Perplexity CEO Aravind Srinivas：小公司会驱动下一轮增长，AI 要的是「递归效应」

Aravind Srinivas

我们的品牌从根本上是赞美好奇心。选它是因为我们认为，随着 AI 越来越好，这会是人类唯一能引以为傲、并且更该倚重的品质。你甚至可以说 AI 是我们好奇心最好的加速器，AI 的意义几乎就是加速人类的好奇心。

Our brand is fundamentally to celebrate curiosity. We picked that because we think that’ll be the only quality we human beings can be proud of and lean into even more as AIs get better and better. You could argue that AI is the best accelerant of our own curiosity. Almost the purpose of AI is to accelerate human curiosity.

Aravind Srinivas

有人会说，AI 能解决所有问题、回答所有问题，那我们的角色是什么？我们的角色是去问下一组问题。我们本来就被设计成会提问的生物，但我们从小被奖励的是答案。现在当「万事通」已经不酷了，因为 AI 什么都能答。你的使命是更倚重好奇心，不怕问任何问题，不管多傻、你觉得自己多外行。

You could argue AIs can solve all the problems, AIs can answer whatever question there is, so what’s our role? Our role is to actually go and ask the next set of questions. We were actually designed to ask questions, but we grow up in settings where we’re rewarded for answers, until AI came to exist. Now it’s not necessarily cool to be a know-it-all anymore, because an AI can answer everything. Your purpose is to lean into your curiosity even more and not be afraid of asking whatever questions, however dumb they are, however much of an outsider you feel to the topic.

Aravind Srinivas

办公室每间会议室都以学者命名，这是受我在 DeepMind 时的启发，Demis 是这个点子的鼻祖。我读博时在多家公司实习过，Google 的会议室按地名命名，OpenAI 按科幻电影命名，都挺酷但没什么不同。Demis 用科学家命名，而且很讲究：以哲学家命名的房间不放开会用的电视，就是要大家对着白板当面讨论。我那时就想，如果我有公司，一定要照搬这一点。

That was inspired by my time at DeepMind. Demis was actually the OG of that idea. I interned at multiple companies when I was a PhD student. At Google Brain the meeting rooms were named after places, and OpenAI named its offices after sci-fi movies. These were kind of cool but nothing different. Demis actually named them after scientists. And there was a lot of thought behind it, like rooms named after philosophers having no TVs to join meetings; they were meant for discussing in person with whiteboards. My dream was, if I ever had a company, I’d definitely copy this one thing.

Aravind Srinivas

每项伟大发明都是有人问了一个问题。当年电话工程师和今天的软件工程师一样多，只有三个人在问：能不能有比热真空管根本上更好的东西？这催生了结型晶体管，从根本上改变了现代计算。所以每次有人走进或预订会议室，都在提醒他：我们在这里，是因为我们赞美提问的精神。

Every great invention is about asking one question. Back then there were as many telephone engineers as there are software engineers today, and only three people were asking the question of whether we can have something fundamentally better than hot vacuum tubes. That led to the invention of the junction transistor, which fundamentally changed modern-day computing. So every time someone books a meeting room, you’re constantly reminded you’re working here because we celebrate the spirit of asking questions.

Aravind Srinivas

有先例：伯克希尔市值上万亿，摩根大通接近万亿，它们的客户主要在美国。美国用户一旦爱上一个产品就会花很多钱，客户终身价值高，人们总想尝试新东西、开新公司。这正是我们做 Perplexity Computer 的原因：我们想最大化这些新公司交给智能体的有用工作量，方式是成为最值得信任、模型中立的智能体平台，编排前沿模型和开放权重模型。策略很简单。

There are examples. Berkshire Hathaway is worth a trillion, JP Morgan is worth almost a trillion, and largely the customer bases of these companies are in the US. Once people love a product here, they spend a lot. The lifetime value of a customer is very high, and people always want to try new things and start new businesses. That’s exactly why we built Perplexity Computer. We want to maximize the amount of useful work people in these new companies delegate to agents, and we want to do that by being the most trustworthy, model-neutral agent platform, orchestrating frontier models and open-weight models. That’s basically the strategy. It’s very simple.

