Dario Amodei 谈 AI 未来：从生物医学到企业数据

太棒了。非常高兴能请到 Anthropic 的 CEO Dario Amodei。我们有一个非常棒的合作关系要谈。感谢你抽出时间。我想问你一个问题。你谈了很多关于你对 AI 未来的愿景。你写了《Machines of Loving Grace》。我想请你稍微总结一下你对 AI 未来的愿景。

Awesome. So, super excited to be here with Dario Amodei, CEO of Anthropic. We have a really awesome partnership that we're going to talk about. Thanks for taking time. I wanted to ask you a little bit. You've talked a lot about your vision of the future for AI. You wrote 'Machines of Loving Grace'. I thought if you could just a little bit summarize what's your vision of the future for AI.

首先，感谢你的邀请。关于 AI 的未来，我几个月前写了一篇文章叫《Machines of Love and Grace》。实际上，这源于一点挫败感：当人们谈论 AI 的好处时，我觉得他们不够有想象力或不够具体。很多人说‘耶，这很激动人心，我们干吧，有各种可能性，冲啊。’但当你真正深入细节时，我认为潜力是不可思议的。我写的一些内容受到我过去是生物学家的影响。所以我认为生物医学创新是能取得最大进展的领域之一。我逐一讨论了身心健康的各种领域，我们面临的许多疾病都是非常复杂的疾病，复杂性疾病，我认为 AI 真的会帮助我们治愈这些疾病。当我想到制药公司拥有的所有数据和所有问题时，我认为 AI 确实有潜力解决这些问题。这并不是说它不会在整个经济中做出惊人的事情。如果我们看看我们合作或将要合作的公司名单，我认为它会应用于整个经济。但我认为它会给社会带来难以置信的改变，甚至超越这些，人类将变得更健康，更有能力解决困扰我们几个世纪的问题。

Well, first of all, thanks for having me. In terms of the future of AI, I wrote this essay a few months ago called 'Machines of Love and Grace'. It was actually motivated by a bit of frustration where when people talked about the upsides of AI, I felt like they weren't being imaginative or specific enough. Lots of people were like, 'Yay, this is exciting. Let's do this. There's all kinds of let's go.' But when you actually get into the specifics, I think the potential is incredible. Some of the things I wrote about were colored a bit by the fact that I have been a biologist in the past. So I think of biomedical innovations as one of the areas where you can make the greatest progress. I just went through the various areas of physical and mental health and a lot of the diseases that we face now are really complex diseases, diseases of complexity, and I think AI is really going to help us to cure those diseases. When I think of all the data and all the questions that a pharmaceutical company has, I think there's really a potential for AI to address these issues. That's not to say that it won't do amazing things throughout the economy. If we look through the list of companies that we're working with or are going to work with in our partnership, I think it's going to be applied across the economy. But I think it's going to be an incredible change to society beyond even that, where humanity gets healthier and much greater in its capacity to address problems that have plagued us for centuries.

这太令人兴奋了。期待那个未来。你估计有多远？

That's super exciting. Looking forward to that future. What's your estimate? How far out is that?

我的意思是，这一切都非常不确定。我认为 AI 的特点是它发展得非常快，但没有根本原因说明它不会突然停止。所以我的猜测是，基础技术将在未来几年内实现，甚至可能就在未来一两年。问题在于，这些想法和技术需要多长时间才能推广到社会，真正产生所有好处。而这实际上取决于企业和市场推广的合作关系，比如我们现有的合作，让那些在某些情况下远离这项创新的经济部门赶上它，并将其应用到他们的领域。

I mean, it's all very uncertain. I think the thing about AI is it's moving very fast, but there's no fundamental reason that it couldn't just stop. So my guess really is that the fundamental technology is going to be there in the next few years, maybe even the next couple years. And there's a question of how quick is it to get those ideas and that technology out to society where it really produces all the benefits. And that honestly is going to come down to enterprise and go-to-market partnerships like the ones we have, getting the parts of the economy that in some cases are far away from this innovation to catch up to it and apply it to their areas of the world.

