AI 智能体或将推动网络活动增加 10,000 倍 | Twitter 前 CEO
- 如果他最初提出的1,000倍判断成立，Agrawal认为Agent使用网页的规模还有5,000–10,000倍增长空间。他估计目前Agent的使用量相当于人类的10–20%，并将其与爬虫流量和媒体报道中的机器人流量区分开来。他自己的Agent每天已经进行数千次搜索：“1,000倍还不够。”
- Parallel以Agent推理需求作为增长基准，Agrawal目前按10倍作规划估算。他预计，全球每投入10块GPU用于模型推理，就会有1–2块GPU用于网页搜索。算力供给取决于具体硬件、地点和交付时点；部分GPU必须部署在距离Parallel网页索引5毫秒以内。
- Parallel主打搜索质量、模型无关性和更低的推理成本。节目开场介绍称，公司上个月完成1亿美元B轮融资，估值达20亿美元；Agrawal表示，公司员工人数从年初接近30人增至约120人。他称，高信噪比搜索结果能让模型“少搜一半”，减少Token消耗并更快完成任务。
- 消费级Agent能否留住用户，取决于是否兑现足够简单的产品承诺。Agrawal看好留存，因为用户只需给Agent发消息，几乎不需要学习成本；但他承认自己没有内部留存数据。他不认同OpenClaw的理念已经碰壁：配置和持续维护门槛较高，早期模型能力也有限，限制了产品的普及。
- Agrawal认为，Agent浏览网页会打破注意力经济，即使购买行为仍会保留。“注意力已经无法变现”，因此他认为广告和按席位订阅都难以持久。Parallel正在搭建向内容所有者支付Agent阅读费用的系统，并由模型决定补偿金额，试图为这种新型网络打造一个“AdSense”。
- Agent访问网页，既需要商业利益协调，也需要身份识别和安全保障。Agrawal希望实用的Agent能够表明身份，并继承所有者的权限，但目前表明身份反而容易招致封锁。安全仍需分层负责；他预计攻击者和防守方都会投入推理算力，造成影响重大的网络安全事件将“每隔几周”发生一次。
- Agrawal对下一阶段需求的判断是：“网页将从拉取转向推送。”如今常驻运行的Agent大多仍是被动响应；他预计Agent会监测变化，并响应用户明确或隐含的需求。在一个拥有万亿页面、机器生成噪声不断增加的网页上，瓶颈将变成如何分配算力，识别值得采取行动的变化。
1. 尽管流量数据醒目，Agent使用网页仍处早期
- Agrawal最初的判断是，Agent使用网页的规模将达到人类历史最高水平的1,000倍。如今他说，1,000倍还不够（“1000x is too low”）：普通人每天使用Google搜索约10次或更少，而他追踪的Agent每天会搜索数千次。他粗略估算，自己的使用量略低于1,000倍。大多数人还远未达到这一水平；要走到那一步，可能还需要几年。
- O’Shea提到，Vercel观察到编码Agent的流量已达到60%。Agrawal提醒，自动化流量还包括预训练爬虫，而且Agent完成同等任务时的搜索次数多于人类。他估计，Agent目前的网页使用量约为人类的10–20%。如果他最初提出的1,000倍判断成立，那么从现在起还要增长5,000–10,000倍。
- O’Shea追问：Agent变得更聪明后，会不会减少搜索次数？Agrawal认同这一点，但不认为可以把旧有搜索基础设施扩容1,000倍。他估计Google每天大约会进行100亿次搜索，也不认为Google能建成支撑千倍搜索量的基础设施。Parallel则用GPU支撑搜索索引；这类系统过去受CPU和内存限制，公司还把排序能力提炼进更小的模型。
- Agrawal的信念来自早期尝试：他用“GPT-3.5之类的模型”运行BabyAGI和AutoGPT。那些产品当时并不好用，但发展方向似乎清晰可见：模型能力会变强、运行时间会变长，而在能力不变的情况下，模型会变得更小、更便宜。他没料到编码Agent会率先普及，也没预料到CLI会成为它们的交互形态。
2. 推理需求决定增长门槛，也带来产能难题
- 节目开场介绍称，Parallel上个月完成1亿美元B轮融资，估值达20亿美元。Agrawal以推理需求为基准：“全球每投入10块GPU用于模型推理，就会有1–2块GPU用于网页搜索。”他目前按10倍估算增长，高于去年给出的数字。他希望Parallel的增速能跟上甚至超过Agent推理需求，同时也在追问10倍增长还能持续多少年。
- Parallel通过租赁获得GPU，过去3个月里多次陷入紧张的算力调配。所谓可用，不只是有GPU，还要型号、数量、地点和交付时点都合适；部分GPU必须部署在距离网页索引5毫秒以内。对于某些大型集群，Agrawal还要确认这部分实体算力能否锁定足够长时间。他会联系每一家超大规模云厂商和新云厂商。供给情况有时会先在iMessage上谈妥，再由财务团队处理手续。到目前为止，GPU短缺还没有迫使Parallel拒绝客户。
- Parallel当天上线了BrowserBase集成，为BrowserBase提供搜索能力。这也体现了“读取”和“行动”的区别：Agent先通过Parallel的API搜索和访问网页，之后才通过浏览器登录、购买或写入内容。Agrawal所说的“Agent云”整合了推理、可能还有沙箱、网页搜索、浏览器和API操作、个人或企业数据连接器，以及可观测性、反馈闭环和安全能力。
3. 留存和企业采用取决于兑现产品承诺
- Agrawal看好Instinct和Muse等消费级Agent的用户留存，但承认自己没有内部留存数据。消息交互本身就能自然完成引导；关键在于产品能不能兑现承诺。他的检验标准是：“它能不能足够频繁地兑现这个非常简单、概括性的承诺，进而成为你生活的一部分？”
- 被问到OpenClaw是否碰壁时，Agrawal不认同：“OpenClaw各方面的方向都是对的。”大多数人既无法完成配置，也无法随着能力变化持续调整产品。早期表现还高度依赖Opus这样的强模型。托管产品则能在用户看不见的地方持续升级，不必让用户理解具体改了什么。
- 早期客户Harvey和Notion推动了搜索质量提升；语音Agent如今则在推动延迟下降。搜索响应需要约150–200毫秒，而人类搜索大约需要600–700毫秒。用户预期很快就会固化：Agrawal打赌，即使换上新模型，今天关掉ChatGPT的网页搜索功能，90%的请求都会让用户失望。“你会觉得自己像个穴居人。”
