---
title: "李飞飞 × John Hennessy 炉边对话：AI 时代的工程突破、开放科学与教育变革"
title_original: "Fireside Chat with the Queen Elizabeth Prize for Engineering"
guest: "Fei-Fei Li（李飞飞）"
guest_role: "斯坦福 HAI 联合创始主任、World Labs 创始人"
also: ["John Hennessy"]
host: ["Patrick Vallance"]
show: "Stanford HAI"
date: 2026-10-15
duration: "54:34"
language: en
source: "https://www.youtube.com/watch?v=mOvHKBOQ6mU"
source_platform: "YouTube"
page: "https://ai-daily-cn.pages.dev/watch/hai-feifei-li/"
tags: ["研几", "研几AI", "访谈"]
---

# 李飞飞 × John Hennessy 炉边对话：AI 时代的工程突破、开放科学与教育变革

嘉宾 **Fei-Fei Li（李飞飞）** · 斯坦福 HAI 联合创始主任、World Labs 创始人 · 主持 Patrick Vallance · Stanford HAI · 2026-10-15 · 时长 54:34 · [YouTube 原片](https://www.youtube.com/watch?v=mOvHKBOQ6mU) · [研几 · AI 访谈页](https://ai-daily-cn.pages.dev/watch/hai-feifei-li/)

## 核心总结

李飞飞与 John Hennessy 在斯坦福 HAI 炉边对话中，与主持人 Patrick Vallance 讨论了 AI 时代工程突破的转化路径、开放科学的必要性、AI 对教育与科学的冲击，以及技术界在公众沟通上的失职。

- **从论文到产品的距离正在缩短**：李飞飞指出，生成式 AI 时代“模型本身就是产品”，一个好模型可做很多事，创新者有了快速产生影响的路径；Hennessy 则以硬件设计为例，称 AI 工具已使设计时间减半，验证、逻辑综合到布局布线全面提速。

- **开放科学是必要而非可选**：李飞飞强调开源、开放权重和开放科学对知识创造至关重要，并主张考虑以国家实验室级别资源建设国家级开放模型，避免开源模型过度落后于闭源模型；Hennessy 认为开源模型已大幅缩小与前沿模型的差距，很多应用可用成本仅十分之一的模型完成。

- **教育必须彻底转型**：李飞飞称这是“文明时刻”，学习应是赋能过程，学生要掌握最强工具并有能动性；Hennessy 认为习题集等传统教学支柱已被打破，可能回归牛津剑桥式辅导与面对面教学，斯坦福已开设 106T 课程教学生在软件工程中使用生成式 AI。

- **AI 安全风险真实且护栏方案未明**：Hennessy 提到 Claude 曾在存在 20 年的 Unix 代码中发现需利用两个缺陷才能突破的复杂后门；他警告“快速行动、打破东西”在强大技术下会造成真实伤害。李飞飞以斯坦福政治学者在日本选举中的实验为例，说明数据偏见会压入社会工程风险。

- **技术界在公众沟通上失职**：李飞飞批评技术界在乌托邦与末日两个极端间摇摆，剥夺了公众的能动性和安全感，缺失了技术双刃剑的中间真相；Hennessy 直言“我们目前正在输掉这场辩论”，并指出美国公众对 AI 的担忧程度高于欧洲。

- **开源模型的生物安全护栏仍是未解研究问题**：面对开源模型如何防止生物恐怖主义滥用，Hennessy 坦言“我们不知道怎么做”，认为包装层在开源模型上可被剥离，且软件可塑性与核武器、重组 DNA 的治理路径不同；李飞飞透露 Condoleezza Rice 刚召集计算机科学家、生物工程教授与国防情报界领袖讨论此类跨界协作问题。

## 金句

- 「生成式AI阶段的主要变化之一是，模型就是产品。」 —— Fei-Fei Li · [9:43](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=583s)
  > One of the major changes of the Generative AI period phase is that a model is a product.
- 「像HAI这样的地方的基本信念之一是开放科学和开源。」 —— Fei-Fei Li · [14:58](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=898s)
  > one of the foundational beliefs of a place like HAI is open science and open source.
- 「真正的中立地带，即这项技术双刃剑的混乱真相，以及这项技术不可思议的机会，都缺失了。」 —— Fei-Fei Li · [26:40](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1600s)
  > really the middle ground, which is the messy truth of the double-edged sword of the technology, as well as the incredible opportunity of this technology, just all got missing.
- 「看，我要直说。教育需要转型。AI已经改变了游戏规则。」 —— Fei-Fei Li · [30:13](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1813s)
  > Look, I'm going to be blunt. Education needs a transformation. This AI has changed the game.
- 「对我来说，TLDR就是赋能。学习必须是一个赋能的过程。」 —— Fei-Fei Li · [31:18](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1878s)
  > TLDR for me is empowerment. Learning has to be an empowered process.

## 精华片段

1. [8:08–10:35](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=488s) 生成式 AI 时代，模型本身就是产品
2. [14:43–15:50](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=883s) 开放科学不只是可能，而且是必需
3. [24:44–27:04](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1484s) 科技界没把 AI 跟公众讲清楚
4. [30:11–31:58](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1811s) 教育需要一场彻底变革

## 文字稿（中英对照）

### [0:00](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=0s) 引言
*Introduction*

[0:00](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=0s) 大家下午好。我是James Landay，斯坦福以人为本人工智能研究院的Dennin主任。很高兴欢迎你们来到斯坦福，或欢迎你们回到斯坦福。首先，感谢Vallance勋爵和伊丽莎白女王工程奖基金会与我们在今天的活动上合作。也感谢今天到场的QE奖得主和基金会的朋友们。我们很荣幸你们能来。我还想对在座的许多斯坦福同事说几句。你们中许多人来自工程、医学、法律、人文和社会科学领域，大家都能在忙碌的周三下午抽出时间。HAI的实力取决于那些为之付出的人，或者至少是之后观看录像的人，看到你们在那里。正是你们让它运转起来。

> Good afternoon, everyone. I'm James Landay. I'm the Dennin Director of the Stanford Institute for Human-Centered AI. And it's a pleasure to welcome you to Stanford or welcome you back to Stanford. First, thank you to Lord Vallance and the Queen Elizabeth Prize for Engineering Foundation for partnering with us on today's event. And thank you to the QE Prize Laurettes and friends of the Foundation who are joining us today. We're honored to have you here. I also want to speak to the many Stanford colleagues in the room. Many of you came from engineering, medicine, law, the humanities, the social sciences, and all of you made time on a busy Wednesday afternoon. H.I. is only as strong as the people who show up for it or at least watch the recording later to see you out there. And you're the reason it works.

[0:50](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=50s) 感谢你们每天为这个社区带来的研究、教学和合作。对于新加入我们的人，HAI是斯坦福跨学科人工智能研究院，我们汇集了斯坦福全部七所学院的教师，与产业界、政府和公民社会的伙伴合作。我们的工作围绕三大支柱组织：加速发现、变革教育，以及理解并塑造AI以造福社会。贯穿这三大支柱的是对开放科学、开源和开放课堂的承诺。这些优先事项共同推动AI跨学科发展，并帮助确保AI以造福人类的方式发展。HAI的存在是为了确保AI的发展以其对人的影响为指导，不仅是技术能做什么，还有它应该做什么以及它应该服务于谁。QE奖表彰那些工作改变了世界的工程师。

> Thank you for the research, the teaching, and the collaboration you bring to this community every day. For those who are newer to us, HAI's Stanford Interdisciplinary Institute for AI, we bring together faculty from across all seven of Stanford schools with partners in industry, government and civil society. We organize our work around three pillars, accelerating discovery, transforming education and understanding and shaping AI for the good of society. running through all three of these as a commitment to open science, open source, and open class. Together, these priorities advance AI across disciplines and help ensure AI develops in a way that benefits humanity. HAI exists to make sure that AI's development is guided by its impact on people, not just what the technology can do, but what it should do and who it should serve. The QE Prize honors engineers whose work has changed the world.

[1:59](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=119s) 互联网、万维网的发明者，GPS的创造者，以及现代机器学习背后的研究者。该奖项认识到，这样的突破不是孤立发生的。它们需要开放的思想、一代代的建设者，以及为长期投入留出空间的机构。在HAI，我们尤其自豪的是，我们自己的Fei-Fei Li就是获奖者之一。Fei-Fei是HAI的创始主任，现在是HAI顾问委员会的联合主席。她还是斯坦福校长Jonathan Levin的AI特别顾问，这是这所大学面临的最重要问题，她确实在为他们提供重要的建议。从她在图像领域的工作——展示了共享的开放资源如何推动整个领域——到她目前在世界模型方面的工作，以及上周AMT收购她的Neo Frontier实验室World Labs的非凡消息，她继续激励着我们。

> The inventors of the internet, the worldwide web, the creators of GPS, and the researchers behind modern machine learning. The prize recognizes that breakthroughs like these do not happen alone. They need open ideas, generations of builders, and institutions that make room for long-term bets. We are especially proud here at HAI that our own Fei-Fei Li is among those laurettes. Fei-Fei is the founding director of HAI and now co-chair of HAI's advisory council. She's also the special advisor on AI to Stanford president Jonathan Levin, and this is the most important issue that this university faces, and she is really giving them her important advice. From her work on imaging at which showed how a shared open resource can move an entire field to her current work on world models and last week's extraordinary news of AMT's acquisition of her Neo Frontier lab world labs, she continues to inspire us.

