搜索的未来：代理网络与 Parallel 的愿景

Parag，非常感谢你今天来参加我们的节目。我们很高兴能邀请到你。对于那些不认识你的人，Parag 曾担任 Twitter 的 CEO，后来把 Twitter 卖给了 Elon。现在他又回到了创始人的轨道上。你创立了一家名为 Parallel Web Systems 的公司，正在为智能体网络扩展智能体式搜索。我们非常激动今天能和你一起聊聊搜索的未来和互联网的未来。那么，我们开始吧。Parallel 是做什么的？

Parag, thank you so much for joining us today. We're delighted to have you on the show. For those who don't know, Parag of Twitter CEO fame was CEO of Twitter before selling it to Elon. And is now back on the founder arc. You founded a company called Parallel Web Systems, which is scaling up Agentic Search for the Agentic Web. We're very excited to have you here today to talk about the future of search and the future of the internet. So, let's get started. What's Parallel?

谢谢你，Sonia，感谢你的邀请。也谢谢 Andrew 加入我们。我对这次对话非常期待。在 Parallel，我们正在构建一系列技术，让智能体能够搜索和使用网络。就像人类一直以来学会使用浏览器、搜索引擎，通过点击和浏览网页来获取信息一样，智能体也需要做同样的事情。我们创立 Parallel 时押注智能体在这方面的使用量会是人类的一千倍。因此，我们需要重新发明能够为智能体提供搜索支持的技术，以及与之配套的商业模式。这就是我们在 Parallel 所做的事情。

Thank you, Sonia, for having me. Thanks, Andrew, for joining us. I'm really excited about this conversation. At Parallel, we're building a bunch of technology in order to allow agents to search and use the web. So just like humans forever have figured out how to use browsers and search engines and clicking and browsing around the web to access information, agents need to do the same things. We started Parallel with the bet that agents would do it a thousand times more than humans ever have. And as a result, we need to reinvent the technology that can power search for agents, but also the business models that go alongside it. And that's what we're doing at Parallel.

好的，我想回到人类搜索和智能体搜索到底有什么不同这个话题。但在那之前，你告诉我们，你在构建 Parallel 的过程中，正在摒弃很多在运营 Twitter 时学到的经验。这是为什么？

Okay, I want to go back into what makes human and agent search so different. But before we get there, you've told us that you're unlearning a lot of the lessons that you learned from running Twitter as you build Parallel. Why is that?

听我说，当我在 Twitter 担任领导职务时，Twitter 是一个已经找到产品市场契合点、规模极其庞大的业务，你的反馈循环来自数亿每天使用产品超过 30 分钟的用户。对吧？在那个世界里，你的运营方式与一个尚未找到产品市场契合点的公司不同，后者基于这样一个前提：几年后会有新客户出现在互联网上，让我们为那些尚未出现的客户构建技术，而我们每天都在、每周都在学习。

Listen, when I was at Twitter in leadership roles, Twitter was a post product market fit extraordinarily scaled business where your feedback loops were from the hundreds of millions of customers using the product for 30 plus minutes every day. Right? In that world, you operate differently than a pre-product market fit company based on the premise that in a few years a new customer is going to show up on the internet and let's go build technology for the not yet here customer that we are all learning every day and every week.

太棒了。好的，我们来聊聊 Parallel。

Fantastic. Okay, let's talk about Parallel.

我其实有个问题。我能插一句吗？

I have a question actually. Can I jump in?

不，不，你不能，Andrew。

No, no, you cannot, Andrew.

我是 Andrew。我没被介绍，但我也很高兴能成为 Training Data 播客的首位嘉宾。

I'm Andrew. I didn't get an introduction, but I'm also happy to be an inaugural guest on the Training Data podcast.

Andrew 在 Training Data 上完成了他的播客首秀。我们很荣幸你终于来了。

Andrew is making his podcast debut on Training Data. We are honored to have you finally here.

谢谢。我觉得我们硅谷的所有人对搜索都有一个非常表面的理解，比如搜索是什么？要做的工作就是，你知道，得到答案，这说得通。但我们都知道有爬取、有索引、有排序。但也许让我们退一步，解释一下网络搜索的问题是什么，无论是对于人类还是智能体，然后我们再深入探讨差异。

Thank you. I feel like all of us in Silicon Valley have a very surface level idea of search like what is search? Like the job to be done is, you know, getting the answer makes sense. But we all know there's crawling and there's an index and there's ranking. But maybe let's take a big step back and explain what is the problem of web search be it for a human or an agent and then we can dive into the differences.

网络搜索的问题，我们都知道并经历过。当我们想找某样东西，却不知道它在网上的哪里时，我们会去 Google 这样的搜索引擎，然后搜索引擎希望能以最方便的方式把答案呈现给我们。这是基本问题。现在，为了做到这一点，搜索引擎在做什么？搜索引擎会去爬取网络，也就是找到所有存在的 URL，尝试读取它们，尝试将这些信息组织成所谓的索引。这样，通过把所有这些东西存储在内存中的一个位置，当查询到来时，你就不必花时间加载页面，因为你已经有了。你已经做了预处理来组织它们，然后当查询真正到来时，你快速理解查询，找到最相关的结果，然后有很多很多，因为你本质上是在从数千亿个页面中缩小范围，找到人们现在正在寻找的五到十个，或者理想情况下一个结果。所以你经历这些多个阶段的检索和排序，以呈现最相关的结果。这就是大问题。一种思考方式是，这是一个十亿对十亿的匹配问题，对吧？所以你有数千亿个页面和随时间累积的数千亿个查询，你需要弄清楚如何在这两者之间进行匹配。

The problem of web search and we all know it and experience it. When we want to find something and we do not know where it is on the web, we go to a search engine like Google and the search engine then hopefully surfaces the answer to us in the most convenient of locations. That's the base problem. Now, in order to do this, what is the search engine doing? The search engine is going and crawling the web, which is finding every URL that's out there, trying to read it, trying to organize that information in what might be called an index. So that by having all of this stuff in memory in one location, you don't have to once the query comes in, you don't have to spend time loading up pages because you already have them. You've already done pre-processing to organize them and then when the query does come in, you quickly understand the query. You find the most relevant results and then there are many many because you're essentially taking hundreds of billions of pages and narrowing it down to what five or 10 or ideally one in terms of what someone is looking for right now. So you go through these multiple stages of retrieval and ranking in order to surface the most relevant result. So that's the broad problem. One way to think about it is it's a billion to billion matching problem, right? So you have hundreds of billions of pages and hundreds of billions of queries over time and you need to figure out how to match make across these two.

这听起来像是一个极其昂贵的基础设施挑战，而且我认为在很长一段时间里，基本上只有 Google 和 Bing 完成了全网络规模的爬取和索引。你为什么认为你作为一家年轻的公司能够去解决这个问题？这似乎是巨头们的问题。

That sounds like an enormously expensive infrastructure challenge and I think for the longest time basically only Google and Bing had done full web-scale crawls and indices. Why did you think about you as a young company could go off and tackle that problem? It seems like a problem of the giants.

所以长期来看，这确实昂贵。但真正有趣的是，当我三年前创办这家公司时，你可以想象这样一个世界：其他人难以竞争的原因之一是无法获得好的反馈，比如这个搜索结果是否比那个更好？要大规模收集这种反馈，存在关于人工评分的问题，还有关于点击数据的问题。你需要访问这些。当然，爬取整个网络并建立索引是昂贵的。但当你考虑到智能体和我们现在可以使用的大模型所带来的可能性，评分数据现在可以由专家以更低的成本创建。我们在 Parallel 的观点是，人类点击数据是一个 bug。使用搜索的智能体应该依赖智能体反馈，而不是人类反馈。我们相信这些模型非常擅长压缩信息，我们可以受益于构建模型的大量研究，并将其应用于搜索索引和排序。所以你现在可以提出很多论点，这就是我们当时提出的论点，实际上现在这个问题变得更加容易处理，因为智能体的存在，不仅作为技术，而且作为一个独特的客户。然后我们找到了一种构建这项业务的方法，不需要我们在服务客户之前预先投入所有基础设施资金。对吧？所以如果你能描绘出一条逐步构建越来越大、越来越复杂的索引的路径，随着你为越来越多的客户解决问题，这个洞察实际上说服了我，这是一个我们可以承担的问题。

So it is expensive in the long term. So what's actually interesting is that when I was starting the company 3 years ago, you could imagine a world where one of the reasons it was difficult for others to compete was not having access to great feedback in terms of is this a better search result than this? To collect that feedback at scale, there was a problem around human ratings. There's a problem around click data. You need to access those. Now it is of course expensive to crawl the entire web and index it. But as you think about what agents and the large models that we now have access to have enabled, the ratings data can be now created by experts way more cheaply. Our view at Parallel is that human click data is a bug. An agent doing work with search should rely on agent feedback, not human feedback. We believe that these models are really good at compressing information and we can benefit from a lot of the research that have gone into building models and apply it to search indexing and ranking. And so you can now make many many arguments and that's the arguments we made back then that actually now it's way more tractable as a problem because of the existence of agents not just as in technology but as a distinct customer and then we figured out a way to go about building this business which did not require us to go spend all of the money on infrastructure upfront before we could service a customer. Right? So if you can paint a path of incrementally being able to build an increasingly larger and more sophisticated index over time as you solve problems for more and more customers that's the insight that actually convinced me that this was a viable problem for us to take on.

