How Companies Are Building Their Own Intelligence | Sonya Huang, Sequoia Capital
频道: Sequoia Capital
视频: https://www.youtube.com/watch?v=bMMv0bZzONg
原文语言: en
统计: 共 15 轮 · Sonya Huang 14
[0:00] Sonya Huang
Okay, good morning everyone. Thank you all so much for being here. We have about 80 portfolio company founders and AI leaders in the room to explore a very timely topic, owning your intelligence or sovereign AI. Uh today's event is meant to be half a rallying call and half technical how-to. And so we have stacked the agenda with, I think, high high substance technical talks and demos so that we don't just talk the talk of building your own AI, but actually learn how to get there together. Uh so thank you all for taking the time out of your mornings to join us. I know your time is incredibly precious, and let's dive right in. So let's start with the opportunity. What is sovereign AI? Uh a sovereign is an independent state that has total self-governance. And sov- sovereign AI refers to companies owning their own intelligence without external dependencies down to the weights. An important nuance here is we are definitely not telling our companies to get off Opus or GPT. That is definitely not the message. Uh for coding agents, for desktop work, uh for frontier level APIs, the closed model APIs are wonderful. But what we've observed is that more and more of our companies are going down the path of wanting to build their own AI capabilities in parts of their products and vertically integrating towards owning more of this intelligence. And so today's session is meant to equip companies that are starting to go down that journey.
好，各位早上好。非常感谢大家今天到场。现场大概有 80 位被投公司的创始人和 AI 负责人，我们要一起聊一个特别应景的话题——owning your intelligence（把智能握在自己手里），也就是 sovereign AI（主权 AI，指公司自己拥有并掌控模型能力，而不是完全依赖外部）。今天这场活动，一半是动员令，一半是技术操作手册。所以我们把议程排得很满，全是高浓度的技术分享和 demo，不只是嘴上讲讲「自建 AI」，而是真的一起搞明白怎么走到那一步。谢谢大家挤出一早上的时间过来，我知道你们的时间极其宝贵，那我们就直接进正题。先说机会在哪。什么是 sovereign AI？sovereign（主权）指的是一个完全自治、不受外部支配的独立国家。sovereign AI 指的就是：公司拥有自己的智能，一路往下拥有到模型权重那一层，不依赖任何外部。这里有个很重要的细微差别——我们绝对不是在劝大家弃用 Opus 或者 GPT，完全不是这个意思。做 coding agent、做桌面端的活儿、需要 frontier（前沿）级别 API 的场景，闭源模型 API 依然非常好用。但我们观察到的是，越来越多的被投公司正在走上另一条路：在产品的某些部分自建 AI 能力，通过垂直整合，把更多的智能握到自己手里。所以今天这场，就是为已经开始踏上这条路的公司准备装备的。
[1:24] Sonya Huang
We're obviously not alone in this idea. In the last month, uh the rhetoric around sovereign AI has escalated sharply with folks like Alex Karp and Satya speaking up in support of companies owning their own intelligence. Just last week, Jensen led the charge in making sure that open weight models remain available in the US. And it was awesome to see the near unanimous wave of support. I think the message is very clear that companies want to own their intelligence. They want to own, not rent, their weights. And I think it's simply because intelligence is too core, too fundamental of a property to just outsource. We're glad that sovereign AI is in the zeitgeist right now uh because we think it's a good thing for the world. On one hand, you have centralized intelligence where a single all-powerful AI powers more and more of the GD the world's GDP as a black box sucking in all the data exhaust, all the data flywheels from the rest of the world. On the other hand, you have decentralized intelligence where the whole world builds on top of a solid core, but every individual person, company builds their own intelligence on top bespoke to their own data, their own industries, their own personalization, uh their own way of working, their own taste.
