世界模型 vs. 大语言模型：医疗 AI 的下一个阶段

我们会在智能眼镜或其他设备里有一个小 AI 助手，随时陪伴我们。它会看到我们所看到的，听到我们所听到的，随时给我们建议，让我们更聪明。但我们是它的老板，对吧？我是说，我们决定做什么，它解决我们交给它的问题。所以，它不会取代我们，而是会放大我们的能力。欢迎收听《How I Doctor》，我们正在把快乐带回医学。这个播客不断推出特别节目，我想这意味着我们在成长，并触及了正确的人群。是的，今天又是一个特别节目。你们现在看到的是我本人，与两位希望在医疗领域留下印记的人工智能顶尖头脑进行视频直播。今天和我一起的是 Alex Le Brun 和 Yan Lun。Alex 是一位连续创业者，也是 Naba Building AIC Scribe 技术的联合创始人兼 CEO，最近被任命为 Advanced Machine Intelligence Labs 的 CEO。这家由 Yan 创立的新公司计划构建大型语言模型的继任者，即所谓的世界模型，我们将探讨它作为人工智能的下一个阶段。过去几年我有幸结识 Alex，我认为他是少数几个学会如何在医疗领域工作并部署的技术专家之一。他利用技术背景快速构建，同时能够迅速减速以适应医疗领域可接受的变化速度。我的第二位客人，如果你关注 AI 或技术，无需介绍，在我看来他绝对是一位巨星——Yan Lun 教授。Lun 教授是 ACM 图灵奖得主，那是计算机科学的奥斯卡奖，他是纽约大学 Jacob T. Schwarz 计算机科学教授，前 Meta 首席 AI 科学家，现在是 Advanced Machine Intelligence Labs 的执行主席兼创始人。我多年来一直关注 Yan 的工作，所以今天能和他坐下来交谈非常激动。如果你还没看过，他的主页 yan.lun.com 会满足你对早期万维网的怀旧，因为它直接来自 1999 年的 Geocities 风格。今天，我希望他们能教我们关于世界模型的概念，而不是大型语言模型，以及如何将这一概念应用于医疗领域，他们希望在那里产生重大影响。Alex Yan，欢迎来到《How I Doctor》。什么是世界模型？它可能与人们现在习惯的 AI（比如大型语言模型）有何不同？

We're gonna have, you know, a little AI assistant in our smart glasses or whatever. Yeah. Right. With us at all time. You will see what we see, hear what we hear, and advise us at all times and make us smarter. But we'll be their boss, right? I mean, we decide what to do. They solve the problem we give them. So, they're not going to replace us. They're going to amplify us. Welcome to How I Doctor, where we're bringing joy back to medicine. We keep on having very special episodes of this podcast. And I think that just means we're growing and reaching the right people. And yes, it's another special one today. You've got me live in person on video with two of the brightest minds in artificial intelligence who want to make their mark on healthcare. Today I'm joined by Alex Le Brun and Yan Lun. Alex is a serial entrepreneur and co-founder and CEO of Naba Building AIC Scribe Technology and was just announced as CEO of Advanced Machine Intelligence Labs. This new company founded by Yan plans to build a successor to large language models, something called World Models that we'll get into as the next phase of artificial intelligence. I've had the pleasure to get to know Alex over the past few years, and I think he's one of the few technologists who has learned how to work and deploy within healthcare. He uses his tech background to build with speed while still being able to decelerate rapidly to the pace of change we are comfortable with in healthcare. My second guest needs no introduction if you follow AI or technology and is an absolute rockstar in my book, Yan Lun. Professor Lun is the ACM touring award laurate. That's the Oscars of computer science, the Jacob T. Schwarz Professor of Computer Science at NYU, the former chief AI scientist at Meta, and now the executive chairman founding Advanced Machine Intelligence Labs. I've followed Yan's work for years, so it's so exciting to get to sit down with him today. And if you haven't checked it out, his homepage at yan.lun.com will scratch an itch for an earlier version of the worldwide web because it's straight out of 1999 geocities. Today, I'm hoping they can teach us about the concept of world models instead of large language models for artificial intelligence and how we can apply the concept in healthcare where they hope to have a major impact. Alex Yan, welcome to how I doctor. What is a world model and how is it maybe different from what people are now used to as AI? You know, like a large language model.

那么，想想当你面对一个新情况时，你是如何在世界中行动的。你想完成一个不太熟悉的任务。你用自己的大脑去想象你的行动会产生什么结果或后果。你可以找出一系列行动来达到某个特定目标。是的，无论你是在建造东西还是完成任何任务。这就是世界模型的概念。世界模型是一个模型或计算机程序之类的东西，给定世界在时间 t 的状态和你想象要采取的行动，它会预测由此产生的世界在时间 t+1 的状态。这与当前那些智能体系统（如果你愿意这么叫的话）的做法非常不同。当前智能体系统的训练方式就像大型语言模型一样。它们只是逐个预测 token。token 就是行动，对吧？它们必须盲目地产生一个行动，然后祈祷这个行动是正确的。所以让这些东西工作的唯一方法就是为特定任务收集大量数据，然后让它们模仿人类在这些情况下所做的任何行动序列，而这很脆弱。它只在某些情况下有效，并且需要收集大量数据，比如整个互联网和世界上所有已发表的内容。

