中文精读
Mercor 的切口：把找到专家，变成可信的持续交付
嘉宾：虞快，Mercor 首位中国出生工程师，曾在 Google、Two Sigma、Citadel 从事工程工作（当期身份）
主持：曲凯
原发布：2025 年 9 月 6 日｜时长：50 分 13 秒
整理：2026 年 10 月 18 日｜官方单集
外界先把 Mercor 看作 AI 招聘平台，后来又把它归为数据标注公司。虞快的解释更具体：长期想解决劳动市场匹配，当下最急、也最愿意为结果花钱的，是需要专家反馈的 AI 模型公司。目标未必换了，切入点却明显收窄。
这期节目值得研究的不是“哪个标签更性感”，而是平台怎样把需求、专业供给、评估与履约接在一起。官方简介称其很快达到一亿美元年收入；正文说的是从一百万美元到一亿美元的年化收入运行水平用了十一个月。年化运行水平是把当时收入速度折算成年，不能理解为这十一个月已经累计收到一亿美元。相关数字均为当期口径，未审计。
一、专家不仅执行标准，也帮助客户知道什么算好
回听：01:57—06:39、18:48—19:43。 虞快列出的需求很细：医生、律师、特定编程语言工程师，甚至某一语言和学科的交叉人才。模型公司可能知道模型有问题，却未必知道怎样系统判断医学或法律回答。他们需要的不只是人力，而是判断力。
主持一度理解为专家先定标准，再交给另一批普通标注者执行。虞快纠正说，这可能是同一批专家：当模型已能回答一般问题，评估者自身的能力也要跟上。工作从机械贴标签，转向指出哪些答案为何不够好，以及模型应该学会完成什么。
他把 evaluation，即评估，类比模型的产品需求文档：研究人员需要一个目标，知道下一步能力向哪里改。这个比喻解释了价值所在，但不等于专家意见自动正确。任务定义、争议复核和交付后的实际效果仍须检验。
同时，保留评估资产与跨客户复用不是同一件事。虞快说，一些项目与专家需要按客户分开。不能把“平台留下标准”理解为可以把私有数据和整套题目随意卖给另一家公司。
二、匹配只是入口，客户还在买协调成本的下降
回听：06:44—08:47、15:10—15:28。 节目录制时，业务主要面向企业，并非任何用户登录后就能自助下单。客户向 Mercor 付款，Mercor 再向专家付款。平台除了找人，还要处理工时、质量、跨国支付与争议。
如果一百位工程师同时做一个项目，客户需要知道谁实际完成工作、谁的质量不够、谁要续约。来自不同地区的专家还可能遇到付款或记录问题。这些协调环节解释了为什么客户不一定在见到候选人后就绕开平台：买到的是一段持续运行的服务。
复购也并非简单抢快。虞快说，过去找到的人质量好，会建立下一次合作的信任。速度与可信质量共同组成产品，而不是“更快把简历发过去”。节目没有给出完整合同与毛利，因此不能据这种结构直接认定利润丰厚。
编辑推论： 双边平台的关键漏斗应延伸到任务接受、完成、争议与付款，而不只看注册和匹配。注册增长很快但履约大量失败，得到的可能只是更大的协调负担。
三、AI 面试需要岗位标准，真实性与质量还有别的检查
回听：09:01—15:28。 面试问题来自简历与职位要求，人可以增加问题和评分规则。约二十到三十分钟的视频形成可定位的转写，招聘者能从某句话跳回原视频，观察表达是否清楚等文字难以完整保留的信息。系统再按预设要求评估和排序。
这个过程不是用一段视频证明全部职业能力。分数是否真正预测交付，需要任务后的反馈；表达风格也不能直接等同能力。节目还列出非 AI 自动化：身份文件、公开履历与外部记录可以帮助发现冒名申请。例如有人借真实奥赛获奖者的名字提交材料，简历“看起来真的”并不足够。
多人回答相同问题、比较异常，是另一种质量方法。但不同答案不自动是错误，尤其专业任务可能存在合理分歧。把共识当真值，会把少见但正确的解法一起排除。
供给端也并非全靠自动获客。推荐占注册的重要部分，稀缺人才还要到专业论坛或活动中找。节目所举外科医生时薪四百美元、推荐费五千美元，是当时特定岗位的例子，不是当前招聘报价。价格还受到过去寻找难度、紧急程度与双方协商影响；高价格本身不证明高质量。
四、快决策的前提，是能在一周后看见并修正错误
回听：24:26—35:13。 虞快谈到创始人的沟通能力、成熟公司早期人才和强增长带来的吸引力。增长吸引人，人才又改善执行，两者可能相互促进。不能把结果归因成“年轻、辍学所以成功”，也不能拿一家公司的自述推出普遍因果。
他说团队未必对每个小决定都做严密实验，但会提前讲清：未来一周想看到什么；没有出现时怎么办。创始人关注什么有效、什么无效。“对错不重要”在完整上下文中，有避免重大错误、能意识到错误并调整的条件，不能扩展到不可逆、高影响决策。