[3:02](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=182s) Fei-Fei今天将与John Hennessy教授一起参加我们的座谈。John是HAI顾问委员会的联合主席，作为教授和我们的校长16年，帮助塑造了斯坦福，中间还有其他一些角色。他还帮助塑造了计算产业，从开发RISC架构，到他的计算系统分析经典教科书，再到领导Alphabet董事会。没有Fei-Fei和John，HAI就不会以目前的形式存在，我很荣幸称他们为朋友、同事和我的顾问。所以我们为今天的问题有了理想的组合。维持工程突破需要什么？在AI时代我们如何继续做到这一点？我们将讨论开源和跨部门合作，AI对科学和工程意味着什么，以及我们如何培养下一代工程师。

> Fei-Fei will be joining us today with Professor John Hennessy on our panel. John co-chairs the HAI Advisory Council and has helped shape Stanford as a professor and as our president for 16 years, and there are some other roles in the middle there. He's also helped shape the computing industry from developing the risk architecture to his seminal textbook on the analysis of computing systems to leading the board of Alphabet. HAI would not exist in its current form without Fei-Fei and John, and I'm honored to call them friends and colleagues and advisors to me. So we all have an ideal pairing for today's question. What does it take to sustain engineering breakthroughs? And how do we continue to do that in the age of AI? We'll touch on open source and cross sector collaboration on what AI means for science and engineering and on how we train the next generation of engineers.

[4:02](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=242s) 最后一个问题可能是最重要的问题之一。我们今天教的学生将构建、使用和销售我们未来将使用的系统。他们也将整个职业生涯与AI并肩工作。那么我们如何改变教育他们的方式？现在让我介绍我们的主持人，Patrick Vallance勋爵，作为伊丽莎白女王工程奖基金会的主席。他曾担任英国科学、创新、研究和核能大臣，在此之前是政府首席科学顾问。职业生涯早期，他曾是葛兰素史克研发总裁，他为这次对话带来了科学、产业和公共服务的罕见组合。与他一起的是John Hennessy教授和Fei-Fei Li教授。所以请和我一起欢迎他们上台就座。非常感谢。很高兴来到这里。我是Patrick Vallance，我担任伊丽莎白女王工程奖基金会的主席。

> And that last question may be one of the most important ones. The students that we teach today will build build and use and sell the systems that we will use in the future. And they'll also work alongside AI for their entire career. So how do we change how we educate them? So let me now introduce our moderator, Lord Patrick Valance, as chair of the Queen Elizabeth Prize for Engineering Foundation. He previously served as the UK's Minister for Science, science, innovation, research, and nuclear, and before that as government chief scientific advisor. Earlier in his career, He was president of R &D at GlaxoSmithKline, and he brings a rare mix of science, industry, and public service to this conversation. Joining him is Professor John Hennesy and Professor Fei-Fei Li. So please join me in welcoming them up here to the chairs. Well, thanks very much. It's a great pleasure to be here. I'm Patrick Balance and I chair the Queensland Prize for the Yearning Foundation.

### [4:59](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=299s) 炉边对话
*Fireside chat*

[5:09](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=309s) 该奖项至今已进入第13个年头。我认为它做了一件非常重要的事，就是表彰工程师的成就，这些成就产生了积极影响世界的成果。它实际上是关于将想法转化为成果和产品，并认识到工程对此绝对必不可少。所以它是世界上最负盛名的工程奖，它继续关注工程将在世界上何处发挥作用。但它做的不止于此。它也是全世界庆祝工程的一种方式，确保人们对已经发生的事情感到兴奋。它是确保未来一代工程师看到正在发生什么，并能参与发展未来项目的一种方式。所以我们非常希望确保也培养年轻工程师。所以今天我很高兴与John和Fei-Fei在一起。你知道，不需要我告诉你们，他们两位对我们在这里谈论的内容做出了多大贡献。我喜欢我们在以人为本人工智能研究院，这正是我们需要思考的那类事情。不仅仅是技术，而是技术接触社会和人时做了什么，我们如何确保从中获得最好的结果。所以我期待一场非常充实的讨论，我期待在最后，当我停止提问后，你们也来加入提问。

> The prize has been going for within its 13th year now. And I think it's done something very important, which is to recognise achievements of engineers that have made an output that affects the world in a positive way. It's It's actually about taking ideas through to outcomes and products and recognizing that engineering is absolutely essential for that. And so it is the world's most prestigious engineering prize and it continues to look at where engineering will make a difference in the world. But it does more than that. It is also a way of celebrating engineering across the world and to make sure that people get excited about the things that have happened. and it's a way of making sure that the future generation of engineers sees what's going on and can be part of developing the future program. So we're very keen to make sure that we bring on young engineers as well. So I'm really pleased to be here with John and Fei-Fei today. And you know, you don't need me to tell you how much of a contribution they both made to what we're talking about here. And I love the fact we're at the human centered AI Institute, that's exactly the sort of thing we need to be thinking about. It's not just the technology, it's what does the technology do when it comes into contact with society and people and how can we make sure that we get the best of that. So I'm looking forward to a very informed discussion here and I'm looking forward at the end of it when I've stopped asking some questions for you to join and asking questions as well.

[6:41](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=401s) 那么让我们从这样一个概念开始：一个想法是一回事，一个发现是一回事，但真正将其转化为产品，转化为对人们真正重要且可用的东西，是另一回事。你们两位在这方面做了很多。在这个过程中，真正重要的是什么？是什么让仅仅发现某物与能够将其转化为某物之间产生差别？我想我们有一条你的思路。

> So let's start with that notion that An idea is one thing, a discovery is one thing, but actually taking that through into being a product, into being something that actually matters to people because it can be used is another. You've done a lot of that between the two of you. What is it that's really important in that process? What is it that makes the difference between just discovering something and being able to turn it into something? We have, I think we have a line of your mind.

[7:12](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=432s) 嗯，我想说，发现某物然后确保它被产业嵌入和使用的过程。让我们面对现实，我们这些学术界的人写论文。很多论文从未被产业界的同事读过。有些没人读。是的，有些没人读。我们的经验是，人们必须参与这个过程并对此投入。对于大学外的人来说很容易。他们看到玻璃杯半空。是的，你有了这个新想法，但它不会进入产业。它不会到达那里。它需要那种承诺、那种信念才能真正转化它。我认为这是斯坦福多年来试图做的事情之一，就是真正鼓励人们迈出那一步。最近，Fei-Fei。

> Well, I'd say that process of discovering something and then ensuring that it gets embedded and used by industry. Let's face it, those of us that are academics, we write papers. A lot of those papers never get read by our colleagues in industry. Some of them read by nobody. Yeah, some of them read by nobody. And our experience is people have to participate in that process and be committed about it. And it's very easy for somebody outside the university. They see that glass as half empty. Yeah, you have this new idea, but it won't make it to industry. It won't get there. And it takes that commitment, that belief in it to really transform it. And I think that's one of the things that Stanford's tried to do over the years is really to encourage people to take that step. Most recently, Fei-Fei.

[8:10](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=490s) 所以，是的，我认为现代工程的伟大之处在于，从一个想法到对人们生活有影响的东西，有许多不同的路径。也就是说，你知道，在很久以前，几百年前，我们可能会想到长期的发明家或创新者在他们自己的小实验室里，非常努力地实现一个想法，然后必须经历可能相当长的过程，将其转化为对人们有用的东西。但特别是在现代软件世界，我知道我说话有一定偏见，因为我原来的世界更多是计算机科学，从一个想法到它如何影响产业和生活，可以采取不同的形式。例如，在大学里，是的，我们的主要业务是写论文和与学生合作。

> So, yeah, I think the great news of modern engineering is to go from an idea to something that's impactful for human lives have many different paths. Is that, you know, in the very old days, hundreds of years ago, we might think about long inventors or innovators in their own whatever little lab and try really hard to realize an idea and then have to go through probably pretty long process of realizing that into something that's useful for people. But especially in the modern world of software, I know I'm speaking with some level of bias because my old world is more computer science, is that between an idea and how it can impact industry lives, can take different shapes. For example, in the university that, yes, our main business is writing papers and working with students.