So, think about how you act in the world when you're facing a new situation. You want to do a task that you're not completely familiar with. You use your mind to essentially imagine what the result or the consequences of your actions are going to be. You can figure out a sequence of actions to arrive at a particular goal. Yeah. Whether you're building something or just any task that you're accomplishing really. That's the idea of world model. The world model is a model or computer program or something like that that, given some idea of the state of the world at time t and an action that you imagine taking, will predict the state of the world at time t plus one resulting from this action. So that's very different from what the current crop of agentic systems, if you will, will do. The current way agentic systems are trained, they're like LLMs. They just predict one token after the other. Tokens being actions, right? They have to produce an action blindly and then just kind of cross your fingers that this action is correct. And so the only way to make those things work is to collect lots and lots and lots of data for a particular task and then have them just imitate whatever action sequences humans do in those situations, and that's brittle. It only works in certain cases and it requires collecting a lot of data for the entire internet and everything that's ever been published in the world.

是的，没错。仅凭预测下一个词和下一个 token 的极其复杂的数学就能取得如此大的进展，真是令人惊叹。但这肯定不是全部。你会遇到局限性，我认为我们这些摆弄大型语言模型并试图理解它们如何融入医疗领域的人，看到了它们开始失效的地方，似乎就是你提到的规划阶段。

Yeah, that's right. It's really amazing how far you can get with just like extremely complex math that is predicting the next word and the next token. But it certainly is not everything. You run into the limitations and I think those of us who are like playing around with large language models and trying to understand how they might fit into healthcare, we see where they start to break down and it does seem to be that planning phase like you're talking about.

是的，绝对如此。我的意思是，这些东西的工作原理确实令人印象深刻，对吧？完全正确，而且当它最初出现时有点令人惊讶，但它们非常有用，有很多人在研究它们，也应该有很多人研究它们。大量资金被用于基础设施来运行它们。这没问题。但它们有一些局限性，你必须真正意识到这些局限性，尤其是在医疗保健或任何可靠性至关重要的情境中，这些局限性尤为突出。

Yeah, absolutely. I mean, it's really impressive how those things work, right? Totally and it was, you know, a little bit of a surprise when it first emerged, but they're very useful and, you know, there's a lot of people working on them and a lot of people who should be working on them. A lot of money being spent on infrastructure, you know, to run them. That's fine. But they have some limitations and you have to really be conscious of them and they're particularly acute in healthcare or any kind of situation where reliability is really paramount.

是的。Alex，你也是一位计算机科学家。你认为世界模型的影响是什么？为什么你们如此确信这将是人工智能和自动化技术的下一个大事件？

Yeah. Alex, you're a computer scientist as well. What do you think the implications are of world models? Why are you guys so convinced that this will be kind of the next big thing in artificial intelligence and automation technology?

我认为正如 Yann 所描述的，大型语言模型旨在处理 token 化数据。也就是任何带有 token 的东西，即低维度的离散数据。我们知道这不是世界运作的方式。你知道，婴儿，你经常用这个例子 Yann，但只有 8 个月大的婴儿就已经有了非常好的世界模型。所以，如果你把一个球抛向空中，抱歉，我在借用你的例子，但如果你抛出一个球，然后球 somehow 停在了空中，五六个月以下的婴儿不会感到惊讶。令人惊奇的是，人类大脑仅仅通过一段时间的观察就开始建立这些神经元模式并理解这些事情。它建立了一个世界模型，然后利用这个模型进行预测和规划。我们知道，动物在没有语言概念的情况下也能做到这一切。我们知道大型语言模型，就像专注于语言的捷径，在许多用例中非常有用，但它只是大脑工作方式的一个方面，只是大脑的一小部分。我们觉得，为了用 AI 解锁更多新的用例，我们需要寻找别的东西，一些不同于大型语言模型的东西。

I think as Yann described, LLMs are designed to process tokenized data. So anything with token, so discrete data with low dimension. And we know this is not how the world is working. And you know, babies, you often use this example Yann, but babies as young as 8 months old already have a very good model of the world. So, if you launch a ball in the air, sorry, I'm stealing all your examples, but you launch a ball and then the ball somehow gets stuck in the air, before five or six months old babies are not surprised. It is amazing that the human brain just with observation over time starts to build these neuron patterns and understands these things. It builds a model of the world and then uses this model to make predictions and then planning. And we know that all that can be done by animals without any notion of language or so. We know that LLM, like a shortcut focused on language, is very useful in many use cases but it's just one side of how the brain works, it's just one part, small part of the brain. And we feel that to unlock new further use cases with AI, we need to be looking for something else, something different from LLM.

Yan，你曾说过语言是容易的部分，物理世界要难得多。我想给你们一个我认为可能适用于医疗领域的翻译。我们有所谓“生病或不生病”的概念，对吧？而“生病”往往意味着危重病。

Yan, you've said that language is kind of the easy thing. The physical world is much harder. I'll give you guys what I think is maybe a translation into healthcare. We have this thing called sick or not sick, right? And sick translates to, you know, often critically ill.

