吴恩达：AGI 是营销术语，提出新图灵测试

AGI 已经成了一个营销术语。如果 AI 能像熟练的专业人类一样完成有用的经济工作任务，那对我来说似乎是一个更合理的 AGI 定义。但我认为我们离那还很远，可能不止几十年。

AGI has become a marketing term. If AI could do a useful economic work task as well as a skilled professional human, then that seems to me like a more reasonable definition for AGI. But I think we're very far away from that, maybe more than decades.

吴恩达是现代人工智能的关键架构师之一，常与杰弗里·辛顿、扬·勒昆和德米斯·哈萨比斯相提并论。

Andrew Ng is one of the key architects of modern artificial intelligence, mentioned in the same breath as Geoffrey Hinton, Yann LeCun, and Demis Hassabis.

现在，请坦诚相待，完全真实。只有少数工作岗位被 AI 完全自动化。所以，坦白说，我认为过度炒作，然后是失望，接着是所谓的泡沫破裂。这对世界不好，对 AI 领域也不好。很多最好的开源开放权重模型来自中国。开源时代结束了吗？许多大学课程更新缓慢，仍在为学生培训 2022 年的工作。很多那些工作实际上已不存在。雇主不想招聘。

Now, just be transparent, just be completely truthful. There is a small number of job roles that fully automated by AI. So, candidly, I think the excessive hype, then there's a disappointment, then there's a, you know, collapse of the so-called bubble. That would not be good for the world, not good for the field of AI. A lot of the best open-source open-weight models are coming out of China. Is the open-source era over? Many universities are slow to adapt curricula and are still training students for the jobs of 2022. Many of those jobs don't really exist. Employers don't want to hire.

程序员会失业吗？今年我们终于能实现 AGI 吗？你几小时前的报道以这个问题开头。对于任何合理的 AGI 定义，我认为答案是否定的。我们不会在 2026 年获得 AGI。在 2026 年获得 AGI 的最佳方式是有人设法大幅降低 AGI 的实际门槛。然后也许有人能跨过这个障碍。那么你如何解释并定义 AGI？

Will programmers be unemployed? Will this be the year we finally achieve AGI? You are opening your expose from a few hours ago with this question. For any reasonable definition of AGI, I think the answer is no. We will not get AGI in 2026. The best way to get AGI in 2026 is someone manages to dramatically lower the bar for what AGI actually means. Then maybe some come to clear the hurdle. How do you explain then define AGI?

我最熟悉的 AGI 定义是 AI 能完成任何人类能完成的智力任务。今天，一个人可能花几十个小时学会如何——我不知道——在森林里开卡车，我以前从未做过，但我想如果你训练我几个小时，我可能就能做到。一个人可以学会接听联络中心的电话，并按照某个业务需求的方式回答问题。有些事是人可以学会的智力任务，而非体力任务。所有这些任务仍然感觉需要大量工作来构建定制的 AI 工作流来执行。我们确实在做，而且结果非常有价值，但需要工程师花那么多时间构建这些东西，并不是公众通常对 AGI 的理解。尽管我希望有一天我们能达到那个目标——计算机在各方面都和人一样智能——但我认为我们离那还很远。

The definition of AGI that I'm most familiar with is AI that could do any intellectual task that a person can. And today a human can learn in maybe tens of hours how to, I don't know, drive a truck through a forest, which I've never done before, but I think if you train me for a few hours, I could probably do it. A human could learn to take calls from contact center and answer questions consistent with how certain business needs it answered. And there's certain things a person can learn to do intellectual tasks as opposed to physical tasks. And all of these tasks still feels like it's a ton of work to build custom AI workflows to perform. Which we do do and which turns out to be really valuable, but needing an engineer to spend so much time building these things is not how the public typically thinks about AGI. And as much as I hope we'll get there someday it's a computers that are as intelligent as people in every sense, I think we're very far away from that.

你今天刚刚提出了一个新版本的图灵测试。你能解释一下你的提议吗？

You just proposed today a new version of the Turing test. Can you explain what you propose?

我对这个很兴奋。所以，我觉得由于围绕 AGI 的炒作，AGI 已经成了一个营销术语，而不是精确的技术含义。它作为营销术语使用时，实际上误导了很多人。但既然人们想要 AGI 或对它感到兴奋，我也很兴奋，我们为什么不提出一个测试，看看我们是否真的接近实现 AGI 呢？所以，想法是这样的。在最初的图灵测试中，人类裁判会通过打字与 AI 或人类来回交流。最初的图灵测试是看 AI 能否欺骗裁判，使其无法分辨是在与 AI 还是人类对话。这在 1950 年是一个伟大的发明，我认为艾伦·图灵当时提出了这个测试，但它并没有像我们今天这样衡量智能。我认为我们可以构建一个图灵 AGI 测试，我称之为，让人类裁判设计一个多天的体验，可能是一些入职、培训环境、计算机。测试对象要么是 AI，要么是人类，可以访问计算机和普通软件，比如网页浏览器，可能还有 Zoom 或其他类型的软件。如果在几天内，在多天的体验中，AI 能像熟练的专业人类一样完成有用的经济工作任务，那么对我来说这似乎是一个更合理的 AGI 定义。我提出这个的原因是，这实际上更接近广大公众对 AGI 的看法。所以，当人们认为 AGI 可能到来时，他们会想，‘哇，AI 真的会做人的工作。’确实，如果 AI 能像远程工作者一样连续多天工作并产出成果，那将非常有价值，我认为这更接近人们认为的真正 AGI，而不是一些企业出于公关、政策或筹款目的推广的其他标准。

I'm excited about this. So, I feel like because of the hype around AGI, AGI has become a marketing term rather than something of precise technical meaning. And the way it's used as a marketing term, it actually misleads a lot of people. But since people want AGI or excited about it, I'm excited about it, why don't we come up with a test to see if we're actually, you know, getting close to achieving AGI. So, here's the idea. In the original Turing test, a human judge would get to type text back and forth to either an AI or a human. And the original Turing test was can an AI fool the judge into thinking of being unable to tell if they're talking to AI or to human. And it was a great invention for 1950, which is when I think Alan Turing came up with this test, but it's not really measuring intelligence the way we think about it today. Here's I think we can build a Turing AGI test, is what I called it, which is have a human judge design a multi-day experience that could be some onboarding, training environment computer. And the test subject will be either AI or a human with access to computer and normal software like a web browser, maybe Zoom, maybe other types of software. And if over a few days, if in a multi-day experience, if AI could do a useful economic work task as well as a skilled professional human, then that seems to me like a more reasonable definition for AGI. And the reason I'm proposing this is because this is actually much closer to what the broader public thinks of as AGI. So, when people think AGI could be here, they think, 'Wow, AI would really do people's jobs.' And indeed, if AI is able to function like a remote worker for multiple days and do productive work, that would be incredibly valuable and I think it's much closer to what people think of as real AGI as opposed to some of these, you know, alternative standards that some businesses are promoting for PR policy or fundraising purposes.

我记得我对诺贝尔奖得主罗杰·彭罗斯爵士的采访。数学是一个封闭系统。所以，AI 可以像玩游戏一样处理它。你不担心吗，Andrew，一个有效的 AI 测试并不存在？

I remember my interview with Sir Roger Penrose, Nobel Prize winner. Math is a closed system. So, AI can play it like a game. Aren't you afraid, Andrew, an effective AI test doesn't exist?

AI 测试或基准测试的一个挑战是，当有人预先固定测试集时，你只是在衡量 AI 的一个特定维度。这就是为什么 AI 模型，大型语言模型，有很多标准基准，比如 Sweep Bench、GPQA 等等。很难避免团队针对已知测试集进行优化。即使他们不直接这样做，最终也会如此。而 AI 是一种锯齿状的智能。它在某些方面很出色，在其他一些方面则非常糟糕。但 AGI，如果你认为它是能完成任何人类能完成的智力任务的 AI，那么它应该涵盖人类能完成的所有智力任务。所以，对我来说，固定测试集基准与人类裁判实时探测之间的一个巨大区别是，人类裁判可以查看当前技术水平，并探测 AI 哪里强、哪里弱。而对于 AI 成为 AGI，大多数人认为它应在每个维度上与人类匹配，这意味着人类裁判应该无法探测到 AI 在那些我们希望 AI 完成的经济价值工作任务上，比人类明显薄弱的地方。

One of the challenges with tests or benchmarks of AI is when someone fixes a test set in advance, then you're just measuring one specific dimension of AI. This is why AI models, large language models, there are a lot of standard benchmarks, you know, Sweep Bench, GPQA, and so on. And it's very difficult to avoid having teams optimize for the known test sets out there. Even if they don't do it directly, it ends up. And AI is a jagged form of intelligence. It's great at some things, really awful at some other things. But AGI, if you would think of it as AI that could do any intellectual task that a person can, should cover all intellectual tasks that people can cover. So, to me one big difference between a fixed test set kind of benchmark versus the human judge probing real time is a human judge can take a look at the state of the art and can probe to see where is AI strong, where is AI weak. And for AI to be this AGI, you know, most people think of it as matching humans in every dimension, which means the human judge should be unable to probe to find places where the AI is materially weaker than what a human can do in terms of these, say, economically valuable job tasks that we want AI to do.

