Sequoia's Huang Urges AI Startups to Own Model Weights Over APIs

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

Sequoia Capital

Open-weight models combined with domain-specific fine-tuning now outperform closed frontier APIs while lowering long-term compute costs and platform risk.

AI-assisted summary of How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Sequoia Capital on August 11, 2026, this video features partner Sonya Huang opening the "Own Your Intelligence" event. According to the page description, Huang presents the strategic and technical case for why companies are increasingly choosing to own their AI intelligence down to the model weights, driven by cost, speed, performance, and controlling their own destiny.

## TL;DR Sonya Huang outlines an opinionated framework for tech companies to transition from renting closed AI models to owning custom, sovereign model weights. She explains how decentralized intelligence and open-weight models allow domain-specific applications to achieve better-than-frontier performance. The talk details a practical four-step execution roadmap covering strategic selection, specialized staffing, research legibility, and technical stack architecture. Ultimately, Huang argues that owning the intelligence layer is essential for controlling product experience and building long-term competitive advantage.

## Atmosphere & style The presentation carries an energetic, rallying, and highly pragmatic tone, blending the conviction of an industry manifesto with the rigor of an executive technical briefing. Sonya Huang speaks with direct, confident authority, setting a high-substance mood focused on actionable strategy rather than empty hype. The experience feels like an exclusive insider masterclass for founders, balancing strategic vision with grounded technical blueprints.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) Sovereign AI means a company owns its own intelligence down to the model weights without external dependencies, though it coexists alongside closed frontier APIs used for general desktop or coding tasks. - [01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) Industry leaders like Alex Karp, Satya Nadella, and Jensen Huang advocate for corporate AI sovereignty under the belief that intelligence is too core to be completely outsourced. - [01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) Decentralized intelligence represents an optimistic future where companies build bespoke AI tailored to their unique data and workflows rather than relying on a single black-box provider. - [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) The four main drivers for transitioning from renting to owning model weights are reducing compute/inference cost, improving speed and latency, maximizing domain performance, and controlling company destiny. - [04:22](https://www.youtube.com/watch?v=bMMv0bZzONg&t=262s) The competitive landscape has shifted from controlling user interfaces to competing for the intelligence layer itself, summarized by the mantra 'Not your weights, not your product.' - [05:32](https://www.youtube.com/watch?v=bMMv0bZzONg&t=332s) Applied AI application startups (such as Harvey, Factory, Glean, and Ramp) are becoming the new 'neo labs,' conducting frontier-level research in domain-specific AI. - [07:05](https://www.youtube.com/watch?v=bMMv0bZzONg&t=425s) Deciding what to own versus rent depends on evaluating four variables: cost sensitivity, latency demands, domain-tuned performance potential, and proprietary data confidentiality. - [09:51](https://www.youtube.com/watch?v=bMMv0bZzONg&t=591s) Sovereign AI teams should be built de novo as small, dedicated research units rather than retrofitting existing platform infrastructure teams. - [11:17](https://www.youtube.com/watch?v=bMMv0bZzONg&t=677s) Legibility—making technical research visible through high-taste papers, custom benchmarks, and dedicated branded labs—is crucial for winning enterprise customers. - [12:33](https://www.youtube.com/watch?v=bMMv0bZzONg&t=753s) A complete technical roadmap progresses from defining strategy and building evals to implementing prompt harnesses, post-training open weights, and setting up automated data feedback loops. - [12:33](https://www.youtube.com/watch?v=bMMv0bZzONg&t=753s) Recent open-weight models like Kimi k3 and GLM-52 enable companies to start near frontier capability and achieve better-than-frontier performance through domain customization. - [13:56](https://www.youtube.com/watch?v=bMMv0bZzONg&t=836s) Moving to a sovereign AI stack unlocks a significantly higher performance ceiling than closed APIs by leveraging proprietary production data, despite higher operational complexity. - [15:36](https://www.youtube.com/watch?v=bMMv0bZzONg&t=936s) Advanced context and memory management involve vector databases, enterprise knowledge graphs, open connectors like MCP, and novel research into encoding context directly into model weights. - [15:36](https://www.youtube.com/watch?v=bMMv0bZzONg&t=936s) Establishing a production development stack requires robust evaluation frameworks, real-time drift monitoring, and rich post-training pipelines using expert trajectories, synthetic data, and RL environments.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) What Is Sovereign AI (and What It Isn't)

Sonya Huang, partner at Sequoia Capital, opens the "Own Your Intelligence" event welcoming approximately 80 portfolio company founders and AI leaders. She frames the gathering as half a rallying call and half a technical "how-to," structured around high-substance talks and demos to move beyond theoretical discussions into practical execution. Setting the stage, Huang defines **sovereign AI** as a company owning its own intelligence down to the model weights without relying on external dependencies—drawing an analogy to a sovereign state that maintains total self-governance.

Huang stresses a critical nuance: Sequoia is *not* advising companies to abandon closed frontier models like Claude 3.5 Opus or GPT-4. Closed-model APIs remain wonderful for tasks like general desktop work or coding agents. However, Sequoia has observed a growing trend of companies wanting to vertically integrate and build proprietary AI capabilities into core parts of their products.

- **Event scale:** ~80 portfolio company founders and AI leaders in attendance. - **Sovereign AI definition:** Owning complete intelligence capabilities without external dependencies, down to the model weights. - **Coexistence of models:** Closed APIs (e.g., Opus, GPT) remain ideal for general desktop work and coding agents; sovereign models are built for core proprietary product features.

[01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) Centralized vs. Decentralized Intelligence

The shift toward sovereign AI has gained massive momentum across the tech industry. Huang points out that over the past month, leaders like Alex Karp (Palantir) and Satya Nadella (Microsoft) have voiced support for corporate AI sovereignty. Furthermore, Jensen Huang (NVIDIA) recently led the effort to ensure open-weight models remain accessible in the United States, receiving near-unanimous industry support. The consensus is clear: intelligence is too fundamental a property for a company to outsource entirely.

Huang contrasts two visions of the future of AI:

1. **Centralized Intelligence:** A single, all-powerful AI operating as a black box that powers an increasing portion of global GDP while absorbing all the data exhaust and flywheels from the rest of the world. 2. **Decentralized Intelligence:** The world builds on top of a shared solid core, but every individual and company creates bespoke intelligence tailored to their own data, industry requirements, workflows, and taste.

Huang argues that decentralized intelligence offers a far more optimistic world view where individuality triumphs and no single tech giant swallows the entire market.

- **Key industry figures supporting sovereign AI:** Alex Karp, Satya Nadella, and Jensen Huang. - **Core philosophical argument:** Intelligence is too core and fundamental to be treated as a purely outsourced property.

[02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) Four Reasons Companies Own Their Models: Cost, Speed, Performance, Destiny

Working with dozens of companies building their own AI, Sequoia highlights the four primary drivers behind the transition from renting to owning model weights:

* **1. Cost:** For low-, zero-, or negative-margin businesses, sovereign AI is a necessity rather than a luxury. Ironically, as an AI product becomes more successful, its compute and inference costs skyrocket. Consequently, the most advanced AI deployers are often the first to build custom models. * **2. Speed and Latency:** In latency-sensitive domains such as interactive coding or real-time cybersecurity, a small, distilled custom model can consistently outperform a large general-purpose model simply because inference speed is paramount. * **3. Performance:** Unlike last year, when companies rarely chose open weights for superior performance, open-weight models tuned on proprietary domain data can now outperform closed frontier APIs on specific tasks. * **4. Controlling Your Own Destiny:** While foundation model labs like Anthropic and OpenAI have been great partners, companies increasingly want independent technical foundations rather than total platform risk.

> "Sovereign AI isn't a nice-to-have, it's a must-have... Ironically, the more successful your AI product is, the higher your AI COGS tend to be."

[04:22](https://www.youtube.com/watch?v=bMMv0bZzONg&t=262s) "Not Your Weights, Not Your Product"

Referencing a slide presented by Pat Grady at Sequoia's AI Ascent event regarding the "race for the application layer," Huang explains how the competitive landscape is shifting. Historically, foundation model labs (building top-down) and application software companies (building bottom-up from the user experience) were competing to own the end-user interface.

Today, that battleground has evolved into a **race for the intelligence layer itself**. Product and intelligence have merged: the product *is* the intelligence. To truly own the user experience and shape product capabilities, companies must control the underlying intelligence rather than just offering a UI wrapper or go-to-market layer.

Drawing a parallel to the classic crypto DeFi meme—*"Not your keys, not your crypto"*—Huang introduces the AI equivalent: **"Not your weights, not your product."**

[05:32](https://www.youtube.com/watch?v=bMMv0bZzONg&t=332s) Application Companies Are the Newest Neo Labs

Huang asserts that the hottest new research labs are not just the traditional AGI research institutions, but rather the applied research groups emerging within leading AI application companies. These teams are conducting frontier-level work across evals, custom benchmarks, harness engineering, and novel algorithmic fine-tuning techniques.

Examples of companies operating as these "neo labs" include: * **Harvey** (legal AI) * **Factory** (software engineering agents) * **Glean** (enterprise search and knowledge) * **Open Evidence** (medical AI) * **Semgrep** (code security) * **Ramp** (financial automation)

These democratized intelligence efforts demonstrate that applied domain focus can yield research breakthroughs competitive with major foundation labs.