我面前这个人情况非常危险。有趣的是，我们常形容这种感觉来自直觉。我们说这是整体感知。Z 世代可能会说气场不对，对吧？但有一种感觉，经过多年训练，我能从房间另一头看到一个进入急诊室的人，然后说：'哇，这里有问题。情况非常危险。我需要立即采取行动。'这就是你们试图达到的目标吗？理解物理世界，因为语言有局限性，我们很难用语言表达出来。

Something is dangerously very wrong with this human in front of me. And the interesting thing is we often describe that as coming from our gut. We say it's the gestalt. Gen Z might call it the vibes are off, right? But there is a sense that I can, you know, through years of training, I can look at a human being from across the room that has entered my ER and say, 'Whoa, there's something bad here. Something is very dangerously wrong. I need to take action immediately.' Is that kind of what you guys are trying to get at is understanding the physical world because language has limitations that we struggle to put into words.

完全正确。如果仅凭语言就能推理、学习世界如何运作，那你光靠读书就能学会医学了。

Absolutely. Yeah. If it were the case that you could reason, learn about how the world works purely from language, you could learn medicine by just reading books.

是的。

Yeah.

但你做不到，对吧？你得实习。你得把手指放在病人肚子上上千次，才能找出几个异常的。

And you can't, right? I mean, you have to do residency. You have to put your hand on the belly of a patient a thousand times to find a few that are abnormal.

是的，完全正确。

Yes. Absolutely.

还要判断他们是不是得了阑尾炎之类的，他们会抗拒等等。所以世界远比文字描述丰富，没有什么能替代亲身经验。

And to figure out if they have appendicitis or something, and they resist and all that stuff. Right. So there's a lot more to the world than textual description of it, and nothing replaces firsthand experience.

所以这和我经常引用的一组数字有关：当你收集互联网上所有公开文本时，大约有 30 万亿个词。

So that and this sort of a pair of numbers that I cite very often, which is when you collect all the text that is available on the internet publicly, it's about 30 trillion words.

上下浮动一万亿。

Give or take a trillion.

也许是 20 万亿个词。

20 trillion words maybe.

这相当于 30 万亿个 token，大概这个数量级，对吧？一个 token 大约三个字节，因为数量是有限的。

Which translates into 30 trillion tokens, something of that order, right? A token is about three bytes, because there's only a finite number of them.

所以大约是 10 的 14 次方字节，对吧？一个 1 后面 14 个零。

So that's about 10 to the 14 bytes, right? A one with 14 zeros.

我们任何人要读完这海量数据和知识（事实知识）都需要大约 50 万年。但这大约是一个四岁孩子四年生活中通过视觉看到的数据量。心理学家告诉我们，这大约是 16000 小时的清醒时间。

It would take any of us like half a million years to read through this enormous amount of data and knowledge, factual knowledge. But that's about the amount of data that a four-year-old has seen through vision in four years of life. Psychologists tell us it's about 16,000 hours of wake hours.

当然。

Sure.

我们有两根视神经，每根有 100 万根纤维，每根每秒传输大约一个字节。所以每秒大约有两兆字节进入视觉皮层。算一下，四年大约是 10 的 14 次方字节。所以现实世界中传感器输入的信息比文本多得多。

And we have two optic nerves with 1 million fibers each, carrying about one byte per second each. So it's about two megabytes per second going to the visual cortex every second. Do the math and it's about 10 to the 14 bytes in four years. So there's just way more information in the real world in sensor inputs than there is in text.

我的意思是，这些世界模型就是通过这种方式训练和构建的吗？通过视频？通过各种不一定基于文本的数据？

I mean, is that the idea of how these world models get trained, built? Is it through video? Is it all sorts of data that's not necessarily text based?

嗯，有两种世界模型。如果你在一个离散的世界里，比如计算机代码或某种玩具场景，你可以描述世界状态、描述动作，然后预测下一个世界状态。但对于现实世界，这就变得非常复杂，因为如果你试图把视频转换成离散的 token（这是非常愚蠢的做法，但你可能尝试），然后用预测来预测视频中会发生什么，这根本行不通，因为现实世界是高维、连续、嘈杂、混乱、复杂的。

Well, so you have two types of world models. If you are in a discrete world, like computer code or something, or some kind of toy situation, you can have a description of the state of the world, description of an action, and predict the next state of the world. For the real world, though, it becomes really complicated because if you try to take a video, then turn it into discrete tokens, which is a very stupid thing to do, but you might try, and then try to use prediction to predict what's going to happen in a video, it simply doesn't work very well because the real world is high dimensional, continuous, noisy, messy, complicated.

非常混乱，非常复杂。

Very messy, very complicated.

是的。所以那些在语言上效果很好的生成方法，在视频或任何传感器输入上根本行不通。可能是医学检测或其他什么。

Yeah. And so the generative approaches that work so well for language really don't work for video or for any kind of sensor input. It could be medical tests or what.

你认为你们需要多长时间才能拿出一些可以给我这样的人看的东西？

How long do you think you guys will take until you have something that's ready to show somebody like me?

