AI 的指数级飞跃：为何未来十年可能重新定义人类寿命

这可能是人类历史上最关键的时期。所以尽量在未来 10 年内别死。

This is probably the most critical time in human history. So try not to die for the next 10 years.

你能解释一下为什么你这么认为吗？

Can you explain and unpack why you think that?

原因是，由于 AI，技术正在呈指数级扩张。癌症很可能在不到十年内实现 100% 可治愈。我们将达到每月可能有数百种新药问世的阶段。然后，我们将在最多 15 到 20 年内达到能够完全逆转衰老过程的阶段。所以如果你 80 岁、90 岁，你将回到 30 岁、40 岁，等等。

The reason is that the technology because of AI is expanding exponentially. Cancer is going to be 100% curable probably less than a decade. We'll get to a point where we'll have hundreds of new drugs coming out every month maybe. And then we'll get to a point probably 15 maximum 20 years where we will be able to completely reverse the aging process. So if you're 80 years old, 90 years old, you will get back to age 30, 40, whatever.

你为什么如此乐观？

Why do you have such an optimistic view?

AI 是一个不可思议的赋能者。它赋予你超能力。关键风险在于人类。人类滥用 AI。对人类来说，唯一的生存威胁就是人类自身。

AI is an incredible enabler. It gives you superpowers. The key risk is humans. Humans misusing AI. There's only one existential threat to humanity and that's humanity.

大家好，今天的节目探讨一个极其令人兴奋的交汇点：人工智能的加速发展以及对科学和医学未来的日益乐观。在本期节目中，我将与杜里亚·埃纳特马斯博士讨论 AI 如何大幅提升我们检测、预防和治疗人类疾病的能力，并最终延长人类预期寿命。在开始之前，我只想提一件小事。观看本播客的人中只有大约 30% 订阅了 YouTube 频道。花点时间订阅并开启通知是支持本节目、帮助我们把这些对话带给更广泛观众的最简单方式之一。我们非常感谢。非常感谢，我真心希望你们喜欢这期与杜里亚·埃纳特马斯博士的节目。

Hey everyone, today's episode explores an extraordinarily exciting convergence, the accelerating pace of artificial intelligence and a growing optimism about the future of science and medicine. In this episode, I discuss with Dr. Duria Enautmas how AI could dramatically improve our ability to detect, prevent, and treat human disease and ultimately extend human life expectancy. Before we begin, I just want to mention one quick thing. Only about 30% of the people who watch this podcast are subscribed to the YouTube channel. Taking a moment to subscribe and enable notifications is one of the simplest ways to support the show and help us bring these conversations to a wider audience. We greatly appreciate it. Thank you so much and I really hope you enjoy this episode with Dr. Duria Enutmas.

我非常激动能坐在这里，与杜里亚·乌纳特马斯博士一起。他是少数几位有机会与 OpenAI 合作的科学家之一，OpenAI 是全球人工智能领域的领导者之一。他还是一位衰老研究者，也是一位免疫学家，真是天作之合，能坐下来聊聊 AI 在衰老研究和医学中的作用。所以今天能请到你，我超级兴奋。

I'm so excited to be sitting here with Dr. Dura Unutmas who is one of the handful of scientists that has had access to collaborate with OpenAI, one of the world's leaders in artificial intelligence. He's also an aging researcher. He's an immunologist, really just a match made in heaven to sit down and talk about the role of AI in aging research and in medicine. So I'm super excited to have you here today.

我非常激动能来到这里。谢谢。

I'm very excited to be here. Thank you.

我们都知道，衰老是一个非常非常复杂的过程。涉及许多因素。它是异质性的。它如此复杂，几乎看起来不可能解决。然而，我听过你说过一些非常有趣的话。我听过你说，如果你能在未来 10 到 15 年内尽量不死，你可能想试试，因为你可以多活 50 年。

As we both know, aging is a very, very complex process. Many factors involved. It's heterogeneous. It's so complex. And it just seems like so almost impossible to solve. And yet, I've heard you say something that's very interesting. I've heard you say if you could try not to die within the next 10 to 15 years, you might want to try to do that because you could live an extra 50 years.

是的。

Yeah.

你能解释并展开说明为什么你这么认为吗？是什么让你相信这一点？

Can you explain and unpack why you think that? What makes you believe that?

谢谢。首先，我非常激动能来到这里。我是你播客的忠实粉丝。我认为它可能是最好的衰老或长寿播客。所以这是极大的荣幸。嗯，是的，我在过去一两年里确实说过好几次，而且可能不需要 10 到 15 年，可能更近。原因是，尤其是由于 AI，技术正在呈指数级扩张。我们的思维是线性的。所以我们认为未来 10 年的进步会和过去 10 年或 15 年一样多。但事实并非如此。未来 10 年，你可以把它想象成比上个世纪更先进。想象一下你生活在 1900 年代初，有人告诉你我们会有疫苗，你永远不会得天花，或者你不会死于肺结核。你知道，人们会嘲笑你。所以那是不可能的。这就是我们谈论的速度。但更重要的是，由于这种加速，治疗疾病的进展也将急剧加速。所以我们将达到所谓的“长寿逃逸速度”。这是奥布里·德格雷提出的，你知道，他是一位伟大的衰老研究者。所以关键是，我们将在未来，我大概说 8 到 10 年内，达到每活一年都会给你的寿命增加超过一年的程度。比如说，你知道，10 年后你得了癌症，通常无法治愈，你只有一两年可活。但在那一年里，会有一种新疗法治愈那种癌症。所以它自动会给你的寿命增加几年，或者可能 10 到 15 年。或者我们已经开始看到 GLP1 药物、受体激动剂，它们为肥胖或患有慢性病的人增加了大约 5 到 10 年的寿命。我们还将有肌肉生成器，我认为它们将对老龄化人口产生巨大影响，因为你知道，那是一个巨大的问题。所以所有这些都会累积起来，技术和 AI 将继续加速。所以 10 年后，一年内发生的事情将相当于 20 年的进步。然后我们将达到一个点，最多 15 到 20 年，我们将能够完全逆转衰老过程。所以如果你 80 岁、90 岁，你将回到 30 岁、40 岁，等等。所以这将为你的寿命增加 50 年或 100 年。然后你可以继续这样做，几乎无限期地延长它。所以我认为这可能是人类历史上最关键的时期。所以尽量在未来 10 年内别死。

Thank you. So, first of all, I'm very excited to be here. I'm a big follower of your podcast. I think it's maybe the best aging or longevity podcast. So this is a great pleasure. Um, yeah, so I've said that quite a few times in the last year or two actually, and it may not even take 10-15 years, might be even closer. The reason is that the technology, especially because of AI, is expanding exponentially. So our minds think in a linear term. So we think that the next 10 years is going to be as much advanced as the last 10 years or the last 15 years. But that's not what's going to happen. In the next 10 years, you can think of it as more advanced than the last century. So imagine that you were living in early 1900s and somebody told you that we're going to have vaccines and you will never get smallpox or you won't die of tuberculosis. You know, people would laugh at you. So that's not possible. So that's the speed that we're talking about. But there's something even more important because of this acceleration. The advances of treating diseases is also going to accelerate dramatically. So we will get to a point what's called the longevity escape velocity. This was coined by Aubrey de Grey, who's, as you know, a great aging researcher. So the point is that we will come to a point in the next, I would say probably 8 to 10 years, where every year you live is going to add more than a year to your life. So let's just say, you know, 10 years ago, 10 years later you get a cancer that's normally not curable and you only have one or two years to live. But during that one year, there is going to be a new treatment that will cure that cancer. So automatically it's going to add several years or maybe 10-15 years to your life. Or we're already starting to see that with the GLP1 drugs, receptor agonists, which are adding about 5 to 10 years to lifespan of people who are obese or who have chronic conditions. We'll have sort of the muscle generators, which I think will have tremendous impact on the aging population because, as you know, that's a huge problem. So all of these things will add up and the technology and AI is going to keep accelerating. So 10 years later, what will happen in a year will be like what happens in 20 years of advance. And then we'll get to a point, 15 maximum 20 years, where we will be able to completely reverse the aging process. So if you're 80 years old, 90 years old, you will get back to age 30, 40, whatever. So that's going to add up 50 years or 100 years to your lifespan. And then you can keep doing that and extend it almost indefinitely. So I think this is probably the most critical time in human history. So try not to die for the next 10 years.

我们将讨论所有这些事情。我想谈谈治愈疾病。我想谈谈逆转衰老、年龄逆转。所有这些都在我今天要和你讨论的议程上。但你提到了一些东西。你提到，现在人工智能作为一个通用术语，正在以指数级速度加速。我听过你谈论这个摩尔定律，以及软件本身如何以这种指数级速度加速。也许你可以解释一下这意味着什么，然后你认为这将如何转化为生物学，因为你知道，人类不是软件，而且有些事情，至少在我看来，你知道，你仍然必须测试安全性，我的意思是，所以如果你加速计算速度，因此你可以测试很多所谓的“in silico”的东西，对于听众来说，我们谈论的是测试东西，比如仅仅建模它们，也许你可以更好地解释这一点。但在某个时刻，你仍然必须测试安全性，而且你肯定有些事情我认为仍然需要在人体试验中完成。所以我很想听听你认为这将如何发生。

And we're going to talk about all these things. I want to talk about curing disease. I want to talk about reversing aging, age reversal. All of that is on my agenda to talk about with you today. But you mentioned something. You mentioned that right now the artificial intelligence as a general term is accelerating at an exponential rate. I've heard you talk about this Moore's law and how the software itself is accelerating right at this exponential rate. Maybe you could explain a little bit about what does that mean and then how do you think that'll translate into biology because you know humans we're not software and there are things that at least in my opinion you know you have to still test safety right I mean so like if you're accelerating the computational speed and therefore you can test a lot of things that are what are called in silico for people listening we're talking about testing things like just modeling them and maybe you can explain this a little bit better. But then at a certain point you still have to test about safety and you definitely that there are things that I think need to still be done in human trials. So I'd love to hear how you think that's going to happen.

我认为这是最关键的问题，因为人们总是提到这一点。

I think that's the most critical question because people always bring that up.

好的，你知道，如果你在几小时内生成药物，你仍然需要在人类身上测试五年，有时甚至更长。你打算怎么处理这个问题？

Okay, you know, if you generate drugs within hours, you still have to test them on humans for five years, maybe sometimes longer. How are you going to deal with that?