[07:05](https://www.youtube.com/watch?v=bMMv0bZzONg&t=425s) Step 1: Deciding What to Own vs. Rent

Huang outlines an opinionated framework for going from zero to one in building sovereign intelligence, emphasizing that sovereignty is not binary (0% or 100%). Companies must strategically draw lines around which AI capabilities to own versus rent based on four primary variables: **Cost**, **Speed/Latency**, **Performance**, and **Proprietary Data**.

``` ┌─────────────────────────────────────────────────────────┐ │ SOVEREIGN AI STRATEGY FRAMEWORK │ └─────────────────────────────────────────────────────────┘ │ ┌───────────────────┬──────────────┴──────────────┬───────────────────┐ ▼ ▼ ▼ ▼ [ COST ] [ SPEED ] [ PERFORMANCE ] [ PROPRIETARY DATA ] How critical is Is low latency Can domain-tuned open Does the domain rely this line item to a P0 requirement weights outperform on highly secret or overall COGS? for the UX? closed APIs? bespoke data? ```

- **Coding (Agents vs. Auto-complete):** Coding agents are largely rented via closed APIs today because out-of-the-box reasoning is critical and latency is less constrained. Conversely, **Tab auto-complete** models rely on sovereign intelligence because sub-100ms speed is essential and API volume makes closed models cost-prohibitive. - **Cybersecurity:** Companies (such as stealth portfolio startups like Alon's) own their models primarily for extreme speed, tailored post-training, and domain-specific detection performance. - **Bio & Life Sciences:** Companies build sovereign models due to the extreme value and confidentiality of proprietary biological and chemical data.

[09:51](https://www.youtube.com/watch?v=bMMv0bZzONg&t=591s) Step 2: Build the Team (and Don't Shoehorn Your Platform Team)

When staffing a sovereign AI initiative, there is no single archetype for a labs leader. Huang highlights two different profiles: * **Nico** (Research-led background): Former researcher at Apple and Google Brain. * **Alex** (Engineering-led background): Engineering background across Microsoft and Ramp.

Crucially, Huang warns against forcing an existing **AI Platform Team** (which traditionally operates in a hub-and-spoke model servicing various product features) to handle sovereign model development. A infrastructure platform team is structured to service internal requests, whereas a sovereign AI team needs to play offense and drive frontier research.

- **Team structure advice:** Start de novo (from scratch) with a small, focused unit rather than retrofitting platform engineers. - **Efficiency example:** Harvey achieved significant research output and published benchmarks with a focused team of just **seven people**.

[11:17](https://www.youtube.com/watch?v=bMMv0bZzONg&t=677s) Step 3: Legibility – Why Your Research Has to Be Visible

Huang stresses that **legibility**—making internal technical capabilities visible to the outside world—is deeply underrated. She cites Winston Weinberg's perspective that a CEO has two main responsibilities: driving substantive results and controlling the narrative around what is being built.

In the current enterprise landscape, buyers are actively selecting their long-term AI vendors. Because every startup pitches similar claims, buyers look for clear markers of technical depth to determine who can actually execute. Achieving legibility requires: * Executing high-quality technical marketing. * Establishing dedicated, branded research groups or labs (e.g., "Company Labs"). * Publishing high-taste research papers, evals, and engineering blogs.

[12:33](https://www.youtube.com/watch?v=bMMv0bZzONg&t=753s) Step 4: The Technical Roadmap

The technical journey toward owning intelligence generally follows a structured multi-step sequence:

1. **Strategy:** Define which model capabilities to build versus rent. 2. **Evals:** Build robust, domain-specific evaluation suites. *(Huang notes this is unglamorous, tedious work, but foundational to all downstream success).* 3. **Harnesses & Model Routers:** Implement routing logic and specialized prompting harnesses around models. 4. **Training (Post-Training $\rightarrow$ Mid/Pre-Training):** Fine-tune open weights via post-training; transition into mid-training or pre-training only if required by the domain. 5. **Data Feedback Loops:** Set up automated pipelines so live customer interaction data continuously feeds back into model retraining.

Recent open-weight models—specifically **Kimi k3** and **GLM-52**—have shifted what is possible. Because open weights are fully malleable, teams can start near frontier-level quality and layer on domain post-training, prompt harness engineering, and online learning to ultimately **exceed frontier API performance**.

> "You start with a baseline that's already close to frontier, and then with a good enough technical roadmap... you can actually reach better than frontier performance by owning your stack. This is new for 2026."

[13:56](https://www.youtube.com/watch?v=bMMv0bZzONg&t=836s) The Stack: Production vs. Development

Huang illustrates the infrastructure architecture required for sovereign AI, dividing it into two parallel stacks:

* **Production Stack (User-Facing):** The real-time inference pipeline that generates every token seen by the end user. At its core, it consists of a **harness on top of a model**. * **Development Stack (Offline/Research):** The internal tooling, evaluation frameworks, and data pipelines used to train, refine, and test intelligence before deployment.

``` ┌────────────────────────────────────────────────────────────────────────┐ │ CLOSED MODEL STACK (Renting) │ │ High floor, lower performance ceiling due to lack of data customization│ ├──────────────────────────────────┬─────────────────────────────────────┤ │ PRODUCTION STACK │ DEVELOPMENT STACK │ │ • Closed API (e.g., Opus, GPT) │ • Basic prompt engineering │ │ • Basic out-of-the-box harness │ • Standard eval scripts │ └──────────────────────────────────┴─────────────────────────────────────┘

┌────────────────────────────────────────────────────────────────────────┐ │ SOVEREIGN AI STACK (Owning) │ │ Requires higher complexity, but unlocks uncapped performance edge │ ├──────────────────────────────────┬─────────────────────────────────────┤ │ PRODUCTION STACK │ DEVELOPMENT STACK │ │ • Open-source base weights │ • Fine-tuning & post-training │ │ • Custom open/configurable │ • Synthetic data generation │ │ harness (logic, tools) │ • Continuous feedback loops │ └──────────────────────────────────┴─────────────────────────────────────┘ ```

Moving from a closed API to owning the stack is like "opening Pandora's box": while closed APIs offer a simple setup with a high floor, they capped performance because teams cannot utilize proprietary online production data to continuously improve the core weights. Owning the stack requires managing base model selection, post-training pipelines, harness logic, tool integration, and context window optimization—unlocking a significantly higher performance ceiling.

[15:36](https://www.youtube.com/watch?v=bMMv0bZzONg&t=936s) Opening Pandora's Box: Context, Memory, and the Development Stack

Sonya Huang transitions into discussing the foundational technologies that power modern AI architectures, specifically emphasizing that context management requires going beyond simple prompt engineering. She highlights components like Turbopuffer for vector database infrastructure, Glean for enterprise knowledge graphs, and open-source connectivity enabled through the Model Context Protocol (MCP). Huang also calls attention to cutting-edge developments in model memory, citing Dan Biederman from N-Gram, who is present at the event, for his company's novel research into encoding context directly into model weights rather than relying solely on context windows. She encourages attendees interested in memory architecture to seek out Biederman during break sessions.

Beyond the real-time production stack, Huang underscores that owning your intelligence elevates the development stack to a critical strategic asset. Companies must implement rigorous evaluation frameworks to measure their custom intelligence against baseline benchmarks and actively monitor for model drift in production environments. Fueling these models demands high-quality post-training data, which can take the form of expert trajectories, synthetic data, or specialized reinforcement learning (RL) environments. Ultimately, the end goal is establishing continuous online learning feedback loops, ensuring that the system's underlying intelligence improves natively with every user interaction.

- Core components of the expanded context and memory architecture include vector databases like Turbopuffer, enterprise knowledge graphs like Glean, and open-source MCP connectors. - Advanced research from N-Gram involves encoding context directly into model weights to evolve memory management. - Production readiness requires real-time monitoring of evals and model performance drift. - Model post-training relies on three main data pipelines: expert trajectories, synthetic datasets, and RL environments. - Continuous online learning is highlighted as the mechanism that guarantees compounding intelligence from everyday user behavior.

[16:35](https://www.youtube.com/watch?v=bMMv0bZzONg&t=995s) Agenda and Workshop Lineup for Building Sovereign AI

Huang concludes her opening presentation by walking through the event's hands-on workshop agenda, which has been explicitly designed to cover every technical layer of owning an AI stack short of pre-training models from scratch. She introduces a sequence of technical deep dives delivered by key industry leaders and ecosystem partners. Lin Qiao from Fireworks AI leads a workshop dedicated to post-training techniques, while Harrison Chase from LangChain hosts a session on orchestration harnesses and evaluation strategies. Brendan Foody from Mercor covers the intricacies of RL environments and synthetic data generation, and the team from Trajectory leads a deep dive into implementing online learning systems.

To tie all these individual architectural pieces together into a cohesive strategy, legal AI company Harvey presents a masterclass on how they designed and built their end-to-end intelligence stack. Huang notes that Harvey had announced "Harvey Research" just the day before the event, with several of their primary technical partners participating directly in the day's workshop schedule. She wraps up her remarks by expressing pride in assembling top-tier speakers from both within and outside the Sequoia Capital portfolio to give founders a practical blueprint for owning their intelligence layer.

- Lin Qiao (Fireworks AI) leads the session on post-training models. - Harrison Chase (LangChain) runs the workshop covering execution harnesses and evals. - Brendan Foody (Mercor) guides the workshop on synthetic data and RL environments. - Trajectory conducts the deep dive into setting up online learning frameworks. - Harvey delivers a comprehensive case study on full-stack AI strategy, following the fresh announcement of Harvey Research.

## Notable quotes > [04:22](https://www.youtube.com/watch?v=bMMv0bZzONg&t=262s) Not your weights, not your product — Sonya Huang

> [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) Sovereign AI isn't a nice-to-have, it's a must-have... Ironically, the more successful your AI product is, the higher your AI COGS tend to be. — Sonya Huang

> [12:33](https://www.youtube.com/watch?v=bMMv0bZzONg&t=753s) You start with a baseline that's already close to frontier, and then with a good enough technical roadmap... you can actually reach better than frontier performance by owning your stack. This is new for 2026. — Sonya Huang

## Mentioned Sonya Huang, Sequoia Capital, Claude 3.5 Opus, GPT-4, Alex Karp, Palantir, Satya Nadella, Microsoft, Jensen Huang, NVIDIA, Pat Grady, Anthropic, OpenAI, Harvey, Factory, Glean, Open Evidence, Semgrep, Ramp, Apple, Google Brain, Winston Weinberg, Kimi k3, GLM-52, Turbopuffer, Model Context Protocol (MCP), Dan Biederman, N-Gram, Lin Qiao, Fireworks AI, Harrison Chase, LangChain, Brendan Foody, Mercor, Trajectory

[Source](https://www.youtube.com/watch?v=bMMv0bZzONg)

