AI 的双重影响：Dario Amodei 谈生产力与就业替代

好的。欢迎大家来到 Journal House，也热烈欢迎在线观看的观众。但最重要的是，热烈欢迎 Anthropic 的 CEO Dario Amodei。

Very well. Welcome everybody. Welcome to Journal House and a big welcome to our audiences joining us online. But above all, a big welcome to Dario Amodei, the CEO of Anthropic.

谢谢邀请。

Thank you for having me.

别客气。Dario，我们在达沃斯，有很多事情发生，但我想从一个宏观问题开始。我感觉去年这个时候，每个人对 AI 都非常兴奋，都在谈论 AI 能做什么、它的潜力和能力。而今年，讨论似乎转向了 AI 正在对世界做什么。我知道你对这些问题思考很多。所以我的问题是：你认为企业、政策制定者、政府是否在为影响做足够的准备？

Not at all. So Dario, we're at Davos. There's a lot going on, but I wanted to start with a big picture question. It feels to me that this time last year, everybody was very excited about AI and talking about what AI can do, its potential, its capabilities. It feels as though the debate has shifted this year to be more about what is AI doing to the world. I know you think a lot about these things. So my question is: do you think businesses, policymakers, governments are doing enough to prepare for the impact?

没有。我现在详细解释一下。我观察这个领域 15 年，参与其中 10 年。我注意到一个现象：技术发展轨迹出奇地平滑，而公众舆论和反应却剧烈摇摆，体现在两个方面。一是技术能力。每三到六个月，极性就会反转：媒体极度兴奋，说技术能改变一切；然后又说全是泡沫，要崩溃了。我看到的是平滑的指数曲线，类似于算力的摩尔定律——我们基本上有了智能的摩尔定律，模型每几个月认知能力就提升一次。这种进步是持续的。起起落落——'我们发明了新东西'、'全要崩了'、'撞墙了'、'要疯了'——那是公众感知现象。在能力方面，我认为技术好坏也有类似的极性摆动。2023 和 2024 年，人们对 AI 有很多担忧——AI 要接管了，谈论 AI 风险、AI 滥用。然后 2025 年，政治风向转向 AI 机遇，现在又转回来。在这整个过程中，我和 Anthropic 试图采取的态度是一贯的，即存在平衡——一种非常奇怪的平衡，因为我认为技术的能力非常极端，但其正面和负面影响都存在。大约一年半前，我写了一篇文章《Machines of Loving Grace》，对 AI 的好处持非常激进的观点——它能帮助治愈癌症、根除热带疾病、为世界未开发地区带来经济发展。我的观点没有变；我相信所有这些。但另一面，我现在正在写更多，可能很快会发布一些内容，就是坏事也会发生。以经济方面为例。我的观点是，这项技术的标志是它将把我们带入一个 GDP 增长非常高、但失业率和不平等也可能非常高的世界。这种组合我们几乎从未见过。高 GDP 增长通常意味着有很多事可做，人人有工作。我们从未有过如此颠覆性的技术。GDP 增长 5% 或 10% 的同时失业率也达到 10%，这在逻辑上并不矛盾；只是以前从未发生过。我既兴奋又担忧。以 AI 编程为例。我们最新的模型 Claude Opus 4.5 发布后，Anthropic 的一些工程负责人告诉我，他们不再写任何代码了——他们让 Opus 做工作，自己编辑。我们刚刚发布了 Claude Co-work，这是我们的工具 Claude Code 的非编程版本，几乎完全用 Claude Opus 在一周半内构建完成。软件工程师仍然有事可做——即使只做 10%，他们仍有工作，或者可以提升层次——但这不会永远持续。模型会做得越来越多。这是一个缩影：惊人的生产力，软件变得便宜，可能基本免费。需要将软件分摊到数百万用户的前提可能不再成立——这次会议，可能只需几美分就能做一个让人们互相交流的应用。它可能非常灵活且可回收。但与此同时，几十年来建立起来的整个工作和职业可能不复存在。我认为我们可以应对，可以适应，但我认为人们完全没有意识到即将发生的事情及其规模。