- Parallel员工人数从年初接近30人增至约120人，主要是过去6个月扩充了市场拓展团队。Agrawal总结了3类购买触发因素：捆绑搜索无法胜任高要求任务；客户需要在专有模型和开放权重模型之间保持独立性；或者业务规模扩大后需要更好的成本效益。他称，高信噪比的搜索结果能让模型“少搜一半”、减少Token消耗并更快完成任务。
4. 要保持网页开放，需要支付机制和继承权限
- Agrawal的经济逻辑从人类“随机漫步式浏览网页”开始。出版商可以在这段浏览路径中通过订阅、广告、引荐或交易获利。到了Agent时代，“购买行为还在，其他所有行为都断了。”他认为广告和按席位订阅难以持久，因此，让内容所有者的利益与Agent生态对齐，也是基础设施问题的一部分。
- Parallel搭建了一套系统：每当Agent读取内容，就运行一个模型计算应向相应内容所有者支付多少费用；公司也在建立内容合作关系。Agrawal把它称作一种“AdSense”，能够“轻松、自动地差异化定价”。他的判断是，为优质内容付费可以让网页保持开放，以补偿机制取代出版商封锁Agent的动机，让他们有理由欢迎Agent。
- 被问到机器人检测是否仍有意义时，Agrawal区分了爬虫、有用的Agent和恶意机器人。他说自己有“两种相互矛盾的想法”：好的Agent应该表明身份，但在利益机制尚未对齐时，表明身份可能先招来封锁。Instinct和Muse无法访问许多他本人能访问的网站。他希望所有Agent，包括大型实验室之外的Agent，都能继承所有者获准的访问权限和操作权限。
5. 安全是贯穿各层的持续攻防
- Agrawal认为网络安全风险真实存在：Agent既能帮助攻击者，也能帮助防守方。他认为，尽早部署防御型Agent的组织会没事，因为攻击者会瞄准最薄弱的一环；前沿模型权重则是另一个更敏感的问题。他预计编码Agent会被广泛用于网络安全，造成重大影响的网络安全事件将“每隔几周”发生一次，攻防双方也都会投入大量推理算力。
- Agrawal用Twitter作比，把网络安全称作一场“刀战”：攻击可以被遏制，却几乎无法彻底消除；最棘手的是防守方无法识别的行为。网页内容可能进入Agent上下文，进而引发提示注入攻击。过去2年里，模型能力确实取得了实质进展，但“不能说‘模型会解决一切’”。Parallel、模型构建者和应用开发者都必须守住各自的层面；偶尔仍会有失守。
6. 主动式Agent改变搜索和算力分配
- Agrawal最近最关注的是“网页将从拉取转向推送”。Instinct、Muse和Dot已经展现出一些苗头，但尽管一直在线，它们仍以被动响应为主，还谈不上主动。他预计Agent会理解用户明确表达和隐含的需求，在全网设置高信噪比、高精度、高召回率的提醒，并判断哪些变化需要采取行动。智能成本下降后，早期的提醒产品思路可能会重新可行。
- 算力分配面临的对象是一个拥有万亿页面的网页，其中大量内容毫无价值，而且每天都有很多页面发生变化。“无论把算力成本降得多低，我们都没有足够算力处理一切。”资源必须优先投向值得采取行动的变化。Agent生成内容的发布增加了变化频率，也可能放大噪声，让信号提取变得更难。
- O’Shea区分了面向聊天机器人的GEO和面向Agent的AEO；Agrawal起初把两者视为一回事，随后接受了她的区分。Parallel从早期隐身阶段起，就同时提供面向人类和机器的网站视图。不过他认为，呈现形式并非首要问题：“即使出版商还没把网站做成适合Agent使用的样子，我们也能改造整个网页。”
- Parallel围绕“开发者身边有编码Agent”来设计文档和引导流程。客户可以把提示词粘贴到Agent里，由人类授权访问、添加信用卡并接受条款，剩下的交给Agent处理。Agrawal认为，围绕Agent重新设计完整的客户旅程，是根本性的工作。
7. 内在标准塑造Agrawal的经营理念
- Agrawal认为，自己的妻子——一位医生，也是Andreessen Horowitz的投资人——让他的 ambition 超出原先想象。小时候，他在印度的同伴都是核科学家的孩子，身边的人都期待自己做到最好。他把这种环境与招聘联系起来：“不招出众者，就招不到出众者；只招出众者，最后手下就全是出众者。”
- 被问到如何承受压力时，他回答：“什么压力？”Twitter收购期间，他的睡眠也一直很好；他把压力和每天的工作表现区分开来。“我只想做那些我不会雇自己来做的工作”：不安全感会让他保持谦逊、一点点偏执、好奇心和努力，也会促使他寻求帮助。除妻子外，他说自己的要求一向高于别人对他的要求；相比外部标准，内在标准更能驱动他。
完整逐字稿
Parallel Web Systems closed a $100 million Series B funding round last month, valuing the AI startup at $2 billion.
You were at Twitter for 11 years. You were CTO, then you became CEO.
I was also an engineer before that. We’re building a collection of APIs that agents use to search and access the web. Parallel is an almost three-year-old company now. I was the lunatic walking around three years ago telling people, “Agents will use the web 1,000x more than humans ever have.” Most people didn’t believe me at the time. I was wrong.
Why?
1,000x is too low. My deck literally contained a section that said, “Are we too early?” Turns out we were not too early. We’re building a parallel web for agents. We can transform the web. Even if you don’t publish your website to be agent-friendly, we’ll do that for you.
