看不见数据也要造 Agent：Legora CTO 谈法律 AI 的评测、原语与计时收费裂缝

信源：YouTube 视频 bJBEFzlKPoQ（Latent Space 频道，2026-11-02 发布，时长 4042 秒 ≈ 67 分钟；Latent Space 官方博客标题为《Building AI for Lawyers - Jacob Lauritzen, CTO of Legora》）。本文依据英文 ASR 转写整理，转写共 34 个 [Ns - Ms] 时间段、完整覆盖 0s→4042s，两遍通读无遗漏。嘉宾身份已核实：Jacob Lauritzen，Legora CTO——经 Latent Space 官方节目页、Slush 2026 演讲者页与 Dealroom 三方交叉确认：2024 年以工程负责人身份加入，2025 年 10 月升任 CTO，此前联合创办 Firmnav、创办 YC 公司 Unhaze（YC23 届）。主持人在节目中自述「Cognition（Devin 母公司）顾问」并系统阐述 agent lab 论述，与节目页 show notes 中「Swyx on the agent lab playbook he brings up at the start」互证，推断为 Swyx（Shawn Wang，Latent Space 主编；推断级）；转写中另有第二位主播少量插话，身份无法确认，按「主持人」泛称处理。节目页另载官方口径背景：Legora 于 2026 年 3 月融资 5.5 亿美元（估值 55.5 亿）、10 月再融 5.5 亿美元 E 轮（估值 88 亿），4 月 ARR 破 1 亿美元（上线不足 18 个月）、客户超 1800 家，年内已完成五笔收购（Wexler 为第五笔）——此类数字属官方披露，非播客内陈述。归属口径：播客内测试分数、公司数据一律标「嘉宾认为/节目称」，主持人观点标「主持人」，写作者判断标「写作者推断」，分歧并列不强合。ASR 疑错统一汇总于文末校正表，确认项已在附录中修正、存疑项保留原文。同域旧文（背景不复述，仅互链）：原告端法律 AI 与计时收费→结果定价的机制见《a16z 对谈 Eve：美国法律界的另一半》；律所数据资产化与 Crosby 意向收购 Weil 一案见《AI Daily Brief：AI-native playbook》。本文只写本场增量。

先说结论

这场访谈的真正主题不是「AI 颠覆法律」，而是一个工程悖论：法律是最依赖保密数据的行业，而 Legora 作为这个行业的 Agent 平台，从头到尾看不到任何一条真实用户 trace——嘉宾原话，这是「这生意一个超级烦人的怪癖」，因为律所的数据「甚至不是律所的，是律所客户的」。他们的答案是绕开数据本身：合成「评测世界」（一万份文件的数据室尽调）、把失败归类成十几种模式分桶、用超廉价 LLM 跑在线评测，用这套替代体系去逼近看不见的真实用法。产品端最重要的界面决策是「表格」：Tabular Review 单表最高做到百万格，把 fan-out 检索变成律师本来就懂的二维审阅动作，而 Lists 这个「Notion database 式原语」让「agent 造列表、列表造 agent」成为组织一切工作的粘合剂。商业模式端，嘉宾认为计时收费已在低端出现裂缝（项目费、甚至免费做关系活），钱向并购、诉讼和客户关系集中；更结构性的判断是合伙人制的深层矛盾——人力资本无法沉淀为公司资产，他想让律所把「本所做法」codify 成可积累的 IP。对做垂直 Agent 的团队，本场最可迁移的三件事：按业务领域×任务追踪模型热力图（回退比想象常见）、fan-out 前先跑前十行迭代、以及在 DPA 供应商极简的约束下做模型套利。

一、从「ChatGPT + 瑞典法条」到 Agent 操作系统

Legora 的起点被嘉宾一句话概括：「带瑞典法条搜索工具的 ChatGPT，字面意义上这就是全部」。2024 年这家公司还叫 Leya（嘉宾当年帮 CEO Max Junestrand 改 YC 申请——两家随后先后进了 YC：嘉宾自己的公司是 23 届，Legora 是 24 届），产品从法律研究（律师花大量时间找「长得像本案的先例」）起家，逐步长成两层：一层是做工作——起草、改 Word 文件、整理归档、跑复杂分析；另一层是编排工作——lists（他自己的类比是「我们的 Linear」）、intake（与客户协作）、portal（律师↔律师、客户↔律师、律师↔agent、agent↔agent）。产品叙事里一个值得注意的细节：Tabular View 这个由 Legora 推广的术语已经溢出成行业通用词，嘉宾明确表示乐见 Anthropic 的 Claude for Legal「用我们的术语」——命名权即品类定义权，这是垂直 Agent 公司少被讨论的一种护城河形态。

从 assistant 到 agent 的切换发生在约一年前，嘉宾将其归因于推理模型「改变了一切」。今天的形态是「真的很好用的律师版 ChatGPT」，加上侧栏里一组定制工具。规模感（节目页官方口径 + 嘉宾口述交叉）：100 名工程师（他曾在入职半年的幻灯片里放《300 勇士》宣称「永远 20 个工程师打天下」，现在自嘲 100 人还嫌小）、纽约 100 人并在继续扩张、工程招聘集中在纽约/斯德哥尔摩/伦敦。文化上他承认作为斯德哥尔摩地下室里的 underdog，「谦逊是底色」，进美国后不得不学会 bold。

二、看不见 trace 的评测学：合成世界、失败分桶与热力图

本场信息密度最高的部分是评测体系，因为它直接回应标题里的矛盾——「造 Agent 却看不见数据」。约束是三重的：律所数据高度保密；很多数据所有权根本在律所的客户手里；Legora 与客户之间也不存在「可以看数据」的协议（主持人猜「你们总签了什么协议吧」，嘉宾答「我倒希望」）。于是连「匿名化改写」都要划一条苛刻的线：改写到只剩三个候选值再任选其一，可以；仅仅去掉人名，「那可能不行」——他们的原则是主动把自己限制到最安全的一侧。能做的是结构化信号：已知失败模式大概十到二十种，把线上案例分桶归类，再叠加任务类型与业务领域分类，就得到了一个不需要看内容也能优化的坐标系。

在这个约束下，评测团队（三位工程师 + 三位律师混编成一个小队，另有约百人的 legal engineer 池——「大所干过三四年、通常有 CS 学位的律师」，他称之为最难招但力量倍增率最高的人）设计了「世界」：一万个文件的模拟数据室，目标是产出一份包含全部红旗、措辞得当的尽调报告，任务再拆成「起草」与「找红旗」两层。所有模型按业务领域×任务跑出一张大热力图——嘉宾给出的反直觉观察是：模型发布的「部分回退」远比外界想象的常见，经常是这几项涨了那几项跌了。这也解释了评测团队的历史时机问题：直到不久前做复杂评测都不划算，因为模型本身在狂飙——他曾有一个引用格式问题，什么都没做，下一次模型发布把指标从 70% 干到 99%。真正持续的工作是三件事：harness、治理（审计日志、数据保留），以及在线评测——用「超廉价 LLM 消息」判断每次会话是否解决了问题，配合 feature flag 对比新功能的「改善还是灾难性下滑」。常规产品指标则漂亮得反常：对一款企业软件，DAU/WAU/MAU 与每月使用小时数都「极高」。

与「看不见数据」配对的是幻觉对策。Legora 有一套引用系统：agent 产出的每句话都带可点击链接，定位到源文档 PDF 里的 bounding box 高亮，而且可以多层下钻——agent 基于 Tabular Review 写出的结论能一路引到那张表的具体单元格，「从底往上全部可验证」。律师的使用行为则像 code review 的两极：有人根本不点（嘉宾复述主持人圈的观察：有人对幻觉出的最高法院链接毫不在意），有人逐个逗号地读；专业人士真正的行为模式是「知道哪几处必须查」，如同审一个初级律师的产出。

三、界面即护城河：百万格表格、Lists 原语与 agent 群

Tabular Review 的交互是本场的具体锚点：一万份雇佣合同做尽调，每一行是一份合同，每一列是一个问题（终止通知期多长？有没有 IP 转让条款？解约赔偿多少？），整表自动填充、全部带溯源引用、可标记「已审/有旗」、可评论协作。更关键的是 agent 可以自行编排它：往 agent 里丢一万个文件说「做尽调」，它会自己拉起 Tabular Review、跑完、分析结果、标出红旗（「CEO 薪酬高得离谱」「这个奖金结构很怪」），再据此给客户起草备忘录。单表规模最高到过百万格。工程上的安全阀同样朴素：Tabular Review 默认先跑前十行，让用户确认「显示的是不是我们以为会显示的东西」，快速迭代到结果对了「再放开跑」——这是对所有做 fan-out 检索的团队都成立的渐进式放量原则。

为什么是表格？嘉宾的答案毫不玄学：总览最快、天然带引用、最好审阅，「比我们试过的任何其他形态都好用」，用户甚至不愿意导出到 Excel（前端用的是做了十几年表格的 AG Grid）。主持人补充了形态学的观察：表格形态在 GTM（Clay）证明过拉力，但奇怪地没有扩散——V7 试过，金融领域也没有头部玩家。写作者推断：fan-out + 同构抽取 + 人工复核的场景，二维表格可能是最短路径，这个形态在更多垂直行业仍有空位。

Lists 则是编排层原语：像 Notion database 一样可配置、带触发器和自动化，是「粘合一切」的胶水——列表可以触发 agent，agent 回写列表字段，「agent 造列表、列表造 agent、agent 读列表，agent 群就活了起来」（他用「用 Linear 协调一群 coding agent」作类比）。intake 是它的店面化应用：一个法律专用 Slack bot，把请求变成列表工单，法务团队坐在 Legora 里管理 agent，办完自动回 Slack——「没人需要换工具」。对内，这套原语还兼职低代码平台：legal engineer 能基于 lists 加生成式 UI 快速搭出定制产品（他举了基金管理仪表盘），「你要自己的？很简单」——主持人概括为「Legora 里的 Lovable」。年内第五笔收购 Wexler（英国，fact intelligence）也挂在这条思路上：百万份诉讼文件抽取成带实体和日期的事实，再互相比对出「佐证 vs 矛盾」——转写一场庭审后，「他说他不在场，但这封邮件说明他在」当场对质。嘉宾把公司级战略说得很直白：客户要感觉你建了很多东西，实际上你只建了很少的原语。

组织上与产品咬合的是双层交付：第一线 legal engineer 以超级用户身份配置产品、不写代码、驻场教律所把工作流结构化；第二线 forward deployed engineer（FDE）写进生产代码、做集成——纽约的第一名员工就是一名 FDE（前 YC 创始人，嘉宾以「两届 YC、做过 Meta 搜索」的高门槛标准招的她）。FDE 同样不碰客户数据，只建系统；API 面则刻意收紧：他们曾长期拒绝开放 API，因为产品迭代太快、不想维护旧版本对已变数据模型的兼容——20 个工程师时「做不到」，100 人时才付得起这个税。嘉宾最后抛出的开放问题也值得记录：「agent 能不能本身就是 API」——用触发 agent 干活替代确定性的 API 调用，代价是效率与确定性（他的估计：约百分之一二的概率干出计划外的事）。

四、模型策略：Pareto 前沿套利与 harness 里的低垂果实

模型选择器的存在本身就是一段妥协史。主持人（以 Cognition 顾问身份）给出 agent lab 论述：垂直 Agent 公司本质是「领域专用 GPT wrapper」，职责是替用户守住模型 Pareto 前沿——「你去问律师：这合同该用 Sonnet-high 还是 Opus-mid？他不知道。我们的时薪是一千美元，让它能跑就行」。嘉宾完全同意这个论点，但现实是用户要尝鲜：他们抗拒了很久，最终还是给了五档模型选择器（含 Fable、Opus、Astra），明知多数法律工作用不上 Mythos 级智能、明知尝鲜者积分烧得快十倍。能力分布上他给出一个粗刻度：约六成工作现有模型已经做得很好，剩下的博弈是成本与延迟；真正需要更强智能的是复杂分析，而且法律工作总量本身会随智能扩张。

选型逻辑被主持人称作 model arbitrage：在能力、延迟、成本、治理、区域可用性构成的空间里逐任务优化——某些开源权重模型适合特定任务就部署，某些任务后训练（直到最近开源变强才开始做，且有严格边界：只做「有界任务」，比如法律研究这种 fan-out 检索，可训出快、便宜、低延迟的专用模型；「训练自己的 Fable 当主 agent？对 Anthropic 没胜算」），其余用前沿模型。约束侧的故事很能说明问题：全栈在 Azure，不是技术选型而是信任继承——三年前律所对 AWS 和 Google 说「绝无可能」，却因 365 已通过安全审查而接受 Microsoft；on-prem 的讨论三年内彻底消失。DPA（数据处理协议）上的供应商清单「非常少」：不用 ElevenLabs，一站式厂商（OpenAI/Google/Microsoft）「不必样样最强、样样够用加治理」就有真实溢价。转写链条上的细节也值得抄：先用 LLM 层抽取关键词、再带着关键词重跑转写模型，对抗领域专名（他举的翻车例子：澳大利亚法庭某称谓被稳定转写成「condom」；主持人的例子：Whisper 和 ElevenLabs 至今把 Claude Code 转成 Cloud Code）。

在 harness 一侧，他排出的优先级是 search、analysis、drafting 三件定制：大合同起草本质是判例检索（「similar」是多维的：同行业、同类型公司、同交易结构），从多份先例文档决定「哪段取自哪份」仍是搜索问题。起草模型的质量评价很具体：Fable 和 Open「极啰嗦」，Astra 简洁、写得好——「实验室都有 Slack 通道，这些不是新闻」。长文档推理仍是模型缺口：百万 token 合同的跨文编辑、记住上方改过的定义与顺序，他们的做法是分块加编排式编辑，且「scaffolding 随模型代际重写」（Word 编辑的 map-reduce 方案改过好几轮）；主持人对应给出自己新闻工作的迁移：从 map-reduce 转向全塞进上下文。嘉宾补了一个更细的坑：分块摘要会丢「块间」信息——某件事跨两页展开，单块与检索词都不相关、合起来才是关键，这需要步骤间推理。最后是主持人抛出、嘉宾背书的一条实用建议：RAG 重的场景别只看 MTEB 榜首，「embeddings 上还有大量 alpha」，看 Voyage 或定制嵌入——唯一前提是多语言能力（「欧洲是多语言之王」）。

五、计时收费的裂缝：合伙人制、初级律师与「谁的数据」

商业模式部分的起点是一个好故事：嘉宾当年帮 Max 改 Leya 的 YC 申请，给出的唯一反馈是「这事别人试过一百万次都没成——计时收费下，效率等于减收，律师凭什么买单」。他现在认为驱动反转的是竞争压力：律师们「知其严重性」，明白商业模式在承压、不变不行。变化已经在美国发生：低端工作（小型 review 之类）转向项目费计价，一部分干脆免费做——律所历史上就愿意为维护客户关系白干活，现在可以不亏本地做。钱则向并购（「最容易举的例子」）、诉讼和客户关系时间集中。这与 Eve 一场（见旧文）从原告端看到的「结果定价外溢」互为印证：两端 squeezing 的是同一条计时收费曲线。

更结构性的判断关于合伙人制。嘉宾的解释框架：律所唯一资本是人力资本，「你不能拥有人，但可以租」——合伙制就是让人力资本当股东从而留下来的结构。他想帮律所建的是无形资本：一个学习「本所做法」的系统、自改进的知识图谱加判例库，让方法论 codify 成可积累的 IP。但激励是错配的：「eat what you kill」奖励拉客户，不奖励沉淀方法论；合伙人离职，建过的资产带不走一分分红——「整个商业模型都在被重新审视」。至于主持人反复挑逗的「paralegals 没了、初级律师没了」，嘉宾的回应是「我们实际看到的不是这样」：如同工程界，好的初级更值钱了（产出翻了几倍），市场在两极化而非一刀切消失；那个「一百万文件、paralegal 锁屋里干到凌晨四点」的场景确实消失了，但角色在升维成 legal engineer——这个词也从 Legora 内部叫成了行业词（主持人补了个梗：CSJD）。