是的。所以我们必须让所有这些企业基于 AI 进行构建，在每个行业中利用它，然后他们还必须让他们的客户日常使用、采用它，并产生内部影响。所以所有这些事情可能需要一点时间，但希望我们能加速。

Yeah. So we have to get all these enterprises to build on that AI, leverage it in each of their industries and then they have to get also their customers to use it day-to-day, adopt it and then have the impact that it has inside. So it might take a little bit of time for all of those things to happen but hopefully we can accelerate it.

尽可能加速。是的。

Accelerate as much as possible. Yes.

那么数据对这些企业有多重要？

So how important is data in this for these enterprises?

数据非常重要。有不同类型的数据。有我们用来训练模型的数据，这对制作基础模型显然非常重要。但还有特定于某个行业、公司或客户的数据，我认为这也非常重要。有很多东西，再次以制药公司为例，但金融、科技公司和生产力软件也是如此。他们拥有大量专有数据，其中包含其他地方没有的信息。模型无法知道它无法访问的事实。所以我认为它在很多方面都很重要。有时你会想要在数据上进行微调。有时你会希望模型能够对数据进行操作。有时你会想通过 RAG 等方法将数据拉入上下文。有时你会希望有智能体对数据进行操作。所以数据在很多方面都是必不可少的，我认为它对于在我们这样的 AI 公司和拥有数据的企业之间建立独特价值至关重要。这是一种建立不可替代价值的方式。我认为我们合作的重点之一就是帮助更有效地将这两者结合起来。

So data is super important. There are different kinds of data. There is the data we use to train the models, and that is obviously very important for making the base model. But then we get to data that is specific to a particular industry or a particular company or customer, and I think that's super important as well. There's a lot of stuff, again if we go to our example of pharma companies, but it's just as true of finance, it's just as true of tech companies and productivity software. They have a lot of proprietary data where there's information that's simply not contained anywhere else. The model can't know facts that it doesn't have access to. So I think it's important in a number of ways. Sometimes you'll want to fine-tune on the data. Sometimes you'll want the model to be able to act and operate on the data. And sometimes you'll want to pull the data into context as is done with methods like RAG. Sometimes you'll want to have agents that operate on the data. So there's a whole bunch of different ways in which that data will really be essential, and I think it'll be essential for building unique value between the AI companies like us and the enterprises which have that data. That's a way to build unique value that isn't substitutable on either side. And I think a lot of what our partnership is about is helping to bring those two things together more efficiently.

是的，这很有道理。那么在未来，如果 AI 会产生这么大的影响，组织拥有数据有多重要？当他们想要为未来做准备，想要建立能够支撑他们的创新时，他们应该如何思考数据本身？

Yeah, that makes a lot of sense. So in the future, if this is going to have this much impact with AI, how important is it for organizations to have data? How much should they, as they want to prepare themselves for the future, they want to build innovations that can sustain them, how should they be thinking about data themselves?

是的，再次强调，我认为这绝对至关重要。随着 AI 模型在几乎所有方面都变得更好，我认为数据代表了公司收集的知识和智慧，这是与 AI 模型能力协同的主要因素之一。无论你有多聪明，如果你什么都不知道，无法接触世界，你能做的也有限。所以是的，我认为这将绝对关键。

Yeah, again, I think it's absolutely essential. As the AI models get better at just about everything, I think that data represents collected knowledge and wisdom from the company that is one of the main things that is going to be synergistic with what the AI models are capable of doing. It doesn't matter how smart you are, if you don't know anything and you don't have access to the world, there's only so much you can do. So yeah, I think it's going to be absolutely critical.