In such a competitive career, at every stage, how do you not break from the pressure?
What pressure?
You don’t feel pressure?
No.
Parag, welcome to Sorcery.
Thank you for having me.
We’re in Index’s office in San Francisco. It’s gorgeous here. They’re one of your investors. I’m not sure if you’ve ever been asked this question before, but what have you gotten done this week?
I’ll tell you what I got done today.
Okay.
We just shipped a really exciting integration with BrowserBase.
Mm-hmm.
We now power search in BrowserBase. If you think about every AI agent people use, what are they doing? They’re looking at your internal information. They’re looking at web information. They’re searching a lot more than we do. Then they’re going into a browser, logging in, and doing something—buying something, writing something.
Mm-hmm.
So that’s what we got done today.
Congratulations. We were talking a lot before this, and we were having a really great discussion. Every interview I’ve had recently has been about some sort of explosive growth in agents, and it’s going to be really fun to unpack all of that. From your point of view, you were talking about read versus write—what you’re doing versus BrowserBase. So I guess to start, let’s talk about that. What is Parallel? How is it different? How has it evolved? And how are you learning as things are changing so fast?
Yeah. Parallel is an almost three-year-old company now. I was the lunatic walking around three years ago telling people, “Agents will use the web 1,000x more than humans ever have.”
Mm-hmm.
There were no agents. Agents didn’t work. Most people didn’t believe me at the time. Fortunately, some did, including Index. At the time, it felt like a science-fiction hypothesis: there would be a lot of agents at work and in our personal lives, and they’d be doing web searches for some reason. I believed it because I’d been obsessed with it for a long time. But today, if you look at what’s happened in the last three years, I was wrong—but on the other side.
Why?
1,000x is too low. At the time, I thought 1,000x was the crazy number to make up, to talk about.
Okay.
Mm-hmm.
Right? At the time, my deck literally contained a section that said, “Are we too early?” Turns out we were not too early. Today, if you force me, 1,000x is too low a number. If you look at some high-level math, an average person does on the order of 10 or fewer Google searches a day. I was checking how many I do: several thousand now. The agents I track—all the agents that I use today—do several thousand per day.
Wow.
So I am now sitting at roughly 1,000x, a little bit less than 1,000 personally. Most of the world is not. Most of the world will get there in due time. It might take us a few years, but 1,000 is a low number.
But is that efficient? Do you think that’ll come back down at a certain point because they’ll just get smarter and know where to go?
Yes, but it is efficient because the other part of the story is that you can’t build old-school search and do 1,000x of this. Think about all the infrastructure that Google might have built to serve on the order of 10 billion searches a day. My best guess is that Google will do 10 billion a day. They have a lot of infrastructure to pull that off. I don’t think we get to build 1,000x that infrastructure to do 1,000x the searches, right?
Mm-hmm.
So part of what we do is build extremely efficient and optimized search, which benefits from models getting better and allows us to distill intelligence for ranking into smaller and smaller models, and to use GPUs as the underpinning of search indexes. Historically, search was done on CPUs.
Mm-hmm.
It was memory- and CPU-bound. Now you get to architect and re-architect search for a price-performance band that the world never needed until today.
One of the things we were talking about before: I recently interviewed G from Vercel, and this is, you know… Maybe this is in parallel, but AI coding agents have reached 60% of traffic on their indexes, from what they see. You had a hot take.
Generally, on the web, the statistic being shared these days, in the last few months, is that agent traffic is exceeding human traffic on the web. That’s roughly the characterization. Now, there are 2 things to remember. An agent doing roughly the same work that a human does will likely do a lot more searches because it can, and a human is time-constrained. An agent isn’t, really.
Yeah.
Right now, they’re not even that dollar-constrained because people are willing to spend on frontier models. For a human to read or search the web to create equivalent value, it takes way more agent traffic. A lot of agent traffic is also people crawling for pre-training. When you look at automated activity versus human activity as a website owner, automated activity is ripping. But my best estimate of where we are right now is that agents are using the web at 10% to 20% of human levels. Which is to say, if my 1,000x number is right, we have at least 5,000x or 10,000x growth left from here. That’s another way of saying we are very early with agents.
Mm.
I think in the Bay Area and Silicon Valley, we’re all living in a little bit of a bubble.
Yeah.
I think most people’s agents aren’t doing the number of searches my agents are doing, but they will.
So Muse just came out with Joly, and I really like Joly, the character for Muse.
Mm-hmm.
We see Dots coming out for OpenAI. We’ve been seeing a lot of these agents pop up for individuals, for consumers. That growth is faster than any we’ve ever seen—even ChatGPT, I think, maybe, possibly.
Yeah.
We’ll see. But how do you expect that to grow, and where does the retention stay?
Two things. One, I have to make a note. We should call out Instinct too. I am—
Yes, I almost forgot.
The reason is, listen, I’m on the rebel alliance here. There are all these big people, and then there are us startup founders who are running at solving really hard problems and pushing the frontier of what’s possible and innovating. So we should talk about Instinct.
Yeah.
They are growing too. Amazing product. So yes, the category is amazing. It’s interesting. I think it’s growing really fast. I don’t have any insider information around retention. As you can imagine, it’s a professional hazard: I use every one of these agents.
Yeah.
I try to deeply understand what they can and cannot do, and what problems we can solve for them.
Mm.
We’ll get into this category and how it changes our product strategy later. But I think my expectation is that these agents’ retention will be good, because for the first time, the product onboarding feels natural. It’s not someone designing a bunch of UX click targets and flows to get you to understand the product. The product is natural. All of these products are somewhat natural, and the question is, can they deliver on the promise created by the UX? The UX is: you can text this thing, and it can do anything.
Yeah.
That’s the UX. It’s not complicated.
Yeah.
Now, retention is entirely a function of whether it can actually deliver on this very simple, high-level promise often enough to become a part of your life. A lot of other products struggle with the onboarding UX barrier to entry. But with this new generation of products, that’s just not a thing. So I have a feeling that retention will get solved here.
