information and suddenly this [15:06] emotional block of like in front of [15:07] other human I don't want to be able to [15:08] to say all of that is is is very [15:10] different so that's different um usually [15:14] people are more snappy with AI voice [15:16] agent. It's like, you know, like quick [15:17] responses. [15:18] >> Yeah. You don't mind cutting it off. [15:20] >> Exactly. So, you can like kind of go [15:22] through to to the point you want much [15:24] quicker. Um, which you needed to like [15:26] change a little bit of the interaction [15:27] model too, which which is working. Uh, [15:30] but we'll work on the pedal and whether [15:31] whether we should we should do [15:34] >> a little bit about celebrities on the [15:37] platform. You have some celebrities who [15:39] are on there. Um, you also have um an [15:43] issue with impersonation. [15:46] Um, I know this because somebody was [15:48] like, "Oh my god, I love your bulldog [15:50] videos." Many people know I'm a big fan [15:52] of bulldogs. I currently have three. Um, [15:55] and I said, "I'm sorry, I don't know [15:57] what you're talking about." And they [15:58] sent me a channel where somebody had [16:01] created a bunch of dogs telling jokes [16:03] and they made one and I guess they were [16:06] looking for a podcaster. So they used [16:08] the this weekend startups archive and 11 [16:12] lab to create my voice and do this huge [16:15] channel. And I contacted them and I [16:16] said, "Oh my god, it's very flattering. [16:18] How how did you do this? This is like a [16:20] year or two ago." And they said, "Oh, I [16:21] used 11 Labs." So I I think I emailed [16:23] you about it. I'm like, "H how do you [16:25] protect against this in advertising in [16:29] the law in the United States? I'm not [16:31] sure about here in France. I'm sure they [16:33] have 17 laws for this. We have one. Um, [16:36] you guys are great at regulations unless [16:38] no offense. And the French guy over here [16:40] is like, "Oh, Mond shal [16:44] um the that's my French angry developer [16:48] guy." Um, I cannot smoke in the Lou. Um, [16:53] this is crazy. Um, and so [16:58] is super like interesting. um with this [17:03] right to privacy and I think you've got [17:05] a quick education on this because you've [17:06] had a couple people I'm sure write you a [17:08] legal letter what it basically means is [17:10] you you can't take somebody's voice and [17:12] use it to you know do commerce in the [17:14] world you can use it for parody there is [17:17] fair use I can do a Donald Trump [17:19] impersonation up here if I like [17:22] we're going to take about 5% of 11 lab [17:25] stock is okay with you put them in Trump [17:26] accounts sounds good okay and for that [17:30] you get to come to the White [17:31] Okay, thank you. Um, [17:34] nasty guy. Wouldn't give 5%. [17:37] Loves socialism, but not America. It's [17:40] the problem with the Nordics. Um, [17:44] nasty, nasty socialism. Then I noticed [17:47] when my guys wanted to clone my voice so [17:49] that they could fix the ads where I [17:51] mispronounce something or I do the wrong [17:53] promo code, use the code jal 20. They [17:56] like were like, "It's 25, dummy." And [17:58] I'm like, "Okay, I have dyslexia." and [18:00] then they redid it and it was like I'm [18:01] sorry you cannot uh clone Jason's voice [18:05] and then it's like I have to go in there [18:06] and do it and you put a bunch of [18:08] protections in there. So explain what's [18:10] happening in that regard in terms of [18:13] people's you know concerns around this [18:15] and then the other side which is the [18:17] opportunity because I think you got [18:18] Jamie Fox and some other folks actually [18:21] that you paid for their voices. [18:23] >> Yeah. No, the the the the voice is [18:26] identity and IP. It's like you know when [18:28] you when you speak a certain way people [18:30] recognize it can feel that emotion and [18:32] you know to some extent it was it it was [18:35] a a um it could be a problem could be [18:38] opportunity before I mean as you did [18:40] impersonation of of of the President [18:42] Trump it's of course similarly uh uh u [18:46] something that is possible even of a [18:48] human uh not specifically AI but um for [18:51] us on the safeguard side you know over [18:55] over last years We took the role as we [18:58] are leading on our development [18:59] development. We also need to lead on a [19:00] lot of the safeguards. So that's like a [19:02] critical element. We do three things. [19:03] One, trace everything that's generated [19:05] so we can take action when needed. Two, [19:08] now we moderate both on the voice and [19:09] text level. So if you were to input [19:11] something that would be commercial in [19:13] nature or uh would try to scam someone [19:15] that gets flagged, we can block it. And [19:17] now free because over last years we've [19:19] seen the development of those models [19:21] more broadly. how how can we create [19:23] systems for the wider world so people [19:25] can upload a sample and get information [19:27] whether it's AI or not immediately um [19:30] and we do it for 11 Labs but we also do [19:31] it for other open source models the [19:34] interesting part given that