执行力则有更清楚的定义：本周承诺的事情，到周五基本交付；不同组也能给彼此同样的信心，因此公司才敢向客户承诺。节目提到很长工时，但真正可迁移的机制是承诺可预测、依赖可协调、反馈进入下一步。不能仅用增长证明工作制度合理，或把凌晨下班当作成功配方。
编辑推论： 把“动作快”翻译成一个可检查的周期：决定—承诺—交付—观察—调整。没有后两步的快，只是更快积累未知问题。
五、工程师多承担问题选择，组织就得接住跨角色协作
回听：35:28—49:59。 虞快招聘时看主动性：一个目标要解释多细，候选人能独立推进多久，何时主动求助。他也没有完全排除偏执行型同事，团队需要不同角色组合，不能用一个词给人排固定等级。
停车题用来观察是否能从逐个模拟转向寻找失败条件；笔算乘法题则暴露另一种行为：忘记规则可以问，不必耗半小时假装推进。这里有用的是抽象与沟通过程，而不是宣称一道 IQ 题足以筛选全部人才。电脑是否装有开发工具，是他的个人筛选偏好，也不应变成通用规则。
学习新概念时，他喜欢与熟悉概念比较，寻找相同与不同。MCP 与 API 的比较可以帮助入门，但类比不能替代正式定义和实际测试。快速产生框架与准确掌握实现，是两件需要分别完成的事。
结尾谈代码智能体：AI 可以先起草代码，结果未必正确，工程师更多时间应转向选择问题、沟通客户与确认方向。虞快回忆在 Google 的早期项目：资深者也未必已有答案，自己的工作就是把未知弄清，再帮助别人。若组织仍把客户交流当某个职位的专属领地，主动性很容易变成“抢活”。管理者需要识别实际贡献，让跨角色合作有激励和清楚责任。
应用练习：给一项专家服务补齐六个环节
画出一次真实订单，从供给招募开始，依次写：资格怎么验证、岗位怎么匹配、质量怎么确认、工时怎样记录、钱怎样到达、争议由谁处理。每一项都填写一个失败例子，不用“平台会解决”跳过。
再选一个可逆小改动，提前写下一周后的预期、最低质量和停止条件。周末复盘时，把“上线了”“客户采用了”“交付合格了”分别记录。若要做推荐奖励，优先观察有效供给和后续交付，别只奖励新增注册。
最后找一个跨组承诺：你敢向客户承诺哪天完成，是因为谁给了什么保证？这能暴露执行力究竟来自稳定协作，还是少数人临时救火。
来源边界
依据官方介绍与完整逐字稿。收入运行水平、增长、市场规模、报价及组织情况均为 2025 年访谈自述；对竞争者和交易动机的猜测不作为独立事实。录制时“首位中国员工”按正文限定中国出生，嘉宾明确有其他华裔同事。不把面试分数、类比或数学题当已验证的公平性与能力测量。编辑推论及练习另行标识。
回听导航
- 01:57—06:39：专家需求为什么成为切口。
- 06:44—08:47：匹配之后的履约工作。
- 09:01—15:28：面试、真实性、招募与价格。
- 18:48—24:20：评估目标、私有边界与长期愿景。
- 30:54—35:13：快决策与跨组交付。
- 35:28—49:59：主动性、学习与工程角色变化。
相关主题：专家评估、双边市场、组织执行。延伸收听：周健谈企业 AI 交付、汪华谈能力与机会阶段。
English edition
Mercor's opening: turn finding experts into dependable, continuing delivery
Guest: Yu Kuai, Mercor's first engineer born in China, previously an engineer at Google, Two Sigma, and Citadel; roles at the time
Host: Qu Kai
Originally published: September 6, 2025 | Duration: 50 minutes 13 seconds
Edited: October 18, 2026 | Official episode
Mercor was initially described as AI recruiting, then as data labeling. Yu's explanation is more specific: its longer ambition concerns labor matching, while AI model companies currently have particularly urgent, well funded demand for expert feedback. The objective need not have changed, even though the entry point narrowed.