Is that where you think OpenClaw maybe hit a wall? Because that was really the first.
No, I think OpenClaw was right in all of the ways.
Mm.
It’s just not for the masses, right? Most people don’t have the ability to set it up and then keep up with change. I think the other factor we discount is that it’s not just a one-time cost to set something up.
Things are changing so fast—
Mm.
—that you have to adapt your instantiation of it alongside that change, and adapt your use of it alongside new capabilities. Part of it might have been that the setup time, because of the flexibility and you controlling it, was far too much. Part of it is also that it was perhaps a little too early, right? In the beginning, there was 1 model that was really good, like the Opus model.
Mm.
As soon as you moved to other models, it wasn't as great in the very beginning, when people started using OpenClaw late last year. So you had to adapt alongside it. But today, as evidenced by Instinct and Muse—
Yeah.
—you kind of don't have to know or think about it, and the product will change underneath. As the customer, you don't need to know what changed in the last month. You're oblivious to the capability improving underneath.
I can't discount that this was really just invented. You mentioned this at the start of the interview, but you've been obsessed with this for a while?
Yeah.
So what does that mean, and how?
I think I got obsessed with it almost 3.5 years ago. I don't know if people remember, but there was this moment of BabyAGI and AutoGPT. Developers in the open-source world were often ahead of everything else.
Mm.
People were building with, I don't know, what was the model of the day? GPT-3.5 or whatever.
Yeah.
They were old models that, relative to today, were really bad. People were trying all these things, and I was trying all of them to make things work. They didn't work. But if you spent any time with the labs back then, you internalized how models would scale and improve.
In some sense, today it feels like everything changes every month. In some sense, it's—
I don't know if it's every month. It might be every week.
But if you look at it a different way, almost nothing that's happening is diverging from what you could have predicted in a straight line 3 years ago.
Yeah.
Models will keep getting bigger and better. They will run for longer. Holding capability constant, models will get way, way, way smaller and cheaper. People will be able to productize around them to disrupt every category of work that can be done on a computer.
These were very simple things that we believed and said 3 years ago, and we've only seen evidence week on week and month on month supporting these few simple things.
Yeah.
Almost everything follows. So now you can say, “Yeah, I didn't predict that the big category would be coding agents first.” I didn't predict the CLI form factor. Those were all bits of information and news to me along the way.
Yeah.
But the macro thesis is that agents are going to get more capable. You're going to do a bunch of inference, and for all the inference you do, you're going to have to do web search alongside it. I've now been on the record everywhere saying that the 1 invariant I believe is that we'll use a lot of GPUs for model inference for agents, whether they be at work or consumer agents. For every 10 GPUs the world spends on model inference, we'll spend 1 to 2 on web search.
Mm-hmm.
That's the business I'm in.
1. The GPU Scramble
So are you aggregating GPUs? You've reportedly raised $230 million, I don't know. Are you aggregating GPUs?
I wouldn't use the word aggregating.
Mm-hmm.
We're renting GPUs to build the scale of search traffic that we're seeing and expecting. As I said, we grow alongside inference. Inference is growing, so we have to do a proportionate amount. I don't know if it's already a core competency of ours. I don't think it is. We've been in multiple mad GPU scrambles in the last 3 months, like, “Oh, shit, we need GPUs. We don't have GPUs. Can someone figure it out?” And you're making a bunch of calls—
Yeah.
—to be like—
Who are you calling?
Sorry?
Who are you calling?
Everyone. Every hyperscaler, every neocloud.
And what's the typical response back? Because this comes up in pretty much every conversation now. How do you get the allocation of GPUs?
I think, at least for us—and I think it's true for many people—the question is, “Which GPU is it?”
Mm.
Physically, where is it? Sometimes we need GPUs within 5 milliseconds of where our large web index is sitting.
Okay.
Other GPUs, we're sometimes more flexible.
Right.
For multi-GPU workloads, it's not just, “Okay, do you have GPUs for me?” It's, “Do you have the right GPU, at the right scale, in the right physical location, available at the right time?” That's what makes the problem much harder.
Mm.
So far, we've managed to solve every problem we've encountered. We've not had to turn away a customer because I didn't have a GPU.
Right.
But we've had a bunch of GPU acquisition fire drills.
I mean, that's a really interesting line to go down. What's the sourcing process like for GPUs? What are the main questions that you're asking? What do you need?
Different workloads are different. For the bulk of our workload, it's how many GPUs, which types are you open to, when do you need them, and where must they be? There isn't too much more to it. For certain large training clusters, there's a little bit more to it.
For us, are you willing to commit to that physical thing for long enough? But that's not complicated. The reality is that I wouldn't say the market is entirely efficient right now. A lot of it—
No.
—is on iMessage, right?
Really?
Yeah. But it works. The market works.
Well, that's crazy. So you're doing deals on iMessage for GPUs?
Not exactly. You're figuring out availability on iMessage, and then the finance person calls the finance person, and the paperwork gets done.
Oh my God. This episode is brought to you by Brex. My favorite. You become what you spend on, and I refuse to spend my time on work that shouldn't exist. Expense reports, receipt chasing, and manual closes. The companies building what's next from Vercel, OpenAI, Anthropic, Granola, and Deepgram all made the same call. They all run on Brex. Brex is the intelligent finance platform that combines cards, expenses, and banking into a single stack with agentic finance built in. AI agents that handle expenses automatically, enforce policy before spend happens, and close your books in minutes. That's why Sorcery runs on Brex, so I can spend time on building and not busy work. It's time to get Brex AF. Learn more at brex.com/sorcery. That's B-R-E-X.com/S-O-U-R-C-E-R-Y. Bye. From handshake to signature faster, Harvey, the AI platform built for legal work is trusted by 80% of the AMLAW 100 and over 20% of the Fortune 500. Learn more today at harvey.ai. That's harvey.ai. AI needs more than chips. It needs power, land, and infrastructure. Zone develops next-generation data center campuses, partnering with AI companies, site developers, and technology leaders to bring compute online faster and at scale. Zone is building the foundation of the AI frontier. Visit zonefrontier.com to learn more. That's zonefrontier.com to learn more.
