20VC: Mercor CPO on Revenue Concentration from Frontier Labs | Why Large Enterprise is Scared to Partner with Frontier Labs | Why Small Specialised Models is the Future with Osvald Nitski

以下是将内容翻译为中文： 播客节目邀请了 McCaw 公司的首席产品官 (CPO) 奥斯瓦尔德·尼茨基，讨论该公司在快速发展的 AI 领域中的独特地位，特别是在为前沿模型提供数据方面。 **McCaw 的商业模式与开源 AI** 尼茨基解释说，McCaw 的核心业务是提供评估和训练数据，这项业务并未被开源模型的进步所“蚕食”。相反，开源模型“抬高了”可能性的“下限”，从而增加了对模型性能“前沿”数据的需求。客户寻求数据来填补当前模型能力的空白，这促使 McCaw 转向更复杂、需要长期规划的任务，而这些任务是当前模型尚未能完全处理的。尼茨基反驳了“90% 的企业工作流程可以由开源模型处理”的观点，他认为这忽视了对高级、多步骤任务的“潜在需求”。 **企业 AI 采用与投资回报率 (ROI)** 针对企业在数据共享和 AI 投资回报率（ROI）方面的疑虑，尼茨基指出，敏感性取决于工作流程的关键程度。敏感度较低的领域，如人力资源或采购，对专有模型的接受度更高；而核心业务功能需要更多的控制，这通常导致企业选择本地部署的开源解决方案。他认为 AI 目前还没有出现“投资回报率问题”，因为公司正处于探索阶段，但他强调需要更好地衡量成果。McCaw 公司自身在 AI 工具上的投入巨大，甚至超过了给开发人员支付的薪水，认为这对于增长和提高效率至关重要，旨在满足“永不满足的客户需求”。 **AI 时代的产品管理** AI 深刻地重塑了 McCaw 的产品管理。尼茨基强调了两项关键变化： 1. **工具整合：** 从使用 Figma 等多样化工具转向基于云的设计和编码智能体，这减少了工具的多样性，简化了采购流程。 2. **侧重业务影响：** 随着 AI 自动化了大部分执行工作，产品经理现在更多地关注“判断力”和“业务价值”。 随着编码智能体提高了工程速度，这也导致了更高的产品经理与工程师比例。他回顾了过去的一个错误，即支持了过于多样化的人工数据工作流程，导致了工具过于复杂。吸取的教训是设定“护栏”并专注于通过持续市场互动确定的“持久需求”。 **规模化、挑战与数据的未来** McCaw 在扩大供给方面的成功归因于良好的专家体验、及时且公平的报酬以及强大的推荐计划。虽然薪资是一个因素，但长期留住人才则依赖于持续的工作机会、技能发展和良好的沟通。尼茨基澄清说，McCaw 的内部重点是为客户和专家提供价值，利润率是之后再计算的。 针对收入集中化（即前沿模型提供商是 McCaw 最大的客户）的问题，尼茨基表示，公司正在积极“下沉市场”以实现多元化。将人工数据项目民主化，面向更广泛的企业受众，是具有挑战性的，原因在于管理边缘案例的运营强度、确保清晰的沟通，以及适应快速变化的数据类型（从有监督微调到偏好排序，再到现在的“强化学习环境”）。 尼茨基认为“环境”数据是增长最快的类型，它模拟应用程序并为智能体设置复杂的“初始状态”，这反映了模型训练正转向更接近实际部署条件的方向。他预测“物理数据”，尤其是用于机器人领域的数据，将在三年内变得非常重要，尽管目前演示还存在局限性。他将其与自动驾驶汽车漫长而缓慢的进展相提并论。 **招聘与文化** 在 AI 驱动的世界中，招聘重心已转向更资深、懂得业务影响的候选人，因为基本的工具熟练度变得不再那么关键。McCaw 专注于寻找 25-35 岁之间，“渴望成功”并“处于最佳状态”的人才。面试流程现在测试的是良好的实验设置能力、统计学知识和明智的判断力，而不仅仅是 AI 工具的熟悉程度。尼茨基强调要招聘具有“高能动性”和“主人翁精神”的人，即使这意味着要容忍一些个性怪癖，因为这些品质比团队合作技能更难培养。他为“McCaw 黑手党”感到自豪——那些离开 McCaw 后创立自己公司的前员工。 **结语** 尼茨基相信 McCaw 可以通过持续解决模型性能的主要瓶颈——评估和训练数据，从而成为一家 2000 亿美元的公司。他建议有抱负的计算机科学学生在快速发展、处于前沿的公司获得“真正的实习机会”，因为学校的知识很快就会过时。他更尊重客户而不是竞争对手，因为他发现竞争对手经常模仿 McCaw 的创新。他对机器人的长期潜力感到最为兴奋，认为这是一个必然的、尽管缓慢的进展，类似于自动驾驶汽车。

The podcast features Oswald Nitsky, CPO at McCaw, discussing the company's unique position in the rapidly evolving AI landscape, particularly concerning data provision for frontier models. **McCaw's Business Model & Open-Source AI** Nitsky explains that McCaw's core business, providing evaluation and training data, is not cannibalized by the advancements of open-source models. Instead, open models "raise the floor" of what's possible, increasing demand for data at the "frontier" of model performance. Customers seek data to fill gaps in current model capabilities, pushing McCaw towards more complex, long-horizon tasks that current models cannot yet fully handle. Nitsky disputes the notion that 90% of enterprise workflows can be handled by open models, suggesting this overlooks "latent demand" for advanced, multi-step tasks. **Enterprise AI Adoption & ROI** Addressing skepticism from enterprises regarding data sharing and AI ROI, Nitsky notes that sensitivity depends on the workflow's criticality. Less sensitive areas like HR or procurement are more open to proprietary models, while core business functions require more control, often leading to on-premise open-source deployments. He believes there isn't an immediate "ROI problem" for AI, as companies are in an exploration phase, but emphasizes the need for better accounting of outcomes. McCaw itself spends aggressively on AI tools, even exceeding developer salaries, viewing it as essential for growth and efficiency to meet "insatiable customer demand." **Product Management in the AI Era** AI has profoundly reshaped product management at McCaw. Nitsky highlights two key changes: 1. **Tool Consolidation:** A move away from diverse tools like Figma towards cloud-based design and coding agents, reducing tool diversity and making procurement simpler. 