Glean：全球首家企业级生成式 AI 公司

嘿，Arvind，感谢你来做客。你是我们这个时代的技术巨擘之一。你创办了许多大型公司。Rubrik 上市了，现在又有 Glean 和 Composer。我知道你和 Karan 有渊源。他在 Rubrik 时曾与你共事，后来促成了我们与 Glean 和 Composer 的合作。你能讲讲你是怎么认识 Karan 的，以及这一切是怎么开始的吗？

Hey Arvind, thank you for joining us. You're one of the technology luminaries of our time. You started many massive companies. Rubrik went public, and now Glean and Composer. I know you and Karan have a history together. He worked with you at Glean at Rubrik, and then this led to this partnership that we have with Glean and Composer. Can you tell me like how you met Karan and how that started?

是的，正如你所说，首先感谢你的邀请。我很高兴来到这里。我和 Karan 从 Rubrik 时代就认识了，他是我们早期的明星工程师之一。我们就是这样认识的。当他创办这家公司，并且开发的技术对我们来说有合作意义时，这显然是个明确的选择，因为我们信任他。

Yeah, I mean, as you said, first of all, like thanks for having me. Excited to be here. Karan and I have known each other since our Rubrik days, where he was one of our early star engineers. So that's how we got to know each other. And when he started this company, and he was building technology which kind of made sense for us to partner on. So it was like a clear choice, because we believe in him.

谢谢，Ajeet。我的意思是，显然我认为 Rubrik 是我加入的第一家公司。感谢 Aakrit 让我加入。我觉得那是我学习如何构建系统、如何组建工程团队的地方。所以，是的，也感谢 Aakrit 第一次就相信我们。这也在某种程度上塑造了我们的产品。

Thanks, Ajeet. And I mean, obviously I think I got Rubrik was the first company that I joined. Thanks Aakrit for having me there. I think that's where I essentially learned how to build systems, how to build the engineering team. So kind of yeah, and thanks Aakrit for believing in us for the first time. And kind of like that has shaped our product as well.

你能简要解释一下 Glean 是做什么的，以及它的公司历史吗？我知道你们最初是一家企业搜索公司，有点像向量搜索系统。后来你们扩展到了更广泛的领域。在最近一期播客中，你把它描述为 ChatGPT 和 Claude Code 的超集。那么，你能讲讲今天的产品是什么样的，以及你们是如何走到这一步的，历史是怎样的吗？

Can you briefly explain like what Glean does and its history as a company? I know you guys started as an enterprise search company, sort of vector search bag systems. You moved to something much broader. In a recent podcast you described it as something like a super set of chat GPT and Claude Code. So yeah, like can you talk me through like what the product is today and like how you sort of got here, what the history looks like?

是的。Glean 始于 2019 年初，这使我们成为全球首家企业生成式 AI 公司。嗯，我的意思是，需要说明的是，当时“生成式 AI”这个词还不存在。但我们是第一批使用 Transformer 和语言模型，并将这些技术引入企业的人。我们这样做是为了我们构建的第一个产品，即面向员工的企业内部搜索。我不知道有多少人知道，Transformer 和语言模型的起源最初是在 Google 构建的，其唯一目的是让 Google 搜索更智能、更聪明。所以当我们创办 Glean，并且也在为企业构建搜索产品时，正如我所说，使用这项技术完全合理。这就是我们成为第一批将 Transformer 引入企业的公司的原因。我们想解决这个问题的原因是我们在 Rubrik 的经历。事实上，在此之前我在 Google 也有过这样的经历，你知道，我曾在那里工作。要找到你需要的信息，得到你问题的答案，从而让你在正在做的事情上取得进展，这总是极其困难。这非常令人沮丧。在 Rubrik 时，这是我们脉搏调查中得分最低的问题。人们抱怨说，找到他们工作所需的信息太难了。所以我们就是这样开始的。现在，我们已经从那个阶段进化了。所以，除了拥有一个看起来像公司内部的 Google 或 ChatGPT 的产品，一个人们进来提问或给我们分配工作的地方之外，Glean 所做的是使用现有的最佳模型，这些模型实际上可以帮助解决这个任务或回答这个问题。这些模型来自 OpenAI、Anthropic、Google 等公司，以及所有开源模型。但更重要的是，Glean 所做的是连接到贵公司的所有内部数据和系统。它知道贵公司内部每一块知识或数据的位置。它知道谁在内部拥有使用这些信息的权限。它还深入了解哪些内容、哪些数据是新鲜的、最新的、相关的、高质量的。所以，我们拥有那种关于你的业务、数据、知识以及公司内部工作方式的上下文。我们能够利用这些上下文和知识，以及所有这些优秀的 AI 模型和技术，开始为你工作。你与 Glean 合作的方式是提问、给我们任务。你可以把我们看作是 ChatGPT、Claude、Cohere 或 Gemini 的超集。我们会利用公司的所有上下文来实际解决这个任务。事实上，我们会为必须执行的每个子任务挑选合适的模型、最佳模型，以完成该任务。所以这就是我们的一个产品，非常简单，用户就像使用 ChatGPT 一样。但在这个产品之下，是一个平台。这个平台连接到你的所有企业系统，我们能够读取信息、采取行动，并真正与你合作。我们能够理解所有这些数据的安全性和治理，并能够构建企业的深度上下文和知识图谱。而且，这个平台就像我们提供给客户的产品一样，也提供给客户，让他们直接集成到他们的其他 AI 系统中。

Yeah. So Glean started in early 2019, which makes us the world's first enterprise generative AI company. Well, I mean, just to be clear, the term generative AI did not exist at the time. But we were the first ones to play with transformers and language models and bring these technologies to the enterprises. And we were doing it for our first product that we'd built, which was search for employees within their own company. I don't know how many people know this, but the origins of transformers and language models is initially these things were built at Google, with the goal with the single purpose of making Google search smarter, more intelligent. So when we started Glean and we were also building a search product for businesses like I said, it made all the sense for us to use this technology. So that's how we became the first ones to bring transformers to the enterprises. The reason we wanted to solve that problem was our own experience at Rubrik. And in fact actually my experience is at Google before that, you know, where I used to work. It's always been incredibly hard to actually find the information that you need to get answers to questions that you have that will allow you to then go make progress on the things that you're working on. It's deeply frustrating. It used to be at Rubrik it was our lowest scoring question in our pulse surveys. People complaining that, you know, it was so hard to find information that they needed to do their work. So that's how we got started. Now, we've evolved from that. So, in addition to actually of course, you know, having this product that looks like Google or ChatGPT inside your company, one place where people come in and ask questions or give us some work to do. And what Glean does is it uses both like, you know, the best models that are out there that can actually help solve this task or answer this question. And these models come from companies like OpenAI, Anthropic, group, you know, Google and all the open source models. But more importantly, what Glean does is it's connected to all of your company's internal data and systems. It knows where every single piece of knowledge or data is inside your company. It knows who has permissions to use that information internally. It's also understands deeply like, you know, what content, what data is fresh, up-to-date, relevant, high quality. So, we sort of had that, you know, that context of your business, your data, your knowledge, and how work happens inside the company. And we're able to use that context and that knowledge along with all these great AI models and technology to then start to work for you. And the way you work with Glean is you ask questions, you give us tasks. You can think of us as a superset of ChatGPT, Claude, Cohere, or Gemini. Like, we will actually solve that task using all of that company's context. And in fact, you know, picking the right models, the best models for the each of the subtasks, you know, that we will have to execute to complete that task. So that's our one product, like very simple users just like how you use ChatGPT. But, underlying that is this platform. This platform that's connected to all of your enterprise systems, where we are able to read information, where to take actions, and where we actually partner with you. We are able to understand the security and governance of all of that data, and we are able to build this deep context and knowledge graph of an enterprise. And all of that platform is also, just like how it's available to our own product that we deliver to our customers, that platform is also available to our customers to directly integrate into their other AI systems.