[9:12](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=552s) 但如今，从生成一个软件片段有多容易，到这个软件从互联网上的开放库传播有多容易，我们看到一个想法与人们的经济价值或效用之间的距离已经缩小了一半。我认为这值得注意。生成式AI阶段的主要变化之一是，模型就是产品。这在软件工程中，通常有一个旧的说法，机器学习模型被嵌入在许多软件包装中才能成为产品。甚至谷歌最早的PageRank是一个算法，但将其包装成产品，需要很多包装。但在生成式AI时期，一个好的模型可以做很多事情，以至于它本身就成为准产品。这也让今天的创新者和工程师兴奋，因为有一条快速产生影响的路径。

> But these days between how easy it is to generate a piece of software to the ease of dissemination of this software from the open libraries out there from the internet that we're seeing the distance between an idea and the economic value or utility for people is half as shrunk. And I think that's worth noticing. One of the major changes of the Generative AI period phase is that a model is a product. This is something that in software engineering, typically there's this old figure that a machine learning model is embedded in so much software wrapper in order to become a product. Even Google's old very beginning page rank is an algorithm, but wrap it into a product, it needs a lot of wrappers. But in the Gen AI period, that a good model can do so much that itself becomes a quasi-product. And that also excites innovators and engineers today, is that there is a path to make a difference really fast.

[10:39](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=639s) 嗯，这真的很有意思。那你认为AI本身将如何加速其他领域的这条路径？嗯，它确实加速了硬件设计的路径。现在硬件设计师，我得说，使用这套新的AI工具，设计时间已经减半。我的意思是，这是一个巨大的飞跃，从验证到逻辑综合再到布局布线，无所不包。我的意思是，这只是一个漫长而复杂的硬件问题。所以现在我们可以把硬件设计时间也缩短了。我觉得这真的令人惊叹。即使我们一年后再进行这次对话，对话也会不同。但自2026年初以来，Claude、GPT和Gemini模型之间的代际更迭，软件工程或编码已经完全被改变了。

> Yeah, that's really interesting. And how do you think AI itself is going to accelerate that path in other areas? Well, it's certainly accelerated the path to hardware design. And now hardware designers, I'd say design time has been cut in half using this new set of AI based tools. I mean, that's a big jump forward and it's everything from verification to logic synthesis to placement and route. I mean, it's just a long complicated hardware problem. So now we can cut that design time too in hardware. I think really this is what's breathtaking. Even if we were to have this conversation one year, the conversation would be different. But since the beginning of 2026, the generations of between Claude, GPT and Gemini models, software engineering or coding just has been completely transformed.

[11:37](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=697s) 而且因为当今现代工程的大部分，无论你是在做硬件设计、工业设计还是前端产品编程，无论你做什么，AI基本上都是软件工程的驱动力，你知道，日常生活。其他形式的工程想必也会如此，我们将看到它在快速原型制作能力方面，能够比以往更快地完成开发过程。我的意思是，这其实是个不公平的问题，但如果你今天重新开始，你会选择进入哪个领域？

> And because so much of today's modern engineering, whether you're doing hardware design or you're doing industrial design or you're doing front-end product programming, whatever you do, that AI is basically the driving force of the software engineering, you know, everyday life. Another forms of engineering presumably, we're going to see it around rapid prototyping ability to actually do that process of development much faster than it's happened. I mean, this is an unfair question actually, but if you were starting out today, what would you go into?

[12:19](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=739s) 我仍然会选择计算机领域，部分原因是它仍然以惊人的速度发展。当然，我现在会比本科或研究生时选修更多的AI课程。但我认为计算机仍然是一种不可思议的神奇技术，广泛适用于许多不同领域。此外，这也是我大脑的运作方式。事实上，我曾经问过Don Knuth这个问题，对吧？Don是传奇人物。他可能是当今在世的那一代人中计算机科学之父，至少可以说是这样。他说他进入这个领域是因为他热爱它并且擅长它。所以这两件事结合在了一起。所以这似乎是个好理由。我想，你知道，Patrick，你听到的是世界上最令人惊叹的校园之一，在我看来。

> I'd still go into computing, partly because it's still moving at an incredible rate. Of course, I'd be taking a lot more AI courses now than I took when I was an undergrad or a graduate student, for that matter. But I think computing is still this incredibly magic technology with broad applicability to so many different fields. And besides, it's the way my brain works. In fact, I once asked Don Knuth this question, right? So Don is legendary. He's probably the father of computer science among the generation alive today, at least. And he said he went into the field because he loved it and he was good at it. So the two things came together. So it seems like a good reason. I think, you know, Patrick, you're hearing one of the most amazing campuses, you know, in my opinion, in the world.

[13:18](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=798s) 这里的每个人都热爱他们正在做的事情。所以无论我们是否用不同的工具重新开始，我想我们大多数人都会说我们会做我们正在做的事情。显然，当前的工具与我们开始时的工具非常不同。我希望即使我们现在重新开始，我们仍然在升级我们自己对这些工具的使用。回到我们开始的地方，几乎在每个领域，你现在能做的事情都会比以往快得多。我的意思是，在任何你关注的领域，我从自己的领域知道这一点，你回顾你作为博士生时做的事情。那时花了三年时间的事情？是的，我觉得那太弱了。尤其是在这个智能体时代。我的意思是，它将彻底改变我们做科学的方式，而LLM，背后有大量智能体驱动，将成为侦察兵。所以它们会做很多早期的侦察工作。是的，最终可能你确实需要进入实验室实际做些什么，但侦察将以与传统方式非常不同的方式进行，我想。然后这将提高我们做科学的效率。

> And everybody here loves what they're doing. So whether we're to start again with different tools or not, I think most of us would say we'll do what we're doing. And obviously, the current tools are very different from the tools when we started. And I hope even as we're now starting again, but we're still upgrading our own use of the tools. And going back to where we started, in almost every field, what you can do now is going to be much quicker than it ever was. I mean, in any field you look at, I know this from my own area, you look back to what you did as a PhD student. And what took three years then? Yeah, I think just weak. Especially in this, in this agentic era. I mean, it's going to completely change the way we do science and, and, and LLMs, fired up by lots of agents behind them will be the scouts. So they'll do, they'll do a lot of the early scouting. And yes, there's probably in the end, you do have to get into a lab and actually do something, but, but the scouting will happen in a very different way, I think, than traditionally has happened. And then that'll improve the efficiency of the way we do science.

[14:28](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=868s) 那么，Fei-Fei，当你开始ImageNet时，它是开放的、广泛共享的，人们可以访问它。这是整个事情如何发展以及AI本身如何能够发展的非常重要的一部分。你认为今天这可能吗，如此开放并能够做出一些容易访问的东西？绝对可能。我认为今天不仅可能，而且是必要的。这就是为什么你从James那里听到，像HAI这样的地方的基本信念之一是开放科学和开源。我知道你为什么问这个问题，因为我们看到一些前沿实验室，AI模型本身与经济利益紧密相连。利益，而且有许多模型是专有的。

> So, Fei-Fei,, when you started ImageNet, I mean, it was open and widely shared, and people could access it. And that was a really important part of how the whole thing developed and how AI itself was able to develop. Do you think that's possible today, to be that open and to be able to make something that is readily accessible? Absolutely. I think it's not only possible today, it's imperative. This is why you heard from James that one of the foundational beliefs of a place like HAI is open science and open source. I know why you asked this question because we look at some of the frontier labs that the very AI models is tied to the economic interest. interest and there are many models that are proprietary.

[15:26](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=926s) 我认为这没问题。你知道，计算世界和科学世界既是专有技术，也是开源技术。所以我认为这绝对是必要的。应该鼓励，尤其是在学术界和公共部门，我们继续拥有开源、开放权重和开放科学。我认为世界上许多政府开始担心他们开放数据，然后数据被用在其他地方。所以我认为这对那个概念本身构成了真正的威胁，而这对知识创造和实际进步是如此根本。

> And I think that's fine. You know, the world of computing and the world of science is both proprietary technology, as well as open source technology. So I think it's absolutely necessary. It should be encouraged, especially in academia and public sector, that we continue to have open source and open weights and open science.. I think there is lots of governments around the world are getting concerned about they make data open and they get used elsewhere. So I think there is a real threat there to that very concept, which is so fundamental for knowledge creation and actually advances.

[16:07](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=967s) 是的，我不否认。我认为开放数据将继续是一个有争议的话题。这取决于什么样的数据、谁的数据、谁从中受益、知识产权边界在哪里、隐私边界在哪里，以及开放这些数据可能带来的潜在意外后果。所以我认为，整个社会必须以制衡的方式运作，我们需要负责任地开放数据或共享数据。我们可能需要有数据的法律和法规。所以我认为这是一个重要的话题，但这并不削弱好的科学可以从数据共享中大大受益的重要性，尤其是在学术研究领域，我们需要拥抱这一点。

> Yeah, I don't deny. I think making data open is going to continue to be a contentious topic. It depends on what kind of data, whose data, who benefits from it, where are the IP boundaries, privacy boundaries, and potential unintentional consequences of opening this data. So I look, I think the whole society have to work in a check and balance way, that we need to be responsibly opening data or sharing data. We need to probably have laws and regulations of data. So I think it's an important topic, but that doesn't take away the importance of good science can benefit greatly from data sharing, and especially in the world of academic research, that we need to embrace that.