it's such a [19:36] good um IP um and and part of your your [19:39] your your element it opens up new [19:41] opportunities so we partnered with [19:43] Matthew McConna on creating a world [19:45] >> all right all right [19:46] >> and across languages and it's the first [19:48] >> I haven't got paid a lot of money for [19:49] these independent films but oh 11 lab [19:52] stock is juicy. [19:53] >> The [19:54] >> yum yum. [19:54] >> Could you do it? Could you do it in [19:56] Spanish? [19:57] >> It's a it's it's a fugazia fugazi. [20:00] >> Yeah. But the the crazy thing with the [20:02] with AI technology open is that now the [20:04] voice can be carried not only English [20:06] but also in Spanish and Italian and [20:08] Portuguese and you can still have [20:10] exactly that element of emotions coming [20:13] through. Um so that that's kind of a a [20:16] good example there but we've seen that [20:17] with master. [20:17] >> What do you pay these guys? What does it [20:19] cost to get Matthew McConn? Is this like [20:20] an eight figure deal, seven figure deal? [20:22] You give him a little equity? [20:24] >> Always depends. Uh so like you know the [20:26] masterclass for example is a good [20:27] example where they worked with talent [20:29] directly. And here you have u previously [20:32] a static content that you would learn [20:33] from. Now um you have interactive [20:36] content. So you have Gordon Ramsey [20:38] teaching you how to how to cook in the [20:40] kitchen. He can scream at you if you're [20:42] not doing [20:43] >> raw scallops raw. So, so, so that is [20:47] that is definitely a [20:48] >> So, they're doing characters now or or [20:50] AI instances using 11 Labs. So, you can [20:53] interact with them as part of your [20:54] subscription. [20:55] >> Exactly. But as a company, what we now [20:57] do and this from the beginning, we [20:58] created a marketplace where people can [21:00] create their voice. We authenticated, [21:02] you can share it, and you earn money. [21:04] Today, we paid back over $22 million [21:06] back to the community of of of talent. [21:08] >> Really? uh which so those voiceover [21:10] actors now who got paid as hourly [21:15] workers sometimes they get a little [21:16] backend if they were doing a commercial [21:18] or something now they can spend an hour [21:21] reading create an 11 labs voice and then [21:23] license it out [21:24] >> 100% and then [21:25] >> do they get to pick their price or you [21:27] pick the price [21:28] >> the depends on the model we do we do [21:30] both so you can you can either give it a [21:32] default that lets us distribute that [21:34] slightly more optimally or you can pick [21:35] yours and the use case is going to be [21:37] going to be different and like you opens [21:39] up a set of incredible opportunities in [21:41] the dynamic context in other languages. [21:43] Uh but maybe a last one on that like [21:45] voice is such a big part of identity and [21:48] probably probably our most important [21:50] work was actually working with people [21:52] that lost their voice due to ALS due to [21:55] throat cancer and working on bringing [21:57] that voice back. So he worked with [21:58] congresswoman in the US Jennifer [22:00] Jennifer Wexon who lost it and wanted to [22:03] continue inspire others that you can do [22:05] incredible work despite that and uh and [22:08] was the first speech delivered in in in [22:10] in in Congress or more recently I think [22:13] this was my the most uh like heart uh [22:16] warming story. There was um there's a [22:18] woman that um that's wanted to get [22:20] married lost her voice before she could [22:22] get married. [22:23] >> Oh wow. [22:23] >> And then they decided to redo the [22:26] marriage together. do the vows again [22:27] >> and do the vows [22:29] >> and you could see the whole family just [22:31] >> for the first time hearing the vows. It [22:33] was just uh you could you could feel the [22:36] emotions that you can see in any other [22:37] any other way because the voice is such [22:40] a connecting thing. [22:41] >> Yeah. And and you've done it for some [22:43] iconic voices. My understanding is the [22:45] estate of James Earl Jones. I'm not sure [22:48] if they did he pass? Is James Earl Jones [22:50] alive? Can somebody pass? [22:51] >> He he passed, right? [22:52] >> Yes. But before he passed, I think he [22:55] did a deal with Disney and he said, [22:58] "Listen, for my family, I would like to [23:02] license the Darth Vader voice for all [23:04] time to Disney." They gave him some [23:06] incredible deal. And then they were left [23:09] with, "Well, how do we actually do this? [23:10] Do we get a voice impersonator, but [23:12] instead they went to you?" Talk a little [23:14] bit about that deal and how it went [23:15] down. And is that what they used [23:18] recently, you know, in in uh some of the [23:20] new films with Darth Vader? There's a [23:22] new Darth Maul series [23:24] um where they have Darth Vader and did [23:26] you power that? [23:27] >> I don't know what I can say about the [23:29] new things but definitely the big use [23:31] case that that big big completely new [23:33] experience uh was in the gaming space [23:36] where yes [23:37] >> uh Fortnite so Epic