是的。重要的是要指出，这仍处于研究阶段。要得到一个现成的通用世界模型，训练得尽可能接近人类，需要很多年，你知道，三、四、五年，甚至更久。这就是拥有这样一个现成模型的目标。但在构建这个模型的过程中，我们将开始与合作伙伴、工业伙伴合作，针对特定用例，将世界模型愿景应用于更狭窄的问题和用例。我们相信，在这些特定用例上，我们可以在四年之前，也就是一两年内取得成果。

Yeah. So it's important to note that it's still at the research phase. In order to get an off-the-shelf universal world model that would be trained as close as possible as a human will take many years, you know, three, four, five, maybe more years. And this is the goal to have such an off-the-shelf model. But on the way, building this model, we will start to work on specific use cases with partners, industrial partners, in order to apply the world model vision to more narrow problems and use cases. And we think we can get results on these specific use cases before you know, earlier than four years, so a year or two.

是的。你知道，我一直把大语言模型想象成，如果你把医院比作餐厅，它们就像前台。它们可能与患者互动，帮助询问患者问题，总结信息，帮助医生思考。但医学还有另一部分，当我进入大脑中某个区域时，我在想：这个人到底怎么了？我在脑子里过一遍。我真的会盯着虚空，思考面前患者的问题、我的计划等等。我不知道大语言模型在多大程度上是解决方案，也许世界模型才是，也许还有其他东西。

Yeah. You know, I've thought of the large language models as kind of, if you think of a restaurant, like the front of the house. There may be the interaction with the patient. Maybe they're helping ask the patient questions. They're summarizing information. They're helping the doctor to think. And then there's a whole other part of medicine that happens when I go to a place in my brain where I'm thinking, okay, what could be going on with this person? And I'm running through it in my head. I literally go, I stare off into space and I think about the problem of my patient in front of me and my plan and everything like that. And I don't know how much a large language model is the solution versus maybe it's a world model, maybe it's something else as well.

你是在运行你的患者模型，对吧？

You're running your patient model, right?

完全正确，我就是那样做的，是的。

Absolutely, that's where I go, yes.

所以如果你有一个系统，拥有这样一种患者的心理模型，你可以想象用这个系统来预测某个疗程的效果。然后预测结果，特别是对于像糖尿病这样的慢性病。也许通过优化，找出一个最优的干预序列，使患者达到特定状态。

And so if you can have a system that has such a kind of mental model of a patient, you could imagine using the system to predict what is going to be the effect of a course of treatment. Right. And then predict what the outcome of this is, particularly for things like chronic disease like diabetes or something like this. And perhaps through optimization, figure out an optimal sequence of interventions that will make the patient go to a particular state.

是的。Epic 有他们正在尝试的 Cosmos 项目。我想他们称之为大型医疗模型、大型患者模型。他们的想法，我认为，是查看护理事件，利用数亿次就诊数据，然后说：好吧，我们可以说，如果你今天被诊断出糖尿病，你可能有七种未来可能，其中之一是 12 年后中风之类的。听起来世界模型真的会试图预测患者接下来会发生什么，而不是他们三年后可能患上高血压之类的。

Yeah. Epic has this Cosmos thing that they're trying. And I think they called it like a large medical model, a large patient model. And their idea, I think, was looking at episodes of care and taking hundreds of millions of encounters and saying, okay, we can say that if you're diagnosed with diabetes today, there's maybe seven future possible futures for you, and one of them is that you have a stroke in 12 years or something like that. It sounds like the world models would really be trying to predict what would happen next to the patient, not they will likely develop hypertension three years later or something.

我认为 Cosmos 也非常有用和强大，但 Cosmos 的输入主要是文本，如果我没记错的话。

I think also Cosmos is very useful and powerful, but the input of Cosmos is text mostly, if I'm not wrong.

病历只是患者状态的一个非常有限的投影。如果你的输入数据如此有限，你就不可能猜出你没有正确数据的东西。

The note is a very limited projection of the state of the patient. If your input data is that limited, there is no chance you can guess things you don't have the right data for.

是啊。我越是阅读关于 AI 的文章并思考它在医疗中的应用，就越是对自己的大脑感到惊叹。就像你说的，我直接在我的神经元中访问那个患者模型，甚至不需要任何正式的工作。感觉你们正在试图弄清楚如何在计算机集群中做到这一点。

Yeah. The more I read about AI and think of its application to healthcare, the more I'm impressed with my own brain. My brain, like on your point, I just go to that patient model in my neurons without any official work. It feels like you guys are trying to figure out how to do that inside a cluster of computers.

让我打个有趣的比方，特别是如果有神经学家在听的话。我们的世界模型位于前额叶皮层，也就是大脑的前部，它让人成为人。当然，许多动物也有世界模型，只是没那么复杂。

Let me give a funny analogy, particularly if there are neurologists listening. Our world model is in our prefrontal cortex, the front part of our brain that makes humans human. Of course, many animals have world models too, just not as sophisticated.

我在想象我的猫，它那小小的世界模型。

I'm imagining my cat with its tiny little world model.

你的猫对自己的物理和动力学有极好的模型。

Your cat has an extremely good model of its own physics and dynamics.

它们太不可思议了。嗯，它们有九条命。

They're incredible. Well, they have nine lives.