数据获取的讨论直接衔接上周的行业新闻（详见旧文里 Crosby 意向收购 Weil 与 Google 约 1000 万美元买 Spirit Airlines 运营数据的部分；嘉宾本场回忆的数字是「一千万、七百万，反正差不多」）。他披露 Legora 的常规做法：请愿意配合的客户匿名化文档加问题描述，「能复现就能修」。至于「买一家律所拿数据」这个版本，他指出两个未解问题：数据到底属律所还是属客户（「客户都走了也许就行——我真的不知道」），以及输入好拿、输出难用——只买到终稿，得逆向工程「他们怎么走到的」；真正的金矿是真实用户的 trace，「极高价值」。呼应他看不见 trace 的日常约束，这个闭环颇为讽刺：最缺的数据恰好是最买不到的形态。

结尾的格局判断：主持人问实验室会不会像截胡编码公司那样（「已经发生两次」）对法律动手，嘉宾认为不会——实验室要通吃整个企业知识工作（数据分析、销售、营销、财务都得做），法律只是其中一块，且「永远不会做得像我们这么深」；Claude 接一些小客户的法律需求「很好，没关系」，但「四亿美元的并购交易，你还是会要 Legora」。两年预测：大企业可能用 Claude co-work 调用 Legora（如他本人通过 Claude 用 Linear、但还是会打开 Linear），律所直接住在 Legora 里。主持人收尾的两个断言值得原样保留：「每个行业——医生、律师、码农——最后都会有两三家 agent lab 统治」；嘉宾的回应则是「现代社会某种程度上就跑在法律上」。另外两个组织侧信号：Legora 在推动法学院教学使用（嘉宾坦承「不知道进展到哪」），主持人则现场出主意——直接资助 CS 背景的人读 JD（JD-MBA 常见，JD-CS 罕见）。

机制、二阶影响与读者启发

链条一：数据不可见 → 评测体系重构。 变化：无 trace、无数据协议、匿名化受限；原因：数据的终极所有权在律所的客户；得失：失去数据飞轮，换来「零风险」的合规位与被迫建起的失败模式本体；二阶影响：Agent 公司的 QA 职能从「看日志」变成「设计合成世界 + 归类失败」的环境工程师，评测能力本身成为垂直壁垒；观察指标：合成评测与真实客诉的命中率、失败分桶覆盖率；启发：做不了数据飞轮的行业，先把「失败模式本体论」建起来——它是数据飞轮的替代资产。

链条二：表格成为 agent 的原生界面。 变化：fan-out 检索被装进百万格表格且用户不导出；原因：二维审阅匹配律师（及一切专业审阅者）的既有肌肉记忆与逐格溯源需求；得失：牺牲对话式炫技，换来审阅效率与信任；二阶影响：每个「同构任务 × 海量对象 × 人工复核」的垂直行业都可能长出自己的 Clay（表格 Agent 层），AG Grid 类基础设施与百万格稳定性成为隐性需求；观察指标：其他垂直 SaaS 是否出现「表格+溯源+agent 编排」三件套；启发：选界面形态等于选用户的肌肉记忆，不是选技术最优解；fan-out 必配「前十行迭代」安全阀。

链条三：计费模式反转采纳激励。 变化：低端转项目费/免费、高端向并购与诉讼集中；原因：计时收费下效率自伤收入，竞争压力强制反转；得失：律所从效率抵制者变成采纳者，软件商的 TAM 从「省下的工时」变成「承接的工作」；二阶影响：合伙人制的资产结构被追问——人力资本若能 codify 成系统 IP，「离开无分红」的激励必须重设计，律所资产负债表可能首次出现「方法论资产」科目；观察指标：美国低端业务平价计费比例、大所初级招聘结构、法学院 AI 课程渗透率；启发：向专业服务卖效率，先问「谁的 KPI 被效率伤害」，计费模式不变则效率卖不动。

链条四：合规约束重塑模型排序。 变化：DPA 供应商极简、全栈 Azure、转写弃 ElevenLabs；原因：信任继承（Microsoft 365 已过安全审查）与数据出境治理；得失：单一云绑定 vs 拿下保守行业；二阶影响：「够用 + 治理 + 一站式」的厂商获得合规溢价，榜单名次不等于可用性排序，开源/自托管组件在合规敏感链路价值上升；观察指标：DPA 名单长度、开源模型在敏感任务中的部署比例；启发：受监管行业选型的第一张表不是模型榜单，而是「我的 DPA 允许谁」。

转写校正表

账本高优先级 18 条均在正文有落点（见各小节），未采用 2 条：姓名词源（-sen＝之子，纯寒暄趣味）、主持人「付钱请律师当朋友/retainer」玩笑（喜剧噪声，无行业信息）。

转写原文（修正版）

[0s - 120s] People were asking us about Mythos and everything, even though I think realistically a lot of legal work does not require mythos level intelligence. Paralegals gone. Yeah, junior lawyers gone. I don’t need them. That’s actually not what we’re seeing. We had one problem where like it wouldn’t cite properly, and so like next month we resolved. It went from like seventy percent to like ninety nine. It’s like we did nothing and like great. But if you’re doing a you know four hundred million dollar M and A transaction, you’re going to want Legora. We’re just going to be twenty engineers for the rest of the company’s history, and like we’re going to beat everyone who’s much bigger. And we’re a hundred engineers now, and it feels like we’re way too small. But it’s something that I’m toying with is like how much of our product can just be the agent versus all the other stuff. Okay, we’re here in the studio with special guests Jacob Lauritzen, CTO of Legora. Did I get it right, Lauritzen? That’s fine. Yeah, that’s okay. I’m not too fancy about how it’s Danish. Yeah, exactly. Does it have a special history? It just no. Sen is like son. Son of. Son of Laurits. Just like you have son and yeah. So son of Laurits. That’s my name. Nice. We’ve crossed paths here and there. You you you spoke it. Yeah, a few times, and obviously Legora is one of the fastest growing legal tech in history. Yep. And you said something at the start in a pre-show that I thought was really interesting. You said that your origin was very European, even though obviously you’re expanding in New York and SF. What feels different about your culture? I think we are just at the core probably very humble about what we’re doing because we’ve always been the underdog from the very beginning. No one believed the smallest startup in a basement in Stockholm was going to you know outcompete this huge behemoth that was like way ahead of us in the US. So I think we’ve always been underdogs, and so we’ve always been very humble. We’ve always sort of had to fight for our existence. I think it’s something we’ve had to learn as we’ve also become more and more as we’ve entered the US and been in the US a lot more is to be a little bit more bold and like tell the world how great we are because it doesn’t come naturally to I think some of us. I kind of want to start by going over your guys’s product because I feel like with a lot of these like legal products.

[120s - 240s] You know they’re kind of like under wraps, like you know you’re not supposed to like show it. But yeah, if you could just talk about like what does the product look like, and maybe like how did you guys get there? Like like what have been the product iterations that got you to where you are today? I think that would be really interesting. Yeah, absolutely. The way everything started was ChatGPT with Swedish laws search tool in it. That was literally the entire everything, and so that’s where we started. We started with legal research, which is like a huge market basically in the world where lawyers spend a lot of time figuring out what are previous cases that look like this, what’s the legal precedent, like is this you know what have previous people done in these situations? That’s extremely important. And so that has always been part of our product. Basically, we started from like an an agent. You could ask questions and it would query laws. That was the very simple beginning, and then from that it’s expanded. And so it’s expanded in like what can the agent do? It will now you know draft a lot of stuff really well. It will edit word files. It will you know move your files away, organize everything, run really complex analyses, and we’ve added just a bunch more features to it. So one of our features called Tabular View, which is now a name that’s being used in industry, and we’re very happy to see Claude for Legal using our terminology. We got to talk about that. We got to talk about that. Which is essentially like a way to just bulk extract and bulk review many, many, many agreements, which you do in due diligence and when you raise money or whatever. So that’s part of it. It’s like we help lawyers just do their work much more efficiently, much much better. And then outside of just doing the work, we’re also now helping orchestrate the work a lot more than we were just even a year ago. So we have different sort of we have a thing called lists, which is sort of like linear for us, and we have intake, which is a way to collaborate with clients, and we have portal, which is also sort of a way to collaborate. You know, lawyer to lawyer, client to lawyer, lawyer to agent, agent to agent. Maybe we’re spending a lot of time there because there’s just frankly the way that.

[240s - 360s] You work with a lawyer ten years ago, or maybe even today, is is really painful, and there’s loads of room to do that in a much better way, to be honest. And so those are like, I think the two things: it’s doing the work, and it’s also like orchestrating the work. When did you guys start doing or building the agent? So I mean, we’ve all we basically had this switch from the assistant to the agent. Yeah, we call it right. I guess that the agent. I don’t know. I want to say a year ago, probably. Yeah. So this is a rough timeline for people who are understanding YC twenty twenty three. Yes. And then you were you were YC founders. YC twenty four. YC twenty. Yeah, I was in YC twenty three. Yeah, Legora was twenty four. Okay. And then, so so there is this is like a little bit pre pre like agents really taking exactly, and then like reasoning models obviously change everything. Yes. I mean, you know, in one in one sense, today, you know, we had a preview of the products. It does look well. Preview. We had a private thing, but we will put up a video on the on the YouTube. It’s mostly like ChatGPT for lawyers, right? Like a very like very fair to say that. But then, like the sidebar has a lot of nice custom tools, like the tabular view, like the intake, which I really liked, and you know a bunch of other yeah things. Exactly. Yeah, that’s that’s exactly what it is. It’s it’s co work or ChatGPT, just like really really good for legal work. And yeah, the the the workflow for like either like in house corporate lawyers, like people like looking for upcoming regulations, all these things, and we can talk about more of the the products there. Yeah, what models are you guys using underneath? Like, can you talk about that? I mean, we’re using a bunch of different ones depending on task, right? That’s okay. And then that’s not available to the lawyer to see. You guys for the agent. Yeah, well, we’ve been back and forth on this. Yeah, yeah, just like everyone else, I think. In the agent, there’s a model selector for the primary model, but there are certain tools that we use different models in, and there’s like tabular view and like we try. I think we found it most effective to subdivide the problem and and and route on like as we subdivide the task, but we do right now have a a picker in the agent. Hmm.

[360s - 480s] I’m not sure if that’s going to stay there always, to be honest. Ah, so yeah, yeah. I mean, so I bring a little bit of context to this because I’m advisor for cognition. Yeah. And when I joined, you know, part of my my thesis for what an agent lab is, an agent lab is, you know, ideally a domain specific GPT wrapper. Yeah. Right. Like we we we run the we are the domain experts. We’ll pay attention to you. We’ll serve your needs way way more than the frontier labs, and we’ll give you the best of the frontier labs according to our judgment. Yeah. So our job is to decide for you because you’re not your job is not to exactly. We’ll always keep you on the Pareto frontier. Yeah. And make sure that it just it’s great all the time, and you don’t have to dive into all the details. Because like one of the things that I’ve exactly to this point, if you go to a lawyer and you ask them, oh, you’re gonna you know draft this contract based on this precedent, do you think that’s a Sonnet high or an Opus mid task? Actually, you know, they they won’t know. Just give me the farthest model. Exactly. Just like make sure it works. Our hourly rate is a thousand dollars an hour. Exactly. Just like what make it work. Um. So yes, I originally came in with that. I was like, oh yeah, this is this is what Devin does differently, and now we have five things in the model picker, including Fable and Opus and Astra. Yes. And uh, people want to experiment with the fancy new model. Yeah. Do you? Is that the same? I think so. I think it’s probably it’s it’s definitely way less illegal than it is in um coding engineering. Yeah. But there are definitely people that just want Fable One right away, and you know, people were asking us about Mythos and everything. Yeah. Um, even though I think realistically a lot of legal work does not require Mythos level intelligence. Yeah. Yeah. So exactly right. Like I think basically you’re like, okay, we’ll give you the option. You you ask for this. Exactly. But you’re gonna run down your credits like ten times faster than. Yeah. I would have loved. We we fought introducing it for a very long time, and then in the end we we just kind of caved in and were like, okay, if you want this level of control, yes, you know, then we’ll give it to you. Yeah. So you think that models are already like as good as lawyers need? Like there is no need for them to get better. It’s a I think it’s a power distribution. So like there’s probably like sixty percent of the work. Yeah. We can do really uh.

[480s - 600s] Well, right now, and the you know from now on, it’s like a cost and latency thing. But then there’s the rest of the the thing where, especially for like very complex analysis, we we still need more intelligence. And then I think there’s also just the work expands kind of like that’s the work that we have right now, but also the work before the law to thus probably expands with intelligence. Yeah, I imagine early on, like when you guys first started, I guess transitioning to the agents or building the agents, like what were like what models were you guys using? I guess it would have been like Opus four six or something. That would have been yeah, Sonnet four six, Open four six is probably right. What were the agents like? Not good at at that point, that they’re much better at now, even though it’s only like nine months later. Sub agent delegation was one of them that we’ve gotten working really well now for like fan out type retrieval queries. Works extremely well for us now. That was difficult before. We basically track so we have a team that’s just doing evals, and so we track all of the models’ performance in different practice areas, which is like subgroups of legal and different tasks across all of those. We basically have like a big heat map, and we’re like, and then we can see them move. And it’s like usually they improve most stuff, but there’s actually a lot of model releases where they improve some and regress in others. Okay, and that actually happens a lot more than you think. Like, what are some of those tasks? Like, are you familiar with how you guys are running a lot of the evals? Like, like what are those different sub segments? I guess that you look at. So like, there’s um, there’s different just like framing stuff. So we have different basically like worlds or like our environments essentially, which is like here’s a because because lawyers work in projects always more or less. So it’s like here’s ten thousand file data room. You’re doing a due diligence. In the end, a grid report looks like this. Like it contains all these right red flags, and like it doesn’t contain whatever weird stuff, and like we want it to be concise, etc., etc. That’s kind of how we’ve built it. So there’s different tasks, which is like part of it is drafting a report. Like does the draft look good? Does it contain all the right stuff? Is the wording sound? The analysis sound? There’s also the

[600s - 720s] Part of finding the red flags, like the analysis before that, so we can kind of like subdivide it a little bit, and and that’s how we run the evals. Interesting. Like, what does your evals team look like? Like, do you hire lawyers to like do that? So yes, it’s a mix. We have engineers, and then we have lawyers, and it’s one team. So there’s like three engineers and like three lawyers right now, and then we have a pool of what we call legal engineers, which is people that have worked in big law for like three or four years. Typically, maybe they have a CS degree and like they’re very tech forward, and then they come to us because they want to like help change the industry. Are they IT in big law? Or no, they’re lawyers in big law. They’re lawyers in big law with CS degrees. Yes. Oh my god. It’s a Venn diagram. Yeah, exactly. That’s like our hardest person to hire for, but they’re also such a force multiplier because we can pull on them for the eval stuff and for the product development. It’s just like you have a pool of a hundred experts that use the product every day and go and talk to clients every day, and also like they live this and have lived this, and we just pull on them to do evals and like help us shape the product. And it’s such a such an advantage. And curious, like when did you guys start that team? Eval’s grew out of the assistant agent team at some point. I think there was a point up until recently where it was like honestly just not really worth it because like all models just got better. So it was like, do we want to spend a lot of our time doing really complex evals, or do we just like want to build? Because there was so much to be gained from building a better harness, and like just sort of leaning into everything’s going to improve, and everything kind of just improved. Yeah, like every two three months you get a different year model. Yeah, exactly. Just pass it through, right? Exactly. I mean, we had one problem where like it wouldn’t cite properly, and so it would be difficult to follow. And it was like next model release solved. It went from like seventy percent to like ninety nine. It’s just like we did nothing and like great. Yeah, maybe the the the real work is uh you know keeping all the data private, making it taggable, sortable, adding new features like the the tabular view. Is that the you know you mean over and above ChatGPT? Like yes, but obviously now you’re also doing evals.