The useful question is how demand, professional supply, evaluation, and fulfillment connect. The official description mentions reaching $100 million in annual revenue; the conversation more precisely describes moving from a $1 million to a $100 million annualized revenue run rate in eleven months. Annualizing the current pace does not mean collecting $100 million over those eleven months. These contemporary figures were not audited.
1. Experts help define good work, as well as perform it
Listen: 01:57–06:39 and 18:48–19:43. Yu lists doctors, lawyers, engineers specializing in particular languages, and people combining unusual disciplines or languages. A model company may know something is wrong without knowing how to evaluate medical or legal answers systematically. It needs judgment as well as labor.
Qu initially imagines experts setting standards for less specialized labelers. Yu explains that the same experts may do both: when models can answer ordinary questions, evaluators need stronger capabilities too. The work shifts from mechanical labeling toward explaining which answers are inadequate and what a model should learn to do.
Yu compares evaluation with a model's product requirements document: researchers need a target for further improvement. The analogy explains value without making expert opinions automatically correct. Task definitions, disputed cases, and subsequent performance still need checks.
Retaining evaluation assets also differs from reusing them across customers. Yu says some projects and experts must remain separated. Standards remaining with a platform do not grant unrestricted permission to resell private data or entire question sets.
2. Matching is an entry point; customers also buy lower coordination costs
Listen: 06:44–08:47 and 15:10–15:28. At recording, the business primarily served enterprises rather than offering universal self service. Customers paid Mercor, which paid experts. Beyond recruitment, it handled hours, quality, international payments, and disputes.
With a hundred engineers on a project, customers need to know who completed the work, whose quality is inadequate, and whom to renew. Experts in different countries may encounter payment or record problems. This coordination helps explain why customers might continue using the platform after meeting candidates: they are buying an ongoing service.
Repeat business is not simply a speed contest. Previous successful placements create confidence in the next engagement. Speed and trustworthy quality together form the product, rather than faster delivery of résumés. The episode does not provide complete contracts or margins, so the structure alone cannot establish high profits.
Editorial inference: A marketplace funnel should extend through task acceptance, completion, disputes, and payment. Rapid registration with unreliable fulfillment can merely enlarge the coordination burden.
3. AI interviews need job criteria; identity and quality need further checks
Listen: 09:01–15:28. Questions reflect résumés and role requirements, with people able to add questions and scoring criteria. A roughly twenty to thirty minute video produces a navigable transcript. Recruiters can jump to a statement and observe communication details that text does not fully preserve. The system evaluates and ranks candidates against defined requirements.
A video does not prove every professional capability. Whether scores predict delivery needs subsequent task feedback; speaking style cannot simply stand in for ability. Yu also describes ordinary automation checking identification and public records. An applicant borrowing a real mathematics competition winner's name illustrates why a plausible résumé is insufficient.
Comparing multiple answers can help detect anomalies. However, different answers are not automatically wrong, especially where professionals reasonably disagree. Treating consensus as truth risks rejecting unusual correct solutions.
