OpenAI Codex lead on the new shape of product work | Andrew Ambrosino

以下是内容的中文翻译： OpenAI Codex 应用的产品与工程负责人安德鲁·安布罗西诺（Andrew Ambrosino），就人工智能如何从根本上重塑产品开发和工作本身的性质，分享了深刻的见解。随着 Codex 呈现指数级增长，并在 OpenAI 内部实现了近乎 100% 的采用率，安布罗西诺的团队正处于这场转型的最前沿。 **产品开发的“颠覆”** 安布罗西诺认为，人工智能“颠覆”了传统的产品开发流程。过去，实施成本高昂，需要在构建之前进行大量的研究、构思和原型设计来规避风险。而现在，有了人工智能，实施成本变得低廉；借助强大的模型，“任何人都可以从零开始构建任何东西”。这导致了原型的激增——安德鲁开玩笑说，有“90个不同且互不协调的团队”正在构建同一个功能的不同版本。 在这个新范式中，挑战从“构建”转向了“筛选”。昂贵的部分不再是实施，而是“品味”——即知道什么是好的、如何组合，以及如何框定和完善无数的探索成果。安布罗西诺澄清说，产品需求文档 (PRD) 并未消亡；相反，沟通的“媒介”必须与目标相匹配。文档可能最适合澄清模糊区域，而原型则用于对交互进行压力测试。危险在于，错将一个精美的原型误认为是可投入生产的功能，而它可能仍然只是一项早期探索。 **“品味”的细微之处与人工智能的设计挑战** 对安布罗西诺而言，“品味”超越了美学。它包含系统思维，即理解一个功能如何融入更广泛的产品，并与战略目标保持一致。尽管人工智能模型在代码生成方面表现出色，但它们目前在设计能力方面有所滞后。这部分是因为设计更难评判（依赖人类品味），而且人工智能研究最初优先关注那些能加速人工智能自身发展的领域（如编码）。更深层次的原因是，人工智能难以满足设计对新颖性的需求，以及对抽象层次（即视觉元素如何与底层代码结构相关联）的深刻理解。人类对于创造性飞跃和理解驱动真正有影响力设计的用户心理仍然至关重要。 **演变中的角色与团队结构** OpenAI 呈现出显著的“角色融合”，设计师编写代码，产品经理拥有强大的技术技能。安布罗西诺认为角色之间有更多重叠，其定义取决于个人工作的“平均”性质而非严格的界限。尽管他乐见“这不关你的事”这种心态的瓦解，但他警告不要完全取消这些角色，因为这可能会放弃每个学科所培养的专业知识和最佳实践。例如，产品经理现在像在打“区域防守”一样运作，筛选混乱，指导零散的工作，并确保高度“自主能动”的员工队伍保持一致。招聘优先考虑具有“高度自主能动性、高品味”的个人，他们能够将一个想法从构思到完成。 **人工智能优先世界中的产品战略** 规划周期已大幅缩短。长期计划必须保持模糊，因为模型能力演进如此之快，以至于一个精确的9个月计划将是“虚假的精确性”。一个关键策略是构建那些可能“尚未”完全奏效但预见到未来模型改进的雄心勃勃的功能。安布罗西诺引用了早期 Codex 网页版因模型智能不足而失败的例子，而后来一个形状相同但本地部署的版本却成功了。教训是：不要固执；一个功能可能只是“尚未”为当前模型做好准备。 **Codex 作为安德鲁的个人助手以及对应用的愿景** 安布罗西诺积极地将 Codex 作为他的主要工作工具，并随着角色变化而调整其使用方式。最初，它是一个用于编写代码的开发工具。现在，它协助进行产品发现、团队协调和信息综合，从数千个 Slack 频道和电子邮件中创建自动化的每日简报。他“指导”人工智能优化这些流程。 Codex（及其最终与 ChatGPT 的结合）的愿景不是一个取代所有其他应用的“超级应用”，而是一个“根据地”。它是一个中心枢纽，用户在此开始和结束工作，自动化任务，并与专业工具协调。安德鲁分享了一个非凡的故事：一位视频编辑师使用 Codex 编辑视频，随后它“自行构建了一个”Adobe Premiere Pro 扩展，以促进更深层次的交互。这阐明了目标：与现有工具的无缝交互，无论是通过直接集成、计算机操作还是自定义扩展，使人工智能成为所有知识工作中的一个积极主动的伙伴。 **从失败中吸取教训** 安布罗西诺坦诚地分享了他作为创业公司创始人的过往经历，描述了在取得当前成功之前“持续失败”的那些年。他强调韧性、持续学习和适应自身流程至关重要。OpenAI 直截了当、常常是批判性的内部反馈文化（例如，在 Slack 上长达2000条消息的讨论中，有人直言某个想法“很愚蠢”），被视为完善产品和加速进展的关键机制。 归根结底，从 OpenAI 的视角来看，未来的工作将是：人的自主能动性、品味以及适应不断变化的人工智能能力至关重要，从而培育一个动态的环境，让个人在所有领域推动产品创造和自动化。

Andrew Ambrosino, Product and Engineering Lead for the Codex app at OpenAI, shares profound insights into how AI is fundamentally reshaping product development and the nature of work itself. With Codex seeing exponential growth and nearly 100% internal adoption at OpenAI, Ambrosino's team is at the forefront of this transformation. **The Inversion of Product Development** Ambrosino argues that AI has "inverted" the traditional product development process. Historically, implementation was expensive, necessitating extensive research, ideation, and prototyping to de-risk before building. Now, with AI, implementation is cheap; "anybody can build anything" from scratch using powerful models. This leads to a proliferation of prototypes—Andrew jokes about "90 different uncoordinated teams" building variations of the same feature. In this new paradigm, the challenge shifts from *building* to *curating*. The expensive part is no longer implementation, but "taste" – knowing what's good, what to combine, and how to frame and refine the myriad of explorations. Ambrosino clarifies that PRDs (Product Requirements Documents) are not dead; rather, the *medium* of communication must match the goal. A document might be best for clarifying a vague area, while a prototype serves to stress-test an interaction. The danger lies in mistaking a polished prototype for a production-ready feature, when it might still be an early exploration. **The Nuance of "Taste" and AI's Design Challenge** "Taste" for Ambrosino extends beyond aesthetics. It encompasses systems thinking, understanding how a feature fits into the broader product, and aligning with strategic goals. While