OpenAI 曾一度公开表示你是他们的主要竞争对手之一。但同时，你能够利用并与之合作。

OpenAI at some point had stated directly publicly that you were one of their main competitors. But at the same time, you were able to leverage and partner with them.

所以，我实际上并不担心竞争。我从来没有担心过。

So, I don't actually worry about competition. I've never worried about it.

许多拥有数据优势的应用公司也在进入模型业务。Cursor 有 Composer 等。这也是某种……

A lot of application companies with their data advantage are also coming to the model businesses. Cursor has Composer, etc. Is that also something...

是的。事实上，今天，如果我们所有人都联合起来，如果所有模型公司、所有应用公司以及类似的公司合并成一家公司，并且我们开始在没有竞争的情况下向所有企业提供产品和服务，我们仍然会不足。我们仍然无法满足那部分需求。

Yeah. In fact, today, if all of us came together, if all the model companies, if all the application and companies like and became one company and we start to deliver products and services without competition to, you know, all the enterprises, we will still fall short. We will still won't be doing that person of the demand.

我们正处于这场 AI 革命的非常非常早期的阶段。我们将用 AI 做的事情的数量，与 5 年后、10 年后相比，是微不足道的。

We're in very very early stages of this AI revolution. The amount of things, you know, that we're going to be doing with AI is minuscule compared to, you know, where it's going to be 5 years from now, 10 years from now.

所以，例如，我们的许多客户将他们的 Claude Code 或 Claude Cowork 或 Cursor 与 Glean MCP 网关连接起来，从而将你公司的智能和上下文带入这些工具，让它们变得更好。所以，这就是你应该如何看待 Glean 的方式——一个企业 AI 平台和一个同事。

So, many of our customers, for example, connect their Claude Code or Claude Cowork or Cursor with the Glean MCP gateway that brings that intelligence and context of your company to those tools and make them better. So, that's how you should think about Glean, an enterprise AI platform and a co-worker.

所以，我记得，我想即使在 Rubrik 时代，我特别记得你的愿景是为企业打造像 Google 一样的东西。

So, I remember, I think even in Rubrik days, I particularly remember your vision was to build like Google for enterprises.

是的。

Yeah.

而且据我了解，你创立 Glean 时也是带着这个愿景。鉴于当时语言模型逐渐兴起等等，这个愿景发生了怎样的变化，还是保持不变？

And I understand, you started Glean with also that vision. How is that given that during that time, language models came along the way, etc. How has that vision changed or has it remained?

它已经扩展了很多。所以，就像，我们当时都是，我的意思是，对我个人来说最激动人心的第一件事是，为自己打造一个产品。我的意思是，那实际上是驱动力。也是为我认识的每个人。我知道每个人都在寻找信息上挣扎。所以，我们其实很乐意先解决那个用例，即每个员工，每当你需要某些信息时，我们都会让你超级快速地获得。员工所有时间的三分之一都花在试图寻找东西上。所以，我们想解决这个问题。你知道，当时的 AI 有点，你知道，有能力做到这一点。它能够帮助我们理解数据、知识、内容。它能够帮助我们理解人们在问什么问题，并在它们之间做好匹配。但是，那时的 AI 还不能自己写作或进行深度推理或思考。但是，它仍然是一个阶跃变化。就像，你知道，当我们创立 Glean 时，我们使用了 Transformer，我们的搜索已经是这个，你知道，后来 2 或 3 年后行业开始称之为语义搜索或向量搜索。我们在行业创造这些术语之前很久就构建了它们。我们内部过去称之为嵌入搜索。你知道，那是我们为构建的这个概念选择的名字。并且记住，有时，你知道，人们认为 Google 是一种体验，ChatGPT 是另一种。Google 是，你知道，我问一个问题，你给我展示 10 个链接。ChatGPT 是我问一个问题，你给我一个答案。但是，这种二分法实际上是不正确的。Google 几十年来实际上也一直试图，当你来问问题时，给出一个答案。而且它很长一段时间不需要生成式 AI 模型就能做到。所以，我们过去也这样做。当我们创立 Glean 时，很多时候当人们来问问题时，我们会向那个用户展示一个答案。但是，那个答案是从你的语料库中提取的，不是由 AI 生成的。然后随着 AI 变得更好，我们，你知道，显然现在能够开始使用生成能力来真正为人们回答问题。所以，随着时间的推移，你知道，我们的目标和使命保持不变，那就是我们希望人们做出非凡的工作。我们想释放，你知道，帮助他们释放，你知道，他们做出非凡工作的潜力。那是我们的使命宣言。为此，你知道，从帮助他们找到东西，到回答他们的问题，到实际上现在为他们做一些工作，到真正成为企业层面的整体平台，你知道，让公司能够以新的智能水平运营。我们觉得，你知道，这只是我们使命的演变和扩展，你知道，这项伟大的技术让我们能够实现。

It has expanded so much. So, like, we were all, I mean, I think the most exciting for me personally was number one, to build a product for myself. I mean, that was actually the driving factor. And for everybody who I know. And I knew that everybody struggled with finding information. So, we were actually quite happy first just actually solving that use case, which is that every employee, whenever you need some information, we're going to make it super fast for you. 1/3 of all employee time is spent just like finding trying to find things. So, we wanted to solve for that. You know, the AI of the day was kind of like, you know, capable of doing that. It was capable of helping us understand data, knowledge, content. It was capable of us helping understand what questions people are asking and do a good matching between them. But, the AI in that time was not capable of writing on its own or doing deep reasoning or thinking. But, it was still a step change. Like, you know, when we started Glean, we used transformers and our search was already this, you know, what later on like 2 or 3 years later the industry started to call it semantic search or vector search. We built them like well before these terms were coined by the industry. We used to call it embedding search internally. You know, that was the name that we chose for this concept that we had built. And remember that sometimes, you know, people think of Google as one type of experience and ChatGPT as other. Where Google is, you know, that I ask a question and you surface, you know, 10 links to me. ChatGPT is I ask a question and you give me an answer. But, that dichotomy is actually not true. Google for decades have actually also tried to, when you come and ask questions, try to give an answer. And it didn't need generative AI models for a long time to do that. So, we used to do the same. When we started Glean, often times when people came and ask questions, we would surface an answer back to that user. But, that answer was extracted from your corpus, not generated by AI. And then as AI got better, we, you know, obviously were able to now start to use the generative capabilities to in fact answer questions for people. So, over time, like, you know, our sort of goal and mission remains the same, which is we want to make people do extraordinary work. Like, we want to untap, you know, help them untap, you know, their potential to do extraordinary work. That's our mission statement. And for that, like, you know, graduating from helping them find things to answering their questions to actually now doing some work for them to actually becoming that holistic platform at an enterprise level, you know, that allows a company to actually operate with a new level of intelligence. We feel like, you know, it's just an evolution and expansion of our mission which, you know, this great technology has allowed us to actually enable.