那你怎么做到这一点？因为我想这是那种客户希望第一天就有全面覆盖的事情。所以你打算怎么实现？

And how do you make that happen? Because I imagine this is one of those things where customers want full coverage day one. So how do you go about making that happen?

我们最初发布产品时，并没有先发布搜索产品，而是先发布了搜索智能体产品。我们的搜索智能体在收到查询后，基本上可以爬取整个网络。所以如果你在做深度研究，你有一分钟的耐心。我们有些产品有时会花 10 分钟做研究。只要有足够的地图知道优先爬取什么，这段时间足够爬取大量页面了，对吧？所以你可以弥补索引之类的不足，很多时候你可以把它看作延迟优化。如果你放弃那个维度，如果你和人类竞争，这就是为什么我们的搜索智能体是在和外包给人类来整理优质数据这种替代方案竞争。所以我们说，坐在搜索引擎前的人类比一个从零开始的搜索引擎更容易竞争。所以通过构建一个在网页数据之上做实际工作的搜索智能体，我们得以逐步构建我们的索引。

When we first launched the product, right, we launched and we did not launch a search product first. We launched a search agent product first. Our search agent could go essentially crawl the web after a query arrived. So if you're doing deep research, you have patience to the extent of a minute. And we've had products which sometimes take do 10 minutes of research. That's a lot of time to be able to crawl a lot of pages if only you have enough of a map to know what to prioritize crawling, right? So you can make up for shortcomings like index is often times you can think of it as a latency optimization. So if you give up on that dimension, if you're competing with humans, that's why our search agents were competing with the alternative being outsourcing to humans to curate amazing data. Right? So we said it seems like humans sitting on search engines are far way easier to compete with than a search engine on day zero. So by building a product that was a search agent to do real work on top of web data. We were able to incrementally go build our index.

有没有一些例子，人们用你们的搜索智能体构建了什么？

Were there some examples of what people built with your search agents?

在最初的日子里，有人用我们构建搜索智能体，用于某些保险承保流程和理赔处理流程。销售领域的人在做各种销售数据增强。金融领域的人，以前会去构建模型，通过把数据隔夜发送给一组人来收集数据，这些人会整理数据，然后输入到他们做的建模练习中，现在他们开始用我们来替代。所以我们最初的目标是取代那些在网页数据之上的外包人工工作，本质上是为了收集评估数据来运行智能体，用实际用例而不是理论评估来弄清楚搜索智能体应该是什么样子。

In the very early days, there were search agents being built with us for doing some kinds of insurance underwriting workflows and claims processing workflows. People in sales were doing all kinds of sales data enrichment. People in finance, who would previously and historically go to build a model, collect data by sending it overnight to a set of humans who would curate data that would feed into a modeling exercise they would do, would start using us to do that instead. And so we were going initially after replacing where there was outsourced human work on top of web data in order essentially to collect evals to run agents to figure out what search for agents should look like in the first place with empirical use cases instead of theoretical evals.

我明白了，你实际上是在用爬取换取推理时的算力。

I see you're trading off the crawl for inference time compute effectively.

是的，你们在 Neolab 这个词出现之前就成立了，你们有网络系统产品，但你们也有很多来自外部的 AI 人才。在我们第一次见面之前，我不太清楚你们到底是一家网络系统公司还是 AI 公司。我们相处的时间越多，就越清楚，按照某些定义，这就是一个 Neolab。你能谈谈你们所做的研究方面吗，不仅仅是基础设施和系统方面？

Yeah, you guys were started before the term like Neolab came into existence and you have, you know, web systems product but you also have like a bunch of AI people from the outside. Before we met for the first time, it wasn't obvious to me how much of a web systems company or an AI company that you've built. The more we spent time together, the more clear it's like by some definitions this is a neolab. Do you want to talk about where the research side of what you're doing comes in, not just the infrastructure and system side of what you're doing?

所以我们不把自己归类为 Neolab。

So we don't categorize ourselves as a neolab.

嗯，你当然不会。

Well, of course you would not.

不，因为我觉得我们的输出不是模型。也许我的定义有问题。我对 Neolab 的定义是输出是模型。我认为我们的输出是模型的补充。我们构建的东西是在模型之上倍增的，要么你可以说模型变得更好，要么用模型构建的智能体变得更好，拥有超能力。对吧？所以我们一直希望处于这样的位置：每当有人发布更好的模型，比如 Neol，如果别人发布了更好的模型，他们就有更高的山要爬。对我们来说，有人发布更好的模型，他们就解锁了四个更多我们可以发挥价值的用例。至于我们是否需要做可能被定义为研究的工作，那仍然存在。所以从这个意义上说，我们确实必须做研究，但我们也不是在预训练特别大的模型。事实上，我们的工作是把东西压缩成微小的排序模型，对吧？我想如果你回到我关于这个 1000 亿对 1000 亿匹配问题的框架，每个查询本质上是从网络上的一万亿个网页中给我一千个词元，并确保它们是正确的一千个词元。这就是我们实际上从搜索引擎那里得到的查询。所以现在你要做的是，直到上周，我们给自己 3 秒时间在读取时投入大量算力来完成。这周，我们发布了一个产品，现在在 200 毫秒内完成。对吧？所以你现在有那么多时间去弄清楚如何挑选正确的。所以我们要做的是，我们试图在内存层次结构中组织信息，以便快速访问。我们试图弄清楚如何训练什么样的模型来选择最佳的一千个词元。所以有大量的研究投入到如何获取某种预训练模型，调整其架构以适应这个新问题，并在有限的算力和延迟预算内有效地分配算力，以产生最佳输出。

We no, because I think I don't think our output is a model. Like I somehow maybe my definition is broken. My definition of a neolab is an output is a model. I think our output is a complement to a model. What we build is something that multiplies on top of a model in order to give either you can call it the model gets better or the agent built with the model gets better and has superpowers. Right? So we've always want to be in a place where whenever someone ships a better model, a Neol for example, if somebody else ships a better model, they have a higher hill to climb. For us, somebody ships a better model, they have now unlocked four more use cases where we can be valuable. Now whether we need to do work that might be framed as research that remains. So in that sense we do have to do research but we're also not pre-training extraordinarily large models. In fact, our job is to compress things down into tiny ranking models, right? Like I think if you go back to my framing around like this 100 billion to 100 billion matching problem, every query is essentially give me a thousand tokens from a trillion web pages on the web and make sure they're the right thousand tokens. Like that's the query that we are getting effectively from our search engine. And so now what you have to do and until last week we would give ourselves 3 seconds to throw a bunch of compute at read time to do it. This week last week we shipped a product which now does it in 200 milliseconds. Right? So you have now that much time to go figure out how to pick those right. So you will do we're trying to organize information in memory across the memory hierarchy in a way that we can access it fast. We're trying to figure out how to train what model to select the best thousand tokens. And so there's a large amount of research that goes into figuring out how to take a pre-trained model of some kind, adapt its architecture for this new problem and allocate effectively compute in a way that produces the best output with a limited compute/latency budget.

所以很多这些用例看起来像是深度研究类型的用例。当我想到深度研究，至少在早期的版本中，它实际上就像一个智能体循环，模型进行推理，然后基本上只是调用搜索引擎，调用 Google 或 Google 的某个代理。是的。

So a lot of these use cases seem like the deep research kind of shaped use case. And when I think deep research, at least in the earlier incarnations, it was effectively like an agentic loop with the model reasoning and then basically just calling a search engine, calling Google or some proxy to Google. Yeah.

对。

Right.

为什么在你看来那是不够的，而我听到你说的是智能体搜索是一种全新的能力。所以在深度研究中为每个查询调用 Google，它能让你达到某种程度，对吧？

Why is that insufficient in your eyes versus what I'm hearing from you is agentic search is kind of a net new capability. So calling Google for every query in a deep research thing, it gets you somewhere, right?