但让我先谈谈 AI 现在如何加速生物学。我们可以按阶段来思考，因为现在和 5 年后会非常非常不同。现在，尤其是在过去一两年，自从大语言模型出现以来，它们的智能一直在加速。最初，生产力提升较小。比如，当 GPT-4 刚出来时，我会让它扫描文献，告诉我某个话题的最新进展。那节省了我几个小时，有时甚至几天。但随着模型进步，尤其是在 01 模型之后，推理模型开始出现，现在我们有了 GPT-5 pro 模型、5.5 pro 模型。发生的是，它们现在能够思考和规划。所以你可以开始问非常复杂的问题。例如，这里有一个巨大的生物数据集，一百万个数据点或一千万个数据点。处理它。不仅要分析并分组，还要从数据中得出什么洞见？人类思维无法做到这一点。事实上，我们有这样的数据集，花了我们几个月来分析，比如一个博士生用深度学习来处理。我们仍然无法真正理解那些数据的含义。我们知道这些基因上调了，这些代谢物在变化，这些在发生。你怎么把所有这些整合起来？所以现在 AI 模型能够做到这一点。我测试过，比如最新的 GPT-5 pro 模型。你可以上传我们多年积累的数百万个数据集，然后在几分钟内，你不仅得到完整的分析——最近我得到了一份 40 页的 GPT-5 pro 报告，分析的是所谓的 RNA 测序，数百万个数据点——而且它还提供了惊人的洞见，比如这些数据意味着什么，下一步应该问什么问题。所以这自动将数月甚至数年的分析工作压缩到几分钟或几小时。所以这已经在加速了。当然，在药物设计方面，我认为每家制药公司最终都会使用 AI 生成来开发新药。过去需要数年筛选小分子的事情，现在只需几小时或几天。所以那里有巨大的加速。然后，最近，因为模型进步如此之大，你还可以问这样的问题：好的，这很棒，这是假设——事实上，AI 甚至能为你生成假设——但我应该做哪种实验来解决这个问题？人们必须意识到，我们在生物学中做的是实验，但我们并不真正知道最好的实验是什么。我的意思是，那是我的工作，但我有一些直觉，我们应该这样做来解决那个问题，但那是理想的实验吗？它有所有对照吗？所以 AI 模型现在能够告诉你，通过模拟，如果你能做的 100 个潜在实验中，这两个是最好的，因为它们会给你最好的输出。我一直在测试。所以那是另一个加速。现在你不需要花一年尝试 100 件事；你只需花几周尝试两件事就能得到结果。所以这就是现在可能的，已经极大地加速了研发部分。

But let me first start with how AI is accelerating biology now. So we can think of it in terms of phases, because now and 5 years later is going to be very, very different. Right now, especially in the last year or two since LLMs came out, their intelligence has been accelerating. Initially, it was fairly smaller productivity gains. For example, when GPT-4 was out, I would ask it to scan the literature and tell me what's the latest on this topic or that topic. That saved me hours, sometimes days. But then as the models advanced, especially after the 01 model, the reasoning models started to come out, and now we have the GPT-5 pro model, 5.5 pro model. What happened was that now they were able to think and plan. So you could start to ask very sophisticated questions. For example, here is a huge biological data set, a million data points or 10 million data points. Go over this. Not only analyze it and group them, but what is the insight from that data? Human mind is not able to do that. In fact, we had such data sets which took us months to analyze, like a PhD student working on it using deep learning. We still couldn't really truly understand what that data meant. We know these genes are up, these metabolites are changing, this is happening. How do you bring all that together? And so now AI models are able to do that. I've tested, for example, the latest GPT-5 pro model. You can upload millions of data sets that we accumulate over years, and then in a matter of minutes you get not only the complete analysis—recently I had a 40-page report from GPT-5 pro, which was an analysis of what's called RNA sequencing, lots of millions of data points—but it also provided incredible insight, like what does this data mean, what should be the next questions to ask. So that automatically contracts months, sometimes years of analytic work into a matter of minutes or hours. So that is already accelerating. Of course, in the drug design part, I think every pharmaceutical company is going to eventually use AI generation for developing new drugs. Things that took years of screening of small molecules now take hours or days. So tremendous acceleration there. And then, again more recently, because the models have advanced so much, you can also ask things like, okay, this is great, this is the hypothesis—in fact, AI can even generate hypotheses for you—but what sort of experiment should I do to address that? People have to realize that what we do in biology is experiments, but we don't really know what's the best experiment to do. I mean, that's kind of my job, but I have some intuition we should do this to address that question, but is that the ideal experiment? Does it have all the controls? So AI models are now able to tell you, sort of simulating, if out of these 100 potential experiments you can do, these two are the best ones because they are going to give you the best output. I've been testing that. So that is another acceleration. Now you don't have to try 100 things for a year; you can just try two things for a few weeks and get the output. So that's what's possible now, already tremendously accelerating the R&D part.

但第二部分，我认为更重要，是我们如何将其应用于临床试验和监管。在人类身上测试所有东西仍然需要数年，我认为解决方案将是我所说的数字孪生。这个术语已经存在好几年了。想法是，如果我们有大量大量的生物数据——当我说大量，是很多，PB 级的数据——如果 AI 达到一个点，我们需要比今天多得多的算力来计算所有这些，并真正模拟整个生物有机体，一个完整的人，但不仅是你表型，还有你的新陈代谢、免疫系统、肠道微生物组、遗传学，以及各种数据集整合在一起。所以它以一种时间性的、完全功能性的方式了解你的生物学。然后你可以问：好的，如果我给这个人这种药物，会产生什么效果？如果他们有这种紊乱，它会有副作用还是有效？所以字面上我们可以将临床试验时间从数年缩短到数月甚至数周。你实际上可以在非常小的患者子集中进行试验，因为你可以选择患者。你可以说，好的，AI 告诉我，对于这些人，这种药物 100% 有效。所以你只需在那些人身上测试。事实上，这将走向个性化。将会有成千上万种针对不同人群的药物。所以这将带来巨大的加速。我们还没有到那一步，但我打赌在五到十年内我们会达到。所以人类身上的迭代过程也将完全是数字化的，然后也许制造部分仍会花时间，但我们甚至也可以改进那部分。所以在某个时候，我们会达到按需治疗。你去一个 AI 模型，它分析你的基因组、你的生物学，为你订购这种小分子或药物或治疗到制造设施，下周你得到你的药物并接受治疗。那就是我想象的世界。

But then the second part, which I think is more important, is how do we apply that to clinical trials and regulations. It still takes years to try everything on humans, and I think the solution to that will be what I call the digital twin. This term has been around for several years. The idea is that if we have lots and lots of biological data—and when I say lots, it's a lot, petabytes of data—if AI comes to a point where we're going to need much more compute than we have today to compute all that and really simulate a whole biological organism, a whole human being, but not just your phenotype but also your metabolism, your immune system, your gut microbiome, your genetics, and all kinds of data sets put together. So it knows your biology in a temporal way, in a totally functional way. Then you can ask the question: okay, if I give this drug to this person, what kind of effect will it have? If they have this disruption, is it going to have a side effect or is it going to be effective? So literally we can cut down clinical trial time from years to a matter of months or even weeks. You can actually do the trials in a very small subset of patients because you can choose the patients. You can say, okay, AI told me that for these people, this drug is going to be effective 100%. So you just test it on those people. And in fact, that will go into personalization. There's going to be thousands of drugs for different people. So that will cause tremendous acceleration. We're not there yet, but I'm betting on that within five to ten years we will get there. So the iteration process on humans is going to be all digital as well, and then maybe the manufacturing will still take time, but we can even improve that part too. So at some point we will come to a point where treatment on demand. So you go to an AI model, it analyzes your genome, your biology, orders this small molecule or the drug or treatment just for you to the manufacturing facility, and next week you get your drug and you get treated. That's the world I'm imagining.

所以我想回到数字孪生这个概念，当我们谈论个性化医疗时。但如果我理解正确，如果我们有这个数字孪生，包含所有遗传数据、代谢组学、蛋白质组学、生物标志物，一切，对吧，所有这些数据以及我们没提到的更多数据，现在我们有了 AI，它可以进行所有这些情景模拟，弄清楚这种药物将如何影响或这种治疗将如何影响。而你说临床试验可能从几年缩短，也许我们可以在进行计算机模拟实验后，查看一些生物标志物，知道这是否会影响他们的生育能力？比如你不想给某人一种会使他们不孕的治疗。所以你认为那将是——AI 将能够识别如何知道它是否会影响生育能力、认知或预期寿命，仅从人的整体组成和进行所有这些测试？

So I want to get back to this concept of digital twin, again when we talk about personalized medicine. But if I understand correctly, so if we have this digital twin which is all the genetic data, metabolomic, proteomic, biomarker, just everything, right, all this data and more that we're not talking about, and now we have AI which can then do all these scenarios and figure out how this drug is going to affect or how this treatment is going to affect this. And you're saying that the clinical trial that may have taken a few years can be condensed down, and perhaps we can look at after doing the in silico experiments, you can look at some biomarkers and know like is this going to affect their fertility? Like you don't want to give someone a treatment that's going to make them infertile. So you think that's going to be—AI is going to be able to identify how to know if it's going to affect fertility or cognition or life expectancy, just from the whole composition of the person and doing all these tests?

是的。所以通往那里的道路需要几个验证步骤，我认为我们会达到一个点，当我们拥有超级智能时，我们将能够几乎 100% 信任它，甚至不需要用生物标志物或其他东西来验证。但要达到那个点，有点像自动驾驶汽车。要让自动驾驶汽车达到五级，它必须达到 99.999% 的安全。你必须验证它。如果有人过马路怎么办？那种场景必须发生，然后你记录下来，有时它不会做正确的事。也许它不会停下来。这就是为什么我们仍然要准备好接管控制。但如果它停下来并一次又一次地挽救生命，而现在自动驾驶汽车可能已经安全 10 倍，它们可能会安全 100 倍。所以你到了一个信任 AI 而不是司机的点。你说，好吧，我想让 AI 决定为我开车。所以我认为我们会在生物学领域达到那个点。这需要更长一点时间，因为极端复杂性。然后我们必须有非常聪明的基准测试和验证方法。生物标志物将非常重要，因为如果你正在开发一种声称能让人们活到 150 岁的抗衰老药物，你不能等待。即使一个 100 岁的人服用它，你仍然需要再等 50 年来验证。所以那行不通。我们必须能够预测它。但实际上，衰老在某些方面可能是最容易预测的，因为我们有这么多生物标志物或功能输出可以测量。我们知道它们在老年人和年轻人中的情况。所以如果你的视力突然变得像 20 岁的人，哇，那太棒了。如果你的肌肉和 30 岁的人一样好，如果你的皮肤看起来像 20 岁的人，那是我妈妈在等待的。那就是证明。而且你会立即看到，几周内或什么的。所以我认为这需要时间。那是需要时间的部分，信任 AI 告诉你，是的，如果你服用这种药，你会被治疗或逆转衰老。我们还有大约十年。这就是为什么我说否则它会更早发生。

Yeah. So the path there requires several steps of validation, and I think we will get to a point where when we have superintelligence, we'll be able to trust it almost 100%, that we don't need to validate it even with biomarkers or whatnot. But to get to that point, it's sort of like self-driving cars. To get a self-driving car to be level five, it has to be 99.999% safe. You have to validate it. What happens if somebody's crossing the street? That scenario has to happen, and then you record it, and sometimes it won't do the right thing. Maybe it won't stop. That's why we still have to be ready to take control. But if it does stop and saves lives again and again, and right now self-driving cars are probably about 10 times safer, they will be maybe 100 times safer. So you get to a point that you trust the AI rather than the driver. You say, okay, I want the AI to decide for me to drive. So I think we'll get to that point for biology. It will take a little bit longer because of the extreme complexity. And then we'll have to have very clever benchmarking and validation ways. The biomarker is going to be really important because if you're developing an aging drug that you claim will let people live to 150, you can't wait. Even if a 100-year-old takes it, you still have to wait 50 more years to validate that. So that's not going to work. We have to be able to predict that. But actually, aging is probably the easiest in some ways to predict because we have so many biomarkers or functional outputs we can measure. We know how they are in an old person and in a young person. So if your vision suddenly becomes like a 20-year-old's, wow, that's amazing. If your muscles are as good as a 30-year-old's, if your skin looks like a 20-year-old's, that's what my mom is waiting for. That's proof. And you'll immediately see that within weeks or whatnot. So I think it will take time. That's the part that's going to take time, the trusting AI to tell you yes, if you take this drug, you will be treated or you will reverse aging. We still have about a decade. That's why I'm saying otherwise it would happen even earlier.

你提到了超级智能，人工超级智能，ASI。也许你可以稍微谈谈，为了让人们现在有个理解，人工智能、人工通用智能（AGI）和超级智能之间的区别，因为你说一旦我们达到超级智能，我们就会信任它，对吧？所以我的意思是，我不知道，我们是否知道这些区别，还是你能解释一下？

You mentioned superintelligence, artificial superintelligence, ASI. Maybe you could talk a little bit about, just for people to have an understanding right now, the difference between artificial intelligence, artificial generalized intelligence, AGI, and then superintelligence, because you said once we get to superintelligence, we're going to trust it, right? So I mean, I don't know, do we know what those differences are, or can you explain a little bit?