中文简报

红杉资本 Sonya Huang 呼吁 AI 初创公司掌握模型权重而非依赖 API 开放权重模型结合特定领域微调，目前表现已超越闭源前沿 API，同时降低了长期算力成本和平台风险。 Sonya Huang 提出了一个明确的框架，指导科技公司从租用闭源 AI 模型转向拥有定制的自主模型权重。她阐释了去中心化智能与开放权重模型如何使特定领域的应用实现超越前沿模型的表现。演讲详细介绍了一份涵盖战略选型、专业人才配置、研究理解能力以及技术栈架构的四步实用执行路线图。Huang 最终指出，掌握智能层对于掌控产品体验和构建长期竞争优势至关重要。 红杉资本合伙人 Sonya Huang 提出“主权 AI（Sovereign AI）”概念，指企业在模型权重层面拥有独立智能而不依赖外部；行业共识正趋向去中心化智能，认为智能过于核心而不能完全外包给单一黑盒服务商。 (0:00) 推动企业从“租用”转向“拥有”模型权重的四大核心驱动力为成本、速度与延迟、领域性能及掌控自身命运；当下的竞争焦点已从前端交互层转向智能层本身，即“没有模型权重，就没有真正的产品”。 (2:54) 应用层企业（如 Harvey、Glean、Ramp 等）正成为开展垂直前沿研究的“新一代实验室”；企业在决定自建还是租用模型时，需综合评估成本敏感度、延迟要求、专属调优性能潜力和私有数据机密性四项维度。 (5:32) 自建主权 AI 应从零搭建小而精的研究团队，并依循“评测套件、外壳与路由、后训练、在线数据反馈闭环”的路线图推进，借助开源可控权重打造超越闭源前沿 API 的垂直性能壁垒。 (12:33)

Reading summary

Sonya Huang of Sequoia Capital argues that AI application companies must transition from renting closed foundation models to building 'sovereign AI' by owning their model weights on core features. While closed APIs remain practical for generalized tasks, Huang asserts that controlling the underlying intelligence layer is essential to manage inference costs, optimize latency, and prevent long-term platform dependency. Economic and Architectural Drivers for Weight Ownership (2:54) Huang identifies four primary drivers compelling companies to build sovereign AI rather than relying entirely on rented APIs: cost, latency, task-specific performance, and controlling platform risk. In high-volume or margin-sensitive environments, inference costs for closed frontier models scale steeply alongside product adoption, turning sovereign models into a financial necessity. Furthermore, latency-sensitive applications like interactive code auto-completion demand small, distilled architectures capable of sub-100-millisecond execution times that large general-purpose models cannot deliver. Historically, open-source models lagged behind frontier APIs in raw capability, but Huang notes that domain-tuned open weights can now outperform closed generalist models on specialized tasks. Adapting the cryptocurrency maxim 'Not your keys, not your crypto,' Huang argues 'Not your weights, not your product,' asserting that as product functionality and intelligence converge, true product differentiation requires direct control over model weights. Framework for Owning Versus Renting Model Capabilities (7:05) AI sovereignty is not an all-or-nothing proposition; companies must determine which capabilities to own versus rent by assessing four variables: cost, speed, performance, and proprietary data sensitivity. Generalized workflows that require expansive reasoning but lack strict latency limits—such as complex coding agents—are still best served by closed frontier APIs like GPT-4 or Claude 3.5 Opus. Conversely, features requiring extreme responsiveness or involving high query volumes require sovereign infrastructure. In coding tools, tab-complete functions are developed in-house to satisfy sub-100-millisecond constraints and avoid runaway API expenses. Similarly, domains reliant on highly confidential assets, such as cybersecurity startups or life sciences organizations handling proprietary chemical and biological structures, build bespoke models to maintain data privacy and maximize domain detection accuracy. Infrastructure Bifurcation: Production and Development Stacks (13:56) Transitioning to sovereign AI requires companies to manage two interconnected operational layers: a real-time production inference pipeline and an offline development stack. While closed API implementations typically consist of simple prompts and standard evaluations, sovereign production systems require specialized model harnesses, dynamic routing logic, and advanced context retrieval layers using tools like Turbopuffer vector databases, Glean knowledge graphs, and the Model Context Protocol (MCP). The development stack represents the primary strategic asset for continuous model improvement. Teams must implement rigorous evaluation frameworks, monitor for production drift, and generate fine-tuning inputs via expert trajectories, synthetic data, and reinforcement learning environments. Huang points to contemporary open-weight architectures, such as Kimi k3 and GLM-52, as foundations that allow teams to begin near frontier-level quality and subsequently surpass closed commercial APIs by closing feedback loops with live production data. Building Dedicated Research Teams and Establishing Legibility (9:51) Huang highlights that leading vertical applications—including Harvey, Glean, and Ramp—now operate as applied 'neo-labs,' conducting frontier-style research in domain fine-tuning and evaluation harnesses. To execute this work effectively, Huang cautions against delegating sovereign AI initiatives to existing internal platform teams, which are conventionally structured as service hubs rather than offensive research units. Instead, she recommends staffing small, dedicated research groups, noting that legal AI company Harvey achieved published research breakthroughs with a team of seven people. In addition to technical execution, Huang emphasizes 'legibility'—the practice of openly communicating technical depth to the market. Because enterprise buyers are actively committing to long-term vendors amid an influx of undifferentiated AI claims, application companies must validate their technical authority by branding dedicated research entities, releasing formal benchmarks, and publishing technical research. Huang envisions a decentralized future where individual companies build proprietary intelligence tailored to their specific workflows and proprietary data, rather than relying entirely on a centralized set of closed frontier models. However, realizing this sovereignty requires engineering organizations to accept the added complexity of custom post-training pipelines, harness orchestration, and ongoing evaluation systems.

精炼阅读摘要

红杉资本（Sequoia Capital）合伙人 Sonya Huang 指出，随着人工智能与产品体验深度融合，企业单纯依赖闭源模型 API 面临成本激增、延迟受限与平台绑定等风险；企业应当战略性地掌握模型权重并发展“自主智能”（Sovereign AI），以在特定垂直场景建立超越通用前沿模型的专有壁垒。 自主智能与去中心化范式 (1:24) Huang 将“自主智能”定义为企业在不依赖外部黑盒的前提下，完全掌握包括底层模型权重在内的技术能力。她明确澄清，这并不意味着企业应彻底放弃 Claude 3.5 Opus 或 GPT-4 等闭源前沿模型，闭源 API 在日常桌面办公和通用代码智能体等领域仍具显著优势；然而，当 AI 成为产品的核心体验时，“产品即智能”，企业必须掌控其核心层。借用加密领域的概念，Huang 提出了 AI 领域的竞争原则：“没有模型权重，就没有产品所有权”（Not your weights, not your product）。 在宏观格局层面，Palantir 的 Alex Karp、微软的 Satya Nadella 以及英伟达的黄仁勋（Jensen Huang）等业界高管近期均公开表态支持企业智能主权。Huang 对比了两种未来图景：一种是由单一黑盒模型统揽全球数据并垄断经济价值的“集中式智能”；另一种则是全球基于开放核心、由各企业结合专有数据、业务流与专业审美构建定制化能力的“去中心化智能”。Huang 认为，去中心化路径更利于保障商业实体的独立性与技术多样性。 驱动自研的四大因素与决策框架 (2:54) 基于红杉资本对数十家自建 AI 企业的观察，Huang 总结了促使企业从“租用”转向“拥有”模型权重的四大核心驱动力：成本、速度/延迟、垂直性能以及掌控自身命运。对于低毛利或零毛利业务，随着用户规模扩大，推理算力相关的销售成本（COGS）会急剧上升，迫使领先团队率先自研以压降单位成本；在代码实时补全或网络安全威胁检测等场景，小型蒸馏专有模型的毫秒级响应能力显著优于大型通用模型；此外，基于行业专有数据微调的开源权重，在特定任务上的表现已经开始反超闭源 API。 Huang 强调，构建自主智能并非非黑即白的抉择，企业应依据四大变量进行战略划分：成本比重、延迟敏感度（是否为 P0 级需求）、垂类性能表现以及专有数据的机密性。例如，复杂的编程智能体因侧重复杂推理且对延迟不敏感，目前主要采用租用闭源 API 的方案；而 Tab 键自动代码补全为了达成 100 毫秒以内的极速响应并控制调用成本，则普遍采用自研专有模型。在生物医药领域，高度机密的分子与化学数据也促使企业必须将模型权重与数据留存在自身体系内。 组织重塑：应用型公司向新实验室转型 (5:32) Huang 认为，当前最具活力的 AI 前沿研究不仅发生在传统通用人工智能（AGI）机构中，更涌现于 Harvey（法律）、Factory（软件工程）、Glean（企业搜索）、Open Evidence（医疗）、Semgrep（代码安全）与 Ramp（财务自动化）等垂直应用公司内。这些团队在评测基准构建、测试框架工程以及创新微调算法等领域展现出极高的科研水准，正转变为“新型实验室”（neo labs）。 在团队组建方面，Huang 告诫管理层不要将构建自主智能的任务强加给传统的内部 AI 平台团队。现有的平台团队通常采用“中心辐射”架构来被动响应各业务线的日常需求，而自主智能开发则需要主动出击的前沿研发特质。她建议从零开始搭建小而精的专属团队，并举例说明 Harvey 仅凭一支 7 人的精干研究团队便完成了大量底层科研突破与评测基准发布。同时，企业必须重视技术“能见度”（Legibility），通过成立专属品牌实验室、发布高质量研究论文和工程博客，向审慎的企业采购方证明自身的技术深度。 双层技术栈与在线反馈闭环 (12:33) 自主智能的基础设施被拆分为两个并行体系：负责实时生成 Token 的用户端“生产栈”（Production Stack），以及用于线下训练、评测与精炼模型的“开发栈”（Development Stack）。租用闭源 API 的方案配置简易且拥有较高底线，但由于无法利用真实生产交互数据反哺核心权重，其性能上限被锁定；自研技术栈虽然引入了底座选择、后训练与上下文优化的复杂性，却能释放出无上限的迭代潜力。 技术落地包含结构化流程：从明确自研范围、构建严格的垂类评测套件（Evals）、开发模型路由器与提示词 Harness，逐步推进到后训练（Post-Training）乃至中训练。Huang 指出，借助 Kimi k3、GLM-52 等开源权重底座，团队可以在接近前沿质量的起点上，叠加专有数据后训练与工程 Harness，实现超越前沿闭源 API 的表现。结合 Turbopuffer 向量数据库、Glean 知识图谱、MCP 协议以及 N-Gram 将上下文直接编码至权重的记忆技术，企业通过专家轨迹、合成数据与强化学习（RL）环境，能够建立起让模型随着用户交互持续在线学习的自增强闭环。 Huang 总结指出，掌控底层智能已成为应用层企业构筑长期防御工事的关键路径；但这一路径伴随着系统工程复杂度的激增，企业不仅要重构离线评测与数据反馈流水线，还必须在闭源 API 的即时便利性与自研权重的定制化上限之间动态评估资源分配。