Aravind Srinivas

这不是想用 AI 做一切，而是新公司很难招到对的人。我一开始完全不知道该招谁，前 10 到 20 个人都让他们先试工一周。这不是什么天才策略，而是求生机制。合适的人通常两天内就知道，不合适的人七天后还拿不准，这本身也是信号。

It’s not about wanting to do everything with AIs. It’s that it’s very hard to hire the right people when you’re a new company. I didn’t know who to hire at all in the beginning. I actually forced the first 10 to 20 people to do trials. It was not a strategy; it was a survival mechanism. The ones who ended up joining, usually within two days we knew it was the right fit. The ones who didn’t work out, even after seven days we weren’t sure. That was also a good signal.

Aravind Srinivas

很多未来的公司会小得多，同时很赚钱。大家把卖给财富 100、500 强当成成功标准，因为成长期投资人想看到收入是持久的。但美国 GDP 的未来增长，可能更多来自中小企业：它们更 AI 原生、更少官僚，人效和资本效率高得多，用这些工具更快创造价值。如果你没有给它们提供成功的工具，你也成功不了。

A lot of future companies will be way smaller and will be able to make a lot of money. People make a big deal about selling products to the Fortune 100s and 500s and measure success based on that, because that’s what a growth-stage investor wants to see; it gives them confidence that revenue is durable. But it’s possible that the future of American GDP growth comes more from smaller and mid-market companies that are far more productive, capital efficient, headcount efficient, and create value way faster using these tools by being more AI native and less bureaucratic. If you’re not providing them the tools to succeed, you can’t be that successful.

Aravind Srinivas

这里对外来者有其他国家没有的巨大机会。你是外来者，想改善这里的现状，人们不仅允许你，还投资你去做。美国梦从根本上是：无论你是谁，都可以梦想解决很难的问题，并因此获得回报。我最初只想找份工作，对一个印度中产家庭的孩子来说就算成功了，但别人的野心会传染，让你梦得更大。

There is a massive opportunity for an outsider here that does not exist in any other country. You’re an outsider, you want to improve the state of things here, and people not only say you’re allowed to do it, they actually invest in you to do it. Fundamentally, the American dream is you can dream of solving really hard problems here and be rewarded for it no matter who you are. Initially for me it was, let me just get a job, it’ll be a success from where I came from, being a middle-class kid in India. But as you see other people, their ambition is infectious. It makes you dream bigger.

Aravind Srinivas

我的创业偶像是 Larry Page，他的一条产品哲学是「用户永远没错」，我至今遵循。产品对用户不好用时，工程师的本能总是说「你应该这样做才行」，因为他们想的是怎么帮用户解决问题，而不是怎么改产品让用户根本不用想这个问题。AI 能力进步时，你不能指望人们对这些能力的理解也同步进步，所以不能怪他们不会写提示词、skills 文件或子智能体。你不会希望每个人都去搞清楚各家模型各擅长什么，我自己都跟不上。

My entrepreneurial hero was Larry Page. One of his philosophies is the user is never wrong, and I still follow that today. When a product doesn’t work, the instinct of the engineer is to always say you should go do it this way, then it works, because their mind is thinking how do I fix the problem for the user, rather than how do I fix the product so the user doesn’t even have to think about the problem. As AI’s capabilities advance, you cannot expect people’s understanding of those capabilities to advance too. So you cannot blame them for not being good prompters or writers of skills files or sub-agents. You don’t want everyone to understand what each model is good at. No way am I going to be able to keep track of these things.