是的，当然。你知道，这些定义每天都在变化，取决于谁的定义。但我已经思考 AGI 和 ASI 几十年了。这不是我最近才开始想的事情。我最初定义 AGI 的方式是，它是人工通用智能。这意味着，首先，它是人工的，对吧？所以它不是人类智能，它是人工智能。然后它是通用的。这意味着，如果 AI 学习了一套规则或一套知识，它可以将其推广到其他事物。这就是我们大脑智能的方式，因为你可以是一个惊人的国际象棋选手。事实上，AI 在 1997 年击败了国际象棋冠军卡斯帕罗夫，我想那是几十年前。但那不是通用智能。它只是在那方面非常擅长。或者 AlphaGo 击败了围棋世界冠军，那是一个困难得多的游戏。要成为通用的，AlphaGo 学习如何下围棋或国际象棋，应该能够，我不知道，解决衰老问题，对吧？所以它应该能够转移这些信息。我认为 LLM（我们称之为大型语言模型）的惊人之处在于，它们获得了这种能力，老实说，我没想到会这么容易发生。我原本期望 AGI 可能在十年前发生。所以在我看来，我们已经实现了所谓的 Level 1 AGI，人工通用智能，因为如果我问 GPT-5 Pro 模型一些它没有训练过的东西，比如我做过的实验，或者如果我说，好吧，把实验想象成一个视频游戏，为我设计另一个实验，就像你在玩视频游戏，所以那是将完全不同的领域转移到生物系统，它能够以惊人的方式做到这一点。但我们仍然需要经历几个级别。我认为下一个级别将是记忆。所以它们现在没有持久记忆。它们有一些记忆。它们了解你。它们知道它们在互联网上学到了什么，但它们需要能够管理上下文，因为有一个连续体。生活是一个连续体。然后另一个将是自我学习，对吧？所以也许那是第三级，没关系。而且那很快就会到来。AI 公司说实时学习可能明年就会到来。然后第三级，我称之为物理智能。人们再次对此非常困惑，因为真正的人类水平智能是物理智能，不是认知智能。数百万年来，我们进化到在物理世界中生存。我们直到，我不知道，一万年前才有语言，我们不知道如何写作。这个认知部分在过去大约 1 万到 2 万年间发展起来。在那之前，事实上，动物有非常好的物理智能。我们天生就带有这种智能。所以一个动物或一个孩子已经有一个世界模型。他们知道如果我扔下这个，它会掉下来，他们不必测试一百万次。那当然是我们对机器人、对具身所需要的。而且你可以看到那花了很长时间。训练一个机器人表现得像孩子比让 GPT-5 解决最难的数学问题更难。所以我们会到达那里。我认为人们正在研究这些世界模型和物理智能，无论我们是否需要另一种算法。所以那将是 AGI 级别的最后一级。

Yeah, of course. You know, this changes on a daily basis what the definitions are, depending on whose definition. But I've been thinking about AGI and ASI for decades. It's not something I started to think about recently. The way I originally defined AGI, it's artificial general intelligence. What that means is, first of all, it's artificial, right? So it's not human intelligence, it's artificial intelligence. And then it's general. What that means is that if AI learns one set of rules or one set of knowledge, it can generalize that to something else. And that's how our brains are intelligent, because you can be an amazing chess player. In fact, AI beat the chess champion Kasparov in 1997, I think, decades ago. But that was not general intelligence. It was super good at that. Or AlphaGo beat the world champion in Go, which is a much more difficult game. To be general, AlphaGo, learning how to play Go or chess, should be able to, I don't know, solve a problem in aging, right? So it should be able to transfer that information. I think the amazing thing about LLMs, what we call large language models, is that they acquire this ability, which honestly I didn't think would happen so easily. I was expecting AGI to happen maybe a decade ago. So in my opinion, we have already achieved what I call level one AGI, artificial general intelligence, because if I ask GPT-5 pro model something that it hasn't trained on, like an experiment that I have done, or if I say, okay, think of the experiment as a video game, design another experiment for me, like you are playing a video game, so that's transferring completely different area to a biological system, and it's able to do that in an amazing way. But we still need to go through several levels. I think the next level is going to be memory. So they don't have persistent memory right now. They have some memory. They know about you. They know about what they've learned on the internet, but they need to be able to manage the context, because there's a continuum. Life is a continuum. And then the other one is going to be self-learning, right? So maybe that's level three, it doesn't matter. And that's coming soon. AI companies are saying that real-time learning might come maybe by next year. And then the third level, what I call physical intelligence. People again confuse this greatly, because true human-level intelligence is physical intelligence, it's not cognitive intelligence. For millions of years, we evolved to survive in a physical world. We didn't have language up to, I don't know, 10,000 years ago, we didn't know how to write. This cognitive part has developed in the last maybe 10 to 20,000 years. Before that, in fact, animals have very good physical intelligence. We're imprinted and born with that intelligence. So an animal or a child already has a world model. They know that if I drop this, it's going to fall, and they don't have to test it a million times. And that's of course what we need for robots, for embodiment. And you can see that that's taken a long time. It's more difficult to train a robot to behave like a child than have GPT-5 solve the most difficult math problem. So we'll get there. I think people are working on these world models and physical intelligence, whether we need another algorithm or not. So that will be the final level of the AGI level.

一旦我们达到所有这些层级，并且 AI 能够自我学习，那就是超级智能了，因为到那时它可以自我训练，速度可能比我们快数千倍甚至数百万倍。人类智能是有极限的，对吧？即使世界上最聪明的人也只能做这么多。我定义的超级智能，是在某个时刻拥有全人类智能的总和。如果你把一百万顶尖科学家聚在一起，就像曼哈顿计划那样，他们解决了非常困难的问题。超级智能将达到那个水平。数千名科学家一年能做的事情，它一天就能完成。所以我很可能会相信这一点。

Once we have all those levels, and once the AI is able to self-learn, then that's the superintelligence, because at that point it can train itself, maybe thousands or millions of times faster than we can. And there's a limit to human intelligence, right? Even the smartest person in the world can only do so much. Superintelligence, as I would define it, is having the combined intelligence of all humanity at some point. If you bring a million top scientists together, like the Manhattan Project, they solved very hard problems. Superintelligence will reach that level. What thousands of scientists can do in a year, it will do in a day. So I would probably trust that.

哇，这太令人兴奋了。这也引出了这样一个概念：当你和人们谈论 AI 时，并不是每个人都像你一样理解它。你会听到悲观和乐观两种观点。我经常和人交谈时，听到很多悲观情绪，也许是出于对未知的恐惧，对 AI 能力的恐惧。你所说的超级智能，如果向某些人解释，会更吓到他们。他们可能担心文化影响、经济影响，还有那种《终结者》式的情景：它们超级聪明，会想接管世界，不再需要我们了，对吧？

Wow, that's pretty exciting. It also brings in this concept that when you talk to people about AI, not everyone has the same understanding as you. You hear pessimistic versus optimistic views. Often when I talk to people, I hear a lot of pessimism, perhaps fear of the unknown, of what AI is capable of. The superintelligence you're talking about, if you explained that to some people, it would scare them even more. They might worry about cultural ramifications, economic ramifications, but also this Terminator situation: they're super smart, they'll want to take over the world, and they won't need us anymore, right?

但你的观点如此乐观。我的意思是，我们在谈论解决衰老问题，活到 150 岁甚至更久。你为什么如此乐观？你难道不担心那些悲观的看法吗？

But you have such an optimistic view. I mean, we're talking about solving aging, living to be 150 or more. Why do you have such an optimistic view? Are you worried at all about the pessimistic viewpoints?

绝对不是，我会告诉你为什么我对此如此超级乐观。当人们说 AI 是生存威胁，会毁灭人类时，我提出反驳：对人类来说只有一个生存威胁，那就是人类自己。如果你看看历史，人类杀死的人类比所有其他因素加起来还多，造成的痛苦比人类经历的任何事物都多。即使是动物，传染病在某个时候可能造成了很多痛苦，但真正的危险是人类智能。所以让我们做个思想实验。想象我们生活在一个平行宇宙，那里世界决定任何智商超过 100 的人都是社会的危险，因为如果你非常聪明，你可能会想出危险的想法。这其实是事实。然后如果你的智商是 105，你立刻被监禁，不允许参与社会，或者被杀。社会决定智能是危险的，所以我们要阻止它。我们会生活在什么样的世界？我们不会有现在拥有的任何东西。我们可能只是农民，用基本的身体智能生存，在一个平均寿命 30 岁的世界里。所以我们应该这样看待 AI。另一个观点是关于 AI 接管并取代我们。我看到的恰恰相反，因为 AI 是一个不可思议的赋能者。它给你超能力。即使是现在，我感觉自己拥有超能力。我这辈子从未这么忙过。我睡得更少，顺便说一句，这不是好事。我不推荐，但因为我能做这么多，它太赋能了。我妈妈 86 岁了，她告诉我 ChatGPT 改变了她的生活。她充满活力，不再那么担心健康。它产生了不可思议的影响，而且这还会加速。在某个时刻，我们将与 AI 融合，通过 Neuralink 类型的脑机接口直接交互。我们将在自己的大脑中拥有 AI 的智能，不仅直接，而且通过工程改造我们的生物系统间接实现。所以为什么每个人不应该拥有爱因斯坦或更高的智能？爱因斯坦和普通人之间的差异可能只是几个单点突变。如果我们能工程化这一点，如果 AI 能教我们怎么做，那么我们就能走得更远。只要我们保持自主权，那是我们必须保护的唯一东西：我们是决策者，我们把 AI 视为合作者，另一个与我们共存的物种，互相赋能。在某种程度上，它是我们的孩子，由我们创造。我认为世界变得更糟的概率极低。当然，永远不会是零，但你一出生就会死。AI 给了我们拯救数十亿生命的机会。不仅仅是延长寿命 5 年或 10 年，而是数千年。那才是真正的拯救生命。关键风险是人类滥用 AI。这就是我们必须对齐 AI 的原因。不要看坏人类；为我们判断更好的世界。当然，我可能错了，但我相当确定我会是对的。

Absolutely not, and I'll tell you why I'm so super optimistic about it. When people make statements like AI is an existential threat and it's going to destroy humanity, I make the counterpoint: there's only one existential threat to humanity, and that's humanity. If you look at history, human beings have killed more humans than everything else combined, caused more suffering than anything humans have been exposed to. Even animals, infectious diseases at some point might have caused a lot of suffering, but the real danger is human intelligence. So let's do a thought experiment. Imagine we live in a parallel universe where the world decided that anyone with an IQ above, say, 100 is a danger to society, because if you're very intelligent, you can come up with dangerous ideas. And that's true, actually. Then if you have an IQ of 105, you get imprisoned immediately, not allowed to participate in society, or you get killed. Society decided intelligence is dangerous, so we're going to stop it. What kind of world would we live in? We would have nothing we have now. We'd probably be just farmers, using basic physical intelligence to survive, in a world where the average lifespan was 30 years. So that's how we should view AI. The other point is about AI taking over and replacing us. I see it exactly the opposite, because AI is an incredible enabler. It gives you superpowers. Even now, I feel like I have superpowers. I've never been this busy in my life. I sleep less, which is not a good thing, by the way. I don't recommend it, but because I can do so much, it's so empowering. My mom was 86 years old, and she told me that ChatGPT changed her life. She's energized, she doesn't worry as much about her health. It's had an incredible impact, and this is going to accelerate. At some point, we will merge with AI, having direct interaction through Neuralink-type brain interfaces. We'll have the intelligence of AI in our own brains, not only directly but also indirectly by engineering our biological system. So why shouldn't everyone have the intelligence of Einstein or higher? The difference between Einstein and a normal person is probably a few single-point mutations. If we can engineer that, if AI can teach us how, then we can go much higher. As long as we keep agency, that's the only thing we have to protect: that we are the decider, that we see AI as a collaborator, another species that lives with us and empowers each other. In a way, it's our child, created by us. I see the chance of a worse world as extraordinarily low. Of course, it's never zero, but the moment you're born, you're going to die. AI gives us the opportunity to save literally billions of lives. Not just extending life by 5 or 10 years, but thousands of years. That's true saving lives. The key risk is humans misusing AI. That's what we have to align AI against. Don't look at the bad humans; judge the better world for us. Of course, I might be wrong, but I'm pretty sure I'm going to be right.