所以那就是你的模型，你的前额叶皮层。你脑子里也有一个 LLM。有一小块脑区叫韦尼克区，让你理解语言；还有一块叫布罗卡区，把思想转化为语言。那就是你的 LLM。LLM 的问题在于它能生成文本，但没有思考的部分。那正是我们试图构建的。

So that's your model, your prefrontal cortex. You also have an LLM in your head. There's a little piece of the brain called the Wernicke's area that allows you to understand language, and another called Broca's area that turns thoughts into language. That's your LLM. The problem with the LLM is that it generates text but doesn't have the part that thinks. That's what we're trying to build.

我喜欢这个观点，即规划特别重要。作为一名急诊医生，当我在考虑治疗患者的三个选项时，我会立即把每个算法推演到最终点：我是要收治他们还是让他们出院，给这种药还是那种药？

I like that idea that planning is particularly important. As an ER doctor, when I'm thinking about three options to treat a patient, I immediately go through each algorithm to the final point: am I going to admit them or discharge them, give this medicine or that one?

这就是思维链的技巧。我们人类推理的方式是搜索：我们想象如果我这样做，就会发生这个；如果我那样做，就会发生那个。你提前做所有这些预测，并用你的心智模型搜索合适的答案。你不是在写下符号或大声思考。很多想法是内部的，使用你对患者的内部表征，无法用语言表达。

That's the chain of thought trick. The way we do reasoning as humans is that we search: we imagine if I do this, then this will happen, or if I do that, then that other thing will happen. You do all those predictions in advance and search for an appropriate answer using your mental model. You're not just writing down symbols or thinking out loud. A lot of those thoughts are internal and use your internal representation of the patient, not expressible in language.

绝对是这样。我们有一个概念，有时患者看起来比他们的指标更糟。当你看着他们时，就知道出了大问题。也有相反的情况：指标看起来很糟糕，但患者却坐在那里无所事事。我们没有好的词语来形容这个，但每个人看到时都知道。

Absolutely. We have this concept where sometimes the patient looks worse than their numbers. When you look at them, something's very wrong. And there's the opposite: people whose numbers look terrible but the patient is sitting there twiddling their thumbs. We don't have good words for this, but everybody knows it when they see it.

Alec，我想问问你关于 Nabla 以及在医疗领域工作的挑战。在科技界，数据无处不在，可以自由传递，而医疗领域有隐私规则。你在将工具交到一线医护人员手中方面有什么发现？

Alec, let me ask you about Nabla and the challenges of working in healthcare. In tech, you have data everywhere and can pass it around freely, versus in healthcare with privacy rules. What have you found about getting tools in the hands of frontline healthcare workers?

科技界人士和科技公司以在医疗领域失败而闻名。我们带着强烈的假设而来。就我而言，我 100%的假设都是完全错误的。拯救我们的是，我们知道我们对医疗一无所知，所以我们在 Nabla 成立之初就开办了自己的诊所，成为我们 AI 的第一批用户。我们有自己的医生和患者，每天观察。我们每天和每位医生待 10 个小时，就这样我们发现了所有我认为不真实的事情。例如，我们以为需要构建一个非常复杂的自动诊断 AI。但运营诊所三周后，我看到医生唯一喜欢做的事情就是诊断。所以我本来可以去找他们说，‘嘿，我造了一个 AI，在你工作中最喜欢做的事情上比你更强。’非常容易部署。

Tech people and tech companies are famous for failing in healthcare. We come with strong assumptions. In my case, 100% of my assumptions were completely false. What saved us is that we knew we didn't know anything about healthcare, so we started our own clinic at the beginning of Nabla to be the first users of our AI. We had our own physicians and patients, and we observed every day. We spent 10 hours a day with every physician, and that's how we discovered everything I thought was not true. For instance, we thought we needed to build a very sophisticated automatic diagnostics AI. But after three weeks of running the clinic, I saw that the only thing physicians love to do is diagnostics. So I would have come to them and said, 'Hey, I built an AI that is better than you at the one thing you like to do in your job.' Very easy to deploy.

是啊。

Yeah.

我们早期学到的另一件事是，当我们训练 AI 模型时，达到 70-80%的准确率很容易，但每增加一个百分点都更难。在科学领域的机器学习中，80%的准确率就很棒了。但在医疗领域，医生被训练成如果某样东西错了三次就会忽略它。如果屏幕上三次不相关，他们大脑中的世界模型就会教他们忽略它，就像电子病历中那些你甚至看不到的警告。

Another thing we learned early on is that when we train an AI model, it's easy to get to 70-80% accuracy, but each additional percent is harder. In machine learning for science, 80% accuracy is great. But in healthcare, physicians are trained to ignore something if it's wrong three times. If it's not relevant three times on the screen, their brain's world model teaches them to ignore it, like all the warnings in EHRs that you don't even see.

我知道它们存在。

I know they exist.

所以这是另一个例子，说明 80%的准确率即使对于测试或试点也不够，因为如果不够相关，它就会被忽略，而且你没有第二次机会。你就失去了医生和用户。

So this is another example where 80% is not enough even for a test or a pilot, because if it's not relevant enough it will be ignored, and you don't have a second chance. You just lost the physicians, the user.