我猜有一个关于技术三选一的问题，对吧？比如你们开始深入研究你们内部称为嵌入搜索、后来行业称为语义搜索或向量搜索的东西。然后后来我们看到了，既从你描述的智能体做事的能力来看，也看到了智能体式搜索，调用工具，作为沙箱或循环中的智能体，通常似乎更成功，你能用它做更多事情，或者它在某种程度上使向量搜索产品民主化。你们在 Glean 内部是怎么考虑这个的？你们如何看待这个转变，因为你们在这方面有一个非常成功的业务，然后你们转向了另一件事，也极其成功。

I guess there's a question about like a tech three choice, right? Like you guys started going very deep on what would you call embedding search within the industry later called semantic search or vector search. And then later we sort of see this like both in terms of what you're describing as the ability for agents to do work, but also you're seeing agentic search sort of calling the tools and as an agent in a sandbox or in a loop generally seems to be much more successful and you're able to do much more with it or it sort of democratizes this like vector search product in some way. How did you guys think about that inside Glean and how do you think about that transition because you guys had a very successful business in this you transitioned to this other thing also extremely successfully.

嗯，我的意思是，我认为在内部对我们来说，有时我们因为将公司从 A 转向 B 再转向 C 而获得很多赞誉，但在内部我们认为，我不认为我们得到了比应得的更多的赞誉。我认为我们只是利用了周围正在构建的技术。而且在某种意义上，这是一个非常自然的进展，从搜索作为我们为人们解决的第一个问题，到为他们回答问题，再到真正为他们做事，因为这一切都，你知道，都与 AI 模型和技术的进步相关。

Well, I mean I think internally for us like sometimes we get a lot of credit for pivoting the company from A to B to C and internally we think about that like I don't think we're getting more credit than we deserve. I think like we were simply taking advantage of the technology that was being built around us. And it's been like, you know, in some sense very natural progression of starting with search as the first problem that we solved for people to answering questions for them to actually doing work for them because it's all like, you know, it's all correlated with the advancements in the AI models and technology.

我猜在构建产品时，你总是会面临这种权衡曲线，比如专注于有效的东西，专注于新事物。我认为这就是赞誉的来源，问题是是否存在创新者的困境，或者公司内部是否有推拉，如果这个东西运作得非常好，我们还要去下这个新赌注。这是……

I guess there's always a sort of this like, you know, this trade-off curve that you're making when you're building products about like focusing on what's working, focusing on the new thing. I think that's where the credit is sort of coming out of question about like was there like an innovator's dilemma or sort of a push and pull within the company if this thing is working super well and we're going to go down this new bet. This is the...

我认为那是显而易见的。对我们来说那是显而易见的。

I think that was obvious. That was obvious for us.

好的。

Okay.

就像，你知道，就像世界，然后我认为在消费领域有一个很好的类比。就像你说的，我们实际上不必等待那个。对我们来说也是显而易见的。就像，你知道，如果我有一个问题，我不想真的去读一份长文档来获得答案。我确实更喜欢，你知道，类似 ChatGPT 的体验。所以，就像，我们很清楚我们也必须提供它。就像，你知道，是的，ChatGPT 最初只面向消费者推出。但是，你知道，在企业中，当然，是同样的人在个人生活中使用那个产品。所以，我不认为他们有选择。我们有一个，你知道，问题或嗯，你知道，在那个意义上。就像，你知道，我们基本上必须做这些，你知道，嗯，推进我们的产品。而且这完全合理，人们也很兴奋，我的意思是，这就是我们研发团队的工作方式。就像，你知道，他们总是在寻找，你知道，我们能做的新事物。在公司内部，我们也有一个概念，你知道，我们从 Google 学到的。这个 70/30 的概念，我们团队 70% 专注于短期和中期交付物。

Like, you know, like the world and then I think there's a good parallel in the consumer world. Like you said, and we didn't have to actually wait for that. It was also obvious to us, too. Like, you know, if I have a question I don't want to actually go read a long document to get an answer for that. Like I do prefer, you know, a ChatGPT-like experience. So, it's like and it was clear to us that we also have to deliver it. Like, you know, yes, ChatGPT was initially launched only for consumers. But like, you know, in an enterprise, of course, it's the same people who use that product in their personal lives. So, I don't think they had a choice. We had a, you know, question or um, you know, in that sense. Like, you know, we were basically we had to do these, you know, um like advance our product. And it actually made all the sense and people were excited and I mean, that's how our R&D team works. Like, you know, they are always looking at like, you know, what are the new things you know, we could be doing. Within the company, also, we have a this concept, you know, that we learned from Google. This concept of 70/30 where 70% of our team is focused on short and medium-term deliverables.

比如，你知道，思考当前的产品路线图，如何演进。然后我们还有另外 30% 的人，你知道，在做登月式的赌注，尝试思考我们可以构建的新东西。所以，这就是我们的旅程。所以，我知道对于这种智能体式搜索和行动转型，也许这不是一次转型，而是一个自然的路线图。我知道我们作为 Compose AI 在其中发挥了重要作用。我们想谈谈我们在哪些方面提供了帮助，我们是如何介入的，以及这种合作关系是如何形成的。

Like, you know, thinking about the current product roadmap, how to evolve it. And then we have another 30% who are, you know, working on moonshot bets, trying to think about new things that we could be building. So, that's been the kind of journey for us. So, I know for this sort of agentic search and actions transition, maybe it's not a pivot, it was a natural roadmap. I know we played a huge part as Compose AI in helping in that. We'd love to talk about where we were helpful, where we sort of came into the picture, how that partnership came about.

是的。所以，你看，其中一部分也在于你所服务的用户。所以，当我们开始真正推出这些智能体式能力，当人们可以用 Glean 做越来越高级的分析、深度研究时，我们所有客户自然会问，好吧，把信息带给人们很好。给他们这些工具来分析数据、让他们为你写内容很好，但最终，如果每个人都开始看到其中的潜力，现在 AI 已经准备好真正做一些工作了，比如完成整个循环，不仅仅是做任何给定任务的第一部分，即研究和数据收集与分析，而是现在你实际上可以完成工作，比如你实际上可以把这些工作保存到你的某些系统中。所以，这要求我们开始思考行动。我们实际上也构建了行动，可能大概和你开始构建的时间差不多，但对我们来说，我们在 Glean 的模式一直是，我们想要真正做别人不做的事情。而且我们是一家非常注重合作伙伴的公司。所以，如果你看到其他地方有伟大的创新，并且你有能力真正拥抱这种创新，那么我们总是会做出那个选择。所以，当你想到你们时，就像我们与你们的合作关系，类似于我们与 OpenAI、Google 或 Anthropic 的合作关系。他们构建了伟大的技术，并将其作为平台提供。所以，当他们这样做时，我们想要真正利用它，而不是重新发明。我们与你们合作的动机也是如此。你们深度专注于行动这个特定领域。我们觉得你们会走得快，会真正深入。你记得，行动是一个漫长的旅程，虽然构建一个能快速调用 REST API 并执行行动的系统很容易，但以正确的方式、用正确的参数去做，并且处理一个需要调用四个不同 API 的任务，最终你会遇到很多很多复杂问题，而且很长一段时间内 LLM 都不擅长这个。所以我们知道，对于一家专注于实现这一点的公司来说，他们会做得很好，我们愿意与他们合作。

Yeah. So, look, the part of it is also the users you serve. So, as we started to actually launch these agentic capabilities, as people could do more and more advanced analysis, deep research with Glean, the natural question from all of our customers was that, well, it's great to bring information to people. It's great to actually give them these tools to analyze data, to get them to write some content for you, but ultimately, if everybody starts to see the potential of that, now AI was ready to actually do some work, like complete the cycle, not just do the first part of any given task which is research and data collection and analysis, but now you can actually complete the work, like you can actually save that work in some of your systems. So, that required us to start to think about actions. We actually built actions also, probably roughly the same time that you started to build it, but for us, our model at Glean has always been we want to actually work on things that others are not. And we are very much a partner-focused company. So, if you see great innovation that is happening elsewhere and if you have the ability to actually embrace that innovation, then we'll always make that choice. And so, when you think about you, it's very much like the partnership that we have with you is similar to the partnership we have with OpenAI or Google or Anthropic. There's great tech that they've built and they make it available as a platform. So, when they do, we want to actually leverage it as opposed to having to reinvent that. And the same motivation we had was with you. You were deeply focused on that one specific area of actions. And we felt that you would go fast. You would actually go deep. You remember this, actions have been a long journey and while it's easy to build a quick system that can actually make a REST API call and make an action, like execute a REST API call, but doing it the right way with the right parameters and taking a task which requires you to make four different API calls, ultimately you start running into lots and lots of complications and for a long time LLMs were not good at it. So we knew that for a company that's focused on making that happen, they'll do a good job and we would like to partner with them.