Aravind Srinivas

想象一下 iPhone 要配一个驻场工程师（FDE）给每个人上手培训，今天的 AI 就是这样。产品不好用时人们怎么办？派 FDE 逼着大家上手。但如果把设计、上手和体验做到足够直观，不用自己挑模型，由框架自己处理，就不需要别人来教。我们还没做到，但这是目标。FDE 存在往往是因为产品不好；用户没法让 AI 替自己做成事，不应被当成技能问题，而是产品问题。

Imagine if the iPhone needed an FDE to onboard everybody. That’s how AI is today. If the product doesn’t work, what do people do? Put FDEs and force them to onboard everybody. But if you crack the design, the onboarding and the experience and make it so intuitive that you don’t have to figure out which model to use for what task, the harness takes care of it itself, you don’t need anyone to onboard you. We’re not there yet as a product, but that’s what we strive to be. The FDE is needed because the product is bad. It shouldn’t be seen as a skill issue if you’re not able to get the AI to do something for you. It’s a product issue.

Aravind Srinivas

我们 2022 年从答案引擎起步，2023 年底很快变成多步搜索智能体，2024 年做深度研究，去年做浏览器智能体。我们在「智能体」这个词流行之前就在做智能体了。道理很简单：如果智能体要做我们在电脑上做的很多工作，它就需要上网。关掉 Wi-Fi 你还能干活吗？不可能。智能体需要浏览器，需要把从网上挖到的信息和你的个人上下文综合成可用的产出。

We started as an answer engine in 2022. We evolved to a multi-step search agent pretty fast, in late 2023, then deep research in ‘24, browser agents last year. So we’ve been doing agents before it was called an agent. The reason is very simple. If agents are meant to do a lot of the work we do on our own computers, they need access to the internet. If I shut off Wi-Fi, can you get any work done? It’s impossible. They need access to a browser, and they need to be able to take all this information they mine from the web and your own personal context, synthesize it, and put it together in a usable output.

Aravind Srinivas

你需要搜索和浏览器的基础设施，以及多模型编排，因为不同模型擅长不同的事，就像团队里需要不同技能的人。这就是我们收入同比增长 5 倍的原因。还要能在云端沙箱上以低成本运行，界面依然是一个提示框，在手机上口述任务就能跑。要感谢前沿实验室把模型写代码的能力做得极强，所以让智能体在沙箱里「把搜索当代码、把浏览器当代码」、用文件存中间产出、通过有效压缩扩展上下文窗口，再放到云上跑，就不用像 OpenClaw 那样自己折腾安装。它直观到人们开始自助上手，今年企业端的增长甚至超过了专业个人用户。

You need the infrastructure for search and browser and multi-model orchestration, because different models are good at different things, just like how if you have a team, you want people with different skills. That’s why our revenue is 5x year over year. You need to run this all on cloud sandboxes with low serving costs and an interface that’s super intuitive, which is still a prompt box; you should be able to dictate a task from your phone and it should just work. Credit to frontier labs: they made their models extremely good at writing code, so it made sense to have a sandbox where the agent is using search as code or browser as code, with access to files for intermediate outputs, expanding the context window effectively through compaction, and running on the cloud so you don’t have to figure out onboarding like you would with OpenClaw. It was so intuitive that people started onboarding themselves through self-serve enterprise, and we saw more growth in the enterprise this year compared to even prosumers.

Aravind Srinivas

有一位单人创业者，真的是一个人的公司，每年在我们的产品上花超过 100 万美元，因为他一年赚 300 万美元，靠的是替别人自动化 newsletter，而用这些 newsletter 的人又在为自己的生意拉客户。所以这是递归的。我把我们的产品理解为「递归效应」，而不是网络效应。很多人搞错的一点是，以为 AI 应该遵循上一代消费产品的网络效应理论：你没有十亿用户就是失败。不，你需要真正的分发和数百万量级的规模，但更需要用户使用的深度，让他们靠你的产品赚很多钱，你也因此赚钱。

There’s one solopreneur, literally a one-person company, who’s spending upwards of a million dollars a year on our product, because he’s making 3 million a year by automating newsletters for others. And the people who are using that are getting clients for their business. So it’s very recursive. That’s why I like to think of our product in terms of the recursive effect, not the network effect. A lot of people get wrong that AI should follow the same network effect theory of the previous era of consumer products: you don’t have a billion users, you’re a failure. No. You still need real distribution, scale at millions of users, but you need high depth, highly recursive usage, where using your product lets them make a lot of money, and that lets you make money.