显然有这个想法的不止我们。过去一个月里，围绕 sovereign AI 的声浪急剧升温，Alex Karp、Satya 这样的人物都公开发声，支持公司拥有自己的智能。就在上周，Jensen 带头呼吁，要确保 open weight（开放权重）模型在美国继续可得。看到几乎一边倒的支持声，真的很振奋。我觉得信号已经非常清楚了：公司想拥有自己的智能，他们想「拥有」权重，而不是「租用」权重。原因很简单——智能这个东西太核心、太底层了，不该就这么外包出去。我们很高兴 sovereign AI 现在成了时代情绪的一部分，因为我们认为这对世界是件好事。一边是中心化的智能：一个无所不能的 AI 驱动世界上越来越大比例的 GDP，它像个黑箱，把全世界的数据尾气、数据飞轮全都吸进去。另一边是去中心化的智能：全世界都建在一个扎实的内核之上，但每一个人、每一家公司都在这个内核上搭建属于自己的智能——贴合自己的数据、自己的行业、自己的个性化需求、自己的工作方式、自己的品味。
[2:38] Sonya Huang
And so the ecosystem flourishes and individuality triumphs. No single company swallows the rest. I think this is a much more optimistic view of the world. And so we work with dozens of companies that are going down the journey of building their own AI. Um here are the biggest reasons that we've seen people move. Reason number one is cost. Uh especially for low, zero, negative margin companies, sovereign AI isn't a nice-to-have, it's a must-have. Um ironically, the more successful your AI product is, the higher your your AI cogs tend to be. And so it's actually the companies that have been most advanced in the deployment of AI that have been the first to go on this journey of of owning their own models. Reason number two is speed. And so in certain domains, coding is one of them, uh security is another one, a small distilled custom model can beat a large general one because speed is so important. Num- Reason number three is performance. Um this is a relatively newer one. I would say last year, most companies were not choosing to own their intelligence to generate better performance. Uh but we're now at the point where open models can outperform closed ones on your domain. And we're going to spend a lot of today's agenda talking about how to get there. And then reason number four, controlling your own destiny. Uh Anthropic and OpenAI, I actually think to their credit, they've been really wonderful partners to a lot of the ecosystem. But companies are increasingly finding that they want their own set of independent legs to stand on as well.
这样生态才能繁荣，个体性才能胜出，不会有哪一家公司把其他所有人吞掉。我觉得这是一个乐观得多的世界图景。我们现在和几十家正在自建 AI 的公司一起走在这条路上。下面说说我们看到的、大家决定动手的几个最主要理由。理由一：成本。尤其对于低毛利、零毛利甚至负毛利的公司，sovereign AI 不是锦上添花，而是生死攸关。有点反讽的是：你的 AI 产品越成功，你的 AI 成本（COGS）往往越高。所以恰恰是那些 AI 部署走在最前面的公司，最先踏上了自建模型这条路。理由二：速度。在某些领域——coding 是一个，安全是另一个——一个小的、经过 distillation（蒸馏，把大模型的能力压缩进小模型）的定制模型，是能打赢大的通用模型的，因为速度在这些场景里太关键了。理由三：性能。这一条相对新一些。我会说，去年绝大多数公司自建智能并不是为了拿到更好的性能。但我们现在已经走到了这样一个节点：在你自己的领域里，open model 是可以跑赢闭源模型的。今天的议程里我们会花很多时间讲怎么做到这一点。理由四：掌控自己的命运。说句公道话，Anthropic 和 OpenAI 其实一直是整个生态非常好的合作伙伴。但越来越多的公司发现，自己也需要有一副能独立站住的腿。
[4:11] Sonya Huang
Um does anybody here come from the crypto days or remember the crypto days? Okay. Um do you guys remember this meme? Okay. Uh in the crypto of the days there was this meme for the DeFi degens. Uh not your keys, not your crypto. And so if somebody was custodying your crypto for you, it fundamentally wasn't yours. I hereby present the AI version of this meme. Not your weights, not your product. Um I think that for product to be truly yours, I think it's reasonable to think that you need to be able to control control and custody your own weights. Pat shows this slide at AI Ascent talking about the race for the application layer. Uh the He looks so proud of yourself. Uh the punchline is that both the AGI labs and the application companies are racing to be the user-facing products from different ends. The foundation model labs from the model side and the application companies from the user back. I think we're seeing a new dynamic emerge now, which is actually that battleground is increasingly becoming not just the race for the application layer, but the race for the intelligence layer. And so this this battleground is no longer just about who gets to control the product, the UI, the go-to-markets, the wrapping. It's actually about who can own the intelligence itself and shape better intelligence in the product. So the product is the intelligence, and the newest battleground is for not just the product surface, but for the intelligence layer itself.