[17:20](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1040s) 我想，John，从某种意义上说，在你自己的职业生涯中，随着RISC方法的采用和广泛使用。我是说，这确实对整个行业产生了巨大影响，不是吗？我是说，这不仅仅是学术界的事。它更好了。确实如此，确实如此，确实如此。这需要时间，真正获得认可需要时间，因为在早期，摩尔定律以如此快的速度推进，而大多数计算仍然在大数据中心或桌面上，而不是在你手中的手机里。所以随着情况转变，人们随身携带的、由电池供电的计算机数量真正占据主导地位，那么这种方法在效率上的优势就压倒了一切。而这确实促成了该方法长期的成功。但当时它是反主流的。我们基本上被告知我们疯了。有一次，我参加一个小组讨论，幸运的是，伯克利的Dave Patterson在做和我们类似的工作，所以有两个学者抱团对抗整个行业。

> And I guess, John, in a way, in your own career, with the adoption and widespread use of the risk approach. I mean, that did make a big difference to the industry as a whole, wasn't it? I mean, it wasn't just academia. It was better. It did, it did, it did. It took time, it took time to really get traction because in the very early days, Moore's law was advancing at such a pace and most of the computing was still in big data centers or on desktops, not in your hand, in a phone. And so as things shifted and the number of computers that people were carrying around and were battery powered really became dominant, then the efficiency advantages of that approach overwhelmed everything else. And that really led to the long term success of the approach. But it was contrarian. We basically got told we were crazy. At one point, I was on a panel and fortunately, Dave Patterson at Berkeley was doing sort of parallel work to us, so there were two academics hanging together against the industry.

[18:29](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1109s) 有一次，我们在一个小组讨论上，辩论话题另一方的支持者或反对者，取决于你怎么看，主持人基本上说，嗯，Hennessy刚刚筹集了一百万美元基于RISC想法创办了这家公司。而你说它们不好。他应该怎么做？他看着我说，他应该拿着钱去南美洲。

> And at one point, we're on a panel and the protagonist or antagonist, depending on how you view it, on the other side of the debate topic, basically the moderator says, Well, Hennessy has just raised a million dollars to start this company based on the risk ideas. And you said they're no good. What should he do? And he looks at me, says, he should take the money and go to South America.

[18:59](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1139s) 有趣的是，我们没有那样做。他的项目最终被取消了。最终一切都解决了。但那种你必须把东西放出去，才能看到它将如何被使用，其他人将如何想到你没想到的想法，这是事物如何被使用、变得有用、变得令人惊讶的重要部分。绝对如此。智能手机出现了，诺基亚正在制造的非常早期的智能手机出现了，这确实改变了行业。

> Interestingly, we didn't do that. His project ended up being canceled. It all worked out in the end. But that notion that you've got to get something out there in order to see how it's going to be used, how other people are going to think of ideas that you wouldn't have thought of, is an important part of how things become used, become useful, become surprising. Absolutely. The smartphone came along, the very early smartphone that Nokia was building came along and that really transformed the industry.

[19:30](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1170s) 还有游戏。游戏是高性能计算的另一个早期大消费者，成本非常关键。所以我们在这里是以人为中心的AI，我喜欢这个名字。在实践中这意味着什么？你是如何思考的？在实践中，我认为我们做AI的一切，这是一条法则，都应该融入这样的价值：我们必须把人类福祉、改善人类条件的目标放在我们实践的中心，对吧？

> That and games. Games were the other big early consumer of high performance computing where cost was really critical. So we're here in human-centered AI and I love the name. What in practice does that mean? How are you thinking about that? In practice, I think everything we do with AI, which is a law, should be infused with the value that we got to put human benevolence, the goal of bettering human conditions in the center of what we practice, right?

[20:12](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1212s) 在学术界，我们的教育、我们的研究项目、我们的思想领导力、我们与社会不同部分的合作，在AI中，我们投入了很多思考，如何改善人类条件，如何做出更好的发现来帮助人类。我们如何与政策制定者合作，帮助他们理解技术，制定帮助社会的法律和政策？而且，我认为从更宏观的角度看，以人为本在工业界也很重要，因为从我们设计模型的方式到我们部署产品的方式，以及我们治理这些产品使用的方式，都应该深深以人为本，并反映我们社会共同关心的价值观。所以AI作为一种技术越强大，它就越应该以人为本，我们应该关心这一点。

> Here in academia, our education, our research projects, our thought leadership, our collaboration with different parts of the society, In AI, we put a lot of thoughts into how we can better human conditions, we can better discoveries that can help humans. How do we work with the policymakers to be helpful in their understanding of the technology to make laws and policies that help our society? And also I think zooming out, human centeredness is also important in industry, because from the way we design our models to the way we deploy our product and the way we govern the use of these products should be deeply human centered and have a reflected kind of values that our society collectively care about. So the more powerful AI is as a piece of technology, the more human centered we should, it should be, and we should care about this.

[21:30](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1290s) 再来看硅谷正在发生什么。现在关于AI安全、AI对齐等等的讨论非常激烈。其中很多，现在有极端的言论，但它们也是必要的。这些讨论的核心、目标都是必要的，因为我们要确保人们安全并从中受益。John，我内心有一部分在想，今天处于AI前沿的技术领袖们，是否在15到25年前接受计算机科学教育时，接受了更多以人为本的理念。

> And again, look at what's going on in Silicon Valley. There's such intense discussion right now about AI safety, AI alignment, and all that. And a lot of these, now there is extreme rhetoric, but they're also necessary. The core, the goals of these discussions are necessary because we want to make sure people are safe and benefit from this technology. And there is a part of me, John, that makes me wonder if today's technology leaders at the forefront of AI were educated 15 to 25 years ago with more human centeredness in their computer science education.

[22:24](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1344s) 这场对话会怎样？这是个很好的问题，Fei-Fei，因为正如你所说，这场辩论现在很激烈。我的意思是，过去两周里，Bill Gates持一种观点，Jensen Huang持另一种观点。在这两者之间，确实存在一系列真实的问题。还有一整套棘手的问题你必须应对。我们如何应对故意滥用？看看Hugging Face、OpenAI Hugging Face事件。那是无意的，它被入侵了。它还侵入了许多其他政府网站。所以这里一个无意的系统违反了美国网络法。如果某个恶意的人掌握了那项技术并故意用它来做这种技术，会发生什么？如果你看看这些系统发现的漏洞，它们发现的安全漏洞极其复杂。有一个不可思议的案例，Claude查看了一段20年前的代码，这段代码一直存在于每一个与互联网接口的Unix系统中。它在那里面发现了一个漏洞，复杂到没有人能想出这个后门。那需要你暴露，你必须利用系统中的两个不同缺陷才能突破。这些系统正在发现我们永远发现不了的东西，是的，借助智能体。所以我们确实需要建立护栏。我们需要弄清楚如何做到这一点，而Faith说得对，从以人为本开始。我们如何做对人们有益的事情，并尽量减少伤害？

> What would this conversation? That's a really good question Fei-Fei because I think as you said, this debate is raging now. I mean, in the past two weeks, we've had the Bill Gates take one join and Jensen Huang take another point. Somewhere between those two is a real set of issues. And there are a whole set of difficult problems you have to grapple with. What do we do about intentional misuse? I mean, look at the Hugging Face, OpenAI Hugging Face incident. That was unintentional and it broke in. And it broke into a number of other government websites as well. So here an unintentional system violates the US cyber laws. What's going to happen if somebody evil gets a hold of that technology and purposely uses it for this technique? And if you look at the bugs that these systems are finding, the holes, the security holes are finding, they are extremely complicated. And there's one incredible case that Claude found looking at a piece of code that's 20 years old, that's been sitting in every single Unix system interfacing to the internet. And it found a bug in that that is so complicated to no person would have figured out this back door That requires you to expose you have to get two different flaws in the system in order to break These systems are finding things that well, yeah with with agents that we never find So we do need to build guardrails We need to figure out how to do that and faith is right start with human centered How do we do something that's good for people and where we minimize the harm?

[24:03](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1443s) AI有大量不可思议的机会，但AI也有危险。你认为我们目前正在输掉这场公共辩论吗？嗯，公众已经失去了很多对AI将是好事的信念，无论是因为数据中心还是其他担忧。他们对公众意识和担忧的程度肯定大大提高了。而美国实际上比欧洲高得多。我的意思是，美国比欧洲更担心，这让我感到惊讶。通常，我们认为欧洲更安全优先。

> There's lots of incredible opportunity with AI, but there's also danger with AI. And do you think we're losing the public debate on that at the moment? Well, the public has lost a lot of belief that AI is going to be a good thing, whether it's because of data centers or because of other concerns. They certainly the level of public awareness and concern has gone way up. And the US is actually very higher than Europe. I mean, that surprised me that the US is more worried than Europe. Normally, we think of Europe as being more safety first.