好的。

Okay.

如果你使用并行搜索，你大部分情况下会使用智能体中不到一半的词元，它会变得更准确，端到端更快。所以你可以把它想象成，但每次如果你只使用一半的词元，如果你的模型受上下文或内存限制，你现在可以解决更多问题，你可以更便宜或更快地解决同样的问题。所以对我来说，当你对信息和相关信息有无穷的胃口时，所有工作的核心都是一个围绕质量、成本和延迟的优化问题。每一个模型进步都是关于如何挤出更多智能，然后如何将其细化，以十分之一的成本保留大部分。对吧？这也适用于搜索。每次你都能在搜索结果中产生更少噪音的所有信号给模型，你就让模型有能力做更多。所以这里很多赌注是在模型层、智能体层一直到搜索层之间进行智能算力分配。

If you use parallel search, you will for the most part use under half the tokens in your agent, it will become more accurate and to be faster end to end. And so you can think of it as but every time if you use only half the tokens if your model is context limited or memory limited you can now do more problems you can do the same problems cheaper or faster. So everything to me when you have a infinite appetite for information and relevant information for all kinds of work it's at its core an optimization problem around quality cost and latency. Every model advancement is about how do you squeeze out more intelligence and then how do you detail it down into keep most of it at a tenth of the cost. Right? And that applies to search as well. And every time you can produce all of the signal with less noise in your search results to give to a model, you now let give the model the ability to do more. And so a lot of the bet here is intelligent compute allocation across the model layer and the agent layer down to the search layer.

另一个有趣的观察是，当你考虑服务智能体与人类时，界面实际上会改变。人类依赖关键词搜索。对于智能体，我们不得不在智能体告诉我们的搜索引擎什么方面进行大量创新。我们知道得越多，我们就能做得越好。

The other interesting thing to observe is the interface actually changes when you think about serving agents versus humans. Humans rely on keyword search. With agents, we've had to innovate quite a lot on the what does an agent tell our search engine. The more we know, the better we can do.

有趣。所以基本上人类只能在记忆中保留几个词，而智能体则不同。

Interesting. So like humans can only hold a few words in memory basically versus an agent.

不，我觉得我们可以。我们只是懒。

No, I think we can. We're just lazy.

对。没错。就像我们能进行对话，能记住很多词，但一旦开始打字，我们只想给 Google 输入两个不完整的词，里面还带个拼写错误，然后指望最好的结果，或者依赖某种下拉自动补全功能来避免打出那三个词。

Yeah. Right. Like we can have a conversation. We can hold a lot of words, but when you start typing, we want to tell Google like two incomplete words with a typo in there and hope for the best or rely on some form of a drop-down autocomplete thing to avoid typing those three words.

对。

Yeah.

首先，我们本质上很懒。结果发现模型不那么懒。所以更少的拼写错误、更明确的查询、可能更长的查询，搜索引擎就不需要去猜智能体想要什么。这样你就能解决另一类问题。关于人类，还有一点是，我觉得我们知道如何解读那些充斥着 Google 搜索结果前列的“AI 前垃圾内容”，你知道，比如“X、Y、Z 的最佳产品”，然后全是联盟广告。但显然它们排名很高。

In the first place. So, we're fundamentally lazy. Turns out like models less. So fewer typos, better specified queries, perhaps longer queries, less for the search engine to guess what the agent might want. So you get to solve a different class of problems. One other thing about humans is I feel like we know how to decipher the pre-AI slop that populates many of the top ranking Google results, you know, for best product for X, Y, and Z, and you get all the affiliate advertising things. And yet obviously they rank highly.

我很容易受其影响。

I'm very susceptible to it.

是吗？是啊。我觉得我有种不可思议的能力，能一眼看穿。智能体怎么处理这类东西呢？这些页面流量很高，看似有很好的答案，但你就是知道它们不真实。

Are you? Yeah. I'm like I feel like I have this incredible ability to see it coming a mile away. How do agents deal with things like that, right? Where these are well-traffic pages. They have seemingly good answers, but you just know they're not real.

也许通过并行处理，智能体需要处理这类问题的程度会少一些。让我先解释这些页面最初存在的原因。我们来看个简单例子。如果你问一家上市公司的最近财务数据，比如只是标题收入数字，我们可以坐在这里知道，SEC 有一份权威文件，里面有这个数字，也许在 PDF 的第 73 页。但那是权威数字。现在想象一下，如果 Google 决定“我在乎权威性”，每当你问那个查询，第一个结果就是那个 PDF，你点击它，加载要 3 秒半，而另一个页面只要 1 秒，你已经不耐烦了，然后你看到一个财报页面。

Perhaps with parallel, agents have to deal a little bit less with that. Let me tell you why those pages exist in the first place. Let's work through a simple example. If you ask for a public company's most recent financials, like just the headline revenue number, we can sit here and know that there exists an authoritative filing with the SEC which has that number, perhaps on page 73 of a PDF. But that is the authoritative number. Right now imagine Google decided like I care about authority and whenever you ask that query that's the first result you see, right? And then you click that, this PDF takes what, like 3 and a half seconds to load instead of the second that the another page does, you're already frustrated, and then you see an earnings page.

那是另一种选择，但你现在得翻到第 77 页才能找到答案。

That's the alternative, but here you have to now grab your way to page 77 to find your answer.

我们很懒，不会去做那个工作。对吧？所以现在网络上存在一整类内容，就是“好吧，很多人需要这个信息，值得把它放在一个加载快、信息在首屏的页面上”。你们都可以去那里。它 99.99% 是对的。所以你去的时候不会那么怀疑它 100% 是错的。而且它在这个过程中增加了真正的价值，因为它从 300 页的财报中识别出 20 条应该放在首屏的信息。

We're lazy. We're not going to do the work. Right? So now there exists an entire class of content on the web which is like okay, this information is needed by a lot of people. It is worth putting it on a page that loads fast where this information is above the fold. All of you can go there. It's 99.99% right. So you're not so skeptical when you go there that this is 100% wrong. And it's added real value in the process because it identified out of the 300-page earnings report 20 bits of information that should be above the fold.

对。

Yeah.

对。所以你可以称之为“垃圾内容”，AI 前的人类垃圾内容，或者你可以称之为迎合懒惰的人类并在 SEO 上取得成功，对吧？好消息是，对于智能体，我们不会让智能体到处点击、摸索、抓取 PDF，对吧？我们从网络上最权威的地方提取一段摘录，并试图把它带到智能体的上下文窗口。所以我们不必被迫陷入这种奇怪的权衡。而这种权衡最初是因为人类，因为标准浏览器和协议等等，我们就是没能想出办法，让我们在网络上任何地方都能精确一致地指向那个被高亮的微小段落。对吧。有了智能体，我们可以把它带到智能体的上下文中，然后让它决定下一步。

Right. And so you can call it slop, pre-AI human slop, or you can call it catering to a lazy human and being successful at SEO, right? The good news is with agents, we're not making the agent click around and fumble around and grab a PDF, right? We're taking an excerpt from the most authoritative place on the web and trying to bring it to the agent's context window. And so we aren't forced into this weird trade-off. And this trade-off existed with humans in the first place because with standard browsers and protocols and everything, we just didn't figure out how to have us point exactly consistently across everywhere on the web to the exact right highlighted tiny paragraph. Right. With agents, we get to bring that to the agent's context and then let it figure out what's next.

你能给我们讲讲，当你的开发者向你的智能体式搜索 API 发送查询时，实际会发生什么吗？

Can you walk us through what actually happens when your developer sends a query off to your agentic search API?

从某种意义上说，这相当标准。我们运行一些模型来理解这个查询是什么，并丰富它，以确定它将如何流入我们系统的其余部分。我们有一堆索引，以不同方式组织网络的不同子集。所以第一层基本上会为这些不同系统各自构造查询。这些系统中的每一个可能是我们的大型索引，一个是我们的新鲜索引，还有一个可能像，有些人会描述为知识图谱，有些人会描述为结构化索引。有很多这样的索引，对吧？所以你现在要决定这个查询需要去哪些索引。你要为每个索引确定查询重写。然后每个索引都有一个大型检索层，然后是排序层，再是更多排序层。所以你要尝试把数百亿的 URL 或文档缩减到数千、数万，再缩减到那些数万中的特定摘录和段落，每个阶段运行越来越大的模型，采用不同架构，提取更多特征，最终得到“这是我想带回给这个 AI 的、信号最强的 1000 个 token”。

It's in some sense pretty standard. So we run some models to figure out what this query is and enrich it to figure out how it will flow into the rest of our system. We have a bunch of indexes which organize different subsets of the web in different ways. And so the first layer will essentially craft queries for each of these different systems. Each one of these systems perhaps is our big index. One of these is perhaps our fresh index. One of these is perhaps like, some people will describe it like a knowledge graph, some people will describe it like a structured index. There are a bunch of these, right? So you're now deciding which ones this query needs to go to. You're figuring out what is the query rewrite for each of these. Then each of these has a big retrieval layer and then a ranking layer and then more ranking layers. So you're going to try to boil down tens, hundreds of billions of URLs or documents down into thousands, tens of thousands down into specific excerpts and paragraphs in those tens of thousands with more and more bigger models running at each stage with different architectures pulling more features to ultimately get down to here are the thousand tokens I want to bring back to this AI, which has the highest signal.