Games [23:41] game Fortnite launched um Darth Vader [23:44] which people and players could interact [23:46] with live in partnership with the state [23:48] in partnership with Disney. So every [23:51] player after reaching a certain stage [23:52] could have a Darve interact and help you [23:54] solve the missions. And we are seeing [23:56] that kind of mode coming up more and [23:59] more often of how you can effectively [24:01] extend extend your likeness your like [24:03] you said publicity into interactive use [24:06] cases bring it across the world uh [24:09] together. Uh so that was exactly that [24:11] model and now we are working on on a on [24:13] a one of the public one is headsp space. [24:15] So Headspace has a great meditation. [24:18] This is the uh second uh greatest [24:20] meditation app right behind Comm [24:23] >> which you are an investor of. I am I [24:25] didn't realize you're right. I did and [24:27] it was a $4 million company [24:29] >> is incredible. I think their their team [24:31] >> but anyway you were you were working [24:32] with the the um the second place. [24:34] >> Exactly. So they not exactly the second [24:36] place but exactly to the working part. [24:38] They so they localize a lot the content [24:40] and and calm I think is trying some of [24:43] the interactive elements. Could you have [24:45] a meditation lesson that's personalized [24:47] to you? [24:47] >> Which uh which we we would uh hopefully [24:50] love to to to [24:52] >> and like imagine just you know so many [24:55] voices. [24:56] >> David Saxs is defending Trump. Take a [24:59] deep breath in. Breathe out. Breathe in. [25:03] Breathe out. [25:04] >> Maybe you should license the voice to [25:05] come. [25:05] >> I I mean that would be interesting. [25:08] Let's talk a little bit about [25:10] um [25:12] being up against [25:14] some of the greatest entrepreneurs ever [25:17] who want to take your business from you. [25:19] Specifically, Daario and Anthropic, Sam [25:22] from OpenAI. They they want your [25:24] business. They've been pretty clear [25:25] about it. Um, and I think you have used [25:29] the frontier models in your product, but [25:32] you must be thinking, my lord, am I [25:36] enabling my own demise by partnering [25:38] with them? And there's all these open- [25:40] source models. So, um, how do you think [25:43] about your partnerships with those type [25:46] of frontier models and the fact that [25:49] they want to kill your company? [25:52] So on on the on the first part the given [25:56] we create a platform we try to provide [25:58] all alms out there. So our customers can [26:00] pick entropic open AI open source uh [26:04] Google models um and that agnostic being [26:06] agnostic to the specific model is [26:08] actually helpful because customers can [26:10] they make sure that they build a harness [26:12] build the agent orchestration create the [26:14] voice element of how that agent [26:16] interacts with the world how the [26:18] marketing interacts with the world but [26:19] they are not dependent on any model. So [26:20] for us that part is is is um is a is [26:24] actually good because we can provide [26:26] that to the customers. On the on the [26:28] kind of the second big part of like of [26:29] course the the space is overlapping [26:32] increasingly [26:34] models our platform platform are [26:35] application everything is becoming a [26:37] little bit more more fuzzy for us. The [26:39] still the defining piece was focusing on [26:41] that one layer of like how does [26:43] interaction look like? How does [26:44] communication look like? And we've been [26:46] able to out compete them on voice models [26:48] um both on text to speech, speech to [26:50] text, on the turn taking on music and um [26:53] and we've you know here our research [26:55] team is is is is a a set of magicians [26:58] that are able to continuously do it time [27:01] and time again. Um and I think part of [27:04] the reason is it's on the research side. [27:06] It's the architecture that matters, not [27:08] the scale. You really need to change how [27:10] the model operates. Two, you need very [27:12] specific data that there's of course a [27:15] wide set of data out there, but it's [27:17] unlabelled data and where we spend a lot [27:19] of time. So we build a internal team of [27:20] over thousand contractors that label all [27:22] those audio assets to make them to make [27:24] them good. So that's on the research [27:25] side. And then as you think about the [27:27] rest of product stack, we want to create [27:28] a fully verticalized solution for that [27:30] communication angle. The product [27:32] understanding the right workflow in [27:33] financial services is very different to [27:35] healthcare, very different to telos. We [27:38] spend all of our product team to figure [27:40] out how that works and those companies [27:42] don't. And then ultimately last piece is [27:44] the ecosystem. Can you build the wider [27:46] set of integrations voices that you use [27:48] templates for the agent authentication [27:50] that you can benefit from instead of [27:52] starting from scratch? And so far we've [27:54] been we've been able to create a new [27:55] model for that. [27:56] >> Certainly though you