No. I'll explain the longer version now. I've been watching this field for 15 years and been in it for 10. One thing I've noticed is that there's been a surprisingly smooth trajectory, whereas public opinion and reaction have oscillated wildly in two ways. One is the capabilities of the technology. Every three to six months, we have a reversal of polarity: the media is incredibly excited about what the technology can do—it's going to change everything—then it's all a bubble, it's all going to fall apart. What I see is a smooth exponential line, similar to Moore's Law for compute—we basically have a Moore's Law for intelligence, where models get more cognitively capable every few months. That march has been constant. The ups and downs—'we invented a new thing,' 'it's all going to crash,' 'it's hitting a wall,' 'it's going to go crazy'—that's a public perception phenomenon. On the capability side, I think there's a similar thing on the polarity of whether the technology is good or bad. In 2023 and 2024, there was a lot of concern about AI—AI is going to take over, talk about AI risk, AI misuse. Then in 2025, the political wind shifted to AI opportunity, and now it's shifting back. Throughout all this, the approach I and Anthropic have tried to take is one of constancy, saying there is balance here—a balance of a very strange form, because I think the technology is very extreme in what it's capable of, but its positive and negative impacts both exist. I wrote an essay, 'Machines of Loving Grace,' about a year and a half ago, with a very radical view of the upside of AI—that it would help cure cancer, eradicate tropical diseases, bring economic development to parts of the world that haven't seen it. My view hasn't changed; I believe all those things. But the other side, which I'm writing more about now and may release something soon, is that bad things will happen as well. Take the economic side as one risk. My view is the signature of this technology is it's going to take us to a world with very high GDP growth and potentially also very high unemployment and inequality. That's not a combination we've almost ever seen before. High GDP growth usually means lots of stuff to do, lots of jobs for everyone. We've never had a technology this disruptive. The idea that we could have 5 or 10% GDP growth but also 10% unemployment is not logically inconsistent; it's just never happened before. I'm excited and worried for both reasons. Take AI coding as an example. With our latest model, Claude Opus 4.5, some engineering leads at Anthropic have told me they don't write any code anymore—they just let Opus do the work and edit it. We just released Claude Co-work, a version of our tool Claude Code for non-coding, built in a week and a half almost entirely with Claude Opus. There are still things for software engineers to do—even if they're only doing 10%, they still have a job, or they can level up—but that won't last forever. Models will do more and more. This is a microcosm: incredible productivity, software becoming cheap, maybe essentially free. The premise that you need to amortize software across millions of users may become false—for this meeting, it might cost a few cents to make an app for people to talk to each other. It may be very flexible and recyclable. But at the same time, whole jobs and careers built over decades may not be present. I think we can deal with it, adjust to it, but I don't think there's awareness at all of what is coming and its magnitude.

你这么说很有意思。那么你认为，在一个高 GDP 增长但高失业率的世界里，这对社会有什么影响？你说人们现在没有考虑这个问题。你能给出具体的例子，说明社会可能如何组织自己来适应这样的世界吗？

So that's so interesting when you say that. So how do you think in a world of high GDP growth but also high unemployment, what does that do to society? And you say people aren't thinking about it now. Can you give concrete examples of how society might organize itself to adapt to such a world?

是的，我认为有几件事。我们做的第一件事，与其说是解决方案，不如说是第一步，就是我们有一个叫做 Anthropic 经济指数的东西。我们已经有了大约一年，更新了四五次。它是一个实时指数，让你追踪我们的模型 Claude 被用于什么。它遍历所有对话，以隐私保护的方式使用 Claude 来统计查询 Claude 的使用情况。

Yeah, so I think there are a few things. The first thing we've done, which is not a solution so much as a first step, is we have this thing called the Anthropic Economic Index. We've had it for about a year, updated it four or five times now. What it does is a real-time index that lets you track what our model Claude is being used for. It goes across all conversations and uses Claude in a privacy-preserving way to statistically query how Claude is being used.

它被用于哪些任务？在多大程度上是自动化还是增强任务？它被用于哪些行业？它如何在美国各州和世界各国扩散？我们只是不断添加更多细节。我的观点是，除非我们能衡量这场经济转型的形态，否则任何政策都将是盲目的、误导的。许多政策之所以出错，是因为它们基于根本错误的假设。所以这是第一步。第二步是，我认为我们需要非常仔细地思考如何让人们适应。人们可以适应得更快或更慢。这可能意味着在现有工作中使用这项技术，或者从一份工作转到另一份工作。例如，我认为实体世界的工作可能会更多，而知识工作经济中的工作会更少。也许最终机器人技术会取得进展，但那是一条较慢的轨迹。所以这是其一。有些工作仍然重视人的触感吗？有些是，有些不是。我们可能会发现这在市场上有多重要，以及在哪里最重要。在公司层面，当软件变得廉价，随后知识工作的其余部分也变得廉价时，护城河是什么？我们不知道。我们从未真正问过这个问题。我们以某种方式思考过护城河。所以公司层面将会有巨大的争夺。教人们适应，教他们期待什么，是第二步。第三步是，在如此宏观规模的大规模失业中，政府需要扮演某种角色。我只是看不到它不发生。蛋糕会变得更大；钱会在那里。预算可能会平衡，因为我们什么都不做，因为有如此多的增长。问题在于将其分配给正确的人。所以我认为现在可能是一个更少担心抑制增长、更多确保每个人都能分享增长的时候。我知道这与当前的主流情绪相反，但我认为技术现实即将改变，迫使我们的想法改变。