在座有经历过加密货币那波的吗？或者还记得那个年代的？好。你们还记得这个梗吗？好。当年在 DeFi degens（玩去中心化金融的那帮老炮/赌狗）圈子里有句名言——not your keys, not your crypto（私钥不在你手上，币就不是你的）。意思是，如果你的加密货币是别人替你托管的，那它本质上就不属于你。今天我在这里正式发布这个梗的 AI 版本：not your weights, not your product（权重不在你手上，产品就不是你的）。我觉得，一个产品要真正属于你，你得能掌控、能自己托管自己的权重，这个要求并不过分。Pat 在 AI Ascent 上放过这张片子，讲的是「应用层争夺战」。他自己看着还挺得意的。这张片的意思是：AGI 实验室和应用层公司，正从两头往中间冲，都想成为直接面向用户的产品——基础模型实验室从模型那一端往上打，应用公司从用户那一端往回打。而我觉得现在出现了一个新的态势：主战场已经不只是应用层的争夺，而是智能层的争夺。这场仗不再只关乎谁能掌控产品、UI、go-to-market、外面那层包装，而是关乎谁能拥有智能本身、谁能把产品里的智能打磨得更好。产品即智能，最新的战场不只是产品界面，而是智能层本身。
[5:37] Sonya Huang
And so the hottest new labs, in my opinion, are actually uh the applied research that we see coming out of companies right now like Harvey, like Factory, Glean, Open evidence, Semgrep, Ramp. The list goes on and on, and I think the research is spanning everything from evals and benchmarks to harness engineering to new algorithmic techniques for fine-tuning and a lot, lot more. And we started this uh we started this morning talking about centralized versus decentralized intelligence. I think it's really wonderful to see the amount of innovation that is happening in these democratized intelligence world. Like I actually think the application companies are the newest Neo labs. Okay, so I assume that everyone here today is pretty bought into this journey. Let's assume that you want to build your own lab, build your own models. How do you go from zero to one to 100? We're going to do something a little bit different today here. Um I'm going to lay out an opinionated framework and technical roadmap. And so take that with a giant grain of salt. Uh every company is different. And and I'm not technical. And so take this with a giant, giant grain of salt, but I hope it provides a useful starting point for how to think about building your own intelligence.
所以在我看来，现在最热的新一代「实验室」，其实是我们在 Harvey、Factory、Glean、OpenEvidence、Semgrep、Ramp 这些公司里看到的 applied research（应用研究，指贴着真实产品问题做的研究）团队。这个名单还能一直列下去。而且这些研究覆盖的面很广：从 eval（评测集）、benchmark，到 harness engineering（模型外壳/调度框架的工程），再到 fine-tune（微调）的新算法技巧，还有很多很多。今天早上一开场我们聊的是中心化智能 vs 去中心化智能。我觉得，看到「民主化的智能」这个世界里涌现出这么多创新，真的非常了不起。我甚至认为，应用层公司才是最新一代的 Neo lab（新型实验室）。好，我假定今天在座各位对这条路都已经买账了。假设你就是想自建实验室、自建模型，那怎么从 0 到 1、再到 100？今天我们来点不太一样的。我会摆出一套带鲜明观点的框架和技术路线图。这套东西请务必打个大大的折扣来听——每家公司情况都不一样，而且我本人也不是技术出身。所以真的，打一个非常非常大的折扣。但我希望它至少能给你提供一个有用的起点，帮你想清楚该怎么自建智能。
[6:55] Sonya Huang
Uh step one um to owning your intelligence is strategy. What parts of your AI do you want to own? What parts do you want to rent? Step two is team, figuring out how to staff and organize people towards the production of intelligence. Step three is legibility. I really think this gets glossed over um and is incredibly important. So more on this later. And then finally step four, we're going to talk about a technical roadmap. What are the building blocks you need to assemble in order to build your own intelligence? So let's dig in. Um step one, defining which capabilities you want to own versus rent. Sovereign AI isn't binary. Uh you're not 0% or 100% sovereign. Um an important part of the strategy is to draw the lines for which intelligence do you want to own and which you're comfortable outsourcing. And so here's a useful framework to think about um what parts you want to own versus rent. I think there are four important factors that go into this. Uh one is cost. Like how how important is this cost line item zero overall COGS? Uh two is speed and latency. Is it a P0 or not? Um factor three is performance. And this is where it gets very interesting. It used to be that you would choose open weights when you didn't care about performance. Now we're getting to the point where in certain domains you may be able to get better performance by tuning models on your own data.