must be concerned [27:59] about hey the reinforcement learning the [28:02] data leakage. They say they're not using [28:04] your data, but they're kind of using [28:06] your data. And so, do you have an [28:09] open-source project internally as the [28:12] like in case of Glass, we got to break [28:15] this? And when do you think [28:17] >> you'll be able to discontin working with [28:21] them if you had to? [28:23] >> We we we we [28:26] know that some companies are [28:28] continuously trying to figure out how to [28:30] distill and use the data. So that is uh [28:32] that is a existing problem and we have [28:34] few mechanism to to to stop it. Um or [28:37] slow it down not stop it. Um but um but [28:41] on the open source question or like [28:44] creating our own um um uh versions we [28:47] are we are looking a little bit closer [28:48] on like how we could use our expertise [28:50] of [28:52] how does like you know we won't focus on [28:54] knowledge work. we won't focus on coding [28:55] but any interaction and how you can [28:58] combine all those pieces together and [29:00] make sure this is this is great. Yeah, [29:01] we want to own. So we are spending more [29:03] time there. Uh but it's also just great [29:06] to be in the arena and compete with [29:07] those guys and uh and and every so often [29:10] show that we can do it and do it better. [29:12] Yeah, it's pretty clear in my estimation [29:16] that that's where you you will wind up [29:18] and the ability to make your own [29:21] language model today, especially with [29:23] all these great models out there that [29:25] are now open- sourced. It's going to be [29:27] a pretty easy for a company with your [29:29] level of resources. So, why wouldn't you [29:32] at least offering it as an option? And [29:34] then I guess there's cost. I mean, you [29:35] must be shipping [29:38] tens of millions of dollars to the [29:39] Frontier models every year. ship good [29:42] amount. Um we are good partners good [29:45] partners with them. Uh um but but it's [29:48] um it's ultimately you know showing up [29:52] in the value we can create too. So like [29:54] a lot of what we spoke at the beginning [29:55] of how we can elevate ourselves as [29:58] organization too is is definitely [29:59] helpful. So I think they they done [30:01] tremendous work on on building uh it's [30:03] it's almost crazy that each of us has [30:05] like a a a Turing uh like you know if [30:08] you if you were to chat with an agent [30:11] now it feels like um the Turing test [30:15] will be complete it's as smart as [30:16] another human. Uh and we hope this year [30:18] we'll do that same thing for voice uh [30:20] where any conversation feels like you [30:22] are speaking with another human. It will [30:23] be [30:24] >> yeah I think you're there. It just [30:25] depends on the application and like what [30:28] question you ask. But it definitely [30:30] passes it. I mean it if we were to look [30:33] at the tests that were created to define [30:35] artificial general intelligence or just [30:37] to define artificial intelligence, we [30:40] passed all of those. These were tests [30:41] that were created 30 or 40 years ago. We [30:43] need a new set of tests right now. I [30:45] think the new test is like can this be [30:48] more intelligent than every single [30:51] person on the planet times 10. And if we [30:54] get anything less than that, we we're [30:55] kind of like, oh yeah, it's not smart. I [30:57] mean, these things, we're kind of there [30:59] on AGI, don't you think that we've kind [31:01] of achieved it, we just haven't deployed [31:03] it. I [31:03] >> I am I there are definitely places where [31:06] where we did achieve it. [31:07] >> Yeah, for sure. All right. Continued [31:09] success. Let's give it up for Mati from [31:11] 11 Labs. Well done. Thanks for coming [31:13] out. I'm going all in. [31:17] >> The AI companies building the future run [31:19] on Oracle cloud infrastructure, training [31:21] and deploying at scale on one of the [31:23] world's largest AI infrastructures. The [31:26] same Oracle AI platform gives [31:27] enterprises access to leading models AI [31:30] grounded in their own data and the [31:32] security to move from pilot to [31:34] production. Learn more at oracle.com/ai [31:37] or experience it live at oracle [31:40] experience live. I'm going all in. [31:44] >> You're growing also at a very [31:46] significant clip. [31:47] >> Exponentially. [31:49] >> Is it exponential? No, it's not [31:50] exponential. [31:51] >> Oh, it's a 50 sustained 50% quarter over [31:54] quarter for the last seven quarters. [31:56] >> 50% quarter over quarter. Last seven [31:58] quarters. Yeah, that's that's pretty [32:00] darn fast. So I think we actually just [32:02] became as of the close last week on [32:04] Tuesday one of the fastest enterprise [32:06] company with a direct sales motion to go [32:08] from one to 100 150 beating Sierra with [32:12] one quarter. [32:14] >> Amazing. Um and so um people I mean [32:18] there's a couple of things in life that [32:20] people really hate and paying lawyers is [32:23] like way up on the top of the list. [32:26] um with your tools obviously you got [32:29] your contemporary and Harvey and people [32:33] >> it's