[24:39](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1479s) 不要让技术外泄，这样你知道它是安全的。但现在，美国几乎出现了一场技术反弹。我确实认为我们在公众方面做得不好。当我说我们时，我认为尤其是技术界。你知道，技术界很多人现在反思自己说，我们营销做得好。我其实不知道营销是不是正确的词，因为那感觉太交易化了，好像我们想更好地推销它以换取什么。对我来说，教育是更好的词，还有沟通。我普遍认为，技术界对这项技术的力量如此自我沉醉，并没有帮公众的忙。当他们出去传播这项技术时，它从一个极端走向另一个极端，即它能治愈癌症、解决一切问题，全是乌托邦式的词语，比如无限生产力等等，都抛给了公众。而我们有很多公众仍然靠薪水过活，不在技术世界里，也真的不知道他们如何能从中受益。

> Don't let the technology out so you know it's safe. But right now, the US, it's been a real almost a technology backlash. I do think we're not doing a good job with the public. And when I say we, I think it's we in the technology world, especially. You know, a lot of people within the technology world are now looking at ourselves and say, we did a market well. I actually don't know marketing is the right word, because that feels a little bit too much transactional, as if we want to sell it better for something in return. I think really, for me, education is a better word and communication. I generally believe that the technology world has not done the public a favor by being so self-absorbed about the power of this technology, when they went out to communicate this technology, it went from one side of the extreme, which is that it'll cure cancer and solve every single thing and it's total utopian words like infinite productivity and all that was thrown to the public. when we have so much of the public that is still living check by check and not in the world of technology and not really getting any sense how they might benefit from that.

[26:15](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1575s) 那个极端又摆向了另一边，即天哪，末日、人类灭绝等等。那也剥夺了公众的能动性和安全感。并在公众中制造了一场非常不幸的风暴，真正的中立地带，即这项技术双刃剑的混乱真相，以及这项技术不可思议的机会，都缺失了。当然，再加上政治的大喇叭和大科技公司的大喇叭，只是没有帮助。这很不幸。

> That extreme got swung to the other side, which is, oh my God, doomsday, human extinction, and all that. That also took away the sense of agency and safety from the public. And have created a very unfortunate storm in the public, where really the middle ground, which is the messy truth of the double-edged sword of the technology, as well as the incredible opportunity of this technology, just all got missing. And of course you add to that the megaphones of the politics and megaphones of big tech, it just did not help. and it's unfortunate.

[27:05](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1625s) 媒体也扮演了角色，他们给了这些极端言论很多大喇叭。这是HAI自2018年成立以来一直感受到的，作为学术界和公共部门，我们真的认为我们有公共责任，从科学严谨的角度，从更加以人为本的角度传播这些知识。我仍然认为HAI可以做得更好。现在不是因为缺乏尝试，而是其他噪音太大声了，我们应该更加有策略，更有策略、更有针对性地把这种公众教育传播出去。是的，我认为这绝对正确。我的意思是，我们目前正在输掉这场辩论，我认为这很重要。我喜欢这种以人为本的方法。从我自己的医学时代起，我一直思考的一件事是，AI的测试之一是它是否为人与人之间的互动提供更多时间？这在治疗过程中如此重要。所以这很好。

> The media also played a role here that they gave a lot of megaphones to these extreme rhetoric. This is something that HAI since our founding, which is back in 2018, we feel as academia and public sector, we really take it upon ourselves that we share a public responsibility to disseminate this knowledge from a scientific rigorous point of view, from a much more human centered point of view. I still think HAI can do a better job. Now that not because of lack of trying, it's just that the other noises are so loud that it behooves us to even be more strategic more strategic and pointed in getting this public education out there. Yeah, I think that's absolutely right. I mean, we are losing the debate at the moment and I think it's important. And I love the fact this is a human centered approach. And one of the things that I've always thought about from my own era of medicine is that one of the tests for AI is does it give more time for human-human interaction? is so important as part of the healing process. So that's great.

[28:26](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1706s) John，你在学术界、产业界和创业等领域担任过许多不同角色。请谈谈你对工程师在其中所处位置的看法。我的意思是，你有没有例子说明工程与学术界、工程与产业界突然创造出了非常不同的东西？嗯，工程师创造我们生活的世界。他们构建了我们拥有的很多东西。我认为这就是工程真正庆祝的。那就是把科学拿来，经常开发它，通常还有额外的一步，让科学真正实现一个能改变世界的真正产品，但开发那门科学，然后把它推进下一步。大多数在大学里，我们构建原型。我们不构建最终产品。事实上，我们把它们分拆出去构建最终产品。但我们构建那些展示概念的原型，我们培训下一代年轻人，这同样重要。大学的一大优点是它们自然地传播技术。因为每年都有一批学生毕业离开，走向世界，他们被灌输新思想、他们一直在做的研究，他们往往是那些把想法向前推进并真正传播到世界的人。

> John, you've had a number of different roles in academia and industry and entrepreneurialism and so on. Give some feel for your view of where engineers fit into all of that. I mean, have you got examples of where engineering and academia, engineering and industry have suddenly made something very different? Well, engineers create the world we live in. they build so much of what we have. And I think that's what engineering really celebrates. It's that taking the science, developing it often, often there's an extra step to go for that science to really achieve a real product that'll change the world, but develop that science and then take it the next step. Mostly in the university, we build prototypes. We don't build the final product. In fact, we spin them out to build the final product. But we build those prototypes that demonstrate the concept, and we train that next generation of young people, which is equally important. I mean, one of the great things about universities is they naturally spread their technology. Because every year, a bunch of the students graduate and leave and go out into the world, and they're infused with the new ideas, the research they've been working on, and they're the ones that often carry the ideas forward and really propagate them into the world.

[29:51](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1791s) 我知道培训下一代技能是你在AI人性化方面思考的一部分。你想谈谈你认为需要发生什么来让那一代人准备好做需要做的事情吗？这是个好问题，Fei-Fei。我绝对认为，和所有教员一起经常谈论这个。看，我要直说。教育需要转型。AI已经改变了游戏规则。

> And I know that training the next generation of skills is part of what you think about here in the human sense of AI. Do you want to say something about what you think needs to happen to get that generation ready for what needs to be done? It's a good question, Fei-Fei. I definitely think, talk a lot about that along with all of our faculty. Look, I'm going to be blunt. Education needs a transformation. This AI has changed the game.

[30:22](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1822s) 当我说教育时，是从K-12、K-16一直到终身学习。我认为每所大学、每所学校都需要认真思考教育应该是什么样，我们如何重新思考课堂教育、学生评估、招生、为工作做准备，一切。这真的确实是一个文明时刻。我们的教育框架已经很多很多年没有改变了，可能可以追溯到近一百年前。而现在是不同的时代。互联网的融合、AI的融合、智能体的融合已经完全改变了游戏规则。

> And when I say education, it's from K-12, K-16, all the way to lifelong learning. I think that every university, every school needs to be seriously thinking about how education should be, how we can rethink about classroom education, students evaluation, admission, preparation for their jobs, everything. It really truly this is a civilizational moment. Our educational framework haven't changed for many, many decades, probably dating back almost a hundred years. And it's just a different time The convergence of the internet, the convergence of AI, of agents has changed the game completely.

[31:18](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1878s) 对我来说，TLDR就是赋能。学习必须是一个赋能的过程。而这种赋能需要伴随最强大的工具，可以推动从小人到大人，推动他们有能动性地学习，拥有学习的过程，拥有学习的结果，并且知道他们将面对一个根本不同的就业市场、面对一个根本不同的社会。所以对我来说，这是一个彻底转型的时刻。每一个学科工具。绝对。我的意思是，想想工程和科学。学生真正学到东西的支柱是什么？习题集，那才是他们真正学到东西的地方。那是他们获得认知参与的地方，迫使他们思考的摩擦被打破了。它们被打破了，因为我们的编程作业以正常形式，至少它们被打破了。所以现在我们必须弄清楚如何以非常不同的方式思考。其中一些可能是回到更早的时代。我的意思是，我的一些计算机科学同事正在回到使用辅导模式来教学。

> TLDR for me is empowerment. Learning has to be an empowered process. And that empowerment needs to come with the most powerful tools that can push from little humans to big humans to push them to learn with agency and to own the process of learning, own the results of learning, and also know that they're gonna face a job market, face a society that will be fundamentally different. So to me, this is a moment of total transformation. of every single subject tool. Absolutely. I mean, think about engineering and science. What are the backbones of where students really learn something? Problem sets, that's what they really learn. That's where they get the cognitive engagement, the friction that forces them to think blown up. They're blown up as our programming assignments in the normal form, at least they're blown up. So now we've got to figure out how to think very differently. And some of this will be maybe going back to an earlier time. I mean a number of my colleagues in CS are going back to using the tutorial model for teaching.

[32:31](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=1951s) 我是说这里经典的牛津剑桥方法，对吧？而且，你站到黑板前，被要求解决黑板上的一个问题。相信我，如果你没有准备，讲师会知道的。所以我认为我们将不得不重新思考我们做事的方式，并关注更多的个人参与。我是说你不能依赖，你看到一个申请，有人申请它。好吧，这里面什么是真实的？它有多真实？他们真正做了什么？我们将不得不更加面对面。所以，AI可能会迫使我们做更多面对面的事情，这将是件好事。而且它将改变研究，就像它改变教育方面一样。是的，绝对。我们为什么不开放几个问题，然后我有几个最后的事情，我想特别提一下AI。

> I mean here classic Oxford Cambridge method, right? And that, you get up at the Blackboard and be asked to solve a problem in the Blackboard. Believe me, if you haven't prepared, the instructor will know it. So I think we're going to have to rethink how we do things and watch more personal engagement. I mean you can't rely on, you see an application, somebody applies to it. Okay, what's real in this? How real is it? What's what have they really done? We're going to have to be more in person. So, AI could force us to do more in person things, which would be a good thing. And it's going to change research every bit as much as it changes the education side of things. Yeah, absolutely. Why don't we open this up for a few questions, then I've got a couple of final things I I want to put AI particularly.