AI models excel at code generation, they currently lag in design capabilities. This is partly because design is harder to grade (relying on human taste) and AI research initially prioritized areas that accelerate AI itself (like coding). More profoundly, AI struggles with design's need for novelty and its deep understanding of abstraction layers – how visual elements relate to underlying code structure. Humans remain crucial for creative leaps and understanding user psychology that drives truly impactful design. **Evolving Roles and Team Structure** OpenAI exhibits significant "role collapse," where designers write code, and product managers possess strong technical skills. Ambrosino sees more overlap, with roles defined by the *average* of one's work rather than strict boundaries. While welcoming the breakdown of "not your lane" mentalities, he cautions against eliminating roles entirely, as this risks discarding specialized knowledge and best practices that each discipline has cultivated. Product managers, for instance, now function in a "zone defense," curating chaos, guiding disparate efforts, and ensuring alignment across a highly "agentic" workforce. Hiring prioritizes "high agency, high taste" individuals capable of driving an idea from inception to completion. **Product Strategy in an AI-First World** Planning horizons have drastically shortened. Long-term plans must remain hazy because model capabilities evolve so rapidly that a precise 9-month plan would be "false precision." A key strategy is to build ambitious features that may not fully work *yet*, anticipating future model improvements. Ambrosino cites the example of an early Codex web version that failed due to insufficient model intelligence, while a later, identically shaped local version succeeded. The lesson: don't be stubborn; a feature might simply not be "ready yet" for the current model. **Codex as Andrew's Personal Assistant and the Vision for the App** Ambrosino actively uses Codex as his primary work tool, evolving his usage as his role shifts. Initially, it was a development tool for writing code. Now, it assists with product discovery, team coordination, and information synthesis, creating automated daily briefs from thousands of Slack channels and emails. He "coaches" the AI to refine these processes. The vision for Codex (and its eventual combination with ChatGPT) is not a "super app" that replaces all others, but a "home base." It's a central hub where users start and end their work, automate tasks, and coordinate with specialized tools. Andrew shares a remarkable story of a videographer using Codex to edit videos, which then *built its own extension* for Adobe Premiere Pro to facilitate deeper interaction. This illustrates the goal: seamless interaction with existing tools, whether through direct integration, computer use, or custom extensions, making the AI a proactive partner across all knowledge work. **Lessons from Failure** Ambrosino candidly shares his past experiences as a startup founder, describing years of "constant failure" before reaching his current success. He emphasizes that resilience, continuous learning, and adapting one's process are crucial. OpenAI's culture of direct, often critical, internal feedback (like 2,000-message Slack threads calling an idea "stupid") is seen as a vital mechanism for refining products and accelerating progress. Ultimately, the future of work, as seen from OpenAI, is one where human agency, taste, and the ability to adapt to ever-changing AI capabilities are paramount, fostering a dynamic environment where individuals drive product creation and automation across all domains.