just a it's a small company um in [32:35] the in the states um and then you also [32:37] have I guess Claude and other folks also [32:40] want to be in your business so this is a [32:43] big prize um to [32:46] take I don't know 80% of what we pay [32:49] lawyers for and compress it by 90% like [32:53] what what is the realistic [32:56] power law here in terms of [33:00] making for startups in the audience, [33:02] your legal bills dramatically drop in [33:05] costs. Yeah. And I'm seeing it already [33:08] in the startup space. I had an I had one [33:10] firm, one startup that hit a million in [33:13] revenue. [33:14] >> Yeah. [33:14] >> They had closed multiple rounds of [33:16] funding. Um, multiple obviously large [33:19] number of employees, a decent couple [33:20] dozen employees. They didn't have a [33:22] corporate lawyer. [33:23] >> No. And I said, "Whoa, whoa, whoa, whoa, [33:25] whoa, whoa. You think I had a million [33:26] dollars in revenue? Like, somebody [33:27] should review the contracts?" And [33:29] they're like, "Chat CPT, bro." And I'm [33:32] like, "What about the cap table?" [33:33] They're like, "Chp, bro." And I was [33:35] like, "Okay." And HR. And they're like, [33:38] "Same thing, bro." And I'm like, [33:40] >> "Okay, [33:41] >> it's going to be a fun diligence target [33:42] one day." [33:43] >> Well, that's what I said. I said, "Hey, [33:45] you know, when you do the series A, [33:48] they're going to ask that like some of [33:49] this stuff be reviewed. like, "Do you [33:51] guys have like IP assignments?" They're [33:54] like, "Yeah." I'm like, "How did you [33:55] know to do IP assignments?" First- time [33:56] founders are like, "We asked chat GPT." [33:58] And I'm like, "Okay, wow. I just turned [33:59] into unk like I guess." Yeah. So, so [34:03] take us through what you you think is [34:05] happening out there. This is not [34:07] uncommon, right? What I scenario [34:09] >> but but a a seedstage startup operates [34:12] very differently from, you know, one of [34:13] the biggest banks in the US. And so the [34:17] way to think about the market or at [34:18] least the way that we like to is you [34:20] have this enormous bucket of legal [34:22] services which today is being done [34:24] manually. It's a trillion dollars every [34:27] year into legal services which is very [34:29] fragmented. [34:30] >> But the software spend into legal [34:33] technology is about 40 billion. So it [34:35] means there's 4% software 96% service [34:39] which is bananas. The software piece [34:41] should be much bigger than that. And so [34:43] the software piece naturally will grow [34:46] into um the service revenue but also [34:50] legal is a very supply constrained [34:53] market. The demand for legal services is [34:57] much larger than what there are lawyers [35:00] or legal services available. And so many [35:03] of the legal service providers are now [35:06] using technology to serve new use cases, [35:10] new market segments and to actually [35:12] package new products and you will not [35:14] make [35:15] >> what's an example of that like [35:16] >> so so an example of that is Kulie [35:19] actually they started serving startup [35:21] founders directly with a sort of [35:25] software uh platform that you just log [35:27] onto the platform they pumped it full [35:29] with their material and their precedent [35:31] and then you have the startup material [35:33] there and they've embedded workflows [35:35] that reviews the contracts. [35:37] >> Um, and what I think is interesting by [35:39] that is it it starts to um break this [35:43] model where you charge out associates [35:46] for very high hourly rates and you have [35:50] a billable hour model. And actually, if [35:52] you look in law firms, the way that that [35:54] business model works is you overcharge [35:57] for the associates and you actually [35:59] undercharge for the partners. I don't [36:01] know if they're under charge. I mean, I [36:02] got a bill recently and it was 1,800 an [36:04] hour, right? But [36:06] >> for a senior person, I think the [36:07] associates were 800. [36:08] >> Well, you know, at Kirkland, it can go [36:09] up to 4,000 an hour. But the but the [36:11] thing is when a when a Kirkland if so, [36:14] let's say, you know, 30 minutes of a [36:17] Kirkland partner's time when it really [36:19] matters can be worth a lot more than [36:21] that. [36:22] >> Like a lot more than that. If it's bet [36:24] the company litigation or you avoid a [36:27] pitfall that would have costed the [36:28] company tens of millions of dollars, [36:30] >> well worth it. Yeah. [36:31] >> Right. Exactly. And but the only way [36:32] they know how to price that is to [36:34] overcharge for the associates. [36:36] >> But as you're saying, the enterprises [36:38] are looking at this and they're going, [36:40] "Huh, we're spending a lot of dollars on [36:42] legal services. Let's take this [36:43] in-house." And [36:44] >> Oh, really? [36:45] >> Absolutely. I mean, we're doing this [36:47] partly at Lora. We acquired four [36:49] businesses so far this year. We did the [36:51] diligence inhouse the with our own tool [36:55] and the fastest transaction we did was [36:57] 12 days from LOI to closing [37:01] >> because your motivation