基准测试还在衡量真实的东西吗？基准测试达到 90%，但用户说感觉更笨了。基准测试在衡量真实的东西吗？

Are benchmarks still measuring something real? Benchmarks hit 90% and the users say that it feels dumber. Are benchmarks measuring something real?

哦，我认为基准测试在衡量真实的东西，但它只是我们想衡量的东西中非常狭窄的一部分。许多基准测试的一个弱点是，我们更擅长设计客观的基准测试。所以，像数学这样有对错答案的问题，或者事实性问题，比如今年奥运会上谁赢得了 200 米自由泳之类的。这些问题有客观的对错答案，所以很多基准测试衡量的是 AI 是对是错。这非常黑白分明。生活中很多事情很难设计一个完全客观的最佳答案。所以，即使在我们这样的人类对话中，你知道，我说什么才是正确的？我不知道，但我可能说一些更好的或更差的东西。

Oh, I think benchmarks are measuring something real, but it's a very narrow slice of the sorts of things we want to measure. One weakness of a lot of benchmarks is we're much better at designing objective benchmarks. So, things like math where there's a right or wrong answer or a factual question, you know, like in the this year's Olympics, who won this, you know, 200 m freestyle swimming or whatever. So, those questions have objective right and wrong answers and so a lot of benchmarks is on measuring is the AI right or wrong. It's very black and white. A lot of life, there is very difficult to design one completely objective best answer. So, even in a human conversation like we're having, you know, what's the one right thing for me to say? I don't know, but I could probably say stuff that is better or some stuff that is worse.

总的来说，我们在设计衡量主观或好坏有程度差异的事物的基准测试方面做得并不好。很难编写一个测试集来明确说这是正确答案、那是错误答案。不是我们不尝试，但我觉得很多现有的基准测试并不能很好地捕捉这些方面。现实是，很多人类工作，比如写研究报告，并没有唯一正确的答案——只有不同质量等级。我们编写基准测试来评估这些灰色地带的能力远不如评估正确解答数学题或编写能正确运行的代码的能力。那些我们更擅长，但它们并不能代表大量实际有用的人类工作所需。

On average, we've been not very good at designing benchmarks to measure subjective things or where there are gradations of good and bad. It's very hard to write a test set saying this is a right answer, this is a wrong answer. Not that we don't try, but I feel like a lot of the benchmarks we have don't capture those things well. The reality is a lot of human work, like writing a research report, has no single right answer—there are different gradations of quality. Our ability to write benchmarks for these gray areas is much weaker than for tasks like solving a math problem correctly or writing code that runs correctly. Those we know how to do better, yet they don't represent what a lot of actual useful human work requires.

如果不是 AGI，2026 年我们可以期待什么？OpenAI 的 Greg Brockman 说两大主题将是智能体采用和科学加速。你怎么看？

If not AGI, what can we expect in 2026? Greg Brockman from OpenAI says two big themes will be agent adoption and scientific acceleration. What's your take?

我认为 AGI 是一种干扰。我们离它还很远。我们不会很快达到。但即使没有实现 AGI，我们在 2026 年及以后也在做并将继续做很多非常有价值的事情。我创造了“智能体式 AI”这个词来描述一个日益增长的现象。尽管我一直说构建 AI 来捕捉有用的业务流程需要大量工作，但有时这样做会带来巨大的价值。所以在 2026 年及以后，将会有很多激动人心的工作来构建 AI 智能体或智能体式工作流，以完成大量非常有价值、经济上重要的工作。在我的 AI Fund 团队，我们的团队一直在使用智能体式工作流来编写代码，就像许多其他团队一样，但也用于处理文书工作、检查关税合规、阅读复杂的法律文件以帮助律师更好地工作、执行医疗辅助任务或支持客户服务。他们发现，将人们目前在这些任务（如法律合规等）中经历的思维过程编码为 AI 智能体，放入智能体式工作流中，让 AI 为你完成，这非常有价值。这将持续很长时间。我们将在未来许多年里构建这些智能体式工作流，但其价值将非常巨大。

I think AGI is a distraction. We're very far from it. We're not going to get there anytime soon. But even without achieving AGI, there's so much incredibly valuable stuff we are doing and will continue to do in 2026. I coined the term agentic AI to describe a growing phenomenon. Even though I've said there's so much work to build AI to capture useful business processes, sometimes when you do that, it's incredibly valuable. So in 2026 and beyond, there will be a lot of exciting work to build AI agents or agentic workflows to do a ton of really valuable, economically important work. Here at my team at AI Fund, our teams have been using agentic workflows to write code, like many others, but also to look at paperwork, check for tariff compliance, read tricky legal documents to help lawyers do their work better, perform medical assistance tasks, or support customer service. They find that taking the mental processes people currently go through for these tasks—legal compliance or whatever—and coding them as an AI agent in an agentic workflow so the AI can do it for you, is very valuable. This will go on for a long time. We'll be building these agentic workflows for many years to come, but the value will be very large.

但我们正在讨论智能体式工作流。Rich Sutton 教授有句名言：原始算力总是胜过人类聪明才智。专注于智能体式工作流，难道你不是在反对这一点吗？

But we are talking about agentic workflows. Professor Rich Sutton famously said raw computing power always beat human cleverness. By focusing on agentic workflows, aren't you against that?

Rich 关于苦涩教训的文章非常有影响力且写得很好。明确地说，我喜欢 Scaling（规模扩张）。当我创办 Google Brain 时（后来与 DeepMind 合并形成 Google DeepMind 并创建了 Gemini），我的首要任务就是 Scaling。我们构建了非常大的神经网络，并向它们投入大量数据。我是 AI 界最早推动 Scaling 的人之一，当时其他人都觉得这很奇怪。我喜欢 Scaling。因为我建立 Google Brain 的方式，它的 DNA 就是 Scaling。这就是为什么 Google Brain 团队发明了 Transformer，这是历史上最具可扩展性的神经网络架构，它推动了生成式 AI 革命。所以我真的相信 Scaling。然而，在不同的时刻，我们 Scaling 的能力和注入其他形式知识的能力之间需要平衡。我不认为仅靠 Scaling 就能实现我们想要的一切。Scaling 非常强大，但有时企业会为了融资或公关而炒作某个现象。确实，由于缩放定律，当你扩展系统时，你可以相当可预测地预测性能。这由我在百度的团队以及后来的 OpenAI 展示过。这是一个很好的融资理由：‘给我更多钱，我会扩展机器和数据，性能会变得更好。’但由于这个巨大的真理内核，Scaling 被炒作得超出了它实际的价值——它非常有价值，但不如炒作所说的那么有价值。智能体式工作流使我们能够利用 Scaling 的 AI 模型、大语言模型，并额外将其他类型的知识注入系统，以构建更可靠、性能更好的工作流。

Rich's article on the bitter lesson was very influential and well written. To be clear, I like scale. When I started Google Brain, which later merged with DeepMind to form Google DeepMind and created Gemini, my number one mission was to scale. We built really big neural networks and threw lots of data at them. I was one of the earliest in AI to push for scale when everyone else thought it was strange. I like scale. Because of how I set up Google Brain, its DNA was to scale. That's why the Google Brain team invented the Transformer, the most scalable neural network architecture in history, which powered the generative AI revolution. So I really believe in scale. However, at different moments, our ability to scale and inject other forms of knowledge creates a balance. I don't think we'll achieve everything we want only by scale. Scale is incredibly powerful, but sometimes businesses hype up one phenomenon for fundraising or PR. It's true that because of scaling laws, when you scale systems, you can fairly predictably forecast performance. This was shown by my team at Baidu and later by OpenAI. This is a wonderful argument to raise funds: 'Give me more money, I'll scale up machines and data, performance will get better.' But because of that huge kernel of truth, scale has been hyped beyond what it actually is—amazingly valuable, but not as valuable as the hype says. Agentic workflows allow us to take advantage of scale AI models, large language models, and additionally inject other types of knowledge into the system to build more reliable, more performant workflows.