Supply acquisition is not entirely automatic either. Referrals account for an important portion of registrations, while scarce specialists may require professional forums and events. The cited surgeon rate of $400 per hour and $5,000 referral reward were examples for particular roles then, not current job offers. Historical search difficulty, urgency, and negotiation also influence prices. High prices alone do not establish quality.
4. Fast decisions require noticing and correcting mistakes a week later
Listen: 24:26–35:13. Yu discusses founders' communication skills, early hires from established firms, and growth attracting talent. Growth and better execution may reinforce one another. This does not establish that youth or leaving college caused success, or that one company's account proves a general rule.
He says not every small decision receives a rigorous experiment, but expectations are explicit: what should happen during the next week, and what if it does not? Founders examine what works and what does not. “Being right does not matter” comes with conditions: avoid major mistakes, recognize errors, and adjust. It cannot justify irreversible decisions with large consequences.
Execution has a clearer definition: commitments made for the week are largely delivered by Friday. Other groups provide similar confidence, enabling customer promises. The conversation mentions long hours, but the transferable mechanism concerns predictable commitments, coordinated dependencies, and feedback. Growth does not validate a working regime, and leaving after midnight is not a success formula.
Editorial inference: Translate speed into a cycle of decision, commitment, delivery, observation, and adjustment. Without the final two steps, speed can simply accumulate unknown problems faster.
5. Engineers choosing more problems need support for collaboration across roles
Listen: 35:28–49:59. Yu assesses agency through how much explanation a person needs, how long they can advance independently, and when they seek help. He does not exclude people who prefer execution: teams need combinations rather than a fixed hierarchy based on one word.
A parking problem tests movement from simulating cases to identifying failure conditions. A multiplication problem reveals another behavior: forgetting the rule is acceptable, but spending half an hour pretending to progress instead of asking is costly. The useful observation concerns abstraction and communication, not a claim that one IQ question evaluates all talent. His preference about installed development tools is also personal rather than a universal hiring rule.
He learns concepts through comparisons with familiar ones. Comparing MCP with APIs may provide an entry point, but analogies do not replace formal definitions and practical testing. Building an initial framework quickly and understanding implementation accurately remain separate tasks.
The closing discussion concerns coding agents. AI can draft a beginning that may be wrong, while engineers spend more time choosing problems, speaking to customers, and checking direction. Yu recalls an early Google project: senior colleagues did not necessarily know the answer either. His job was to investigate, solve it, then teach others. If customer contact remains another role's exclusive territory, initiative can become a territorial dispute. Management needs incentives and clear responsibility that recognize actual contributions.
Exercise: complete six stages of an expert service
Map a real order, recording how qualifications are checked, jobs matched, quality accepted, hours recorded, money delivered, and disputes resolved. Add a failure case for each rather than writing “the platform handles it.”
Choose a reversible small change and record the expected result a week later, minimum quality, and stopping condition. Distinguish shipping, customer adoption, and acceptable delivery in the review. Referral rewards should consider effective supply and later performance rather than registrations alone.
Finally, examine one cross-team promise. Whose guarantee allows you to promise a completion date to a customer? This reveals whether execution rests on dependable coordination or recurring rescue by a few people.
Source boundary
This reading uses the official introduction and complete transcript. Revenue run rate, growth, market size, prices, and organizational conditions are accounts from the 2025 conversation. Competitor assessments and transaction motives are not independent findings. “First Chinese employee” is limited to birth in China because Yu explicitly mentions other colleagues of Chinese ancestry. Interview scores, analogies, and puzzles are not validated measures of fairness or ability. Editorial inferences and the exercise are identified separately.
Listening navigation
- 01:57–06:39: Expert demand as an entry point.
- 06:44–08:47: Fulfillment after matching.
- 09:01–15:28: Interviews, authenticity, acquisition, and prices.
- 18:48–24:20: Evaluation goals, private boundaries, and the longer vision.
- 30:54–35:13: Decisions and delivery across teams.
- 35:28–49:59: Agency, learning, and engineering roles.
Related themes: expert evaluation, marketplaces, organizational execution. Further listening: Zhou Jian on enterprise AI delivery and Wang Hua on capability and opportunity stages.