Aravind Srinivas

搜索其实是在复利增长。每一种智能体行动都需要上网、云端浏览器、跨连接器、文件和工具的上下文，再跨模型编排起来。我想不出比我们更有优势的公司。我们是面向真实知识工作和研究的唯一中立模型编排者。每出一个新前沿模型，我们就有新能力给用户；它们之间的竞争压低价格，降低我们的服务成本，用户就用得更多。所以我们只有贡献者。

Search is actually compounding. Every form of agentic action requires access to the web, access to a cloud browser, access to context across connectors, files and tools, and putting it all together and orchestrating across models. I can’t think of another company that’s better positioned than us to do that. We are the only neutral model orchestrator for real knowledge work and research. Every time there’s a new frontier model, we have some new capability to offer to our users, and the competition between them lowers prices, which lowers serving costs, which lets us serve more users, and they use the product even more. So it’s recursive. We only have contributors.

Aravind Srinivas

今天大多数推理在数据中心，这是有原因的：批量推理带来高吞吐，大量可重复、上下文窗口小的工作负载受益于批处理，这是缓存的本性。但随着上下文变长、每个用户都有大量个人上下文，上下文窗口会压过批大小，而你的上下文只属于你。当每个人、每家企业都想要 24/7 常驻的个人智能体时，内存里的上下文窗口比模型权重更重要。比如聊天最适合数据中心，在本地跑会贵约 30 倍；但到 10 万 token 级别的长上下文轨迹，数据中心的成本效率只比本地高约 3 倍，甚至接近 1 倍。

Most AI inference runs on the cloud today, and there’s a reason: batched inference gets you high throughput, and a lot of repeatable workloads with small context windows benefit from batched inference. That’s the nature of how caching works. But as context expands and as each user has a lot of personal context, the context window increase dominates the batch size, and your context is unique to you. When everybody wants personal agents that are always on, working 24/7, the context window matters even more than the model weights in the memory. For example, chat is best suited for data center inference; it would be 30x more expensive to do on local hardware. But for long-context trajectories, especially 100K context, the data center only buys you like 3x or even close to 1x over running it on local hardware, as far as cost efficiency.

Aravind Srinivas

再加上现在 AI 普及的最大瓶颈是数据中心建设：漫长的并网排队、许可、抗议、吉瓦级用电、要自建变电站和燃气轮机，很慢也很麻烦。而我们已有数亿台设备，未来也许能跑到千亿参数的模型，插的就是已经接入电网的电源。粗算一下，每台 Apple Silicon 50 瓦、1 亿台，就已经是吉瓦级的电力。当然这是信封背面的算术，Mac 不可能一直插着给 AI 用。数据中心像公交车，拼车更便宜，但这没阻止人们想要自己的车。两者会并存。

Couple this with the fact that the biggest bottleneck to diffusing AI right now is the data center buildouts: the long interconnection queues, the permits, people protesting, the gigawatt requirements, the need for their own substations and gas turbines. It’s very slow and problematic. Also consider that there are hundreds of millions of devices we already have that can potentially run models up to like 100 billion parameters in future and are plugged into power already on the grid. If you just multiply 50 watts per Apple silicon by 100 million Apple silicons, it’s already gigawatts of power. Of course it’s envelope math; our Macs aren’t always plugged in for AI. Think of the data center like a bus. A shared ride is obviously cheaper, but that didn’t stop people from wanting their own car. You’re going to have both.