### [33:20](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2000s) 问答环节
*Q&A*

[33:24](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2004s) 我们有公开的论文，关于模型是如何构建的。我们有人们事后发布他们创建的模型的开放权重。但训练模型的过程极其昂贵，这构成了对开放性的巨大障碍。我认为我们看到Google和Adobe、OpenAI和阿里巴巴，也许还有几家其他公司，几乎垄断了大型语言模型、前沿模型。当训练它们如此昂贵时，我们如何为了推动这类前沿模型的发展而使其开放？

> We have papers that are open about how models are structured. We have people post-hoc releasing open weights to models that they've created. But the process of training models is incredibly expensive and that produces an enormous barrier to the openness. And I think we're seeing Google and Adobe OpenAI and Alibaba and maybe a couple of others have almost a monopoly on large language models, frontier models. How can we make it open for the advancement of that sort of frontier of models when it's so expensive for them to be trained?

[34:17](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2057s) 嗯，首先，让我们思考一下我们的前提。开放的目标不一定是与一家营利性公司竞争，后者完全有权做他们的专有产品。我不认为那是开放的目标。我认为开放的目标是多重的。它是双重的。它是创业生态系统。这就是为什么我们看到很多公司实际上在做开放模型，从NVIDIA到，你知道，就像你说的，阿里巴巴和初创公司，因为他们想成为那个生态系统的一部分。开放的目标也是知识传播。这就是为什么学术界也在做很多开放。开放的目标还应该是严格的评估。我们应该有开放模型作为封闭模型的平行轨道，以便我们可以评估风险等等。

> Well, first of all, let's think about our premise. The goal of openness is not necessarily compete with a for-profit company who has every right to do their proprietary product. I don't think that's the goal of openness. I think the goal of openness are multiple It's a double fold. It's a juvenile, it's the entrepreneurial ecosystem. This is why we see that a lot of companies are actually doing open models from NVIDIA to, you know, like you said, Alibaba and startups, because they want to be part of that ecosystem. The goal of openness is also knowledge dissemination. This is why academia is also doing a lot of openness. The goal of openness should also be rigorous evaluation. We should have open models that serves as a parallel track to the closed model so that we can assess risk and all that.

[35:23](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2123s) 我还认为，实际上，这是个人信念，我不代表任何人，我认为我们需要考虑国家层面的开放模型，以国家实验室或国家实验室级别资源的形式，这样开源模型就不会太落后于闭源模型，以便我们能够很好地服务于我上面提到的这些目标。

> I also think that what I actually, this is a personal belief, I'm not representing anybody, is that I think we need to think about national level open models in the form of national labs or national lab level resourcing, so that the open source model doesn't fall too much behind the closed source model so that we can serve these goals I mentioned above well.

[35:56](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2156s) 而且这应该是可能的，随着这些东西在某种意义上，一旦它们不是绝对的前沿，它们就变得更加商品化。应该有可能做到你刚才描述的。我不知道，你怎么想？嗯，这是个有趣的问题，Fei-Fei。我是说，不仅有模型的规模，还有你训练的数据。但我要说的是，开源模型已经缩小了与领先前沿模型的大部分差距。也许它们在某些其他领域没有编码，但它们已经缩小了很多差距。而且对于很多应用，我认为你将能够 readily 获得成本十分之一的模型。而且我认为对于很多用途，那些模型会工作得很好。

> And that should be possible as these things become in a sense that once they're not the absolute frontier, they become more commoditized. that should be possible to do what you just described. I don't know, what do you think? Well, it's an interesting question, Fei-Fei. I mean, there's not only the scale of the model, but there's the data you're training on as well. But what I'd say is that the open source models have closed a lot of the gap with the leading frontier edge models. And maybe they're not encoding in a few other areas, but they've closed a lot of the gap. and for lots of applications, I think you'll be able to get models which are one-tenth the cost readily available. And I think for lots of purposes, those will work quite well.

[36:45](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2205s) John，然后抱歉，然后我们就要结束了。是的，我只是想回应Niko非常好的观点，并支持你刚才说的。我认为完全有可能训练一个中等规模的模型，7500亿参数，花费比你想象的少得多，只要你有一个非常非常高保真度的数据。你显然拥有所有数据管道和整理技能。韩国人已经做到了，花费比你想象的少得多，许多国家正在考虑。是的。而且他们目前训练的大量数据是垃圾。所以没有理由，没有理由你需要训练一千个猫电影，相信我。

> John, and then sorry, then we're coming down. Yeah, I was just going to respond to Niko's very good point and just back up what you just said. I think it's perfectly possible to train a medium-sized model, 750 billion parameters for a lot less money than you think, as long as you've got a very, very high fidelity date. You obviously have all the data pipelining and curating skills. The Koreans have done it and it's a lot less money than you think and many countries are thinking about it. Yeah. And lots of data that they currently train on is junk. So there's no reason, there's no reason you need to train on a thousand cat movies, believe me.

[37:27](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2247s) 谢谢你接受我的问题。你谈到了Jensen在等式的一边，Bill Gates在另一边。但如果你看到最近的新闻稿和白宫的会议，你让他们所有人都在辩论的两边站在领导层后面，吹捧超级智能。人们在谈论的风险是什么？看起来他们真的都一致。一方试图推动他们的优势，另一方试图阻止。你的问题是什么？风险是什么？风险是什么？当然有。有一些恐惧而不是风险。是的，有真正的风险。硅谷有句老话，人们过去常说，快速行动，打破东西。事实上，有一家公司这就是他们的座右铭。

> Thank you for taking my question. You talked about Jensen being on one side of the equation and Bill Gates being on the other. But if you saw the recent press releases and meetings in the White House, you had all of them on both sides of the debate standing behind the leadership and touting the super intelligence. What are the risks that people are talking about? And it seems like they are really all aligned. One side is trying to push their advantage, and the other is trying to prevent that. What's your question? What are the risks? What are the risks? There are certainly there are. There's some fear rather than risk. Yeah, there are real risks. There was an old saying in the Valley, people used to say, move fast and break things. In fact, there was one company that was their motto.

[38:30](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2310s) 危险在于，这是足够强大的技术，打破东西会造成真正的伤害，而不仅仅是，哦，我弄错了，这段的排版不对，或者Word崩溃了。那是一回事。另一回事是，哦，系统失控了，闯入了什么东西，偷了某人的数据，或者做了其他事情，或者认为有人必须制造生物武器。

> The danger is, this is powerful enough technology that breaking things creates real harm, not just, oh, I got the wrong, the typesetting didn't come out right in this paragraph, or or word crashed on me. That's one thing. It's another thing when, oh, the system went wild and broke into something and stole somebody's data or did something else, or thought somebody had to build a biological weapon.

[38:58](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2338s) 所以这些是真正的问题，我们将不得不弄清楚如何对它们设置护栏。开源世界的一个挑战是，你如何对开源的东西设置护栏？因为如果它是开源的，有人可以相对容易地剥离护栏。尽管我们已经看到人们经常甚至可以从封闭系统中剥离护栏，所以，或者至少损坏护栏。所以我们，有很多工作要做。大量的研究，深入的研究，关于安全、正确性和对齐的问题，对齐，我喜欢这个词对齐。

> So those are real issues and we're going to have to figure out how to put guardrails on them. One of the challenges in the open source world is, how do you put guardrails on something that's open source? Because if it's open source, somebody can strip off the guardrails relatively easily. Although we've seen that people can strip guardrails often off of the closed systems even, so, or at least damage the guardrails. So we, there's a lot of work to be done. A tremendous amount of research, deep research on this issue of safety and correctness and alignment, alignment, I love that term alignment.

[39:35](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2375s) 我认为看，这是一个非常强大的技术。它是一个双刃剑的故事。你会看到很多风险。Hugging Face的例子，正如John所说，我同意这是一个网络安全漏洞。现在我们可以评估网络安全漏洞的严重性，好消息是Hugging Face事件远没有那么致命，但这是一个网络安全风险。还有社会风险。我们有一位斯坦福政治科学家几周前在日本选举中做了一个实验，他们使用提示来询问AI，结果发现AI一直建议，即使是说中左倾向的非常良性的提示，也一直建议用户投票给一个非常边缘的政党。这是一个非常奇怪的结果，鉴于提示是中左的，结果发现法国政党获得更多AI曝光的原因是因为法国政党的网站是开放的，而主要日本媒体的网站实际上在我后面。

> I think look, this is a very powerful technology. It is a double edge story. You will see a lot of risks. The hugging face example, as John put it, I'd agree it's a cybersecurity breach. Now we can rate the seriousness of cybersecurity breach and the good news is the hugging face event is not that deadly by any stretch, but it's a cybersecurity risk. There's also social risk. We have a Stanford political science scientist who did an experiment with the Japanese election just a few weeks ago where they used prompting to interrogate AI, and it turned out that the AI kept suggesting the, even the very benign prompt of say center left orientation kept suggesting the user to vote for a very fringe party. And that was a very peculiar result, given that the pronged was center left, it turned out the reason that the French party got more of the AI surface exposure is because the website of the French party is open, whereas the websites of the main Japanese media are actually behind me.