完全同意。所以我看到很多企业正在做的用例，他们想用 AI 做很多事情，但我觉得最有趣的是那些拥有一些特殊数据、这些数据对客户来说是独一无二的、他们一直在收集的。这将非常有价值，我认为这能保护他们，也是他们应该进行创新的地方。他们应该使用你们提供的模型，结合他们的数据，然后提出那些创新。我认为这就是你们将产生影响的地方。你提到了一堆东西。你提到了 RAG，你提到了智能体。

Yeah totally. So I see a lot of use cases that enterprises are doing and they want to do so much with AI but the ones that I think are the most interesting are where they have some special data that's unique to their customers that they've been collecting. That's going to be really valuable and that's I think what protects them and that's where they should do the innovation. They should use the models that you guys provide together with their data and then come up with those innovation. I think that's where you're going to have this impact. You mentioned a bunch of things. You mentioned RAG, you mentioned agents.

我很好奇智能体的未来是什么？它会是什么样子？以及如何让这些智能体在 Databricks 中的企业专有数据上运行？

I'm curious what's the future of agents? What's that going to look like? And how can you get these agents to operate on this enterprise proprietary data that they have in Databricks?

是的，我的看法是 AI 的未来主要是智能体。所以我们会越来越多地让我们的模型以智能体的方式运行。我们已经发布了一些早期的智能体。去年年底有 Computer Use，模型可以接收电脑截图并执行操作。我们最近发布了 Claude Code，这是一个执行代码的早期智能体。当然，我们通过 API 提供的很多模型也可以连接起来并教会它们作为智能体运行。所以随着模型变得更聪明，我们构建更多这类专门化的东西，我认为智能体会变得更加普遍。智能体会使用工具，其中一个工具可以理解为以各种方式访问数据。通过数据库访问数据，通过各种搜索访问数据。我认为我们的合作在这方面会有很多协同效应。

Yeah, I mean my view is the future of AI mostly is agents. So we are increasingly going to have our models operate as agents. We've already released some early agents. There was computer use late last year where the models can take in screenshots from a computer and take actions. We recently released Claude Code, which is an early agent that performs code. And of course, a lot of the models we make via API can be linked up and taught to operate as agents. So as the models get smarter and we build more of these specialized things, I think agents are going to be much more of a thing. Agents are going to use tools, and one of the tools they use can be thought of as accessing data in various ways. Accessing data through databases, accessing data through various kinds of search. I think there's going to be a lot of synergy in our partnership with that.

是的，非常兴奋。特别是，我知道我们有客户同时关注我们两家。AT&T 很令人兴奋，因为他们无线系统上有大约 1.8 亿用户，他们正在用我们来检测欺诈。所以这是一个令人兴奋的用例。

Yeah, super excited. And in particular, I know we have customers that are looking at us both together. AT&T is kind of exciting because they have I think 180 million subscribers on their wireless systems and they're using us to detect fraud. So that's an exciting use case.

另一个非常棒的用例是 Square，或者叫 Block 的公司。他们以前有那些 UI 界面，你必须配置如何设置你的商店以及你要买什么。现在有了生成式 AI，他们可以全部配置好。你只需对着它说话，它就能自己设置好，让他们的客户轻松很多。所以对我们将看到的这些用例感到非常兴奋。

Another really awesome use case is Square, or the company called Block. They used to have these UI interfaces where you have to configure how you want to set up your store and what you're buying. And now with generative AI, they can just configure it all. You can just speak to the thing and it can just set itself up, making it so much easier for their customers. So excited about all these use cases that we're going to see.

关于这次合作，我非常兴奋能原生地让我们的客户在 Databricks 内部访问 Claude 的模型。当然，治理和安全对他们来说非常重要，所以他们会使用 Unity Catalog 并确保安全。但我很好奇，你认为客户对此以及同时使用我们两家有什么看法？

So about this partnership, I'm super excited to natively let our customers access Claude's models inside Databricks. Of course, governance and security are super important for them, so they'll be using Unity Catalog and securing it. But I'm curious, what do you think customers are saying about this and using us together?