[720s - 840s] I assume your your use of open models is also increasing. We can talk about that. So yeah, I don’t know what I’m saying here apart from like you know what is the job. The job is maybe not in the early days. It was not so much to do do emails because it was mostly exposed model capability to lawyers. Yeah, the huge part of it is just figuring out what does the legal profession look like in five years, ten years with all of this stuff. Like coding is maybe a year or two ahead, so like we can kind of look to that a little bit. But like, what are the primitives? How do people work? How do you collaborate? If it’s not a 200 thread email with like V5 word files back and forth, like what does that look like? And that’s like not even necessarily purely AI. That’s just like build a much better product experience. Yeah, and make sure you do it with like audit logs and data retention and all of those sort of really important things in regulated industry. And then there’s the the AI part, which is like make the work faster and better and proactive agents that react to stuff when it comes in, so you don’t have to you know kick off everything. So I think a lot of the work is just figuring out what do we think is going to look like in five years. If you know model capabilities keep increasing, if price keeps coming down, what could we do? Can you log in and it’s like command center where it’s just like you need to react to these five things that the agents have been doing lately and like review and check and improve the system. Maybe that’s where we’re going, or is where we’re going. Yeah, yeah, yeah. I was going to ask, like, could you go into that in a little more depth? Like, yeah, what what is the like law profession going to look like in a few years? I should preface this by saying I don’t know. I mean, who knows really? But I think we can look to coding a lot. I think all lawyers are going to be way more productive than they were before. They’re going to be able to do a lot more stuff at a higher quality for lower cost. I think the profession itself, the work itself, is going to improve a lot. It’s going to be a lot more. I think lawyers are going to spend more of their time doing the stuff that they became lawyers to do.

[840s - 960s] You know, figuring out what’s the right risk profile, which manageable risk, what’s like all of their judgment, strategic judgment, and like working with clients—they’re going to do that a lot more, and much less of the like trolling through big documents to like find the smoking gun or like do all this menial, super annoying work that a lot of lawyers have been doing. What do you think is going to happen in the partnership model? Like, do you think that stays, or like what happens? I gave this—I actually gave this a lot of thought, just like last week. Yeah. The the partnership model—I don’t want to get into too much of the details, but like the partnership model kind of exists because the only type of capital that these companies have is human capital. So like it’s not like you know us or like a software engineering firm or like a manufacturer or whatever. You you have physical assets or like some sort of intangible assets that that you accrue over time and and they pay dividends to the shareholders. But if you work in an industry like professional services, and it’s the same for like consultancies and and whatever else, all of your value is in your people, and you can’t really own human capital. You know, you you can’t—you know—slavery is abolished. You can’t own people, but you can rent them. You can pay them money, and so partnerships essentially is a structure where you make the human capital the owners of the company so that they stay. Yeah. Will that change? Like, I don’t. It doesn’t have to, but I think what we really want to do with the lawyers, we actually want lawyers to be able, law firms to be able to build some sort of intangible capital, some durable capital, in like you can build a system that learns of how you work. A lot of firms right now think about like, well, there’s the X law firm way of doing things, and we can codify that, I think, and we can build a system, and whether that’s like a self-improving knowledge graph or whatever it is, and like precedent files and documents, where there’s like a way of working that produces a really good out output, and that can be their IP. So it’s like a way to codify the human expertise and judgment into a system. That I think is now. I’m just rambling completely, but like that’s something that we want to help empower.

[960s - 1080s] And does that change the partnership model then? Because then incentives are different. In the partnership model, you just got to bring more clients, and you don’t want to take away time from bringing in clients to codify your information, basically. Because like you get rewarded from you know eat what you kill or make this large step, whatever. That needs to change a little. We need to figure out an incentive structure where partnerships, partners, are incentivized to codify things for to build an asset for the company. But right now there aren’t because the sum is they leave the partnership, they don’t get any dividends from the assets that they built. Yeah. So it’s like a yeah. I think the the the whole business model is is we’ll we’ll see. Yeah, yeah, it’s all it’s all in question right now. Yeah. My my actual other take, you know, so on the partnership level, that’s like the top end of the the industry. Actually, my take is more on the bottom end, right? Yeah. What happened there? Paralegals gone. Yeah. Right, like junior lawyers gone. I don’t need them. Yeah. That’s actually not what we’re seeing. Okay. I think it’s. I know. Would you Would you say that we’re seeing that in engineering? Ah, yeah. Yeah. So junior engineers are really kind of cooked. Yeah. You kind of have to come out of drop out of university fully baked, or else you’re. Yeah. You’re not gonna make it. I feel like I’m not. I’m being a bit dramatic for for comedic purpose. But like, yeah, I know a lot of lawyers. They started as paralegal. They go to law school. All these things, right? And now they don’t. They’re not needed. They’re just seriously just not needed. Yeah. Yeah. Yeah. It’s interesting what happens with the partnership on the like the pyramid. I think generally a lot of lawyers, you still need you need the next generation, and at least for engineering, the way I think about it, you have a good engineer today. They’re worth a lot more, even than they were two years ago, because they can produce so much more stuff. Like if they’re good, and I think it’s the same thing in legal, where like they can actually do a lot of really good work if they’re good. So I think it’s more that like really good lawyers or junior lawyers are worth a lot more. Yeah, it’s it’s always been the case, right? Yeah, yeah, exactly. Even more so. But it’s like it’s yeah, it’s polarizing even more so.

[1080s - 1200s] I’m actually curious about that. So, like, the whole like law business is basically like charging per hour. So, you know, if each project now can be done a lot faster, is it going to be like way cheaper? Like, this was actually so fun story. I was I was reading Max’s YC application and like helping him getting into YC. Leya, as it was called back then. It’s YC, and this was my one thing. I was like, and curious. Leya is a Danish thing for something. No, a Swedish. Swedish. No, also it means nothing. Okay, okay, okay. There’s this really funny story, and I think it was a Spanish newspaper that wrote about us in the early days, Leya, and then they they explained that the abbreviation was legally efficient Yoda assistant. Why Yoda? Yeah, I don’t know. I’m like, why? Why is the hallucinations are not just from LLMS? No, I think journalists hallucinate. But no, it doesn’t mean anything. So he wrote the application. He he wrote. Yeah, my feedback to him was, you know, people have tried this a million times. Legal tech has never been a thing because like there’s no incentive for lawyers to to. Why would they be more efficient? They earn less money. Like it doesn’t make any sense. But I think the reason that lawyers do want to adopt this is that they understand the gravity of the situation. They know that like the business model and everything is under pressure, and it’s like they need to change because the competitive pressure is so high. The billing model itself, I think, is also going to change. We we’re already seeing that, particularly in the US, that a bunch of stuff is actually being priced on like a project fee and not on an hours fee. Particularly like sort of lower end stuff on the spectrum, like whatever it is, like small reviews and whatever. Some of it honestly is being done for free, and some of it is just being priced on a flat rate. So where is the money being made then? Is it because the strategic work? Yeah, I feel like a lot of what you guys talk about is M and A specifically. Like that’s where all the money is. It’s just a very easy example. Yeah, yeah. Litigation is also the big work still is there, and it’s very valuable, and so they’ll they’ll spend lawyers will spend more of their time there. Client relationships also super important. Like a.

[1200s - 1320s] No, just like spending more time with the client. Like a lot of law firms will, or historically have, basically done stuff for free just to have the relationship with the client. Yeah, yeah. And so now they can do that, but they don’t have to lose a lot of money on doing it. I thought I thought I just have to pay the lawyer money just to be my friend, like isn’t that the retainer thing? Yeah, yeah, that you tell me I can do it for free, like yeah, you just have to find the right lawyers. Yeah. Okay, cool. Or land space has to grow much bigger. I don’t know. Yeah, yeah, we’re trying. You know, yeah. So okay, so so so you’re saying, you know, obviously the the strategic work and all these high value add is still there, but our legal fees are not going down. But you also said paralegals are not affected, or like they’re they’re just shifting, and and this is something that we also did with worry about with cognition, like people worry about the phrasing of like we will replace your junior engineer because nobody likes that kind of statement. But in actual fact, yes, like some some part of it does happen. Other junior engineers embrace it and they become more productive. Exactly, exactly, like you just said. So I guess like is is there a new term for this role or this this job that is taking that is having traction with your customers? You mean the the super AI built junior legal? Nothing I found yet. Maybe we should coin one. Isn’t it legal legal engineer? That’s our and and also now an industry thing. Yeah. CSJDs exactly. Yeah, yeah, yeah. Man, maybe that’s the those other people that are best positioned to do this kind of stuff. I think the the the derogatory term of meet proxy, where you just like sit in front of the machine and type in something that someone else tells you to type, and then you repeat what the machine tells you to say. But like I do think a very good meet proxy is just like was you know agents very well, and you you know Legora very well, and you know how to like like this is my toolkit, and you know every industry has a toolkit, and I can use my tools to do the thing that you need me to do. But I I guess then you’re not really a proxy, right? You’re actually just doing meaningful work. Yeah, it’s it’s just like I no longer like you know.

[1320s - 1440s] Well, I’ll talk about Tabular Review because that that’s a thing that happens a lot. I have hundreds, thousands, millions, thousands. We do up to a million cells, I think. In a in a single that’s impressive. Yeah. We should talk about the the the table and the windowing. Yeah, you know, I’ve like I’m working on a big case. I have a million documents. I type it in. Usually, I’ll throw a throw like a team of paralegals locked in the room. And they’ll sit there till like 4 a.m. Like that that doesn’t need to happen, right? Because now they can extract it, and you can go and re prompt everything. So, like, yeah, I think now I don’t need a paralegal for that. But like a a a good like super AI pilled paralegal would do something else, you know. Yes. Yeah. Exactly. Whatever. Yeah. Yeah. They can do more stuff or like spend more time on the things that they you know didn’t have time for before or actually sleep. Who knows? No. I have a question. So you guys bought a company Kira. Yes. Is that how you say it? And you you guys used your own product to do the M and A. Yes, we do that. So how did that work? Like, can you talk a little bit about that? Oh, dogfooding. Yeah, no, that works great. We’re doing that a lot. Our legal team is is I mean, Intake has been built with our legal team because right now we have a Slack workflow where everyone flows their stuff in, and they’re like, “This is annoying.” Yeah, yeah, right, and inefficient. So it’s like great. We should do Legora for that. Yeah, yeah. Can you talk a little bit about that? Like, you know, what that process looked like? Yeah, absolutely. So it’s it’s we do a lot of the due diligence ourselves because we can do it with Tabular Review and other tools. We do still have counsel coming in to like assess certain risk or like you know, there’s payroll and so the career was simple because I was all in Sweden, but you know, we’ve done one in Canada and one in Australia, so there’s like payroll in Australia. Okay, do we want someone to look at this like a bit more strategically? So there, we will still use Lawless for the strategic part where they can actually be sparring partners and help us a lot on like how to figure out how to structure this the right way. But a lot of the more menial stuff, we could actually just do ourselves. So like a lot of yeah, a lot of the early things we would find, and we have had cases where we found something that’s been like a red flag, and then we’ve we’ve pulled the brakes. Can you talk a little bit about any? No, I don’t think so. Yeah. All right.

[1440s - 1560s] Early on, you said you your early customers were Swedish, like Swedish law firms. Have you seen a difference between working with like Swedish law versus like American law firms, that sort of thing? And you guys are mostly working with like big, big law firms, right? Well, we’re also smaller law firms, also. Oh, no, we we definitely started really, really big law firms, but fifty lawyers also. Okay, and and get a lot of value. Yeah, so have you seen differences between working with like the Swedish ones versus American ones? Maybe like even I don’t know Asian ones. Yeah, yeah. I think the European ones have been a bit slower to move. The American ones are very forward, move really quickly. I think it’s also very high on their agenda of the American companies, more so than the European ones. I think it’s like a strategic, super important thing we need to just like do a lot here. The American companies generally, both corporates and law firms, also have a much higher appetite to like custom build stuff. That’s like a thing that they all want to do. Like, hey, we want to build this like huge thing with you guys and like structure in whatever way. Whereas Europeans are much more like, we’ll just take the off the shelf and. What’s an example of something they would want to build themselves? Integration into their own, like some of these customers build their own CRM and their own document management system and their own all sorts of stuff, and they want to build that, and then they want to build triggers that, like, you do a review in in Salesforce and goes to this one person for whatever reason, and they do something, and then it goes into Legora, and then it goes back in, and there’s that’s like the simple end, and then there’s some that are extremely ambitious, where it’s like a whole new product that they want to build with us, and I, you know, we can’t really do that. Yeah, yeah. Um, but but that’s definitely a thing where like American companies really want to build their own more than European ones do. And the European ones are like, hey, like, yeah, they give you product suggestions or. Yeah, for sure. I mean, we have great collaboration with all of them. Yeah, so they’ll they’ll say, “Hey, this sucks. This is great. We want more of this, and we’ll just build a product that does that.” Okay, so they don’t see like an existential threat here in their business. I mean, whereas the American ones do, or I don’t know if that’s the reason.

[1560s - 1680s] That they want to custom build versus not custom built. I think they both know that this is like extremely important and key to their firm’s success and their corporate success. So they’s both extremely important. I think it’s just a different way to achieve it. Okay. Another question I think maybe you want to bring this up is like product metrics. Like, how do you know that like you know we’re actually doing something good here? Yeah. How do you think about that? I’m assuming that’s also related to evals. So we have the regular product metrics that we care a lot about. Are people spending a lot of time in our product? Did they seem to get a lot of value out of it? Do they renew it? Do they come back a lot? We have really good numbers for an enterprise software of like our DAU/WAU/MAU and our like you know hour spent per month in the product are extremely high. And then we do some sort of we do online evals also, where it’s like try to figure out generally like did we solve the problem here? Just with like super cheap LLM messages, try to just like have an online metric, and then we can use that also when we do roll ups of new things. We can just like compare feature flag wise. Okay, does it did it improve or like did it catastrophically decline? Because one thing actually that’s might be interesting that’s difficult for our businesses. We can’t look at any traces basically. We can’t look at any data because it’s all super confidential, and that’s like extremely frustrating to work with because you you basically don’t know what people are doing. So it’s like here’s an agent and it can do so many amazing things, and I have no idea what people are actually. I would have assumed you guys like sign some sort sort of agreement with the law firm, and they’re like, oh yeah, I wish. Yeah, I wish. No, but the thing is like a lot of the data that they have is not even even theirs, right? It’s their clients, and it’s super confidential, so we can’t look at it. And so basically, we have an evals team that’s like trying to approximate what we think everyone is doing and like build that into eval sets. So that’s just like a super annoying kink of the business. Can you do paraphrase? So for example, ChatGPT has like a slash feedback thing that directly fully says full trajectory, but obviously that that is not possible for you guys. But you can use an LM to paraphrase like from this real thing, yeah, make a synthetic thing that is like.

[1680s - 1800s] Enough, but change all the names, change all the details, but like still, yeah, yeah. We can’t. There’s, I mean, there’s always a a range, right? Of like, if you if you paraphrase all the way down to like three values and you pick one of them, then obviously that’s fine. If you if you just remove the names, then that’s maybe maybe not fine. Yeah, it’s like we are um generally pick the the most safe way. Yeah, because like we we restrict ourselves as most possible because we you know there should be no risk at all. Yeah. Um, but there are things where it’s like okay, we know that the failure modes are probably these ten or twenty, so we can classify it into one of those, and that’s a helpful signal. We can classify it into the types of tasks and the practice areas also, and then like we have an idea. It’s something that we’re always trying to to figure out how to do in a smart way. Also, like on the privacy front, I guess I’m assuming even two years ago, it must have been like insanely difficult to get the law firms to be okay with this. But like, how do they think about their data being sent over to Anthropic or yeah, yeah, yeah? It must be like a huge thing. All of our infrastructure is on Azure. Yeah. Um, primarily for this reason, because like when we started, law firms were like, AWS. No, no way. Law firms don’t trust AWS. Yeah, yeah. It’s different now. No, but there were like Amazon, Google. We can’t trust them with any of our data, so we basically we had to host everything on Microsoft because they already have 365. Oh, so they’re okay with Microsoft exactly because they’ve already done like whatever security review, etc. Of all of their documents and all of their Word and whatever, so like Microsoft is an approved vendor in their mind, and so we had to host everything on on Azure, and that’s changed over the past few years. So like now people understand, okay, we need to get on cloud, we need to do all this kind of stuff. Um, there was also a lot more people talking about on-prem back then. That’s also disappeared completely. People understand that like, you just do it in the cloud, and and it’s safe, and it’s probably safer than what you’re doing. Yeah. Um, so that’s changed. Yeah. But and by the way, it’s only been three years, so like it’s actually like yeah, yeah, yeah, it’s a huge shift. It’s it’s great. I mean, good for you guys, but yeah, I mean, I don’t expect the like you know the legal firm legal industry has somehow ignored this for the last like ten years. Yeah. Yeah. I think I think we needed the AI.