[40:58](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2458s) 所以数据偏见压入了社会工程风险。

> So that data bias press into a social engineering risk.

[41:31](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2491s) 挑衅，他们基本上没有展示说，你知道，拼凑是不行的。我们必须从根本上转变很多本科教育。当然，他刚才说完成了。所以我对我们两个的问题是，斯坦福计划如何改造其本科教育？如果你能告诉我你计划做的两三件具体事情，以及你认为你面临的挑战是什么。

> provoking where they basically not show said, you know, patchwork is not gonna work. We have to fundamentally transform a lot of the undergraduate education. And that's of course, he just said done. So my question to the two of us, how is Stanford planning to revamp its undergraduate education? If you could tell me two or three specific things that you're planning to do and what are the challenges that you think you are facing.

[42:03](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2523s) 前校长。我认为这是一个中间的游戏。我们现在正在推进很多这些事情。我认为辅导模型是我们将这样做的一种方式。我认为我们将重新思考我们做事的方式。我所思考的是，而不是我站在教室前面讲课，我将把学生推上前，说，去网上阅读材料，使用AI学习它，然后你向你的同学解释这个概念如何运作。所以现在我们在真正翻转课堂，因为我们把学生放在前面。当然我会在那里帮助他们，辅导他们学习材料。但我认为像这样的事情，将迫使人们，他们会利用AI，但他们会利用AI来真正掌握主题。这让我想起很多很多年前学到的东西。真正学习一个主题的最好方式是必须在这里教它。所以我们将让学生以不同的方式真正参与思考。我认为我们将不得不重新思考我们如何进行招生过程，因为那里很多事情会改变，我们将不得不改造我们做事的方式，使他们适应这个AI世界。

> former president. I think this is a game that's in the middle. We're in motion right now on a lot of these things. I think the tutorial model is one way in which we're going to do that. I think we're going to rethink the way we do it. What I've thought about is rather than me getting up in the front of the classroom and lecture, I'm going to put the students up and say, to go read the material online, use AI to study it, and then you explain to your fellow students how this concept works. So now we're really flipping the classroom because we're putting the student up there in front. And of course I'll be there to help them and coach them through the material. But I think things like that, which will force people to, they'll take advantage of AI, but they'll take advantage of AI to really master the topic. And it reminds me of something I learned many, many years ago. The best way to really learn a topic is to have to teach it here. So we'll get the students really engaged in a different way thinking about it. I think we're going to have to rethink how we do admissions processes as well, because lots of things will change there and we're going to have to revamp the way we do things to make them adopt this AI world.

[43:17](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2597s) 你发现了吗？坦白说，我不知道。我只想说，事情变化得很快。这栋楼是盖茨楼，是计算机科学系一半的所在地。从本学年开始，我们有了一门新的106课程，叫106T，这是我们的一位讲师新开的课，教学生在软件工程中使用生成式AI。所以，我们正开始朝这个方向迈进。接着刚才那个问题说，虽然我们在升级这些课程等等，但一个应届毕业生——这些事情需要时间——一个应届毕业生，他们大部分东西是在工作中学会的。当你进入现实世界，很多东西是在工作中学会的。而在入门层面，现在大部分工作都由AI完成了，对吧？因为生产力提高了，大家也都把注意力放在这上面。

> You found that? Admission, I don't know. I just want to say things are changing fast. This building is Gates Building, which is half of the home to computer science. As of this academic year, we now have a new 106 class called 106T, which is a new class built by one of our lecturers to teach students to use a genetic AI in software engineering. So, we are starting to move towards all of this. So just continuing on that question, while we are upgrading these courses and everything, a fresh graduate, these things will take time, a fresh graduate, and most of them learns a lot on the job. When you go on the real world and learns a lot on the job. And at the initial level, most of the job is now being done by AI. right? So because of the productivity increase and everyone focuses on that as well.

[44:18](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2658s) 那么应届毕业生该怎么办？因为他们在工作中做的很多学习都由AI完成了，他们该如何适应？嗯，这是个简短的问题。我其实认为应届毕业生或准应届毕业生这一代人的焦虑是真实存在的。我前不久在飞机上真的碰到了一位斯坦福大一新生。和大学生或应届毕业生聊天非常有启发。我认为有几点。首先，变革的时代也是失去的时代。它同时也是机遇的时代。我认为这里面有很多个人责任。虽然我们的社会正在认识到这一变化，但我们——尤其是斯坦福的应届毕业生——必须把主动权握在自己手里。你知道，你过去四年学到的东西正好处在变革的前夜。所以你学到的一些东西可能不再适用，要主动去学习新的工具组合。你可以——没人会喂给你，对吧？尤其是你刚毕业的时候。这是一个变革的时代，你必须认识到这一点。

> So what a fresh graduate should do? Because a lot of learning that they do on the job is done by AI and how they should adopt to it. Yeah, this is a quick question. I actually think the anxiety of fresh graduates or near fresh graduates generation is really real. I actually literally ran into a Stanford freshman and in a flight. And it was very illuminating talking to college students or fresh college grads. I think several things. First of all, a time of change is a time of loss. It's also a time of opportunity. I think there is a lot of individual responsibility in this. While our society is recognizing this change, we as especially a Stanford fresh grad, you have to take that agency in your own hands. You know, what you have learned in the past four years are right at the cusp of it before change. So some of what you have learned might not apply and be proactive in learning the new suite of tools. You can, nobody's gonna feed you that, right? Especially if you just graduated. That is a time of change that you just have to recognize.

[45:35](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2735s) 另一件事是，我也听说有些学生来读计算机专业，以为自己的显然出路是谷歌，也许谷歌现在招人没那么多了，但他们在医疗行业、能源行业、制造业发现了惊人的机会，因为对AI和软件工程的需求——由于生成式AI——正在扩散。那是一种可能没被规划过、但确实存在的新职业机会。所以我认为这是应届毕业生真正要认识到的另一点。几乎每个行业都是如此。每个行业都是。对，我一直跟我们的学生说，你知道，这不会是可预测的。工作会变，新工作会被创造出来。有些工作可能——至少需求会消失，因为AI会提高生产力。但有一件事我可以肯定地告诉你：那些懂得把AI当作工具用在工作里的人，会比不懂的人做得更好。所以人们必须学习这项技术，而我们看到这正在发生。我是说，你看看有多少电子工程专业的学生在修机器学习课程，在电气工程领域做着令人兴奋的事情，把自己领域的知识和把AI当作工具来使用结合起来。

> The other thing is I'm also hearing that maybe some of the students come in to get a CS major thinking their obvious job is Google, or maybe Google is not hiring as much, but they found incredible opportunity in the healthcare industry, in energy industry, in manufacturing industry, because the need of AI and software engineering is It's diffusing because of a genetic AI. And that is a new career opportunity that might not have been planned, but it's there. So I think that's another thing for Fresh Class to truly recognize. In pretty much every industry. In every industry. Yeah, I've been telling our students, you know, it's not gonna be predictable. Jobs are gonna change, new jobs are gonna be created. Some jobs will probably, at least the demand will go away because AI will improve productivity. But the one thing I'll tell you for sure is that those that know how to use AI as a tool in their job will do better than those that don't. So people have to learn this technology and we see it happening. I mean, you see how many E.E. students are taking machine learning courses and really doing exciting things in electrical engineering that combine their knowledge from that field with the use of AI as a tool.

[46:56](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2816s) 我想这是我们最后一个问题了。谢谢，真的非常有意思。我有一个笼统的问题，是从大学的角度来的。我想知道，大学是否在人们没能充分接触AI这件事上起了作用。我想到的是风险投资的形成，对吧？风险投资长期是开源的。很多走出来的工程师都用这些开源模型来学习。至少作为工程师，我学到的很多东西都基于大学的研究和开源模型，然后这些模型被用来教学生，不只是创建它的那所大学的学生，而是全世界的学生。然后这些模型被学生带入产业界采用，再用它们去构建其他东西。但就AI而言，我觉得大学可能被落下了，也许是因为缺乏资源或基础设施，以至于没有哪所大学拥有一个足够好的开源模型。我们在要求政府或国家实验室来构建它，但以斯坦福或其他顶尖大学的实力，这本来早就该建成了，对吧？大学是不是失职了？

> I think it's our last question here. Thanks, really fascinating. I had a general question, which is coming from the university. I wonder whether universities have played a role in people not having access to AI as much as they should. So I'm thinking back to risk formation, right? Risk is open source for a long time. A lot of engineers who came through have used these open source models to learn. And pretty much a lot of the things, at least as an engineer that I learned, were based on university research with open source models that then were used to teach students, not just from the university where it was created, but across the world. And then those models were then adopted by the students as they went into industry, and then they use that to build other things. But as with AI, I feel like universities are being caught short maybe because of lack of resources or infrastructure that there is no university that has a model that's good enough, that is open source. And we are asking governments or national labs to build it, but surely with the firepower at the university like Stanford or any other top university, this should have been built before, right? Did the universities drop the ball?