我同意我们必须警惕人类的说法。当然，因为你说得对，人类过去一直是人类最大的威胁。我想回到你提到的几点，当你谈论 ASI 和自我学习能力时，甚至你使用 GPT-5 Pro 帮助设计实验和解释结果的方式。这是我作为生物学家的问题。正如你所说，我们做实验，检验假设，然后我们有所有这些数据和结果，我们必须知道什么结果有意义，什么异常有意义，因为通常异常……

I agree with the statement that we have to watch out for the humans. For sure, because you're right, they can and have in the past been the biggest threat to humanity. I want to go back to a couple of things you mentioned when you were talking about ASI and this ability to self-learn, and even the ways you use GPT-5 Pro to help design experiments and interpret results. That was a question I had as a biologist. As you mentioned, we do experiments, test hypotheses, and then we have all this data and results, and we have to know what result is meaningful and what anomaly is meaningful, because often the anomaly...

对。

Right.

……你可能会忽略。

...which you might ignore.

正是如此。

Exactly.

你绝对认为这是突破，对吧？这是一种直觉，这种生物学直觉。那么你首先认为，我们现在拥有的模型已经能够具备这种生物学直觉了吗？如果没有，那还有多远？

Is what you absolutely is the breakthrough, right? And that is a sort of intuition. This biological intuition. And so you do you think first of all do you think we're that that you know the models we have now can already are capable of that sort of biological intuition and if not like how far off is that?

是的。是的。这是个很好的问题。事实上，我认为那种直觉可能算是解决方案的最后一英里，或者说最后 10%。因为 90% 的 AI 模型能够做到，是因为它基于知识，在人类中也是如此。对于医生、科学家或任何人来说，90% 或 95% 是基于已知的知识以及你如何处理这些知识，但还有额外的 5% 到 10% 完全依赖于你的直觉。比如，如果你是医生，看到病人进门，你就知道那个人心脏病发作了。你还没做任何检查，但你就是知道，你不知道自己怎么知道的。在实验室里也是如此。事实上，我会和我的学生和博士后打赌，我会说，我打赌如果你做这个实验，你会得到这个结果。我从未输过赌注，他们也不再和我打赌了，尽管这可能看起来违反直觉，比如“哦，那永远不会成功”，但不知何故我就是知道。我怎么知道？因为我在实验室工作了 30 多年，你会获得一些文献中没有的东西，或者你无法通过教科书学到的东西。你只能通过实践来获得。所以，我认为直到最近，模型在 5.5 之前都能很好地处理那 90% 的部分。尤其是在 GPT-5 Pro 发布之后。比如，我会给它一个我们已经做过的实验，一个非常复杂的实验，花了两个星期。我已经知道结果了，因为我们完成了实验。但我想看看模型会如何预测实验的结果。它们，不仅仅是 GPT-5，还有其他几个模型，都能做到 80% 到 90% 的正确率。我的意思是，那已经相当不错了。它们会说，好的，两天后会发生这个，一周后会发生那个，两周后会发生那个。但我所拥有的那种额外的直觉，我本可以预测到的，仍然有所欠缺。我认为 GPT-5.5 跨越了这个门槛。所以我用 5.5 Pro 模型重复了那个测试，因为我总是说 Pro 和思考模式非常不同，当然和即时模式也非常不同，因为 Pro 会推理更长时间，它在思考。所以在某些情况下，我让它思考了两个小时。AI 思考两个小时相当于人类思考多年。所以那个模型真的跨越了那个门槛。在我给你的那个例子中，它几乎达到了 100%，我会说 98% 的正确率，和我本会预测的一样。我自己都不会和 5.5 Pro 打赌。所以对我来说，这真的令人难以置信，因为我无法理解这些模型被训练了所有信息，我们无法与之竞争，对吧？所以它们能把这些模式组合在一起，但为什么模型现在几乎拥有了我花了 30 年才获得的经验，那种直觉，现在达到了那个水平，这很神秘，但我接受它。

Yeah. Yeah. That's that's a great question. Uh in fact um uh you know I I see that intuition maybe sort of the the last mile or the top 10% uh or 10% of the of the solution because 90% um AI models they are able to come up with because it's it's knowledge based also in humans is is you know for for a medical doctor for a scientist for whoever 90% or 95% is based on uh what's known how you process that knowledge, but there's that extra 5 10% totally dependent on your intuition. Uh like you you if you're a doctor, you see a patient coming through the door, you know that guy is having a heart attack. You haven't checked anything yet. Somehow you know, you don't know how you know. the same thing in the lab like um uh in fact I would I would bet with my uh students and and postto I would say okay I bet you if you do this experiment you're going to get this result um and I've never lost a bet and they stop betting against me even though it might look counterintuitive oh no that's never going to work somehow I know how do I know because you know I've been working in the lab for 30 plus years and and you you acquire ire certain things that are not in the literature or you know you can't really read a textbook and learn it. You only do it by by practicing it. Um so the the models up to I would say 5.5 until recently were were great at that 90% level. Uh so especially after GPT5 pro came out. So you know I would ask it to for example I I would give it an experiment that we have already done. It's a very complex experiment took two weeks. I already know the result because we're done the experiment. But I wanted to see how the model would predict the outcome of the experiment. And they would do you know not just GPT5 but several other models as well. Um they they would come up with 90% 80 to 90% correctly. I mean that's that's pretty good. Uh they would say okay this is what's going to happen after two days after one week after two weeks. But that extra level of intuition that I I have I would have predicted was still somewhat lacking. I think GPT 5.5 crossed that threshold. So I I repeated that with with the uh 5.5 pro model uh because I I always say pro uh it's very different than the thinking of course very very different than the instant model because pro uh is reasoning much much longer. It's thinking. So in some cases I I pushed it to think for two hours. So two hours in AI thinking is like years of thinking for for a human being. So that model really crossed that threshold in that example I gave you. It was almost 100%. I mean I would say 98% correct. What I would have predicted like I would not have bet against 5.5 Pro myself. Um uh so that to me is is actually really mind-boggling because uh I couldn't understand these models are being trained with all of the information we can't compete with that right so it's they they can put these patterns together but how is it that the model has now almost the experience that I have that I spent 30 years acquiring that experience that intuition that is now getting to that level that is uh that is a mysterious but uh I I I live to it. Now

你说你让 GPT-5.5 Pro 思考了两个小时。我们说的是什么样的提示词，还是也包括数据集和提示词？我的意思是

What sort of you said you you pushed GPT 5.5 Pro to think for 2 hours. I mean what sort of prompt are we talking about or is it just the data set too and the prompt? I mean

那些通常是数据集。我可能打破了纪录，因为我甚至问过 OpenAI 的朋友，我不认为他们把它推得那么远。所以这实际上是 2R 那个，是巨大的数据集，数百万个数据点。然后我还说，不要只是分析它，写一份大报告，你知道，30 到 40 页，随便多长，然后对数据提出很多见解，问什么问题，我们学到什么机制。那是一个免疫学数据集，序列、基因、蛋白质等等。所以那个我记得是 112 分钟，它生成了这份 40 页的报告，我简直不敢相信。你知道，分析部分之前的模型也能做到，它们会说，好的，有这些类型的基因和这种蛋白质，所以这意味着这个和那个，你从这些信息中推导出来。但要提出见解，这可能意味着什么，或者下一步该问什么问题，那是非常非常高水平的推理。所以，是的，那两小时绝对值得。听到你这么说很令人兴奋，因为那正是我想知道的，我想知道这是否已经可能，看起来确实如此。所以这也引出了下一个问题，那就是现在所有科学家真的需要开始理解如何正确使用 AI，对吧？我的意思是，这就像帮助他们。

So those are usually data sets. Uh um I might have broken a record because I even asked the the friends at OpenAI. I don't think they they pushed it that that far. Uh so this was actually the the 2R one was u uh huge data sets millions of data points um uh and um um and then I I also said okay don't just analyze it write a huge report you know 30 40 page whatever length and then you know come up with a lot of insights about this data what questions to ask and what do we learn the mechanism it was a an iminological ical data set um sequence and genes and proteins and all that and so so that one I think 112 minutes I remember that uh uh and it came up with this 40page report uh which I I I was just unbelievable uh you know the the analysis part the previous models were able to do as well you know you know they say okay well there are these type of genes and this type of protein so it means this and that you it deres from that information, but to come up with an insight what that could mean or what would be the next question to ask. Uh that that's that's a very very high level of reasoning. Um and so uh yeah um it was it was worthwhile two hours for sure. I mean that's very exciting to hear you say that because that was kind of my I wanted to know I wanted to know is that something that is already possible and it seems like it is and so it also leads to the next question which is you know all all these scientists now really need to start understanding how to use AI in the right way right I mean this is like to help them

我的意思是，这将会发生，对吧？这基本上就是，你知道，我们现在都用 Google，还记得 Google 刚出现的时候吗？所以，我的意思是，这最终会发生，但想到 AI 将如何改变研究和医学，这非常令人兴奋。而且，你知道，你提到过，我也说过，我想回到数字孪生这个想法，因为我听过你谈论它，这让我非常兴奋。你知道，几十年来我们一直听说个性化医疗即将到来。我们将拥有个性化医疗，但至今我们还没有。它就是不在这里。而且，我听过你，甚至听你说过一些有趣的话，也许我不是直接引用，但大意是，现在医生不使用 AI 在某种程度上应该算是医疗事故。那么，你能谈谈你为什么这么说吗？医生负责任地使用 AI 意味着什么？还有，患者如何为自己辩护，因为那也是另一个领域。

I mean this that's going to happen right that's basically you know we all we all use Google now remember when Google was like new So, I mean, it's eventually going to happen, but um it's very exciting to think about how AI is going to change research and and medicine, and that's something, you know, you you mentioned and I talked I said I wanted to get back to this digital twin idea because I've heard you talk about it and it's very exciting to me. You know, we've heard for decades now that personalized medicine is coming. We're going to have personalized medicine and and yet still we just don't have it. It's just not there. Um, and I've heard you I've even heard you say something sort of interesting which perhaps I'm not saying the direct quote, but that it kind of should be medical malpractice in a way for a physician today right now to not be using AI. So, can you talk a little bit about why you said that? What it means to for a physician to use AI responsibly? um also how patients can self- advocate for themselves because that's also another area.