是的，我认为我们听到的，可能你们也听到了，就是将所有东西放在一个地方带来的巨大便利。所以想法是，你拥有这些模型和你的数据。客户在很多方面都面临这个问题。他们可能在不同的云上，或者这些东西以不同的方式存储。没有一个边界让数据和进入数据的模型都包含在这个边界内，管理起来非常头疼。所以我认为你们所做的，是让我们的模型更容易与现有的其他组件配合，并让一切在这个数据边界内运行。所以我认为这对人们来说是一个巨大的好处。

Yeah, I think what we're hearing, and probably you're hearing it as well, is that there's this incredible convenience to having everything all in one place. So the idea is you have these models and you have your data. Customers face this in a whole bunch of ways. They might be on a different cloud or these things are stored in a different way. The idea that there's not one boundary where the data and the models that go into the data are all contained within the boundary, it's just a huge headache to have to manage that. So I think what you guys are doing is making it easier to take our model and fit it in with the other ingredients that are present, and make everything run within this data boundary. So I think that's a huge benefit for folks.

是的，非常兴奋能让数据和迭代之间的切换变得超级无缝和简单。因为如果你想提高质量，让智能体基于那些数据进行推理，你可能需要迭代。所以那是内部开发循环。

Yeah, super excited to make that super seamless and simple, going between the data and iterating. Because if you want to improve the quality, make the agents reason on that data, you probably want to iterate through that. So that inner dev loop.

就是迭代循环，对吧？就是开发循环时间。每个开发者都知道，缩短迭代循环时间，很多事情都会更快发生。

It's the iteration loop, right? It's the dev loop time. Every developer knows that you get that iteration loop time down and lots of things happen faster.

完全正确。是的。所以对此非常兴奋。你们刚刚发布了 Sonnet 3.7、Claude Code 以及许多其他创新。能跟我们聊聊这些吗？

Exactly. Yeah. So super excited about that. So you guys just released Sonnet 3.7, Claude Code, and lots of other innovations. Can you tell us a little bit about these?

是的。关于 Sonnet 3.7，我们发布了它。截至发布时，距离我们发布已经大约两个月了。反响真的非常惊人。Sonnet 3.5 和后来追溯命名的 Sonnet 3.6 在编程方面都代表了显著的进步，而 Sonnet 3.7 代表了又一次显著进步。但我们觉得，通过这一步，模型已经超过了一个阈值，能够端到端或自主地完成某些编程任务，这比之前可能的要强。所以我们看到，在广泛的客户以及 Claude Code（我稍后会谈到）中，使用量大幅增长，也出现了以前没有的新用例。它甚至催生了一个术语，我相信是 Andre Karpathy 提出的，叫做“氛围编程”，就是你只需跟模型说话，让它编程。所以不必精确，你给个氛围。我认为这仍然有很多不完美之处，还有很大的改进空间，但我认为从这一点开始的曲线会非常快，人们会对即将推出的模型的能力感到惊讶。

Yes. So Sonnet 3.7, we released it. As of time of release, it'll have been about two months since we released it. The response has really been incredible. Both Sonnet 3.5 and the retroactively named Sonnet 3.6 represented a significant step up in coding, and Sonnet 3.7 represented another significant step up. But we felt like with this step, the model had passed a certain threshold where it was able to do certain coding tasks end to end or on its own that was greater than what was possible before. And so we saw with a wide range of customers and with Claude Code, which I'll get to, we saw huge upticks in usage as well as new use cases that hadn't been present before. It even led to the coining of this term by, I believe it was Andre Karpathy, called "vibe coding," which is you just talk to the model and you let it code. So instead of having to be precise, you give the vibes. I think there are still a number of imperfections to that and there's a large amount of improvement, but I think the curve from this point onward is going to be pretty fast and people are going to be surprised at the power of the models that will come out soon.

是的，我们确实看到了。所以跟我们聊聊 Claude Code 吧。

Yeah, and we're seeing it right. So tell us a little bit about Claude Code.