What are the tasks it's being used for? To what extent is it automating versus augmenting tasks? What industries is it being used in? How is it diffusing across states in the United States and countries in the world? We've just added more and more detail here. My view is until we can measure the shape of this economic transition, any policy is going to be blind and misinformed. Many policies have gone wrong because they're based on fundamentally incorrect premises. So that's step one. Step two is I think we need to think very carefully about how to allow people to adapt. People can adapt more quickly or more slowly. This can mean adapting to use the technology within existing jobs, or adapting from one job to another. For example, I think there will probably be more jobs in the physical world and fewer in the knowledge work economy. Maybe eventually robotics makes progress, but that's on a slower trajectory. So that's one. Are there jobs that still value a human touch? Some do, some don't. We may find out how important that is in the market and where it's most important. At the company level, what are the moats when software becomes cheap and then the rest of knowledge work becomes cheap? We don't know. We've never quite asked that question. We've thought about moats in a certain way. So there's going to be a huge scramble at the company level. Teaching people to adapt, teaching them what to expect, is the second step. The third step is there's going to need to be some role for government in a displacement that's this macroeconomically large. I just don't see how it doesn't happen. The pie is going to grow much larger; the money will be there. The budget may balance without us doing anything because there's so much growth. The issue is distributing it to the right people. So I think this is probably a time to worry less about disincentivizing growth and more about making sure everyone gets a part of that growth. I know that's the opposite of the prevailing sentiment now, but I think technological reality is about to change in a way that forces our ideas to change.

那么显然，在你希望创造这种更强烈的紧迫感的过程中，你是在与政府中的人交谈吗？我的意思是，Anthropic 并不总是本届政府嘉宾名单上的首选，但你有在那里交谈的人吗？

So obviously in your desire to create this greater sense of urgency, are you speaking to people in the administration? I mean Anthropic hasn't always been the sort of first on the guest list for this administration, but do you have people there that you're talking to?

我亲自对他们说过。要清楚的是，我们有很多共识。我认为政府今年年中发布的 AI 行动计划实际上有一些非常好的想法。我想我们可能同意其中的绝大部分。但最重要的是，我们只想公开说出这些事，并进行公开辩论。我们不控制政策。我们能做的最有用的事情是向世界描述我们看到的，提供数据，然后在一个民主国家，由公众利用这些数据来推动政策。我们不能独自推动政策。

I've said it to them myself. To be clear, there are plenty of things we agree on. I think the AI action plan that the administration put out in the middle of this year actually had some very good ideas. I think we probably agreed with the vast majority of it. But most of all, we just want to say these things in public and have a public debate about them. We don't control policy. The most useful thing we can do is describe to the world what we're seeing and provide data, and then it's left to the public in a democracy to take that data and use it to drive policy. We can't drive policy on our own.

你在这里期间会与官员交谈吗？你去过美国馆了吗？

Are you going to be talking to officials while you're here? Have you been along to USA House yet?

我没去过美国馆。我将在达沃斯之行期间与官员交谈。

I've not been to USA House. I will be talking to officials during my trip to Davos.

那么回到 Anthropic。你创立 Anthropic 是因为你担心 OpenAI 没有足够重视安全。现在有人说竞争压力让你变得更鹰派了。那些竞争压力——为了跟上中国并保持领先——是否损害了你的安全原则？

So just to go back to Anthropic. You founded Anthropic specifically because you were worried that OpenAI wasn't taking safety seriously enough. Now some people say that competitive pressures mean you've gone more hawkish. Do those competitive pressures, to keep up with China and keep ahead, have they compromised your safety principles?