所以，我想，是的，这让我们进入下一个问题，我一直在思考很多。你提到了与 OpenAI 和 Anthropic 这些公司的合作。你如何看待这个领域的竞争？因为当我看到 Glean 的产品时，显然你们做得非常好。产品进展非常顺利。我听到一些推测性的收入数字，我不一定需要你评论，但你们增长非常快，做得非常好。但在 Anthropic 或 OpenAI 这样的公司的产品路线图中，他们曾表示，我认为 OpenAI 在某个时候曾公开表示你们是他们的主要竞争对手之一。但同时，你们能够利用并与他们合作。如果你看看你们在编程领域的同行，比如 Cursor 或 Cognition，他们做得非常好，而他们面临的竞争可能甚至更激烈。那么，作为企业家，你如何看待应对这个问题？你建议我们如何思考构建产品？

So, I guess, yeah, that sort of leads us, thank you for trusting us, betting on us so early. But it leads us into the next question that I've been thinking a lot about. You sort of mentioned partnering with OpenAI and Anthropic, these people. How do you think about competition in this space, right? Because when I see Glean's product, clearly you guys are doing exceptionally well. The product is going really well. I've heard speculative revenue numbers that I don't necessarily want you to comment on but like, you guys are growing really fast, doing really well. But very much in the product roadmap of something like an Anthropic or OpenAI, the way they stated, I think OpenAI at some point had stated directly publicly that you were one of their main competitors. But at the same time, you're able to leverage and partner with them. And if you look at your contemporaries in coding, like a Cursor or Cognition are doing exceptionally well while they have all maybe even more fierce competition. So, how do you as an entrepreneur think about navigating this and how would you suggest we navigate this in terms of thinking about building products?

我认为我们首先需要意识到，我们正处于这场 AI 革命的非常非常早期的阶段。我们现在用 AI 做的事情数量是极小的。我的意思是，这很多，我想在某种程度上，有时 AI 在过去几年取得了很大进展。但与 5 年后、10 年后相比，这实际上是微不足道的。所以，对于所有利用 AI 为客户构建有价值产品的公司来说，这是一个非常非常大的机会。而且我实际上觉得，以一种方式思考是荒谬的，即嘿，一个模型公司或另一个模型公司基本上会让所有其他公司都没有存在的必要。事实上，今天，如果我们所有人都联合起来，如果所有模型公司、所有像我们这样的应用层公司、像你们这样的平台公司，如果我们所有人都联合起来成为一家公司，并且开始在没有竞争的情况下向所有企业提供产品和服务，我们仍然会不足。我们仍然无法满足企业需求的 10%，比如他们想用 AI 做什么。这是每个初创公司都需要真正意识到的第一件事。不要过多关注竞争和其他参与者会做什么。你最大的竞争或最大的挑战将是你自己。你能执行吗？你能构建高质量的产品吗？你能真正谈论它吗？你能把它推向市场吗？你能赢得客户的信任吗？你能为他们做好事吗？这些才是真正的挑战，而不是别人是否也在做你正在做的事情。因为并不是说只有 10 家公司，我们都把产品卖给他们，客户方面实际上是几乎无限的。所以这就是我的看法。所以我实际上并不担心竞争。我从来没有担心过。

I think we first need to have this awareness that we are in very, very early stages of this AI revolution that we are in. The amount of things that we're going to be doing with AI right now is minimal. I mean, it's a lot, I guess in some way, sometimes AI has made a lot of progress over the last few years. But it's actually minuscule compared to where it's going to be 5 years from now, 10 years from now. So there's a very, very big opportunity for all companies that are leveraging AI to build valuable products for customers. And I actually find it kind of ridiculous to think in a way that hey, one model company or another model company will basically make it unnecessary for all other companies to even exist. In fact, today, if all of us came together, if all the model companies, if all the application layer companies like us, platform companies like yours, if all of us came together and became one company and we start to deliver products and services without competition to all enterprises, we will still fall short. Like we still won't meet even 10% of the demand that enterprises have, like what they want to do with AI. That's the first realization that every startup needs to actually have. Don't focus too much on competition and what other players are going to do. Your biggest competition or your biggest challenge is going to be just yourself. And can you execute? Can you build a high-quality product? Can you actually talk about it? Can you take it to the market? Can you earn the trust of your customers? Can you do well for them? Those are the real challenges, not whether somebody else is also doing what you're doing. Because it's not as if there are only 10 companies that we're all selling the products to, the customer side is actually almost infinite. So that's sort of my take. So I don't actually worry about competition. I've never worried about it.

很多人说我们直接对标，就像你提到的，OpenAI 也提过，市场也认可我们是企业 AI 和搜索领域的先驱。每当有大型公司（这很常见）来构建像 Glean 这样的产品时，我们就会得到认可，说他们在构建类似我们的产品。所以确实如此。有很多公司在做我们做的事情。作为创业者，我唯一合理的回应就是专注于我们的客户和产品，睁大眼睛做得更好。如果市场上有其他玩家想竞争，那么首先看看我们能否尝试与他们合作。这就是我们与 OpenAI 等公司合作的方式——他们构建了伟大的技术，作为平台公司提供给像我们这样的公司。所以这就是我们的策略：你不可能与所有人合作，也不能说别人在构建的东西我就不构建。但总的来说，采取一种策略是有意义的，即尝试做那些独特适合你优势的事情。对我们来说，那就是搜索，构建深度知识图谱、上下文图谱，以及理解你的业务运作方式。那个特定领域是我们最擅长的。我们专注于那个；其他人不关注。而围绕这一切的，我们将采取合作伙伴优先的策略。如果市场上有什么我们可以用的，我们就会用。这让我们能够竞争。

A lot of people talk about us being directly, like you mentioned, OpenAI has mentioned it, and the market recognizes us as the pioneers in enterprise AI and search. And whenever another large company, which is quite often, comes and builds a product like Glean now, we get the recognition that they're building a product like ours. So that is there. There are a lot of companies doing things that we do. And the only reasonable response as an entrepreneur that I could have is to focus on our customers and our product, and just do better with our eyes wide open. And if there are other players in the market and they want to compete, well, first let's see if we can try to partner with them. That's what we do with OpenAI and others—they build great technology that they make available to companies like us as a platform company. So that's our strategy: you can't partner with everybody, and you can't say that if someone else is building something, I won't build it. But in general, it makes sense to have a strategy where you try to do things that are uniquely suited to your strengths. For us, that is search, building deep knowledge graphs, context graphs, and understanding how your business works. That particular area is something we are the best in. We focus on that; others don't. And everything surrounding that is where we'll have a partner-first strategy. If we can get something in the market, we'll use it. That allows us to compete.

那么，有一种情况是，很多应用公司凭借数据优势等也在一定程度上进入模型业务，比如 Cursor 有 Composer 等。这也是某种……

So, there is a kind of where a lot of application companies with their data advantage and other things are also coming into the model business as well to a certain extent, like Cursor has Composer, etc. Is that also something...