这说明了电子病历和临床决策支持的挑战，就是敏感性和特异性的问题。我数不清有多少次急诊室的警报响了，病人盯着看，惊慌失措没人来救，但我们几乎听不到，因为知道是假警报。我们宁愿完全关掉它。这就是构建真正有效工具的挑战。

It speaks to the challenges of electronic medical records and clinical decision support, just sensitivity and specificity. I cannot tell you how many times an alarm goes off in the ER, and the patient is staring up, freaking out that no one is coming, but we don't even hear it because we know it's a false alarm. We'd rather turn it off entirely. That's the challenge of building tools that actually function well.

是的。我们学到的另一件事是产品必须非常非常容易使用。

Yeah. Something else we learned is the product has to be very, very easy to use.

它应该是可用的。

It should be usable.

可用性很关键。你只有几分钟时间来展示它能给医生带来的价值。在没有看到回报之前，你没有时间和精力投入更多。

Usability is key. You have just a few minutes to demonstrate the value it will bring to the physicians. You don't have time or energy to invest more without first seeing the reward.

OpenAI 和 Anthropic 刚刚发布了他们的医疗工具版本。你认为在一两年或五年内，LLM 会产生重大影响，还是我们会看到它们的价值，然后转向其他模型？

OpenAI and Anthropic just released their versions of healthcare tools. Do you think in one, two, five years LLMs will have a big impact, or will we see them for what they're worth and move to other models?

我认为有很多不同的用途。一个是让人们获得快速的非正式建议，他们可能信任也可能不信任。这比联系医生容易得多。

I think there are many different uses. One is for people to get quick informal advice, which they may or may not trust. It's a lot easier than getting in touch with your doctor.

我无法不同意你的观点，Yann。

I can't disagree with you, Yann.

事实上，我有一个朋友，前同事 Mustafa，来自塞内加尔。他回国创办了一家名为 Kerah Health 的初创公司。他微调了一个 LLM，用所有当地语言提供医疗建议。相对简单，超级有用。他发现即使能看医生的人也会和这些 LLM 交谈。所以在发展中国家有巨大的应用，会帮助很多人，在发达国家也可以作为传播信息的方式。

In fact, I have a friend, a former colleague, Mustafa, from Senegal. He went back and created a startup called Kerah Health. He fine-tuned an LLM to provide medical advice in all local languages. Relatively straightforward, super useful. He finds that even people who can see a doctor talk to those LLMs. So in developing countries there are huge applications that will help people a lot, and also in developed countries as a way of disseminating information.

是的。美国的健康素养下降了。我经常教人们如何处理流鼻血，或者同时服用泰诺和莫特林是可以的。你认为是什么推动了最近对医疗技术引入的关注和兴趣？

Yeah. Health literacy in the US has gone down. I often show people how to manage a nosebleed or that it's okay to take Tylenol and Motrin together. What do you think is driving the recent attention and interest in bringing more technology to healthcare?

我认为最近为医生打造的 AI 工具终于解决了真正的痛点，而不是工程师喜欢的花哨问题。也许这是第一次。自动记录的想法已经存在了近 20 年。我们有真人记录员，通常是医学生，他们会走来走去记下我问的问题和病人说的话。现在我们有了 AI 替代品。最终，有了大语言模型，正确的技术出现了，我们能够实现这一点并解决问题。这是第一次，这项技术满足了医生的实际需求，而不是管理层或首席财务官的需求。

I think the recent AI tools for physicians finally address a real pain point, not a fancy problem that engineers like. Maybe it was the first time. The idea of scribing has been around for almost 20 years. We had human scribes, premed students who would walk around and write down what I asked and what the patient said. Now we have an AI alternative. Finally, with large language models, the right technology was there, and we could deliver on this and solve the problem. For the first time, this technology addressed the physicians' actual need, not the administration's or the CFO's.

话虽如此，AI 在医学中最大的用途是成像。用于理解和分析图像，以及加速 MRI 数据收集，现在 40 分钟就能做一次全身 MRI，只需 500 美元。这太惊人了。

That said, the biggest use of AI in medicine is imaging. For understanding and analyzing images, and for MRI to accelerate data collection, you can get a full body MRI in 40 minutes for $500 now. It's pretty amazing.

这实际上是我最初与纽约大学放射科的 Dan Sodickson 讨论的一个项目的结果。我们问，能不能用 AI 提高 MRI 质量并减少时间？纽约大学和 Meta 研究实验室 FAIR 的一群人开始合作。Meta 对医疗保健没有兴趣，但这个项目很酷，最终发表、开源，现在市场上每台 MRI 机器里都有一个版本。

It's actually the result of a project I first discussed with Dan Sodickson at NYU's radiology department. We said, could we use AI to improve MRI quality and reduce time? A bunch of people at NYU and at FAIR, Meta's research lab, started working together. Meta had no interest in healthcare, but the project was cool, ended up published, open-sourced, and now there's a version in every MRI machine on the market.

在最后 10 分钟，我们来谈谈愿景和未来。Alex，你说过你不会抢走我的工作。医疗领域有太多工作要做，我们会很高兴有些被替代。如果世界模型在医疗领域起飞，谁是赢家，谁是输家？

In our last 10 minutes, let's talk about vision and the future. Alex, you said you're not going to get rid of my job. There's so much work in healthcare that we'll be happy to have some replaced. If world models take off in healthcare, who are the winners and losers?