对吧？现在如果你看看我们搜索 API 的各种版本，它们只是针对不同的延迟和成本约束，在这条链的不同点投入不同数量的算力来达到这些限制。所以抽象来看很简单，对吧？有趣的是我描述的这些模型，你如何为它们策划和收集训练数据，对吧？你如何构建这些模型并优化它们。索引本身，你如何使用内存层次结构来存储它，以便达到特定的成本、质量、延迟阈值。

Right? And now if you look at our various versions of our search API, they just throw for different latency and cost constraints, different amounts of compute at various points in this journey to hit those limits. So in the abstract, it's simple, right? What's interesting is what each of these models that I described, how you curate and collect the training data for them, right? How you build those models and optimize them. The index itself, how you use the memory hierarchy to store it to be able to hit certain cost, quality, latency thresholds.

从质量角度看，北极星指标是什么？比如在 Google，有那种“你在前三个答案里得到你想要的结果了吗”之类的。对你们来说，对应的北极星是什么？

What's the north star from a quality perspective? Like you know in Google there's that, you know, did you get the result you wanted in the first three answers or something. What is the equivalent north star for you?

所以我想，我不知道我这样想对不对，但我的看法是，十亿对十亿的匹配问题是一个永远的问题，所以真正的问题是，在什么时候增量优化不值得付出努力，对吧？所以我不认为存在“我们完成了改进”这种事。问题是，在某个时候，改进这个东西会越来越难，就不值得了，对吧？但我希望我们不会到那一步，因为如果你想想我们在 AI 上做的事情，我们会拥有更多智能，而且每几个月就会变得更便宜，结果就是，归结为拥有又好又便宜的模型，拥有出色的信息匹配能力，跨越需求和所有可用的信息，无论是你自己的还是网络上的，并用它做独特而有差异化的事情，产生更多知识，对吧？任何时候你能比别人好 20%，那可能就会给你带来优势。

So I think of, I don't know if I'm right on this, but my take is that a billion-to-billion matching problem is a forever problem, and so the real question is at what point incremental optimization isn't worth the squeeze, right? So I don't think there is a thing as like okay we're done on improving this. The question is at some point it's going to get harder and harder to improve this thing and it just won't be worth it, right? But I'm hoping that we don't get there actually, because if you think of what we're doing with AI, we will have more intelligence that gets cheaper every few months, and as a result of it, it'll come down to having great models which are cheap, having great information matchmaking across a need, and all of the information available to you, whether it's your own or on the web, and doing something unique and differentiated with it to produce more knowledge, right? And anytime you can do something 20% better than somebody else, that might give you an edge.

那为什么不做呢？在超级 AGI 构建的世界观里，如果你能在网络搜索和模型层都提升质量，为什么不做呢？我听到一种观点，说这对模型公司来说太根本了，他们最终会自己掌控，部分原因是他们在收集预训练数据时，那是一个非常昂贵的基础设施工程，基本上就是爬虫的真相来源。你怎么看？

So why wouldn't you, right? So in the super AGI build worldview, it feels like if you can push on quality across web search and in the model layer, why wouldn't you? I've heard a point of view that this is so fundamental to the model companies that they're just going to own it and in part because as they're collecting data for pre-training that's a very very expensive infrastructure exercise that kind of is your source of truth for the crawl. What do you think of that?

我没有看到实际数据支持这个观点，即构建一个全新的网络索引。我认为那种爬虫并不是特别有用，让我解释一下原因。如果你想想我们在构建什么，我们是在构建模型的补充。我们喜欢爬取那些人们不愿意为预训练而爬取的内容，因为如果模型已经训练过这些数据，那就没用了。模型公司在训练时没有耐心以完成主义的方式去等待非常慢的随机 JavaScript 加载，因为你投入的算力与获得的 token 数量相比，差了一个数量级甚至两个数量级，所以对他们来说不值得。但对我们来说，额外的努力是值得的，因为我们是完成主义者，而对他们来说，就是拿几万亿 token。这是第一点。现在我认为这是每个智能体的核心部分，对吧？我的世界观是，如果你购买 LLM 来工作，在 10 个用例中有 9 个，你会希望它们能访问网络，并创建针对智能体优化的搜索基础设施。所以这对所有类型的 LLM 推理来说都是一个真正的邻近领域，这支持了你的观点，即模型公司可能应该拥有最好的基础设施。所以他们可以自己构建，也可以购买，这就是对话，我们会看到谁构建、谁购买、谁合作，以及事情如何演变。

I don't see empirical data on the ground to support that view to build a fresh web index. I don't think that crawl is particularly useful and let me frame why. So if you think of what we are building, we're building a complement to models. We like to crawl things that people don't like to crawl for pre-training because if the model already was trained on it, it's not useful. Model companies for the training aren't patient enough to go in a completionist way try to wait for really slow random JavaScript to load because the number of tokens you get per amount of compute you throw at it is like one order of magnitude so two order of magnitude too bad and so like is it worth the extra effort to get these tokens for us it is because we're completionist right for them it's I'll take x trillion tokens. So that's one. Now I do think this is a core part of every agent, right? I my worldview is that if you're buying LLMs for doing work for nine out of 10 use cases, you will want them to have access to the web and create search infra optimizing for agents. So it is a real adjacency for all kinds of LLM inference and that supports your view that model companies could should have the best-in-class infra for it. So now they can build it or they can buy it and that's the conversation and we will see who builds and who buys and who partners and how things evolve.

你们有没有和哪些模型公司合作，可以透露一下？

Are you partnering with any of the model companies that you can share?

我们想合作，我们会合作，但我不能透露任何具体信息。我们确实宣布了，虽然我不知道你对模型公司的定义是什么，我们今天宣布了与 Google Cloud 合作，成为他们企业智能体 API 的搜索和接地提供商。所以如果你考虑在 GCP 上构建智能体或聊天应用，或进行任何 LLM 推理时，为 Gemini 模型或其他模型提供接地，当你附加网络搜索时，你的选择是 Google 搜索或 Parallel Search，而 Parallel Search 是产品集成的，集成已经优化，我们与技术团队、训练团队、产品团队和商业团队花了时间，确保人们使用 Gemini 模型与 Parallel 结合时能获得卓越的结果。所以是的，这些合作正在出现。我打赌会有好几个这样的合作，它们都会看起来有些独特。

We want to we will I can't share anything on that. We did announce and I don't know how your definitions of model companies we announced today actually that we are working with Google Cloud to be a search and grounding provider for their enterprise agent APIs. So if you think of grounding Gemini models or other models available on GCP when you build agents on GCP or chat apps on GCP or do any other inference with LLMs on GCP when you attach web search to it your options are Google search or parallel search and parallel search is product integrated the integration is optimized we've spent time with technical teams and training teams and product teams and commercial teams to make sure that when people use Gemini models with parallel like they get exceptional and great results. So yes, there are these partnerships now emerging. I bet that there'll be several of these. They will all look somewhat unique.

这是搜索巨头的大动作，恭喜。

That's a big deal from the search king. Congratulations.

Google 是最初的 Neolab。我不知道如何将 Google 定位为模型实验室还是超大规模云服务商，我也不知道从这个案例中能学到什么具体教训，是否适用于其他实验室或超大规模云服务商，所以我们会拭目以待。

Google's the original Neolab. I don't know how to frame Google as a model lab versus a hyperscaler and I don't know what precise lessons to learn from this one whether it applies to other labs or to hyperscalers and so we will see.