Detailed notes — zh

## Background 该视频由红杉资本（Sequoia Capital）于 2026 年 8 月 18 日发布在 YouTube 频道。视频内容为红杉资本合伙人 Sonya Huang 在“掌控你的智能”（Own Your Intelligence）闭门活动上的开场演讲，系统阐述了为何越来越多的企业正选择从顶层应用垂直贯通至底层模型权重、自建主权 AI 的核心动因与落地框架。 ## TL;DR Sonya Huang 指出，随着开源权重模型逼近前沿水平，应用层的竞争正演变为对智能层本身的争夺，“没有权重就没有产品”已成为新时代的技术格言。企业决定自建模型主要受成本失控、实时速度延迟、垂直领域性能超越以及掌控自身命运四大动力驱动。为了实现自主智能，企业需遵循四步落地框架：严密权衡自研与租用的业务边界、组建独立的攻坚型实验室而非硬塞给传统平台团队、打造高辨识度的技术叙事，以及从扎实的评测（Evals）和 Harness 工程起步构建生产与开发双栈。通过将开源基座与深度后训练、专有数据反馈闭环相结合，垂直企业完全能够在特定场景反超通用闭源大模型。 ## Atmosphere & style 整场演讲呈现出一种兼具战略动员与硬核实操的鲜明氛围，节奏紧凑、观点犀利且极具感染力。Sonya Huang 表现得从容干练，既有宏观行业洞察的远见，又毫无投资人的高高在上，而是以创业者同盟的姿态倾囊相授具体的战术路线。现场传递出一种处于技术历史拐点时的紧迫感与兴奋感，既打破了“必须完全依赖闭源黑盒”的迷思，又坦诚指出了自研过程中工程复杂度与脏活累活的真实挑战，给在场的技术领袖和创始人带来了切实的战术底气。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) 红杉资本指出主权 AI（Sovereign AI）正成为行业共识，应用层竞争已演变为对底层智能层与模型权重所有权的争夺。 - [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 驱动企业自主掌控模型的四大核心推力为成本、推理速度、垂直领域性能超越以及掌控自身商业命运。 - [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) “没有权重，就没有产品”：顶尖应用层公司正在转型为新型实验室（Neo Labs），以开源权重基座结合后训练实现超越通用闭源 API 的表现。 - [06:25](https://www.youtube.com/watch?v=bMMv0bZzONg&t=385s) 构建内部智能体系需遵循四步落地路径：划定自研与租用边界、组建独立的攻坚型研究团队、对外建立高品位技术叙事，以及以扎实评测为基石搭建生产与开发技术栈。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) 主权 AI 意味着企业从架构直至模型权重垂直贯通并掌控核心智能，但并不排斥在代码生成与通用任务中继续使用顶尖闭源 API。 - [01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) AI 发展面临集中式智能与分布式智能两条路径的抉择：前者是全知全能的超级黑盒，后者则允许多样化的专属模型根据私域数据与偏好生长。 - [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 高频调用业务的推断成本（COGS）容易呈指数级失控，促使激进的应用落地拓荒者率先走向自主控权模型。 - [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 在代码补全等极度敏感的高频交互场景中，紧凑蒸馏的高性能小模型能以毫秒级响应速度击败庞大的通用闭源模型。 - [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 行业已迎来自建模型的性能拐点，基于企业专有数据微调的开源权重模型在垂直任务表现上已能超越前沿闭源基座。 - [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) 借鉴加密货币理念，“没有权重，就没有产品”（Not your weights, not your product）成为企业避免单点依赖、掌握自身命运的核心法则。 - [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) 应用层与模型层的界限正在打破，Harvey、Factory、Glean 等领头羊企业已蜕变为深入微调算法与 Harness 工程的新型应用实验室（Neo Labs）。 - [06:25](https://www.youtube.com/watch?v=bMMv0bZzONg&t=385s) 企业在自研与租用之间并非非黑即白，应依据成本影响、时延要求、垂直性能增益及私有数据敏感度四个维度进行动态权衡。 - [06:25](https://www.youtube.com/watch?v=bMMv0bZzONg&t=385s) 软件工程体系呈现出明确的自研与租用分工：高频低延时的代码补全已普遍自建轻量模型，而复杂的多步推理 Agent 仍倾向租用闭源 API。 - [09:34](https://www.youtube.com/watch?v=bMMv0bZzONg&t=574s) 搭建智能团队切忌硬塞进传统 AI 平台团队，主权模型研发需要白手起家、具备主动进攻思维的精悍攻坚小团队。 - [10:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=654s) 研究成果的可辨识度（Legibility）是企业级销售的核心胜负手，企业应成立专属 Labs 品牌并发布高品位技术报告以建立客户信任。 - [12:03](https://www.youtube.com/watch?v=bMMv0bZzONg&t=723s) 技术演进路线图的核心基石是严密繁琐的评测体系（Evals），后续通过模型路由、Harness、后训练以及生产数据闭环实现持续进化。 - [13:52](https://www.youtube.com/watch?v=bMMv0bZzONg&t=832s) 自建智能体系将系统拆解为面向用户的“生产栈”与打磨能力的“开发栈”，告别单一 API 模式，主动拥抱后训练、上下文工程与在线学习的复杂度。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) 主权 AI 的定义与大会初衷：拥抱自主智能而非完全摒弃闭源 红杉资本（Sequoia Capital）合伙人 Sonya Huang 向在场的约 80 位投资组合公司创始人及 AI 领袖致以热烈欢迎。她指出，这次闭门集会旨在探讨当下极具时效性的话题——“掌控属于你自己的智能”（Owning Your Intelligence），亦即“主权 AI”（Sovereign AI）。今天的议程设计兼具号召性动员与实操指南的性质，充斥着硬核的技术分享与现场演示，目的不仅在于空谈自建 AI 的理念，更在于共同探索从理论迈向工程落地的具体路径。 究竟何为主权 AI？在传统政治学语境中，“主权”指拥有完全自主治理权力的独立实体；而在商业与技术维度，“主权 AI”意味着企业能够完全掌控自身的核心智能，消除外部技术依赖，甚至垂直贯通并掌控底层的模型权重（weights）。Sonya 随即澄清了一个极其关键的认知误区：倡导主权 AI 绝不代表企业必须立刻抛弃 Claude Opus 或 GPT 系列模型。对于代码生成 Agent、日常桌面办公助理以及前沿能力的通用 API，闭源模型依然具有无可替代的卓越表现。然而，红杉在被投企业中观察到一个鲜明的趋势：越来越多的企业开始在产品的特定业务模块中自研 AI 能力，逐步走向垂直整合，将智能资产的所有权牢牢握在手中。 - **会议定位与参会规模**：汇聚约 80 位红杉被投企业的创始人与顶尖 AI 研发负责人，定位于“动员大会+技术实操教学”。 - **“主权 AI”的核心定义**：企业无须依赖外部黑盒，能够从架构直至模型权重层面自主掌控智能资产。 - **与闭源生态的关系**：并非要求企业完全与 GPT/Opus 切割；前沿闭源 API 在高通用推理任务中依然扮演重要角色，但垂直领域的自主化已是大势所趋。 --- ### [01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) 集中式智能 vs. 分布式智能：时代浪潮下的战略抉择 主权 AI 概念的兴起绝非孤立现象。在过去一个月里，整个科技界对“拥有自主智能”的讨论急剧升温。Palantir 首席执行官 Alex Karp 与微软首席执行官 Satya Nadella 等行业巨头纷纷公开发声，力挺企业掌控自身智能；英伟达首席执行官 Jensen Huang（黄仁勋）更在关键政策节点带头行动，极力确保开源权重模型在美国市场的可获取性，并获得了行业近乎一致的强烈声援。这一连串事件传递的信号极为纯粹：企业迫切希望“拥有”（Own）而非仅仅“租用”（Rent）模型权重，因为智能本身是商业构架中最底层、最根本的核心属性，绝不能轻率地完全外包出去。 Sonya 指出，主权 AI 成为时代精神（Zeitgeist）不仅是一种商业选择，更是关乎全球技术格局走向的关键节点。如果世界走向“集中式智能”（Centralized Intelligence），未来将由极少数全知全能的超级 AI 统治全球 GDP 的底层运转，它们如同吞噬一切的黑盒，不断抽吸来自全行业产生的数据废气与数据飞轮；反之，“分布式智能”（Decentralized Intelligence）则描绘了一幅更为乐观多元的图景：全世界在稳固坚实的基础能力之上，由每个个体、每家公司依据自身的专有数据、行业知识、个性化偏好以及独特品味，量身定制专属的智能体。在这一生态下，多样性战胜了同质化，没有任何单一寡头能将全行业吞并殆尽。 --- ### [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 驱动企业自主构建模型的四大核心动力 在陪伴数十家企业探索模型自研之路的过程中，Sonya 总结出了推动企业走向主权 AI 的四大底层驱动力： ```mermaid flowchart TD accTitle: 驱动自主构建模型的四大动力 accDescr: 展示推动企业自建 AI 模型的四大要素：成本、推理延迟与速度、垂直领域性能超越、掌控自身命运。 subgraph CostEfficiency["1. 成本考量 (Cost)"] C1["边际利润薄弱业务"] --> C2["高调用量导致 COGS 激增"] C2 --> C3["自建小模型大幅削减推断支出"] end subgraph LatencySpeed["2. 速度与延迟 (Speed)"] S1["高交互场景 (代码补全/安全)"] --> S2["通用超大模型延迟难以忍受"] S2 --> S3["紧凑蒸馏模型实现极速响应"] end subgraph DomainPerf["3. 领域性能 (Performance)"] P1["专有数据与业务流对齐"] --> P2["开源权重微调超越闭源"] P2 --> P3["特定场景构建精度护城河"] end subgraph Sovereignty["4. 掌控自身命运 (Destiny)"] D1["闭源厂商存在单点依赖风险"] --> D2["业务演进自主不受外部限制"] D3["'Not your weights, not your product'"] end ``` 第一大考量是**成本（Cost）**。尤其是对于本身毛利率较低甚至为负的业务，主权 AI 已经不再是锦上添花的加分项，而是关乎生死存亡的生存刚需。行业内存在一个颇具讽刺意味的现象：你的 AI 产品越成功、用户量与调用量越大，吞噬利润的推断成本（COGS）就越是呈指数级失控。因此，那些在 AI 应用落地方面最激进的前沿企业，往往也是最先被迫走向自主控权模型道路的拓荒者。 第二大考量是**速度（Speed）**。在代码编写、网络安全防御等高度敏感且需要即时交互的领域，通用超大模型的响应延迟是致命的短板，一个经过紧凑蒸馏（distilled）的高性能定制化小模型，往往能凭借毫秒级的极速体验彻底击溃庞大的通用模型。 第三大考量是**性能（Performance）**。在上一年度，几乎没有企业会单纯为了追求更高的性能表现而去自研或调整开源模型，普遍认知是闭源 API 的性能无可匹敌；然而行业已经迎来了拐点——在深度绑定的细分领域中，基于企业自身数据微调的开源权重模型，其任务表现已经能够超越通用闭源模型。 第四大考量是**掌控自身命运（Controlling Your Own Destiny）**。尽管 OpenAI 和 Anthropic 等模型实验室是非常出色的生态合作伙伴，但任何具有长远野心的企业都愈发清醒地认识到，自身业务不能永远寄生于他人的 API 之下，必须拥有独立自主的立足根基。 --- ### [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) “没有权重，就没有产品”：应用层与智能层的合流 为了具象化这种对所有权的反思，Sonya 幽默地唤起了加密货币领域的经典切口。她询问在座听众是否还记得 DeFi 玩家口耳相传的著名迷因——“没有私钥，就不是你的加密资产”（Not your keys, not your crypto）。如果资产由交易所或第三方代管，从本质上讲它就不属于你。Sonya 将这一逻辑提炼为 AI 时代的全新格言： > “没有权重，就不是你的产品（Not your weights, not your product）。” 要让一个技术产品真正属于你自己，企业必须能够对模型权重拥有实际的控制权与托管权。 紧接着，Sonya 提到了红杉合伙人 Pat Grady 此前在 AI Ascent 大会上展示的经典观点——“应用层争夺战”。原本的认知模型认为，基础模型实验室自顶向下扩展，力求将模型直接做成面向用户的产品，而应用层企业则自底向上，立足用户界面与业务场景向下渗透，双方在中间层展开碰撞。 然而，当前的行业演变正在彻底重塑这一战场：这场战争不再仅仅局限于谁能夺取产品界面、UI 交互包装或分销渠道，而是迅速蜕变为**对智能层（Intelligence Layer）本身的争夺**。在当下，“产品即智能，智能即产品”。最前沿的应用层公司并不甘于只做外包装，而是正在演变为新型应用实验室（Neo Labs）。以 Harvey（法律智能）、Factory（工程智能）、Glean（企业搜索与知识系统）、OpenEvidence（医疗决策）、Semgrep（代码安全）以及 Ramp（金融科技）为代表的领头羊企业，正在内部全面展开前沿的应用研究，涵盖了评测体系构建、模型框架工程（harness engineering）以及创新的微调算法等深水区。 --- ### [06:25](https://www.youtube.com/watch?v=bMMv0bZzONg&t=385s) 自建 AI 的四步落地框架与步骤一：权衡自主研发与租用 面对希望构建内部模型实验室、实现从 0 到 1 再到 100 演进的企业，Sonya 提供了一套极具实操导向且带有鲜明主观视角的路线图。她谦逊地强调自己并非技术研发背景，但希望通过梳理顶尖企业的实践，为创业者提供可落地的战术地图。整个演进框架共分为四大步骤： 1. **战略决断（Strategy）**：明确界定哪些能力必须自主掌握，哪些可以继续外购租用； 2. **团队搭建（Team）**：探索针对“智能生产”的全新组织架构与人才配置方式； 3. **研究透明化（Legibility）**：将模型研究过程结构化暴露给全公司，避免研发孤岛； 4. **技术路线图（Technical Roadmap）**：系统化拼装模型自研所必需的基础技术积木。 在**步骤一（权衡自主研发与租用）**中，Sonya 强调，主权 AI 并非非黑即白的二元对立，绝非只有 0% 或 100% 两个极端。企业的战略重心在于根据以下四个核心维度绘制边界： | 评估维度 | 决策考量要点 | 倾向自主研发（Own）的情境 | 倾向外部租用（Rent）的情境 | |:--- |:--- |:--- |:--- | | **成本影响 (Cost)** | 该项模型调用在整体销货成本（COGS）中的占比 | 高频次调用、规模化消耗，直接威胁产品毛利空间 | 偶发、低频调用，对整体成本几乎没有实质性影响 | | **速度与延迟 (Latency)** | 端到端推断时延是否属于核心业务指标（P0） | 实时交互场景，用户对几十毫秒的卡顿极度敏感 | 异步任务或复杂的后台 Agent 长链条推理流程 | | **性能收益 (Performance)** | 垂直领域内专用模型微调能否显著超越通用大模型 | 业务逻辑深度垂直，可通过特定数据与后训练大幅提效 | 依赖开箱即用的通识推理、跨领域常识判断能力 | | **专有数据 (Proprietary Data)** | 用于迭代模型的训练数据是否具有极高的商业敏感度 | 数据涉及核心商业机密、强监管合规或高壁垒专有资产 | 依赖公开互联网通识数据，缺乏独特的企业私域沉淀 | 为了直观呈现上述维度的权衡，Sonya 列举了几个典型场景：在软件工程（Coding）领域，开发体系分化为两组截然不同的路径： - **代码补全（Tab Autocomplete）**：对响应延迟有着近乎苛刻的极致要求，且开发者的击键触发频率极高，API 计费会迅速成为天文数字，因此各大代码助手公司已全面转向自主研发并部署极度轻量化的主权模型； - **编程 Agent（Coding Agents）**：目前绝大多数团队仍选择外部租用前沿闭源 API，原因在于 Agent 依赖极其强大的开箱即用复杂推理能力，且其执行多步任务时对短时延迟的容忍度相对较高。 