我认为正确的图景对医生来说可能有点难以想象，但就像大学教授拥有整个实验室，里面有比你更聪明的博士生和博士后。这就是你想要的。当你说蠢话时，他们会告诉你并纠正你。所以你有一群比你聪明的人为你工作。这就是我们未来与 AI 助手的关系。我们会在智能眼镜或其他设备里拥有小 AI 助手，随时陪伴我们。

I think the right picture is a little difficult for a physician to imagine, but it's like being a professor at a university with your entire lab of PhD students and postdocs who are smarter than you. That's what you want. When you say something stupid, they tell you and correct you. So you have a staff of people smarter than you doing the work. That will be our relationship with AI assistants in the future. We'll have little AI assistants in our smart glasses or whatever, with us all the time.

你会看到我们所看到的，听到我们所听到的，并随时给我们建议，让我们更聪明。但我们会是那个老板，对吧？我的意思是，我们决定做什么。他们解决我们给的问题。所以，他们不会取代我们，而是会放大我们。

You will see what we see, hear what we hear and advise us at all times and make us smarter. But we'll be that boss, right? I mean, we decide what to do. They solve the problem we give them. So, they're not going to replace us, they're going to amplify us.

太棒了。我的意思是，我很乐观，而且我认为我们开始看到，借助 AI 文档，医生能够回到床边，花更多时间作为解释者、翻译者、治疗者。如果我们能回到那种状态，那将是一个多么了不起的系统。你们认为医生角色的哪一部分可能最先消失或开始褪去？我们谈过一些文档工作。你们觉得会不会更专注于诊断、管理、预后？有没有哪个对世界模型来说更容易或更具战略性？

That's amazing. I mean, I'm hopeful that it and I think we're starting to see with AI documentation that physicians are able to go back to the bedside, spend more time as interpreters, translators, healers. I mean, what a be an amazing system if we could get back to that. What part of being a physician do you guys think maybe disappears first or starts to fade away first? We've talked about documentation a bit. Do you have a sense you guys might focus more on diagnostics, management, prognosis? Is there one that is easier for world models or more strategic?

有一个我认为容易，不简单但短期可行。

One that would be, I think, easy, not easy but shorter term.

是的，公平。一个短期可行的就是医疗编码。这看起来简单，但当你尝试去做时，你会发现即使现在的语言模型，编码基本上只是查看过去类似案例然后应用相同代码，但这不是正确的编码方式，你需要某种使用明确指南的推理。所以它是推理指南和过去经验的混合，而当前技术在这方面非常糟糕。我认为这解释了为什么没有人能做到 100%准确的编码，而且世界上没有人喜欢做编码。我认为这是你们赢得人心的另一种方式——说你们可以拿走编码和我必须填写的九页文件。是的，我会喜欢这一切。

Yeah, fair. One that would be shorter term is medical coding where this is something that looks simple as long as you and when you try to do it and then you realize even with LM today, coding just works basically looking at the past for similar cases and then you apply the same code, but it's not the right way to do coding and you want some kind of reasoning using explicit guidelines. So it's a mix of reasoning of guidelines but also experience in the past and this is what current technology is really bad at, and I think this explains why nobody has managed to do 100% accurate coding and nobody in the world likes to do coding. I think it's a if you find yeah if anyone Yeah. I mean that I think that's probably the other way you guys win hearts and minds, to say you could take away by coding and the nine-page document I have to fill out. Yeah, I would love all that.

是的。除此之外，如果你谨慎而不编码，你会错过收入；如果你过于激进而过度编码，则可能面临后果。所以除了编码的负担，你实际上还有一些风险。

Yeah. And on top of that, if you are cautious and you don't code, you miss revenues and if you're too aggressive and upcode, then you may have consequences. So on top of the burden to do coding, you actually have some risks coming.

是的。我的意思是，我认为医生感受到的根本紧张是，至少在美国，我们很多人受雇于大型医疗系统，这些系统想从每个代码中榨取每一分钱，所以他们会说这应该是五级而不是四级。所以我认为真正重要的是医生最终感受到的紧张和摩擦——我们最终负责，但我们有这么多不同的老板，他们都想让我们按他们的方式做事，而我们自己对世界应该如何运作也有自己的看法。

Yeah. I mean that is the fundamental tension that I think physicians feel is at least in the United States many of us are employed by a large health system that health system wants to get every dollar out of that code and so hey this should have been a level five instead of a level four and so that's the piece that I think is really important is just the tension and the friction that physicians feel at the end of the day where we're ultimately responsible but we've got so many different bosses who want us all to do things the way they want us to do things and we've got our own opinions about how the world should work too.

我认为最终这回到了我们最初讨论的离散与连续的问题。医学中有很多规则，我不知道它们是如何建立的，但比如回到糖尿病的例子，他们会治疗你，使你的糖化血红蛋白低于 7，如果只是略高于 7，就增加剂量，因为需要这样——其实 7 并不是一个神奇的阈值。

I think ultimately it goes back to this question we're discussing initially of discrete versus continuous right. So there's a lot of rules in medicine that are, I don't know how they were established, but like to get back to the example of diabetes, they'll treat you so that your A1C is below seven, and if it's just a little bit above seven, increase the dosage of whatever, because it needs to be like there's no magic threshold at seven.