今天，你真的相信 Scaling（规模扩张）时代已经结束了吗？

Today, do you truly believe that the era of scaling is over?

不，我不认为 Scaling（规模扩张）时代已经结束。它变得越来越难，但很有趣。AI 可能一直在以指数级进步。推动这种指数级进步所需的资金数量也是指数级的。所以我们花指数级的钱来推动指数级的快速改进。这并不坏；它非常有价值，因为构建成本可以分摊到大量用户身上。我认为通过 Scaling 还能获得更多收益。但在这一点上，它并不是我们改进 AI 的唯一途径。

No, I don't think the era of scaling is over. It's been getting harder and harder, but it's interesting. AI has been progressing maybe exponentially. The number of dollars needed to drive this exponential progress has also been exponential. So we spend exponentially many dollars to drive exponentially rapid improvement. It's not bad; it's very valuable because the cost of building can be amortized over many users. I think there's still more to be gained by scale. But it's not the only avenue we have for improving AI at this point.

什么会让你改变对 Scaling 的看法？

What would change your mind about scaling?

什么会让我放弃 Scaling？如果在一段较长的时间内，进一步扩展模型的努力没有回报，那会让我改变想法。但有一个重要的附带说明：就像几十年来摩尔定律推动进步但需要不断变化的技术一样，Scaling 的具体方法已经改变。在大语言模型的早期，方法是获取更多数据，创建更大的模型。这正是 Google Brain 在 15 多年前开始做的。但自那以后，AI 模型已经阅读了几乎整个开放的互联网。所以那个简单的方法不再有效，这就是为什么许多团队现在在合成数据生成上做更多工作，这需要更多的人类工程，以及在不同的强化学习方法上做更多工作。所以尽管 Scaling 仍然有回报，但推动 Scaling 的具体方法在过去几年里已经发生了显著变化。

What would make me give up on scaling? If for an extended period of time, further efforts to scale up models are not paying off, that would make me change my mind. But one important asterisk: just as over many decades Moore's law drove progress but required different technologies over time, the recipe for scaling has changed. In the early days of large language models, the recipe was get more data, create a bigger model. That's exactly what Google Brain set out to do over 15 years ago. But since then, AI models have read pretty much the entire open internet. So that simple recipe doesn't really work anymore, which is why many teams are now doing much more work on synthetic data generation, which requires more human engineering, and more work on different reinforcement learning recipes. So even though scaling still pays off, the specific recipes for driving that scaling have shifted significantly over the last few years.

那么从另一方面来说，一个更智能的模型能否直接胜过一套智能工作流？

So, from the other side, can a smarter model simply beat a smart workflow?

我倒希望事情这么简单。理论上是的，但实际上比大多数人想象的要难得多。为什么？一个很棒的现象是，随着更智能的模型出现，比如 Claude 4.5.5 Opus、Gemini 3、GPT 5.1、5.2，我们越来越能够给大语言模型一套工具，然后放手让它去做。例如，给它读写文件系统的工具，然后让它执行清理硬盘上多余文件之类的任务。这些模型的表现非常出色。然而，对于许多工作流来说，它们的可靠性还不足以投入生产。因此，虽然更智能的模型很好，但我在许多实际商业用例中看到的是，团队会仔细思考工作流，确定关键步骤，然后逐步实现，以获得几乎每次都能正常工作的可靠性能。随着模型变得更智能，我们让它们更加自主，并移除护栏。我的团队经常发现，六个月前构建的系统有更多护栏和脚本化步骤，而我们通常会减少这些脚手架。我们不再给出非常详细的逐步指令，而是更倾向于说：“你自己决定要做什么。”例如，过去我们构建深度研究工具时，会明确步骤：进行网络搜索、执行多少次查询、下载多少页面、总结等等。现在，AI 模型更能自主决定是否继续搜索或进行总结。因此，我经常发现自己把一年或一年半前构建的原型系统中的指令剥离出来，让模型自己做出决策。但这还有很长的路要走。对于深度研究这样的用例，如果遗漏了引用，问题不大。但对于高风险的企业用例，让 AI 模型完全自主行动，其可靠性差距虽然正在缩小，但仍然比一些人想象的要大。

I wish it were that easy, though. In theory, yes. In practice, way harder than most people think. Why? One thing that has been fantastic is with more intelligent models, like Claude 4.5.5 Opus, Gemini 3, GPT 5.1, 5.2, we've been increasingly able to give a large language model a set of tools and just let it loose. For example, give it tools for reading and writing from a file system, then tell it to do a task like looking for extraneous files and cleaning up my hard disk. The level of performance is fantastic. However, for many workflows, it's just not reliable enough to be production ready. So, while smarter models are great, what I see for many practical business use cases is teams thinking through the workflow, identifying key steps, and implementing them piecemeal to get reliable performance that works almost every time. As models become more intelligent, we are letting them be more autonomous and removing guardrails. My team often finds that systems built six months ago had more guardrails and were more scripted, and we routinely reduce the scaffolding. Instead of giving very detailed step-by-step instructions, we are more likely to say, 'Go decide for yourself what you want to do.' For example, we used to build deep researchers with explicit steps: do a web search, run this many queries, download this many pages, summarize, etc. Now, AI models are much more capable of deciding whether to keep searching or summarize. So, I find myself frequently taking systems prototyped a year or a year and a half ago and ripping out the instructions, prompting the model to figure out decisions by itself. But it's still a long way to go. It works for use cases like deep research where missing a citation isn't the end of the world. But for high-stakes enterprise use cases, the reliability gap to letting an AI model just do whatever it wants, while closing, is still bigger than some people think.

我想问问你关于 Yann LeCun 和 Demis Hassabis 之间的争论。Yann LeCun 有句名言说人类智能是专门的，而非通用的。Demis Hassabis 则认为，从理论计算的角度看，大脑是通用学习器。你对此怎么看？

I need to ask you about the dispute between Yann LeCun and Demis Hassabis. Yann LeCun famously says human intelligence is specialized, not general. Demis Hassabis argues that brains are general learners under theoretical computations. Where do you stand on this?

我看不出矛盾之处。也许我漏掉了什么。对我来说，人类大脑——假设人类本身就是 AGI，是通用智能——最神奇的地方在于它的可塑性或学习能力。我认为 AGI 不应该是已经无所不知的 AI，那看起来非常困难且不切实际。人类大脑在经济任务中如此有价值的原因，在于它能够学习做任何需要的新事物。通过学习，我们获得了极其专门的智能。例如，人类大脑通过攻读数学博士来学习解决极其困难的数学问题。这是作为终点的非常专门的智能，但它是通过学习获得的。理论上，同一个大脑，经过不同的训练，可以成为国际象棋大师或网球高手。因此，人类大脑之所以如此通用，并不是因为它已经无所不知，而是因为它能够适应并学习极其广泛的事物。

I see no contradiction. Maybe I'm missing something. To me, the amazing thing about the human brain—assuming that human is AGI itself, it's general intelligence—is its plasticity or ability to learn. AGI to me should be less about AI that already knows everything under the sun; that seems very challenging and impractical. What makes the human brain so valuable for economic tasks is its ability to learn to do new stuff, whatever is needed. Through learning, we gain incredibly specialized intelligences. For example, the human brain learns to solve difficult math problems by getting a PhD in math. That is very specialized intelligence as the endpoint, but it was learning. Theoretically, that same brain, given different training, could have been a chess master or amazing at tennis. So, what makes the human brain so general is not that it already knows everything, but its ability to adapt and learn a huge range of things.

这难道不是隐含着站在 LeCun 一边吗？

Doesn't that implicitly side with LeCun?

在我开始组建大脑团队时，真正激励我的一个想法是，人类学习的很多方面可能源于一个单一的学习算法。我们的 DNA 并不长，包含的信息量非常有限。然而，我们的 DNA 却编码了大脑的生物学结构，而大脑是一个相当通用的学习算法。这就是为什么大脑可以学习攻读数学博士、骑摩托车、在电脑上打字或在呼叫中心工作。它拥有这个非常通用的学习算法，通过这个算法，它可以专门化到令人眼花缭乱的各种领域。正是这种学习能力让智能显得通用——它能够获得几乎任何领域的专长。

One thing that really motivated me when I was starting to grow the brain team was the idea that a lot of human learning may be due to one learning algorithm. Our DNA is not that long; it contains a very limited amount of information. Yet, somehow our DNA has encoded the biology of the brain, and the brain is a fairly general learning algorithm. That's why the brain can learn to do a PhD in math, ride a motorcycle, type on a computer, or work in a call center. It has this very general learning algorithm that allows it, through learning, to specialize in a bewildering range of things. That's what makes intelligence seem general—this learning ability that lets it gain almost any specialty under the sun.