是的。是的，我认为对 Glean 来说这也很重要。在谈论像我们这样的企业级公司之前，我们每天都会与很多企业交流。他们今天最大的担忧之一是对模型提供商的过度依赖。作为企业，你需要这种独立性——你要确保模型是一种你可以使用的技术。但你的业务运作方式、专有业务流程、独特差异化优势——所有这些你都要完全掌控在自己手中。这个概念是，随着时间的推移，AI 将在企业中做越来越多的工作。但 AI 如何变得更好？这些学习是复合的。谁拥有这些学习？企业需要完全掌控这一点。所以即使在单个客户层面，也有很多关注点在于如何不完全依赖模型提供商。对于像我们这样的应用型 AI 公司来说，也是一样的——我们不能对任何单一提供商的技术栈有巨大依赖。开源模型实际上非常棒。所以 Glean 是多模态的，我们实际上与所有不同的模型合作，包括开源模型。

Yeah. Yeah, I think for Glean that's important too. Before even talking about companies like us for enterprises, we talk to a lot of them every day. One of the biggest concerns they have today is overdependence on model providers. And this independence you need as an enterprise—you want to make sure that the model is a technology you can use. But the context of how your business works, your proprietary business processes, your unique differentiators—all of those things you remain in full control of yourself. This is the concept that over time AI is going to do more and more work in the enterprise. But how is AI getting better? Those learnings are compounding. And who owns those learnings? Enterprises need to be in full control of that. So even at an individual customer level, there's a lot of focus on how to not fully depend on model providers. For applied AI companies like ours, it's the same thing—we cannot have a huge dependency on any one provider for our technology stack. Open source models are actually fantastic. So Glean being multimodal, we actually work with all the different models that are out there, including open source.

我想这很自然地过渡到，如果你要说 Glean 上的 token 有多少百分比来自前沿美国模型，又有多少来自这些中国开源模型？

I guess this sort of transitions very nicely into like what if you had to say what percentage of tokens on Glean are on a frontier American model versus some of these Chinese open source models?

嗯，今天主要是美国模型。但拐点实际上刚刚发生。

Well, today it's dominantly on the American models. But the inflection point actually just recently happened.

那是像 Chimera 3 还是 GLM 52？

Was that like a Chimera 3 or a GLM 52?

是的，我会说 GLM 52 是——我的意思是 Chimera 3 在那之后出现，对吧？但 GLM 52 确实让事情变得非常清楚，现在你实际上可以完全依赖开源模型来处理超过 90% 的企业 AI 任务。

Yeah, I would say GLM 52 was—I mean Chimera 3 came after that, right? But GLM 52 was really when it became super clear that now you can actually fully rely on open source models for over 90% of all enterprise AI tasks.

我想我有两个相关的问题。似乎每 8 到 12 个月就有这样一个循环，开源赶上前沿，然后前沿又拉开差距，然后开始扩散，对吧？比如回到 DeepSeek-Coder-1 或者更早的 Mistral-7B，就像“哦，我们正在接近前沿，前沿差距很小”，然后前沿又推得更远。我想一个相关的问题是，你作为这些语言模型的买家，同时与企业谈论购买和销售这些模型，你如何看待，作为企业或作为 token 的购买者？我能和 Anthropic 签一年合同吗？因为事情变化太快了。所以这有点像两个问题：现在我们正处于开源赶上前沿的时刻。前沿会再次拉开差距吗？

I guess I have two sort of tangential questions. There seems to be this cyclic every 8 to 12 months where open source catches up to the frontier, then frontier pulls ahead and then starts to spread, right? Like if you take back to DeepSeek-Coder-1 or even further back to Mistral-7B, it's like oh we're coming up to the frontier, the frontier delta is small, and frontier pushes much further ahead. And I guess a tangential question to that is how do you, as both a buyer of these language models and talking to enterprises buying and selling these models, how do you think about, as an enterprise or as a purchaser of tokens? Can I sign a one-year contract with Anthropic because the thing is changing so quickly? So it's sort of two questions: right now we're in this moment where open source has caught up to the frontier. Will the frontier again just pull ahead?

是的，是的，我明白了。所以首先关于开源，这不像我们六个月前或十二个月前的情况。我认为这是第一次。比如，我们自己的工程团队——我们一直在评估开源模型，我们团队的决定总是它们还没准备好。他们当然是在与当时的前沿模型比较。结论是它们还没准备好，我们从未费心去使用它们。这现在才第一次改变。所以很多其他公司也在经历同样的事情——他们现在觉得，是的，他们可以用开源做很多工作。所以这是一个变化，而且你会看到，事实上，我自己的信念是，在接下来的 12 到 18 个月里，我们会看到企业中的大部分推理转向开源。当然，你也会有美国模型的开源版本。我的意思是，我认为有些东西你甚至不能——我认为这不是一个选择；我们必须拥有这个，公司正在为此努力。

Yeah, yeah, I got it. So first on the open source, it's not like we're in the same moment as we were six months back or 12 months back. I think this is the first time. And I say, for example, our own engineering team—we've been always evaluating open source models, and our team's decision was always that they were not ready. They were of course comparing with the frontier at the time. And the conclusion was that they were not ready, and we never bothered to use them. That only changed for the first time now. And so a lot of other companies are going through that same thing—they're now feeling that yes, they can do a lot of work with open source. So there is a change, and you will see, in fact, my own belief is that in the next 12 to 18 months, we'll see majority of inferencing shift to open source in enterprises. And of course, you'll have US-based models in open source. I mean, I think some of it is you cannot even—I think it's not an option; we have to have that, and companies are working hard at that.

这里有很多有趣的点。但首先我想回到的是，我们谈了很多关于企业如何分配预算和使用模型。在企业内部，企业 AI 的使用中，哪些是有效的，哪些是无效的？比如你发过一条推文，引用了 All-in 播客里 Chamath 的说法，说 45% 的代币支出每个月或每两个月增长，但生产力只增加了 5%。大家都在问，投资回报率到底有没有？你觉得这些模型在哪些地方渗透得比较广，尤其是除了编程之外，你觉得哪些地方渗透得比较弱，为什么？

So, many interesting points here. But the first thing I wanted to sort of come back to is like, we're talking a lot about how enterprises are thinking to allocate budgets and use models. What's actually working and not working in the usage of enterprise AI inside enterprises, right? Like you had this tweet where you talked about you'd sort of cited the All-in thing where Chamath is talking about how 45% tokens spend is going up every month or every 2 months and there's like 5% extra productivity. There's all these questions about is the ROI there? Is the ROI not there? Where do you think these models are quite diffuse, especially like I'm curious in general, but especially outside coding, where do you think the diffusion is weak and why the diffusion is weak?

目前每个企业面临的头号问题就是成本。大家都以相当激进的方式把 AI 推广给了全体员工。人们都在用 AI 做事。而 AI 是一项非常广泛的技术，你可以用它做各种各样的事情。企业越来越难以衡量 AI 的实际投资回报率。大多数企业根本不知道发生了什么。过去两年他们并不在意，因为很明显每家公司都应该投资 AI。首先，即使只是为了实现员工队伍的现代化，确保他们为 AI 将越来越占主导地位的未来做好准备，这也是完全合理的。但现在成本已经达到了他们无法支撑的水平，除非他们能证明这能带来利润或营收的增长。所以这就是行业的现状。45%、5%，这些数字我不知道，都是编出来的。现实是企业对此没有可见性。解决这个问题有两种方法。一是必须降低这项技术的成本，我认为开源模型将在这方面帮我们很多。事实上，Glean 的使命，现在当我们与客户交谈时，我们首先告诉他们的是，我们理解你们的 AI 成本非常高。我们是企业 AI 平台，是帮你们最小化成本的解决方案。

The number one issue right now with every enterprise is cost. Everybody rolled AI in a pretty aggressive manner to all of their workforce. People are doing things with AI. And AI is such a broad technology you can use it to do all kinds of things. It's becoming very hard for businesses to measure actual return on investment on AI. Most of them have no idea what's happening. And they didn't care for the last 2 years because it was obvious that every company should be investing in AI. First, just for the purposes of even modernizing your workforce, making sure that they are ready for that future where AI is going to be more and more dominant. So it kind of makes all the sense, but it has reached the level where now they don't have the money to actually support this, unless they can actually prove it with bottom line or top line improvements. So that's the state of the industry. 45%, 5%, all of that, I don't know, they're all made up numbers. The reality is that enterprises don't have visibility into it. And there are two ways to solve for that. One is that you have to bring down the cost of this tech, and I think open source models are going to actually help us a lot with that. In fact, Glean's mission right now, when we have conversations with customers, that's the first thing that we tell them, that we understand your cost is so super high with AI. We are the solution for you as an enterprise AI platform that's going to minimize your costs.