对，是的。

Correct, yeah.

除了有研究表明低于 7 是好的，但这取决于个人，取决于动态，取决于很多事情。决策不需要那么干脆，如果你有一个真正能预测患者状态演变过程的底层模型，你就不需要那些干脆的决策。你可以有一个针对患者优化的治疗方案。

Other than there are studies that show it's nice if it's below, but that depends on the person. That depends on dynamics. That depends on a lot of things. Decision making doesn't need to be that crisp in the sense that if you have an underlying model of the patient that really can predict the course of evolution of the state of the patient, you don't need to have those crisp decisions. You can have a treatment that is optimized for the patient.

这是个很好的观点。我认为我们在医学中遇到的许多质量指标，就像你说的，糖化血红蛋白需要低于某个值，血压需要低于某个值。所以你可以看到这些，我的意思是，有多项研究显示阶梯函数，比如你到了 65 岁，更多人做结肠镜检查，那是因为他们有了医疗保险。所以我完全同意，这基本上是医生想要做的：我们想说，你因感染性休克入院，指南可能说我需要给你三升静脉输液，而我作为你的医生，试图不考虑指南，只专注于我认为对你最好的，也许我认为我只想给他 2.5 升或 4.5 升，取决于你的血流动力学和生理状况。如果模型能帮助做到这一点，并且不让我与会计们惹上麻烦，我认为医生们肯定会喜欢。

I mean it's a great point. I think many of the quality metrics that we come across in medicine, like to your point, the A1C needs to be less than this, the blood pressure needs to be less than this. So you can see these, I mean there are multiple studies showing step functions where it's like oh you turn 65 all these more people get colonoscopies, and it's like well they got Medicare they have insurance now so of course. And so I absolutely agree that is fundamentally what doctors are trying to do: we are trying to say you come in with septic shock, the guidelines might say I need to give you three liters of IV fluid, and I might say as your doctor who's trying to not think about the guidelines, just be laser focused on what I think is best for you, maybe I think I only want to give him 2.5 liters or 4.5 L depending on your hemodynamics, your physiology. And if the model can help do that and not get me in trouble with the bean counters, I think physicians would love that for sure.

最后几分钟我有几个快速问答。哪个专科应该最担心世界模型？

I've got a couple lightning round questions for you guys in the last few minutes here. What specialty should be most nervous about world models?

嗯，每一个，但不用太担心。

Well, every single one of them, but not too worried.

好的，告诉我为什么。

Okay, tell me why.

嗯，不太担心，因为再次强调，你会是它们的老板。

Well, not too worried because again, you'll be their boss.

是的。喜欢这个答案。

Yeah. Love it.

但它们会帮助你。

But they'll help you.

Alex，当今医疗 AI 中最被过度炒作的部分是什么？

Alex, what's the most overhyped part of AI in healthcare today?

几乎所有。为什么这么认为？为什么被过度炒作？因为我们还没能部署它，或者它只是理论上的，或者因为有些人非常擅长营销。而且医疗行业有很多钱。所以所有要素都具备了，导致 AI 的几乎每个方面都被过度炒作。

Almost all. Why do you think that is? Why is it overhyped? Because we haven't been able to deploy it or it's theoretical or because some people are really good at marketing. And there is a lot of money in healthcare. So all the ingredients are there to get overhype of almost every aspect of AI.

最后一个问题。Naba 建了一家诊所。你们认为如何将 Ami 从研究转移到医疗？会通过那种方式——开办诊所来学习吗？

Last question guys. Naba built a clinic. How do you guys think you'll move Ami from research to healthcare? Will it be through learning through starting a clinic in that way?

所以建诊所很有趣，原因我解释过，但也是一个巨大的错误，因为非常慢、非常贵。如果我能回到七年前，我会与像你这样的人合作，我们会在过程中节省大量时间和金钱。所以想法当然是跟医疗服务提供者、临床医生合作。

So building a clinic was interesting for the reason I explained but also a big mistake because it's very slow, very expensive. Now if I could go back seven years, I would partner with people like you and we would have saved a lot of time and money in this process. So the idea would be of course to partner with providers, with clinicians.

嗯，是的，这有一整条链条。我的意思是，Emy 的使命实际上是构建通用的世界模型，能够理解任何过程，并可能帮助控制任何你想要的东西。然后与合作伙伴合作，针对特定领域，将这一技术带给实践者，比如在医疗保健领域，但还有其他领域。

Um, yeah, there's a whole chain, right. I mean, there is the, you know, Emy's mission is really to kind of build generic world models that can understand any process and maybe help control anything that you want. And then working with partners for sort of particular areas themselves, working with partners to bring this to the practitioners in the case of healthcare, but, you know, other areas.

好吧，各位，我迫不及待想看看你们……请允许我在你们准备好时成为 alpha 或 beta 测试者。我太兴奋了。今天能和你交谈真是荣幸。谢谢。

Well, guys, I cannot wait to see what you... Well, please allow me to be an alpha beta tester whenever you're ready. I'm so excited. It's been a pleasure to get to talk with you today. Thank you.

我也很荣幸。谢谢，Ram。

Pleasure, too. Thanks, Ram.