此外，网络安全领域的隐形孵化公司（Sonya 在现场致意了参会的 Alon）为了追求极致响应速度并对模型进行高度定制化的后训练，毅然走上模型自研道路；生物科技企业（Bio）则因专有实验数据的极端商业机密性与唯一性，全员加速将智能资产收归自身主权掌控。 --- ### [09:23](https://www.youtube.com/watch?v=bMMv0bZzONg&t=563s) 步骤二：搭建智能团队的组织雏形 在厘清战略边界后，核心诉求立刻转向人才与组织阵型的搭建（Assemble a team）。如何寻找并组建能够驾驭“智能生产线”的研发队伍，是决定自建模型成败的前提。Sonya 展示了两位不同路径的实验室负责人画像作为代表案例，其中一位代表人物是 Nico，其职业背景深度植根于经典学术与工业界研究体系，曾先后在苹果（Apple）与谷歌（Google）的核心研究部门主导前沿算法研发，代表了从纯正研究基因切入企业内部智能体系构建的典型路径。 ### [09:34](https://www.youtube.com/watch?v=bMMv0bZzONg&t=574s) 组建团队：切忌将自主 AI 硬塞给传统平台团队 在探讨实验室负责人的选拔时，通往成功的路径绝非只有一条。以 Ramp 的 Alex 为例，他在加入 Ramp 之前曾在微软担任多个工程技术职位，更偏向深厚的工程背景；而另一些负责人则专注于基础理论研究。究竟选择哪种人选，取决于公司打算开展的研究偏向基础理论还是偏向落地应用。 更关键的痛点在于组织架构的设计。传统上，AI 团队通常采用“中心辐射”（hub-and-spoke）架构，由一个中心化的平台团队向各个应用产品团队提供支持。然而，许多公司正试图强行将这种平台团队塞进主权 AI（Sovereign AI）的研发任务中。Sonya 极力反对这种做法，呼吁企业从零开始组建全新团队。因为主权模型研发根本不是一项“平台服务式”的工作，它需要团队能够敏锐应变、身处技术前沿并产出前沿水准的研究成果。企业需要的是主动进攻的攻坚力量，而不是被动响应各业务线需求的客服型支持人员。事实证明，白手起家的精悍小团队往往能走得更远——例如法律 AI 独角兽 Harvey，仅凭一支 7 人的初创研究团队，就发布了极其丰硕的前沿成果。 - **组织陷阱**：避免让负责基础设施支持的传统 AI 平台团队兼顾底层模型研发，服务支持心态与前沿攻坚思维存在本质冲突。 - **团队规模**：精简的独立实验室完全能够创造奇迹，Harvey 仅用 7 名研究员就撑起了全套底层研究体系。 --- ### [10:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=654s) 打造可辨识度：用高品位技术叙事赢得客户心智 研发的可辨识度（Legibility）是一项被严重低估的战略能力。许多公司在内部默默完成了惊艳的研究，但外界却对此一无所知。正如 Harvey 联合创始人 Winston Weinberg 所言，CEO 的职责核心有两点：一是“拿出实质性的过硬结果”（drive substantive results），二是“掌控叙事与外在可辨识度”（control the narrative, control legibility around what you're building）。 在构建自有 AI 栈的过程中，可辨识度直接决定了商业成败。如今每一个企业级买家都在急切寻找属于自己的“AI 冠军”，但他们每天听到的销售话术千篇一律，根本无法分辨哪家供应商才具备真正带他们抵达应许之地的硬核技术实力。买家渴望选择那些真正懂底层机制的团队，而最强有力的证明就是公开发布自身的研究成果。保持高度的可辨识度意味着要做好顶尖的技术营销，例如成立独立品牌的 Labs 研发部门，并以高审美和高品位（high taste）对外发布技术报告。 - **买家痛点**：企业客户面临严重的供应商同质化疲劳，亟需通过底层研究成果来验证团队的技术成色。 - **行动倡议**：成立专属的 Labs 品牌，将技术营销提升至战略高度，向行业输出代表自身技术水准的白皮书与研究论文。 --- ### [12:03](https://www.youtube.com/watch?v=bMMv0bZzONg&t=723s) 技术路线图：从扎实评测到超越前沿闭源基准 构建自有智能的技术路径因公司而异，但大体遵循一套严谨的演进路线：首先确立战略；第二步是定义评测体系（Evals）。编写评测是一项极其枯燥、繁琐且毫无光环的工作，但前期在此投入越多，后续所有环节就越扎实，这是整个路线图中最至关重要的基石。 在此之后，团队会逐步探索模型路由器（Model Routers）和 Harness（运行线束与代理框架）。部分企业发现开箱即用的基座模型已能满足需求；另一些企业则通过后训练（Post-training）获得了大幅性能跃升，极少数场景甚至会切入中训练（Mid-training）或预训练。终极形态则是建立起闭环系统，利用生产环境中实时产生的真实客户数据反哺模型，让智能水平随着每一次用户交互持续进化。得益于如 Kimi k3、GLM 52 等新一代顶尖开源权重模型的涌现，自主可控不再意味着性能妥协——这些开源权重具有远超闭源 API 的可塑性，企业以接近前沿的基座为起点，结合后训练、Prompt Harness 工程与在线学习，完全能够在特定领域实现超越闭源前沿模型的卓越表现。 ```mermaid flowchart LR accTitle: 技术演进路线图 accDescr: 展示从设定战略、构建评测到反馈飞轮的技术落地全流程。 subgraph Phase1["基础奠基"] A["明确业务战略"] --> B["构建评测体系 (Evals)"] end subgraph Phase2["架构与训练"] B --> C["模型路由与线束 (Harness)"] C --> D["后训练定制 (Post-training)"] end subgraph Phase3["数据飞轮"] D --> E["生产数据反馈与在线学习"] end ``` - **开源红利**：以 Kimi k3、GLM 52 为代表的开源权重模型打破了闭源壁垒，为企业提供了无限接近前沿基准的底座。 - **超越闭源**：通过“开源基座 + 深度后训练 + 针对性线束工程 + 生产反馈闭环”，垂直模型在特定场景的表现能够反超通用顶尖 API。 --- ### [13:52](https://www.youtube.com/watch?v=bMMv0bZzONg&t=832s) 架构剖析：生产栈、开发栈与被打开的“潘多拉魔盒” 从基础设施的角度审视自有智能，技术栈可以清晰划分为两大平行板块：左侧是面向用户的“生产栈”（Production Stack），它驱动着呈现给用户的每一个 Token，其本质是“构建在模型之上的 Harness 与上下文结构”；右侧则是“开发栈”（Development Stack），包含所有用于打磨模型能力的工具与供应商。 在纯闭源生态中，这套架构极其单薄：底层调用 Claude Opus 或 GPT 等闭源模型，顶层搭载开箱即用的 Harness，虽能通过 Prompt 或 RAG 注入上下文，但企业无法掌控底层数据，更无法用生产数据精调权重。这种模式“下限很高，但上限很低”。而一旦决定掌控核心智能，就如同打开了“潘多拉魔盒”：简洁优雅的单一 API 调用不复存在，企业必须在开源基座上亲手操刀复杂的后训练；在生产端必须自主配置 Harness 逻辑、工具调用与上下文体系；在开发端更要严密监控评估指标与线上性能漂移，并采集海量高质量数据（专家操作轨迹、合成数据、强化学习环境）以支持在线学习。 ```mermaid flowchart TD accTitle: 生产栈与开发栈架构图 accDescr: 对比面向用户的生产推理栈与支撑模型迭代的开发训练栈。 subgraph Production["生产栈 (Production Stack)"] direction TB P1["应用交互层 / Token 生成"] P2["Agent 线束 (Harness) & 业务逻辑"] P3["多维上下文 (向量库 / 知识图谱 / 权重记忆)"] P4["自选开源基座模型 (Open Base Models)"] P1 --> P2 --> P3 --> P4 end subgraph Development["开发栈 (Development Stack)"] direction TB D1["严密指标监控与评测 (Evals & Drift)"] D2["高质量数据源 (专家轨迹 / 合成数据 / RL)"] D3["在线学习系统 (持续数据飞轮)"] D1 --> D2 --> D3 end Development -. "持续后训练与调优".-> Production ``` - **多维上下文注入**：性能的突破高度依赖上下文供给，包括 Turbopuffer 等向量数据库、Glean 等企业知识图谱、标准的 MCP 开源连接器，甚至是 N-Gram 创始人 Dan Biederman 所探索的“将记忆与上下文直接编码进权重”的创新架构。 - **闭源 vs 自主**：闭源调用是省心的黑盒，而自主可控虽然打开了工程复杂度的魔盒，却真正赋予了企业掌控上限、利用数据飞轮拉开差距的自由度。 --- ### [16:35](https://www.youtube.com/watch?v=bMMv0bZzONg&t=995s) 全明星工作坊矩阵：全方位落地自主 AI 为了帮助在场的创业者和技术领袖系统性掌握自建 AI 所需的全部能力，大会绕开了极其昂贵且通用性较低的基座预训练，针对自建技术栈的关键节点精心设计了全套实战工作坊： - **后训练实战**：由 Fireworks 的 Lynn 主持，深入解析开源基座的模型微调与特定任务对齐。 - **Harness 与评测**：由 LangChain 的 Harrison 领衔，剖析 Agent 逻辑骨架、上下文管理与评测工具链的搭建。 - **强化学习与合成数据**：由 Mercor 的 Brendan 主讲，详解 RL 环境搭建、专家轨迹收集与合成数据生成技术。 - **在线学习机制**：由 Trajectory 团队指导，分享如何打通生产交互数据并让模型持续自主进化。 - **端到端全栈复盘**：最后由刚刚正式宣布成立 Harvey Research 的 Harvey 团队压轴登场，全面公开他们如何从零梳理整套 AI 战略，并将各层合作伙伴技术融会贯通为行业标杆。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) 掌控属于你自己的智能 — Sonya Huang > [00:00](https://www.youtube.com/watch?v=bMMv0bZzONg&t=0s) 倡导主权 AI 绝不代表企业必须立刻抛弃 Claude Opus 或 GPT 系列模型 — Sonya Huang > [01:24](https://www.youtube.com/watch?v=bMMv0bZzONg&t=84s) 企业迫切希望“拥有”（Own）而非仅仅“租用”（Rent）模型权重，因为智能本身是商业构架中最底层、最根本的核心属性，绝不能轻率地完全外包出去 — Sonya Huang > [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 你的 AI 产品越成功、用户量与调用量越大，吞噬利润的推断成本（COGS）就越是呈指数级失控 — Sonya Huang > [02:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=174s) 在深度绑定的细分领域中，基于企业自身数据微调的开源权重模型，其任务表现已经能够超越通用闭源模型 — Sonya Huang > [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) 没有权重，就不是你的产品（Not your weights, not your product） — Sonya Huang > [04:11](https://www.youtube.com/watch?v=bMMv0bZzONg&t=251s) 产品即智能，智能即产品 — Sonya Huang > [06:25](https://www.youtube.com/watch?v=bMMv0bZzONg&t=385s) 主权 AI 并非非黑即白的二元对立，绝非只有 0% 或 100% 两个极端 — Sonya Huang > [10:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=654s) 拿出实质性的过硬结果 — Winston Weinberg > [10:54](https://www.youtube.com/watch?v=bMMv0bZzONg&t=654s) 掌控叙事与外在可辨识度 — Winston Weinberg > [12:03](https://www.youtube.com/watch?v=bMMv0bZzONg&t=723s) 编写评测是一项极其枯燥、繁琐且毫无光环的工作，但前期在此投入越多，后续所有环节就越扎实，这是整个路线图中最至关重要的基石 — Sonya Huang > [13:52](https://www.youtube.com/watch?v=bMMv0bZzONg&t=832s) 下限很高，但上限很低 — Sonya Huang ## Mentioned Sonya Huang, 红杉资本 (Sequoia Capital), Claude Opus, GPT, Alex Karp, Palantir, Satya Nadella, 微软 (Microsoft), Jensen Huang (黄仁勋), 英伟达 (NVIDIA), OpenAI, Anthropic, Pat Grady, AI Ascent, Harvey, Factory, Glean, OpenEvidence, Semgrep, Ramp, Nico, 苹果 (Apple), 谷歌 (Google), Alex, Winston Weinberg, Kimi k3, GLM 52, Turbopuffer, Dan Biederman, N-Gram, Lynn, Fireworks, Harrison, LangChain, Brendan, Mercor, Trajectory, Harvey Research ## Video description Sequoia Capital partner Sonya Huang opens our Own Your Intelligence event with the case for why more companies are choosing to own their intelligence, down to the weights. She lays out the four forces driving the shift: cost, speed, performance, and controlling your own destiny, and explains why the race for the application layer is becoming the race for the intelligence layer itself. Sonya shares an opinionated framework for deciding what to own vs. rent when assembling your AI stack, and another for how to get going from zero to one. With today's open-weight models near the frontier, she argues, owning your stack is no longer a performance sacrifice. It may be your performance edge. 00:00 What is sovereign AI (and what it isn't) 01:24 Centralized vs. decentralized intelligence 02:54 Four reasons companies own their models: cost, speed, performance, destiny 04:22 "Not your weights, not your product" 05:32 The application companies are the newest neo labs 07:05 Step 1: Deciding what to own vs. rent 09:51 Step 2: Build the team (and don't shoehorn your platform team) 11:17 Step 3: Legibility – why your research has to be visible 12:33 Step 4: The technical roadmap 13:56 The stack: production vs. development 15:16 Opening Pandora's box – base models, harnesses, context [Source](https://www.youtube.com/watch?v=bMMv0bZzONg)