[1800s - 1920s] Kick in the butt for everyone to understand that we need to change our ways a little bit. Okay, okay. So also then on like open source models now are getting really good. So yeah, how are you thinking about that with the? Yeah, we’re just thinking about to your point. Just bring our clients along on like the Pareto frontier. Yeah. And like the of the the map of all the models and what they’re good at and their latency and their cost and their governance and and and their regional availability. How do we just optimize across all of that? So some open weights models are great for certain tasks, and then we’ll use them there. Some will post train, and and some will just use the frontier models. Yeah, you called you called it model arbitrage. Yeah, you have different favors. You did some research on this, where you are firmly Anthropic at the start of the year. Yeah. Grok is improving a lot. Yeah. You want to give an example of a model that wasn’t ranked highly overall, but like you want to shout it out as surprisingly good. Yeah, I think not of all good. Gemini, one of the early Geminis, is yeah extremely good for certain types of transcription and speaker attribution. So we use that for transcription. Also, you have a yeah. Are you guys transcribing? Like a granola, we don’t have. We actually probably will build like a meeting thing. Yeah. But this is just like talk to speech, speech to text, text to speech, just like in and um any any other app that has this. And we do actually also transcribe files if you upload hearings or whatever. Oh. We’ll we’ll then like transcribe them for them into text, and so we can do all of our analysis on it. No Eleven Labs. No Eleven Lab for DPA reasons. Same thing is like the data processing agreement. I should say so. Like our data processing agreement has very few vendors on it. Yeah. Because it’s really difficult. Exactly. We need to keep it super tight on where any data goes, and so. I think no, that that is one of the benefits of like a Gemini and OpenAI, where they’re like you come to us, we have most of the things that you need, and that is actually very very valuable. Even if it’s like it doesn’t have to be like state of the art, as long as it’s like mostly good enough for all tasks. Yeah. Then who cares? Yeah.

[1920s - 2040s] Yeah, no, absolutely. I should actually shout out also Microsoft’s transcription model. They made one, my transcribe. Yeah. Okay. It’s great. Okay. It’s better than the old Gemini. Yeah, yeah, yeah. They they they did some good results. I don’t know, don’t even know. Like, I think like usually all of them are like top two, top three, top four in terms of like the the arena scores, and that’s good enough. Yeah, exactly. And it’s like you have all the governance and it’s super scalable, and it’s like then that’s probably good enough. Yeah, I was gonna say like at this point, are pretty much the speech to text models like good enough to where you’re like one word here there if it’s off. No, no, no, for sure. Like I’ve done this a lot, particularly in legal. There’s like some we we had this thing where I think it’s I forget if it’s like in Australia where there’s like a certain name for like a a person in court, and it would consistently transcribe it as condom. Oh. So there, yeah, there’s there’s certain things that you don’t want there. Yeah, but so usually you run an LLM layer on top of it, and then you you extract some keywords, and then you run the transcription model again with the keywords. Like this is kind of my my workflow so far. Yeah. But yeah, if you if if you’ve just done this on any large scale, you will pick up tons and tons of transcription errors on any domain that is not like average day to day conversation. Yeah. Yeah, absolutely. If there’s any sort of jargon in there, yeah, it’s gonna be tough. Yeah, like I think Whisper and Eleven Labs still transcribe Claude Code as Cloud Code. Yeah. And like yeah, I’ve actually had very very common. Yeah. Uh, so so like that’s that’s how you know there’s like a lot more to improve. Yeah, absolutely. I do want to cover a little bit on the the agent product side. Uh, so we’ve covered a few like the of the top level stuff, right? We’ve already name dropped tabular review. Uh, we name dropped intake. Yeah. And you’ve you’ve got some other uh stuff that we haven’t even touched on yet, but I just want to cover the the big two. Let’s call it like is it is that is is that fair to say they’re the big two agent and tabular review uh and projects yeah yeah and our word add in um okay is also something because that a lot of people were working.

[2040s - 2160s] Inside of Word, and so Word is a huge an outlook add-in, also. Yeah. So I just want to cover like the the big ones that everyone must know. Yeah. So Tabular Review is basically like fan out of like a lot across each document. You can ask the same question across all all documents. It’s collected on the table. Yeah. And then work back and forth. And then yeah. A million is impressive. What’s any other like engineering considerations or performance things or like what what’s hard about this? Before we do that, can you just like walk through like what is the user experience for a lawyer and then like okay, so you want to use a top level flow, yeah, yeah, yeah. Like so, like what are they doing like with Tabular Review? Yeah. And then we can go into that. Sure. Yeah. So they it’s typically I mean a good example is just a due diligence right where it’s like we need to review all employment contracts, all supply contracts, all of whatever. And so employment contracts, you put ten thousand employment contracts in there, and then you like each cell is one employment contract. Each row is one employment contract, and then for cells you ask questions basically. So like the first question, the first, the sorry, per column. So the first column might be what’s the termination period, and then it’s gonna like fill out for everyone. It’s thirty days or whatever it is. And the next question is like, is there an IP assignment clause or whatever? Next one is like, how much you know, how much do we have to pay them if we fire them or whatever it is? Like how much are they getting paid? And then it’s gonna fill out the entire thing. You can also just ask the agent to do this. Like if you just drop ten thousand files into the agent and say like do a due diligence on this, it will spin up a Tabular Review and run all of them and wait for it to be done, and then analyze the results and it can then highlight like red flags. Okay, this the CEO is getting paid an outrageous amount or there’s like this weird bonus structure. Okay, then like obviously we need to look at that, and then the agent from that can draft a memo to send to the client, or it can just do the review and then have you know lawyer the use or just be like okay let’s let’s look at this here are the red flags. It’s also collaborative so that you can comment on it for other people to look at, etc. And why did you choose that like spreadsheet format? Yeah, it’s just a lot better to it’s it’s super easy to work with work in. It’s really easy to get an.

[2160s - 2280s] Over you super quickly of what’s going on and to review it. So all of the stuff that we produce also has citations back to source documents, and so you can basically review it super easily. If you want to just make sure that everything’s right, you can also like highlight cells. These are reviewed or whatever. These are the ones that are flagged, and you can comment and collaborate. That’s much easier than any other form factor that we’ve tried. Yeah, I would say that the spreadsheet form factor has has a lot of pull in some areas, but somehow hasn’t really spread. I uh, it’s I find it very weird. Like Clay is another one on the GTM side that you know very known for the spreadsheets. Yeah. You guys, I don’t really know if anyone else does a large spreadsheet AI thing. No, no. There was like V7 that like try to launch. There’s like sheets. There’s even in finance domain. I don’t really have like a top name that like does spreadsheets. Does this really well? Yeah. No, I I agree. I think it’s like a great if you if you have fan out and you do the exact same thing and you’re extracting stuff. It’s it’s such a great form factor to like understand what’s going on. Do you find like you must export to Excel or like people are happy to live in your spreadsheet? People are happy to live there. Do you use AG Grid or your own? We do use AG Grid. AG Grid that seems to be like the leading one. I mean, they’ve done this for like I don’t know, 15 years. Yeah, yeah, they know how to. Yeah, yeah. So, so yeah, I mean, I’m happy on that that spreadsheet angle. I mean, I think the the only other interesting thing for me, engineering wise, is the fan outs. Yeah. Engineers maybe can are well equipped to understand the trade offs of like, well, okay, before you you fan this thing out on a million documents, maybe test it on like two. Yeah. Yeah, yeah. Does that work before you go nuts? Yeah. Do they do that? Yeah, they will. So our Tabular Review actually will run the first 10 rows. Yeah, just check right. Yeah. Does this make sense? That it did it show us what we thought it was going to show us? Yeah, that’s that’s really good. Oh, but also for people, when people build this, it’s just really good for them to iterate quickly on the first 10 rows until like they have a they’re getting the results that they want, and then they can go ham. Yeah, I’m happy to move on to the next thing, intake, which is your Slack bot. Yeah. Everyone has a Slack bot. What’s different?

[2280s - 2400s] About yours, so ours is just legal specific, and what what happens with ours is it’s not actually. It might also we will probably also build one that’s just an interface to the agent, but it’s actually specifically to basically kick off an agent that then adds a task to our list. So we have this lists thing, which is like linear for you know that’s the closest equivalent, and so it will add it to a specific intake list, and then lawyers and agents can go into Legora and just be like, okay, these are the things that you know people are sending to us as our law firm or as our legal team in a big corporate. We can kick off agents for them. We can collaborate on them, and then we can just like mark them as done, and then it gets sent back in Slack. So it’s more of a. It’s not just like kicking off an agent. It’s more of a like, I don’t storefront thing where it’s like, hey, here’s the ticket. I want this thing reviewed. Okay, now it’s now here it is. It’s back. No one needs to change anything, and the legal team just sits in Legora and like manages the agents and everything there. I think that the data structure for this kind of thing is really interesting because it’s like basically the inbox, right? For all these kinds of queries that can potentially translate into a thing, into something else, into something else, and like I don’t know, is it hard to integrate all these with all the other features, or do you just not worry about it? It’s no, it’s not. I think we have some primitives that work really well. So like lists is a great primitive because it’s essentially it’s like extremely flexible. It’s also like a notion database. Yeah, a notion database. It’s like a notion database, and so you can basically configure it for whatever you want. And we of course invest a lot into making this super scalable and work really well. And you can have triggers on them, automations on them, and it’s like a great glue to glue everything together. Because then you can kick off an agent run based on a list if something happens, and then that comes back and populates a field or whatever, adds new documents or whatever it is. And so it’s like just like you might coordinate a bunch of coding agents with linear or some sort of task management tool. We can do it similarly with lists, and so like agents can create lists, and lists can create agents, and agents can read lists, and so it’s like the agent swarm comes alive. Yeah, do people get excited or overwhelmed by that? You know, like I feel like this is basically end user.

[2400s - 2520s] Programming, you’re, you know, you can let people set up crazy automations. Some people get super obsessed by it, and they become like influencers, like you know, it’d be cool to have like, yeah, like Clay has influencers, yeah, yeah. Is that worth it? Too overcomplicated. We need the the flexibility, even if we don’t actually expose it to the end user. Let’s say this was just all internal. Then it’s a really fast way for us to utilize people, people like our legal engineers, to basically build a lot of products extremely fast. So like, you can build custom like fund management dashboards. You can basically build that on list with like your own a generative UI, and then you basically have a new product that some people are gonna love, and we can have a market place of just like a bunch of specialized products that we can scale super quickly because all of our legal engineers can just build new ones if they’re out with a client. Oh, you want your own? Easy, like we’ll just fix one for you. So it’s like a lovable inside the door a little bit. And if we don’t want to expose it, that’s also how we build our own custom feature. So we we made another acquisition in the UK called Wexler, and they’re doing fact intelligence, which is for litigation mostly. Which is like you upload a million documents for a lawsuit or whatever, and then we extract facts from those with like named entities and dates and whatever. And then when you have all the facts, you can basically hold them up against each other and see which ones are corroborated and which ones are what to call disagreement. They, you know, and that’s extremely powerful because then you can also transcribe a court hearing and you’d be like, oh, he said this thing, but actually the email over here says that he was not there. And it’s like if you see it live, it’s like so fucking cool. And that we can also build on this primitive, and so like we can build a lot of really cool things on this one primitive, basically. That’s how we think about it. Like we have to be, as we scale the company, we just have to solve so many things. And like so, so I think a lot about how do we build a core primitive that we can scale to to many different end products and create very customizable experiences. Okay, so you have what what core primitives do you have? Let’s just restate for for people. Lists is a core one, yeah, super configurable, yeah, project.

[2520s - 2640s] Basically, containers of files, file systems. The agent, of course, you can kick that off and run that in very different configurations. Do you have a name for the agent, by the way? No, we don’t. No, just intake. No. Okay. Yeah. No, you mean like we called it like Leya, or like we gave it a name? No. Yeah, Leya. Yeah, Leya would be cool. Yeah. People used to talk about Leya as like an a named entity. Yeah. But yeah, we we stopped that. We might well bring it back. It might be confusing. I don’t. So list agent tabular review tabular review, which like we’re considering is that actually an instantiation of like you can take this as far down as you want, right? But those are probably the big the big ones, yeah. Yeah. So I actually wanted to talk a little bit about like FTEs because this sounds kind of like that, and I think Max has talked about that a little bit. That like okay, when you onboard a new like big customer, you want to get them value like ASAP so that you know they stay on stay on Legora. Yeah. Yeah. How has that evolved? Like, when did you guys start that forward deploy team? So we hired our first forward deploy in New York. That was our first hire in New York, also a YC founder, nine months ago. Maybe I’m not actually sure. Are they considered legal engineers or no? No. Okay. Pure engineers. But we do. It is something that I think about. Is like how do we exactly enable that at scale? And I think a really core thing is like how can we empower our legal engineers? Like how can we power our power users and our legal engineers to customize the product in a powerful way? And then of course we we will still have forward deploy engineers like help out you know do all the custom integrations and the the scoping and figuring out what’s possible and feed all of that back into core product development. So that’s kind of how I think about is like almost a first line, second line defense. Where it’s like legal engineers should be able to customize a lot of stuff and build a great experience. And then where they can’t, or we need to make core product changes or custom integrations or whatever, we have forward deploy engineers who also then cycle in a little bit behind. Like sometimes they’ll just work on the product, and sometimes they’ll work with clients. Can you walk through like an example? Like, okay, so you just got you just got this new client. Like, what’s step number one? Like, are you guys like okay, hey, we

[2640s - 2760s] A forward deployed engineer for you guys, they’re gonna go build some kind of product for you. Like, how does that work? So it’s very different client to client, right? Yeah. But if it’s a big client and they’re super engaged and they like want, they have all these dreams of things that they would want to do, or they already have. Typically, they have already tried to build something and it’s failed, and they’re like, “This is what we. This is a problem that we’d like to solve.” Then we’ll we’ll basically staff them with legal engineers and a forward deployed engineer potentially, depending on what you know their thing is, and then we’ll just be like, “Hey, these guys will come. They’ll make you successful. They will sit with all your lawyers and teach them how to use it, and like figure out all the workflows that you have right now and build them into Legora. And some of that will just be legal engineer, and some of that will require a forward deployed engineer. What is that separation? I guess if you could like talk about that a little bit. Legal engineers configure the product as a power user would. They don’t write code. Yeah, yeah. And a forward deployed engineer writes code into production, basically. I guess that’s the that’s the. What kind of background do the forward deployed engineers normally have? You said YC founder. Yeah, founders, regular engineers. They’re old. Yeah. D was your yeah exactly. D was our first yeah two YC companies and worked on search at Meta. Yeah, she’s awesome. Yeah, yeah. Okay, that’s a high bar. She’s great. Yeah, yeah, yeah, yeah. Yeah, first first boots on the ground in the US. Yeah. So what are some examples of I don’t know workflows or things that these people are building for the clients? Are do you guys have like oh these are like you know there’s a couple use cases that like all the law firms are like you know this is the first thing that RFDs are gonna do when they when they go in. It’s very different. Some core workflows everyone have, and we can just build in Legora, and that could be anything from like this is how we handle these types of NDAs or whatever, and then it’s just like you activate it, and it’s more like skill building really for at least, and most of the FDs, that’s more like there’s something that we can’t do yet, like we need some more flexibility in some of the primitives or whatever it might be, or we just have this really complex system that we need to integrate into, and then they’ll go out there and they’ll build the integration against our APIs and maybe expand our APIs or whatever to make that work.