是的，Claude Code。Claude Code 基本上是一种使用 Claude 3.7 Sonnet 的特定方式。它有点回到命令行工具的复古时代。所以想法是，它是一个命令行工具，易于与 Git 之类的东西集成，你只需说：“我要和 Claude 对话，让它写我的 PR 并执行这些操作。”它就能直接行动。它会请求你的许可来实际在电脑上执行操作，但之后就会为你行动。所以你可以把它看作一个非常早期的智能体。同样，界面非常简单，但我们发现，在发布 Claude Code 的几天内，就有近 10 万人尝试了它。它确实被采用了。我们与希望使用它的企业进行了交流。所以它真正展示了编程模型和智能体的力量。

Yeah, Claude Code. So Claude Code is basically a particular way to use Claude 3.7 Sonnet. It kind of goes back to the retro era of command line tools. So the idea is that it's a command line tool that's easy to integrate with things like Git, and you just say, "I'm going to talk to Claude and basically ask it to write my PR and take these various actions." And it can just act. It'll ask for your permission to actually take action on the computer, but it'll just act for you. So you can think of it as a very early agent. Again, the interface is pretty simple, but we found very quickly within a few days of releasing Claude Code, we had nearly 100,000 people trying it out. And it's really been adopted. We've talked to enterprises who want to use it. So it really shows the power both of coding models and of agents.

是的，这非常令人兴奋。是的，我们在 Databricks 中也有这些，它已经改变了人们如今使用 Databricks 的方式。他们不再需要自己编写代码。它只是一个基本部分。我的意思是，我现在什么都大量使用助手。你知道，像 Cursor、Kodium 等等。而且我认为很多这些用例都把它隐藏在后台。

Yeah, it's super exciting. Yeah, we have these in Databricks as well, and it's kind of changed how people use Databricks these days. They no longer have to write the code themselves. It's just an essential part. I mean, I never these days leverage the assistant heavily for everything. And you know, it's things like Cursor, Kodium, and so on. And I think a lot of these use cases drop it behind the hood.

MCP，也就是模型上下文协议，看起来是一个非常令人兴奋的方式。我的意思是，这些模型需要能够获取它们的上下文，需要获取数据，需要能够从其他地方获取函数和提示。这背后的想法是什么？有什么选择？

MCP, so Model Context Protocol, seems like a very exciting way. I mean these models need to be able to get their context. They need to get data. They need to be able to get functions prompts from elsewhere. What's the thinking behind it? And what's the option?

我们觉得在模型本身和它们需要做的常见事情之间，存在某种缺失的环节或缺失的组件。模型需要以某种方式集成到某个工具中，使用某个工具，访问某些数据。这是一件重复性的事情，但每次都不完全相同。精神上相同，但字面上总是有些不同。所以我们想，这不正是 Claude 本身可以解决的好问题吗？因此，MCP 就像是连接模型与模型需要访问和使用的东西（比如数据或工具）的胶水。这真的很有趣。我们几个月前发布了 MCP，很多人立刻就看到了它并喜欢上了它。但我要说，就在最近几周，由于我们也不完全清楚的原因，它的兴趣度突然爆发了，几乎成了社交媒体上的一个梗。我认为这说明了人们从中发现了实用性。

We felt like there was kind of a missing link or a missing widget between the models themselves and common things that they need to do. The models need to somehow integrate themselves into some tool, use some tool, access some data. This is one of these things that is repetitive, but it's never quite the same twice. Spiritually the same, but literally always something different. So we thought, isn't that a good problem for Claude itself? And so MCP is kind of this glue that connects the model to things that the model needs to access and use, like data or tools. It was really interesting. We released MCP a few months ago and many people saw it right away and loved it. But I would say just in the last few weeks, for reasons not entirely clear to us why it happened exactly now, it's kind of exploded in interest and almost become like a meme on social media. I think it speaks to the utility that people have found from it.

是的，我们当然在采用它，并通过 MCP 将 Databricks 中的所有信息暴露给大语言模型。所以这很令人兴奋，而且你们把它开源了也很酷。我的意思是，你们本可以把它只作为 Anthropic 的东西。

Yeah, we're certainly adopting it and exposing all the information that's in Databricks to the LLMs using MCP. So, it's exciting and it's also cool that you guys open sourced it. I mean, you could have just had it be an Anthropic thing.