Source quotes

[0:00] Sonya Huang: 掌控属于你自己的智能 [0:00] Sonya Huang: 倡导主权 AI 绝不代表企业必须立刻抛弃 Claude Opus 或 GPT 系列模型 [1:24] Sonya Huang: 企业迫切希望“拥有”（Own）而非仅仅“租用”（Rent）模型权重，因为智能本身是商业构架中最底层、最根本的核心属性，绝不能轻率地完全外包出去 [2:54] Sonya Huang: 你的 AI 产品越成功、用户量与调用量越大，吞噬利润的推断成本（COGS）就越是呈指数级失控 [2:54] Sonya Huang: 在深度绑定的细分领域中，基于企业自身数据微调的开源权重模型，其任务表现已经能够超越通用闭源模型 [4:11] Sonya Huang: 没有权重，就不是你的产品（Not your weights, not your product） [4:11] Sonya Huang: 产品即智能，智能即产品 [6:25] Sonya Huang: 主权 AI 并非非黑即白的二元对立，绝非只有 0% 或 100% 两个极端 [10:54] Winston Weinberg: 拿出实质性的过硬结果 [10:54] Winston Weinberg: 掌控叙事与外在可辨识度 [12:03] Sonya Huang: 编写评测是一项极其枯燥、繁琐且毫无光环的工作，但前期在此投入越多，后续所有环节就越扎实，这是整个路线图中最至关重要的基石 [13:52] Sonya Huang: 下限很高，但上限很低

Source quotes

[4:22] Sonya Huang: Not your weights, not your product [2:54] Sonya Huang: Sovereign AI isn't a nice-to-have, it's a must-have... Ironically, the more successful your AI product is, the higher your AI COGS tend to be. [12:33] Sonya Huang: You start with a baseline that's already close to frontier, and then with a good enough technical roadmap... you can actually reach better than frontier performance by owning your stack. This is new for 2026.