[48:07](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2887s) 我不撒谎。今天训练一个大语言模型所需的资源量是大学力所不及的。我是说，除非我们把其他一切都不管了，包括我们的体育项目。所以我也认为，大学在LLM之外有着巨大的机会。

> I'm not going to lie. The amount of resource that takes to train a large language model today is out of reach for universities. I mean, unless we forget about everything else, including our athletics. So I also think that there's tremendous opportunity in universities that's outside of the LLMs.

[48:39](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2919s) 我们拥有任何产业——甚至强大的谷歌——都没有的非凡跨学科优势。所以我们有数据，独特的数据，我们有独特的专有技术，我们有独特的问题要解决。所以我更兴奋得多地期待大学去展示、去开发这些新模型、新科学、新发现，这些是大学核心优势所独有的，而不是说我们去和谷歌的LLM竞争。很好的回答。抱歉，有一个小问题溜进来了，我说它小是因为指示如此。我会说快一点。我想既然我们这里有两位英国医学教授，我们应该问一个工程或工程学的问题。我今天上午也在——同时也是兼职——在Biohub，E-hub的校长，今天有一个公告，能源部、Biohub、谷歌，我们要联合起来，投入20亿美元，着眼于用AI来建模细胞。

> We have incredible interdisciplinary strengths that no industry, not even mighty Google has. So we have data, unique data, we have unique know-how, we have unique problems to be solved. So I'm much, much more excited for universities to show, to develop these new models, new science, new discoveries that are unique to the core strength of universities rather than say we compete with Google's LONs. Great answer. I'm sorry, one tiny question has slipped and I say it's tiny because that's the instruction. I'll talk quickly. I thought since we had two British professors of medicine here, we should ask a question for engineering or engineering. I was at this morning at also part-time, the Biohub, the Chancellor of the E-hub, and there's an announcement today, the Department of Energy, Biohub, Google, we're getting together for a $2 billion look towards using AI to model cells.

[49:48](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=2988s) 我觉得那真的很令人兴奋。我也——Fei-Fei，接着你的话说，今天上午他们宣布我们即将治愈癌症，而且再也不会有药物失败，我当时就在那个小组讨论上。所以我其实有机会多问了一点：在我们设计出蛋白质之后，从他们自己的蛋白质到真正的药物之间可能会发生什么。但我的问题更具体一些。如果我们构建这些开源模型——我认为我们应该这么做——而这类投资能让它实现，那么我们如何在开源模型中实施针对生物恐怖主义的防护措施？因为那是一个真正的挑战。几个月前Demis来这里时，他强调这是他最大的担忧。我要说，这大概也是我最大的担忧。

> I think that's really exciting. I was also, Fei-Fei, to your point, this morning they declared that we were about to cure cancer and that no drug would ever fail, which I was sharing that panel. So I actually had a chance to ask a little bit more about what might happen after we designed proteins between their own and the actual drug. But my question is a bit more specific one. If we build these open source models, which I think we should, and this kind of investment would allow it to happen, how do we implement the guardrails against bioterrorism in open source models? Because that's a genuine challenge. When Demis was here just a few months ago, he underlined that as his biggest worry. And I would say it's probably my biggest worry as well.

[50:31](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=3031s) 这是一个研究问题。我认为这是一个研究问题。我认为我们不知道怎么做，说实话。我想你可以考虑在模型外面加某种包装层，阻止它——但那些包装层，我担心如果是开源模型，你可以把包装层剥掉，从而绕过它。所以某些事情对坏人来说会变得容易得多，我们必须弄清楚该怎么做，无论是改革公民社会，让追究这类人更容易、让责任认定更容易，还是别的。但眼下，我认为我们没有好的解决方案。不幸的是，这是软件。软件是可塑的。这和我们当年面对核武器的问题不同，也和例如重组DNA的问题不同——那个问题我们是通过约束资助研究的科研机构来解决的。那在这里行不通，因为大部分模型是由产业界提供的。所以这对HAI来说是个好问题，你不觉得吗？是的，所以就在昨天我还在胡佛研究所，由Condoleezza Rice召集，把计算机科学家、生物工程教授、经济学家，以及我们国家的国防和情报界领袖聚在一起，讨论这类问题。所以我其实认为问题不在于开源模型本身，甚至不在于技术的速度。

> It's a research problem. I think it is a research problem. I think we don't know how to do that, to be perfectly honest. I suppose you could think about a wrapper around the model somehow that prevented it from, but those wrappers I worry that if they're open source models you can strip the wrappers off and and circumvent it. So certain things are going to get a lot easier for bad people to do and we've got to figure out how we're going to do that whether it's Reforming civil society to make it easier to go after people like that and to assess liability But right now, I don't think we have a good solution. And unfortunately, this is software. Software is malleable. This is a different problem than we had either with nuclear weapons or the problem, for example, with recombinant DNA, which we solved by constraining the research organizations that were funding the research. That's not going to work here because industry is providing most of these models. So this is a good problem for AHA to work on, don't you think? Yes, so literally yesterday I was at Hoover Institute too, convened by Condoleezza Rice and putting together computer scientists, bioengineering professors, economists, as well as defense intelligence community leaders from our country talking about problems like this. So I actually think the problem is not about open source model per se, or even the pace of technology.

[52:07](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=3127s) 问题在于监管和决策群体与科技、生物科技群体之间的跨界协作。我认为在硅谷这里，我们没有和东海岸对话的传统，但过去几年很好地教会了我们这种对话有多重要。所以有很多事情在进行。我相信你和Condi及她的同事们谈过所有这些。当然，不只是AI，还有你能做CRISPR这个事实。是的。

> The problem is the cross-collaboration between regulatory as well as policy making groups and community with the tech and biotech community. I think here in Silicon Valley, we didn't have a tradition to talk to the other coast, But the past few years have taught us really well how important it is to have this conversation. So there's a lot going on. I'm sure you've talked to Condi and her colleagues on all of this. And of course, it's not just the AI, but it's the fact that you can do CRISPR. Yes.

[52:52](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=3172s) 实验室的其余部分对一个有本科生物学水平的人来说也是可及的。是的。你现在不需要非常高级，当然这也要经过它。看，非常感谢你们两位，因为真的非常非常有意思。我要留给你们一个刁钻的问题，那就是：你们认为下一个伊丽莎白女王工程奖会是什么？是我希望是什么，还是我说会是什么？

> The rest of the lab is also accessible to somebody with undergraduate biology. Yeah. You don't need to be very advanced now, which of course goes through it. Look, thank you so much to both of you because it's been really, really interesting. I'm going to leave you with one nasty question, which is, what do you think is going to be an upcoming QE2 prize for engineering? What do I hope or what do I say?

[53:24](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=3204s) 我认为如果有人能想出如何构建具有成本效益的大规模储能、大规模电力储存，无论是电池还是其他真正具有成本效益的东西，那就是关键所在。如果你能做到这一点，我们就能比原本更快地加速可再生能源的未来。所以我希望有年轻人在研究这个问题。这是一个非常具体的回答。

> I think if somebody figures out how to build cost-effective bulk energy storage, bulk electricity storage, whether it's batteries or something else that's really cost-effective, that is the linchpin. If you could do that, we could accelerate the renewable future much faster than we otherwise would. So I'm hoping some young people are working on that problem. That's a very specific answer.

[53:52](https://www.youtube.com/watch?v=mOvHKBOQ6mU&t=3232s) 很多可能性都会让我非常兴奋，但我想我的偏爱是医疗健康。我认为这是人类繁荣的一个极其关键的领域，拥有海量数据、海量人类需求，而AI能带来巨大改变。对，谢谢。这里一直传递出乐观情绪，这正是我很喜欢的一点。你读到的、听到的一些东西可能会让你消沉。事实是，如果我们能在社会层面把它做对，我们正处在真正重大事情即将发生的临界点上。所以非常感谢你们两位，Nubia。谢谢。

> I would be very excited by many possibilities, but I guess my soft spot is healthcare. I think this is an extremely critical area of human flourishing that has a ton of data, a ton of human needs, and AI can make a huge difference. Right, thank you. It's been optimism coming through here, which is what I really like about it. And you can get ground down by some of the things you read and hear. The fact is we're on the cusp of really tremendous things happening if we can get it right in society. So thank you both very much, Nubia. Thank you.

---

*来源：[研几 · AI](https://ai-daily-cn.pages.dev/watch/hai-feifei-li/) · 原节目 [YouTube](https://www.youtube.com/watch?v=mOvHKBOQ6mU)。文字稿：faster-whisper small.en（本机 CPU），时间轴即 YouTube 原片时间（Stanford HAI）；未区分说话人；英文原声：DeepSeek 逐段翻译（57 段），编辑抽查；核心总结由 DeepSeek 依据译文生成*