我觉得当你说成本是头号担忧时，这隐含的意思是这项技术还不够有用，以至于这个成本……

I think when you're saying cost is the number one concern, that's implicitly saying that the technology is not useful enough that this cost is...

它是有用的。每个人都觉得它有用。

It is useful. Everybody thinks it's useful.

但还不够。

But not enough.

但很难衡量。够不够用并不是人们的抱怨点。难衡量是因为很多都是非常主观的。部分原因只是需要把点连起来。比如，想想你的客服团队。他们用 AI，实际上每天解决的工单可能比以前更多。或者他们解决这些工单的速度比以前快得多。而且你确实有这些指标。客服团队实际上把这两个作为他们的首要指标。很多公司已经看到了明显的改善。例如，美国最大的电信公司之一，他们用 Glean 来运行他们所有的客户服务功能。每次有新工单、新案例进来，Glean 实际上帮助那些客服代表更快地解决这些案例。我说的是去年的数据，远在现在我们能做更多之前。即使去年，我们已经将他们的案例解决时间提高了 48%。所以他们知道这一点，这就是为什么他们愿意为此付费。而且这为他们节省了数千万甚至数亿美元。所以当你用 AI 在部门层面解决问题时，你就能开始衡量了。但很多公司还没有达到那个状态。他们仍然处于第一步，当然，你做的就是给公司里的每个人 AI，让他们用。另一个是编程，当然，在某些方面你可以声称有很多生产力提升，因为你写了更多的代码。

But hard to measure. Whether it's enough or not is not the complaint from people. It's hard to measure because a lot of it is very subjective. And part of it is just connecting the dots. For example, think about your customer support team. They're using AI and they are actually likely resolving more cases every day than before. Or they're actually resolving those cases much faster than before. And you do have metrics for those. Customer service teams actually track these as their top two metrics. And a lot of these companies have been able to see a clear improvement. For example, one of the largest telcos in the US, they use Glean to run all of their customer care function. Every time new tickets come, cases come, Glean actually helps those agents resolve those cases faster. And I'm talking about a stat from last year, well before now we can do so much more. Even last year, we had already improved their time to resolve a case by 48%. So they knew it and that's why they were willing to pay for it. And it generates tens or hundreds of millions of dollars of savings for them. So when you use AI to solve problems at a departmental level, that's when you can start to measure. But a lot of companies haven't reached that state. They're still in the first thing, of course, what you do is just give AI to everybody in the company and let them use it. And the other one, where coding, of course, is where in some ways you can claim a lot of productivity gains because you're writing many more lines of code.

嗯，你的提交数量每天都在增加。解决的 Jira 问题数量可能也在增加。但即便如此，有时人们会觉得，好吧，我们是不是只是把瓶颈从一个地方转移到了另一个地方？你实际上是否在更快地交付产品？这个问题有时对公司来说很难回答。

Um, your number of commits are going up on a daily basis. Number of Jira's resolved probably is also going up. But even there, sometimes people feel like, okay, did we just shift the bottleneck from one place to somewhere else? Are you actually shipping the product faster or not? That is a question that is sometimes hard for companies to resolve.

是的，你可以这样想：token 数量相对于员工人数的比例，以及员工人数如何变化，对吧？因为归根结底，如果……

Yes, one way you can think about this is like percentage of tokens relative to headcount, and then how headcount moves there, right? Because at the end of the day it's like if...

顺便说一句，我不相信那个。我实际上就是不喜欢把这两件事放在同一个句子里。好吧。员工人数和 token 支出或预算。我认为对于 AI 供应商或技术提供商来说，说这是劳动力成本和你能拥有多少 AI 之间的权衡，我认为这是我这三十多年技术生涯中听过的最荒谬的事情。技术支出只是你业务运营中非常非常小的一部分。

I don't believe in that, by the way. I actually just don't like those two things to be in the same sentence. Okay. Headcount and token spend or budgets. I think for AI vendors or technology providers to say that it's a trade-off between labor cost and how much AI you can have, I think it's the most ridiculous thing I've never heard of in my entire over three decades career in technology. Tech spend is a very, very small fraction of how your business runs.

我想问题是，你知道，Salesforce 报告说他们在工程上花费了 300 美元的 token，这大约是工程薪资的 3%。如果你相信 Anthropic 或 OpenAI 所宣传的未来，那么预期是这 3% 可能变成 20%，甚至 150%。所以这不是一个权衡；这来自运营预算，而以前软件几乎没有边际净成本。运行起来相对便宜。而现在它有极高的边际净成本。

I guess the question is, you know, Salesforce reported they spent $300 on tokens they ran on for engineering, and that was like 3% of engineering salaries. And if you are believing the sort of thing that Anthropic or OpenAI is selling about the future, the expectation is this 3% maybe becomes 20%, maybe becomes 150%. So it's not that this is a trade-off; this comes from an opex budget, and before, software had almost no marginal net cost. It's relatively cheap to keep running. Whereas now it has this extreme marginal net cost.

我认为这就是我们目前的情况。

I think that's just what we are doing where it is today.

所以你预期它会再次变成类似无边际净成本的东西。

And so you have expectations that it'll again become something like a no marginal net cost.

我实际上认为做工作的成本将大幅降低。它从来都不是零。就像，你做任何类型的工作，你总是在消耗 CPU 周期来做那项工作。但相对于软件为你提供的效用，它基本上非常低。这就是为什么你会说它没有边际成本。对吧？我认为 AI 也应该如此。为什么它像软件一样？它运行在机器学习模型上。那么是什么让它变得极其昂贵？只是因为今天的硬件成本。所以我认为长期来看，我的猜测，有时没人知道答案，所以这更像是希望而非预测，AI 技术也会变得便宜得多。你可以看到这种情况发生的一个层面是 AI 模型变得更加针对特定任务。一旦你专门化一个模型来做某件事，你今天就已经知道了，我们通过蒸馏、通过后训练来实现这一点。我们已经看到同样的任务，比如今天 Opus 会消耗大量的 GPU 周期来计算。如果你创建一个非常小的模型，只擅长那一件事，你可以达到同样的性能水平，但成本要低两到三个数量级。所以我们知道这是这项技术变得更便宜的一种方式。希望还会有很多其他类似的事情。

I actually think that the cost of doing work is going to significantly reduce. It was never zero. Like, there's always any type of work that you do, you're always burning CPU cycles to do that work. But it is basically so low compared to the utility that software would provide to you. And that's why you would say that it was like there's no marginal cost. Right? And I think AI should be, in my opinion, the same way. Why is it like it is software? It is running on a machine learning model. So what makes it dramatically expensive? It's just because hardware costs today. So I think the long-term, my guess, sometimes nobody knows answers, so it's kind of more hope than prediction, that AI technology is also going to become much cheaper. And you can see one level of how that's going to happen is by AI models becoming more task-specific. Once you specialize a model to do a certain thing, you already know today, and we are achieving it through distillation, through post-training. We're already seeing that the same task, like today Opus is going to consume a lot of GPU cycles to actually compute. If you create a very small model that does only that one thing very well, you can achieve that same level of performance, but at a two or three orders of magnitude less cost. So that's sort of we know that that is one way this tech is going to become cheaper. There are hopefully going to be a lot more other things like that.