[2760s - 2880s] How does the like legal’s structure work there? Because you said you guys yourselves can’t even work on deals. Yes. Yeah. So like, how how do they work on client projects? You mean the law firms? How do they work on client projects? Sorry. How do your FDS work on client projects? So they don’t touch the data still. So they’ll just build the system. Okay. But they will also be onboarded like an employee sometimes. So like they might get a laptop and everything. Like to the lawyer. Yeah. Yeah. Oh. Wow. Interesting. That doesn’t happen very often because like usually they’ll just build an integration and like kind of works. But they yeah they can if they need to. And you only onboard FDS into a customer. I’m assuming if they’re a certain size. Yeah. Yeah. Okay. It needs to make sense because our FDS can also work on core product, right? So it’s like it’s a constant prioritization thing. Is there a rule like I guess does every like you know so Palantir famously had like you know four days you spend on site, one day you come back and contribute back to the products, anything like that? No. We don’t. To be fully honest with you guys, it’s still it’s still you know still early days in a lot of this. I mean, where do you see it? How to maybe over the next year? I do think that it’s going to be more maybe not exactly like Palantir, but like we I think we will have a larger in constant circulation group of FDS that way we sit with a client for like two months or whatever to build out this one thing and then they go to the next one, etc. Etc. Yeah, but I do imagine us rolling in with legal engineers and forward deployed engineers to like work on specific projects. Yeah, the API business is actually interesting. Like like you said, because you know that’s another rule that you can have where you are CTO maintainer of the Legora platform. Yeah. Right. Yeah. We will offer everything that we have in the Legora web app. Yeah. As an API, and then you guys can use whatever. Is is that is that a promise that you make? Or that’s not something we’re promising now. Yeah. Um, we do have APIs that you can build against, but not the entire surface of the product. That’s the curious thing. Like, is that the strategy? Or and obviously you know some people will love it because I don’t have to learn a new tool. I can just use the existing tool I have, right? And then I just like.

[2880s - 3000s] Yeah. Install Legora, for example, with the the word document stuff. Other than that, you know, I I do think like then you have to maintain it. Exactly. That’s that’s so that’s been my thing. That’s why we’ve tried not to do it. We haven’t we haven’t had APIs. I mean, we started our APIs like six months ago or something. We’ve fought it for a very long time, mostly because yeah, I just don’t want to. We we move so quickly in in the product that I don’t want to be like slowed down by having to maintain old API versions for like a data model that we’ve changed. I mean, now we’re bigger and things are different, so now it’s it’s a cost that we can take on. But like when we’re twenty engineers, it’s just like we we move so quickly that we can’t do this. I’ve toyed with this, right? Like, it’s just generally good for all companies to have a platform team and a product team because the product team you have one, but then you can easily spin up a second one as long as the platform is not super tied to product. Yeah. Yeah. Yeah. You know, like a very simple one is web app, mobile app, right? Yeah. Or like web app, mobile app, RPC, and and then like you know whatever like other other integrations. Then your FDEs just are one more client to the overall platform. Anyway. Yes. Yeah. And I think like even the agent, you can think about it this way. There’s like a harness team or like a that runs the agent, and then you can even you can build a product around it with like scales and automations and whatever else. Yeah. This this is like very famously the Amazon web services mandate. Like, Bezos was like, guys, no more backdoor access. You only offer other teams your APIs, the same APIs you offer customers. If you don’t have it, you build it. Yeah. And then the customers also get it. Yeah. I think I think that’s also changing a little bit. A lot of um, a lot of APIs now, like one thing that we’re toying with is like, can the agent be the API? I know it sounds weird, but like for some of this stuff, what they actually want to do is they want to trigger certain things to happen. For some part of this, it could you could just be. I’m not saying that we’re doing this, but it’s like a thing that I’m toying with in my head is could that just be like okay, you trigger the agent and ask it to do this thing? Like, obviously, it’s way less efficient than if you just had like an API that you always hit, and it’s like you know. But it’s something that I’m toying with is like how much of our product can just be the agent, versus all the other stuff. Well, you know, unbounded execution.

[3000s - 3120s] And costs and latency. That’s the that’s is that the only holdback? That’s yeah, and and I think deterministic. Like maybe in you know a hundred out of or in two out of a hundred times, it’s going to do something whack that we didn’t want it to do. We’re going to go into like harness and model fine tuning in a bit, but I think there’s a there’s a product question about restraint about what you don’t do because you could do all that, and obviously there’s a prioritization issue. Like, are you more at the point of let’s say no to most things and then only do a short number of stuff, or no, actually we want to do more things and we just need to hire like 100 people? I was I was very strongly the first thing for a very long time. Let’s just say no to everything. Like keep the engine team. I I think I’ve said this before, but like one of our presentations that like after I’ve been at Legora for half a year, I I pulled up this slide of um you know the 300 Spartans and and for the movie, and I was like, we’re just going to be 20 engineers for the rest of the company’s history, and like we’re going to beat everyone who’s much bigger. And we’re 100 engineers now, and it feels like we’re way too small. I wanted to do that, but I think in the real world, we’re in like fierce competition. People want so much stuff. It’s like I could be a puritan and just be like, no, we have to stay 20 people, and we can only build you know whatever. But then we just lose, and like people would go elsewhere. So we have to build more. And so I think my job is to figure out how do I build a system where we can build a lot, but actually only like where we can the customer feels like we’ve built so much stuff, but we’ve actually only built few things. That’s that’s how I think about my job right now. Yeah, I mean primitives is right, you know. Yeah, exactly. I think that’s going to be that’s really powerful. Do testing new products, just seeing if like hey, there’s like any pull here, and then you just kind of like you’re like no, I just focus on our main thing. And we haven’t made. I mean, we’ve made certain strategic bets. Yeah, but we’ve not tried anything that we don’t have a lot of confidence in already. Mostly because we just have like a mountain of demand. Yeah, it’s like we want all of these things, and so it doesn’t hasn’t really made sense for us to try to be like, hey, what maybe this thing could work because like there’s so many people screaming for the other things, and so we mostly build those.

[3120s - 3240s] And then sometimes we’ll disagree with clients, and we’ll be like, “Actually, I think this thing over here will make a lot of sense strategically.” And then we’ll try to put that anyway, like strategically for you. Yeah, exactly. So, what are those things? So, one of those things was the agent actually. So, like, no one was actually asking for like, “I want it to be.” I want the assistant that I’m using to have more agency and do things without me asking it. Like, no one, no one ever. But we’re just like, this is definitely where the industry is going. This is going to empower you to do so much cool stuff. Collaboration was another one where, like, we think the industry is going in a place where you can collaborate inside of where you’re working, which is Legora, and not in like emails and text messages. So, yeah, there there are there are a few of those. And then how like once you say okay, yeah, we want you guys to use agents. I’m assuming like internally at the law firm, you know, it’s going to take a lot for them to actually learn how to do that. Yeah. And so that’s where you kind of like say, “Hey, FTEs, go in there and teach them what they.” Yeah, legal engineers mostly. Oh, legal engineers. Yeah, yeah. That makes sense. But they’re like some people call them forward deployed legal engineers if you want, but that’s essentially what they are. Then they go in, yeah, and help and help them structure their workflows around the capabilities of the agent. You guys said previously, I think Max said this, but that early on, specifically, you guys weren’t fine-tuning your own models. Is that still the case? Yeah. So we haven’t fine-tuned any models up until now. I think now, in the past few months, open-weight models have become extremely capable, and so we will do it for certain tasks where we think there’s a huge delta between what a frontier model can do at the cost and latency versus what we think a post-trained model can do. Legal research is a great example. It’s like a you have to make it’s it’s a fan out retrieval problem, and you can make a really effective model to those queries super quickly and super cheaply at low latency. So specifically, you’re post-training them to be better at very specific bounded tasks. Yeah, exactly. Okay. Yeah, we’re not trying to like train our own fable. Yeah. To like be the main agent, I don’t think that’s like a we won’t win that you know versus anthropic most.

[3240s - 3360s] Yeah, there’s still a lot of let’s say search and retrieval research or rag research to do. Yeah, that is like maybe less glamorous. It’s not actually even model training. Sometimes it’s just like architecture and pipelines. You you you did a case study with TurboPuffer, but like you know, there’s also like do you do late interaction of yes, absolutely. Yeah, you do that. Yeah, we actually don’t do that, but it’s you’re completely right that this is so. There’s one guy on the team, son, who’s he just really cares about search, and he he basically comes to me all the time, and he’s like, we should do these things, and like this is new embedding models, and like we can do this late and late interaction. He actually talked to TurboPuffer team to get them to build multi vector so that we could do these types of things. But no, there’s loads of low hanging fruit generally even before models, and even in the harness. Honestly, like if you just want to create a better output for our clients, there’s like so much stuff that we can do that has immediate impact before training models. Like what makes a good harness for like the law industry? I guess like you guys are making. I’m assuming like your own custom harness. Yeah, yeah, yeah. The way that we edit files is important. The way that we do an analysis is important. The way you do legal research is all baked into it. So it’s like you might have a problem that comes up or see something, then it knows how to go and do the legal research to figure out what the answer is, what the precedent is. That’s all very important, and then of course there’s all the governance stuff around it as well. Yeah, it seems like the search is probably like would you say the most important tool that you guys are like customizing because it’s if you can’t pull in the right information, yeah, and then you’re screwed. Yeah, yeah, yeah. It’s yeah, I think it’s it’s it’s you guys are searching like a lot of and analysis. Yeah, yeah, it’s search, it’s analysis, and it’s drafting probably because like drafting word, drafting huge contracts is actually also not super trivial. Like what about it is not because you have to let’s say you’re drafting a huge contract from scratch, then you have to find precedents of like how did we do this previously? Do we need this type of clause? A specific law firm? How did they do it previously? Yeah, so I’m I am the law firm, and I’m about to draft a huge whatever NDA is very simple, but like there’s there are contracts that are way more complex, and then you need to find the right precedent.

[3360s - 3480s] First, because like lawyers just don’t don’t like sit and like start typing stuff. They they figure out what clauses have we used in similar deals, and similar can mean a lot of different things, as I’m sure you know. In search, it could be like same industry, same type of companies, same whatever. And so you need to identify the right person. And you might have like multiple documents that you then try to figure out which part should I take from what document, and that’s actually non-trivial. So it’s essentially also a search problem. Which models best at drafting? So I’ll say I told that farfetched yesterday. Definitely, Fable and Open are extremely chatty. I’m sure you guys have known this. They’re very verbose. So that’s not good. That’s generally not good. Okay. I’ve noticed Astra is pretty good at writing. Astra is yeah, pretty good at writing. Much less more concise. Are you running like a hundred different prompts to like yeah be sure we can yeah we can we can do we can do certain things, but I think the chattiness verbosity is actually a little bit difficult for us to do. So this is I think you know let’s see, but probably the next model from Anthropic is going to be much better at this. Astra is actually very concise. Yeah, yeah, and I assume they they all have a Slack channel with you, and you’re telling them this right? Like exactly, yeah, this is not news to them. No, no, no, they know this. Then the other thing that I really care about is hallucination. Do they make stuff up? So we have a system of a citation system where we basically check what they’re writing has a link. So the user has a link they can click, and it highlights with a bounding box, like in the PDF or whatever, where where this information was taken from, and we can sort of make sure that what it says always is citable somewhere, even multiple layers down. So like if the agent writes something based on a tabular review, you know it can cite the cell in the tabular view or the row in the tabular view, so you can actually like click multiple layers down and find you know sort of validated all the way up. Yeah. Are you finding that lawyers are like verifying that these citations are correct or not? Initially, initially, some of them were, but now yeah, exactly the same thing. Are you actually are you actually people shit like hallucinated links to Supreme Court? They don’t care. That happens. Yeah, exactly. That happens. Yeah, I I think it’s like coding. You know, you have engineers that just don’t give a shit and just like.

[3480s - 3600s] Agents and you have people that like sit and read every comma. Same. There’s just like so much that like how can one human do that, right? Yeah, but I think they probably just like we know in code review what to look for. Yeah, what is the system or whatever. I think they also know. Okay, this is the the thing. These are the red. Like this is the thing that’s important. Let me just like double check these things. And like these sound reasonable, so it’s probably fine. Okay, interesting. Just like if you if you receive something from a junior lawyer, it’s the same thing you do, right? Just like I should double check this, and like this probably is fine. Any other elements of let’s call it legal reasoning or what’s the the legal frontier that models are not good enough yet that you want from the next generation of models? Nothing I can think of specifically right now. I think long document reasoning is still difficult. Long like million token token context. Yeah, exactly. You have a huge contract. Yeah. And like you’re making edits across it, it’s a difficult thing to do for a model today. And like remember that you changed all these things up here, and like the definitions changed, and that you the ordering changed, and whatever else. So the simplest would be one shot. You put feed everything in, and then one shot output. Yes. You don’t do that. You break it up into chunks, and then you orchestrate edits. Yeah, exactly, exactly. And this is like one of those things that are funny, like that are fun to be part of in this industry, right? So our word editing was like we we’ve done you know we’ve done map reduce and we’ve changed them. There was one big one, and there was like one big one in multiple times. It’s like we keep changing the the scaffolding that we build around as models change. Yes, I I would say I um I recent I do a lot of like news reporting obviously because of what I do and I’ve recently shifted from map reduce to throw everything into context. Yeah, and because now there it’s better. Yeah, exactly. But it’s also it’s different. I mean, we had this problem with search where it’s like you’re searching for certain things in a huge contract, and you can map reduce it, but then it might if you’re summarizing any sort of thing, then it might lose some important nuance. Yeah. In between the chunks, essentially, if you just summarize the individual parts up, and so we this has been a problem where it’s like we can do two types of search. We can also do the other one where it’s like um you need to have reasoning in in the different steps there in between. So I’m trying to think of a really good example.