我使用语音智能体的方式是，我需要预订晚餐，该去哪个餐厅，然后这个请求由智能体完成。当我们和很多 Parallel 客户交流时，有这种后台智能体，比如监控产品，这些智能体总是在观察世界或网络，当事情发生时，它们就去采取行动。这可能值得探讨，因为如果我们想想一千倍的增长意味着什么，有研究的深度，还有什么是真正发起任务的？是人类发起搜索，还是智能体自己？你想谈谈这个动态吗？

The way that I use voice agents and I need to get reservation for dinner at night what restaurant should I go to and then that request gets fulfilled by the agent. When we talk to a lot of the parallel customers, there's this, you know, background agent, whether the monitor product, these agents that sort of are always watching the world or watching the web and when something happens, then they go off and take actions and do something with it. It might be worth because if we think about what does a thousandx more mean, there's like the depth of research and then there's just like the you know, what is actually initiating the tasks? Is it a human initiating the search or is it agent itself? You want to talk about that dynamic a little bit?

是的，这里有很多维度。让我们先回到搜索智能体，对吧？如果你运行一个典型的搜索智能体，即使不做深度研究，它也会进行 5 到 20 次搜索，即使它在几秒内回答，因为为什么不呢？所以如果你从使用 ChatGPT 的中等水平（所有类别，但不是最高级别，比如即时、低、中等）过渡，每次你在 ChatGPT 中写提示，它会做 5 到 10 次搜索，当你调高时，它会做数百次甚至数千次搜索。所以一个有趣的现象是，人类行动对搜索次数的乘数效应，对吧？仅仅通过使用 AI 应用，你就在某种程度上将搜索次数乘以一个数量级。现在，我们最初对产品和市场的很多看法是追求比这更大的乘数，对吧？所以我们更感兴趣的是，当你说我有一个 10,000 家小企业的投资组合，我每个月都给他们发放信贷，我有一个人工流程来评估风险如何变化，现在我们可以用智能体来做这个，它依赖于大量网络数据。所以这里一个开发者实际上，乘数是数十万或一百万，就发生的网络搜索次数而言，因为人类去编程，而不是现在每月做这个过程，我们可以每周做，对吧？所以当这个智能体每周运行时，它会做大量搜索，为投资组合创建仪表板和一系列需要有人查看的行动项目。对吧？现在你可以转到另一个更有趣的例子。我不知道。你们是否使用某种工具来准备会议文档？那是一个智能体。

Yeah. So there's a bunch of dimensions here. Let's go back to for a moment on search agents, right? If you run a typical search agent and even without doing deep research, it'll do somewhere between 5 to 20 searches even if it answers within a few seconds because why not, right? So already if you transition from using a chat GPT medium or and all the categories but like somewhere not on the high tier like instant low medium every time you write a prompt to it in chat GBT it will do five to 10 searches as you dial it up it'll do hundreds and thousands of searches. So one interesting thing to observe is like a human action to a multiplier on number of searches that happened right so just by using a AI app you're kind of multiplying your way to perhaps one order of magnitude more searches now a lot of our initial takes on the product and the market were to go after bigger multipliers than even that right so we were much more interested when you said I have a portfolio of 10,000 small businesses where I have given out credit to for all of them every month I have this human process that runs to feel out how my risk is going up or down now can we and it relies on a bunch of web data we're trying to use agents for doing this so here a developer is effectively The multiplier there is hundreds of thousands or a million in terms of the number of web searches that happen because a human goes and programs that instead of now doing this process every month, we can do it every week, right? So it's doing a lot of searches when this agent runs every week to create a dashboard on the portfolio and a collection of action items that somebody needs to look at. Right? Now you can go to another example which is even more interesting. I don't know. Do you use something to do meeting prep documents for you all? That's an agent.

有一个 Sequoia 的。有一个 Sequoia 智能体。

There's a Sequoia one. There's a Sequoia agent.

是的，是的。我本来想说 James Flynn，但他是我手下优秀的年轻人之一。所以我使用 Notion 的智能体，你可以构建自定义智能体，它查看我们在 Parallel 拥有的所有内部数据，加上使用 Parallel API 的所有网络数据，来创建会议准备文档。有一次我创建了一个提示来构建这个自定义智能体。现在它为我每次会议进行数十次和数百次网络搜索。每次我为新用例构建新智能体，这个数字就会成倍增加。所以我认为这些后台智能体为我们做越来越多工作的路径，只会受到价值与花费的限制。

Yeah. Yeah. I was going to say James Flynn, but he's one of our great young guys. So I use Notion's agent and you can build custom agents which look into all the internal data that we have at Parallel plus all of the web data using Parallel's APIs to create meeting prep docs. One time I went and created one prompt to build this custom agent. Now it does tens and hundreds of web searches for every meeting I have. Every time I build a new agent for a new use case, that keeps multiplying. So I think the path to these background agents doing more and more and more work all the time for us. It's only going to be bounded by value versus spend.

你知道，我觉得现在这并不理性。我可能在这上面花的钱比应该花的要多，但还不算太多，所以我暂时不在乎。所以所有这些智能体都会有一些合理化调整。但我认为我们会部署后台智能体，只要做这些计算有增量价值，搜索也是如此。所有这些后台智能体都会做大量搜索。

You know, like I don't think it's rational right now. Like I'm probably spending more on it than I should be, but it's not too much. So I don't care yet. So there'll be some rationalization in all of these agents. But I think we will deploy background agents to the extent that there is incremental value in doing that compute and the same thing applies to searches. All of these background agents would do a bunch of searches.

所以我们的第一批产品确实非常专注于这些，部分原因是我们当时在构建和扩展索引，我们决定公司的基础是三个维度：质量、成本和延迟。在最初几年，我们说让我们专注，先忽略延迟，把另外两个做好，因为优化系统和蒸馏到更小的模型，更多是已知的技艺，而不是未知的研究。

And so our first set of products were really obsessively focused on these in part because we were building and growing the index and we decided that our company was based on three dimensions: quality, cost, latency. And for the first couple of years we said let's focus, let's ignore latency and let's just nail the other two, because optimizing systems and distilling to smaller models is much more known art than unknown research.

对吧？所以一旦我们在每个价位都实现了最好的质量搜索和搜索智能体产品，我们就开始着手延迟，这就是我们推出的产品，叫做 Turbo。现在它是市场上最快、最高质量的智能体式网络搜索，遥遥领先。

Right? So once we achieved the best quality search and search agent products at every price point, we started working on latency and that's what we shipped with a product we call Turbo. Now it is the fastest, highest quality agentic web search on the market by a lot.

你觉得现在网络上的智能体式查询比人类查询更多吗？

Do you think there are more agentic queries than human queries on the web now?

我觉得还没有。我觉得我们……

I don't think yet. I don't think we...

考虑到你提到的一些倍数效应，以及后台智能体的存在，似乎是这样。

Just given some of the multipliers you mentioned and given the background agents it seems.

是的。我想我不需要提醒你，但我们还早。我们在智能体采用方面非常早。比如你走出我们的圈子，人们还没听说过“fable”这个词。所以我们非常非常早。我觉得现在像我这样的人可能已经以千倍的速度在操作了。我不知道。你觉得呢？

Yeah. I think I don't have to remind you but we are early. Like we are very early in agent adoption. Like you go step outside of our bubble, people haven't heard the word fable. So we are very, very early. I think there are now people like me who are probably operating at the thousandx. I don't know. What do you think?

三年前，在 ChatGPT 出现之前，你每天做多少次谷歌搜索？

How many Google searches a day did you do 3 years ago before ChatGPT?

是的，我猜大概 20 到 30 次。我觉得今天如果你看看我所有的智能体，我打赌它们做的比那多一千倍。可能 100 到 1000 倍。如果你算上我公司里发生的一些事情，那些不是分配给人类的，可能很容易超过一千倍。但我认为我们是异类，而不是常态。

Yeah, I would have guessed something like 20 to 30. I think today if you just look across all of my agents, like I bet they're doing thousandx more than that. Maybe 100 to a thousand. If you count some of the things that happen at my company, which isn't like assigned to a human, it might easily be more than a thousandx. But I think we are the outliers rather than the norm.

所以我觉得我们在这条路上非常非常早。我确实觉得最近看到，我想是 Cloudflare 说的，在他们对网络流量的监控中，AI 流量在页面读取方面大约和人类流量持平，这跟搜索略有不同，因为它可能包括所有爬虫和其他一些东西。但我认为这会发生。

So I think we're very very early on this journey. I do think I recently saw I think it was Cloudflare that said that in their monitoring of web traffic, the AI traffic is about the same as human traffic in terms of page reads, which is slightly different from searches because it includes perhaps all crawlers that are out there and a bunch of other stuff. But I think it's going to happen.

所以也许这是一个很好的过渡，来谈谈我知道你热衷的话题，那就是我们所知的互联网经济。那里的一些基本假设，你知道，人类眼球、注意力的稀缺，现在似乎就在我们眼前崩塌。互联网经济现在崩溃了吗？接下来会发生什么？

So maybe this is a good segue to talk about a topic that I know you are passionate about, which is the economics of the internet as we know them. Some of the fundamental assumptions there, you know, human eyeballs, scarcity of attention, seem to be falling right in front of us right now. Are the economics of the internet broken now and what's going to happen?