几个月前有位老人说谷歌完蛋了，因为他们把 AI 建立在旧的搜索之上。OpenAI 是从零开始构建的。你认为他们能把新旧结合起来吗？

Some old man said a few months ago Google is doomed because they are building AI onto old search. OpenAI builds from scratch. Do you think they can marry old with the new?

竞赛正在进行中。这难道不令人兴奋吗？Sam 是我在斯坦福的学生，我在谷歌也有很多朋友。所以，我非常支持 OpenAI，也非常支持谷歌。回顾技术颠覆的历史，有时新进入者获胜，有时现有企业获胜。两者都有机会表现良好，游戏仍在继续。例如，在互联网颠覆期间，谷歌是一家随着互联网崛起的初创公司，但像微软和苹果这样在互联网之前就成立的现有企业也做得很好。显然，AI 对谷歌这样的现有企业具有很大的颠覆性。我认为谷歌打出了好牌。Gemini 3 是一个令人难以置信的模型。

The race is on. Isn't it exciting? Sam was my student at Stanford, and I have a lot of friends at Google. So, I'm very pro OpenAI, also very pro Google. Looking back at the history of technological disruptions, sometimes new entrants win, sometimes incumbents win. There's room for either to do well, and the game is still on. For example, during the internet disruption, Google was a startup that rose with the internet, but incumbents like Microsoft and Apple, founded long before the internet, did just fine. Clearly, AI is very disruptive for incumbent businesses like Google. I think Google has played its hand well. Gemini 3 is an incredible model.

比 ChatGPT 更好吗？

Better than ChatGPT?

我经常使用 Gemini、ChatGPT、Claude 以及很多其他模型。

I use both Gemini and ChatGPT and Claude and a whole bunch of models quite often.

你经常谈论 AGI。五年前……

You are talking about AGI a lot. Five years ago...

我能说其实我不喜欢谈论 AGI 吗？但其他人谈论得太多，炒作也太多了。

Can I just say I actually don't like to talk about AGI, but others talk about it so much and there's so much hype.

AGI 是炒作还是被炒作？

AGI is hype or hyped?

它被严重炒作了。至少公众认为 AGI 是 AI 变得像人类一样通用智能。我们离那还很远。我希望我们能实现它。我很想达到 AGI。

It is vastly hyped. At least the public thinks of AGI as AI becoming as intelligent as people in a very general sense. We are so far away from that. I hope we'll get there. I would love to get to AGI.

但现实地说，我认为我们距离那还有几十年，甚至可能更久。所以那种‘哦，我们只需要再几个季度就能实现 AGI’的想法根本不会发生，除非你重新定义 AGI，降低标准，让它更容易实现。坦白说，在 AI 的历史上，我们经历过几次 AI 寒冬，一些善意的人过度夸大了 AI 的前景。这导致了期望过高而未能实现，投资和兴趣也随之崩溃。这对这个领域是不利的。AI 现在运行得很好，非常有价值。我认为能阻碍 AI 势头的事情相对较少。我真正担心的一个问题是过度炒作导致失望，进而导致所谓泡沫的破裂。这对世界或 AI 领域都不好。因此，平息关于 AGI 的炒作是为我们更可持续的增长奠定基础的重要一步。

But realistically, I think we're decades, maybe more than decades away from that. And so this idea that, 'Oh, we just need another few quarters to get to AGI.' That's just not going to happen unless you redefine AGI to lower the bar and make it much easier to achieve. And candidly, in the history of AI, we've seen a few AI winters where well-meaning people overhyped the promise of AI. This led to elevated expectations that were not met and a collapse of investment and interest. And this was bad for the field. AI works really well right now. It's incredibly valuable. I see relatively few things that could derail AI's momentum. One of the things I actually worry about is excessive hype that leads to disappointment that leads to collapse of the so-called bubble. And that would not be good for the world or good for the field of AI. So diffusing hype about AGI is an important thing to do to lay the foundation for us to have more sustainable growth.

五年前，解决任何编程问题都会被称作 AGI。今天，我们有了它，却称之为工具。我们是不是移动了目标？

Five years ago, solving any coding problem would have been called AGI. Today, we have it and we call it a tool. Did we move the goalposts?

我不记得有可信的团队宣称实现了 AGI。我记得有团队宣称 AGI 就在不久的将来。但三年前，没有团队宣称他们实现了 AGI。他们说的是我们很快就能达到。然而到目前为止，还没有人达到。你不觉得我们移动了目标吗？我认为，如果说有什么变化，那就是团队一直在试图降低实现 AGI 的标准。我的意思是，AGI——能做人类能做的任何智力任务的 AI——是一个非常高的标准。我们离那个标准还差得很远。但如果团队提出容易实现的替代定义，那么他们可能更快达到。顺便说一句，我不介意我们如何定义 AGI。我们可以随意定义 AGI，但问题是大多数公众认为 AGI 是普遍非常智能的 AI，基本上就是类人智能。所以世界上大多数人都是这么想的。问题在于，因为人们不断提出替代定义，当很多人用同一个词指代非常不同的事物时，这个词就失去了意义。例如，我们都知道‘蓝色’这个词是什么意思，对吧？我的衬衫是蓝色的。但如果由于某种原因，社会指着各种不同的颜色说‘这是蓝色，这是蓝色’，那么‘蓝色’这个词就会失去意义，因为当有人说蓝色时，人们甚至不知道它是什么意思了。AGI 的情况就是这样。不同的团队提出了不同的替代定义。人们说‘这是 AGI’。因为人们对这个词应用了许多不同的定义，现在当有人说 AGI 时，很难知道他们确切的意思。而问题是公众认为 AGI 是类人智能。所以如果有人提出一些奇怪的狭隘技术定义，并说‘AGI 将在两年内出现’，公众仍然会认为 AI 将在两年内像人类一样智能，这在我看来并不真实。

I don't remember credible teams declaring achieving AGI. I remember teams declaring that AGI was only a short time in the future. But teams were not, three years ago, declaring they got AGI. They were saying we could get there really soon. And then, so far no one's gotten there. Don't you think that we moved the goalposts? I think that, if anything, teams have been trying to lower the bar to what it means to achieve AGI. I mean, AGI—AI that can do any intellectual task that humans can—is a very high bar. We have not gotten anywhere near to that bar. But if teams come up with alternative definitions that are easy to achieve, then maybe they get there sooner. And by the way, I don't mind how we define AGI. We can define AGI however we want, but the problem is most of the broader public thinks of AGI as generally very intelligent AI. Basically, human-like intelligence. So that's what most people out in the world think. And so, the problem with that is that because people keep coming up with alternative definitions, terms lose meaning when lots of different people use that same term to refer to very different things. For example, we all have a sense of what the word blue means, right? My shirt is blue. But if for some reason society points all sorts of different colors and says, 'This is blue. This is blue.' then the word blue will lose meaning because people don't even know what it means anymore when someone says blue. And that's what happened with AGI. Different teams came up with different alternative definitions. People say, 'This is AGI.' And because people are applying lots of different definitions to the term, when someone now says AGI, it's hard to know what exactly they mean. And the problem is the public thinks of AGI as human-like intelligence. So if someone comes up with some weird narrow technical definition and says, 'AGI will appear in 2 years,' the broader public nonetheless thinks AI will be as intelligent as humans in 2 years, which just doesn't seem true to me.

我们来谈谈，我们现在处于什么阶段？是框架重要，还是模型本身重要？

Let's talk, where are we now? Does the harness matter or is it all about the workhorse?

我认为框架非常重要。看到 Anthropic 构建 Claude Code 以及他们的 SDK，确保他们和其他人有一个好的框架来使用驱动它的模型，这真是令人难以置信。框架的细节——如何构建提示词，给模型什么工具——所有这些细节仍然很重要。实际上，举一个小例子：我们当前的模型非常智能，它们在工具使用方面越来越好，比如进行函数调用。话虽如此，如果你给模型太多工具，它会消耗大量输入上下文，并且更有可能出错，做出错误的 API 调用或工具使用调用。所以这种小事情——也许感觉像是 2026 年，为什么我们还需要担心这样的工程细节？事实证明我们需要，因为它仍然对整体性能有很大影响。许多团队使用 MCP，我也经常使用 MCP。一个实际的工程问题是，如果你的 MCP 服务器有太长的工具列表，它会消耗大量输入上下文。而且工具可能太多，模型无法有效判断该用哪个。所以你就进入了上下文工程，有时框架可以帮助更顺畅地做出这些决策。

I think the harness matters a lot. It's been really incredible watching Anthropic build Claude Code as well as their SDK to make sure that they and maybe others have a good harness to use the model powering this. And the details of the harness—how you structure the prompts, what tools you give the model—all those details really matter still. Actually, maybe one small example: our current models are incredibly intelligent and they're becoming better and better at tool use, making function calls. Having said that, if you give a model too many tools, it consumes a lot of the input context and it's much more likely to struggle and make the wrong API call, the wrong tool use call. And so, this little kind of thing—maybe it feels like it's 2026, why do we need to worry about engineering details like these? It turns out we do because it still makes a big difference to the overall performance. Many teams use MCP, I use MCP a lot. One of the practical engineering things is if your MCP server has too long a list of tools, it consumes a ton of your input context. And it may be too many tools for the model to effectively figure out which ones to use. So you get into context engineering, and sometimes a harness can help make these decisions much more smoothly.