Original transcript

[0:00] Okay, good morning everyone. Thank you [0:02] all so much for being here. We have [0:04] about 80 portfolio company founders and [0:06] AI leaders in the room to explore a very [0:09] timely topic, owning your intelligence [0:12] or sovereign AI. Uh today's event is [0:14] meant to be half a rallying call and [0:17] half technical how-to. And so we have [0:19] stacked the agenda with, I think, high [0:21] high substance technical talks and demos [0:24] so that we don't just talk the talk of [0:25] building your own AI, but actually learn [0:27] how to get there together. Uh so thank [0:29] you all for taking the time out of your [0:30] mornings to join us. I know your time is [0:32] incredibly precious, and let's dive [0:35] right in. [0:37] So let's start with the opportunity. [0:39] What is sovereign AI? Uh a sovereign is [0:42] an independent state that has total [0:43] self-governance. And sov- sovereign AI [0:46] refers to companies owning their own [0:48] intelligence without external [0:50] dependencies down to the weights. [0:52] An important nuance here is we are [0:54] definitely not telling our companies to [0:55] get off Opus or GPT. That is definitely [0:58] not the message. [0:59] Uh for coding agents, for desktop work, [1:02] uh for frontier level APIs, [1:05] the closed model APIs are wonderful. [1:07] But what we've observed is that more and [1:09] more of our companies are going down the [1:10] path of wanting to build their own AI [1:13] capabilities in parts of their products [1:15] and vertically integrating towards [1:16] owning more of this intelligence. And so [1:18] today's session is meant to equip [1:20] companies that are starting to go down [1:22] that journey. [1:24] We're obviously not alone in this idea. [1:26] In the last month, [1:28] uh the rhetoric around sovereign AI has [1:30] escalated sharply with folks like Alex [1:33] Karp and Satya speaking up in support of [1:35] companies owning their own intelligence. [1:37] Just last week, Jensen led the charge in [1:40] making sure that open weight models [1:41] remain available in the US. And it was [1:44] awesome to see the near unanimous wave [1:46] of support. I think the message is very [1:48] clear that companies want to own their [1:51] intelligence. They want to own, not [1:53] rent, their weights. And I think it's [1:54] simply because intelligence is too core, [1:57] too fundamental of a property to just [1:59] outsource. [2:02] We're glad that sovereign AI is in the [2:03] zeitgeist right now uh because we think [2:05] it's a good thing for the world. On one [2:07] hand, you have centralized intelligence [2:09] where a single all-powerful AI powers [2:12] more and more of the GD the world's GDP [2:14] as a black box sucking in all the data [2:16] exhaust, all the data flywheels from the [2:18] rest of the world. [2:20] On the other hand, [2:21] you have decentralized intelligence [2:23] where the whole world builds on top of a [2:25] solid core, but every individual person, [2:28] company builds their own intelligence on [2:30] top bespoke to their own data, their own [2:33] industries, their own personalization, [2:36] uh their own way of working, their own [2:37] taste. [2:38] And so the ecosystem flourishes and [2:40] individuality triumphs. [2:42] No single company swallows the rest. I [2:44] think this is a much more optimistic [2:46] view of the world. [2:49] And so we work with dozens of companies [2:51] that are going down the journey of [2:52] building their own AI. [2:54] Um here are the biggest reasons that [2:55] we've seen people move. [2:57] Reason number one is cost. [3:00] Uh especially for low, zero, negative [3:02] margin companies, sovereign AI isn't a [3:05] nice-to-have, it's a must-have. [3:07] Um ironically, the more successful your [3:09] AI product is, the higher your your AI [3:11] cogs tend to be. And so it's actually [3:14] the companies that have been most [3:15] advanced in the deployment of AI that [3:18] have been the first to go on this [3:19] journey of of owning their own models. [3:22] Reason number two is speed. And so in [3:25] certain domains, coding is one of them, [3:28] uh security is another one, a small [3:30] distilled custom model can beat a large [3:33] general one because speed is so [3:34] important. [3:36] Num- Reason number three [3:38] is performance. Um this is a relatively [3:40] newer one. I would say last year, most [3:42] companies were not choosing to own their [3:44] intelligence to generate better [3:45] performance. [3:46] Uh but we're now at the point where open [3:48] models can outperform closed ones on [3:50] your domain. [3:51] And we're going to spend a lot of [3:52] today's agenda talking about how to get [3:54] there. And then reason number four, [3:56] controlling your own destiny. Uh [3:58] Anthropic and OpenAI, I actually think [4:00] to their credit, they've been really [4:01] wonderful partners to a lot of the [4:03] ecosystem. But companies are [4:05] increasingly finding that they want [4:07] their own set of independent legs to [4:08] stand on as well. [4:11] Um does anybody here come from the [4:13] crypto days or remember the crypto days? [4:15] Okay. [4:16] Um do you guys remember this meme? [4:18] Okay. [4:19] Uh in the crypto of the days there was [4:21] this meme for the DeFi degens. Uh not [4:23] your keys, not your crypto. And so if [4:25] somebody was custodying your crypto for [4:27] you, it fundamentally wasn't yours. [4:30] I hereby present the AI version of this [4:32] meme. Not your weights, not your [4:34] product. [4:35] Um I think that for product to be truly [4:37] yours, I think it's reasonable to think [4:39] that you need to be able to control [4:40] control and custody your own weights. [4:45] Pat shows this slide at AI Ascent [4:47] talking about the race for the [4:48] application layer. Uh the He looks so [4:51] proud of yourself. Uh the punchline is [4:53] that both the AGI labs and the [4:54] application companies are racing to be [4:57] the user-facing products from different [4:59] ends. The foundation model labs from the [5:01] model side and the application companies [5:03] from the user back. [5:05] I think we're seeing a new dynamic [5:07] emerge now, which is actually that [5:09] battleground is increasingly becoming [5:11] not just the race for the application [5:13] layer, but the race for the intelligence [5:15] layer. [5:16] And so this this battleground is no [5:17] longer just about who gets to control [5:19] the product, the UI, the go-to-markets, [5:21] the wrapping. It's actually about who [5:23] can own the intelligence itself [5:26] and shape better intelligence in the [5:28] product. So the product is the [5:29] intelligence, and the newest [5:30] battleground is for not just the product [5:33] surface, but for the intelligence layer [5:35] itself. [5:37] And so the hottest new labs, in my [5:38] opinion, are actually uh the applied [5:41] research that we see coming out of [5:43] companies right now like Harvey, [5:45] like Factory, [5:48] Glean, [5:51] Open evidence, [5:53] Semgrep, [5:56] Ramp. [5:57] The list goes on and on, and I think the [5:59] research is spanning everything from [6:00] evals and benchmarks to harness [6:02] engineering to new algorithmic [6:04] techniques for fine-tuning and a lot, [6:06] lot more. [6:08] And we started this uh we started this [6:09] morning talking about centralized versus [6:12] decentralized intelligence. I think it's [6:14] really wonderful to see the amount of [6:15] innovation that is happening in these [6:18] democratized intelligence world. Like I [6:20] actually think the application companies [6:22] are the newest Neo labs. [6:25] Okay, so I assume that everyone here [6:27] today is pretty bought into this [6:28] journey. Let's assume that you want to [6:30] build your own lab, build your own [6:31] models. How do you go from zero to one [6:33] to 100? [6:34] We're going to do something a little bit [6:35] different today here. Um I'm going to [6:37] lay out an opinionated framework and [6:39] technical roadmap. And so take that with [6:41] a giant grain of salt. Uh every company [6:43] is different. [6:45] And and I'm not technical. And so take [6:46] this with a giant, giant grain of salt, [6:48] but I hope it provides a useful starting [6:50] point for how to think about building [6:52] your own intelligence. [6:55] Uh step one um to owning your [6:58] intelligence is strategy. [7:00] What parts of your AI do you want to [7:02] own? What parts do you want to rent? [7:05] Step two is team, figuring out how to [7:06] staff and organize people towards the [7:09] production of intelligence. [7:11] Step three is legibility. I really think [7:13] this gets glossed over [7:14] um and is incredibly important. So more [7:16] on this later. [7:17] And then finally step four, we're going [7:19] to talk about a technical roadmap. What [7:21] are the building blocks you need to [7:22] assemble in order to build your own [7:23] intelligence? [7:25] So let's dig in. [7:27] Um step one, defining which capabilities [7:29] you want to own versus rent. Sovereign [7:31] AI isn't binary. Uh you're not 0% or [7:34] 100% sovereign. [7:36] Um an important part of the strategy is [7:38] to draw the lines for which intelligence [7:40] do you want to own and which you're [7:41] comfortable outsourcing. [7:42] And so here's a useful framework to [7:44] think about um what parts you want to [7:47] own versus rent. I think there are four [7:49] important factors that go into this. Uh [7:52] one is cost. Like how how important is [7:54] this cost line item zero overall COGS? [7:57] Uh two is speed and latency. Is it a P0 [8:00] or not? Um factor three is performance. [8:03] And this is where it gets very [8:04] interesting. It used to be that you [8:06] would choose open weights when you [8:08] didn't care about performance. Now we're [8:10] getting to the point where in certain [8:11] domains you may be able to get better [8:13] performance by tuning models on your own [8:15] data. [8:16] And then finally, proprietary data. Are [8:19] you in a domain where the data you're [8:20] giving the model to improve it is super [8:22] proprietary super proprietary to your [8:24] business or less so? [8:27] Um so here are some examples of how [8:29] companies have decided to make this [8:30] trade-off. In coding, you have both [8:32] agents and you have auto complete. On [8:35] the agent side, this these are still [8:37] mostly rented today because you want [8:39] strong out-of-the-box performance and [8:42] latency isn't a P0. [8:44] Um on the other hand, for Tab auto [8:45] complete models in coding, you really [8:48] really care about speed and these uh [8:50] these API calls are so frequent that the [8:52] cost really rack up. And so most Tab [8:54] auto complete models now run on [8:57] sovereign intelligence. [8:58] I work at a cybersecurity company in [9:00] stealth. Alon, I think I