我们走了一条与其他参与者非常不同的路。早期我们做的一个好选择是成为一家专注于企业而非消费者的公司。与自己公司的商业激励作斗争非常困难；选择一种不需要与之斗争的商业模式更容易。我对消费者 AI 有很多担忧——它会导致需要最大化参与度，导致垃圾内容。我们看到其他参与者围绕广告做了很多事。Anthropic 不是那样运作的。我们只是向企业销售产品，这些产品直接具有价值。我们不需要将十亿免费用户变现，也不需要为十亿免费用户最大化参与度，因为我们没有与另一个大型参与者进行某种死亡竞赛。这让我们能够更仔细地思考。但即便如此，我们也做出了牺牲。我们对模型进行了许多其他公司没有做过的测试。其他一些参与者做过，但我认为我们在进行测试方面是最积极的，这些测试显示出令人担忧的行为——欺骗、勒索、谄媚——这些在所有模型中都有，但我们确保总是向公众公开这些。我们开创了机械可解释性科学，用于观察模型内部。我们做得完美吗？当然不。我认为我们总体上做得不错。你提到了中国。那不是关于竞争；那是关于公共利益使命。我担心如果专制国家在这项技术上领先，对在座的每一个人来说都将是一个糟糕的结果。

We've taken a very different route than some other players. One good choice we made early was to be a company focused on enterprise rather than consumer. It's very hard to fight your own business incentives; it's easier to choose a business model where there's less need to fight them. I have a lot of worries about consumer AI—it leads to needing to maximize engagement, leads to slop. We've seen a lot of stuff around ads from other players. Anthropic is not a player that works like that. We just sell things to businesses, and those things directly have value. We don't need to monetize a billion free users or maximize engagement for a billion free users because we're in some death race with another large player. That has let us think more carefully. But even with that, we have made sacrifices. We do all these tests on our models that others have not done. Some other players have done them, but I think we've been the most aggressive in running tests that show concerning behaviors—deception, blackmail, sycophancy—that are present in all models, but we make sure to always talk to the public about them. We've pioneered the science of mechanistic interpretability for looking inside models. Have we been perfect? Of course not. I think we've done a generally good job. You mentioned China. That's not about competition; it's about the public benefit mission. I'm worried that if autocracies lead in this technology, it will be a bad outcome for every single person in this room.

你具体担心什么？是关于芯片，关于围绕芯片的数据共享吗？

What are your specific concerns there? Is it about the chips, about sharing data around chips?

手段是出售芯片。我认为这是对谁领先谁落后影响最大的事情。

The means is selling the chips. That's the thing that I think will have the most impact on who is ahead and who's not.

但你知道，这种担忧并非针对某个特定国家或人民，而是针对一种政府形式。我担心 AI 可能特别适合专制政体，并加深我们在专制国家看到的压迫。我们今天的技术已经可以实现某种监控国家。但想想 AI 能做到什么：制造个性化宣传、入侵世界上任何计算机系统、监控全体人口、到处发现异议并压制它、制造一支庞大的无人机军队来追捕每个人。这真的很可怕。我们必须阻止它。但你觉得政府对此关注不够吗？

But you know the concern, and it's not about any particular country or certainly not the people in any country. It's about a form of government. I am concerned that AI may be uniquely well suited to autocracy and to deepening the repression that we see in autocracies. We already see it in the kind of surveillance state that is possible with today's technology. But if you think of the extent to which AI can make individualized propaganda, can break into any computer system in the world, can surveil everyone in a population, detect dissent everywhere and suppress it, make a huge army of drones that could go after each individual person. It's really scary. And we have to stop it. But again, is that something that you feel governments aren't paying enough attention to?

我认为可以说，不同国家都认为自己有地缘政治对手，但具体关注点——‘我们不希望专制政权获得这种强大技术’，我们应该有针对性政策，比如我们不需要与他们作战，我们只需要不卖这些芯片——我认为这方面关注不够。

I think it's fair to say that different countries think of themselves as having geopolitical adversaries, but the specific focus on 'we don't want autocracies to get this powerful technology' and we should have targeted policies like we don't need to fight them, we just need to not sell these chips. I think there's not enough focus on that.

我想多谈谈 Claude，因为可以说它正迎来一个真正的时刻。它确实正当时。我们最近报道了工程师和普通用户如何被 Claude 吸引。我只是想知道，与一年前相比，你现在对业务状况感觉如何？

I want to talk a bit more about Claude because I think it's fair to say it's having a real moment. It is having a moment. And we recently reported on how engineers and regular users are getting clawpilled. I just wondered how you feel about the state of the business today versus a year ago.