Anthropic 预测持续学习将在 2026 年解决。我是否也这么认为？

Anthropic predicts continual learning will be solved by 2026. Do I expect the same?

我期待持续学习被解决的那一天。那将是惊人的。如果它在 2026 年完全解决，那将令人愉快。我预计我们会取得进展。我认为持续学习非常重要。一个孩子从几次摔倒中就学会了走路。强化学习需要数百万次模拟。我们难道要暴力学习吗？暴力学习的一个挑战是，人类智能来自于学习算法的通用性，这正是人脑如此强大的原因。它是通用的，能很快学会新东西。这就是为什么与我们合作的人类能做这么多事情。而如果你需要花很长时间暴力训练 AI 去做某个狭窄的任务，在某些情况下它可能仍然很有价值，但对于很多任务来说，这样做就不划算了。所以，能够跟人类打个招呼，聊几句，让他们弄清楚该做什么并完成工作，这非常有价值。如果你需要雇佣一个 AI，然后花一百万美元训练它做这一项任务，那么对于很多任务来说，花那么多精力训练 AI 是没有意义的。如果样本效率不重要，那么不幸的是，地球上只有三家公司负担得起训练前沿模型。

I look forward to when continual learning will be solved. It would be amazing. It'll be delightful if it ends up being completely solved in 2026. I expect we'll make progress. I think continual learning is very important. A child learns to walk from few falls. Reinforcement learning needs millions of simulations. Do we just brute force learning? One of the challenges of brute force learning is that to the extent that human intelligence comes from the generality of the learning algorithm, that's what makes the human brain so powerful. It's general, learns new things really quickly. And this is why humans that work with us can do so much. Whereas if you need to spend a long time to brute force AI to do some narrow task, it could still be really valuable in some cases, but the case for doing it just isn't going to be there for a lot of tasks. So the ability to say hi to a human and talk to them a bit and let them figure out what to do and do a job, that's really valuable. If you need to hire an AI and then spend a million dollars to train the AI to do this one task, then there are a lot of tasks that it just doesn't make sense to spend that much effort to train the AI. If sample efficiency doesn't matter, only three companies on earth unfortunately can afford to train frontier models.

我非常渴望看到开源和开放权重模型与专有模型之间持续演变的动态。AI 作为寡头垄断。我希望我们不要走向那样的未来。坦白说，如果我看移动开发平台，如今它已经不那么有趣了，部分原因是有两个守门人。要在移动设备上做点什么，至少在美国，你目前需要得到 iOS 或 Android 的许可。因此，由于这些封闭平台，我们无法进行某些创新。

I would be really keen to see the continuing evolving dynamics between open source and open weight models versus proprietary models. AI as oligopoly. I hope we don't get to that future. So candidly, if I look at the mobile development platform, it's just not as interesting these days, partly because there are two gatekeepers. To do something on mobile, at least in the United States, you currently need permission from either iOS or Android. And so there are certain innovations that we're just not allowed to do because of those closed platforms.

我们很多 AI 从业者真的希望不会出现两三个守门人垄断前沿 AI 开发的情况，如果有人有想法，我希望人们能够在大语言模型之上进行创新。因此，开源开放权重模型是防止出现少数守门人的关键。如果我们能确保每个人都保留创新的自由——这在今天的 AI 世界比移动世界要多得多——那么我们将看到更多的发明、更多酷炫的应用，社会也会因此更加富裕。

A lot of us in AI really hope that they will not end up with two or three gatekeepers to building cutting-edge AI things, where if someone has an idea, I'd love to have people allowed to innovate on top of large language models. So open-source open-weight models is a key to preventing this small handful of gatekeepers from arising. If we can make sure that everyone preserves the freedom to innovate, which we have much more in the AI world today than say in the mobile world today, then we'll see a lot more inventions, a lot of cool applications, and society will be much richer for it.

我们真的需要真正的持续学习吗？持续学习还有很多工作要做。我认为它是当前 AI 领域重要的开放研究课题之一。目前许多基于文本的记忆系统让 AI 做它该做的事，然后将一堆文本写入某种智能体记忆，问题在于：文本真的是足够好的记忆表示吗？有一些关于非文本表示的研究，但目前大部分还是文本。此外，我们构建的所有这些记忆系统实际上从未更新语言模型的权重，这感觉像是我们遗漏了拼图的关键一块。所以，我的一些团队有一些想法，不确定会走向何方，但目标是改进持续学习。

Do we really need real continual learning? There's a lot of work to be done in continual learning. I think of it as one of the important open research topics in AI right now. A lot of the current text-based memory systems have the AI do whatever it does, and then write a bunch of text into some sort of agentic memory, and questions are: is text really a good enough representation for memory? There's some research on non-text representations, but a lot of it is in text right now. Also, the fact that we're building all these memory systems that don't really ever update the weights of the LM, that feels like we have to be missing a key piece of the puzzle. So, some of my teams have had ideas, not sure where it will take them, but improving continual learning.

当前持续学习最大的瓶颈是什么？

What's the biggest bottleneck to continual learning right now?

哦，我会说我们还没有正确的想法。我们不确定正确的想法是什么。有很多想法，但就是这样。

Oh, I would say we don't have the right ideas. We're not sure what the right ideas are. There are a bunch of ideas. It's just one of those things.

连你也是？

Even you?

有一些我认为有希望的想法。但是……

There are a few ideas that I think are promising. But...

比如？

For example?

也许让我先看看有没有时间尝试一下，如果有效我会告诉你。老实说，我不知道是否有效。但对我来说就是这样。如果我们问……是的，就像我无法说出解决这个巨大未解研究问题的瓶颈是什么，我不知道，因为通往那里的路径并不清晰。所以我甚至无法确切说出瓶颈是什么。我们只是不知道如何做到。

Maybe let me see if I have time to try it out first and I'll let you know if it works. I honestly don't know if it works. But to me it's one of these things. If we were to ask... Yeah, it's like I can't say what's the bottleneck to solving this huge unsolved research problem, and I don't know because the path to get there isn't clear. So I can't even say what exactly are the bottlenecks. We just don't know how to do it.

Eliezer Yudkowsky 说如果有人构建 AGI，所有人都会死，但每拖延一年，就有数百万人死于癌症、衰老、疾病，而 AI 本可以解决。哪个风险更大？

Eliezer Yudkowsky says if anyone builds AGI everyone dies, but every year we delay millions die from cancer, aging, disease AI could solve. Which risk is greater?

我读他的很多论点时完全摸不着头脑。他的许多论点足够循环论证，我甚至不知道如何反驳。我认为 AI 今天在世界上做了很多好事，任何庆祝 AI 进步的事情都会带来更好的生活、拯救更多生命、让更多人更富裕、更幸福、摆脱贫困。AI 的净收益远大于净危害。有一些有害用例我们需要清醒认识并消除它们。但此时此刻，我非常确信任何庆祝 AI 进步的事情对人类都是有益的。

I can't make head or tails of a lot of his arguments when I read them. Many of his arguments are sufficiently circular. I don't even know how to argue with them. I think that AI is doing so much good in the world today that anything we can do to celebrate the progress of AI will lead to better lives, many more lives saved, many people much wealthier, many people much better off, lift a lot of people out of poverty. AI net benefit is so much greater than the net harm. And there are a few harmful use cases that we should actually be clear-eyed on and let's get rid of those. But at this moment in time, very confident anything we can do to celebrate AI progress would be good for humanity.

从你的角度看，未来是不可预测的，对吧？

From your point of view, future is unpredictable, right?

