丹·希珀 (Dan Shipper)，Every 的联合创始人兼首席执行官，提出了一种在他所谓的“移动前沿”上构建产品的策略。他反对在快速技术革命（特别是人工智能领域）中“按部就班的商业模式的不可理喻的低效性”。他强调了近期人工智能进步的变革力量，例如 Fable 5.1 和 Astra 6，展示了其从提示词创建历史准确的 3D 战斗场景、运行复杂的代理模拟以及执行高级视频编辑和演示动画等能力。 希珀指出了产品领导者面临的核心挑战：执行现有产品路线图与同时探索快速发展的技术前沿之间存在固有的冲突。探索是发散的、实验性的，并且涉及高淘汰率的理念，而执行则是收敛的、专注的，并要求可靠的交付。他警告不要仅仅依赖客户反馈，因为大多数客户不熟悉尖端人工智能能力，并期待产品团队提供方向。 他提出的解决方案是将研究实验室的元素整合到产品组织中，具体而言是组建专门的“实验室团队”。这种方法可以明确职责分工：产品团队可以专注于改进和扩展现有产品，而实验室团队则专注于探索新模型可能实现的功能。希珀强调，人工智能大大降低了进入门槛，即使是“一人实验室团队”也能利用强大的新工具进行大量探索。这种结构有效地利用了组织内部的“早期采用者”（那些已经在尝试新技术的个人），同时又不会分散整个产品团队的核心任务。 这两类团队的期望差异显著：实验室团队可能会预期淘汰 90% 的创新成果，而产品团队只会整合约 10% 的成功实验室创新。他引用了 Anthropic Labs 的例子，该公司通过实验团队开发了 Claude Code 和 Cloud Design 等主要产品；以及 OpenAI，其小型 Codex 团队的桌面应用最终并入 ChatGPT，成为其基础并扩展到 8 亿日活跃用户。 希珀概述了运营实验室的关键最佳实践： 1. **超小型团队：** 倡导“两片团队”（1-2人），与传统的“两块披萨团队”（8-10人）形成对比，这得益于人工智能提供的效率。 2. **海盗与架构师：** 将寻找价值的探索性“海盗”与将凌乱原型塑造成有价值、可扩展系统的“架构师”配对。 3. **内部试用 (Dogfooding)：** 为内部使用而构建，以创建最紧密的反馈循环，或者与早期采用者客户密切合作。 4. **并行实验：** 对一个问题采取多种甚至相互竞争的方法，以有效描绘未知的技术前沿。 5. **净正向投资回报：** 确保即使是被淘汰的实验也能提供价值，例如生成外部内容、为早期采用者项目提供信息，或向更广泛的产品团队介绍新功能。 为了将创新从实验室过渡到主要产品，希珀详细描述了一个“研究管道”。创意从仅限于实验室的实验进展到内部使用，然后是早期客户测试，最后到规模化发布。他通过一个内部项目来说明这一点，该项目旨在为其主编凯特 (Kate) 自动化文案编辑。最初，他使用 Fable 来复制她的编辑风格。一旦功能实现，它就转变为一个“Every 代理”供内部团队使用，对草稿进行“凯特式修订”。随后，一位架构师帮助将其完善成一个跟踪有效性的系统，显示凯特在代理介入后的编辑工作量减少了 12%。现在的目标是将这一流程提供给早期客户。 希珀最后推荐了管理此管道的最佳实践：定期审查、定义明确的推进决策标准（例如，内部采用、比现有解决方案提升 10 倍、规模化后的可承受性），以及将成功项目（“赢家”）并入主要产品。他断言，这一策略使组织能够主动构建其产品的下一版本，并以兴奋而非担忧的态度拥抱新的技术变革。

Dan Shipper, co-founder and CEO of Every, presented a strategy for building products on what he termed a "moving frontier," arguing against the "unreasonable ineffectiveness of business as usual" during rapid technology revolutions, particularly concerning AI. He highlighted the transformative power of recent AI advancements, such as Fable 5.1 and Astra 6, showcasing capabilities like creating historically accurate 3D battle scenes from prompts, running complex agent simulations, and performing advanced video editing and presentation animation. Shipper identified a core challenge for product leaders: the inherent conflict between executing an existing product roadmap and simultaneously exploring the rapidly evolving technological frontier. Exploration is divergent, experimental, and involves high rates of discarding ideas, while execution is convergent, focused, and demands reliable delivery. He cautioned against relying solely on customer feedback, as most customers are unfamiliar with cutting-edge AI capabilities and look to product teams for direction. His proposed solution is to integrate research lab elements into the product organization, specifically by forming dedicated "labs teams." This approach allows for a clear separation of concerns: product teams can focus on improving and scaling existing offerings, while labs teams concentrate on exploring what new models make possible. Shipper emphasized that AI significantly lowers the barrier to entry, enabling even a "labs team of one" to achieve substantial exploration with powerful new tools. This structure effectively harnesses "early adopters" within the organization—individuals already experimenting with new tech—without distracting the entire product team from its core mission. The expectations for these two types of teams differ significantly: labs teams might expect to discard 90% of their creations, while product teams would integrate only about 10% of successful lab innovations. He cited examples like Anthropic Labs, which developed major products such as Claude Code and Cloud Design through experimental groups, and OpenAI, where a small Codex team's desktop app eventually merged into ChatGPT, becoming its foundation and scaling to 800 million daily active users. Shipper outlined key best practices for running a lab: 1. **Very Small Teams:** Advocating for "two-slice teams" (1-2 people), contrasting them with the traditional "two-pizza team" (8-10 people), due to the efficiency AI offers. 2. **Pirates and Architects:** Pairing an exploratory "pirate" who seeks value with an "architect" who shapes messy prototypes into valuable, extensible systems. 3. **Dogfooding:** Building for internal use to create the tightest possible feedback loops, or collaborating closely with early adopter customers. 4. **Parallel Experiments:** Pursuing multiple, even competing, approaches to a problem to effectively map the unknown technological frontier. 5. **Net Positive ROI:** Ensuring that even discarded experiments provide value, such as generating external content, feeding early adopter programs, or informing the broader product team about new capabilities. To transition innovations from the lab to the main product, Shipper detailed a "research pipeline." Ideas progress from lab-only experiments to internal use, then to early customer testing, and finally to scaled release. He illustrated this with an internal project to automate copy edits for his editor-in-chief, Kate. Initially, he used Fable to replicate her editing style. Once functional, it transitioned to an "every agent" for internal team use, performing "Kate passes" on drafts. An architect then helped refine this into a system that tracks effectiveness, showing a 12% reduction in Kate's post-agent editing work. The goal is now to offer this process to early customers. Shipper concluded by recommending best practices for managing this pipeline: regular reviews, defining clear decision criteria for advancement (e.g., internal adoption, 10x improvement over existing solutions, affordability at scale), and merging successful projects ("winners") into the main product. This strategy, he asserted, enables organizations to proactively build the next version of their product and embrace new technological shifts with excitement rather than apprehension.