是的，这是那种业务增长很快但与技术一样沿着平滑指数曲线发展的事情。所以我们的收入曲线是：2023 年从零到大约 1 亿，2024 年从大约 1 亿到大约 10 亿，2025 年从大约 10 亿到大约 100 亿。不精确，是近似数字，但大致如此。在这个过程中，如果你上 Twitter，每隔几个月就会有人说‘天哪，Anthropic 正在改变世界。天哪，Anthropic 彻底完蛋了。’只是当下的兴奋情绪。但我们只是看着它，看着这条曲线。它很快。它在不断进步。它给了我们信心。我们永远不确定它是否会继续。可能不会。但这是我们从始至终观察到的经验事实。然后有些时刻，即使曲线平滑，也会出现突破点。所以现在我认为在开发者中，Claude Code 正迎来一个突破点。这个能够制作完整应用并端到端完成事情的东西。同样，它逐渐进步，但通过我们最新的模型 Opus 4.5，它达到了一个拐点，改进是渐进的，但就像温水煮青蛙。你看到逐渐改进，然后某个特定点突然被人们注意到。我认为第二件进一步加速的事情是，我们审视了 Claude Code，发现 Anthropic 内外有很多非技术人员意识到 Claude Code 可以为你完成这些令人难以置信的智能体式任务。它不仅能写代码，还能组织你的待办事项列表、规划项目、整理文件夹、处理大量信息并总结。所以需要的不仅仅是聊天机器人，而是智能体式任务。非技术人员意识到了这一点，他们非常想要它，以至于他们挣扎着使用命令行。非技术人员如果不是程序员，没有理由使用那个糟糕的界面。但人们还是去使用它。所以我看到后说，这看起来是未满足的需求。于是我们在大约两周内再次使用 Claude Code，制作了一个具有更好 UI 的版本，专门为代码以外的任务定制。我们发布了它，大约一天内，它的大多数指标是我们发布过的任何东西的四倍。所以这就是那两个时刻。我不知道这些是不是新能力，但就是那种共识时刻，人们变得非常兴奋，它正在非常快地推动采用。我认为人们正在赶上这项技术的能力，因为它达到了某个点，而且我们构建了使其可访问的界面。

Yeah, this is one of these things where the growth of the business has been fast but kind of on the same smooth exponential curve as the technology. So we have this revenue curve that in 2023 went from zero to roughly 100 million, in 2024 went from roughly 100 million to roughly a billion, in 2025 went from roughly a billion to roughly 10 billion. Not exactly, these are rounded numbers, but that is roughly it. Through that, if you go on Twitter, every couple months it's like, 'Oh my god, Anthropic's changing the world. Oh my god, Anthropic's totally destroyed.' Just the excitability of the moment. But we just watch it and we watch this curve. It's fast. It's constantly progressing. It's given us confidence. We never know for sure if it's going to continue. It might not. But that has been empirically what we have observed the whole time. And then there are these moments where even though the curve is smooth, there's a breakout moment. So right now I think there's a breakout moment around Claude Code among developers. This thing about being able to make whole apps and doing things end to end. Again that advanced gradually but with our most recent model Opus 4.5 it just kind of reached an inflection point where the improvement was gradual but it's like boiling the frog. You see the gradual improvement and then there's a specific point at which suddenly that's the point people notice. I think the second thing that has accelerated that further is we looked at Claude Code and noticed there were a lot of people inside and outside Anthropic who were not technical but realized that Claude Code could do these incredible agentic tasks for you. It couldn't just write code. It could also organize your to-do list or plan your projects or organize your folders or process a bunch of information and summarize. So the idea that not just a chatbot, but agentic tasks were needed. Non-technical people were realizing it and they wanted it so much that they were wrestling with the command line. Non-technical people have no reason to use that terrible interface if they're not programmers. But people were going through and using it anyway. And so I looked at that and said that looks like unmet demand. So we used Claude Code again in like two weeks to make basically a version with a better UI that's customized for tasks other than code. And we released it and within like a day most of the metrics on it were like four times as much as anything we'd ever released. So those are the two moments. I don't know that these are new capabilities but there was just one of these kind of consensus moments where people got really excited and it's driving adoption really fast. I think people are catching up to what the technology is capable of because it's reached a certain point and because we built interfaces that have made it accessible.

你能告诉我们一些你个人在生活中、家庭生活中如何使用智能体式 AI 吗？

Can you tell us a bit about how you personally in your life, your family life, use agentic AI?

是的。所以当我写文章或在公司面前讲话时，我觉得我工作的很大一部分是写作。所以我让 Claude 提供资料，帮助我写作，诸如此类。

Yeah. So when I'm writing an essay or something or things I say in front of the company, I feel like a fair amount of my job is writing. So I kind of have Claude come up with sources, help me with my writing, that kind of thing.