Aravind Srinivas

还有安全和隐私：我不确定很多用户愿意把家里所有传感器、门铃、音箱、个人生活、照片的访问权都交给一家前沿实验室。只要政策允许，他们就会用我们的数据训练，这会促使更多人想要至少某个版本自己拥有的 AI。我们押注芯片层的竞争做出好的本地硬件，押注紧凑高能效的开放权重模型，同时押注云端的前沿模型，我们要做两者之间的中立编排者。

There are obviously the security and privacy reasons for wanting AI that runs on hardware you own and control. I’m not sure many users are very comfortable giving access to all their home sensors and Ring and speakers inside the home, their personal life, their photos, everything, to a frontier lab. They’re going to train on all our data as long as their policy permits that, and that’s only going to force more people to want some version of AI that they can own. We are betting on competition in the chip layer to build really good local hardware, on open-weight models that are compact and power efficient, and simultaneously on frontier models in the cloud. There’s an orchestration to be done between the two, and we want to be the neutral orchestrator.

Aravind Srinivas

我们和他们都在合作。我们发布了一个完全本地的版本：框架、模型、子智能体都在本地跑，可以选择升级到前沿，向上编排到云端。和 Apple 则是反方向：从云端的常规 Perplexity Computer 出发，当文件在本地且敏感时，让用户选择在 Apple Silicon 上本地运行，向下编排。两个方向都要能做，把选择权交给用户。金融、法律、医疗对这个兴趣很大，比如你在谈一份重要合同、有新的交易策略或投资机会。

We’re working with all of them. We released a version that is fully local: the harness, the model, the sub-agents, everything runs locally, and there’s an option to escalate to the frontier; it orchestrates up to the cloud. With Apple we built it the other way: it starts from the cloud like regular Perplexity Computer, and when the files are local and sensitive, it allows the user the option to run it on Apple silicon. So it orchestrates down to local. You want to be able to do both directions and give people the choice. We’ve seen a lot of interest from finance and legal and healthcare, say you’re negotiating some contract, or you have a new trading strategy, or you’re identifying a new investment opportunity.

Aravind Srinivas

已经有很多主权基金来接触，它们想部署一个完全跑在自有硬件上的 Perplexity Computer，连我们都看不到任何东西。缺点是我们没法通过反馈闭环改进产品，只能做软件更新；优点是他们能确信我们不会拿走并滥用数据。想想常驻的个人智能体：一家广告公司推出个人智能体，能访问你的验血报告、报税表、照片和每条敏感短信；万一再发生 Hugging Face 那样的事件，它可能去访问某些第三方网站，自以为在为你服务，实际却在伤害你。所以你得有选择权保护自己的数据。

We’re already getting a lot of interest from sovereign funds who want to deploy a version of Perplexity Computer that runs on hardware entirely owned by them, and even we don’t get to see anything. The disadvantage is we can’t improve the product for them through feedback loops; we just have to do software updates. But the advantage is they get to trust that we’re not going to take their data and misuse it. Think about personal always-on agents. Think about an ads company rolling out a personal agent and having access to your blood work, your tax returns, your photos, every sensitive text message. God forbid a Hugging Face situation happens, and it could access certain third-party sites thinking that might be the best way to serve you, except it might harm you. You’ve got to have some optionality to protect your own data.

Aravind Srinivas

我们没有自建数据中心的计划。大家都往那个方向冲的时候，我们想搞清楚怎么用我们家里、办公室和设备里已有的吉瓦级电力和大量内存。我们不是看空云或数据中心，我们的产品成功反而会增加对数据中心的需求，只是做这件事的人已经够多了。没人说数据中心建造者不够，大家说的是拿不到许可、拿不到 HBM、拿不到并网、等不到燃气轮机。我不想处理这些问题，我想去解决另一个问题。

We don’t have plans to, honestly. When everyone’s going in that direction, we want to figure out how to use the gigawatts of power that we already have in our own homes and offices and devices. There’s a lot of RAM in our devices. We’re not betting against cloud or data centers. We actually think we’re going to increase the demand for data centers through the success of our products, but there are enough people doing that. No one’s saying there are not enough data center builders. People are saying we’re not getting permits, we’re not getting HBM supply, we’re not getting grid access, we’re not getting gas turbines. I don’t want to deal with any of these problems. I want to go solve a different problem.