第一步，把智能握在自己手里的第一步是战略。你的 AI，哪些部分你想自己拥有？哪些部分你愿意租？第二步是团队：搞清楚怎么配人、怎么组织，把人组织到「生产智能」这件事上。第三步是 legibility（可被外界读懂、被看见）。我真的觉得这一点常常被一带而过，但它极其重要，稍后细讲。最后第四步，我们讲技术路线图：要自建智能，你需要拼装哪些积木？我们一个个来。第一步，界定哪些能力自己拥有、哪些租用。Sovereign AI 不是非黑即白的，你不是 0% 主权就是 100% 主权。战略里很重要的一环，就是划清那条线：哪部分智能你要自己拥有，哪部分你可以放心外包。这里有个挺好用的框架来判断自建还是租用，我觉得主要看四个因素。第一是成本：这一项成本在你整体 COGS 里占多大分量？第二是速度和延迟：它是不是 P0（最高优先级）？第三是性能——这一条现在变得非常有意思。过去的逻辑是：你不在乎性能的时候才选 open weight。而现在我们正走到这样一个节点：在某些领域，用你自己的数据去调模型，反而能拿到更好的性能。
[8:16] Sonya Huang
And then finally, proprietary data. Are you in a domain where the data you're giving the model to improve it is super proprietary super proprietary to your business or less so? Um so here are some examples of how companies have decided to make this trade-off. In coding, you have both agents and you have auto complete. On the agent side, this these are still mostly rented today because you want strong out-of-the-box performance and latency isn't a P0. Um on the other hand, for Tab auto complete models in coding, you really really care about speed and these uh these API calls are so frequent that the cost really rack up. And so most Tab auto complete models now run on sovereign intelligence. I work at a cybersecurity company in stealth. Alon, I think I saw you earlier. Uh that owns its models primarily for speed and performance um and the ability to post-train the model in very bespoke ways. Um bio companies are moving that to their own models because of the value of proprietary data in that space. And so I think this is just a useful framework to think about which AI capabilities do I want to own versus rent? Step two, assemble a team. I've shown two profiles of labs leaders here just as examples.
最后一个因素是专有数据。你所在的领域，你喂给模型去改进它的那些数据，是极其专有、极其贴身于你业务的，还是没那么专有？下面举几个例子，看看大家是怎么做这个取舍的。在 coding 里，既有 agent，也有自动补全。agent 这一侧，今天大部分公司还是选择「租」，因为你需要开箱即用的强性能，而且延迟不是 P0。反过来，coding 里的 Tab 自动补全模型，你极其在意速度，而且这类 API 调用太频繁了，成本会像滚雪球一样堆起来。所以现在大部分 Tab 自动补全模型，跑的都是自己的 sovereign intelligence。我们合作的一家还在隐身状态的网络安全公司——Alon，我刚才好像看到你了——他们自建模型主要就是为了速度和性能，以及能用非常定制化的方式做 post-training（后训练，在基座模型之上继续训练，让它更适配特定任务）。生物医药类公司也在往自建模型走，原因是专有数据在那个领域实在太值钱。所以我觉得这套框架挺好用的，帮你想清楚：哪些 AI 能力我要自己拥有，哪些我租。第二步，搭团队。这里我放了两位「实验室负责人」的画像，只是举例。
[9:29] Sonya Huang
Uh Nico comes more from the research side of the house having done uh research at Apple and then at Google Brain. Alex comes more from the engineering side having held multiple engineering roles at Microsoft and then Ramp. And I I show this just to say there's multiple paths to Nirvana and depending on the flavor of research you're going to be doing at your company, um whether it's going to be more fundamental or applied, there are different profiles uh that work for a labs leader. I also think it's important to think about how to design your organization. Um traditionally AI teams have frequently been organized hub and spoke. So you have a single platform team supporting different uh different AI product uh different application uh product teams. And what I've seen is that a lot of companies are shoehorning this AI platform team into doing the sovereign AI stuff as well. Um I'd encourage folks not to do this. Uh I'm for encourage people to start from scratch here because this fundamentally is not a platform services capability. You want people that are able to think on their feet, think on the frontier, and produce frontier level research. And it's such a different flavor of research and you want them to be playing offense, not just servicing teams.