[3600s - 3720s] Here, yeah, there’s there’s some documents where they talk about something at length over like let’s say two pages or whatever, and then each individual chunk is not necessarily relevant to the search term or what they’re trying to figure out, but together it’s extremely relevant. And so, yeah, these are some of the you know problems where it’s and it might be hard, like you said earlier, to actually make the models better at that because you can’t even look at the contracts, right? It’s the real contracts. Yeah. So do you guys ever ask them for like, hey, can you make example contracts that are similar to your real ones? Yeah, yeah. So we have we have feedback and we receive feedback from them, and if they’re willing to engage, yeah, exactly. Or they’re just like, hey, this is like super important to us and it’s not working well, and then we’re like, great, can you like anonymize the documents and like the problem whatever, and we’ll if we can recreate it, then we can fix it. Yeah. I mean, so one version of this is you buy a law firm. Yeah. Which some people are considering. Yes, but there’s some we there’s some stuff there on whether their data is actually their data or if it’s the client’s data. So if the clients are all gone, then maybe it’s fine. I actually don’t know. Interesting. Yeah. But it’s uh yes, it’s it’s a way to get a lot of data. Yeah. Yeah. People like Google is buying Spirit exactly. Yeah. Same thing. Yeah, they yeah they bought the yeah yeah for like ten million seven million or whatever. Yeah. Like you know, if you have a failed company, you still have your Slack. You can sell Slack. I think there’s like there’s like a YC company like buys your failed company. Yes, there is. Yeah. Yeah. Yeah. Because like it’s just like oh, I just want like real people doing work. Exactly. And then just like exactly stuff out exactly the problem with buying law firm data is well you get all the inputs the real inputs which is great the outputs are a little bit more difficult because you just get what their end result was and it’s difficult to figure out if like the way you have to basically like reverse engineer how they got there. So in our product when we get real traces from real users it’s extremely high value, but if we just got this output and we don’t know if it’s good or bad or how it was created it’s a little bit difficult. We’re going to wrap up soon. The last thing I’ll point people to is also there’s a lot of alpha left on embeddings. If you’re if you’re a rag heavy work workload, you’re

[3720s - 3840s] You’re just looking like MTEB, and you’re taking like top model. You’re leaving a lot on the table. You should look at like Voyage or like some of the custom embeddings. You’ve looked into this, yeah, yeah. So basically, don’t take default embeddings, even like OpenAI embeddings, Gemini embeddings. Doesn’t matter. Customize your embeddings. Yeah. Right. Yes. You you support that? Yes, I think that’s right. I think that’s right. The only thing is, need to make sure that you’re multilingual. So for us, we we serve so many languages. So that’s I mean, Europe is like king of this. Yeah. Yeah. Exactly. Mr. O, how old there? No. Until well, until now, they just raised a ton of money, right? So maybe. Yeah. Let’s. I mean, you guys. They also raised a ton of money a year ago. Europe champions. I think we need way more model companies out of the out of Europe. Yes. I just wanted to ask you, like, what predictions you have, maybe like over the next year for like the law industry. We we asked a little bit of that earlier. Yeah. I’ll modify it to your second question, which is: in two years, will lawyers do their work inside Legora? Will they use other tools like Claude or ChatGPT and call Legora, like you know the API? Yeah, yeah, yeah, yeah. I think both. It really depends. I think if you’re a big enterprise, you might have Claude co work, and then you’ll. It’s. I think it’s similar to like how I use Linear through Claude, but I still go to Linear. I think enterprises will do that. I think law firms will just use Legora because that’s simpler and better. Yeah. It kind of reminds me of like OpenEvidence for like medicine. Yeah. Like you’re you’re basically that for law. Yeah. Yeah. Exactly. Coding has gotten to a point where people the labs will cut you off for competing with them, and this has happened twice. Yeah. Do you do you think that will happen to law? No, I don’t think so. I think because like you know they are releasing legal services. Yeah, yeah, yeah. They are. I don’t. I mean, we obviously talk to them a lot, but it’s also it’s like, hey, are you gonna? Your sales reps will say no. Exactly. They don’t make the call. Exactly. That’s not their. Right. And they don’t even know like often what’s actually going on. I don’t. I don’t think that’s the case. They have so much stuff to capture and knowledge work, or in in sort of enterprises knowledge work generally, that.

[3840s - 3960s] Like legal has to be part of that, just like data analytics has to, and sales and marketing and finance. So like they have to do things that are good there, but they’re never going to go as deep as we are. And I think if if Claude takes small clients, small you know corporate clients, and they do some of their legal via Claude, like that’s that’s great, that’s fine. But if you’re doing a you know 400 million dollar M and A transaction, you’re going to want Legora. And kind of an unrelated question, but I’m curious: Are you guys going to law schools and kind of like getting them to like teach Legora like in classes? Like, how does that work? Yeah, it’s a it’s definitely a program that we are. I think we’ve started. I actually don’t know exactly where we are, but we we do want people in law school to learn to use Legora. Like, are you seeing that like law schools are changing their curriculums a lot or not really? I don’t know. I I I wouldn’t know actually. Interesting. I would actually do it a different way. You should do scholarships for for law school for CS majors for engineers. So we will cover your after you’ve done CS. Yes, we’ll put you in law school. CS plus some software experience. Yeah, we’ll put you in law school. Yeah, we’ll cover it. Or you do like the MA like like masters plus JD programs. Yeah, they don’t care about the masters. Yeah, meaning like because there’s a lot of people who didn’t do CS undergrad, so then they do like masters just to kind of like get caught up. Yeah, yeah. I I don’t know. I don’t see that as as often. Yeah, yeah. People do JD MBA. That’s a thing. Yeah, but JD CS, I haven’t seen it. Yeah, happens. Yeah, yeah, yeah. Okay. Look, overall, you know, super impressive progress. I mean, congrats on all your success. Like, are you still going to keep up the pace? Like, absolutely. Yeah, you know, and I think you’re hiring in SF, New York, Sweden, and yeah, engineering. We’re hiring in New York, Stockholm, and London. Yeah, and all other roles we’re hiring everywhere. How many people do you have in New York right now? Hundred. Okay, and you guys are planning on how many in New York? In New York, yeah. For just like generally, let’s just say yeah, generally and.

[3960s - 4042s] A lot, as many as you can. Yeah, it’s it’s aggressive. Yeah, I mean, that that’s the whole corporate law, right? As far as like, yeah, Middle America’s concerned. So, like, yeah, I mean, I’m looking forward to you buying a law firm and like, you know, becoming, yeah, yeah, it’s it’s super exciting to to. I think like you know every industry, especially like the top like doctors, lawyers, coders, you know, they all have to have like you know two or three agent labs that like just yeah, so you dominate the space. Agree. So you guys have like a great market. Yeah, yeah, I think so, and I think it’s only going to get much bigger as like law changes and get becomes more available to more people. Yeah. Also, I think like I mean, we’re already like super over, but like the problems that you work on are very like the frontier knowledge work. Yes. Yeah. They’re okay, like unless you’re literally like doing like science, science, navier Stokes and science. Other than that, like most work is like basically converges to legal. Yeah. Like. Yeah. Yeah. It’s all legal in the end. Yeah, hiring, finance, everything. Like, I just have to consult my lawyer. So. Yeah, I know, I know, which is why I think it’s a really exciting thing to work on because, like, modern society runs in some way on legal. Yeah. Yeah. Cool. All right, that’s it. Thank you. Thank you. Thanks.

中文翻译（全量）

[0s - 120s] （冷开场拼贴）大家都在问我们 Mythos 之类的新模型，尽管 realistically 大量法律工作并不需要 Mythos 级别的智能。「律师助理消失了。初级律师消失了，不需要他们了。」「我们实际看到的并不是这样。」我们曾有一个引用格式的问题，下个月就解决了，从大概百分之七十直接到九十九，「我们什么都没干，挺好」。但如果你在做一笔四亿美元的并购交易，你还是会想要 Legora。「我们就永远保持二十个工程师，然后打赢所有比我们大得多的对手」，现在我们一百个工程师了，反而觉得太小了。还有一个我一直在琢磨的问题：我们产品里有多少可以干脆就是 agent 本身，而不是其他所有东西。——正片开始：我们请到特别嘉宾，Legora 首席技术官 Jacob Lauritzen（转写作 Lauridsen，节目内确认发音随意，主持人问姓氏有无特别来历，嘉宾解释丹麦语里「-sen」就是「之子」的意思，「Lauritzen 即 Lauritz 之子」）。主持人称 Legora 是「史上增长最快的法律科技公司之一」。嘉宾谈到公司起家「非常欧洲」：从斯德哥尔摩地下室里最小的创业公司起家，一直是 underdog，没人相信能赢过美国遥遥领先的巨头，所以底色谦逊、一直要为生存而战；进入美国后不得不学着更 bold 一点，向世界宣告自己有多好——这对他们来说并不自然。

[120s - 240s] 主持人请嘉宾介绍产品与迭代历程（很多法律产品藏着掖着不让看）。嘉宾：一切的起点就是「带瑞典法条搜索工具的 ChatGPT」，字面意义上这就是全部。从法律研究起步——律师花大量时间弄清「有哪些先前的案子像这个、法律先例是什么、以前的人怎么处理」，这是个巨大市场，一直是产品的一部分。最初就是一个 agent：你提问，它检索法条。之后一路扩展：起草能力很强、编辑 Word 文件、整理归档文件、跑复杂分析。其中一个功能叫 Tabular View（表格视图），这个名字现在成了行业用语，看到 Claude for Legal 沿用我们的术语我们很高兴——本质上是对大量合同做批量抽取与批量审阅，尽调、融资时都用得上。这帮律师把工作做得更快更好。而除了「做工作」，我们现在比一年前多做了很多「编排工作」：一个叫 lists 的东西（相当于我们的 Linear）、intake（与客户协作的入口）、portal（律师↔律师、客户↔律师、律师↔agent、agent↔agent 的协作）。

[240s - 360s] 嘉宾：十年前（甚至今天）跟律师协作的体验都非常痛苦，改进空间巨大——所以产品就是两件事：做工作 + 编排工作。主持人问什么时候开始做 agent。嘉宾：我们经历了从 assistant 到 agent 的切换，大概一年前吧。时间线对照（主持人梳理）：YC——嘉宾自己是 23 届（自己的创业公司），Legora 是 24 届。那还是 agent 真正兴起之前，随后推理模型「改变了一切」。主持人（看了产品预览后概括）：它基本上是「律师版 ChatGPT」，这么说是公允的，但侧栏里有很多漂亮的定制工具，比如 Tabular View、intake，还有为企业内公司法务设计的流程（找即将出台的法规之类）。嘉宾确认：就是「真的很好用的、专为法律工作做的 co-work/ChatGPT」。模型层面：按任务混用多个模型，这一点律师看不到；agent 里有一个主模型选择器，但特定工具（比如 Tabular View）内部用不同模型——我们把问题拆分、按子任务路由最有效。

[360s - 480s] 嘉宾：我不确定这个选择器会不会一直在。主持人补充背景：我是 Cognition 的顾问，我对「agent lab」的论点是——理想的 agent lab 是「领域专用的 GPT wrapper」：我们是领域专家，会比前沿实验室更上心地服务你，按我们的判断把前沿实验室最好的能力端给你；我们的职责就是替你做决定，让你永远站在 Pareto 前沿上，不用操心细节。你去问律师「这份合同基于这个先例起草，你觉得该用 Sonnet-high 还是 Opus-mid」，他们不知道——「我们的时薪是一千美元，让它能跑就行」。嘉宾：我本来完全认同这套（这就是 Devin 做得不一样的地方），结果我们现在选择器里有五个模型，包括 Fable、Opus、Astra。大家就是想试新模型——都在问我们 Mythos，尽管大量法律工作并不需要 Mythos 级智能。于是「行，给你选项」，但你的积分消耗会快十倍。嘉宾：我们抗拒了很久，最后还是妥协了——你们要这种控制权，那就给你们。主持人接着追问：那你觉得模型已经「好到律师够用」的程度了？嘉宾：我认为是幂律分布——大概百分之六十的工作，现在的模型就能做得很好。

[480s - 600s] 嘉宾（承上）：就当下与可见的将来而言，剩下的是成本和延迟的问题；但确实还有另一部分——尤其是很复杂的分析——我们仍需要更强的智能。而且工作总量本身会随智能扩张（「法律之前的工作」也会膨胀）。主持人问早期转 agent 时用什么模型。嘉宾：大概是 Opus 4.6、Sonnet 4.6 时代的模型。当时的 agent 不行，现在好多了——虽然才过了九个月。子 agent 委派是进步很大的一项：fan-out 型检索现在做得非常好，以前很难。我们有专门的评测团队，追踪所有模型在不同业务领域（法律的子门类）× 不同任务上的表现，有一张巨大的热力图，能看到模型随版本移动——通常什么都在涨，但其实很多次发布是「这几项涨、那几项跌」，这比外界想象的常见得多。评测环境长这样：我们有不同的「世界」/环境——律师基本都在「项目」里工作，比如「这是一万份文件的数据室，你在做尽调，最终产出一份包含全部红旗、简洁得当的报告」；任务里一部分是起草报告：草稿好不好、内容全不全、措辞与分析对不对。

[600s - 720s] 嘉宾（续）：另一部分是找红旗——即起草之前的分析，评测就这么拆开。主持人问评测团队构成、是否雇律师。嘉宾：是混编的，工程师和律师在同一个团队，现在三个工程师加三个律师；另外还有一个「legal engineer」人才池——在大所干过三四年、通常有 CS 学位、非常 tech-forward 的律师，来我们这因为想改变这个行业。主持人惊讶：他们是律所的 IT 吗？嘉宾：不是，是正牌律师、带 CS 学位。主持人：「这就是个维恩图啊。」嘉宾：这是我们最难招的人，但也是力量倍增器——一个上百人的专家池，天天用产品、天天见客户、亲身活在这个行业里，评测和产品打磨都能拉他们进来，这是巨大优势。主持人：那这支评测团队什么时候建的？嘉宾：从 assistant/agent 团队里长出来的。直到不久前，认真做复杂评测其实不划算——所有模型都在涨，是花大力气做评测，还是直接建更好的 harness、搭模型进步的顺风车？那时 harness 红利大得多。主持人：「每两三个月就换一代模型。」嘉宾：对，直接透传就行。我们曾有个引用格式问题，写出来的东西不好 follow，下一次模型发布就解决了，百分之七十直接到九十九，「我们什么都没干」。主持人：所以现在你们也在做评测了。

[720s - 840s] 主持人：我猜你们对开源模型的使用也在上升（这个我们待会儿展开）。嘉宾：早期的工作与其说是「做产品」，不如说主要是把模型能力直接暴露给律师；现在更大的命题是想清楚：五到十年后，模型能力持续涨、价格持续降的前提下，法律行业会变成什么样。编码行业大概领先一到两年，可以参照。未来的原语是什么？大家怎么工作、怎么协作？如果不是两百封邮件带着 V5 版 Word 文件来回飞，那是什么形态？这甚至不纯是 AI 问题，而是把产品体验做好——并且要带着审计日志、数据保留这些监管行业的关键约束。AI 的部分则是让工作更快更好、加上 proactive agent：事情进来就自动反应，不用事事人工发起。也许最终形态是「指挥中心」：你登录进来，系统说这五个事情是 agent 最近在做的、需要你回应、检查、改进——也许我们正走向那里。主持人：那我能深入问一下吗——几年之后律师这个职业会变成什么样？

[840s - 960s] 嘉宾先声明：「我真不知道，谁能知道呢。」但可以参照编码行业：律师的产出会大幅提高——更低的成本做更多、质量更高的工作，行业本身会变好。律师会把更多时间花在他们当初入行想做的事情上：判断合适的风险偏好、做战略判断、和客户相处；更少时间花在翻大文件找「冒烟的枪」这类苦活上。主持人追问合伙人模式会不会变。嘉宾：我上周还真认真想过。合伙人制的存在，是因为这类公司唯一的资本就是人力资本。不像我们或者软件公司、制造商有实物或无形资产给股东分红，专业服务行业（律所、咨询都一样）的全部价值在人身上，而你没法「拥有」人——奴隶制都废除了——只能「租」：付钱雇他们。合伙制本质上是一种让人力资本成为公司主人、从而留下来的结构。嘉宾（续）：这个结构非变不可吗？不一定。我们真正想做的，是帮律所建立某种无形资本、可持续的资本：建一个「学习你们怎么干活」的系统。很多律所都在想「本所有本所的做事方式」，我认为这可以被 codify——建成一个自改进的知识图谱加判例库、文档库，把「能稳定产出好结果的工作方式」变成律所自己的 IP，也就是把人的专业知识和判断力固化进系统。嘉宾自嘲「我彻底跑题了」，但这是他们想赋能的方向。

[960s - 1080s] 主持人点破矛盾：那合伙人制的激励就不同了——现在的模式下合伙人只需要拉更多客户，没人愿意牺牲拉客户的时间去沉淀知识，因为「eat what you kill（各凭猎杀吃饭）」的分成制度奖励的就是拉单。嘉宾：需要设计新激励，让合伙人愿意为公司沉淀资产；现在没有——合伙人离职时，他们帮公司建的资产一分红利也带不走。所以整个商业模式「都在被重新审视」。主持人顺势抛出底层问题：「那低端呢？paralegal 消失了、初级律师不需要了？」嘉宾：「我们实际看到的不是这样。」主持人拿工程界类比：初级工程师「真的有点熟透了」，你得从大学出来就是完全体，否则就难了——「我为了喜剧效果说得夸张点」。嘉宾：我认识很多律师，从 paralegal 起步、再去读法学院——这些起步路径现在不需要了吗？那些纯执行的工作确实不需要了。但更重要的是：一个好律师（好初级）今天比两年前值钱得多，跟工程界一样——强者恒强，两极化加剧。