我希望我知道如何预测未来。当然不可预测。但仅仅因为我们不确定世界会走向何方，有些趋势我们非常有信心。例如，让计算机更智能在我看来显然是一件好事。让智能的获取民主化，这样它就不只是……世界上最昂贵的东西之一就是智能。雇佣一个聪明的医生、一个聪明的老师要花很多钱，对吧？来照顾你或……

I wish I knew how to predict the future. Of course it's unpredictable. But just because we don't know with certainty exactly where the world will be, there are trends that we have very high confidence about. For example, having computers be more intelligent seems clear to me that's a great thing. Democratizing access to intelligence so that it's not just... it turns out one of the most expensive things in the world is intelligence. It costs a lot of money to hire a smart doctor, a smart teacher, right? To care for you or...

我们需要更多 AI 安全工具吗？

Do we need more AI safety tools?

是的，当然我们应该致力于让 AI 系统更可靠并降低风险。同时，如果我看看所有这些……

Yes, of course we should be working on making AI systems more reliable and reducing risks. At the same time, if I look at all these...

周围每个人都专注于有利可图的事情。

Everyone around is focused on something that is profitable.

我想很多人会惊讶于有一种观点认为硅谷的人只关心钱，其他什么都不管，这完全错误。坦白说，很多朋友在这些公司工作，我亲自认识很多 CEO，是的，有极少数人只关心利润，但那是非常小的一部分。我的很多朋友在这些公司，我认识了几十年或更久，他们真的想做正确的事。所以人们认真对待 AI，认真对待负责任的 AI，他们会坐下来真正集思广益，思考 AI 系统可能出问题或成功的所有情况，并努力降低风险。所以我知道有一种观点，一种错误的刻板印象，认为硅谷有一群牛仔或女牛仔为了利润什么都干，这完全不是真的。不幸的是，确实有少数公司在诱惑面前想赚几十亿美元。这种诱惑很强，但我认为那真的是少数，非常少数，极小部分人必须做出的决定。

I think many people will be surprised at how there's this view of Silicon Valley that people only care about money and nothing else, and that's just absolutely wrong. Frankly, a lot of friends that work in a lot of the companies, I know many of the CEOs personally, and yes, there's a tiny minority that just cares about profit, but it's a very small fraction. A lot of my friends in these companies, people I've known for decades or longer, really want to do the right thing. So people take AI seriously, people take responsible AI seriously, people will sit down and really brainstorm all the things that could go right or wrong with an AI system and try to mitigate your risk. So I know that there's a view, a false stereotype of a bunch of cowboys or cowgirls in Silicon Valley that would just do anything for profit, and it's just completely not true. Now, unfortunately, it is true that there's a small number of companies where there's a temptation to make billions of dollars. That temptation is strong, but I think that's really a small minority, very small minority, tiny minority of the decisions people have to make.

开源时代结束了吗？

Is the open-source era over?

开源做得很好。现在，一个奇怪的事情是很多最好的开源开放权重模型来自中国。回顾过去几年的历史，每年开源或开放权重的选项都在快速增长。所以我认为开源运动非常强劲。同时，专有选项也在快速增长，但这没关系。重要的是开放选项也在强劲增长。

Open source is doing great. Now, one weird thing is a lot of the best open-source open-weight models are coming out of China. Looking back on history over the last few years, every single year the options that have been open-source or open-weight have grown rapidly. So I think the open-source movement is very strong. At the same time, the proprietary options have also grown rapidly, but that's okay. The important thing is the open options are also growing strongly.

当你看到 AI 发展的当前阶段时，你心里在想什么？

What goes through your mind when you look at the current stage of AI development?

我想赋能每个人构建 AI。作为开发者，我再也不想手动编码了。我希望 AI 尽可能多地为我写代码，AI 加速软件工程是非常明显的。但很多人不太清楚的是，许多非软件开发者通过构建软件、用 AI 做事会比不这样做要好得多。我已经在我们的团队中看到了这一点。懂 AI 的营销人员开始甩开不懂的。我的 CFO，在 AI Fund，她会写代码，她比另一个假设不懂用 AI 助手写代码的 CFO 完成得多得多。所以我看到的是，有了 AI 工具，特别是用 AI 构建软件，这感觉像是一项非常重要的新能力，我们需要每个人都拥有。

I want to empower everyone to build AI. So as a developer, I don't ever want to have to code by hand again. I want AI to write as much of my code for me as possible, and acceleration of software engineering because of AI is very clear. But what is less clear to many people is that many people that are not software developers will be so much better off building software, doing things with AI than not. I'm already seeing this among our teams. The marketer that knows how to use AI is starting to run circles around the ones that don't. My CFO, at AI Fund, she writes code, and she gets much more done than some other CFO, hypothetical CFO, that doesn't know how to write code using AI assistants. So what I'm seeing is with AI tools, including specifically building software with AI, it feels like a really important new capability that we need everyone to have.

所以，随着新能力的出现，许多人会拥抱它们，积极向前，真正做得更多，效率更高。然后遗憾的是，会有不拥抱它们的人，不幸地被抛在后面。我对此非常担忧。一个重大挑战是，如果你看看大学体系，许多大学在调整课程方面很慢，仍然在为学生培训 2022 年的工作技能，但很多这些工作实际上已经不存在了。所以，雇主不想雇佣一个 2022 年的学生。相反，许多雇主找不到足够了解 AI、知道如何用 AI 构建的人才。我不是指软件工程师。我找不到足够多的真正懂 AI 的市场人员、招聘人员和财务专业人士。转变教育体系，让学生和成年人都有能力使用这些工具，完成更多工作，这需要教育体系的巨大转变。但如何实现这一点，目前对我来说仍然非常具有挑战性。

So, as new capabilities come up, many people will embrace them, embrace ahead, and really do much more, be much more productive. And then sadly, there'll be people that don't embrace them that unfortunately will be left behind. I'm actually quite worried about that. One of the big challenges is if you look at the university system, many universities are slow to adapt curricula and are still training students for the jobs of 2022, but many of those jobs don't really exist. So, employers don't want to hire a student from 2022. But instead, many employers can't find enough talent that knows AI, knows how to build with AI. And I don't mean software engineers. I can't find enough marketers and recruiters and finance professionals that really know AI. Shifting the educational system to give students, as well as adults, the ability to use these tools and get so much more work done is going to require a huge shift of the educational system. But how to get there at this moment still seems very challenging to me.

要让 AI 取代你作为教育者的角色，需要什么条件？

What would it take for AI to replace you as an educator?

我希望我知道。事实上，我的团队经常尝试编写 AI 来取代我，而且是在我的祝福和强烈鼓励下。但依然不可能，对吧？遗憾的是，他们还没能取代我。我怀疑我是一种通用形式的 AI，而我们不知道学习是如何运作的。我想是的。如果我们有一天达到 AGI，那可能很棒。我可以退休，去做别的事情。我觉得取代很多熟练的人感觉像是一个 AGI 问题，感觉还很遥远。我认为对于狭窄的垂直领域，我们构建 AI 服务这些领域的能力会比通用 AI 增长得快得多。所以，我不知道。驾校其实是一个有趣的例子。事实证明，美国的驾校有点受到监管捕获的影响。所以，很多创新没有被采用，比如驾驶模拟器。在许多国家，驾驶模拟器的时间算作驾驶时长，但在美国不算。因此，驾驶模拟器在美国的使用不如其他国家多。所以，这种奇怪的事情非常令人沮丧。我认为美国应该更多地使用驾驶模拟器，因为这是教孩子的一种非常安全的方式。

I wish I knew. It turns out my team regularly tries to write AI to replace me, with my blessing and strong encouragement. And still it's impossible, right? Sadly, they've not managed to replace me yet. I suspect that I am a general form of AI where we don't know how learning works. I think so. If we ever get to AGI, it may be great. I could retire, go do something else. I feel like replacing a lot of skilled people feels like an AGI problem, which feels like it's still far off. I think for narrow verticals, our ability to build AI to serve narrow verticals will grow much faster than this whole for general AI. So, I don't know. Driving schools are actually an interesting one. It turns out that driving schools in the US have been subject a bit to regulatory capture. So, a lot of innovations haven't been adopted, like driving simulators. It turns out that driving simulators count in terms of number of hours driven in many nations, but not in the United States. Because of that, driving simulators are not used as much in the United States as in other nations. So, there's this weird stuff that is very frustrating. I think America should be using driving simulators more because it's a very safe way to teach kids.

程序员会失业吗？

Will programmers be unemployed?