是的，我们把它看作有点像 USB-C 之类的东西，一种 AI 时代的连接器，一种灵活可塑的胶水。所以我认为它将在我们的合作中发挥重要作用，因为它把我们做的事情和你们做的事情结合在一起。所以我认为它肯定会有价值，我们希望它能在整个行业中被广泛采用。

Yeah, we thought of it as a little bit like USBC or something, like a kind of connector for the AI age, a kind of flexible moldable glue. So I think it's going to play an important part in our partnership because it brings together the stuff that we do with the stuff that you do. So I think it'll definitely be valuable and we hope it's widely adopted across the industry.

我很好奇，想换个话题，谈谈治理、安全和隐私，这是我们两家公司都非常关心的事情。为了促成这次合作，我们讨论了很多。你从客户那里听到了关于治理的什么反馈？

I'm curious about switching gears to governance and security and privacy, which is something both of our companies care a lot about. We've discussed it so much to get this partnership going. What are you hearing from the customers around governance?

是的，数据治理、数据安全、数据隐私在受监管的行业中尤其重要。所以我们一开始谈到的所有生物医学奇迹，为了让模型真正开发药物、使用专有数据和分析临床试验，显然让模型的智力能力达到能够提供帮助的水平是一个巨大的挑战。但正如商业和工程中的许多事情一样，常常阻碍你的是那些对技术人员来说看似愚蠢的事情，但从商业角度来看却非常重要，那就是：我们能确保数据不会泄露吗？我们能确保遵守所有必要的法规吗？通常，技术问题无论看起来多难，我们都能解决。但如果我们不解决这些看似愚蠢的问题，它们可能会阻碍我们多年，从而也将这些好处和疾病治疗方法推迟多年。所以我认为把这些问题处理好，以正确的方式治理数据，非常重要。正如你所说，这是我们两家公司共同关注的事情。这有两个部分：一部分是技术部分，技术解决方案；但另一部分是，我们的客户必须信任我们。

Yeah, data governance, data security, data privacy is particularly important especially in regulated industries. So all the biomedical wonders that we talked about at the beginning, in order to actually get the models developing drugs and using proprietary data and analyzing clinical trials, obviously it's a big challenge to get the intellectual capability of the models to where they can help. But as with many things in both business and engineering, often the things that hold you up are the things that to a technologist seem kind of dumb, but that are very important from a business perspective, which is: can we make sure that this data doesn't leak? Can we make sure that we're complying with all the necessary regulations? Often the technical problems, as hard as they seem, we'll solve them. But if we don't address these seemingly dumb problems, they can be the thing that holds us up for years and therefore holds up these benefits and cures to diseases for years. So I think getting these things right, governing the data in the right way, is really important. As you said, that's something that both of our companies share. There's two parts to this: one is just the technical part, the technical solutions, but there's another part which is that our customers have to trust us.

他们必须相信我们在做正确的事情，并且我们可以被信任处理他们的数据。我认为我们两家公司都建立了值得信赖的声誉。我很兴奋我们能一起，两家拥有这种声誉的公司，不仅提供产品，还提供人们可以信任的组合。

They have to trust that we're doing the right thing and that we can be trusted with their data. And I think both of our companies are companies that have cultivated a reputation of trustworthiness. I'm excited for us to together, two companies with this reputation, present not just a product offering but a combination that people can trust.

是的，对此非常兴奋。实际上，正如你所说，我们作为技术人员可能不会去想这个。我当然没想到我会创办一家专注于治理的公司，但几年前我们推出了一个叫 Unity Catalog 的东西，它帮助你进行治理，现在它已经成为我们做的第一要务。我的意思是，它是我们研发的第一优先级，大部分人力都投入在那里。谁知道呢？但我认为，任何与数据相关的事情，另一面就是隐私和治理。所以从某种意义上说，我们也是一家隐私和安全公司。所以这很有趣。

Yeah, super excited about that. And actually, you said it, us as technologists we might not be thinking about it. I certainly didn't think I'm starting a company that's going to focus on governance, but a few years ago we launched this thing called Unity Catalog that helps you do governance and it's now become the number one thing we do. I mean it's like our number one R&D priority. It's where most of the headcount goes. So who would have known? But I think anything you do with data, the other side of it is privacy and governance. So in a way, we're a privacy and security company too. So it's interesting.