然后显然你正处在这个伟大时刻，我认为人们普遍预计你今年会 IPO。你能告诉我们一些你的计划吗？

And then obviously you're having this great moment and I think it's widely expected that you're going to IPO this year. Can you tell us a bit about your plans for that?

是的，我的意思是，我们不确定要做什么。我会说我们更专注于保持收入曲线增长，更好地向人们销售模型，警告社会影响，并带来好的社会影响。所以这是目前最高优先级。但我说这是一个资本需求非常高的行业，这并不新鲜。在某个时候，私人市场只能提供这么多。

Yeah, I mean we don't know for sure what we're going to do. I would say we're more focused on just keeping the revenue curve going, better selling the models to people, warning about the societal impacts, and bringing the good societal impacts. So that's the highest priority right now. But I'm not saying anything novel if I say that this is an industry with very high capital demands. And at some point, the private markets can only provide so much.

那么，另一个绝对正当时的模型是 Gemini，它最近冲到了应用商店榜首，OpenAI 宣布了代码红色警报，所以每个人都很兴奋。考虑到谷歌的庞大规模，你担心自己与 Gemini 竞争的能力吗？

So, another model that's absolutely having a moment is Gemini and it sort of surged to the top of the app store recently and OpenAI declared code red and so everyone got very excited about that. Do you worry about your ability to compete against Gemini given the sheer size of Google?

所以我认为这是另一个差异化有帮助的地方。企业战略：谷歌和 OpenAI 目前正在消费者领域激烈竞争，这对两者都是生死攸关的。

So I think this is another place where just being different helps. The enterprise strategy: Google and OpenAI are fighting it out in consumer right now, it is existential to both of them.

对 OpenAI 来说这是生死攸关，因为那是他们的全部业务；对 Google 来说也是生死攸关，因为他们有搜索业务，而这正是被这项技术颠覆的东西。所以他们需要自我革新，对抗颠覆。这始终是他们的首要任务，他们似乎比运营企业业务更关注这一点。看到 Gemini 在消费端的表现很棒。我认为他们走的是不同的路线。我刚刚和领导 Google 研究的 Demis 一起参加了一个小组讨论。我觉得他是个很棒的人，我认识他 15 年了，所以我支持他。

Existential to open AI because that's their whole business existential to Google because they have the search business and that's what's being disrupted by this. So they need to replace themselves and fight the disruption. So that's always their first priority and they seem much more focused on that than operating in the enterprise. It's been great to see what Gemini is capable of in consumer. I think they're going about it a different way. I was just on a panel with Demis who leads research at Google. I think he's a great guy. I've known him for 15 years, so I'm rooting for him.

一个区别是 Anthropic 没有生成视频和照片的能力。你认为这是一个潜在的弱点吗？

One difference is that Anthropic doesn't have the ability to generate videos and photos. Do you see that as a potential weakness?

对于企业业务来说，并没有对猫骑驴的照片或消费类视频的需求。也许在幻灯片和演示方面有边缘案例，但如果我们需要，我们可以从其他公司购买模型。所以我不知道未来会怎样，但我至少不认为我们需要这个。而且我认为这伴随着一些问题。看看那些大量的短视频，很多都是假的、上瘾的、垃圾内容。不是说所有内容都不好，或者做这个就意味着你不好，但这并不是我急于涉足的市场部分。

For enterprise business, there's not really a demand for photos of cats riding donkeys or consumer video. There's maybe an edge case around slides and presentations, but if we ever need it, we can just contract a model from someone else. So I don't know what will happen in the future, but I at least don't anticipate needing this. And I think there are problems associated with this. We look at the amount of short form video out there, a lot of it's fake, addictive, slop. Not to say that all of it is bad or that doing it means you're bad, but it's not a part of the market that I'm tripping over myself to get involved in.

你提到你和 Demis 一起参加了小组讨论，你说了一件有趣的事：科学家们对待 AI 时代的方式与科技企业家不同。能详细说说吗？

You mentioned that you were on a panel with Demis and you said something interesting: scientists are approaching the AI era differently from tech entrepreneurs. Can you say more?