我认为不使用 AI 的程序员会有麻烦。但真正懂 AI 的程序员效率非常高。我找不到足够多这样的人。他们是最后被淘汰的吗？最后被淘汰听起来有点可怕。我认为大多数工作岗位不会消失。但那句话‘AI 不会取代人，但使用 AI 的人会取代不使用 AI 的人’在很多情况下都是对的。现在，为了透明和完全诚实，有一小部分工作岗位被 AI 完全自动化了。所以，坦率地说，我认为法律翻译人员有麻烦。翻译人员，我认为配音演员也可能有麻烦。所以，坏消息是有一小部分工作岗位可以被 AI 完全自动化。我确实为他们感到难过，我认为我们有责任做很多事情，确保他们能获得新技能，重新加入劳动力队伍，找到其他有意义的事情做。作为 AI 从业者，我觉得有责任尽我所能确保人们得到照顾。但对于绝大多数工作，AI 可以自动化部分内容。放射科？我认为自动化放射科医生花费的时间比人们想象的要长得多。比早期的预测要长得多。我认为法律有很多监管需要应对。完全取代人类律师将非常困难，但不使用 AI 的律师会比使用 AI 的律师效率低得多。因为 AI 在法律研究方面非常出色。我认为 AI 可以很好地完成法律的部分工作，但这里有个窍门。如果一份工作 AI 能完成 30%，那么剩下的 70%仍然需要人类来做。但那个人最好使用 AI，因为不使用 AI 的人会失去生产力。

I think programmers that don't use AI will be in trouble. But programmers that really know AI are so productive. I just can't find enough of them. Are they last to go? Last to go sounds a bit dire. I think that most job roles are not going anywhere. But that saying, 'AI won't replace someone, but someone that uses AI will replace someone that doesn't,' that's true a lot of the time. Now, just to be transparent, just to be completely truthful, there is a small number of job roles that are fully automated by AI. So, candidly, I think law translators are in trouble. Translators in general, I think voice actors could be in trouble as well. So, the bad news is there's a very small fraction of job roles that AI can automate entirely. Those I actually feel for them and I think we owe it to them to do a lot to make sure that they can gain new skills, rejoin the workforce, find other meaningful things to do. As an AI person, I feel a duty to do whatever I can to make sure people are taken care of. But for the vast majority of jobs, AI can automate parts. Radiology? I think it's taken much longer than people thought to automate radiologists. It's taken much longer than the early predictions. I think law has so much regulation to capture. It'll be very hard to completely replace human lawyers, but lawyers that don't use AI will be so much less productive than lawyers that do use AI. Because AI is very good at legal research. I think AI could do parts of law really well, but here's a trick. If there's a job and AI could do 30% of it, then that 70% you still need a human to do it. But that human had better use AI because one that doesn't use AI is missing out on productivity.

你能预测哪些工作最终会消失吗？

Could you predict which jobs could finally disappear?

嗯，明显的是一些客服工作正在消失。翻译工作、配音演员工作。我认为有麻烦的是那些几乎 100%的工作或接近 100%的工作可以被自动化的工作。但事实证明，很多工作非常复杂、多面，几乎同等包含文本和非文本内容。所以，对于大多数工作，我认为我们看的是基于任务的分析。有一些研究，比如 Erik Brynjolfsson 和其他人做的。拿一份工作，分解成任务，看看哪些任务可以被 AI 自动化。对于很多工作，AI 可以自动化大约 30-40%的工作内容。所以你仍然需要人类来做那 60-70%。还有一小部分工作，AI 几乎可以自动化一切。所以，那些工作有麻烦，但那是非常少数的。

Well, the clear ones are a lot of call center jobs are going away. Translator jobs, voice actor jobs. I think the ones that are in trouble are the ones where almost 100% of the work or 100% almost 100% of the work can be automated. But it turns out that so many jobs are so complex, multifaceted, almost equally text and also equally non-text things. So, for most jobs, I think we look at these task-based analyses of jobs. There are studies that Erik Brynjolfsson and others have done. Take a job, break it into tasks, see what tasks AI can automate. For a lot of jobs, AI can automate like 30-40% of someone's job. And so you still need a human to do that 60-70%. And there is a small fraction where AI can automate almost everything. So, those jobs are in trouble, but that's a very small minority of the jobs out there.

你在 2017 年离开了百度，对吗？你决定离开的原因是什么？

You left Baidu in 2017, right? What was the reason that you decided to leave?

所以，我在百度度过了一段美好的时光。有趣的是，人们总是想知道我从 Google Brain 到 Google，然后从百度到 Running AI Fund 和 DeepLearning.AI，人们总是觉得有什么秘密。但实际情况是，我觉得我在百度领导 AI 团队，一个很棒的团队，我认为团队在建设母舰方面做得很好，对吧？所以，在线广告、改进网页搜索，所有核心业务，团队在那里做得很好。我记得看着组织架构图，然后想，‘你知道吗？如果我不在这里，我们有一个非常好的团队。没有我他们也能做得很好。’然后我也意识到，当时我工作中最有趣的部分实际上是建立新的业务部门。我们为母舰赚钱做得很好，我为团队的工作感到自豪。但我发现，我个人最享受的是建立新业务。例如，我领导了百度的自动驾驶汽车团队，今天在中国仍然做得很好。我领导了百度的智能音箱团队，像 Alexa 或 Siri，今天仍然做得很好。我觉得虽然我可以为母舰赚钱，但我真正投入的是建立新业务。

So, I had a great time at Baidu. And it's funny, people keep wondering when I move from Google Brain to Google and then Baidu to Running AI Fund and DeepLearning.AI, people always wondered there's some secret thing. But what happened is, I feel like I was running the AI team at Baidu, a great team, and I think the team did a wonderful job building the mothership, right? So, online advertising, improving web search, all the core businesses, and the team was doing a great job there. I remember looking at the org chart and then thinking, 'You know what? If I weren't here, we've got a really good team. They'd really do just fine without me.' And then I also realized that the most fun part of my job at the time was actually building new business units. We did a great job making money for the mothership, really proud of the work that the team did. But I found that where I personally had the most fun was building new businesses. So, for example, I ran Baidu's self-driving car team, still doing really well in China today. I ran Baidu's smart speaker team, like Alexa or Siri, still doing really well today. And I felt that while I could make money for the mothership, where I was really engaged was building new businesses.

然后我问自己，虽然在大公司里做这件事也不错，但我觉得也许如果我开始做点别的——后来就成了风险工作室 AI Fund，从零到一打造企业，而不是在大公司背景下——也许我能做得更好。所以这就是我离开百度的原因，开始投入更多精力做教育，赋能人们用 AI 构建，同时还运营 AI Fund，一个风险工作室，我们在这里创建初创公司。

And then I asked myself while doing it within the context of a big company was fine, I felt that maybe if I start something else, which turned out to be the venture studio AI Fund, to build businesses more from zero to one rather than the context of a big company, maybe I could make that work even better. So, that's why I stepped away from Baidu to start doing a lot more work in education, to empower people to build with AI, and then additionally to run AI Fund, a venture studio, where we build startups.

你认为中国现在领先了吗？

Do you think China is now ahead?

我认为 AI 是多方面的，中国在某些方面领先美国，比如开源开放权重模型，而美国在某些方面领先中国，比如专有模型。所以我不后悔自己的选择。我觉得和中国团队一起工作很开心。我在 DeepLearning.AI、AI Fund、最近的 AI Ascend、Landing AI 工作也很愉快。所以我对现在做的事情很满意。没有遗憾。

I think AI is multifaceted, and China is ahead of the US in some places like open source open weight models and the US is ahead of China in some places like proprietary models. So we don't regret your choice. I think I had a great time doing some work with teams in China. I've had a great time doing work with DeepLearning.AI, AI Fund, more recently AI Ascend, Landing AI. So I'm happy with what I'm doing. No regrets.

从纯粹的规模角度看，中国不是更有能量吗？

From pure scale perspective, doesn't China have more energy?

中国有很多优势。美国也有很多优势。我觉得，尽管这个联盟不完美，但我热爱美国和西方世界，即使有些年份民主运作得更好，有些年份更差，这仍然是我们美国做事方式的重要支柱。我认为美国还有很多好工作要做，我很兴奋。

China has a lot going for it. The US also has a lot going for it. I feel like, as imperfect as this union is, I love America and the Western world and even though some years it feels like democracy works better and it works worse in some years, that feels like an important pillar of how we do things in America and I think there's a lot of good work to be done still in America and I'm excited.

我们可以期待你的公司带来什么？

What can we expect from your company?