Nico 更偏研究出身，先后在 Apple 和 Google Brain 做过研究。Alex 更偏工程出身，在 Microsoft 和 Ramp 都担任过多个工程岗位。我放这两个例子只是想说明：通往彼岸的路不止一条。取决于你的公司要做哪一类研究——偏基础还是偏应用——适合做实验室负责人的人才画像是不一样的。另外我觉得，组织怎么设计也很值得想清楚。传统上 AI 团队常常是「轮辐式」（hub and spoke）组织：一个中央平台团队，支撑下游各个 AI 产品线、各个应用产品团队。而我看到的情况是，很多公司在硬把这个 AI 平台团队塞去干 sovereign AI 的活儿。我建议大家别这么干。我更建议在这件事上从零起一支新队伍，因为它本质上不是一种「平台服务」型的能力。你需要的是能随机应变、能在前沿上思考、能产出前沿级别研究的人。这是完全不同味道的研究，而且你希望他们是在打进攻，而不是在给别的团队做后勤支持。
[10:39] Sonya Huang
And so we've seen uh small de novo teams get very far here. Harvey, for example, has published a ton of research. They have just a team of seven. And so start small, I'd say start from scratch, um consider making it uh your own lab. Step three, legibility. I think this is totally underestimated in how important it is because my guess is a lot of people in this room are doing wonderful research in-house internally and that not all of it is very externally legible. And Winston Weinberg talked about how the responsibility of a CEO is twofold. One, drive substantive results, but two, control the narrative, control legibility around what you're building. And I totally agree. Um when it comes to owning your AI stack, legibility really matters because every single buyer right now is choosing their AI champion. And so they're trying to they're they're getting the same pitch over and over again. They're trying to discern which vendor is sophisticated enough to take me to the promised land. Um and you know, they want to they want to choose people that know what they're doing. Increasingly, that means putting out your own research. And so, being legible here means doing excellent technical marketing. Maybe having your own separate branded labs or research group.
我们看到不少全新组建的小团队在这件事上走得非常远。比如 Harvey 发了大量研究，团队只有七个人。所以我的建议是：起步就小，从零开始，认真考虑把它做成你自己的实验室。第三步，legibility（可被外界读懂）。我觉得这一点的重要性被严重低估了，因为我猜在座很多人内部都在做非常棒的研究，但这些研究并不都能被外界看见、看懂。Winston Weinberg 讲过，CEO 的职责有两重：第一，做出实打实的结果；第二，掌控叙事，掌控外界对你在做什么的可读性。我完全同意。在「自己拥有 AI 技术栈」这件事上，legibility 特别重要，因为现在每一个买家都在挑选自己的「AI 冠军」。他们反反复复听到的是同一套说辞，真正想分辨的是：哪家供应商足够有本事，能把我带到应许之地？他们想选的是真正懂行的人。而这越来越意味着，你得把自己的研究公开发出来。所以在这件事上，「可被读懂」意味着：把技术营销做到极致，也许还要有一个独立品牌的 labs 或者研究组。
[11:52] Sonya Huang
Publishing research with high taste. All of this matters a lot. I think it goes overlooked. And so, um for the all the Sequoia companies in this room, I would really really push on us to think about this. And then finally, step four, setting your technical roadmap. Um at a high level, the uh the rough journey that I see companies take and every company goes on a different journey, but first you set your strategy. Um second, defining evals. This is so important. Um it is unglamorous work. It is not fun work, but the more that you do up front, the better positioned you are for everything after. And so, I think this is a really really crucial to get right at the beginning. Um next, we see companies starting to play with model routers, with harnesses. Um some companies find they can get good performance with out-of-the-box models. Um others are finding strong performance gains from post-training, in some rarer cases needing to move into mid-training, pre-training. And then finally, setting that machine up so that live customer data is actually creating a feedback loop where your model, your intelligence, is improving with every customer interaction. And so, this is the rough journey that I see people go on. Um again, every company is very very different. And I'd encourage everyone in the in the audience today is just compare notes with with people around you. Everyone is everyone's somewhere on this journey.