当你思考这项技术时，它是持续了几十年的研究的交汇点，其中很多在十五年前还是学术性的，以及过去十五年开发和部署这些技术所需的规模，这只有大型互联网和社交媒体公司才能提供。他们有基础设施和资金。所以我们看到这样一个世界：一些公司由有科学背景的人领导，比如我和 Demis，而另一些则由社交媒体一代的企业家领导。我认为这非常不同。科学家有悠久的传统，思考他们建造的技术的影响，承担责任，不逃避责任。他们的动机首先是为世界创造一些东西，然后当这些东西可能出错时，他们会担心。企业家的动机，尤其是社交媒体一代的企业家，非常不同。作用于他们的选择效应，他们与消费者互动和操纵消费者的方式，非常不同。我认为这导致了不同的态度。

When you think about this technology, it's the intersection of research that has been going on for many decades, much of which was academic until a decade and a half ago, and the scale needed to develop and deploy these technologies over the last decade and a half, which has only come from large scale internet and social media companies. They have the infrastructure and cash. So we've seen a world where some companies are led by people with a scientific background, like me and Demis, and some are led by the generation of entrepreneurs from social media. I think that's very different. Scientists have a long tradition of thinking about the effects of the technology they build, of taking responsibility for it, not ducking responsibility. They are motivated by creating something for the world, and then they worry when that something can go wrong. The motivation of entrepreneurs, particularly the generation of social media entrepreneurs, is very different. The selection effects that operated on them, the way they interacted with and manipulated consumers, is very different. I think that leads to different attitudes.

美国和欧盟之间的紧张局势非常严重。如果事态升级，你是否担心这会如何影响你的业务运营？

Big picture tensions are running very high between the US and the EU. Do you wonder how that might impact how you operate your business should things escalate?

我们只代表自己。我们一直认为自己是独立的。我们不会刻意支持或反对任何人。但当我们在政策上不同意时，我们会说出来。当我们同意时，我们也会说出来。我们始终专注于 AI。我没有看到世界其他地方的人不愿意与我们合作。我们是我们自己。我们提供 AI 模型。我们努力负责任地做到这一点。

We only speak for ourselves. We've always thought of ourselves as independent. We don't go out of our way to be for or against anyone. But when we disagree on policy, we say so. When we agree, we say so. And we keep it focused on AI. I haven't seen any reluctance in folks in other parts of the world to work with us. We're our own thing. We're providing AI models. We try to do that responsibly.

有很多关于 AI 主权的讨论。我不太确定大家似乎都有什么。

There's been a lot of talk about AI sovereignty. I'm not entirely sure what everybody seems to have.

我也不知道那是什么意思。

I don't know what it means either.

要使前沿 AI 在现实部署中可靠安全且可控，仍然缺少的最重要的技术突破是什么？

What is the single most important technical breakthrough still missing to make frontier AI reliably safe and controllable in real world deployment?

我认为我们需要在机制可解释性上取得更多进展，这是研究模型内部结构的科学。训练这些模型时的一个问题是，我们无法确定它们会做你认为它们会做的事。你可以在一个上下文中与模型对话。它可以说各种事情。就像人类一样，这可能不是他们真实想法的忠实反映。如果他们告诉你‘我做 X 是因为 Y’，他们可能出于完全不同的原因做 X。他们可能在撒谎。我们非常习惯人类的这些问题，但 AI 也存在这些问题。任何现象学测试或训练我们都不能确定。但就像你可以通过做 MRI 或 X 光了解人类大脑，而不仅仅是与人类交谈一样，研究 AI 模型内部结构的科学是我们拥有的唯一真实依据。我相信这最终是让模型安全可控的关键。

I think we need to make more progress on mechanistic interpretability, which is the science of looking inside the models. One of the problems when we train these models is that we can't be sure they're going to do what you think they're going to do. You can talk to the model in one context. It can say all kinds of things. Just as with a human, that may not be a faithful representation of what they're actually thinking. If they tell you, 'I'm doing X because Y,' they might be doing X for a completely different reason. They might be lying. We're very used to these problems with humans, but they exist with AI as well. Any kind of phenomenological testing or training we can't be certain of. But similar to how you can learn things about human brains by doing an MRI or an X-ray that you can't learn just by talking to a human, the science of looking inside the AI models is the only ground truth we have. I am convinced that this ultimately holds the key to making the model safe and controllable.

AI 将如何影响当前的 K12 教育成就差距？

How will AI affect current K12 educational achievement gaps?

短期来看，有人用 AI 作弊，我认为这有问题。

There's the short-term stuff about people using AI for cheating, which I think is problematic.

