How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital

以下是内容的中文翻译： 本次活动汇集了80位投资组合公司创始人及AI领袖，旨在就“主权AI”或“掌控你的智能”这一适时主题，既发出“号召”，也提供技术层面的“指导”。演讲者强调，这并非要放弃像GPT或Opus这样强大且在许多任务中表现出色的闭源模型，而是鼓励公司为特定产品部件构建自己的AI能力，通过垂直整合来掌握更多属于自身的智能。 “主权AI”的概念意味着公司对其智能拥有完全的自主权，甚至包括模型权重，不依赖任何外部因素。这一转变正获得显著的推动力，Alex Karp、Satya Nadella和Jensen Huang等行业领袖都在倡导公司拥有而非仅仅租用其智能。核心信息是：“智能过于核心，过于基础，不应仅仅外包。” 演讲者提出了一个乐观的愿景，将未来由集中式、黑盒AI控制全球GDP的图景，与去中心化智能的图景进行对比。在去中心化模型中，公司在坚实的核心基础上，根据其独特数据、行业、个性化需求和工作流程，开发定制化的智能。这种方法促进了一个蓬勃发展的生态系统，其中个性化取得胜利，防止任何单一公司垄断智能。 公司追求主权AI的四个关键原因如下： 1. **成本：** 特别是对于低利润/负利润公司，随着产品规模扩大，不断上涨的AI成本可能会变得难以承受。 2. **速度：** 在编码或安全等领域，由于速度要求，小型、定制化精炼模型可以胜过大型通用模型。 3. **性能：** 这是一个相对较新的发展，开放模型在专有数据上经过适当微调后，现在可以在特定领域达到甚至超越闭源模型的性能。 4. **掌控自身命运：** 尽管目前的AI合作伙伴通常很优秀，但公司越来越希望拥有独立能力，以避免对外部供应商的依赖。 借鉴加密货币领域的类比（“非你之钥，非你之币”），演讲者提出了其“AI版本”：**“非你之权重，非你之产品。”** 这强调了真正的产品所有权需要对底层智能的掌控。竞争格局正从仅仅“应用层之争”（用户界面、市场推广）转向“智能层之争”，公司内部的应用研究（被称为“新实验室”——如Harvey、Factory、Glean）正在推动前沿创新。 为指导公司踏上这一征程，会议提出了一个四步框架： 1. **策略：定义哪些是自研，哪些是租用。** 这并非二元选择。影响这一决策的因素包括成本、速度/延迟、性能需求以及数据的专有性质。例如，从编程代理（通常因其强大的开箱即用性能而选择租用）到编程自动补全（通常出于速度和成本效率考虑而选择自研）。网络安全和生物科技公司也倾向于自研，以获得速度、性能和利用专有数据。 2. **团队：组建实验室。** 演讲者建议不要将主权AI的努力生硬地塞进现有的AI平台团队。相反，公司应考虑组建小型、专注的“从零开始的团队”（de novo teams），专注于前沿研究，充当内部实验室。这些团队应被授权主动出击，而不仅仅是为其他部门提供服务。 3. **可见性（Legibility）：** 这一常被忽视的步骤至关重要。公司必须通过出色的技术营销、品牌实验室和发表高质量研究来掌控其AI能力的话语权。这有助于他们在寻求先进AI合作伙伴的买家面前脱颖而出。 4. **技术路线图：** 一般的路线包括： * **定义评估（Evals）：** 这是一个不起眼但至关重要的第一步，用于准确衡量智能。 * **模型路由与协调器（Harnesses）：** 试验这些工具以引导查询和整合模型。 * **后训练（Post-training）：** 在特定数据集上微调开放模型（最常见）。 * **中训练/预训练（Mid-training/Pre-training）：** 更高级、更罕见的情况，从头开始训练模型或对其进行重大修改。 * **实时数据反馈循环：** 建立系统，让客户互动持续改进模型的智能。 演讲者强调，随着强大开源权重模型（例如Kimi K3、GLM 5.2）的出现，公司可以从接近前沿的基线开始。结合强大的技术路线图，包括扎实的后训练、提示/协调器工程以及在线学习，公司现在可以通过拥有自己的技术栈，实现“超越前沿的性能”。 技术栈概览区分了生产栈（通过协调器和模型面向用户的智能）和开发栈（用于改进智能的工具）。尽管闭源模型生态系统提供了一个更简单、下限更高但上限更低的技术栈，但主权AI则涉及更复杂的开发栈。这包括选择开源基础、广泛的后训练、定制协调器、各种形式的上下文（向量数据库、知识图谱、新颖编码），以及关键的开发工具，如全面的评估、高质量数据生成（专家轨迹、合成数据、强化学习环境）和在线学习机制，以确保持续改进。 本次活动的议程包括来自多位专家的深度技术研讨会，涵盖后训练、协调器与评估、强化学习环境与合成数据、在线学习，以及来自Harvey公司的全栈案例研究，展示了构建完整AI技术栈的实用方法。演讲者最后感谢各位高水平的演讲嘉宾为这次关键活动所做的贡献。

This event, bringing together 80 portfolio company founders and AI leaders, serves as both a "rallying call and a technical how-to" on the timely topic of "Sovereign AI" or "owning your intelligence." The speaker emphasizes that this isn't about abandoning powerful closed models like GPT or Opus, which remain excellent for many tasks, but rather for companies to build their own AI capabilities for specific product parts, vertically integrating to own more of their intelligence. The concept of Sovereign AI implies that companies possess total self-governance over their intelligence, down to the weights, without external dependencies. This shift is gaining significant momentum, with industry leaders like Alex Karp, Satya Nadella, and Jensen Huang advocating for companies to own their intelligence rather than merely rent it. The core message is that "intelligence is too core, too fundamental of a property to just outsource." The speaker presents an optimistic view, contrasting a future of centralized, black-box AI controlling global GDP with one of decentralized intelligence. In the decentralized model, while building on a solid core, individual companies develop bespoke intelligence tailored to their unique data, industries, personalization needs, and workflows. This approach fosters a flourishing ecosystem where individuality triumphs, preventing any single company from monopolizing intelligence. Four key reasons are highlighted for companies to pursue Sovereign AI: 1. **Cost:** Especially for low/negative-margin companies, where escalating AI costs can become prohibitive as products scale. 2. **Speed:** In domains like coding or security, small, custom-distilled models can outperform larger general ones due to speed requirements. 3. **Performance:** A relatively new development, where open models, when properly fine-tuned on proprietary data, can now achieve or even surpass the performance of closed models in specific domains. 