I want to go back to the topic of scale, because we were chatting about historical scale, what that used to look like, and what you would expect. You were at Twitter for 11 years. You were CTO, then you became CEO.
I was also an engineer before that.
You were also an engineer before that. You've done the whole thing.
Yeah.
You also have a PhD. I've done all this research, but I want to talk about that. Can you set expectations now? What would meaningful scale look like today if, back then, it was “triple, triple, double, double”? What do you have to achieve now?
I don't know how somebody else could do it. For our business, it turns out the answer is relatively easy.
As I've said, we just build alongside inference. If you think about inference growing year on year, to us, that's the anchor. Our goal is to grow faster than inference for knowledge-work agents or consumer agents. If we grow alongside that or faster, I'm feeling pretty good. That's the benchmark for us.
Now, what's the right inference benchmark? That's a question I don't know the answer to long term. But I think in the short term, it's clear that GPUs coming online this year or next year are going to do inference, and search is going to grow alongside it. The short-term benchmark is very clear: our benchmark is the growth of inference.
So what is that in numbers?
I think the easiest approximation is 10X.
You have to think about a 10X goal as the very simple round number.
Mm.
If you think about today, if you’d asked me last year, I would’ve probably said a lower number.
Mm.
But the pattern here is that even when I’ve thought I was being provocative and making up a high number, I’ve somehow managed to underestimate. So now I feel like I can go to 10X, and I feel a little bit uneasy that it feels crazy to grow 10X for too many years.
Mm.
How long can that happen? Maybe I’m wrong again.
2. What Customers Demand
I’m curious, because you get to see so many different market insights internally, externally, all around. What are some surprises that you’ve come across, and what are the signals you’re looking for?
I think it might be useful to tell people about the kinds of customers we end up working with.
Okay.
All of our insights, or at least my insights, come from working with people doing really, really interesting and unique things. We started out working with a lot of the people pushing the limits of building agents. Some of our customers, like Harvey and Notion, were some of the early people building agents to do real work where the quality bar was quite high.
Yep.
Everyone in the adoption chain was asking, “Oh, can we trust this agent to do our work?” That was the question being asked. By working with really demanding customers, or customers who had really demanding customers in turn, we ended up pushing our systems’ capabilities on the quality dimension.
The next phase, I would say, is that voice agents are starting to work. The interesting thing is that, initially, there was some acceptance for your voice agent pausing a little unnaturally or saying, “Let me look that up and come back to you.” That was—
Yeah.
—acceptable for a three-month period, I think.
Yeah.
And then very quickly, today, if that happens to you—
Mm.
—at least people who’ve interacted with multiple voice agents will be like, “That is ridiculous. That’s a really bad product.”
Mm-hmm.
Now you’re seeing all these voice agents that are super natural but also super fast. So what does that mean for web search? Now you’ve got to be able to search the web in 150 or 200 milliseconds. Human web search was a sub-second product, roughly a 600- or 700-millisecond product end to end.
Mm.
We don’t notice it getting any faster, right? For a voice agent, that latency constraint on web search is unacceptable. So now you have to push yourself on the quality dimension, the cost dimension, and the latency dimension. You’ve got to crush all 3.
That’s something we get pushed on by the frontier customer, who says, “Listen, the product experience that we can build today is going to be unacceptable very, very quickly.” So what does every underlying infrastructure piece need to solve—
Right.
—in order to create the market expectation, which changes so fast? Going back to web search, you should at some point try turning off web search in ChatGPT, even with the new model.
Okay.
I think we all used ChatGPT for the first, I want to say, 8 months, maybe 9. It did not have web search, and we were just talking to a model. Try to do that today and see what happens. My bet is: 1, 90% of things you ask, it will disappoint you. 2, you will feel like a caveman or cavewoman using that product. So it’s only 3 years—what? 3 and a half, 4 years ago now.
Mm-hmm.
Our implicit expectations of a product that looks the same have changed so dramatically, and that is what’s mind-boggling to me. I think that is essentially what drives all technical progress—
Yeah.
—is that it becomes table stakes.
Right. And now they might be getting some consciousness. They might have some consciousness to them and start judging us and attacking us.
I’m actually more of an optimist on this.
Okay.
I think they will actually get proactive.
Mm.
They will know us. They will know our explicit and implicit needs, and they will try to proactively do work. That’s the next phase. I think you’re starting to see glimmers of proactivity in Instinct, Muse, and Dot. Glimmers. I think they’re still more reactive than proactive.
Mm-hmm.
They’re always on, but they’re not really proactive. I think the next era, they will be hyper-proactive, which I’m really excited about.
I guess, to peel back—because this is a brand-new economy and layer of the internet, and what could happen? I don’t know. The internet might not exist anymore. Maybe we’ll just use agents. This is what people are saying. What are those components of the underlying infrastructure? You’re one of them. What makes up that base?
There are 2 ways to talk about this, I think. Before going into the internet, let’s first talk about the underlying infrastructure components for building an agent. You need a model, which is inference. You need perhaps a sandbox. I think you need web search, like we provide. You need an action space, which includes a browser and a few other API hooks. And you need access to a bunch of connectors to access your personal data or your company’s data.
You build these together and you can build an agent. The rest of it is all the same stuff we’ve always had: observability, feedback loops, security. And there you have a cloud for agents.
Mm.
Now you go into the web piece of the infrastructure layer. You need to do web search, but you also need to figure out—to your point about the internet shutting down—I do think the primary consumer of the internet is going to be an agent, which means the internet will get re-architected.
The first thing that’s happening is agents are reading the web, which is what we do. What we haven’t figured out is: What is the new business model? Previously, if you put your content out there on the web, you were either building a subscription business—you put some content out there, a bunch of users come, and some fraction of them subscribe for a monthly subscription to a newspaper.
Other internet businesses were monetized by saying, “I’m going to put ads, and I will get paid based on human visitors seeing ads.” Other business models were affiliate business models, or you drive a transaction of some sort. All of these business models have fundamentally relied on us random-walking the web as humans—
Mm-hmm.