我想花点时间问你一些有趣的话题。你以前被问过这些。我很好奇，开放模型与封闭模型。关于这些，有一个永无止境的问题。未来会怎样？哪个会赢？哪个会存在，哪个会消失？

I want to take a second and ask you about some of the interesting topics that are out there. You've been asked about these before. I'm curious, open models versus closed models. There's the ever never ending question around these. How what's the future going to be? Which one's going to win? Which one that we're going to have, which one's going to go away.

是的。所以我认为从商业角度以及安全角度来看，我一直觉得开放与封闭的区别有点被夸大了。我认为今天的模型还没有我担心的那种风险，尽管我认为那些风险正在迅速到来。我认为在那个世界里，会有封闭模型和开放模型的繁荣生态系统。我认为它们都对世界有益。它们都以不同的方式推动科学和行业的发展。

Yeah. So I think from both a business perspective and a safety and security perspective, I've always seen the open versus closed distinction as a little bit overblown. I think today's models don't have the kinds of risks that I've been worried about, although I think those risks are coming quickly. And I think in that world, there's a flourishing ecosystem of closed models, open models. I think they're all good for the world. They all advance science and the industry in different ways.

随着模型越来越强大，这些风险就会出现——我们经常担心的国家安全风险。但归根结底，无论是开源还是闭源模型，模型都非常强大，我们必须以某种方式应对这些风险。也许开源和闭源模型的应对方式略有不同，但我从未认为这是最重要的因素。最重要的因素是，这个领域发展非常迅速，带来了许多好处，也存在一些风险，我们需要应对这些风险。

As we get to more powerful models, these risks will be present—the national security risks we often worry about. But at the end of the day, whether it's open or closed models, the models are very powerful, and we have to contend with these risks one way or another. Perhaps the way we need to contend with them is a little different for open versus closed models, but I've never really seen that as the most important factor. The most important factor is that the field is advancing very quickly, has many benefits, and has some risks that we need to deal with.

所以我们很可能会有开源模型，对吧？

So we probably will have open source models, right?

我认为在市场的不同领域和不同能力水平上，这两种模型都会继续存在。

I think there will continue to be, in different parts of the market and at different capabilities, both kinds of models.

如果我们会有开源模型，就会有人推出它们，比如 DeepSeek 等等。那么，至少这些模型在这里开发可能会更好。

If we're going to have open source models, someone's going to put them out like DeepSeek and so on. Then probably it's better if at least those models are developed here.

在所有条件相同的情况下，我认为这很可能是正确的。

All things equal, I think that's probably correct.

很有道理。

Makes a lot of sense.

推理模型——这实际上是重点。与经典的预训练缩放定律相比，你们在这方面投入了多少精力？

Reasoning models—that's the focus actually. How much of that is a focus for you versus classic pre-training scaling laws?

我认为两者都是我们的重点。我们看到推理模型在缩放定律上表现出强劲的回报，但普通的预训练也在以有希望的方式继续扩展。所以我们实际上同时看到了两者。特别是我们对推理模型的看法——我们在这方面体现了一些特点。对于推理模型，我们不是第一个进入市场的，但我们认为我们做对了。推理模型作为一种独立的东西——有时你调用推理模型，有时调用非推理模型，对于简单的问题你调用推理模型，结果它思考了 20 分钟，而它本来可以在两秒钟内回答——我认为这有点尴尬，不是正确的做法。所以我们引入了我们所谓的混合推理模型，这种模型你可以告诉它思考多长时间。你可以为同一个模型、同一组权重打开或关闭推理。这就像人类一样。我没有一个大脑用于简单问题，另一个大脑用于困难问题。我只有一个大脑，我会思考需要多仔细地回答某个问题。

