Why Claude can’t be your PM (yet) | Anthropic CPO Panel

讨论始于一个紧迫的问题：一年前，人工智能似乎威胁要淘汰产品经理（PM），但现在，每个人似乎都需要成为一名产品经理。来自Anthropic的Ami和Mike就产品经理角色如何演变提供了见解。 Ami强调，产品经理的核心工作保持不变：将现实世界中的人类问题与技术解决方案连接起来。真正改变的是技术进步的速度，现在是每两个月变化一次，而不是每五年或十年。这种快速变化要求产品经理不断适应，挑战他们“抛弃我所知道的大部分过去有效的东西”，并以全新的视角审视问题。她将当前状态比作产品管理早期，当时这个角色定义不明确，要求产品经理成为边学边做的通用问题解决者。 Mike分享了他作为Anthropic的个人贡献者（IC）最近的一个“顿悟时刻”。尽管拥有像Claude这样先进的人工智能，一个即将交付的项目仍然迫切需要一名产品经理。他最初质疑这种必要性，认为Claude可以处理，但产品经理的到来凸显了关键的“粘合剂”工作：让利益相关者参与进来，准备客户成功团队，引入保障措施，并确保卓越运营。Mike总结道，这种“召集者”角色，即确保一切连接起来并满足最终用户需求，由于人工智能带来的速度提升，比以往任何时候都更加重要。他指出，虽然人工智能可以提供洞察力，但它尚未具备有效的组织凝聚力或调度能力来协调人类协作。 对话随后深入探讨了正在过时的特定技能和正在变得重要的技能。Ami反思了她花费数十年时间磨练的详细UI/UX设计技能（例如，“我们应该把这个按钮放在哪里？”）。现在，她发现“构建三个版本并试用”更快速有效。这种个人“创新者困境”强调了抛弃旧能力的需求。新的关键技能包括对变化的更高容忍度、适应能力、强大的判断力和面对不确定性时的不屈不挠。Ami鼓励“框架化混沌”，使其感觉安全且易于参与，从而培养一种将持续适应视为常态的文化。Mike补充说，明确的领导力和“DRIs”（直接责任人）对于驾驭这种混乱至关重要，能在快速变化中提供稳定的指导。 关于*构建什么*，Mike讨论了向“代理原生”软件的转变。最初，人工智能被视为一个侧边栏功能，然后集成到现有功能中。下一个阶段是构建代理可以执行人类所能做的一切操作的产品。他指出，Anthropic内部使用Claude来创建和迭代项目跟踪的用户界面，这表明界面可以被代理“塑形”。这引发了关于数据来源以及如何在保持用户友好的同时使软件对组织变化做出反应的问题。Mike建议公司构建正确的“原语”——基础架构层，允许人类和代理无缝交互，从而实现渐进式演进，而不是简单地“附加”AI功能。 Ami谈到了平衡前沿探索与提供稳定用户体验的挑战。Anthropic的方法是“在每个人所处的位置与他们会合”，鼓励大量的“射门尝试”和并行实验。这使得他们能够在整合为更稳定的产品之前发现哪些有效。Mike补充说，成功往往来自那些信念坚定、拥有强大底层基础设施（如共享内存系统）的团队，这些基础设施能让不同的实验产品感觉相互补充而非脱节。 小组还谈到了如何识别成功并避免“能力盲点”。除了传统的产品市场契合度指标，Mike强调了“搁置”那些不适用于当前模型但随着新模型发布而重新审视的项目的重要性。他举了一个内部计算机使用产品的例子，该产品最初“非常糟糕”，但在Claude 3.7发布后取得了突破，表明模型可以通过意想不到的方式改进。 展望未来一年，Ami希望增强个人和小型团队的赋能，让每个人都能更轻松地进行构建，并放大构建者的影响力。Mike的愿望是缩小人工智能模型的能力与大多数人使用它们的方式之间的差距。他设想的未来是，先进的多代理设置不仅仅供“Claude狂热的软件工程师”使用，而是变得可访问和民主化，赋能更广泛的用户在他们的职业和个人生活中。

The discussion kicks off with a pressing question: a year ago, AI threatened to make Product Managers (PMs) obsolete, but now it seems everyone needs to become one. Ami and Mike from Anthropic offer insights into how the PM role is evolving. Ami emphasizes that the core job of a PM remains unchanged: to bridge real-world human problems with technological solutions. What *has* changed is the pace of technological advancement, now shifting every two months instead of every five or ten years. This rapid change demands constant adaptability, challenging PMs to "throw away most of what I know about what used to work" and approach problems with fresh eyes. She likens the current state to the early days of product management when the role was less defined, requiring PMs to be general problem solvers who learn as they go. Mike shares a recent "realization moment" from his individual contributor (IC) role at Anthropic. Despite having advanced AI like Claude, a project nearing shipment desperately needed a PM. He initially questioned the need, thinking Claude could handle it, but the PM's arrival highlighted the critical "glue" work: bringing stakeholders along, preparing customer success teams, looping in safeguards, and ensuring operational excellence. Mike concludes that this "convener" role, ensuring everything connects and the end-user needs are met, is more essential than ever due to the increased speed enabled by AI. He notes that while AI can provide insights, it doesn't yet possess the organizational pull or scheduling ability to orchestrate human collaboration effectively. The conversation then delves into specific skills that are becoming obsolete and those gaining importance. Ami reflects on spending decades honing skills in detailed UI/UX design (e.g., "where should we put this button?"). Now, she finds it's faster and more effective to "build three versions and try them out." This personal "innovator's dilemma" highlights the need to shed old competencies. New critical skills include a higher tolerance for change, adaptability, strong judgment, and relentlessness in the face of ambiguity. Ami encourages "framing the chaos" to make it feel safe and plausible to engage with, fostering a culture where constant adaptation is the norm. Mike adds that clear leadership and "DRIs" (Directly Responsible Individuals) are crucial to navigate this chaos, providing a steady hand amidst rapid shifts. Regarding *what* to build, Mike discusses the shift towards "agent-native" software. Initially, AI was seen as a sidebar feature, then integrated into existing features. The next stage is building products where agents can perform any action a human can. He notes that Anthropic uses Claude internally to create and iterate on UI for project tracking, demonstrating how interfaces can become "malleable" by agents. This raises questions about data provenance and how to make software reactive to organizational changes while maintaining user-friendliness. Mike advises companies to build the right "primitives" — foundational architectural layers that allow both humans and agents to interact seamlessly, enabling gradual evolution rather than bolted-on AI features. Ami addresses the challenge of balancing frontier exploration with providing a stable user experience. Anthropic's approach is to "meet everyone where they are," encouraging numerous "shots on goal" and parallel experiments. This allows them to discover what works before consolidating into a more stable product. Mike adds that success often comes from teams with high conviction and robust underlying infrastructure (like shared memory systems) that allow different experimental products to feel complementary rather than disconnected. The panel also touches on how to identify success and avoid "capability blind spots." Beyond traditional product-market fit metrics, Mike highlights the importance of "parking" projects that don't work with current models but revisiting them as new models are released. He gives an example of an internal computer-use product that was initially "so bad" but saw a breakthrough with Claude 3.7, demonstrating that models can improve in unexpected ways. Looking a year ahead, Ami hopes for increased individual and small team empowerment, making building easier for everyone and magnifying the impact of builders. Mike's aspiration is to close the gap between the capabilities of AI models and how most people use them. He envisions a future where advanced, multi-agent setups are not just for "Claude-pilled software engineers" but are accessible and democratized, empowering a broader range of users in their professional and personal lives.