—opening pages and taking the next action after the next action on them, some of which are monetizable actions for these participants. If agents are doing this, the buying action still remains. Every other action breaks, so you can’t monetize attention anymore. The ads business model, per-seat subscription business models—these things are not durable. So what replaces them? That’s part of what gets me super, super excited about building Parallel, because I think solving just technology problems is far less interesting than solving a technology and a business model problem.
Mm-hmm.
We have a highly opinionated point of view on how we keep the web open and how we incentive-align content owners with the AI app or AI agent, so that a content owner isn’t sitting there saying, “I’m going to block all AI agents because I don’t know how that becomes business for me.”
Right.
Right? So we need to solve this problem for people: Give me an incentive to make my content available to AI agents. What is that? We built a system that runs a model, not surprisingly, to compute how much to pay each content owner every time an agent reads their content. We’ve now been doing a bunch of partnerships with content owners, and the bet we’re making is that we’ll be able to keep the web open as a result of our version of—we can call it AdSense—which is easy, automatic differential pricing that actually provides content owners value for high-quality content.
That’s a really good point. I was interviewing USV, Union Square Ventures, not too long ago, and Fred Wilson was talking about how Google’s traditional business is being cannibalized by AI, LLMs, agents. It just is. So Gemini is like the Hail Mary to pull it in, and they’re emphasizing that and taking advantage of that.
I’m curious to go deeper into this monetization standpoint. In another interview I did at BrowserBase with John Collison of Stripe, he posed the question: Do we need bot detection anymore? There are obviously good bots and bad bots, but if all the traffic becomes essentially bots—these agents—what is the point?
Yeah. No, I think today there are a lot of people doing a lot of bot detection. The term “bot detection” refers to a few different things: crawlers, actual agents working on behalf of a human or a business, and malicious bots that are trying to do harm.
I think we have to use that word a little more carefully now than we have in the past, because historically people have mostly had negative connotations with that word, as if it isn’t serving anyone.
Yeah.
Today there is a category that is actually highly useful for you or for me.
I would like my agents to inherit the privileges I have. I think we should be careful. The bot detection market is doing well right now because we haven’t figured out how to align incentives.
Hmm.
There’s a separate kind of incentive alignment needed on the action side. Earlier, I spoke about the incentive alignment needed on the content side: why do I let you read my content? The other side is, what happens if I let your agent buy from my store?
You’re starting to see divergent strategies there. I’m of the belief that we need to solve these problems so that all agents—not just the ones built by a large lab—can access all content and take all actions their owner is allowed to take. But we’re not there right now.
One of the biggest topics around growth is cybersecurity and active threats. I know you said you’re an optimist, but realistically, how do you see the biggest concerns around this?
The cybersecurity risks are real. Agents help both defenders and attackers. The people who rush ahead to adopt agents for defending themselves will be fine, because attackers like to find the weakest link, and there will always be a weakest link. Broadly, that’s what I think.
The security risk for something highly sensitive, like the model weights of a frontier model, is a whole different—
Mm.
—conversation.
Yeah.
By and large, my intuition is that we will use coding agents aggressively to secure things. Along the way, there’s going to be a consequential cyber event every few weeks, which will drive action and urgency around getting ahead on defense. We’re going to spend a bunch of inference on both attack and defense.
Do you see any of this as analogous to your time at Twitter in consumer internet? Obviously, it’s one of the largest—
Yeah.
—social media platforms, and so—
Yeah.
—that creates—
No.
—a lot of vulnerabilities.
There are obvious parallels. Not to keep calling out my company’s name again and again, but at Twitter, there was a real incentive to attack, whether it was crypto spam—
Yeah.
—which was popular for a period—or some state-sponsored influence campaign—
Mm-hmm.
—which we used to disclose publicly a lot. Those things always happened, and it took real energy and focus to control them. The thing most people don’t understand is that it’s almost impossible to take them to zero. You have to contain them because it’s a knife fight, and the tactics change. Any time you control some vectors, another new vector will emerge, and then you’ll control that.
The other thing people don’t understand is that most of the bad stuff doesn’t show up on the surface. If you saw an account and thought it was a bot, it was harmless because you could see that it was a bot. The real problem is going to be in the margins, when we lose our ability to differentiate.
Yeah.
Either as people or as the technical people building these systems. When you can’t differentiate a human from a bot or an agent, that’s when you think, “Okay, now I don’t know what I’m up against.” I think that’s very similar to what we used to deal with back then.
Do you think bots should address themselves?
I have 2 contradictory feelings about this.
Okay.
I do think good bots should declare themselves so that others can recognize them as good bots and treat them differently from bots that might be adversarial. And again, we keep using the word “bots.” I’d like to use the word “agents.”
Okay. Well—
Right?
Every time we use “bot,” we just overlay it.
Because I really think “bot” has a more negative connotation, while “agent” is a more appropriate word.
Mm-hmm.
I think agents are perhaps the good bots, mostly. So, yeah, I think they should declare themselves. I know there are a bunch of protocols being developed around declaring identity to the world at large.
The core challenge on the other side is a conceptual question: should your agent have the same access as you? Should it inherit your access, or should it have only part of your access? It would be so much better if we could figure out a way to get rid of the bad bots, but live in a world where we could enable our agents to have the access we have and grant them that privilege.
Today, it’s strange. I can’t get my Instinct[?] or Muse[?] to have the access I have on many websites because they got blocked there, and they’ll get even more blocked if they start declaring themselves.
Yeah.
That’s why I think we’re in this moment of transition, where these products have rough edges partly because we haven’t aligned incentives. If we started declaring them everywhere, and they started declaring themselves everywhere, a bunch of people would start blocking us because we haven’t done the work to align business models.
Yeah.
Then these products would look worse and be less useful, and that sucks. To me, the existential problem is: how do we drive alignment between agents and everything on the web?
Do you think everyone will need alignment and safety on their team? Is that the next higher class?