2. **Focus on Business Impact:** PMs now focus more on "judgment" and "business value" as AI automates much of the execution. This leads to a higher PM-to-engineering ratio, as engineering velocity increases with coding agents. He recounts a past mistake of supporting too many diverse human data workflows, leading to an overly complex tool. The lesson learned was to impose guardrails and focus on "enduring demand" determined by constant market engagement. **Scaling, Challenges, and the Future of Data** McCaw's success in scaling supply is attributed to a great expert experience, timely and fair pay, and a strong referral program. While salary is a factor, long-term retention relies on consistent work, skill development, and good communication. Nitsky clarifies that McCaw's internal focus is on delivering value to customers and experts, with margins calculated afterward. Addressing revenue concentration (frontier model providers being McCaw's largest customers), Nitsky states the company is actively moving "down market" to diversify. Democratizing human data projects for a broader enterprise audience is challenging due to the operational intensity of managing edge cases, ensuring crisp communication, and adapting to rapidly changing data types (from supervised fine-tuning to preference ranking, and now "RL environments"). Nitsky identifies "environments" data as the fastest-growing type, simulating apps and complex "start states" for agents, reflecting a shift towards training models in conditions closer to real-world deployment. He predicts "physical data," especially for robotics, will be highly significant in three years, despite current demo limitations, drawing parallels to the long, slow progress of self-driving cars. **Hiring and Culture** Hiring in the AI-driven world has shifted towards more senior candidates who understand business impact, as basic tool proficiency becomes less critical. McCaw focuses on finding individuals between 25-35 who are "hungry" and "hitting their prime." Interview processes now test for setting up good experiments, understanding statistics, and sound judgment, rather than just AI tool familiarity. Nitsky emphasizes hiring people with "high agency" and "ownership," even if it means tolerating some personality quirks, as these are harder to instill than team-working skills. He is proud of the "McCaw Mafia" – former employees who go on to start their own companies. **Final Thoughts** Nitsky believes McCaw can become a $200 billion company by continuing to solve the primary bottleneck to model performance: evaluation and training data. He advises aspiring computer science students to get "real internships" at fast-growing, frontier companies, as school knowledge quickly becomes outdated. He respects customers more than competitors, finding competitors often copy McCaw's innovations. He is most excited about the long-term potential of robotics, seeing it as an inevitable, albeit slow, progression akin to autonomous vehicles.