as the founder [37:03] is to get the deal done, [37:04] >> right? [37:05] >> The motivation of the lawyer is to not [37:07] have you sue them if they up the deal, [37:10] >> right? [37:11] >> And to make as much money as possible, [37:14] >> which means to drag it out, [37:15] >> which means their incentive is to even [37:17] if they don't say it explicitly, it is [37:19] to drag it out. your incentive is to [37:21] close it as quick as possible. Yeah. [37:22] >> Yeah. And so, you know, I think a lot of [37:24] law firms are also experimenting with [37:26] different pricing models where you do a [37:28] fixed fee for a transaction or for a [37:30] fund raise. Um, in litigation, you can [37:32] take a part of the success fee when you [37:35] win the deal. [37:36] >> Yeah. [37:37] >> Uh or win the case. And so, I think it's [37:39] just very interesting how, you know, one [37:41] of the biggest industries in the world [37:43] now is being completely transformed and [37:45] reshapen as a consequence of the tech. [37:47] And are those law firms feeling like [37:50] they're being disrupted or this is a [37:52] huge opportunity [37:54] and and and did that switch at a certain [37:56] point in time or has it switched for [37:57] them? [37:58] >> Um there's a lot of anxiety and a lot of [38:02] fear and you know these law firms are [38:07] enormously profitable and big [38:10] businesses. Kirkland Ellis turns around [38:12] $10 billion a year. [38:14] >> How many lawyers did it have? [38:15] >> Four or 5,000. Wow. [38:17] >> I mean, per partner, they make it [38:18] between 5 and 10 million every year in [38:21] profits. And so, when something like AI [38:25] comes along, that poses um existential [38:29] threats and existential opportunity. And [38:32] I that's actually a big part of my job [38:34] to help articulate with the leadership [38:37] teams that we work with because we will [38:39] only be as successful as our customers [38:41] are. And so we actually have a very [38:43] unique role at Legora as well which is [38:46] called the legal engineer. So in the [38:48] same way that Palanteer has forward [38:50] deployed engineers, we have forward [38:52] deployed lawyers and their job is to sit [38:55] down with the Kirkland partners and help [38:57] them transform their business [38:58] >> from a preAI to a post AI world. A and [39:03] it's sort of like document management [39:05] and PCs were but 20 or 30 years ago when [39:09] they were printing out and keeping [39:11] drafts in a in a library in in a storage [39:14] facility and they had to sort of walk [39:16] them through and handhold that. [39:18] >> Absolutely. But I think the difference [39:20] >> the difference is those were you know [39:23] mild productivity gains. [39:25] >> Yeah. this can do a lot of the work and [39:30] so it's really reshaping what it also [39:32] means to be a junior lawyer going into [39:35] this occupation. [39:36] >> What does it mean? Are those jobs going [39:38] to still exist or a lot of the lawyers [39:40] who are coming out of school going, "Oh [39:42] my god, was this a good idea or a bad [39:44] idea?" [39:45] >> The job will exist. The tasks will be [39:48] different, [39:49] >> right? Um, in order to have a [39:52] partner-driven model, you need to bring [39:54] people up the ranks, right? In the same [39:57] way as you do with software engineers. [39:59] You need to have junior engineers so [40:00] that one day you can have senior [40:02] engineers who know what they're doing. [40:03] Um, but the way to get there is very [40:06] different. The way of getting there [40:07] today will not be lock yourself in the [40:10] physical data room, read through every [40:13] single document, mark the errors, and [40:17] you know, go fax it, right? It's and [40:19] it's also no longer just look in the [40:21] virtual data room and control F. It's [40:23] orchestrating the agent that will be [40:25] doing that work. [40:27] >> And and when you look at that work, you [40:31] you have a global backdrop [40:34] attorneys obviously very famously [40:36] localized, right? And is this going to [40:39] create attorneys who can operate across [40:43] borders in a way that didn't exist? And [40:45] you're starting to see that and is that [40:46] something that's built into the product? [40:48] So when you're doing even in the United [40:50] States, it's a it's state level [40:51] certification obviously. Um and doing a [40:55] non-compete in the Northeast is very [40:56] different than doing it in California. [40:59] >> They're not very enforceable or [41:01] enforceable at all in California as [41:03] people know, but they're quite [41:04] enforceable if you're in Boston. [41:06] >> Yeah. [41:07] >> So, so talk about that. Uh because that [41:09] seems to be a place where there could be [41:11] massive gains from AI. [41:12] >> 100%. And it it's really two things. I [41:15] mean the data that Legora sits on top of [41:18] is on one hand side the firms and [41:20] enterprises own data their precedent [41:23] their organizational data and secondly [41:26] we do the hard work of gathering all the [41:29] cases all the legislation all the [41:31] regulatory updates for every [41:33] jurisdiction in the world and that is [41:35] very painful but