Yeah. How do you not deal with the problem? If you think about us, we’re the web layer. The web is a dangerous place. Anyone can post anything malicious anywhere, and a lot of that content will flow into an agent’s context. So you can do injection attacks.
To the credit of model builders, in the last 2 years, models have made very real strides in not being very vulnerable. But this whole world will rely on defense in many, many layers. Almost every layer has to obsess about these problems.
Mm-hmm.
You can’t say, “Oh, models will solve it.” We have to focus on solving it at our layer before we get into the models. Then models have to focus on it, and the app builders who are building with models and with us have to focus on it at their layer. We have to deal with adversaries at every layer, and that’s what gives you true defense. Still, the occasional thing will leak out.
How good is AI at real professional work? Mercour's AI Productivity Index, or APEX, tracks it with every major frontier model release. Built with partners like Harvey and Cognition, see where the frontier actually stands at mercour.com/apex. Today's episode is sponsored by VC-X by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. Learn more at getvcx.com. Some of you may not have heard this yet, but our sponsor Public just launched something called Generated Assets, and it brings AI into investing in a way I've honestly never seen before. Here's how it works. You type in an idea like AI-powered supply chain companies with positive free cash flow or defense tech companies growing revenue over 25% year over year. Public's AI then dispatches a swarm of agents that scan every single U.S. stock, evaluates them, and instantly builds a custom index around your thesis. What really stands out is how clearly it explains why each stock is included. And before you invest, you can even back test your idea against the S&P 500 so you're making decisions with real context, not just guessing. And beyond Generated Assets, Public lets you invest in stocks, bonds, options, crypto all in one place. They'll even give you an uncapped 1% match when you transfer your investments over from another platform. If you wanna build a portfolio that actually reflects your thesis, visit public.com/sorcery, paid for by Public Investing. Full disclosures in the description. Founders scale faster on Deel. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/sorcery. That's D-E-E-L.com/sorcery.
3. Building Parallel's Next Phase
How did you recruit for your team, especially as this is moving so fast? You have global offices now.
Yes.
How big is the team, and how have you recruited? What are the main roles?
The team is about 120 people now. I think we started the year close to 30, so this has been a year of explosive growth. A lot of that growth has been because we had no GTM at the beginning of the year.
We were a pure tech play. Over the last 6 months, we’ve built a pretty sizable GTM team, because we’re seeing that almost every business that buys inference needs to buy web search. We’re finding really large companies—the top 100 or 500 companies in the world—starting to spend on tokens and buy from us alongside those tokens to make them more effective.
A lot of our growth is the sales team, which has been for the past—
Well, how do they know to buy you alongside the tokens?
They run into problems. One of the common things people run into is that they start using a model.
That model perhaps bundles itself with a web search.
Okay.
And they start building an application with that.
Hmm.
Sometimes it’s not good enough, and the team realizes that the web search component is causing the problem. Despite the model being great, the search is making it incapable of solving the problem they’re trying to solve. That’s perhaps how we started working with some of the AI legal, AI finance, and AI science customers.
Right.
The second point at which they really need us is when they decide to use multiple models. They might say, “Okay, we want to switch across multiple proprietary models on a dime,” or, “We want to start adopting open-weight models.” At that point, having the highest-quality web search that works with every model enables model independence and gives you the flexibility to build your application. That’s the second reason people buy from us.
The third reason is when people have really scaled use cases and start optimizing for measuring ROI and end-to-end costs. Because our search is so high-signal—
Right.
—and token-efficient, the models do half the number of searches, use far fewer tokens, and are cheaper and faster end to end. That’s the third reason people buy from us. They’re spending too much on tokens, the product works and is useful, and they want us to help them be more efficient.
So it’s not just routing. You also help with token optimization.
Right. For a lot of categories of knowledge work where a lot of work happens on the web, we help a lot with token efficiency.
Wow. When we talked to Gia Vercel, we were talking about the shift in GTM. It was SEO, then GEO, and now it’s AEO.
I think GEO and AEO are the same.
A little bit, but not really.
Okay. Okay.
Because it’s agents now—
They’re more than chat.
Yeah.
GEO was like chat.
Exactly.
Okay. Fair. Fair.
So Vercel has optimized itself for agents coming to its website—
Yeah.
—so agents can read the website and know—
Yeah.
—exactly what products they offer—
Right. Right.
—and how they can use them.
Yeah.
And because of that, Vercel is one of the top tools for building out your products and that kind of coding environment.
Yeah, yeah. No, I think it’s interesting you bring that up. Three years ago, when we put up our first website, we were still in stealth. It didn’t say anything. We had a machine version. To this day, if you go to our website, you’ll see the human version, and then there’s a toggle to switch to the machine version.
Really?
As far as I know, we were the first website to take on this metaphor because we’re building a parallel web for agents.
Yeah.
We’ve always carried this metaphor, which the world is now coming around to. It’s a cute thing to do because the truth is, the presentation layer almost doesn’t matter. We can transform the web even if you don’t publish your website to be agent-friendly. We’ll do that for you. We can understand what the human version is and how to help it become agent-friendly.
Oh, wow.
We don’t do AEO or GEO, but I think your website will become consumable by agents with us.
Hmm.
We even think about our onboarding, documentation, and every part of the developer or customer journey with the primary assumption that it’s a human with a coding agent, or a developer with a coding agent, rather than a developer alone. On our website, you’ll see prompts like, “Put this into your agent,” and then it’ll know exactly how to use Parallel to solve whatever problem you’re trying to solve. With one click, the human just needs to authorize, add a credit card, and accept the terms. The agent can rip through the rest.
Really rethinking how you design every experience to be agent-first is pretty fundamental. Vercel is very agent-built, so people who live in that world have an advantage.
We have probably 10 or 15 minutes left, but I’m curious. You’re doing a lot of press. You’re out there, and you’re used to getting asked a bunch of questions. What are the questions you want to get asked?
I spend a fair amount of time thinking about proactive agents and always-on agents, and how they change the compute world and the web. The question for me is about this notion of a proactive web. I’ve been out there saying the web will go from pull to push.