发布有品味的研究。这些都非常重要，而我觉得它们普遍被忽视了。所以在座的 Sequoia 系公司，我真的想使劲推一把，让大家好好想想这件事。最后，第四步：定你的技术路线图。从大方向上讲，我看到公司们走的大致路径是这样的——当然每家公司走的路都不一样。第一，定战略。第二，定义 eval（评测集），这一步太重要了。它一点都不光鲜，做起来也不有趣，但你在前期下的功夫越多，后面每一步的位置就越有利。所以我觉得，一开始就把这件事做对，是真正关键的。接下来，我们看到公司开始折腾 model router（模型路由）、折腾 harness（模型外壳）。有些公司发现开箱即用的模型就能拿到不错的性能；另一些公司则在 post-training 上拿到了显著的性能提升；还有比较少见的情况，需要一路做到 mid-training 甚至 pre-training。最后，是把这台机器搭起来，让线上真实的客户数据真正形成反馈闭环——你的模型、你的智能，随着每一次客户交互不断变强。这就是我看到大家大致走过的路径。再强调一次，每家公司的情况都非常非常不同。我也建议今天在场的各位，多和身边的人对一对进度，每个人都在这条路的某个位置上。
[13:07] Sonya Huang
The beauty of owning your stack is that you can actually drive frontier-level performance now. And so, this is somewhat new. And in large part, this is thanks to the newest open weight models, uh especially Kimmy K3 and GLM 52 being extremely good. Um because the weights are available, they're actually much more malleable than working with the closed APIs. And so you start with a baseline that's already close to frontier, and then within with a good enough technical roadmap, so with strong post-training, prompt harness engineering, online learning, you can actually reach better than frontier performance by owning your stack. And so this is new for 2026. I think this is very, very important. This is a big part of why people are starting to think about owning their intelligence. Um I like diagrams, and so in an attempt to orient us all, here is how I think about the stack from an infrastructure perspective. On the left-hand side, this is production. This is your user-facing intelligence. This is the stack that drives um every every uh token that your user ends up seeing. And fundamentally, I think of the production stack as a harness on top of a model. Alongside that, you have a development stack. These are the tools and vendors that you use to get your intelligence good.
自己拥有整个技术栈，最妙的地方在于：你现在真的能跑出 frontier（前沿）级别的性能了。这件事其实挺新的。这在很大程度上要归功于最新一批 open weight 模型，尤其是 Kimi K3 和 GLM 5.2，都好得不得了。因为权重是公开的，相比闭源 API，它们的可塑性要强得多。所以你的起点就已经很接近 frontier 了，再配上一份足够好的技术路线图——扎实的 post-training、prompt 和 harness 工程、online learning（在线学习）——你完全可以靠自建这一整套，跑出比 frontier 还好的性能。这是 2026 年才出现的新局面。我觉得这一点非常非常重要，也是很多人开始认真考虑「拥有自己的智能」的一大原因。我这个人喜欢画图，为了让大家有个统一的坐标，这就是我从基础设施视角理解的整个栈。左边这一侧是生产环境，也就是面向用户的智能，是驱动你的用户最终看到的每一个 token 的那套栈。本质上，我把生产栈理解成「一个模型 + 套在它外面的 harness」。与之并行的是开发栈：也就是你为了把智能打磨好而用到的那些工具和供应商。
[14:20] Sonya Huang
In the closed model ecosystem, this entire stack is very simple. You have foundation models like Opus, GPT. Um and you have the harnesses that come out of the box with each. And you can get quite far with the stack, including building your own harnesses, prompts, feeding contexts into the models, doing your own evals. Um but you're not really having to collect a ton of data. You're not really having to train your own models. And so it's much simpler stack. And so I'd say this is a higher floor, but it's a lower ceiling because you don't actually have the ability to take your own data, to take online production data, and then improve your own intelligence. The minute that you start to think about owning your own intelligence, it is like opening a Pandora's Box because that beautiful, clean API call is now um having to train your own models. And so instead of having a single model API and some uh good performance out of the box, you have to choose a open-source base, do a lot of post-training on top. For your harness, you've got a choice of several open harnesses, and then it's up to you to configure the harness, the logic, the tools, the context uh versus taking an agent that just works out of the box.