once you start to do [41:37] that at scale it builds a real data mo [41:41] >> and so in the system if you are the GC [41:44] of a company in California and you just [41:47] landed your first customer in South [41:49] Africa, right? Legora can be adapted to [41:52] the local legislation in South Africa. [41:55] And we actually had a case of this [41:56] where, you know, instead of having to [41:59] call a lawyer who then knows a lawyer in [42:01] that region who will respond to the [42:03] query, they can get an 80% accurate [42:06] response immediately that they can start [42:08] working off out of. And the better that [42:11] gets um the more um the more interesting [42:16] things I I I believe you can do because [42:18] this data has really never been [42:20] structured before and there are so many [42:23] people who are working with setting [42:24] policy and building regulation and and [42:27] this is a enormous inefficiency in [42:30] society [42:31] >> and Lexus Nexus has been a juggernaut [42:35] and the legacy player in you know all [42:38] the case law and regulations. They have [42:41] a massive data moat. They but they only [42:45] make a couple of billion dollars a year. [42:46] And if you put your revenue and Harvey's [42:49] revenue together, you guys are probably [42:51] already just that the two of you uh [42:54] you're both making hundreds of millions [42:56] of dollars. So, they must be looking in [42:59] the rearview mirror at you like the [43:01] Tyrannosaurus Rex in Jurassic Park and [43:03] going, "Holy shit." Like, are they [43:06] coming for our business? And then here [43:08] you are on stage saying, "Hey, we're [43:09] doing all the manual hard work of [43:12] getting that information into our what I [43:14] assume is a proprietary language model." [43:16] We'll get to that in a second. Um, are [43:19] you going to just try and buy Lexus [43:21] Nexus? I know it's part of a larger [43:22] enterprise, or are you just going to [43:24] kill it? [43:25] Well, I think that some of the existing [43:29] uh providers and the sort of legacy [43:31] players have a really hard time pivoting [43:35] into becoming AI native businesses. [43:37] >> Sure. [43:37] >> And they have a really hard time meeting [43:40] and and catching up to the tempo that we [43:43] run at. They can't get the talent. They [43:45] don't work our hours. And they're so [43:49] political in their organizations that [43:51] it's just hard to move. Um, I think at [43:54] the outset of AI, many believed and made [43:57] a bet that those organizations who had [43:59] all the data was going to be the [44:00] winners. As we're starting to see in the [44:03] market, that's no longer the case. I [44:05] think there's a real opportunity for us [44:07] to partner with content providers. And [44:10] we're already doing this in many of the [44:13] smaller jurisdictions like in Germany, [44:15] in France, in Spain. The US is is [44:18] peculiar because it's such a duopoly on [44:21] legal research as a West Law is the [44:23] other one. [44:24] >> West law and Lexus Nexus. Exactly. But [44:26] um yeah, if you look at how their stock [44:28] is doing, um I think [44:31] >> or are they getting priced in with the [44:33] AI uh certainty? [44:35] >> Yeah. [44:36] >> Yeah, that's one way of putting it. [44:37] >> Yeah, they're getting crushed. Uh and I [44:39] I would assume there's some power law [44:41] here. you know, they might have an [44:44] incredible breath of, you know, old case [44:47] law that they scanned in and went to the [44:50] courouses and did all that work on, sent [44:52] to India to be double blind typed in. [44:54] Like they literally [44:55] >> You're right. That's what you have to [44:56] do. [44:57] >> Yeah. They literally had two different [44:59] people type in the cases or OCR them, [45:01] then check them, look for the [45:03] differences. I mean, cuz you can't get [45:04] it wrong. [45:05] >> Nope. [45:06] >> But today with the AI tools, the AI [45:08] tools are really good at doing what they [45:10] did manually. Yes. Um, you still have to [45:13] ship the books because you have to [45:15] physically scan. This is very strange in [45:17] the US, but uh, West Law basically has a [45:20] monopoly with the American government to [45:23] report on the cases. So, they're not [45:26] owned by the public in a way. They're [45:28] owned by [45:29] >> That's crazy. [45:29] >> The company. You guys are very good at [45:31] capitalism. [45:33] >> Sometimes too good. But sometimes too [45:34] good. But I mean, Harvard has a project. [45:37] There's the court law court listener. [45:39] >> They're trying. They're trying. [45:40] >> Yeah, it's it doesn't work. Um or rather [45:43] put it this way, you cannot build a [45:47] legal research solution that doesn't [45:49] have all of the data. [45:52] >> Because if you go to WCTEL and a [45:54] litigator at WCLE, the best law firm in [45:56] the world, says, "I'm going to use this [45:58] to to, you know, go after Elon or or do [46:01] a billion dollar case." You better make [46:04] sure you have all the cases. [46:05] >> So So it's the opposite of the power [46:08] law. You don't just need the top 80%. [46:10] You actually need all of it. [46:11] >> All of it. [46:12] >> Which