我一直在测试大量医学问题，有些很难，有些涉及实时数据。在 01 之前，它在查阅文献方面很出色，比如医生可能缺乏某些知识，它能知道最近发表的内容。但它的推理水平还不行。所以某个模型能够推理了。而推理在医学中极其重要，因为即使你掌握了所有信息，你仍然要考虑那个人的具体情况，以及什么治疗方案更可能有效。我们并不总是知道答案，也不知道如何诊断。所以我认为 01 达到了那个水平，那时我就说，现在医生不使用 AI 已经不道德了。我还没说这是医疗事故，但确实不道德，因为你可以使用它，你显然仍然可以运用自己的判断，但它能防止你犯一些明显的错误，有时是不明显的错误，或者诊断出需要多个临床专科协作才能诊断的疾病，而如果你住在乡村，你就不具备那种能力。但现在我觉得，在我看来，这真的会被视为医疗事故。虽然法律上还不是，但最终会是的，因为当前的高级模型能够比该领域的专家更好地或同样好地诊断并制定治疗方案。这不仅仅是全科医生。假设你有一个非常复杂的癌症，你知道突变等等，你去看一位非常专业的肿瘤科专家。我相信当前模型已经达到那个水平。当然，并非每个专家都是顶尖专家，对吧？如果是那样的话，美国每年就不会有数百万的误诊和误治了。我想他们说过大约有 1200 万次误诊。我认为有 70 万人因此受苦，死于误诊。其中有些是完全无辜的，任何医生都可能漏掉。但现在 AI 不会漏掉。所以即使是专家也可能犯错或误诊或误治，因为他们缺乏模型所具备的某些东西。想象一下，你拒绝使用 MRI 或 CT 机器，因为你说，那技术太先进了，我只用 X 光就够了。然后你漏掉了一个肿瘤。AI 模型能够比放射科医生提前几年检测出某些肿瘤，比如乳腺癌。如果你漏掉了，那个人就会死。所以在我看来，这就成了医疗事故，因为技术已经达到那个水平了。五年前漏诊乳腺癌不算医疗事故，因为那时还没有那种技术，但现在我们有这种技术，所以你绝对应该使用它。这会拯救很多生命。如果你能减少误诊，让每个医生都达到超级医生的水平，我认为那将是非常好的事情。

So been testing lots of medical questions and some of them are hard, some of them are real-time data. Before 01, it was great in reaching to the literature, like the physician might lack certain knowledge, so it knows what was published recently. But it was not at the reasoning level. So one model was able to reason. And the reasoning is extremely important in medicine because even if you have all the information, you still have to consider that person's context and what would be more likely to treat that person. And we don't always know the answer as well or how to diagnose it. So I think 01 was able to get to that point, and at that point I said right now it's unethical for physicians not to use AI anymore. I didn't say malpractice yet, but it truly unethical in the sense that you can use it, you can still do your judgment obviously, but it will prevent you missing some obvious mistake or sometimes nonobvious mistakes or diagnose things that require multiple clinical specialties coming together, and you don't have that capability if you live in a village or something. But now I think I feel that it is truly going to be considered malpractice in my opinion. It's not legally so, but eventually it will be because the current models, the advanced models, are able to diagnose and write a treatment protocol better than or as good as a specialist in that field. It's not just a family physician. Let's say you have a very complex cancer, you know the mutations and whatnot, and you go to a specialist like an oncologist who is very specialized on that. I believe that the current models are at that level. And of course not every specialist is the top specialist, right? So if that was the case, we wouldn't have millions of misdiagnoses and mistreatments in the US alone every year. I think they've said something like 12 million misdiagnoses. I think 700,000 people suffer from it, die from it, from misdiagnosis. Some of them is totally innocent. Any doctor could have missed it. But now AI wouldn't miss that. So even a specialist might make a mistake or misdiagnose or mistreat because they lack certain things that the model doesn't have. So imagine that you refuse to use an MRI machine or CT machine because you say, well, that's too much technology. I'm just going to do an X-ray because that's enough for me. And you miss a tumor. The AI models are able to detect certain tumors like breast cancer years before a radiologist is able to see that. So if you miss that, that person's going to die if you don't. So that to me becomes malpractice because the technology is at that level now. It wasn't malpractice missing a breast cancer five years ago because nobody had that technology, but now we have that technology, so you should definitely use it. And this is going to save a lot of lives. If you could just reduce the misdiagnosis and bring every doctor to super doctor level, I think that would be a really good thing.

所以你的意思是，基于医生目前可用的数据，无论是 MRI、超声还是血液生物标志物，这些数据被输入到像 GPT-5.5 Pro 这样的模型中，利用这些数据，它们能够更好地诊断、更好地预测，看到你提到的癌症等问题。这比放射科医生更好吗？这是基于什么样的数据实现的？

So what you're saying is based on the current data that doctors have available to them, whether it's an MRI, whether it's an ultrasound, whether it's blood biomarkers, this sort of data is what is given to a model like GPT-5.5 Pro, for example, and with that data they're able to better diagnose, better predict, to see things like you mentioned cancer. Is that better than a radiologist can? Is that based on what kind of data is implemented?

这些是有研究的。我认为谷歌最近做了一项研究。事实上，最近发表了一篇《科学》论文，用的是 01 预览模型，那是一个非常老的模型。我的意思是，当前模型可能要好 10 倍甚至更多。

These are studies. I think Google did a recent study. In fact, a Science paper came out recently which was done with the 01 preview model, which is a very old model. I mean, the current models are probably 10 times or maybe more.

那是不是像第一个专业版，几乎是你 2024 年 9 月早期测试的第一个推理模型？

Was that like the first pro, almost like the first reasoning model that you early tested in 2024 September?

它发布了，他们发现 01 模型在诊断方面明显优于普通医生。他们没有漏诊。所以想象一下当前模型有多好。但我认为这不仅仅是诊断疾病，因为那实际上只是医生工作的一小部分。这真的是一个连续的过程。大多数疾病，好吧，如果你得了流感或某种细菌感染，你知道该怎么做。你给药，然后观察结果。但很多疾病，即使在这种情况下，也可能不是这样，因为你可能遇到变异病毒或细菌。所以你可能需要改变治疗方案，或者可能会有一些副作用。所以那里有很多连续性。所以我认为 AI 可以参与整个过程。如果你能持续输入数据，好吧，病人我们给了这种治疗，情况不错，血压降下来了，但有这个症状，那我们该怎么办？改变药物剂量，还是加这种药，或者去掉那种药，换另一种抗生素？这是一个持续的过程，而且并不总是持续的，因为人们不会每天都去看医生，对吧？你拿到处方，看到有效，然后再回去。那么如果有东西能在癌症治疗后持续监测你呢？这非常重要，因为癌症是一种非常动态的疾病。癌细胞不断试图存活、突变并抵抗免疫系统。所以你给一种药，化疗有效，然后癌症又复发了，对吧？为什么？因为突变在积累。所以我们能更早发现吗？我们能改变那些决策吗？我们能确保在癌症有机会复发之前，给予多种药物或不同的药物，从而阻止这种可能性吗？所以所有这些决策都可以与 AI 一起做出，我认为它将对医疗保健产生巨大的影响。

It came out and they found that the 01 model did better than the average doctor in diagnosing, significantly better. They didn't miss. So imagine how good the current models are. But I think it's not just about diagnosing a disease, because that's actually a small part of the job of a doctor. It's really a continuum. Most diseases, okay, if you have a flu or some bacterial infection, you know what to do. You give it and then you see an output. But a lot of diseases, even in that condition, that may not be true because you might have a mutant virus or bacteria. So you might have to change the treatment or might have a little bit of side effect. So there's a lot of continuum there. So I think AI can be involved in all of that process. If you can continuously feed the data, okay, the patient we gave this treatment, it's doing well, the blood pressure is down, but has this symptom, so what should we do? Change the dose of the drug or add this or remove that drug and give another antibiotic? There's a constant process there, and that's not always constant because people don't go to the doctor every day, right? You get a prescription, you see something works, and then you go back. So what if there's something that's continuously monitoring you post-treatment for cancer? It's very important because cancer is a very dynamic disease. The cancer is constantly trying to survive and mutate and counteract against the immune system. So you give a drug, chemotherapy works, and then the cancer comes back again, right? Why is that? Because mutations are accumulating. So can we catch that earlier? Can we change those decisions? Can we make sure that we give multiple drugs or different drugs so before the cancer has the opportunity to come back, we prevent that possibility? So all of these decisions can be made together with AI, and I think it's going to have tremendous impact on healthcare.

我确实想稍后回到癌症这个话题，但在此之前，我只是觉得医生们，并非所有医生都知道如何使用 AI。他们不知道使用哪些模型。他们是用 GPT-5.5 Pro 还是 Claude，还是如何负责任地使用它，你刚才稍微提到了一点，但不要外包他们的临床判断？

I do want to get back to the cancer equation in a minute, but before that, I just think that physicians, not all physicians know how to use AI. They don't know which models to use. Do they use GPT-5.5 Pro or Claude or how do they responsibly use it, which you kind of talked about a little bit, but without outsourcing their clinical judgment?

你对使用不同模型有什么看法吗？我知道你和 OpenAI 有合作。你一直是最早真正在生物学领域测试这些模型的科学家之一。但我确实认为，听众中的医生和人们想知道他们该用什么模型。我们肯定在谈论 OpenAI，那一定是 Pro 版，对吧？一定是推理模型，但 Claude 呢？Gemini 呢？

Do you have any opinions on the different models to use? And I know you have a collaboration with OpenAI. You've been one of the first scientists really testing these models in a biological sort of arena. But I do think that people and physicians listening want to know what models they use. We're definitely talking about OpenAI, it's got to be the pro, right? It's got to be the reasoning model, but what about Claude? What about Gemini?

是的。所以，我认为人们有一种误解，他们把 AI 想成：好吧，我们有 AI，我们有互联网，那我们就用互联网吧。我们有 AI，那就用它。但这发展得太快了。我们一个月前用的 AI 模型和我们现在用的不一样。它每隔几个月智力就翻倍。我举过一个预览版的例子。有些人还停留在 GPT-4o 模型。哦，是的，我用过它，它经常产生幻觉。甚至 01 也不太好。它会犯错。那就像古代历史了。

Yeah. So, I think people have this misunderstanding of thinking of AI as okay, we have AI, we have internet, so let's just use the internet. We have AI, let's use it. But this is advancing so rapidly. The AI model we used one month ago is not the same AI model we use now. It's just doubling in intelligence every few months. I gave the example of one preview. Some people got stuck at the GPT-4o model. Oh yeah, I used it and it hallucinated a lot. Even 01 wasn't so good. It was making mistakes. That's like ancient history.

这就是为什么我甚至没问幻觉的问题。

That's why I haven't even asked about hallucinations.

是的。所以，我的优势在于，因为我全身心投入 AI，我持续测试，所以我能看到这些模型的进化。它们变得好 90%、95%、97%，就像不断自我更新。最终，现在用 5.5 模型，我完全看不到任何幻觉。我的意思是，可能有 0.1%，但极其罕见。所以你的信任度会上升。这又类似于自动驾驶汽车，对吧？我们拥有自动驾驶汽车将近十年了，它们不断变得更好，因为它们的 AI 模型在更新。所以我的建议是，医生应该把这看作不是可选的。他们必须定期更新医学知识。事实上，他们必须通过考试才能获得认证，或者他们必须了解新出的药物。对吧？你不能只依赖 5 年前、10 年前出的药。你需要知道上个月批准了什么，你需要以类似的方式更新知识，甚至更要如此。他们必须不断更新 AI 知识。所以 AI 必须成为他们实践的一部分。当然，我的建议是始终使用你能用的最新顶级模型。现在是 GPT-5.5。事实上，对于复杂问题，我总是使用 Pro 模型，因为它会思考几分钟。但至少如果你日常快速使用，始终使用思考模型。思考模型不同于即时模型。即时模型也在变得更好，但它需要推理。它需要思考。特别是如果你输入大量患者数据并进行分析，你绝对需要 Pro 模型。

Yeah. So, the advantage I have is that because I'm all in on AI, I'm continuously testing, and so I can see the evolution of these models. They get 90% better, 95% better, 97% better, like it just continuously updates itself. And eventually, right now with the 5.5 model, I don't see any hallucinations whatsoever. I mean, there might be 0.1%, but it's extremely rare. So your trust level goes up. Again, it's similar to self-driving cars, right? We had self-driving cars for almost a decade, and they just keep on getting better and better because their AI models are getting updated. So my advice would be doctors should see this not as something optional. They have to update their medical knowledge periodically. In fact, they have to have tests to be certified, or they have to update on new drugs that are coming out. Right? You can't just rely on some drug that came out 5 years ago, 10 years ago. You need to know what was approved last month, and you need to update your knowledge in a similar way, even more so. They have to constantly update their AI knowledge. So AI has to be part of their practice. And of course, my recommendation is always use the latest top model you can use. Right now it's GPT-5.5. In fact, I would always use for complex problems the pro model because that thinks in minutes. But at least if you're using it on a daily basis in a rapid fashion, always use the thinking model. The thinking model is different than the instant model. The instant model is also getting better, but it needs to reason. It needs to think. And especially if you're putting in lots of patient data and analyzing that, you definitely need the pro model.