So far, as humans, our primary way of searching the web has been to go and search on Google. A rare few used Google Alerts, but it didn’t really work. People tried to build things like Google Now and a bunch of other things that would send you notifications. Nothing really worked on the web in terms of being proactive. I think with agents and the cost of intelligence going down, it is going to work.
Yeah. Mm-hmm.
Now imagine an Instinct or a Muse implicitly understanding your world and your life, and being able to set up really high-signal, high-precision, high-recall alerts on everything that changes across the entire web. Then your agent can know when it needs to act on something on your behalf. If something changes in the world or something might be happening, what do you need to do as a result? Does that change your plans or your actions? How we allocate compute in that world is a problem I’m currently obsessed with.
Hmm.
The web is really large, right? There are a trillion pages out there. A bunch of them are changing every day. Most of it is junk. No matter how cheap you make compute, we don’t have enough compute to process everything. So now we need to be really smart about prioritizing our compute over everything that is changing on the web, so more compute can be spent on things that are actually actionable—
Hmm.
—rather than junk changes. The rate of change on the web is also going to increase dramatically. You’re starting to see that. People are publishing with agents. The AEO and SEO businesses are churning out agent-generated content on the web.
Yeah.
Which just means the rate of change on the web is increasing. The job of a search engine is to pull the signal from the noise as potentially more noise gets created.
How do you think about data centers in space?
I wish I knew more about it.
Mm-hmm.
I hope they work.
Have you gone down any of these rabbit holes around compute?
I haven’t gone down the data centers in space rabbit hole. I do think we need compute.
Yeah.
I have gone down the rabbit hole a little bit on how sovereigns will participate in the data center.
Mm-hmm.
And that’s fascinating to me—how that’ll play out.
Well, one of our partners is Brex, and they're all about performance and spending smarter, moving faster. I take this topic in a personal way. I do think performance is a matter of who you spend your time around. Some people say it’s your closest 5 people. Some people think it’s your mentor or the people who inspire you. Your career has been tremendous and very interesting, and I’m curious: who is that person or those people for you?
My wife is Benita. I was an engineer at Twitter when we got to know each other. So, if you think of the part of my career that people know about, where I’ve done well, it’s mostly her. She’s a superwoman. She’s hyper-ambitious, and I was ambitious, but I think she has driven my ambition even higher.
I’ve always aspired to be in the top X in the world at anything I do. That’s the only way. I was driven to do that for math, for physics, and for computer science research. I think I got there for all of those. But I don’t think I would have thought I was capable of this until she pushed me. Then I was like, “Yeah, I think if you care enough about something, you can become top 10 in the world at anything.” The top 1 is a whole different field.
For me, the standard-setter is my wife. She’s a physician and still practices. She’s an investor at Andreessen Horowitz. She’s done a startup. She does healthcare and biotech. She works really hard and is an amazing mom and wife.
That’s amazing. So you guys have a nice little competitive dynamic?
Yes. Yes. Yes. I’m privileged that when I go to a lot of things, people use me to access her.
Really?
Oh, all the time. People will come to me to figure out how to get me to forward an email to her.
Oh my gosh. That’s amazing. Well, that’s a really great call-out. I wasn’t expecting that. Why do you think partner selection, or who you choose as a partner, is so important?
I think it’s probably the most important decision in life. And, as you said, your peer group…
I was fortunate growing up to have a really crazy, weird, outlier peer group. All of my friends were essentially kids of nuclear scientists in India. Everyone was like, “Okay, I’m going to be the best at whatever we try to be.” There was an ambition, a competitive expectation, all around.
Mm.
Then I went to college, and my college was the most competitive. Then I went to grad school, and that was the most competitive. When you surround yourself with that, you kind of surprise yourself by being around high expectations. This is what everyone’s talking about now, right? When you build a team, high talent density is a superpower.
Mm.
If you don’t hire outliers, you don’t get outliers. If you hire only outliers, you get only outliers. There’s a lot of self-selection that happens. I’ve been fortunate to go through that journey. I got started in life with that kind of peer group, and then you kind of won’t ever let yourself not be in that peer group. That percolates into your partner selection—
Yeah.
—which then further drives things. So it’s kind of a self-fulfilling prophecy of sorts.
In such a competitive career, at every stage, how do you not break from the pressure?
What pressure?
You don’t feel pressure?
No.
No? Why?
Why would I?
A lot of people do.
Yeah, I think it’s kind of a made-up thing. I really believe that. Through the Twitter acquisition, I used to track my sleep score.
Really?
I was sleeping fine. If you know what you must do, you do it. You don’t have to feel pressure, right? The world can come at you, but if you know what you’re capable of and you have a slightly long-term view, everything is a week or a month, so you don’t have to feel pressure. You have to perform every day, which is a completely different thing. The core question you have to ask is: are there moments where you must feel pressure to perform? I really like feeling insecure. That’s my word.
Really?
I’ve always told people that I only want jobs I wouldn’t hire myself for. I want to feel that somehow someone’s thrown me into a challenge that, on paper, I’m not supposed to be able to accomplish. One, that drives you to be slightly humble. Two, it makes you a little bit paranoid. It makes you want to learn things. It makes you curious, work really hard, and seek help. If you feel secure in what you’re doing, it doesn’t drive any of the most important behaviors, in my experience. Pressure is a different word, because it suggests that you’re trying to meet somebody else’s expectations instead of your own. I think internal expectations are way, way, way better drivers of success—
Mm.
—than external ones. At least for me, other than my wife, my expectations have always been higher than the world’s expectations of me, even from true believers who believed in me.
What a fascinating turn. I did not expect that. This is like—we… I will say thank you so much—
Okay.
—for such a fantastic conversation.
Yeah.
It was really awesome to learn more about Parallel and your personal life, too.
Yeah.
If only we’d started with this.
Yeah.
Your philosophy and how you think about things is really impressive.
Yeah, that’s fun.
Yeah.
Yeah.
Well, thank you so much, Parag.
Amazing.