所以我认为当企业在决定他们的 AI 预算时，我认为在某种程度上他们仍然在考虑他们愿意为每个人花多少钱。考虑到后台智能体也正在成为现实，它并不完全与公司中的某个人一一对应，这种思考过程是否在改变，或者在考虑年度支出时是否需要改变？

So I think when enterprises are deciding about their AI budget, I think to a certain extent they are still kind of thinking about how much per person they are willing to spend. Is that, given background agents are also becoming a thing, where it's not exactly one-to-one mapping to a person in the company, is that thought process changing or does it need to change while thinking about how much you need to spend in the year?

嗯，我实际上喜欢这种按人建立预算的过程，而且不同部门会有不同的预算。我喜欢这个概念，因为最终企业很难一夜之间改变。想想沃尔玛。沃尔玛不能基本上立即完全改变他们的工作模式、预算规划流程。这些事情非常复杂。所以你必须以那种模式思考。而对于这些新事物，例如后台智能体，可能正确的思考方式是，好吧，我要保留一个自由裁量的预算池，那些是我可以分配给它们的。但我确实喜欢他们已经有一套非常成熟的模式来思考他们的总支出，而且很多支出实际上与工资单成比例。所以这种模式对于大部分 AI 支出来说是合理的，然后对于某些部分，你实际上可以有这个自由裁量的预算。

Well, I actually like this process of establishing budgets on a per person basis, and you're going to have different budgets by different departments. I like that concept because ultimately businesses, it's very hard to change them overnight. Think about Walmart. Walmart cannot basically just immediately completely change their working model, their budget planning process. These things are super complex. And so you do have to think in that mode. And these new kind of things, for example background agents, the right way to probably think about that is that okay, I'm going to keep a discretionary budget pool and those are the ones that I can assign to them. But I do like that they already have very established model of how to actually think about their overall expenditure, and a lot of that actually gets associated in proportion to the payroll. And so that model does make sense for majority of what AI spend, and then for some you can actually have this discretionary budget.

我想……嗯。理解，这很有趣，因为是的，运行一个 CPU 周期有一些边际成本。它太小了，以至于你认为它四舍五入为零。

I guess like... Mhm. Understand, it's very interesting because yes, there is some marginal cost of running a CPU cycle. It's just so small you think about it as it's rounding to zero.

是的。

Yeah.

所以我想如果我对世界模型的理解正确的话，那就是你并不期望 token 支出相对于员工人数变成 20%、50%。它非常非常小。但我们使用更多的 token。这是正确的模型吗？就像实际上我们会比今天花得少得多。

So I guess if I perceive the world model correctly, it's something like you are not expecting that token spend relative to headcount becomes like 20%, 50%. It is like very, very small. But we use way more tokens. Is that sort of the correct model? Like it's actually like we will spend way less than we're spending today.

嗯。

Mhm.

但我们会花得多得多，可能多个数量级……

But we will spend way more like order maybe multiple orders of...

我们都会用 AI 做更多的工作。我希望我们也会停止谈论 token。就像我们从不谈论一个任务消耗多少 CPU 周期。没人关心那个。所以我认为希望这种对 token 的过度关注也会在未来几年内消失。但你的观点是正确的，我们都会用 AI 做更多而不是更少，而且实际上仍然会比今天的成本更低。

We will all do a lot more work with AI. I think we hopefully will also stop talking about tokens. Like we were never talking about the number of CPU cycles that a task takes. Nobody cares about that. And so I think hopefully this over fixation on tokens also is going to go away in the next few years. But your point is correct that we're all going to be doing more not less with AI, and still going to actually cost us less than what it cost us today.

那么我想问的是，公开市场的资本配置是不是错了？因为他们预期收入曲线会这样，比如 Anthropic 年底可能达到 1000 亿美元收入，而且预期还会继续上升。这个模型太棒了，听起来很厉害。你知道，就像“哦，我会做得更多”，然后“是啊，我为什么要考虑 token 或者在乎 token 呢？”

I guess then there's a question about like is the public market capex allocation wrong? Because they're expecting that the revenue curves like Anthropic might hit a hundred billion in revenue at the end of the year. It's expecting to go up. This model is amazing. It sounds like amazing. You know it's like oh I will do more. It's like yeah why should I think about a token or care about a token?

嗯，我觉得投资者比我聪明，所以我不想对他们是错是对发表评论。那是一个非常复杂的问题。但让我们谈谈我能评论的事情，那就是今天，就像我们之前谈到的，AI 的价值，它产生的商业价值，你问大多数企业，他们都会说它远远落后于投资。所以我确实认为行业里有两种叙事。一种是全力押注 AI，非常乐观，认为 AI 将占你运营支出的一半，人们会花掉一半的运营支出。我觉得这种叙事有问题。我知道这有点像，除非你先让企业意识到那个价值并认同、谈论它，否则你不配谈论那些。

Well, I think investors are smarter than me, so I don't want to actually make a comment on whether they got it wrong or right. That's a very complex question. But let's talk about something that I can comment on, and that is that today, as we were talking before, the value of AI, the business value that it generates, you talk to most enterprises and they will say that it's lagging the investment by a very significant margin. So I do think that there are two narratives in the industry. One of them is all in on AI and bullish, that AI is going to be half of your opex, and people are going to be at the half of their opex. And I think that narrative to me is problematic. I know it's kind of like you don't deserve to talk about all of that until you first get businesses to realize that value and agree and talk about it.

所以，我觉得这其实是一个非常有趣的观点，对吧？我们一直在大量讨论这个投资回报曲线的成本端。我想知道，你觉得价值在哪些地方没有成功实现？因为当然在代码方面，它能写很多，也许瓶颈转移到了审查或其他环节。但在其他任务和知识工作类任务中，你觉得如今模型在更有用方面的瓶颈在哪里？因为大多数时候，当你和这些模型对话时，在某个非常具体的领域，它们可能已经比你更聪明了。但在某些方面它们参差不齐，存在瓶颈。你在自己的使用中、Glean 的使用中，还有客户的使用中，发现这些瓶颈在哪里？

So, I guess that's actually a very interesting point, right? We've been talking about the cost side of this return on investment curve a lot. I guess like where do you find that the value is not being as successful? Because of course in code it's able to write a lot, maybe the bottleneck is moved to review or something else. But in sort of like other tasks and like knowledge work tasks, do you find like where are the bottlenecks today in the model just being more useful? Because most of the time when you talk to one of these models, it's like on one very specific domain, they're likely smarter than you already. But it's like in some way they have jagged and they have some bottlenecks. Where do you find those bottlenecks for your own usage, Glean's usage, but also like your customers' usage?