4. **Controlling Your Own Destiny:** While current AI partners are often excellent, companies increasingly desire independent capabilities to avoid reliance on external vendors. Drawing an analogy from the crypto space ("not your keys, not your crypto"), the speaker introduces the "AI version": **"not your weights, not your product."** This underscores the idea that true product ownership requires control over the underlying intelligence. The competitive landscape is shifting from merely a "race for the application layer" (UI, go-to-market) to a "race for the intelligence layer," where applied research within companies (dubbed "Neolabs" like Harvey, Factory, Glean) is driving frontier-level innovation. To guide companies on this journey, a four-step framework is proposed: 1. **Strategy: Define What to Own vs. Rent.** This isn't binary. Factors influencing this decision include cost, speed/latency, performance needs, and the proprietary nature of data. Examples range from coding agents (often rented due to strong out-of-box performance) to coding autocomplete (often owned for speed and cost efficiency). Cybersecurity and bio companies also lean towards ownership for speed, performance, and leveraging proprietary data. 2. **Team: Assemble a Lab.** The speaker advises against shoehorning Sovereign AI efforts into existing AI platform teams. Instead, companies should consider forming small, dedicated "de novo teams" focused on frontier research, acting as internal labs. Such teams should be empowered to play offense, not just service other groups. 3. **Legibility:** This often overlooked step is crucial. Companies must control the narrative around their AI capabilities through excellent technical marketing, branded labs, and publishing high-quality research. This helps them stand out to buyers seeking sophisticated AI partners. 4. **Technical Roadmap:** A general journey includes: * **Defining Evals:** An unglamorous but critical first step to measure intelligence accurately. * **Model Routers & Harnesses:** Experimenting with these to direct queries and integrate models. * **Post-training:** Fine-tuning open models on specific datasets (most common). * **Mid-training/Pre-training:** More advanced, rarer cases of training models from scratch or significantly modifying them. * **Live Data Feedback Loops:** Establishing systems where customer interactions continuously improve the model's intelligence. The speaker emphasizes that with the advent of powerful open-weight models (e.g., Kimi K3, GLM 5.2), companies can start with a near-frontier baseline. Combined with a robust technical roadmap including strong post-training, prompt/harness engineering, and online learning, companies can now achieve "better than frontier performance" by owning their stack. A technical stack overview differentiates between the production stack (user-facing intelligence via harnesses and models) and the development stack (tools for improving intelligence). While closed model ecosystems offer a simpler, higher-floor but lower-ceiling stack, Sovereign AI involves a more complex development stack. This includes choosing an open-source base, extensive post-training, custom harnesses, various forms of context (vector databases, knowledge graphs, novel encoding), and crucial development tools like comprehensive evals, high-quality data generation (expert trajectories, synthetic data, RL environments), and online learning mechanisms to ensure continuous improvement. The event's agenda features deep-dive technical workshops from various experts, covering post-training, harnesses and evals, RL environments and synthetic data, online learning, and a full-stack case study from Harvey, demonstrating a practical approach to building a complete AI stack. The speaker concludes by thanking the high-caliber speakers for their contributions to this pivotal event.