[1080s - 1200s] 主持人问计费：法律行业按小时收费，项目做得快了会不会整体变便宜？嘉宾讲了个故事：我当年帮 Max 改他的 YC 申请（那时公司还叫 Leya——主持人猜名字是不是丹麦语/瑞典语里有含义，嘉宾：都没有，就是个名字；还分享了一家西班牙报纸早年把 Leya 解释成「legally efficient Yoda assistant（合法高效尤达助手）」的缩写——「幻觉不止出在大模型，记者也幻觉」）。我当时给的唯一反馈就是：这事别人试过一百万次都没成，法律科技从来做不起来，因为计时收费下律师没有提效的动机——效率越高收入越少，不合理。但现在律师愿意采纳，是因为他们明白事态的严重性：商业模式整体承压，竞争压力太大，不变不行。计费本身也在变，尤其美国：不少工作已经按项目费而不是小时费计价，尤其是光谱低端的小型 review 之类；有的干脆免费做。主持人：那钱从哪赚？嘉宾：并购（M&A）——「这是最容易举的例子」；诉讼也仍是大头、非常有价值，律师会把更多时间放这上面；客户关系也极其重要。

[1200s - 1320s] 嘉宾（续）：很多律所历史上就愿意免费干活，只为了维系客户关系；现在他们还能这么做，而且不用亏那么多钱。主持人开玩笑：「我还以为 retainer（预聘费）就是我花钱请律师做朋友呢。」（笑）主持人追问：所以战略型工作和高价值工作还在，法律费不会降，但 paralegal 不是消失了，是转移了？我们做 Cognition 时也遇到过这种表述难题——「我们将取代你的初级工程师」没人爱听，但实际上部分发生了，另一部分初级工程师拥抱它、产出暴增。嘉宾：完全一致。主持人：这个新角色有没有个流行叫法？「超级 AI 加持的初级法务」？嘉宾：目前没找到，也许我们该造一个。「legal engineer」算吗？——那是我们的词，现在也成行业词了。主持人：「CSJD（计算机科学+法律博士）」。嘉宾笑认。随后聊到「meet proxy（肉身代理，存疑词）」这个贬义词：坐在机器前、别人让你输什么你输什么、机器说什么你复述什么。但嘉宾认为一个「很好的肉身代理」——真正懂 agent、懂 Legora、懂「每个行业都有自己的工具箱，我能用工具把你要求的事做好」的人——其实就不再是代理，而是在做有意义的工作了。

[1320s - 1440s] 主持人：聊聊 Tabular Review 的规模吧——几百、几千、几百万？嘉宾：单次能做到一百万格。主持人：「这个数字很吓人，我们待会儿得聊聊表格和窗口化。」嘉宾举例：我在办一个大案子，一百万份文件，往常的做法是「把一队 paralegal 锁在房间里干到凌晨四点」——现在不需要了：机器批量抽取，你随时可以整体重新 prompt。所以「这个活我不再需要 paralegal」，但一个「超级 AI 化的 paralegal」会去做别的：做以前没时间做的事，或者干脆去睡觉，谁知道呢。主持人转入收购话题：你们买了家公司叫 Kira？还用自己的产品做了这笔并购的尽调？嘉宾：对，dogfooding，效果很好，我们经常这么干。intake 这个产品就是和自家法务团队一起建的——他们原来用 Slack 工作流收需求，天天抱怨「太烦了、太低效」，那就「拿 Legora 给自己做一个」。主持人：能展开讲讲当时的过程吗？嘉宾：尽调的大部分我们自己用 Tabular Review 等工具做，但仍会请外部律师评估特定风险（比如薪酬体系）。Kira 那笔（转写作「the career」，存疑）因为全在瑞典境内所以简单；之后做过加拿大的、澳大利亚的收购，涉及当地薪酬，就要考虑「这里要不要请人从战略层面看一看」——战略部分我们仍会用外部律所（转写作「Lawless」，存疑，疑为某律所名）当 sparring partner，帮我们想清楚交易结构；而大量琐碎工作完全可以自己干。我们自己就在尽调里发现过红旗，然后直接踩了刹车。主持人问还有没有别的细节，嘉宾说没有。

[1440s - 1560s] 主持人问早期客户是瑞典律所——瑞典所和美国所（甚至亚洲所）有什么差异？你们主要服务大所吧？嘉宾：起家时合作的确实是非常大的律所，但五十人规模的律所也能获得很多价值。欧洲客户整体动作偏慢，美国的推进非常快，而且 AI 在美国公司的议程优先级明显更高。美国的公司（企业法务和律所都是）自建胃口也大得多：他们想「跟你们一起搞个大动作」——自建 CRM、自建文档管理系统等等，然后做触发器：在 Salesforce 里做完一次 review 自动派给某个人、处理完流进 Legora、再回流出去。这是简单端；还有极其野心的——想直接和我们一起造一个全新产品，「我们真做不了这个」。欧洲客户则更偏向「拿现成的」，给产品建议。主持人问：是欧洲客户不觉得生存威胁、美国客户觉得？嘉宾：两边都知道这事关律所/公司的成败，都极其重视，只是实现路径不同。

[1560s - 1680s] 主持人转入产品指标：怎么知道自己做的东西是好的？嘉宾：常规的企业软件指标我们都很在乎——用户在产品里花的时间、有没有获得价值、续不续约、回不回头；对企业软件来说，我们的 DAU/WAU/MAU 和「每月在产品里的小时数」都高得惊人。另外我们做在线评测：用超便宜的 LLM 消息判断「这次会话到底解决用户问题没有」，形成一个在线指标；上新功能时用 feature flag 对比——「是改善了，还是灾难性下滑？」然后是嘉宾认为最有意思、也最难的一点：我们基本看不到任何 trace，看不到任何数据——全部高度保密，这对工程来说是极其憋屈的工作状态：你给用户一个什么都能干的 agent，而你完全不知道用户实际在拿它干什么。主持人：我以为你们总跟律所签了什么数据协议。嘉宾：「我倒希望。」问题是律所手里很多数据连律所自己都无权给——那是他们客户的。所以我们只能靠评测团队「逼近」大家的真实用法，把它做进评测集。「这是这生意一个超级烦人的怪癖。」

[1680s - 1800s] 主持人：能不能做改写（paraphrase）？比如 ChatGPT 有个 /feedback 直接上传完整轨迹，你们显然不行，但可以用 LLM 把真实内容改写成「改掉所有名字和细节、但结构足够像」的合成版？嘉宾：做不了。这里有个区间问题：如果你把内容抽象到只剩三个候选值、随机取一个，那没问题；如果只是去掉人名，那「可能不行」。我们一贯选最安全的路线，把自己限制到最狠，目标是「零风险」。但有些事可以做：我们知道失败模式大概就这十种二十种，可以在线上把会话归类到某个桶里，这就是有用信号；还可以按任务类型、业务领域分类，就有概念了。这是我们一直在想怎么做聪明的事。接着谈隐私史：两年前要让律所接受「数据发给 Anthropic」这种事有多难？嘉宾：我们全部基础设施都在 Azure 上，主要就是为此——起步时律所说「AWS？绝无可能」，Amazon、Google 都「不能把数据交给他们」，我们只能全线架在 Microsoft 上，因为律所已经用着 365：Microsoft 在他们心里是「已通过安全审查的认可供应商」。这几年风向变了：大家明白必须上云，on-prem（本地部署）的讨论也彻底消失了——「云上是安全的，甚至比你现在自己搞的更安全」。嘉宾感慨：这不过是三年内发生的转变。主持人：法律行业此前把这事晾了十年。嘉宾：「我们需要 AI 来踹所有人一脚。」

[1800s - 1920s] 主持人：开源模型现在很强，你们怎么考虑？嘉宾：就像你说的——把客户始终带着沿 Pareto 前沿走。在一张「所有模型 × 擅长什么 × 延迟 × 成本 × 治理 × 区域可用性」的大地图上做优化：某些开源权重模型在特定任务上很好，就用在那里；有些我们会后训练；有些直接用前沿模型。主持人：「你把这叫 model arbitrage（模型套利）。你们年初还是坚定的 Anthropic 阵营。」（另提到 Grok 进步很大。）主持人请嘉宾举一个「总体排名不高但意外地好」的模型。嘉宾：早期某个 Gemini 版本在特定类型的转写和说话人归因上极好，我们就用它做转写。主持人：你们要做会议记录类产品（类似 Granola）吗？嘉宾：我们其实可能会做一个 meeting 功能；这里说的转写还包括用户上传听证会录音等文件，我们转成文本后就能在上面跑全套分析。不用 ElevenLabs——DPA（数据处理协议）原因；我们的 DPA 名单上供应商非常少，因为必须把数据流向锁得极紧。嘉宾由此说到一站式厂商的好处：Gemini、OpenAI 这类「来我这，你要的大多都有」的供应商其实非常值钱——不必样样最强，只要每样都够好，还带治理、可扩展，那就够了。

[1920s - 2040s] 嘉宾：还得夸一下 Microsoft 的转写模型（转写作「my transcribe」，存疑）——他们做的这个比旧的 Gemini 还好，arena 排名通常都在前二到前四，「这就够了」：治理齐全、可扩展。主持人：语音转文本是不是已经好到「错一两个词无所谓」？嘉宾：不不。我在这上面踩坑很多，法律领域尤其：某个国家（好像是澳大利亚）法庭里对某类人的特定称谓，会被一致地转写成「condom（安全套）」——这种错误你绝不想要。所以我目前的 workflow 是：在上面再跑一层 LLM 抽取关键词，然后带着关键词重跑转写模型。只要规模化做过转写就会知道：任何偏离日常对话的领域都会堆出海量转写错误。主持人：有行话就完蛋。对，就像 Whisper 和 ElevenLabs 至今仍把 Claude Code 转成 Cloud Code——（笑声）——「这就是为什么我们知道还有很大改进空间」。随后转向 agent 产品盘点：Tabular Review、intake 已经提过；嘉宾确认大家必须知道的「大件」清单：agent、Tabular Review、projects，另外 Word 加载项很重要——很多人就在 Word 里干活——还有 Outlook 加载项。

[2040s - 2160s] 主持人请嘉宾把 Tabular Review 的用户体验从头走一遍（先律师视角，再谈工程难点）。嘉宾：典型场景就是尽调——要审全部雇佣合同、全部供货合同等。放一万份雇佣合同进去，每一行是一份合同；然后按「列」提问：第一列问「终止通知期是多久」，全部自动填出来（三十天之类的）；下一列问「有没有 IP 转让条款」；再下一列「如果要裁掉他我们要赔多少」「他的薪酬是多少」——整张表全部填满。你也可以直接把一万个文件丢给 agent 说「做尽调」，它会自己拉起一个 Tabular Review 跑完全部，等结果出来后做分析、标出红旗——「这个 CEO 的薪酬高得离谱」「这个奖金结构很怪」——需要重点看的它替你挑出来；然后 agent 可以据此给客户起草备忘录，或者把红旗清单交给律师人工复核。它还是协作式的：可以评论、给其他人看。主持人：一百万格确实吓人；先说说为什么选表格这个形态？

[2160s - 2280s] 嘉宾：表格就是好用——工作的入手门槛极低、对全局的总览极快、审阅极容易。我们产出的所有内容都带回到源文档的引用，你可以很轻松地核验；想确保全对的话，可以把单元格标成「已审」，把有旗的标出来，加评论协作。「这比我们试过的任何其他形态都好用。」主持人：有意思的是表格形态在某些领域拉力很强（GTM 那边的 Clay 就以表格著称），但不知为何没有扩散开——V7 试过，金融领域我想不出一个做「大表格+AI」的头部名字。嘉宾：确实没有。我也觉得奇怪——只要你是 fan-out、每份文档做完全相同的抽取，表格就是理解全局最好的形态。主持人：用户会想导出到 Excel 吗，还是愿意住在你们的表里？嘉宾：愿意住在里面。主持人：你们用 AG Grid 还是自研？嘉宾：用 AG Grid，它应该是这个领域领跑的，人家做了十五年表格，懂行。表格这条线聊完，主持人转向工程侧：fan-out 最有意思——工程师都懂「往一百万份文档上扇出去之前，先在两份上试试」，你们会这么做吗？嘉宾：我们真的会这样：Tabular Review 会先跑前十行，让你检查「它显示的是不是我们以为会显示的东西」。而且对自建这类功能的用户来说这也是最佳实践：在前十行上快速迭代，直到拿到想要的结果，「再放开跑（go ham）」。下一个大件：intake，你们的 Slack bot——主持人打趣「人人都说自己有个 Slack bot，你们的有什么不同」？

[2280s - 2400s] 嘉宾：我们的是法律专用的。它其实不只是「跟 agent 对话的界面」——它专门用来「发起一个 agent」，agent 把任务加进列表。我们有 lists 这个东西（最接近的类比是 Linear）：请求进入专门的 intake 列表，律师和 agent 在 Legora 里看到「这是律所/企业法务收到的请求」，可以为它们发起 agent、协作处理，标记完成后自动回到 Slack。所以它更像一个「店面」：提一张工单、说清要审什么，东西办好了自动送回来，「没人需要换任何工具」，法务团队就坐在 Legora 里管理所有 agent。主持人：这类东西的数据结构很有意思——本质上是个 inbox，所有查询进来再转化成别的形态；跟其他功能做集成难吗？嘉宾：不难，我们有一些真正好用的原语：lists 是个极好的原语——极其灵活，像 Notion database，想配成什么样都行，我们在「让它超级可扩展」上投了很多；还能配触发器和自动化，是把一切粘在一起的胶水：列表上发生什么就发起 agent run，结果回来填进字段、加进文档。就像你可能用 Linear 或任务管理工具协调一堆 coding agent，我们用 lists 做同样的事——「agent 可以建列表，列表可以建 agent，agent 还能读列表」，于是「agent 群就活起来了」。主持人：用户是兴奋还是被吓到？这基本就是终端用户编程了。

[2400s - 2520s] 嘉宾（接「终端用户编程」）：可以让人搭出疯狂的自动化，有些人会超级上头——像 Clay 有自己的网红用户生态，「有 influencers 挺酷的」。值不值要看怎么用——我们需要这种灵活性，哪怕完全不暴露给终端用户：假设它只对内，这也是让 legal engineer 快速量产产品的最快方式——基金管理仪表盘这类东西可以直接用 lists 加生成式 UI 搭出来，等于一个新产品；我们可以做一个「专门产品市场」，legal engineer 出去见客户，「你想要自己的定制版？容易，马上给你配一个」——有点像「Legora 里面的 Lovable」（转写作「inside the door」，存疑）。即便不暴露给用户，我们自己的定制功能也这么建。我们在英国还做了另一笔收购叫 Wexler，做 fact intelligence（事实智能），主要服务诉讼：给一场官司上传一百万份文件，从中抽取带命名实体和日期的「事实」，然后把所有事实互相比对，看哪些互相佐证、哪些互相矛盾——这威力巨大：你可以转写一场庭审，然后「他在庭上说了这个，但那边那封邮件说明他当时不在场」。现场看它跑起来「真的酷到不行」。我们同样能在这个原语上搭很多产品。规模化公司的思路就是：必须解决太多问题，所以要不断想「怎么建一个能延展到很多终端产品的核心原语、支撑高度可定制的体验」。

[2520s - 2640s] 主持人请嘉宾把原语清单重述一遍。嘉宾：lists（超级可配置）、projects（本质是文件容器/文件系统）、agent（可以很多种配置跑起来）。主持人：agent 有名字吗？给它起个名？嘉宾：没有。（旧名 Leya 时代人们曾把它当命名实体谈，后来停了，「也许会捡回来，也许会混乱，不知道」。）Tabular Review 算是这些原语的一个「实例化」——你可以想走多深走多深。那些就是大件。主持人转入 FDE（forward deployed engineer，前置部署工程师）话题：Max 谈过——新客户要尽快拿到价值才能留住。你们的 forward deploy 团队怎么演化的？嘉宾：第一个 forward deploy 招在纽约，也是我们在纽约的第一名员工，YC 创始人背景，大概九个月前（记不太准）。主持人：算 legal engineer 吗？嘉宾：不算，纯工程师。我一直在想怎么规模化地做这件事，核心是：怎么赋能 legal engineer、赋能超级用户去深度定制产品；然后我们仍会保留 forward deploy 工程师做定制集成、做 scoping（界定可行性），并把这一切反馈回核心产品——像「第一道防线、第二道防线」：legal engineer 能搞定大部分定制和体验，搞不定的、需要动核心产品或做集成的，FDE 跟进；FDE 有时做产品、有时跟客户。