是的。是的。不，我觉得这也许是完全说服我需要亲自做这件事并构建 Parallel 的部分原因。我知道大多数人讨厌广告。我以前做过广告，也构建过广告系统。

Yeah. Yeah. No, I think this was perhaps part of the thing that totally convinced me that this needed to be done by me and build parallel. I know most people hate ads. I used to do ads and build systems for ads.

我喜欢广告。

I love ads.

哦，哇。嗯。

Oh wow. Huh.

我喜欢购物。

I love shopping.

好的。是的。所以你会看到好广告。我不知道我是否喜欢广告，但我在理智上喜欢广告，因为广告是我们所有人都能免费获得这么多精彩内容和科技的原因，对吧？比如没有广告，谷歌搜索就不会免费。没有广告，推特也不会免费。这些是真正有用的技术。网络上有大量我们可以免费访问的内容，如果没有广告，它们就不会免费。所以广告是一种非常高效的变现方案。它非常高效，因为它在差异化定价方面非常出色，对吧？所以大多数查询谷歌是亏钱的。有些查询弥补了这些亏损，这是一个利润率非凡的非凡业务。推特也是如此。就像大多数用户。

Okay. Yeah. So you get good ads. I don't know if I love ads, but I intellectually love ads because ads are the reason that so much amazing content and technology is available for free to all of us, right? Like Google search wouldn't be free without ads. Twitter wouldn't be free without ads. And these are truly useful pieces of technology. A lot of content on the web out there that we can access for free wouldn't be free if not for ads. And so ads is a very efficient monetization scheme. It's a very efficient monetization scheme because it is exceptional at differential pricing, right? So most queries Google loses money on. Some of them make it up and it's an extraordinary business with extraordinary margins. Same with Twitter. Like most users.

正是如此。

Exactly.

不客气。我在补贴你们获得的免费信息。

You're welcome. I'm subsidizing the free information you guys are getting.

你在补贴我们所有人。谢谢你。

You are subsidizing all of us. Thank you.

但我认为广告在网络上的差异化定价和变现方面极其高效，这就是为什么它一直是主导的商业模式。现在，正如你指出的，围绕有限的人类注意力转化为结果的核心假设。如果人类不出现，而他们的智能体出现在网络上，这意味着什么？业务如何运作？是的，我觉得这就是……如果我们不找到新的商业模式，我们已经看到了，对吧？比如人们会说，好吧，我不希望我的内容被智能体访问，因为我有针对人类的商业模式。所以我实际上想花钱请人做 SEO 优化，让更多人类出现。但如果他们的智能体出现，我就会切断它，对吧？这看起来令人困惑和不安。比如好吧，最终这个智能体代表人类行事，但我们还没有找到商业模式的对齐。你无法将那次访问变现。

But I think ads is extraordinarily efficient at differential pricing and monetization on the web, which is why it has been a dominant business model. Now the core assumptions, as you noted, going into it around limited human attention to translate into outcomes. If humans don't show up and their agents show up on the web, what does this mean? How does the business work? And yeah, I think this was the... If we don't figure out a new business model, we're seeing it already, right? Like people are going to say, okay, I don't want my content to be accessed by an agent because I have a business model for humans. So I actually want to go pay someone to SEO optimize myself so more humans show up. But then if their agent shows up, I'm going to cut it off, right? Which seems confusing and disconcerting. Like okay, ultimately this agent is acting on behalf of a human, but we haven't found business model alignment in a way. You can't monetize that visit.

所以假设你从事订阅业务，对吧？所以你有一千个用户，你把其中 20 个转化为月度订阅，对吧？你还不知道这个循环如何与智能体一起运作。你无法区分。你没有统计数据。你不知道这些是否会带来订阅，或者它们只会作为无名智能体不断窃取你的内容。所以围绕旧商业模式崩溃和我们没有找到真正可扩展的新商业模式，存在真正的挑战。

So let's say you are in the business of subscriptions, right? So you get a bunch of... you get a thousand humans and you convert 20 of them into monthly subscriptions, right? You don't know how that loop works yet with agents. You can't distinguish. You don't have statistics. You don't know if these will lead to subscriptions or they'll just keep stealing your content as nameless agents. And so there are real challenges around the old business models breaking and us not figuring out real scalable new business models.

对吧？所以如果你拥有高质量内容，并且目前把它放在网上，你在 Parallel 之前唯一可用的商业模式就是成为头部，能够与实验室以某种固定费用合同交易。这实际上包括一定量的训练和责任，然后是推理时访问。那是你唯一的选择。这个选项对网络上大多数内容不可用。这是一个非常头部的现象。即使对于头部，这也是一个破碎的商业模式，因为当 AI 推理增长，比如说今年 7 倍，明年再 7 倍。在这 50 倍的情况下，他们的交易规模并没有增长 50 倍。比如他们在签署两年合同后，没有人相信他们的份额在续约时不会大幅下降。所以这些就像 AI 推理世界中的固定价格构造。这并不能推动所有这些业务的可持续性。现在我们的解决方案是尝试从我在广告方面的工作中吸取所有教训，比如我把 Twitter 数据卖给了 OpenAI，在那一侧进行过交易，来弄清楚什么可能真正有效并且激励对齐。那么什么可能有效呢？

Right? So if you own high-quality content and you currently put it on the web, your one available business model pre-Parallel was to be in the head and be able to transact with the lab on some fixed fee contract. That's literally which includes some amount of training and liability and then inference time access. That was your one option. That option is not available to most content on the web. It's a very head phenomena. And even for the head, it is a broken business model because when AI inference grows, let's say 7x this year and another 7x the next year. On this 50x, their deal size is not growing 50x. Like none of them after signing a 2-year deal believes that their share isn't going to decline materially at renewal. And so these are like fixed price constructs in the world of AI inference. And so which doesn't drive for sustainability for all of these businesses. Now our solution is trying to learn all the lessons from my work at ads, on having transacted like I sold Twitter data to OpenAI, having transacted on that side, to figure out what actually might work and be incentive aligned. So what might work?

高效的差异化定价。为极高价值的内容支付差异化的高额费用。为访问相同内容的高价值工作支付高额费用。所以它在两个维度上都是差异化的：质量和使用它完成的工作的价值。并且要以可扩展的方式做到这一点，而不仅仅是定制交易。

Efficient differential pricing. Paying for differentially a lot for extraordinarily high value content. Paying a lot for high value work accessing the same content. So it's differential on both dimensions: quality and value of work being done with it. And a way of doing this scalably and not just with bespoke deals.

所以这些是任何合理解决方案所需的特性，而其中最重要的是激励对齐，对吧？比如，人们什么时候愿意合作参与这个事业？

So those are the properties needed for any reasonable solution, and the biggest property of all is incentive alignment, right? Like, at what point do people want to collaborate into this enterprise?

为什么模型公司需要支付任何费用？

Why do the model companies need to pay anything at all?

嗯，你应该问他们。我的理解是，第一，你需要在推理时获得新鲜数据，以便在像 ChatGPT 这样的产品中展示；第二，你需要训练数据；第三，你需要为已经进行的训练获得一些责任保护。所以支付是这三者的某种组合。我不能确定他们如何评估这三者各自的价值。

Um, you should ask them. My understanding is one, you want fresh data during inference time to be able to display it in products like ChatGPT. Two, you want training data, and three, you want some liability protection for training that you already did. And so the payments are some combination of these three things. And I can't be sure of how they value each of these three.

你说过，即使一个非常简单的查询也会进行 10 次搜索。那么，对于最终归结为一段回复的所有不同来源，你如何进行归因？

You said, you know, even a very simple query will go do 10 searches. And so how do you do attribution between all the different sources that, you know, boil down to one paragraph response?

是的，在 Parallel，我们喜欢构建模型。

Yeah, at Parallel we like to build models.

我听说了。

I heard.