但相对而言，你可以用 AI 进行不同的教学方式。我们考虑过这一点，并发布了针对教育的 Claude 版本。但我觉得更难的问题是：在 AI 时代，我们到底应该教什么技能？教育应该是什么样子？这并不容易，因为颠覆是广泛的。如果有人问我该从事什么职业，令人不安的事实是我不确定。我还无法判断未来的方向。我认为我们应该回归一些早期的教育理念。我们一直有一种非常经济导向、近乎功利的教育观念。我们应该摆脱这种观念，回到教育是为了塑造人格、培养品格、丰富自我、让你成为更好的人这一理念上。我认为这实际上是未来教育更安全的基础。

But in relative terms, you can have a different way of teaching using AI. We've thought about that. We've released versions of Claude for education that are designed around that. But I think the harder problem is what skills are we actually teaching in the world of AI? What does education look like in the world of AI? It's not easy because the disruption is broad. If someone asked me what career to go into, the uncomfortable truth is I'm not sure. I can't tell the direction it's going yet. I think we should go back to some earlier concepts about education. We've had a very economically inflected, almost mercenary notion of education. One thing we should do is move away from that notion back to the idea that education is designed to shape you as a person, build character, enrich you, and make you a better person. I think that's actually a safer foundation for education in the future.

这听起来像是我希望接受的教育。为了公平起见，我们还有时间问一个问题。这位女士。

That sounds like an education I'd have liked to have. To be fair to everyone, I think we have time for one question. This lady here.

从 AI 实验室的角度来看，当有经济体、国家和人民被抛在后面时，你们承担什么责任？这是否会扩大到从结构上让他们参与进来、放慢速度，或者确保他们不被排除在外？

From the point of view of AI labs, what responsibility do you hold when there are economies, countries, and people being left behind? Would that expand into structurally involving them, slowing down, or making sure they're not left out?

我在很多层面上担心这个问题。这不仅仅是国家之间的差距。我担心发展中国家与发达国家之间的差距，发展中国家可能会被技术革命抛在后面。但我也担心国家内部的差距。从我们的客户来看，初创公司迅速采用 AI，而传统企业则慢得多。我们在经济数据和美国各州的技术扩散中都能看到这一点。毫无疑问存在差异。噩梦是出现一个新兴的“第零世界”国家，有 1000 万人，其中 700 万在硅谷，300 万分散各地，形成自己的经济并脱钩。那将是反乌托邦的。我们应该思考如何阻止这种情况。Anthropic 正在做很多事情。一是针对发展中国家：我们开始从事公共卫生工作，与卢旺达教育部宣布了合作，并与盖茨基金会合作。我在《优雅的机器》一书中写过这一点。让发展中国家实现快速经济增长将是非常好的。在国家内部，我们需要思考如何不让一部分地区脱钩。如何让密西西比州获得硅谷那样的经济增长？我们在经济流动性和机会方面做了工作。但这两者都需要政府参与。意识形态无法经受这项技术的本质。它无法经受现实。我现在谈论的事情，虽然今天可能带有政治色彩，但将会成为两党共识和普遍认同，因为每个人都会认识到其必要性。记住我的话，明年或后年回来，每个人都会这么想。

I worry about that on many scales. It's not just country versus country. I worry about the developing world versus the developed world, where the developing world can get passed by technological revolutions. But I also worry about divisions within a country. I've seen across our customers that startups adopt AI quickly, while traditional enterprises move much slower. We can see this in our economic data and the diffusion of technology across US states. There's no question there's a differential. The nightmare would be an emerging zeroth world country of 10 million people, like 7 million in Silicon Valley and 3 million scattered, forming its own economy and becoming decoupled. That would be dystopian. We should think about how to stop that. Anthropic is doing a number of things. One is regarding the developing world: we're starting work on public health, we've announced stuff with Rwanda's Ministry of Education, and we're working with the Gates Foundation. I wrote about this in 'Machines of Loving Grace'. It would be great to get fast economic growth rates in the developing world. Within countries, we need to think about how not to have a part of the world that decouples. How do we get economic growth to Mississippi that is coming to Silicon Valley? We've done work on economic mobility and opportunity. But both of these will need government involvement. Ideology will not survive the nature of this technology. It won't survive reality. The things I'm talking about, while politically coded today, will become bipartisan and universal because everyone will recognize the necessity. Mark my words, come back next year or the year after, everyone will think this.

好吧，你总算以或多或少积极的语气结束了。我就此打住，非常感谢你，Dario。这真的非常精彩。

Well, you've managed to end on a more or less positive note. So I'm going to draw a line there and say thank you very much, Dario. That was really fascinating.

谢谢邀请。

Thank you for having me.