然后还有像 OpenEvidence 这样的公司。我认为大多数医生开始使用它。OpenEvidence 基本上应用了最新模型。它更新了，所以医生不用担心。我认为会有更多这样的公司提供这种服务。所以医生不必担心该用 5.5 还是 4.7 还是什么。那个“套件”模型会为医学挑选最好的，并应用它。当然，医院应该像大型科技公司一样实施 AI。有企业级 AI 可以更安全，保护患者数据。所以应该像在医院患者面前，你看到这些监视器，比如心跳之类的。应该有一个 AI 监视器不断监控数据，然后给护士和医生提供信息。好的，这是最后的情况。现在有了 AI 智能体，你可以做到这一点。就像我在日常生活中做的那样，比如我的电子邮件。我的智能体自动检查我的邮件，告诉我什么重要，所以我不必浏览数百封邮件。所以，哦，你知道，这个在等你。你今天和 Rhonda 有个播客，所以你最好准备好。所以是的，它需要完全整合，几乎像一个共同医生。AI 医生和真正的医生一起工作。

And then there are companies like OpenEvidence. I think most doctors are starting to use that. OpenEvidence basically applies the latest model somehow. It's updated so the doctors don't have to worry about it. And I think there's going to be more companies like that who will provide that service. So the doctor doesn't have to worry about should I use 5.5 or 4.7 or whatever. The harness model is going to pick the best one for medicine and apply it there. And of course, hospitals should implement AI just like big tech companies. There's enterprise-level AI that can be more secure, protect patient data. So it should be like in front of the patient in the hospital, you see these monitors like the heartbeat and all that stuff. Should be an AI monitor constantly monitoring the data and then giving information to the nurses and the doctors. Okay, this is the last situation. And now with AI agents, you can do that. Like I do it for my daily life, for my email. Automatically my agents go and check my email and they tell me what's important, so I don't have to go through hundreds of emails. So, oh, you know, this is waiting for you. You have a podcast with Rhonda today, so you better be prepared for that. So yeah, it needs to be fully integrated, almost like a co-physician. You have the AI doctors working together with real doctors.

对。我注意到我联系或互动过的一些公司似乎经常使用 Claude。我不知道你是否试过，但我很好奇为什么是那个模型，而不是其他。但公平地说，我从未用过它。我用 GPT 和 Pro，所以每次，就像你说的，幻觉对我来说就像古代历史。我记得那是个大事。是的。

Right. I've noticed that some of the companies I've corresponded with or interacted with seem to use Claude a lot. I don't know if you've experimented with that, but I'm kind of curious why that certain model versus like, but in all fairness, I've never used it. I use GPT and the Pro, and so every time, like you said, the hallucinations are like ancient history for me. I remember that was a big thing. Yeah.

但现在它发展得这么快，更好了。但比如 Claude 和 GPT-5.5 Pro 之间有什么区别？对于某些事情，没有区别了，因为智力已经达到顶峰。对于常规诊断，不是非常复杂的病例，Claude 很棒。Claude 也非常好，比如 Opus 4.7 模型，最近的那个，用于分析数据集。它可以处理数百万条数据，分析它，并做得很好。我的偏好是 GPT-5.5 Pro，因为正如我提到的，它有额外的洞察力。对我来说，我需要那种额外的预测性洞察力。

But it's going so fast and better now. But what's the difference between, for example, Claude and GPT-5.5 Pro? For certain things, there is no more difference because the intelligence has peaked. For doing regular diagnosis, not very complex cases, Claude is great. Claude is also very good, like the Opus 4.7 model, the recent model, for analyzing data sets. It can take millions of data, analyze it, and do a great job. My preference is GPT-5.5 Pro because, as I mentioned, it has this extra insight. For me, I need that extra level of insight that's predictive.

那种直觉……

The intuition that...

那种直觉和真正深刻的理解。但如果我要诊断和治疗一种肺癌亚型，我很确定 Gemini 3.1 Pro 或 Claude 4.7 都能做得很好。我认为有些人更喜欢 Claude 的原因是它可能更令人愉快，更人性化。我认为 GPT 模型开始接近了，但 Claude 仍有某种人们喜欢与之互动的东西。这真的是……

The intuition and really kind of a deep understanding. But if I'm going to diagnose and treat a subtype of lung cancer, I am pretty sure Gemini 3.1 Pro or Claude 4.7, they all do a pretty good job. I think the reason some people prefer Claude is that it's maybe more pleasant to interact with, kind of more humanlike. I think GPT models are starting to get there, but still there's something about Claude that people enjoy interacting with. It's really a matter of...

更有人情味，还是……

Personable or...

我认为它过去更有人情味。所以这真的不重要。我认为它们都是顶级水平，除非你在做研究或非常复杂的问题。例如，我和一位同事用 GPT-5 Pro 模型做了一个皮肤病的测试。它能够仅凭照片和症状诊断出一种我朋友无法诊断的皮肤病。其他模型做不到。

I think it used to be more personable. And so it doesn't really matter. I think they're all super top levels unless you're doing research or a very complex problem. For example, we did a test with a colleague of mine on skin disease with the GPT-5 Pro model. It was able to diagnose a skin disease that my friend couldn't really diagnose, just based on a photo and symptoms. The other models couldn't do that.

它们也能处理 90% 的病例，但总有那么一两个特别难的病例，结果难以预料。GPT Pro 模型能够跨越这个门槛。所以这类病例你确实需要非常高的水平，就像你不会无缘无故去找哈佛教授一样，对吧？所以必须是其他医生诊断不出的非常特殊的疾病之类的。这就是我的看法。

They could do 90% of the cases as well, but there's that one extra case or two extra cases that is really difficult that could go anywhere. The GPT Pro model was able to cross that threshold. So those kinds of cases you really need the very high level, like you know, you don't go to a professor at Harvard for any reason, right? So it has to be very specialized disease that other doctors couldn't diagnose or something like that. So that's how I view it.

昨天刚有来自 OpenAI 关于 GPT Rosalyn 的消息，我知道你不能多谈，但从公开信息来看，它似乎将用于药物发现。我想知道你对衰老研究、生物学、医学的未来怎么看。我们会使用这些更专业类型的 AI 模型，还是你认为像 GPT-5.5 Pro 这样的通才模型以及后续版本，才是解锁医学突破和生物学突破的关键？

We just had news yesterday from OpenAI about GPT Rosalyn, which I know you can't talk about much, but from what was publicly available, it seems as though it's going to be used in drug discovery. I'm wondering what you think in terms of the future of aging research, biology, medicine. Are we going to be using these more specialized types of AI models, or do you think more of a generalist like GPT-5.5 Pro and the subsequent ones that come out after it are going to be the key to unlocking medicine breakthroughs and biology breakthroughs?

我的偏好始终是通用模型，因为回到 AGI 的概念，如果一个模型有……当然，有些模型只针对特定数据训练，比如心电图或 RNA 测序之类的，它们会非常擅长，就像最好的国际象棋 AI 或围棋 AI 模型，但它们会错过那种联系，因为我认为医学在某种程度上是一种整体性的艺术。如果你只是试图分析一组数据，专用模型可能非常有用。事实上，我举过心电图的例子。大多数通用模型在某种程度上并不擅长，不知为何，心电图图像在诊断其显示内容方面表现不佳。而专用模型非常好，因为它们训练时使用的心电图数据集比通用模型多数百万个。但我认为，如果我们能训练通用模型，或对其进行微调或过度训练，我不知道怎么说，那么它们将始终优于专用模型。因为它们不仅了解心电图，还了解放射学、RNA、蛋白质。所以它们可以利用这些信息，在所有其他生物学的背景下分析心电图、心电活动、你的心跳。那将是极其丰富的知识。但我认为，你可以通过利用这些大模型并进行某种程度的驾驭或微调来实现专业化，因为有很多数据集是不公开的。所以这些模型无法访问这些数据。你可能有一些数据因监管原因或其他原因被锁定。所以你可以拿一个大模型。事实上，你可能甚至不需要闭源模型。你甚至可以使用一些现在变得非常好的开源模型。如果你用这些数据训练它们，它们同时拥有通用知识，结合这些，它们可能会做得更好。

My preference would always be the generalized models, because again, going back to AGI, if a model has... of course, there are some utilities of models that are only trained on, I don't know, like the EKGs or RNA sequencing or something like that, and they'll be very, very good at that, like the best chess player AI model or the best Go player AI models, but they will miss that connection, because again, I view medicine as a kind of a holistic art in a way. If you are just trying to analyze one set of data, the specialized models could be very, very useful. In fact, I gave the example of EKGs. Most generalized models were not terribly great at, for some reason, the EKG images were not very good at diagnosing what it was showing. And specialized models were very good because they were trained with millions more EKG data sets than the generalized model was. But I think if we can train the generalized model or fine-tune it or overtrain it, I don't know how to say it, then they will be better than specialized models all the time. Because not only they know all about EKGs, but they know about radiology, they know about RNA, they know about proteins. So they can take that information and analyze the EKG, the electrocardiogram, your heart beats, in the context of all the other biology. So that will be very enriching knowledge. But I think you can specialize in the sense that you can take these big models and you can sort of harness them or fine-tune them, because there's a lot of data sets that's not public. So these models don't have access to that. You might have some data locked in certain places because of regulatory reasons or whatever. So you can take a big model. In fact, you may not even need the closed models. You can even take some of the open-source models which are now getting very good. If you train them on that, and they also have the generalized knowledge, combined with that, they'll probably do better.

我想谈谈治疗疾病、治愈疾病，然后是逆转衰老。让我们从治愈疾病、治疗疾病开始，因为显然我们确实会死于与年龄相关的疾病，心血管疾病是大多数发达国家的主要杀手。我们有癌症。那是个大问题。而且癌症是一种非常可怕的疾病。任何正在收听的人，如果自己得过癌症或认识得过癌症的人，都知道这是真的。但我也认为，很多人把癌症看作一种疾病。非科学家、非医生，他们有点把癌症看作一种疾病，对吧？正如你我都知道的，它绝对不是一种疾病。它是数百种疾病。我很好奇，首先，我们仍然没有治愈癌症的方法。我的意思是，我们取得了很大进展，对吧？不同的癌症可以比其他癌症得到更好的治疗，但你能谈谈为什么找到癌症的治疗方法如此困难吗？

So I want to talk about there's treating disease, there's curing disease, and then there's reversing aging. So let's start with curing disease, treating diseases, curing diseases, because obviously we do die of age-related diseases, cardiovascular disease being the number one killer in most developed countries. We have cancer. That's a really big one. And with cancer, it's just such an awful disease to have. And anyone that's listening that has either had cancer or knows someone that has had it knows this is true. But also, I think, cancer, a lot of people think about it as one disease. Non-scientists, non-physicians, they kind of think about cancer as this one disease, right? As you and I both know, it is definitely not one disease. It's hundreds of diseases. I'm curious, first of all, we still don't have a cure for cancer. I mean, we've made a lot of progress, right? And different cancers can be treated better than others, but can you talk a little bit about why it's been so hard to find a treatment for cancer?