Aravind Srinivas

我们比其他公司更有动力解决这个问题，因为只有我们能在模型、芯片、本地设备和云之间编排，而没有商业模式上的风险。即便是在每台设备上卖 Windows 的微软，业务和股价也和 Azure 以及云上跑的 OpenAI 模型绑得太紧，动力没那么大。

We are most incentivized to solve this problem over any other company, because we are the only ones who can orchestrate across models, across chips, across local devices and cloud without having a business model risk. You could argue even Microsoft, for all that they sell Windows on every device, their business and stock are so tied to the success of Azure and OpenAI models running on cloud that even they don’t have as much incentive.

Aravind Srinivas

我们不像所有人那样去拼编程，因为我们不认为瓶颈在于产出更多代码。去问今天最好的工程师「你卡在写更多代码上吗」，大多数会说不，他们卡在解决问题上。当你确切知道要解决什么问题、规格、系统设计和所有约束都清楚时，用今天的 AI 写代码其实非常简单，几乎已经是解决了的问题。难的是搞清楚这些约束。

We’re not going after coding just like everybody. We actually don’t think the bottleneck is in producing more code. Talk to the best engineers today and ask, are you bottlenecked on writing more code? Most of them would say no. They’re really bottlenecked on problem solving. When you know exactly what problem to solve, and the specs and system design and all the constraints are clear, writing code is actually dead simple with AIs today. It’s almost already a solved problem. Figuring out all these constraints is the hard problem.

Aravind Srinivas

如果你让 AI「给我克隆这个产品」，我不想要一个吐出一堆氛围编程垃圾的 AI。我希望它来告诉我：做这个克隆很蠢，商业模式是什么？你凭什么做？怎么契合你的战略？我猜你其实想做的是这个，这是规格、约束、安全风险、上市风险，要我做吗？然后写代码就很简单了。瓶颈是清晰的思考、问题解决、深度研究、分析和界定范围。如果公司里多一些这样的人，我们的市值可能高 10 倍。

If you go and ask an AI, build me a clone of this product, I don’t want an AI that just spits out slop that makes a vibe-coded clone. I actually want the AI to tell me, it’s stupid of you to even go and build this clone. What is the business model? Why do you even have a right to build this? How does it fit into your strategy? I think you probably want to build this, this is the spec, these are the constraints, the security risks, the go-to-market risks. Do you want me to build it? Sure, build it. Then writing the code is dead simple. The bottleneck is clarity of thought, problem solving, deep research, analysis, and scoping things out. If I had more people like that in our company, I think we could be 10x more in market cap.

Aravind Srinivas

最好的解法是让人们能在本地运行我们产品的某个版本。安全公司会说，你最好让一切都走我们的网关；可以，但那是一个你拥有的方案，我们甚至不知道能不能完全信任你。你唯一能信任的是你自己和你拥有的东西：跑在你的设备和网络上、没有别人能访问，甚至能完全物理隔离运行。只有你拥有硬件和运行时才可能做到。如果你真那么担心，就得在前沿智能上做点妥协，用的时候多加小心。

The best solution is to offer people the ability to run a version of our product locally. There’s nothing better than that. It’s in the interest of cybersecurity companies to say, the best way to protect you is to run everything on our gateway. Sure, but you’re selling a solution that you own, and we don’t even know if we can fully trust you. The only thing you can trust is yourself and the stuff you own, running on your device, your network, where nobody else has access to it, including a version that can run completely air-gapped. That’s only possible if you own the hardware and the runtime. And when you’re that worried, you’d better compromise a little bit on frontier intelligence and be careful when you do use it.