所以，呃，不，我认为这回到了我关于激励对齐的观点。所以，在我们尝试构建模型之前，让我们先试着从理论上、智力上搞清楚，我们该怎么做。所以你必须问这个问题：某人的内容增加了多少增量价值，对吧？所以你可以运行所有这些模拟练习：你从语料库中取出一段内容，然后你说，让我们运行智能体，看看我们输出的质量是否下降。要挽回多少质量？也许如果我投入更多算力或更好的模型，我能否挽回质量？哦，我能。这花了我一分钱。嗯？我可以通过没有这个来源而失去的质量来获得它。我的替代方案是投入更多算力，价值一美分的算力，来获得同样的质量。结果可能不同，但就最终客户价值而言，质量是一样的。所以你会说，好吧，这个来源价值接近一分钱。感觉对，对吧，这很直观。现在，这种直觉的形式化是我们使用的核心框架，叫做 Shapley 值。

So, uh, no, I think it goes back to my point around incentive alignment. So, let's before we try to build a model, let's try to figure out, like, if you were going to try to do this intellectually, theoretically, how do we go do it. So you have to ask the question of, okay, how much incremental value did somebody's content add, right? So you can run all of these simulation exercises: you take one piece of content out of the corpus, and then you say, let's run the agent, let's see if the quality of our outputs declined. How much to claw back that quality? Perhaps if I threw a little bit more compute or a better model in some way, could I claw back that quality? Oh, I could. It cost me a cent. Huh? I could get the quality I lost by not having this source. My alternative was to throw more compute, a cent's worth of compute, to get that same quality. It might be a different result, but as far as the end customer value is concerned, it's the same quality. And so you're like, okay, this source is worth close to a cent. Feels like, right, that's intuitive. Now the formalization of this kind of an intuition is the core framework we use; it's called Shapley values.

是的，继续。什么是 Shapley 值？

Yes, let's go. What's a Shapley value?

这是一个——

It's a—

正合我意。

Music to my ears.

这是一个博弈论——你是博弈论专家。

It's a game theoretic—you're a game theorist.

一堆博弈论。是的。

A bunch of game theory. Yeah.

是的。Shapley 值是一个非常理论化的数学概念。对于那些做机器学习的人来说，它用于特征重要性中的 Shapley 值。假设我们三个人合作做某件事，整体大于部分之和，呃，在那一刻。理论问题是，好吧，我如何分配我们通过合作创造的更大的蛋糕？所以我们三个人都有合作动机，对吧？而 Shapley 值是一种有效回答这个问题的数学方法，对吧？现在这听起来很棒，对吧？比如，如果你是联合创始人，你可以弄清楚如何分配股权。不幸的是，它没那么有用。它没那么有用，因为为了计算 Shapley 值，你需要模拟所有世界，其中我们两个人中的某个子集合作了，但第三个人没有，并演绎这些现实，然后回溯到今天我们应该如何分配蛋糕，这在实践中大多数地方是做不到的，对吧？在机器学习模型中，当你做特征重要性时，你某种程度上可以：你可以保留一个特征，运行你的模型，看看效果如何。在网络搜索中，我们可以模拟如果我们无法访问这个 URL 或这个域名或它们的集合，智能体会表现如何？如果你擅长评估，如果你擅长评估质量，你可以通过运行各种场景来构建这些数据，收集一堆这样的数据，然后你可以训练模型来——

Yes. Shapley values is this very theoretical mathematical concept. It is used in Shapley values in feature importances for those who are ML folks here. Let's say the three of us collaborate on something, and the whole is bigger than the sum of parts, uh, in that moment. The theoretical question is, okay, how do I divide up this sort of bigger pie that we created by collaboration? So all three of us have incentive to collaborate, right? And Shapley values is a mathematical way of effectively answering this question, right? Now that sounds amazing, right? Like, you could, if you're co-founders, you could figure out how to divide equity. Unfortunately, it's not that useful. It's not that useful because in order to compute Shapley value, you need to simulate all worlds where some subset of the two of us collaborated but the third one doesn't, and play out those realities to then impute back to today in terms of how we should divide the pie, which in practice you can't do most places, right? In ML models, when you do feature importances, you kind of can: you can hold a feature back, run your model, and see how well it did. In web search, we can run simulations of if we did not have access to this URL or this domain or this collection of them, how would the agent perform? We can, if you're good at evals, if you're good at assessing quality, you can build that data by running various scenarios, collect a bunch of this data, and then you can train models to—

你最喜欢的东西。

Your favorite thing.

最喜欢的东西。所以 Shapley 值的挑战在于，在我们的情境中计算 Shapley 值比我们花在智能体上的美元要昂贵得多。更不用说我们想付给出版商的金额了，对吧？所以为了计算内容所有者得到一美元，如果我决定做完整的 Shapley 值计算，那可能需要几美元，对吧？所以这没有意义，除非你能通过构建合适的数据和模型来很好地估计它们。但我们相信我们的估计是好的，而且这在理论上是合理的，如果存在完美的信息对称，人们会愿意合作。所以就像过去广告时代，人们使用第二价格拍卖，并相信，无论好坏，人们会在广告拍卖中透露真实出价，最终支付更少。一旦反馈循环在市场建立，就像今天一切都是自动出价，对吧？大多数人在做广告时衡量投资回报率，并使用自动出价而不是手动出价。我认为基于激励对齐的坚实基础，由 ChatGPT 驱动，最终将最大化内容所有者以及通过 AI 的内容寻求者在这个优化系统中的参与。我们有一些积极的证据支持这一点。比如我们有一些有趣的合作，我们能够完成并宣布。是的，想象一下与内容所有者坐在一起，解释 Shapley 数学需要一点时间。但最终，一旦你拿出这些特性，对吧，你参与价值分配，如果你有独特差异化的数据，你会得到更多报酬。如果一个高薪工作的银行家阅读你的数据，而我的退休父亲阅读你的数据，银行家最终会为那次阅读支付更多，对吧？因为它是高价值工作的一部分。而且宏观数学似乎也成立：如果你要在知识工作的 LLM 推理上花费大量资金，如果我们将其中的 2% 到 10% 分配给网络数据，那将比今天除 Facebook 和 LinkedIn 等围墙花园之外的所有网络数据商业模式都要大得多。所以宏观数学支持这一点。这是一种可扩展的方法，随着网络上的智能体每年增长一个数量级，根据我的计算，我们距离这种数学能够为网络上广泛的内容所有者提供有意义的美元还有 12 到 24 个月。

Favorite thing. So the challenge with Shapley values is computing Shapley values in our context is way more expensive than the amount of dollars we spend on the agent. Forget the amount of dollars we want to pay a publisher, right? So to compute that a content owner gets a dollar, if I decide to do the full Shapley value computation, that might take several dollars, right? So it doesn't make any sense, except you can really well estimate them if you build the right kind of data and models around it. But we have confidence that our estimations are good and that this is sound theoretically, that if there was perfect information symmetry, people would want to collaborate. So the same way in the old ads days, people did second price auctions and believed, for better or worse, that people would reveal their true bids in an ad auction and end up paying less than that. Once feedback loops get established in a market, like today everything is like an auto bid, right? Like most people are measuring ROI when they're doing advertising and running on auto bid instead of making bids. I think a solid foundation based on incentive alignment that ChatGPT drives will ultimately maximize participation of content owners into this, as well as content seekers via AI, in an optimized system. And we have some positive evidence to support it. Like we've had some interesting partnerships that we've been able to do and announce. And yeah, like imagine sitting with content owners and explaining Shapley math takes a moment. But ultimately, once you pull out the properties, right, that you participate in the value, if you have unique differentiated data, you get paid more. If a banker in an expensive job reads your data versus my retired dad reads your data, the banker ends up paying more for that read, right? Because it's a part of high value work. And the macro math also seems to work: if you're going to spend a lot of money on inference on LLMs for knowledge work, if we allocate it like 2 to 10% of it to data on the web, that's way bigger than all web data business models today outside of like walled gardens like Facebook and LinkedIn. So the macro math supports it. It is a scalable approach, and as agents on the web grow an order of magnitude year on year, it's like, by my calculations, we're 12 to 24 months from this math being able to give meaningful dollars for a very wide range of content owners on the web.

这非常令人兴奋，尤其是在你看到流量数据、Stack Overflow 暴跌的时候。你看到我们所知的人类互联网因为激励而大量消失。看到你如何思考激励对齐以让人们继续发布内容，这非常令人兴奋。

That's very exciting, especially at a time when, you know, you see the traffic data, the Stack Overflow plummeting. You see a lot of the human internet as we know it going away because of incentives. It's very exciting to see how you're thinking about incentive alignment for people to keep publishing.

是的，这就是我们创办公司的原因。

Yeah, it's why we started the company.

为什么公司叫 Parallel？

Why is a company called Parallel?

公司最初的名字是 Shapley Inc.

The company's original name was Shapley Inc.

真的吗？

Really?

那是当——

It was when—

你知道吗？

Did you know this?

我知道。是的。

I knew that. Yeah.

当我注册公司时，我告诉你，我对问题空间的一切都着迷。

When I incorporated, I'm telling you, I was obsessing about everything to do with the problem space.