是的。嗯，首先让我们谈谈商业价值。在 Glean，我们非常关注这一点。我们专注于部门级智能体，可以用 AI 自动化的业务流程。当我们与客户合作时，我们总是对他们说，看，你要进行这个合作，你会使用 Glean，你会构建所有这些智能体，但在你知道衡量成功的指标之前，不要做所有那些工作。你怎么知道价值被添加了？所以首先真正就如何衡量该特定任务的效率和生产力达成一致。例如，以法律团队为例。如果你与他们合作，他们的一个指标就是一名合同律师能处理多少份合同。对于支持，我们已经谈过你能处理多少案例。对于试图进行外呼开拓的销售代表，你基本上衡量他们每天能产生多少会议。所以当你开始深入时，对于许多这些职能，你确实有生产力指标。然后你建立基线。然后你真正构建智能体，然后你看到基线的改进。这样你就有了非常严格的衡量价值的方法。所以这最终是企业证明 AI 投资合理性的方式。不幸的是，事情并没有那样发生。AI 只是存在这种恐慌。每个 CEO、每个董事会都说你必须投资 AI。给我看看发生了什么。事实上，我们并没有谈论价值。就在 4 个月前，我们还在谈论 token 最大化。科技行业推动的企业成功标准是，只要你的员工在消耗 token，你就没问题。对吧？所以这就是我们现在的处境。所以问题只是要让这件事继续下去，在企业中建立更成熟、更正式的 AI 项目。那基本上会缩小衡量价值的差距。

Yeah. Well, first let's talk about business value. With Glean, we focus a lot on that. We focus on departmental agents, business processes that we can automate with AI. And as we go and work with our customers, we always talk to them about that look, you're going to do this engagement, you will be using Glean, you'll be building all these agents, but don't do all of that work before you even know the metrics to measure that success. Like how would you know that value was added? So first actually agree on how you measure efficiency and productivity for that particular task. So, for example, take legal team. If you work with them, one of the metrics for them would be how many contracts can one contract lawyer handle on a basis. For support, we already talked about how many cases you can do. For a sales rep trying to do outbound prospecting, you basically measure how many meetings they can actually generate on a daily basis. So when you start to go deep, for many of these functions, you do have productivity metrics. Then you establish the baseline. Then you actually build agents, and then you actually see improvements to that baseline. And that way now you have a very strict way of measuring value. So that's the way to actually ultimately for enterprises to justify that investment in AI. Unfortunately, it just didn't happen like that. AI was just there was this panic. Every CEO, every boardroom said that you got to invest in AI. Show me what's happening. And in fact, we were not talking about value. Just 4 months back we were talking about token maxing. The success criteria for enterprises, as pushed by the tech industry, was that as long as your people are burning tokens, you're in good shape. Right? So that's sort of like where we are. So it's a matter of just getting that going, having a more established, more formal AI programs at enterprises. That will basically close the gap between measuring value.

但我想问的问题更像是，它们在哪些方面非常擅长快速产生价值，在哪些方面仍然非常薄弱？比如，它们在哪里受到阻碍？在哪里它们不如其智能所暗示的那样有用？

But I guess the question I'm sort of asking is more like where are they very good at generating value quickly and where are they still very weak? Like, what are like, where are they hobbled? Where are they not as useful as they could be as their intelligence suggests they are?

哪些职能可能还是……

Which functions probably or...

哪些职能？

Which functions?

嗯，我的意思是，我不知道。这就是我所说的，人们在工程、支持、销售方面看到了相当多的价值，在营销方面也是如此。所以他们看到……我想对你说的是，用正确的方法，你在任何地方都能看到成功。就像在所有职能中。我们看到所有职能都在产生价值。当然，对许多公司来说，研发或销售往往是最大的部门，所以最大的价值实现潜力实际上发生在这些部门或客户关怀中。所以我想说，如果你一年前问我这个问题，我会说真正的影响只在工程和客户支持中感受到。但今天不是这样了。今天你实际上可以在所有不同部门产生价值。这些失误实际上并不是说对于这个特定职能，AI 还没有准备好。实际上 AI 现在已经适用于所有知识工作。更多的是你是否遵循了正确的方法？你是否真正得到了正确的工具？那实际上才是决定成功与否的更大因素。

Well, I mean, I don't know. That's what I was saying that people are seeing value across engineering, support, sales quite a bit, also in marketing. So they're seeing like... My point to you was that with the right approach you are seeing success everywhere. Like in all functions. We're seeing value being generated in all of them. Of course for many companies the R&D or sales tend to be the biggest department, so the most value realization potential actually happens in those two or in customer care. So I would say if you asked me this question one year back, I would say that the real impact is only being felt by engineering and customer support. But today that's not the case. Today you can actually generate value across all the different departments. And these stumbles are actually not so much that for this particular function AI is not there yet. Actually AI is there for all knowledge work now. It's more that did you follow the right approach? Did you actually get the right tools? That actually is a more determinant of success versus not.

而且，好的，我很想谈谈这个，我们稍微谈过，好的，如果你想获取价值，你必须衡量它，并方法论地思考它。但那些工具或你需要做的事情是什么，让你能够使模型对特定职能更有价值？是的。你们在做什么来促成这一点？

And like okay like I'd love to speak to that like we talked a little bit about like okay if you want to capture value you have to measure it and think about it methodologically. But like what are those like tools or things that you need to do that where you are able to sort of make the models more valuable for specific functions? Yeah. Like what are you guys doing that sort of enabling that?

嗯，我认为首先，AI 要在企业中发挥作用，为你的不同业务流程增加价值，你必须能够深入了解你的业务是如何运作的。

Well, I think you need first of all for AI to work in the enterprise to add value for your different business processes, you have to actually be able to go deep into understanding how your business works.

你必须能够接入并连接到你所有的企业系统中。而你们在这里扮演着重要角色，因为没有你们的平台，企业要真正构建出能够完成工作并在企业系统中采取行动的智能体，需要做大量的工作。当然，你需要具备基础技术，包括模型、企业搜索和检索技术，以及行动平台。这些是你必须首先落实的核心基础设施。而现在，这将真正让你产生价值。

You have to be able to go and connect into all of your enterprise systems. And you guys come into play there a lot because without your platform, there's a lot of work that a business has to do to actually build that agent that can complete that work and take actions in your enterprise systems. Of course, you need the basic technology in place, which includes models, enterprise search and retrieval technologies, and an actions platform. These are the core pieces of infrastructure that you have to first put in place. And this is now going to actually allow you to generate value.

我们知道的一件事，我觉得如今人们经常构建这些智能体，并使用 MCP 作为收集上下文的主要方法来构建它们，然后它们失败了，因为这些智能体的表现达不到人类员工的水平。原因很简单：这些智能体就像你第一天入职的员工。是的，你给它们指了一些文档，但它们并不真正拥有那种上下文，以及一个在那里工作了两年的员工所拥有的那种沉浸感。

The one thing that we know, which I feel today often times, people build these agents and they use MCP as basically the primary method to gather context to build those agents, and then they fail because these agents are not performing at the level that their humans were. And the reason for that is simple: these agents are kind of like your employee on day one. Yes, you pointed them to some documentation, but they don't really have that context and that immersion that has happened for a person who's been there doing that work for two years.

因此，我们越来越看到对投资于深度上下文图谱的需求。理解你的主题专家，理解这些专家如今如何完成工作，而这些做事方式并没有被妥善记录。所以这已经成为一个大瓶颈，而这就是我们介入的地方。这弥合了智能体像第一天入职的员工与像资深员工之间的差距。

So we're increasingly seeing this demand for investing in that deep context graph. Understanding your subject matter experts, understanding how those experts today get things done, which is not the way of doing those things is not documented properly. So that has become a big bottleneck, and that's where we come in. And that closes the gap between having agents that work like your day one employee versus your tenured employee.

我觉得这很有道理，我想我就此结束，以免时间过长。感谢你抽出时间接受我们的采访。

I think that makes sense, and I think I'll wrap it so we don't go too long. Thank you for spending time with us.

是的，谢谢。感谢你的邀请。

Yeah, yeah, thank you. Thank you for having me.