在闭源模型的生态里，整个技术栈非常简单。你有 Opus、GPT 这样的 foundation model（基础大模型），每家都配好了开箱即用的 harness（套在模型外面的那层调度框架，提示词、工具调用、上下文管理都在里面）。就靠这套栈，你其实能走得相当远——自己搭 harness、写 prompt、往模型里喂上下文、自己跑 eval（模型效果评测），这些都做得了。但你基本不需要去攒海量数据，也不需要自己训模型，所以栈要简单得多。我的判断是：这条路下限更高，但天花板更低——因为你并没有能力拿自己的数据、拿线上跑出来的真实生产数据，反过来让自己的智能变得更强。而你一旦动了「把智能握在自己手里」的念头，基本就等于打开了潘多拉的盒子：原本那个干净漂亮的 API 调用，现在变成了得自己训模型。你不再是调一个模型 API、开箱就有不错的效果，而是要先挑一个开源底座，在上面做大量 post-training（后训练：在开源基础模型之上继续训练，让它贴合你的领域）。harness 这一层，摆在你面前的是好几个开源方案，然后 harness 怎么配、逻辑怎么写、工具怎么接、上下文怎么组织，全得你自己来——而不是拿一个开箱即用的 agent（能自己规划、调工具、多步把活干完的智能体）直接跑起来。
[15:28] Sonya Huang
Um I just on the side here, context is really, really important to driving performance. Um and there's several different flavors of context that that that are driving these these models. Um a vector database like a Turbo puffer, an enterprise knowledge graph like a Glean, um open source connectors obviously via MCP, um and then even novel approaches to context. Uh Dan Biederman from N Gram is here. Uh they're doing novel research around encoding context in in the weights themselves. Um so I'd encourage anyone that wants to chat about memory to go find Dan during one of the breaks. Um and then alongside the production stack, the development stack becomes way more important when you own your own AI. You have to carefully monitor evals to know how your intelligence measures up um and watch how the model performance is drifting in production. You need a lot of high-quality data to post train the models for your domain. Sometimes this is expert trajectories. Uh sometimes it's synthetic data. Sometimes it's RL environments. And then finally you need to think about how to set up online learning for your system so that your intelligence gets better and better with every single user interaction.
这里顺带插一句：上下文对性能的影响真的非常非常大。往这些模型里灌上下文，有好几种不同的路子——像 Turbopuffer 这样的向量数据库，像 Glean 这样的企业知识图谱，当然还有通过 MCP 接进来的各种开源连接器，甚至还有一些很新的做法：Ngram 的 Dan Biderman 今天也在现场，他们在研究怎么把上下文直接编码进模型的权重里。所以想聊 memory（模型记忆）这块的，中场休息的时候去找 Dan 聊就对了。另外，当你真的要拥有自己的 AI，除了这套生产栈，开发栈的分量会陡然变重。你得非常小心地盯住 eval，才知道自己的智能到底是什么水平，也才看得出模型在生产环境里的表现有没有在漂移。你需要大量高质量数据，才能针对自己的领域做 post-training——有时候是专家轨迹（expert trajectory：专家一步步操作留下的完整过程记录），有时候是合成数据，有时候是 RL 环境（强化学习的训练场，让模型在里面反复试错、拿到反馈）。最后你还得想清楚，怎么给你的系统搭起 online learning（在线学习：模型在真实使用过程中持续更新），让每一次用户交互都能把你的智能再往上推一点。
[16:35] Sonya Huang
And so the way that we've set up today is we've picked a series of technical workshops to give you deep dives into everything you need to build your own AI other than pre-training your own models. Uh Lynn from Fireworks is going to lead a workshop on post training. Harrison from LangChain is going to lead a workshop on harnesses and evals. Brendan from Mercor is going to lead a workshop on RL environments, synthetic data, and more. And then finally Trajectory is going to lead a workshop on online learning. And then to bring it all together, Harvey's going to lead a workshop on how they approach building their entire AI stack and strategy. And just yesterday they announced Harvey research. You'll see that um Um, uh, of their technical partners are actually speaking today. And so, we've really gone all out to get the best possible lineup of speakers today, both inside and outside the portfolio.
所以今天的安排是这样的：我们挑了一系列技术 workshop，把「自己造 AI」这件事里，除了预训练模型之外的每一个环节，都给你讲透。Fireworks 的 Lin 来讲 post-training。LangChain 的 Harrison 来讲 harness 和 eval。Mercor 的 Brendan 来讲 RL 环境、合成数据这些。最后是 Trajectory 来讲 online learning。再往上收口，Harvey 会讲他们整套 AI 技术栈和策略是怎么搭起来的——就在昨天，他们刚刚发布了 Harvey Research。你们会发现，他们的技术合作方，今天有好几家都在台上。所以我们真的是拼尽全力，凑齐了今天这份能拿到的最强讲者阵容——投资组合里的、组合外的都有。
[17:25]
[applause]
[掌声]