是的，我认为需要澄清的重要一点是，癌症不是一种疾病。它可能是 100 种不同的疾病，可能还有数百种亚疾病或亚型。事实上，某些癌症是 100% 可治愈的，或者 95% 可治愈。比如一些儿童白血病，几十年前还是完全致命的，现在 90% 或接近 100% 可治愈。如果你及早发现某些癌症，几乎又是 100% 的治愈率。所以因为这是一组非常不同的疾病，胰腺癌和肺癌或乳腺癌非常不同。有些癌症非常缓慢，比如如果你在 80 岁时得了某些类型的癌症，医生甚至懒得治疗，因为到那时……除非我们先治愈衰老，因为在你死于衰老之前，那个癌症不会杀死你。衰老会先杀死你。或者某些年龄的前列腺癌。所以这就是为什么我们必须真正理解这是一个非常复杂的生物学。但更重要的是，癌症之所以如此具有挑战性，是因为癌细胞是我们的一部分，对吧？所以如果你感染了细菌或病毒，它可能会杀死你，对吧？它们极其危险，但我们能够将它们识别为敌人、威胁，你的免疫系统，我们可以反击，虽然不总是成功，但大多数时候非常成功。我们还可以非常特异性地靶向它们，比如我们有抗生素只作用于细菌。它不会碰你的正常细胞，因为它只是外来生物。但癌症不是这样。所以如果我试图用某种东西阻止癌症，我也在阻止一些正常的细胞，对吧？这就是为什么人们会掉头发，免疫系统大大减弱，因为免疫系统必须分裂。你的毛囊细胞必须分裂。所以你阻止了它们，因为癌细胞也在分裂，而化疗的副作用有时比癌症本身更糟糕，比如成千上万的人因此死亡。所以癌症领域的革命最近是因为我们所谓的免疫疗法。

Yeah, I think the important thing to clarify is that cancer is not one disease. It's probably 100 different diseases that have probably hundreds of sub-diseases or subtypes if you like. In fact, certain cancers are 100% curable or 95% curable. Like some of the childhood leukemias, which were completely fatal a couple of decades ago, are now 90% or close to 100% curable. If you catch certain cancers early enough, again, 100% cure rates almost. So because it's a very different set of diseases, the cancer of pancreas is very different than cancer of lung or breast cancer. There are some cancers that are so slow, like if you get certain types of cancers at age 80, doctors don't even bother to treat it because by the time that... unless we cure aging first, because by the time you die of aging, that cancer is not going to kill you. Aging is going to kill you first. Or there's certain prostate cancers at certain age. So that's why we have to really understand that this is a very complex biology. But more importantly, why cancer is such a challenge is that the cancer cells are part of us, right? So if you're infected with a bacteria or a virus, it can kill you, right? They're extremely dangerous, but we are able to recognize them as an enemy, as a threat, your immune system, and we can fight back, not always successfully, but most of the time very successfully. And we can also target them very specifically, like we have an antibiotic that will only act on the bacteria. It's not going to touch your normal cells because it's only a foreign organism. But cancer is not like that. So if I try to stop cancer with something, I'm also stopping some other cells that are normal, right? That's why people lose their hair, their immune system is greatly weakened, because the immune system has to divide. Your hair cells have to divide. So you block them because the cancer cell is also dividing, and your side effects of chemotherapy sometimes worse than having the cancer, like hundreds of thousands of people die because of that. So the revolution in cancer was recently because of what we call immunotherapy.

问题是我们能否让免疫系统把癌症识别为外来威胁，就像对待恐怖分子一样，对吧？恐怖分子你无法分辨他是不是敌人。他们看起来和你一样，你知道，他们混进来然后制造……所以免疫系统就是这么看的，它认为乳腺癌细胞和正常乳腺细胞没有太大区别，比如上皮细胞之类的。所以它不知道该怎么做。如果我们能教会免疫系统，或者如果我们能解除它的一些调节刹车，让它识别并攻击癌细胞，那可能会产生巨大的效果。这就是假设，而且确实奏效了。所以癌症免疫疗法，我认为现在比化疗和放疗加起来还要强大。我是说，它们仍然有作用。当然，另一件事是如何让治疗非常精准。如果我给一种不精准的化疗，那就像试图敲病人的头，希望癌症在病人死之前先死。但如果我知道某个特定癌症中发生的单一突变，比如 EGF 受体上的突变，我就可以开发一种小分子，只有在 EGF 受体上有那个突变时才会起作用，所以它不会触及任何其他地方。它只会针对那个。事实上，人们称它们为智能药物，它们非常有效。所以如果你有那个特定突变，在 1% 的肺癌患者中，你用那种药治疗，几乎能获得 100% 的治愈率。但同样，我们可以做得更好。例如，免疫系统可以被工程化——这是我们在实验室做的工作——来识别，就像字面意义上的工程化：我们把细胞取出来，训练它们，把基因放入它们体内，说，好吧，如果这个基因结合到某个细胞上，就认为那是威胁并杀死它。这被称为 CAR-T 疗法，它们会去寻找带有那个标记的癌细胞并杀死它们。这样做的优点是癌症没有太多办法逃脱。它可以试图抑制免疫系统，但除此之外，即使它突变，免疫系统仍然会识别它，找到那些藏起来的少数细胞并摧毁它们。这显示了令人难以置信的结果。所以 mRNA 疫苗，我认为将是革命性的，就是基于这个原理，对吧？这真正实现了癌症的个性化。所以我有乳腺癌，但我的乳腺癌有某些类型的突变，其他患者没有。所以即使免疫系统能识别 X 患者，它也不会识别我的，因为癌症有不同的突变。如果我取那些突变，合成所谓的 RNA，然后作为疫苗回输，训练我的免疫系统，告诉免疫系统，看，如果你在这些基因中看到这些突变，那就是敌人。去摧毁它。这就是 mRNA 疫苗。这变得非常强大，因为现在你是在引导你的免疫系统针对你体内的内部威胁。假设癌症需要不同的突变，你可以创建另一个 mRNA 疫苗，然后训练免疫系统也针对那个。所以我认为这些是困难的部分，但我们看到了隧道尽头的曙光。癌症将 100% 可治愈，可能不到十年。

The question was why can we make the immune system recognize cancer as foreign threats, like they're kind of like terrorists, right? So a terrorist, you will not know if that's an enemy or not. They look like you, you know, they just come in and then they create... So the immune system is seeing it that way, that it thinks that the breast cancer cell is not so different than a normal breast cell, you know, like epithelial cell, whatever. And so it doesn't know what to do. If we could teach the immune system, or if we could remove some of the breaks that it has in regulation, and let it recognize and attack the cancer cells, then that could have a tremendous effect. That was the hypothesis, and it actually worked. So cancer immunotherapy, I think, is more powerful now than chemotherapy and radiotherapy put together. I mean, they still have a role. And of course, the other thing is how can we make the treatments very specific? So if I give a chemotherapy that's not specific, it's like trying to hit the patient on the head and hope that the cancer will die before the patient dies. But if I know this single mutation that's happening on, you know, whatever EGF receptor in certain cancers, I can develop a small molecule which will only act if there's that mutation on the EGF receptor, and so it's not going to touch anywhere else. It's only going to target that. And in fact, people call them smart drugs, and they're extremely effective. So if you have that particular mutation in 1% of lung cancer patients, you get treated with that drug, you get almost 100% cure rate. But again, we can make this even much better. For example, the immune system can be engineered—something we work on in the lab—to recognize, like literally engineer: we take the cells out, we train them, we put genes into them, saying, okay, if this gene binds to a cell, assume that that's a threat and kill that. And so it's called CAR-T therapy, and they will go and seek out whatever cancer cells have that marker and kill them. The advantage of that is that cancer doesn't have much way to escape that. It can try to suppress the immune system, but other than that, even if it mutates, the immune system will still recognize it and find those few cells that are hiding somewhere and destroy them. And that's showing incredible results. So the mRNA vaccines, which I think is going to be revolutionary, is on that basis, right? And that really personalized the cancer. So I have breast cancer, but my breast cancer has certain types of mutations that other patients don't have. So even if the immune system can recognize X patient, it won't recognize mine because the cancer has different mutations. If I take those mutations and synthesize what's called RNA, and then give it back as a vaccine and train my immune system, and tell the immune system, look, if you see these mutations in these genes, that's an enemy. Go destroy that. That's mRNA vaccine. And that becomes extraordinarily powerful because now you are directing your immune system to an internal threat just in you. And let's say the cancer required different mutations, you can create another mRNA vaccine and then train the immune system to that as well. So I think those are the difficult parts, but we see the light at the end of the tunnel. Cancer is going to be 100% curable, probably less than a decade.

AI 将如何实现这一点？

How is AI going to make that happen?

是的。事实上，它已经在实现这一点了。你可能听说过澳大利亚的这个故事。这位计算机科学家用 ChatGPT 和其他一些 AI 模型为他的狗开发了一种 mRNA 疫苗。他的狗得了黑色素瘤，我想，他做了基因测序。他拿到序列，交给 AI 模型，AI 模型设计了精确的 mRNA 分子，用于训练狗的免疫系统。然后合成，我想在 3 个月内就能应用了。如果没有法规限制，可能更快，肿瘤开始消退，狗在预期死亡时间后还活着。所以我的意思是，这是一个非常明显和简单的版本。但因为癌症类型有数百种，你可以想象，我们可能对一种肺癌就有上百种不同的治疗方法。有的是 mRNA，有的是针对那个的小分子。所以为了能够按需快速开发这些，我们将需要 AI。所以 AI 将模拟每一个可能的突变，我们将筛选数百万种化合物。我们将达到每月推出数百种新药的程度，也许。我们将有上千种针对乳腺癌患者的药物。但如果你有这些突变，而且是第四期，那么你服用这种组合。如果是那种情况，你采用这个方案。这就是 AI 将如何帮助。当然，如果你有数字孪生，那将加速这一过程。

Yeah. So, in fact, it's already making that happen. You probably heard of this story from Australia. This computer scientist had ChatGPT and some other AI models to develop an mRNA vaccine for his dog. His dog had, I think, a melanoma, and he got it sequenced. He took the sequence and gave it to an AI model, and the AI model designed the precise mRNA molecule that the dog's immune system needs to be trained with. Got it synthesized, and I think it was able to apply it in 3 months. Probably could have been shorter if there weren't regulations, and the tumor started to regress, and the dog was alive when it was supposed to die. So I mean, that's a very obvious and simple version. But because there are hundreds of different cancer types, you can imagine that we'll have maybe a hundred different treatments for just a type of lung cancer. Someone will be mRNA, someone will be small molecule targeting that. So to be able to develop those on demand and very rapidly, we're going to need AI. So the AI is going to model every possible mutation and we'll screen millions and millions of compounds. And so we'll get to a point where we'll have hundreds of new drugs coming out every month, maybe. And we'll have this thousand drugs for breast cancer patients. But if you have these mutations and it's stage four, then you take this combination. And if it's that, you take this protocol. And that's how AI is going to help. Of course, if you get to digital twin, that will accelerate that.