I think both are a focus for us. We're seeing that reasoning models are showing strong returns to the scaling laws, but also ordinary pre-training continues to scale up in a promising way. So we're really seeing both at once. Our take on reasoning models in particular—we've embodied a little bit of this. With reasoning models, we were not first to market, but we think we've done it right. The idea of reasoning models being a separate thing—sometimes you call the reasoning model, sometimes the non-reasoning model, and you call the reasoning model for an easy problem and it takes 20 minutes thinking when it could have answered in two seconds—I think that's a little awkward and not the right way to do it. So we've introduced what we've called hybrid reasoning models, which are models that you can tell how long to think. You can turn reasoning on and off for the same model, the same set of weights. It's just like humans. I don't have one brain for easy questions and another for hard questions. I just have one brain, and I think about how carefully I need to answer a given question.

我有多长时间来回答下一个问题？

How much time do I get to answer this next question?

正是如此。没错。

Exactly. Right.

很有道理。我很好奇，推理模型的缩放定律呢？它们不同吗？我们对它们了解多少？现在说还为时过早吗？

It makes a lot of sense. I'm curious, what about scaling laws for reasoning models? Are they different? How much do we know about them? Is it too early to say?

我们在缩放定律中看到的一般动态是，被缩放的具体内容会不时地发生一点变化。我们对 Transformer 架构有一些小的改动。推理模型是一种略有不同的训练类型，但它们基本上是同一回事。我们过去使用基于人类反馈的强化学习；我们仍然进行这种推理训练。这只是目标略有不同的强化学习。所以我们总是在缩放，但缩放的东西会有一点变化。现在，正如我们之前有时看到的，我们经常同时缩放多个东西，并将它们叠加在一起。所以，随着小的变化，缩放定律基本上是不可战胜的。它们继续有效。

The general dynamic we've seen with scaling laws is that exactly what is being scaled changes a little bit from time to time. We have little changes to the transformer architecture. The reasoning models are a slightly different type of training, but they're basically the same thing. We used to do RL from human feedback; we still do this kind of reasoning training. It's just RL with a slightly different objective. So we're always scaling, but the thing we're scaling changes a little. Now, as we've seen at times before, often we're scaling more than one thing at a time and stacking them on top of each other. So with small changes, the scaling laws are kind of undefeated. They continue.

那泛化能力呢？在预训练方面，当你们推出新模型时，它在许多不同的任务上全面表现更好。对于推理模型，它的泛化能力是否和预训练一样好，还是有所不同？

What about generalization? It seemed on the pre-training side, when you got the new model out, it was just better across the board on so many different tasks. On reasoning models, does it generalize as well as pre-training did, or is it different?

这取决于你缩放什么。我们在发布 Claude Sonnet 3.7 时说过一件事，我们对很多人用来通过基准判断的数学和编程竞赛不太感兴趣。原因之一是这些任务相当狭窄和专门化，它们泛化得不多。我们专注于一些更广泛的任务，发现虽然泛化并不完美，但这导致了比我们在数学竞赛中看到的更好的泛化。所以这又是一个我们不是第一个进入市场，但我们认为我们做出了最好东西的领域。

It kind of depends what you scale. One of the things we said when we released Claude Sonnet 3.7 is that we were less interested in these math and coding competitions which a lot of people use to judge via benchmarks. One of the reasons is that those are pretty narrow and specialized tasks and they don't generalize that much. We focused on some broader tasks and found that while the generalization is not perfect, that leads to better generalization than we've seen in the math contest. So that is again another area where we were not first to market, but we think we've produced the best thing.

非常感谢 Dario 抽出时间并投入如此多的精力共同推进这次合作。我们对此非常兴奋。如果你想使用云端模型，现在可以在我们的模型服务、智能体框架、向量搜索和 RAG 中使用它们。所以对此非常兴奋。

Thank you so much, Dario, for taking the time and spending so much energy jointly on this partnership. We're very excited about it. If you want to use the cloud models, you can now use them in our model serving, in our agent framework, in vector search, RAG. So super excited about this.

谢谢你邀请我。我对这次合作非常兴奋。

Thank you for having me. I'm very excited about the partnership.