[2640s - 2760s] 主持人：走一个例子——刚签下一个大客户，第一步是什么？是「给你们派一个 FDE、帮你建产品」那样吗？嘉宾：客户和客户差别很大，但如果是大客户、特别投入、有一堆想法——通常他们已经自己试过造什么并且失败了，「这是想解决的问题」——我们会配 legal engineer、必要时加一名 FDE：「这几位会过来，陪你们所有律师上手，把你们现在的工作流梳理出来、搬进 Legora。」主持人：legal engineer 和 FDE 的分界到底是什么？嘉宾：legal engineer 以超级用户的方式配置产品，不写代码；FDE 写进生产代码，这就是分界。FDE 什么背景？创始人、普通工程师、「他们都老了」（笑）。D 是我们的第一位——做过两家 YC 公司、在 Meta 做搜索，「她非常强」。主持人：「这门槛可不低。」嘉宾：对，她是我们在美国的第一双靴子。FDE 都给客户建什么？差别很大：有些是人人都要的核心工作流，直接建进 Legora——「我们这么处理 NDA」之类，激活就行，更像「技能构建」；多数 FDE 案例是「我们还做不到的事」——原语需要更多灵活性，或者客户有个极复杂的系统要集成，FDE 就过去对着我们的 API 建，必要时扩我们的 API。

[2760s - 2880s] 主持人：法务结构上怎么处理？毕竟你们自己都不能碰交易数据——FDE 怎么在客户项目上工作？嘉宾：他们仍然不碰数据，只建系统。不过他们有时会像「员工」一样入驻客户：发一台电脑什么的（主持人惊讶：给到律所层面？）——对，但很少发生，通常建个集成就搞定了；只有需要时才这样，而且要看客户规模是否值得——因为 FDE 同时也能做核心产品，这是个持续的优先级权衡。主持人：有没有 Palantir 那种「四天驻场一天回总部写产品」的固定规则？嘉宾：没有，老实说这都还在早期。明年我预计会有一个更大的、持续轮转的 FDE 池：驻场两个月把一件事建完，再去下一家；带着 legal engineer 和 FDE 滚动进驻特定项目。主持人转到 API：你现在是「Legora 平台的 CTO 兼维护者」——你们会把 web app 里的全部能力都开成 API 吗？嘉宾：现在不做这个承诺。我们有可用的 API，但不是产品全部表面。有意思的问题是：这是不是战略？有人会喜欢——「我不用学新工具，在现有工具里装个 Legora 就行」（比如 Word 加载项）——但那样你就得维护它。

[2880s - 3000s] 嘉宾（续 API 话题）：这正是我们一直不做 API 的原因：API 是六个月前才开始的，之前抗拒了很久，因为产品迭代太快，我不想被「维护旧版 API 去适配已经改掉的数据模型」拖慢。现在公司大了，付得起这个成本了；二十个工程师的时候真做不到。主持人（Swyx）：我琢磨过这个问题——所有公司都该有平台团队和产品团队：产品团队你有一个，只要平台别跟产品绑死，第二个就很容易拉起来；最简单的例子是 web app、mobile app、RPC（转写作 RCP，已校正）以及其他集成，然后你的 FDE 就只是平台的「又一个客户端」。嘉宾：对，甚至 agent 也可以这么看：有一个跑 agent 的 harness 团队，你可以在它周围建产品——配上规模化和自动化等等。主持人：这就是著名的 AWS 教令（Bezos：不许再走后门，团队之间只准通过 API 交互，给客户的 API 就是你内部的 API，没有就去建），客户因此得到了一切。嘉宾：这个思路现在也在微调。我们在玩的一个想法是：「agent 能不能就是 API？」听起来很怪，但有些场景客户真正想要的只是「触发某些事发生」——不是说我们在做这个，但我在脑子里推演：你触发 agent、让它去做那件事，显然比直接打一个固定 API 低效得多——可这就是我在琢磨的问题：我们产品里有多少可以只是 agent，而不是其他所有东西。

[3000s - 3120s] 主持人：无界执行、成本和延迟——这是仅有的顾虑吗？嘉宾：还有确定性：大概每一百次里有两次，它会干出我们不想让它干的怪事。主持人转问产品克制：你是「对大多数事说不、只做少数事」，还是「做更多事、再雇一百人」？嘉宾：我曾是非常坚定的第一种。入职半年时我放过一张《300 勇士》电影的幻灯片说「我们这辈子就保持二十个工程师，打赢所有比我们大得多的对手」；现在一百个工程师了，反而觉得太小。现实是惨烈的竞争、用户要的东西太多——我可以当清教徒说「我们就二十人、只建这点东西」，但那样我们只会输、用户会去别家，所以必须建更多。我现在的工作定义是：设计一个系统，让客户感觉我们建了非常多东西，而我们实际只建了很少的东西。主持人：「原语化，对喽。」嘉宾：测试新产品、看有没有拉力？我们没怎么做过——只下过几个有高度信心的战略赌注，主要因为需求堆积如山：大家喊着要的东西都做不完，没必要去试「这个也许行」。

[3120s - 3240s] 嘉宾（续）：有时我们会跟客户说「不，我们认为另外那个东西对你更战略」，然后坚持做那个。这种战略赌注的例子：agent 本身——没有任何人要求过「我希望我的助手有更多 agency、不问我就自己干」，但我们判断行业必然走向那里、它会让大家做很多很酷的事；协作是另一个——行业会走向「在你工作的地方（Legora 里）协作」，而不是邮件和短信里。主持人：那律所内部学会用 agent，是靠 FDE 进去教吗？嘉宾：主要靠 legal engineer——也可以叫「前置部署的法律工程师」——他们进驻律所，帮律师把工作流围绕 agent 的能力重新组织。主持人转题：你们说过（Max 提过）早期不做微调，现在还是吗？嘉宾：到目前为止没微调过任何模型。过去几个月开源权重模型变得极强，所以我们会针对特定任务做——前提是「前沿模型在该任务上的成本/延迟表现」与「后训练模型的潜力」之间存在巨大落差。法律研究是绝佳例子：本质是 fan-out 检索问题，可以训出一个执行这类查询极快、极便宜、低延迟的专用模型。主持人：所以是「把模型后训练得更擅长非常具体的有界任务」。嘉宾：对，我们可不想训一个自己的 Fable 当主 agent——那场仗对 Anthropic 打不赢。

[3240s - 3360s] 主持人补充：搜索/检索（RAG）上仍有大量研究可做，不那么光鲜，甚至不是训练模型，而是架构和管道——你们跟 turbopuffer 合作过案例，late interaction（晚期交互）你们做吗？嘉宾：实际上还没做，但你说得对，这正是方向。我们团队有个哥们（转写作 son，存疑，疑为人名）特别痴迷搜索，老来找我说「该上这个了、有新的 embedding 模型、可以搞 late interaction」——他还真去推动 turbopuffer 团队做了 multi-vector 支持，好让我们能做这类事。但在模型之前、甚至在 harness 之前，都有大量低垂果实：想给客户更好的产出，有一堆立刻见效的事可做，轮不到先训模型。什么才是法律行业的好 harness？文件怎么改、分析怎么做、法律研究怎么跑——全烤进去：出现一个问题，它知道怎么去做研究、找先例找答案，这些都非常关键，再加上周边的治理。主持人：那最重要的定制工具是不是检索——引不到对的信息就全完了？嘉宾：是检索、分析、起草三件——起草大 Word 合同其实一点也不简单。从零起草一份大合同，你得先找先例：我们以前怎么写的？这条款需不需要？这家律所以前怎么处理的？NDA 简单，但复杂合同必须先找对先例。

[3360s - 3480s] 嘉宾（续）：律师从不坐下就开敲，他们先琢磨「类似交易里我们用过哪些条款」，而「类似」是多维的：同行业、同类公司、同结构；然后从多份文档里决定「哪一段取自哪一份」——这本身就是个检索问题。主持人：哪个模型起草最好？嘉宾：我昨天刚说过这个——Fable 和 Open（转写如此，疑指 OpenAI 模型）话都极多、极啰嗦，你们肯定也知道，啰嗦一般不是好事；Astra 写作相当好、简洁得多。主持人：你们会跑一百个 prompt 来约束吗？嘉宾：能做一些，但「话痨体质」对我们挺难治。下一代 Anthropic 模型大概率会好很多——Astra 其实就非常简洁。主持人：你们肯定跟实验室都有 Slack 通道直说这些吧？嘉宾：当然，这不是新闻，他们都知道。另一个我在乎的是幻觉：它们会不会编东西？我们有一套引用系统：写出来的东西必须带链接，用户点开，PDF 上一个 bounding box 高亮「这条信息取自这里」，我们确保它说的每句话都可引用——而且多层下钻：agent 基于 Tabular Review 写的结论，可以一路引到那张表的单元格/行，从底往上全部可验证。主持人：律师真的会去核对这些引用吗？嘉宾：一开始有些会，现在嘛——跟代码圈一模一样：有的工程师对 agent 输出满不在乎（有人对幻觉出的最高法院链接都不点），有的人每个逗号都读。

[3480s - 3600s] 嘉宾（续）：但话说回来，一个人怎么可能逐条看完那么多东西？他们其实知道「哪里是必须查的关键」，就像 code review 知道看什么系统；其他「听着合理的，大概没问题」——跟收到初级律师的产出一样：这几处我要核一下，那些应该没事。主持人：那下一代模型之前，法律推理还有什么缺口是你想要的？嘉宾：一时想不出具体的，但长文档推理仍然难：百万 token 的巨大合同，你要在全文各处做编辑，还得记得「上面这些地方你都改过、这个定义变了、这个顺序变了」。最简单的做法是一把梭——全部塞进去、一次性输出——我们不这么做，我们分块、然后「编排式」地做编辑。嘉宾感慨：身处这个行业有趣就有趣在这些地方：我们的 Word 编辑方案做过 map-reduce、又改掉，来来回回——「模型一变，我们搭的脚手架就得跟着重搭」。主持人深有同感：我因为工作做大量新闻整合，最近刚从 map-reduce 切换到「全塞进上下文」，因为现在这样效果更好。嘉宾：是，但也不完全一样。我们在搜索上遇到过一个坑：在一份巨大合同里找某个东西，可以 map-reduce，但只要涉及摘要，就可能丢掉「块与块之间」的重要细节——逐块摘要再拼起来，nuance 就没了。所以我们是两套搜索并用，其中一套需要在中间步骤里做推理。我正想找个好例子。

[3600s - 3720s] 嘉宾（例子）：有些文件用整整两页讲一件事，单独每一块跟你的检索词都不相关，但合在一起却极度相关——这就是难做的地方，而且（呼应前面）想让模型在这种事上变好很难，因为我们根本看不到真实合同。主持人：那你们会不会请客户做「类似真实合同」的示例？嘉宾：我们有反馈机制，客户愿意配合的话：他们说「这个对我们特别重要但现在做得不好」，我们就说「太好了，能不能把文档和问题匿名化给我们，能复现我们就能修」。主持人：这事还有个更激进的版本——买一家律所。嘉宾：确实有人在考虑。但这里有讲究：那些数据到底属于律所还是属于客户？「如果客户都走光了，那也许就没问题了——我是真不知道。」不过这确实是搞到大量数据的一条路。主持人：就像 Google 买 Spirit（航空）那样——他们花了大概一千万、七百万还是多少来着；一家倒闭的公司还留着 Slack，Slack 都能卖；好像还有家 YC 公司专门收购倒闭公司——大家要的就是「真人在真工作里留下的东西」。嘉宾：买律所数据的真正问题是：输入好拿——真实输入很棒；但输出难用——你只拿到最终结果，得逆向工程「他们是怎么走到的」。在我们的产品里，拿到真实用户的真实 trace 价值极高；只有产出、不知道好坏和过程，就很难用。（主持人：我们快到时间了。）主持人最后指路一个方向——embedding 上还剩很多 alpha：如果你的负载很吃 RAG，只看 MTEB 榜拿第一名的模型用，你是在把大量收益留在桌上。

[3720s - 3840s] 嘉宾（续 embeddings）：该看 Voyage 或者定制 embedding；别用默认 embedding，OpenAI 的、Gemini 的都一样，定制你的 embedding。唯一要注意的是必须多语言：我们要服务太多语言了，这方面「欧洲是王者」（此段还聊到 Mistral——转写严重损坏存疑——与「欧洲需要更多模型公司」的看法）。主持人随后改问预测：两年后，律师会在 Legora 里干活，还是用 Claude/ChatGPT 之类的工具调用 Legora 的 API？嘉宾：两者都会有。大企业可能会有 Claude co-work（就像我通过 Claude 用 Linear、但我仍然会打开 Linear），企业会这么用；律所会直接用 Legora，因为更简单更好。主持人：「这让我想起医疗领域的 OpenEvidence——你们就是法律版的它。」主持人又问：编码领域已经到了「实验室会把竞争者掐掉」的地步，还发生过两次，法律会发生吗？嘉宾：我不这么认为。他们确实在发布法律服务，我们也一直跟他们聊，但他们的销售会说「不会」——销售说了不算，他们甚至常不知道实际发生了什么。

[3840s - 3960s] 嘉宾（续）：实验室要吃下的是整个企业知识工作——法律跟数据分析、销售、营销、财务一样，只是其中一块；他们会做出「够好」的东西，但永远不会像我们扎得这么深。如果 Claude 接走一些小客户、让他们用 Claude 做部分法律工作，「那挺好的，没关系」；但你要做一笔四亿美元的并购交易，你还是会要 Legora。主持人顺势问个不相关的问题：你们会去法学院推广、让他们在课堂里教 Legora 吗？嘉宾：确实有这个计划，应该已经启动了——说实话我不太清楚具体进展到哪了，但我们希望法学院的学生学会用 Legora。主持人：你觉得法学院课程变化大吗？嘉宾：「这个我真不知道。」主持人现场支招：我会换个打法——给 CS 背景的学生发法学院奖学金：你读完 CS、有点软件经验，我们出钱送你读 JD；或者搞「硕士+JD」联读。嘉宾：他们不在乎硕士那一截——很多没读过 CS 本科的人读个硕士只是为了补课，这种人我不常见到；JD-MBA 倒是常见，JD-CS 我还没见过。主持人：「会有的。」随后收尾寒暄：整体进展超级亮眼、恭喜——节奏能保持吗？嘉宾：当然。招聘：纽约、斯德哥尔摩、伦敦招工程，其他岗位全球招。主持人：纽约现在多少人？嘉宾：一百。计划扩到多少？

[3960s - 4042s] 嘉宾：很多，能招多少招多少。主持人：「很凶悍啊——反正对美国中部来说，公司法务全在这一块了。我很期待你们买下一家律所。」嘉宾：确实非常令人兴奋。主持人：我觉得每个行业——尤其是顶端的专业工作：医生、律师、码农——最后都会有两三家 agent lab 统治各自的领域。嘉宾：同意。我们市场非常好，而且随着法律服务变得更平价、触达更多人，这个市场只会大得多。主持人：另外（我们已经严重超时了）你们做的问题本身就是前沿知识工作——除非你真在做「科学中的科学」、纳维-斯托克斯方程那种，否则大多数工作最后都收敛到法律：招聘、财务、一切。「到头来全是法律，出事我就得咨询我的律师。」嘉宾：这正是我觉得这事激动人心的原因——现代社会某种程度上就运行在法律之上。好，就到这里，谢谢。谢谢两位。