所以，尽管知道“also one sharing”这个最初的名字对 B2B 产品来说很糟糕，它不会成为长期名称。所以我们以 Shappley Inc. 的名义注册，因为一时找不到更好的词，而 Shappley.ai 恰好可用。Shappley.com 是一个被占用的域名。所以它不会成为正式名称。我到处都不把自己的公司称为 Shappley，在所有活动上我的胸牌都写着 NCO 或“隐形公司”。所以我们花了差不多六到八个月才确定公司的真正名称。最终我们定为 Parallel，部分原因是一，当时我们做的事情很多都是并行的；二，我们开始设想一个为 AI 构建的并行网络，以及它的特性有何不同。这个比喻是，当你发布内容时，你现在要考虑两个受众：好的，我要创建一个页面，我知道人类会读它，它应该看起来像什么样？然后我如何确保智能体也能读懂它？所以感觉就像你在向两个受众双重发布。因此我们有了这个想法，一个为智能体而生的并行网络将会出现。这就是我们开始喜欢 Parallel 的原因。

So while knowing that the first set of also one sharing is a terrible name for a B2B product, it was not going to be the long-term name. So we incorporated as Shappley Inc. for lack of a better word, and Shappley.ai happened to be available. Shappley.com is a parked domain. So it was not going to be a name. I went around not talking about my company as Shappley, and my badges at all events used to call it NCO or stealth company. So it took us almost six to eight months to figure out what the real name of the company would be. We ended up at Parallel in part because one, at the time we were doing a lot more in parallel, and two, we started visualizing this sort of a parallel web built for AIs and how its properties are different. And this metaphor that when you publish, you're now thinking of, as we all are now, of two audiences: okay, I'm going to create a page, I know humans will read it, what should it look like to them? And then how should I make sure that agents can read it too? So it feels like you're dual publishing to two audiences. And so we had this idea that a parallel web for agents will emerge. And that's why we started liking Parallel.

太棒了。

Love it.

我最近其实在想这件事，比如财报电话会议记录，对吧？我非常确信，通过智能体来消费财报记录的人，比真正听音频的人多，说实话可能比读原文的人还多。所以网络上这些高价值的部分已经转向智能体优先了，显然沟通方式还没有转变。但我觉得如果今天我是上市公司 CEO，要做财报发布，我会非常明确地表示，我说的话会被智能体正确转录和解读，而不仅仅是给听众听。

I actually was thinking about this recently with things like earnings transcripts, right? Like I feel very confident that more people are consuming earnings transcripts through agents than are actually listening to the audio, and honestly probably even reading the transcript itself. So there are these very high value portions of the web that have flipped already to being agent first, and obviously the communication has not yet flipped. But I think if I were a public company CEO today and I was doing an earnings report, I would make very clear that the thing that I'm saying will be transcribed and interpreted correctly by the agents, not just by the people listening.

理论上，是的，对我们来说也一样，当我们发布 API 文档时。就像，我们的客户在构建 AI 解决方案，他们用 AI 来做这件事。是他们的智能体在阅读我们的文档和 SDK 中的代码，而不是人类在文档页面里摸索。所以实际上对我们来说，主要受众是智能体，我们就是这样测试文档的。

Theoretically, yeah, same for us when we're publishing docs for our APIs. Like, it's like our customers are building AI solutions. They are using AI to do it. It's their agents reading our docs and code in our SDKs. It's not humans fumbling around docs pages for the most part. So in fact for us the primary audience is an agent, and that's how we test our docs.

为智能体而生的并行网络，非常酷。

A parallel web for agents. It's very cool.

也许最后收个尾。告诉我们，给我们一个 Parallel 现状的快照，如果你希望构建的一切都成真，Parallel 会是什么样？世界会是什么样？你在其中扮演什么角色？

Maybe close us out. Tell us, just give us a snapshot of where Parallel is today, and if everything that you hope to build comes true, what does Parallel look like? What does the world look like? What's your role in it?

我把这段旅程分为三个层次。第一层是人们在构建简单的智能体，把网络更多地当作工具来用，比如网络搜索工具，因为很熟悉，因为我们构建的第一批智能体，正如你所说，就是模型加上同样的工具。因为这些模型已经训练过使用人类习惯使用的所有工具，然后让它们去工作。总的来说，如果你想想大多数工作，今天已经存在了，或者推理，今天已经存在了。我们现在开始看到一部分客户，他们正在构建更复杂的多智能体系统，使用子智能体，让智能体互相唤醒或以有趣的方式编排。人们把子智能体看作最熟悉的一种，最常见的是带有编码智能体框架的子智能体，但我们看到更多类似的情况。比如如果你在构建一个 AI 科学家，有些系统非常有趣、非常复杂、运行时间很长，只是把大量的算力和数据投入到真正困难的问题上。我认为第三层，特别是对于网络来说，是网络从拉取转向推送。所以今天，总的来说，在我描述的两种模式中，要么智能体调用工具或子智能体，但它是在发出请求，说“现在去帮我找这个”。我认为几年后我们会达到的状态是，各种用例将是“并行网络，如果发生这种情况就叫我”，这样我的智能体就能做一些工作，或者人类能做工作。所以我经常对团队说这句话，就像，想象一下智能体无处不在，它们能做很多事情。如果今天有你能做的事情，而且值得做，我们还有可用的 GPU，我们就会直接去做。我们不会说“哦，为了做而做，明天再做吧”。如果今天能完成，就今天做。那我们明天做什么？我们将做明天的工作，以响应某种循环的东西，要么是另一个智能体的工作，要么是世界的变化，比如卫星图像可见的变化，或者发生的客户评论。某个智能体完成了一些计算，人类有了新的见解来触发工作。但会有一些这样的信息流，它们将驱动明天新的智能体工作。其中一个信息流将是网络上所有变化的东西，这对我来说非常兴奋，因为这样你就不只是在定义一个时间点的需求，而是长期来看什么对我可行。如果发生类似的事情，正如网络上不断变化的所有信息所证明的，就叫我，对吧？然后我会让我的智能体去处理它。这样我们就能代表所有客户，随时将算力分配到整个网络上，这真的很令人兴奋。

I think of three levels in this journey. Level one is people are building simple agents that use the web more like a tool, like a web search tool, because it's familiar, because the first set of agents we built, to your point, was just like model and give it the same tools. Because these models have been trained to use all the tools humans have been used to, and let them do work. By and large, if you think of most work, it's there today, or inference, it's there today. We have now started seeing some subset of customers who are in this world where they're building more sophisticated multi-agent systems which use sub-agents and have agents wake each other up or orchestrate in interesting ways. People are seeing that as sub-agents, the most familiar one is sub-agents with encoding agent harnesses, but we see a lot more of that. Like if you're building an AI scientist, some of those systems are very interesting, very sophisticated, very long-running, and just throw large amounts of compute and data at a really hard problem. And I think the third layer, for the web specifically, is the web goes from pull to push. So today, by and large, across the first two modalities I described, either an agent is calling a tool or a sub-agent, but it is telling it a request, saying go find this for me right now. And I think where we will end up in a couple of years is a variety of use cases will be the web parallel, call me if this happens, so that my agent can do some work or a human can do some work. So I use this line with my team all the time, which is like, imagine agents are everywhere and they can do a lot of things. If there is something you can do today and it's worth doing, and we still have GPUs available, we'll just go do it. We won't be like, oh let's just do this tomorrow for the sake of it. If it can be done today, do it. So what will we do tomorrow? We will do tomorrow work in response to something that circles through either another agent's work or something changes in the world as visible in satellite imagery or some customer commentary that happens. Some agent finishing some compute, a human having a new insight to trigger work. But there are going to be a few feeds like this which will drive new agentic work tomorrow. And one of those feeds is going to be everything that changed on the web, which is really exciting for me because then you're framing not a point-in-time need, but a long-term here is what is actionable for me. If something like this happens, as evidenced by all of the information on the web that is changing all the time, call me, right? And then I'll run my agent on it. And so we get to then allocate compute onto the entire web all the time on behalf of all the customers, and that's really exciting.

太棒了。非常有雄心的愿景。你显然对此充满热情，看到你在构建这些真的很令人兴奋。非常感谢你今天加入我们，Parag。也感谢你加入我们，Andrew。

That's awesome. Really ambitious vision. You're clearly extremely passionate about this, and exciting to see you building. Thank you so much for joining us today, Parag. And thank you for joining us, Andrew.

不客气，Sonia。谢谢。

You're very welcome, Sonia. Thanks.

谢谢你，Andrew，这是你的首次亮相。

Thank you, Andrew, for your debut.