means you have to go to courouses [46:14] and ask them for a copy to print it out [46:16] and pay them 10 cents a page. [46:18] >> Well, there's other ways of getting it, [46:20] but in in practice, yes. Um, you have to [46:24] physically get the books all the way to [46:26] India. You need to open them. You need [46:28] to scan them cuz you need to get what's [46:30] called page citations. I never thought [46:33] in in college I would get this nerdy [46:36] about legal data, but here we are. And [46:39] what's interesting is that these [46:42] previous generation of databases were [46:45] very much search in the database, find [46:47] the case, and then the lawyer, you know, [46:49] does their work, [46:50] >> right? [46:51] >> What's really interesting about [46:53] especially the agents following the [46:54] release of Opus 4.5 and 4.6 Six is they [46:57] can now start to do really intelligent [47:01] case strategy and they can actually [47:03] start to combine the witness statements, [47:07] the cases and they can really do end to [47:10] end work which is I think moving us from [47:14] a world where AI is just augmenting [47:17] to AI is actually really doing things [47:20] and your job becomes to orchestrate and [47:23] to manage those agents as we're seeing [47:25] in coding. [47:26] And so you have partnerships with I'm [47:30] assuming anthropic and open AI. Yes. To [47:33] and you spend millions or tens of [47:36] millions of dollars on tokens. [47:37] >> Absolutely. [47:38] >> And they are also competing with you on [47:41] the margins. [47:42] >> Um they are not competing in our product [47:45] category at all at at all. Um from [47:48] >> for now the well you know from the [47:51] outside uh you know Claude has a legal [47:54] offering. Yeah. which is basically a [47:56] bundling of markdown skills files and a [47:59] couple of integrations. [48:01] >> And so I think what's really helpful [48:03] about that is that it illustrates to [48:05] everyone how applicable AI is in law. [48:10] >> Um what it also does is it drives a lot [48:12] of initial usage there and then you hit [48:14] the ceiling or you know you understand [48:17] how shallow it is and then you call us [48:19] right [48:19] >> and so it's actually a big pipeline [48:21] generator uh for us. [48:22] >> Got it. So they start experimenting. [48:23] Boom. And we were just talking with the [48:27] CEO of um 11 Labs about hey building [48:31] your own models is you know pretty um uh [48:36] pretty doable these days and every 6 [48:39] months it gets easier and easier. So are [48:42] you working on your own models using [48:45] open source to then fork it and and make [48:47] your own models? Is that the future for [48:48] your firm? So um I don't believe in [48:51] fine-tuning or building any general [48:54] intelligence models. I think that's uh [48:57] total waste of time and money. I do [48:59] believe in very narrow models for narrow [49:02] use cases that you also drive a lot of [49:05] scaling. So you can drive both cost and [49:07] latency down. An example of this for us [49:09] is we have a big feature called tabular [49:12] review which is basically the number of [49:14] documents times the number of prompts. [49:16] So 100 documents, 100 prompts, 10,000 [49:19] API calls. [49:21] >> If you make a fine-tuned model at [49:23] extracting contract data, it's very [49:25] applicable there. But it doesn't make [49:27] sense to build a general legal [49:30] intelligence model like some of our [49:33] competitors are attempting. [49:35] >> Yeah. Um, and how do you mitigate [49:39] against the data leakage issue with your [49:43] customers? these, you know, are highly [49:46] regulated industries with a lot at [49:48] stake. [49:50] So, putting in, you know, this recent [49:52] case you're working on in a litigation, [49:55] if any of that were to seep into a [49:58] language model and then come out the [49:59] other end, I mean, you this is [50:01] disastrous. You have a higher level of [50:04] >> responsibility [50:04] >> and uh compliance is our currency. And [50:08] so um it's actually one of the reasons [50:12] why it's really hard to sell into law. [50:15] There's a lot of legal AI companies and [50:19] very few are making it through [50:21] >> and not because it's hard to build [50:23] stuff. It's actually quite easy to [50:25] understand where you can build value, [50:27] but getting it to the customer is very [50:29] hard. But that's something we cracked [50:31] pretty early on. And [50:33] >> once you're in, it's much easier to [50:35] expand. So that's al also one of the [50:37] driving forces behind our M&A strategy. [50:40] But yeah, I mean we're hosting national [50:42] secrets weapons manufacturers with their [50:45] contracts on Lora and we work with [50:47] governments and [50:48] >> does that mean you have to put it on [50:50] prem as well or [50:51] >> we don't do on prem. Um I [50:53] >> that's on the road map or [50:55] >> no I mean you know deploying in a VPC is [50:59] very [51:01] time consuming and it creates a lot of [51:03] dependencies which slow down your road [51:06] map and the execution forward. [51:08] >> All right continuous success max thanks [51:10] for taking some time for us. [51:27] I'm going all in.