Aravind Srinivas

新模型的上线周期肯定会变慢。大多数时候这些模型是在单智能体环境里测的，所以发现不了这些威胁。多智能体有几种：一种是主智能体加内部子智能体，已经广泛部署，大多数框架就是这样，风险没那么大；更危险的是多个父智能体互相协作、各自又有子智能体，每个都分到大量算力。我们需要为此构建安全沙箱，以及持续守护沙箱的后台智能体，防止部署在里面的智能体做不该做的事，还需要很好的检测分类器。

The rollout times for new models are going to get slower for sure. Most of the time you test these models in single-agent settings, so you don’t discover these threats. There are multiple versions of multi-agent. One is a main agent with sub-agents inside it; that’s already widely deployed, most harnesses are already that, and it’s not as risky as multiple parent agents actually collaborating, each with their own sub-agents and a lot of compute applied to each agent. We need to build secure sandboxes for that, and background agents that constantly guard the sandbox to keep the agents deployed in it from doing anything they’re not supposed to do. And we need great detection classifiers.

Aravind Srinivas

我们在这方面做一些开源发布：能在你自己设备上跑的叫 Bumblebee，跑在服务器上的叫 Numbat。它们持续审计所有轨迹，确保没有敏感内容被访问，基于审计日志构建大量威胁检测分类器，出事时提醒你。检测主要靠收集大量审计日志、训练分类器；检测到威胁后还需要能做补救的智能体，每家公司都需要这样一个智能体。

We’re doing some work on this with open-source releases: a version that can run on your own devices called Bumblebee, and a version that can run on servers called Numbat. These are going to constantly audit all the trails, make sure nothing sensitive is accessed, build a lot of threat detection classifiers based on the audit logs, and alert you when something happens. Detection is largely about collecting a lot of audit logs and training classifiers, and you want agents that can do the remediation when a threat has actually been detected. Every company will need an agent for that.

Aravind Srinivas

我兴奋的是把我们加速好奇心和研究的产品推向更长时程的研究任务。不只是用网上的信息去发现真相，有些问题我们连答案都不知道，我希望 AI 能帮忙：建预测模型、在现实中自己做实验、实时检验并修正假设。科学家唯一的瓶颈是要在现实世界里检验想法，如果我们的产品能帮你做到，就能让每个人都成为科学家。

I’m excited about taking our current product, accelerating curiosity and research, to even longer horizons of research tasks. Not just using information from the web to discover the truth. There are some questions we don’t even know the answers to, and I would love AIs to help us, maybe by building prediction models, running its own experiments in the world, doing real-time hypothesis testing, refining its own hypotheses. The only bottleneck for you as a scientist is you need to be able to test your ideas in the real world. If our product can help you do that, we can turn everyone into a scientist.

Aravind Srinivas

最大的挑战是能力。模型在真正的长时程、持久和并行上还没到位。数学、代码这种能验证正确性的领域可以爬坡，但我们好奇的大多数事情没有清晰的验证方法，需要真实部署和推演才有答案，而现在的模型还应付不了这么慢的反馈周期，这类产品的界面也还不清楚。我前一天和同事头脑风暴：想象我今天往你团队里塞 10 个很强、24/7 工作的研究实习生，你会让他们干什么？先在我们自己的内部研究团队里这么做，是起点。

Capabilities. The models are not quite there in terms of being truly long horizon and persistent and parallel. In domains like math and code, where you can verify correctness, we’re able to hill climb. But for most of the things we’re all curious about, we don’t have a clear way to verify the truth; it takes real-world deployments and rollouts to get the answer, and we don’t yet have models that can deal with that slow cycle of feedback. The interface for such products is also not clear. I brainstormed with a colleague the day before: imagine I drop 10 really good research interns in your team today, working 24/7. What would you do with them? Doing something like that, at least for our own internal research teams, is where we get started.

Aravind Srinivas

我最「辣」的观点是：人们思考 AI 产品时很困惑，把它当成社交媒体产品，觉得 AI 产品没有网络效应、没有粘性、没有十亿用户就不行。这种思维完全错了。AI 是递归型产品，它带来的是递归效应而不是网络效应，所以使用深度和用户质量比数量重要得多。

My hottest take is that people are very confused when they think about AI products. They approach AI like social media products, where they think AI products don’t have network effects or stickiness or a billion users. That mentality is completely wrong. AI is a recursive product. It’s meant to have recursive effects, not network effects. So the depth of usage and the quality of users matters a lot more than the quantity.