我想赋能每个人用 AI 构建。我们一直专注于帮助开发者获取最新工具，并继续在这方面努力。除了真正支持 AI 开发者建立和发展他们的职业生涯，我实际上想拓宽我们的工作，赋能开发者和所有人用 AI 构建。这就是 DeepLearning.AI 的重点。我离开百度的原因之一是：如果我同时领导 Google 和百度的团队，有些业务在 Google 内部构建有意义，在百度内部构建也有意义，但还有其他事情比如关税合规。互联网公司为什么要关心这个？所以 AI Fund 构建了许多不同的初创企业，我很兴奋于我们在 AI Fund 能处理的各种各样的事情。然后 AI Ascend，一个相对较新的努力，是我和朋友们 Chris 和 Tan 一起为大型企业提供 AI 咨询。事实证明，如果你想真正推动 AI 采纳，开发者很重要，个人消费者也很重要，我们还需要让大型企业参与进来。所以我们花了很多时间与 Bain 和其他公司合作。Bain 是一个很棒的团队。很荣幸与 Kristoff 和他的团队合作。但目的是帮助大型企业建议如何驱动真正价值，如何真正用 AI 取胜。

I want to empower everyone to build with AI. We've been very focused on helping developers get access to latest tools and continue to work hard on that. In addition to really supporting AI developers, build their careers, grow their careers, I actually want to broaden what we do to empower developers and everyone else to build with AI. So that's DeepLearning.AI's focus. And one of the reasons I stepped away from Baidu: if I found both at once leading a team at Google and at Baidu, there's certain businesses that make sense to build within Google, that make sense to build within Baidu, but there are other things like tariff compliance. Why would an internet company care about that? So AI Fund builds lots of different startup businesses, and so I'm excited about just the diversity of stuff we get to work on at AI Fund. And then AI Ascend, which is a relatively new effort, is my friends Chris and Tan and I working on AI advisory for large enterprises. It turns out that if you want to really move the needle on AI adoption, developers are important, individual consumers are important, and we just got to get large businesses there. So, we're spending quite a bit of time in partnership with Bain and others. Bain is a fantastic team. Privilege working with Kristoff and his team over there. But to help advise large businesses on how to drive real value and how to really win with AI.

你也决定成为 AI 传播者，原因是什么？

You've decided to be also, I think, AI communicator. What was the reason?

是的，你触及了非常核心的价值观。我告诉你我如何优先安排我的工作。我认为有两件事是我最优先考虑的，我真正相信的。我认为这是对我时间的好利用。一是让人类更强大。这就是为什么我成为研究者，我觉得。成为斯坦福教授，因为我认为研究、推动技术前沿，通过发明新技术让人类更强大。所以，我相信让人类更强大。我深信的第二件事是帮助他人实现他们的梦想。重要的是帮助他人实现他们的梦想，而不是帮助他人实现我的梦想。这是一个重要的区别。在我的一生中，我觉得如果我们能给别人工具和技能，那就能让他们更好地实现梦想，这就是为什么教育，通过 DeepLearning.AI、Coursera，一直是我觉得非常有动力的事情。

Yeah, you're touching on a very core values kind of thing. I'll tell you how I prioritize what I do. I think there are two things that I prioritize most highly, that I really believe in. I think it's a good use of my time. One is things that make humanity more powerful. That's why I became a researcher, I feel. Became a professor at Stanford, because I think research, advancing the state of the art, that makes humanity more powerful by inventing new technologies. So, I believe in making humanity more powerful. The second thing I deeply believe in is to help others realize their dreams. It's important that it's help others realize their dreams, not help others realize my dreams. That's an important distinction. And all through my life, I felt that if we can give others tools and skills, then it puts them in a better position to realize their dreams, which is why education, through DeepLearning.AI, Coursera, has always been something I found very motivating.

回顾你的故事，有什么遗憾吗？如果时光倒流，你会改变什么吗？

Is there something that you regret when you look at your story? If you could turn back time, would you change something?

嗯，有很多事情我本可以做得不同。所以，坦白说，我不知道。我很幸运做出了一些好的决定，但我也做了很多糟糕的决定。我觉得有很多事情，比如，我不知道，我是否应该雇佣那个人，或者做那个项目，或者我是否应该在那个项目上更努力而不是放弃？

Well, there's so many things I could have done differently. So, frankly, I don't know. I've been fortunate to have made a few good decisions, but I've made so many bad ones as well. I think there's so many things I, like, I don't know, should I have hired that particular person or done that particular project or should I have worked harder on that project rather than giving up?

如果你能找到一个关于现实的答案，那会是什么？

If you could find answer for one question about the reality, what would it be?

我希望我理解智能的本质是什么。

I wish I understood what is the nature of intelligence.

你是指意识吗？物质如何在人脑中转化为意识，还是更多？

You mean consciousness? How matter is transformed into consciousness inside a human brain or something more?

实际上，不是意识。我认为意识是一个重要的哲学问题，但我不知道意识是什么。所以，哲学家谈论意识是指自我意识这个概念，但结果是你实际上不知道我是否有意识，我也不知道你是否有意识，对吧？你怎么知道？在哲学中，有一个概念，也许我只是一个僵尸，我实际上没有意识，我只是在移动我的手，移动我的嘴，假装有意识。所以，因为你无法接触我的内在体验，反之亦然，我们实际上不知道其他人是否有意识，但我认为出于礼貌，我们假装其他人像我们认为自己一样有意识。所以，因为意识不可测量，对我来说，这使它成为一个哲学问题而不是科学问题。虽然哲学很重要，但我更倾向于科学问题。对我来说，智能的本质到底是什么机制让人类大脑或其他生物大脑展示出我们看到的如此广泛的智能行为？比如这到底是如何工作的？顺便说一句，有一件不广为人知的事情。在开始 Google Brain 团队之前，我做的其中一件事是经常和我的神经科学家朋友在一起，我读了大量的神经科学论文，并得出结论，恕我直言，神经科学基本上不知道大脑是如何工作的。所以我放弃了神经科学作为构建智能的路径。但理解智能到底如何工作？智能的本质是什么？我很确定它不是带有缩放定律的 Transformer 网络。我很确定它需要更多。但我希望我能……

Actually, not consciousness. I think consciousness is an important philosophical question, but I don't know what is consciousness. So, philosophers talk about consciousness in the sense that is this notion of being self-aware, but it turns out that you don't actually know if I'm conscious and I don't actually know if you are, right? How do you know? In philosophy, there's this concept that maybe I'm just a zombie and I'm not actually conscious, but I'm just moving my hands, moving my mouth, and pretending to be conscious. So, because you don't have access to my inner experience and vice versa, we don't actually know if anyone else is conscious, but I think out of politeness, we pretend everyone else is conscious as much as we think we are. So, because consciousness is not measurable, to me, that makes it a philosophical rather than a scientific question. And while philosophy is important, I gravitate more to the scientific questions. And to me the nature of intelligence is what on earth are the mechanisms that allow the human brain or maybe other biological brains to demonstrate this huge range of intelligent behaviors that we see. Like how on earth does this work? By the way, something that is not widely known. Before starting the Google Brain team, one thing I did was hang out a lot with my neuroscientist friends and I was reading huge piles of neuroscience papers and concluded, with all due respect to my neuroscience friends, but neuroscience has no idea how the brain works so frankly pretty much. So I gave up on neuroscience as a path to building intelligence. But understanding how does intelligence actually work? What is the nature of intelligence? I'm pretty sure it's not a transformer network with scaling laws. I'm pretty sure it needs more than that. But I wish I would...

也是推理的本质吗？

The nature of reasoning also?

是的。

Yeah.

我认为推理是智能的一个子集，但理解推理的实际运作方式也会很迷人。为什么在西方世界我们如此不快乐？我不同意他关于西方世界迷失的说法。我不从喝的水中寻找快乐，但我喝的水让我更健康。我认为从 AI 中寻找快乐是个错误，但 AI 确实能帮助人们。它帮助我们整体，帮助我们构建事物，赋予我们技能。所以 AI 很重要，但从 AI 中寻找快乐是个错误，就像从我们建造的器物中寻找快乐一样。我认为快乐更多来自内心，而非我们建造的器物，但正是那种努力帮助他人的过程。我希望很多人能在其中找到快乐，我也希望通过我在 AI 领域的工作帮助他人，并且我在能够做对他人有用的事情中找到了很多快乐。

I think reasoning is a subset of intelligence, but understanding how reasoning actually works would also be fascinating. Why in the Western world are we so unhappy? I really disagree with his characterization of the Western world as lost for these reasons. I don't look for happiness in the water I drink, but I drink water that keeps me healthier. I think it'd be a mistake to look for happiness in the AI, but AI really helps people. It helps us as a whole. Helps us build things. Helps give us skills. And so AI is important, but it's a mistake to look for happiness in AI in the same way that it would be a mistake to look for happiness in the artifacts that we build. I think happiness comes much more from within than from the artifacts that we build, but it is that process of striving to help others. I hope a lot of people can find happiness in that and I hope I can help others through my work in AI and I find a lot of joy in being able to do work that could be useful to others.

非常感谢您的时间。

Thank you very much for your time.

谢谢。

Thank you.