saw you [9:02] earlier. Uh that owns its models [9:04] primarily for speed and performance um [9:06] and the ability to post-train the model [9:08] in very bespoke ways. [9:10] Um bio companies are moving that to [9:12] their own models because of the value of [9:14] proprietary data in that space. And so I [9:16] think this is just a useful framework to [9:18] think about which AI capabilities do I [9:21] want to own versus rent? [9:23] Step two, assemble a team. [9:25] I've shown two profiles of labs leaders [9:28] here just as examples. [9:29] Uh Nico comes more from the research [9:31] side of the house having done uh [9:33] research at Apple and then at Google [9:34] Brain. [9:35] Alex comes more from the engineering [9:37] side having held multiple engineering [9:39] roles at Microsoft and then Ramp. And I [9:41] I show this just to say there's multiple [9:43] paths to Nirvana and depending on the [9:45] flavor of research you're going to be [9:46] doing at your company, um whether it's [9:48] going to be more fundamental or applied, [9:50] there are different profiles uh that [9:52] work for a labs leader. [9:54] I also think it's important to think [9:55] about how to design your organization. [9:57] Um traditionally AI teams have [10:00] frequently been organized hub and spoke. [10:02] So you have a single platform team [10:04] supporting different uh different AI [10:06] product uh different application uh [10:09] product teams. [10:10] And what I've seen is that a lot of [10:11] companies are shoehorning this AI [10:13] platform team into doing the sovereign [10:16] AI stuff as well. Um I'd encourage folks [10:19] not to do this. [10:20] Uh [10:21] I'm for encourage people to start from [10:22] scratch here because this fundamentally [10:24] is not a platform services capability. [10:27] You want people that are able to think [10:30] on their feet, think on the frontier, [10:31] and produce frontier level research. And [10:33] it's such a different flavor of research [10:35] and you want them to be playing offense, [10:37] not just servicing teams. [10:39] And so we've seen uh small de novo teams [10:41] get very far here. Harvey, for example, [10:43] has published a ton of research. They [10:45] have just a team of seven. And so start [10:48] small, I'd say start from scratch, um [10:50] consider making it uh your own lab. [10:54] Step three, legibility. I think this is [10:56] totally underestimated in how important [10:58] it is because my guess is a lot of [11:00] people in this room are doing wonderful [11:02] research in-house internally and that [11:04] not all of it is very externally [11:06] legible. [11:07] And Winston Weinberg talked about how [11:09] the responsibility of a CEO is twofold. [11:12] One, drive substantive results, but two, [11:15] control the narrative, control [11:16] legibility around what you're building. [11:18] And I totally agree. [11:20] Um when it comes to owning your AI [11:22] stack, legibility really matters because [11:24] every single buyer right now is choosing [11:26] their AI champion. And so they're trying [11:29] to they're they're getting the same [11:30] pitch over and over again. They're [11:32] trying to discern which vendor is [11:34] sophisticated enough to take me to the [11:35] promised land. Um and you know, they [11:38] want to they want to choose people that [11:39] know what they're doing. Increasingly, [11:41] that means putting out your own [11:42] research. And so, being legible here [11:44] means doing excellent technical [11:46] marketing. Maybe having your own [11:48] separate branded labs or research group. [11:52] Publishing research with high taste. All [11:54] of this matters a lot. I think it goes [11:56] overlooked. And so, um for the all the [11:58] Sequoia companies in this room, I would [11:59] really really push on us to think about [12:01] this. [12:02] And then finally, step four, setting [12:04] your technical roadmap. Um at a high [12:06] level, the uh the rough journey that I [12:08] see companies take and every company [12:10] goes on a different journey, but first [12:11] you set your strategy. [12:13] Um second, defining evals. This is so [12:15] important. Um it is unglamorous work. It [12:18] is not fun work, but the more that you [12:20] do up front, the better positioned you [12:22] are for everything after. And so, I [12:23] think this is a really really crucial to [12:25] get right at the beginning. [12:26] Um next, we see companies starting to [12:28] play with model routers, with harnesses. [12:31] Um some companies find they can get good [12:33] performance with out-of-the-box models. [12:35] Um others are finding strong performance [12:37] gains from post-training, in some rarer [12:40] cases needing to move into mid-training, [12:41] pre-training. And then finally, setting [12:44] that machine up so that live customer [12:46] data is actually creating a feedback [12:48] loop where your model, your [12:50] intelligence, is improving with every [12:51] customer interaction. And so, this is [12:53] the rough journey that I see people go [12:55] on. Um again, every company is very very [12:58] different. And I'd encourage everyone in [13:00] the in the audience today is just [13:01] compare notes with with people around [13:03] you. Everyone is everyone's somewhere on [13:04] this journey. [13:07] The beauty of owning your stack is that [13:09] you can actually drive frontier-level [13:10] performance now. And so, this is [13:12] somewhat new. And in large part, this is [13:14] thanks to the newest open weight models, [13:17] uh especially Kimmy K3 and GLM 52 being [13:21] extremely good. Um because the weights [13:23] are available, they're actually much [13:25] more malleable than working with the [13:27] closed APIs. And so you start with a [13:30] baseline that's already [13:32] close to frontier, and then within with [13:34] a good enough technical roadmap, so with [13:36] strong post-training, prompt harness [13:39] engineering, online learning, you can [13:41] actually reach better than frontier [13:43] performance by owning your stack. And so [13:44] this is new for 2026. I think this is [13:46] very, very important. This is a big part [13:48] of why people are starting to think [13:49] about owning their intelligence. [13:52] Um I like diagrams, and so in an attempt [13:54] to orient us all, here is how I think [13:56] about the stack from an infrastructure [13:58] perspective. On the left-hand side, this [14:00] is production. This is your user-facing [14:02] intelligence. This is the stack that [14:05] drives um every every uh token that your [14:08] user ends up seeing. And fundamentally, [14:10] I think of the production stack as a [14:12] harness on top of a model. [14:14] Alongside that, you have a development [14:16] stack. These are the tools and vendors [14:18] that you use to get your intelligence [14:19] good. [14:20] In the closed model ecosystem, this [14:22] entire stack is very simple. You have [14:24] foundation models like Opus, GPT. Um and [14:27] you have the harnesses that come out of [14:28] the box with each. And you can get quite [14:30] far with the stack, including building [14:32] your own harnesses, prompts, feeding [14:34] contexts into the models, doing your own [14:36] evals. Um but you're not really having [14:38] to collect a ton of data. You're not [14:40] really having to train your own models. [14:42] And so it's much simpler stack. And so [14:44] I'd say this is a higher floor, but it's [14:46] a lower ceiling because you don't [14:47] actually have the ability to take your [14:49] own data, to take online production [14:51] data, and then improve your own [14:52] intelligence. [14:55] The minute that you start to think about [14:57] owning your own intelligence, it is like [14:59] opening a Pandora's Box because that [15:01] beautiful, clean API call is now um [15:04] having to train your own models. [15:06] And so instead of having a single model [15:08] API and some uh good performance out of [15:11] the box, you have to choose a [15:13] open-source base, do a lot of [15:15] post-training on top. [15:17] For your harness, you've got a choice of [15:18] several open harnesses, and then it's up [15:20] to you to configure the harness, the [15:22] logic, the tools, the context uh versus [15:25] taking an agent that just works out of [15:26] the box. [15:28] Um I just on the side here, context is [15:30] really, really important to driving [15:31] performance. [15:32] Um and there's several different flavors [15:34] of context that that that are driving [15:36] these these models. Um a vector database [15:39] like a Turbo puffer, an enterprise [15:41] knowledge graph like a Glean, um open [15:43] source connectors obviously via MCP, um [15:46] and then even novel approaches to [15:48] context. Uh Dan Biederman from N Gram is [15:51] here. Uh they're doing novel research [15:53] around encoding context in in the [15:55] weights themselves. Um so I'd encourage [15:57] anyone that wants to chat about memory [15:58] to go find Dan during one of the breaks. [16:01] Um and then alongside the production [16:03] stack, the development stack becomes way [16:05] more important when you own your own AI. [16:08] You have to carefully monitor evals to [16:10] know how your intelligence measures up [16:12] um and watch how the model performance [16:14] is drifting in production. [16:16] You need a lot of high-quality data [16:18] to post train the models for your [16:19] domain. [16:20] Sometimes this is expert trajectories. [16:23] Uh sometimes it's synthetic data. [16:25] Sometimes it's RL environments. [16:27] And then finally you need to think about [16:28] how to set up online learning for your [16:30] system so that your intelligence gets [16:31] better and better with every single user [16:33] interaction. [16:35] And so the way that we've set up today [16:36] is we've picked a series of technical [16:38] workshops to give you deep dives into [16:40] everything you need to build your own AI [16:42] other than pre-training your own models. [16:44] Uh Lynn from Fireworks is going to lead [16:46] a workshop on post training. [16:48] Harrison from LangChain is going to lead [16:50] a workshop on harnesses and evals. [16:53] Brendan from Mercor is going to lead a [16:55] workshop on RL environments, synthetic [16:57] data, and more. And then finally [16:59] Trajectory is going to lead a workshop [17:01] on online learning. [17:04] And then to bring it all together, [17:05] Harvey's going to lead a workshop on how [17:07] they approach building their entire AI [17:09] stack and strategy. And just yesterday [17:11] they announced Harvey research. You'll [17:13] see that um Um, uh, of their technical [17:16] partners are actually speaking today. [17:18] And so, we've really gone all out to get [17:20] the best possible lineup of speakers [17:21] today, both inside and outside the [17:23] portfolio. [17:25] >> [applause]