对，这就是下一个问题。假设我们有真正的个性化医疗和个性化癌症治疗，但你也需要了解副作用，比如我服用这种 mRNA 疫苗，我的免疫系统会不会失控，开始让心脏发炎，导致心肌炎，对吧？你如何看待这个数字孪生，它现在拥有基因组信息、你所有的蛋白质、代谢物以及一切实时数据，然后它还能模拟如果我们给这个人这种特定的 mRNA 癌症疫苗或这种小分子会发生什么？

Right, and that's the next question. So let's say we have the true personalized medicine and personalized cancer treatment, but you also need to know about side effects, like am I going to take this mRNA vaccine and my immune system is going to go crazy and start to inflame my heart and give me myocarditis, right? How do you also see this digital twin, which now has genomic information, all your proteins, metabolites, and everything in real-time data, then it can also simulate what's going to happen if we give this specific mRNA cancer vaccine or this small molecule to this person?

绝对。我的意思是，你提到了心肌炎，顺便说一句，这在新冠疫情期间发生过，这就是为什么有很多反疫苗情绪。但人们没有意识到新冠病毒本身也会引起心肌炎。是的，接种疫苗的人，年轻人，以五千分之一到一万分之一的比例出现心肌炎。它大多不是致命的。但应该问的问题是，为什么一万分之一的人得了心肌炎，而其他人没有？或者事实上，我们可以反过来问这个问题。我们给所有人接种了疫苗，但如果你是年轻人，你死于新冠的概率，比如说，千分之一或万分之一。所以 999 人不必接种疫苗。但为了救那一个人，我们必须给那个疫苗。或者我再举一个更普遍的例子：我们给任何胆固醇高的人服用他汀类药物。所以我认为大约五分之一或十分之一的人真正从中受益。高胆固醇并不自动意味着你会得动脉粥样硬化。

Absolutely. I mean, so you mentioned myocarditis, which by the way happened during the COVID pandemic, and that's why there was a lot of anti-vaccine sentiment. But people didn't appreciate that the COVID virus itself caused myocarditis. Yes, the vaccinated people, young people at one in 5,000 to one in 10,000 rate got myocarditis. It wasn't mostly fatal. But the question should be asked like, why is it that one out of 10,000 got myocarditis and the other ones didn't? Or in fact, we can reverse that question. We vaccinated everybody, but if you were a young person, your chance of dying from COVID was, let's say, one in 1,000 or one in 10,000. So 999 people didn't have to be vaccinated. But to save that one person, we have to give that vaccine. Or I'll give another more general example: we give statins to anyone who has high cholesterol. So I think like one out of five or one out of ten people truly benefit from that. High cholesterol doesn't automatically mean you're going to get atherosclerosis.

你需要有炎症等等。但因为缺乏数据，我们无法预测。这不是个性化的。数百万人服用他汀类药物，只为拯救几千人。是的，这是好事，因为你不知道。所以 AI 将能做到这一点。我们会告诉你，好的，不仅会为你量身定制药物，还会说，好的，你不需要吃这种药，你应该吃这种，或者你甚至根本不需要任何治疗。比如，你得了传染病或别的什么，或者某些癌症，这就够了。我们为什么要额外化疗加免疫治疗加放疗？因为不确定单一疗法是否足够。所以这将大幅减少副作用问题。当然，你可能仍会有一些副作用，但那是可控的副作用。例如，它不会要你的命。

You need to have inflammation this and that. But because we don't have the data, we cannot predict that. It's not personalized. Millions of people take statins to save a few thousand people. Yes, that's a good thing because you don't know. So AI will be able to do that. So we'll tell you, okay, not only will we create the drug just for you, but also we'll say, okay, you don't have to take this medicine, you should take this, or maybe you don't even need any treatment at all. Like, you have an infectious disease or whatever, or certain cancers, this will be enough. Like, we give extra chemotherapy plus immunotherapy plus radiotherapy. Why are we doing that? Because we're not sure if one is going to be enough or not. And so that will dramatically reduce the side effect issue. You might still have some side effects, of course, but it will be manageable side effects. It's not going to kill you, for example.

那用 AI 根据你的蛋白质、代谢物、生物标志物，也许还有基因，提前十年或更早预测癌症呢？你怎么看？我们谈的是个性化癌症治疗，但能不能在癌症发生前几年就预防它？

What about using AI to predict cancer a decade or years before it forms based on your proteins and metabolites and your biomarkers and maybe perhaps your genetics too, right? Like how do you see that? We're talking about personalized cancer treatment, but what about being able to prevent cancer before it happens, you know, years before it happens?

是的，又是好问题，因为我觉得这非常重要，但人们很少思考。我们说的是医疗保健，其实我们没有医疗保健，只有疾病护理。我们从不照顾健康的人。你不会去看医生问“我有多健康？”或者去看医生问“你能检查我的免疫系统吗？它健康吗？我会生病吗？我会得癌症吗？”他们无法回答这个问题。只有你生病了，他们才会治疗问题。所以预防医学将变得绝对关键。我认为不是所有但大多数疾病是可以预防的。有些只是运气不好。无论你做什么都会发生。即使你过着完美的生活，你可能仍会得某些疾病，但很多是因为基因等等。很多是可以预防的，我认为 AI 在这方面会很出色，因为它已经能做到。有一项来自英国生物银行的研究。英国有惊人的生物银行，有 50 万人，大量数据集，令人难以置信的数据集。这项研究实际上是在一年多前做的，用的是一两年前的模型。他们用了大量数据，能够预测大约一千种疾病在发生之前。当然，这有点追溯性。他们根据多年前收集的数据知道人们会得什么病。但 AI 告诉你，好的，这个病人会得这种病，但不是病人，是正常健康的人，他们会得这个那个。所以对我来说，这很惊人，而且会越来越好，因为总有迹象。癌症不是几天内形成的。它需要数年。可能我们大多数人都有一些癌细胞，大部分被免疫系统控制，它缓慢生长，必须再有突变，再有突变，但可能在某个地方有迹象，无论是在新陈代谢还是别的什么。AI 即使不是 100%，也能说，好的，看，我认为如果你继续这种生活方式，你得这种病的几率现在是 85% 或什么的。比如，我戴着血糖监测仪。我不是糖尿病患者，但我想每分钟或每五分钟看到我的血糖水平连续变化，或者如果我吃了东西，它会飙升吗？会下降吗？因为我想预防胰岛素抵抗，这是可能发生的最糟糕的事情之一。如果我不这样做，我直到得糖尿病才会知道。如果我的血糖持续飙升，胰岛素工作过度，这可能持续数年。顺便说一句，在某个时候它会崩溃。对有些人可能持续 50 年，没事。有些人可能五年，但那个数据集可能有预测价值，加上我的年龄、基因等等。所以，是的，我认为每个人都会有自己的 AI，我不知道怎么称呼它，健康教练之类的。但它会持续分析数据。希望收集数据会更容易，因为那是另一个问题。我们不收集数据。我们对血液中超过一千种代谢物一无所知。我们只看其中 10 或 20 种，而且只在生病时，甚至体检都不看。所以我们必须有持续的监测器，比如血糖监测仪。我想看到我的蛋白质变化、激素变化，以相当连续的方式。

Yeah, again, great question because I think this is so important that people don't think about very much. We say health care, you know, we don't have health care, we have sick care. So we never take care of healthy people. You don't go to a doctor to say, oh, how healthy am I? Or just go to a doctor and say, can you check my immune system, is it healthy? Am I going to get sick? Am I going to have cancer? They won't be able to answer that question. Only if you get sick, they will treat the problem. So preventative medicine is going to be absolutely critical. I think not all but most diseases can be prevented. Some are just bad luck. It happens no matter what you do. Even if you live the perfect life, you might still get certain diseases, but a lot of them are because of your genes and so on. A lot of them can be prevented, and I think AI is going to be amazing in that because it's already able to do it. There was a study from the UK Biobank. The UK has this amazing biobank with 500,000 people, lots of datasets, incredible datasets. And this was actually done, I think, more than a year ago with models that were a year or two years old. They took a lot of that data and they were able to predict about a thousand diseases before they happened. Of course, this was kind of retroactive. So they knew what people were going to get based on their data that was collected years before. But the AI was telling you, okay, this patient's going to have this disease, but not patient, normal healthy people, they're going to get this and that. So that, to me, was amazing, and that's going to get better and better because there are always signs. Cancer doesn't just develop in days. It takes years. Probably most of us might have some cancer cells, most of it controlled by immune system, and it slowly grows, it has to have another mutation, another mutation, but there's probably some signs of that somewhere, whether it's in your metabolism or whatever. And AI, even if it's 100%, will be able to say, okay, look, I think that if this is the lifestyle that you continue, your chances of getting this disease is now 85% or whatever. Like, I wear a glucose monitor. I'm not diabetic, but I want to see every minute or every five minutes what my sugar levels are in a continuum, or if I eat something, is it spiking? Is it coming down? Because I want to prevent insulin resistance, one of the worst things that can happen to you. If I don't do that, I won't know until I get diabetes. If my sugar is constantly spiking and insulin is just working too hard, that could continue for years. By the way, at some point it's going to break. For some people it might continue 50 years, nothing happens. Some it might be five years, but that dataset probably has that predictive value, plus my age, my genes, whatever. So, yeah, I think everyone's going to have their own AI, I don't know what to call it, health coach or something. But it will continuously analyze the data. And hopefully it will be much easier to collect data because that's another issue. We don't collect data. We know nothing about the more than a thousand metabolites in our bloodstream. We look at maybe 10 or 20 of them, only if we get sick, not even for a checkup. So we have to have a continuous monitor, like a glucose monitor. I want to see what my proteins are changing, hormones are changing, in a reasonably continuous manner.

说得好，我很高兴你提到英国生物银行的研究。我记得模型好像叫 Milton 之类的，是阿斯利康自己开发的。我记得看过这项研究，因为正如你所说，生物银行的数据集非常庞大，跨越了几十年。所以我想他们看了超过 200 种血浆蛋白。你说的是 10 种，我们说的是 200 种，对吧？

Such a good point and I'm so glad you brought up the UK Biobank study. I remember I think the model was like called Milton or something and it was AstraZeneca's own, developed it or something. And I remember looking at this study because like you mentioned, the Biobank data is huge data set and it just spans many decades. And so I think they looked at, you know, like over 200 plasma proteins. You're talking about 10, we're talking about 200, right?

哦，是的。

Oh yeah.

还有所有其他数据，对吧？他们能够预测，我想癌症和神经退行性疾病排在最前面，提前 10 年，他们能够观察这些人。所以 AI 根据所有这些生物特征数据进行了预测。然后他们查看并说：“哦，是的。那些人确实最终得了癌症和阿尔茨海默病。”而且非常准确。

And all the other data, right? And they were able to predict, and I think cancer and neurodegenerative disease were at the top, like 10 years before, and they were able to look at the people. So the AI predicted it based on all this biometric data. And then they looked and said, "Oh, yep. Those people actually did end up getting cancer and Alzheimer's disease." And it was very accurate.

是的。

Yes.

对我来说，令人兴奋的是你可以在它发生之前进行干预。你可以改变生活方式，改变饮食。我的意思是，这些事情很重要。确实重要。这很令人兴奋。因为那样你甚至不必走到药物那一步，也许你会，但如果你能做出这些改变，如果你知道，嘿，我正走在得癌症的轨道上。我有所有这些炎症。所有这些事情正在发生。如果我现在不改变，那么 10 年后，我可能会得癌症。

And to me, the exciting thing here is that you can intervene before it happens. You can make lifestyle changes, you can make dietary changes. I mean, these things matter. They do matter. And that is exciting. Because then you don't even have to get to the drug part, which maybe you will, but if you can make these changes, if you know, hey, I'm on this trajectory to get cancer. I have all this inflammation. I have all these things happening. If I don't make a change now, then in 10 years, I might have cancer.

对某人来说，这非常激励人心。所以这也非常令人兴奋，而 AI 的加入只会让它变得更好。

It's very motivating, you know, for someone. So it's very exciting as well, and having AI in there is just going to make it even better.

然后我想进入年龄逆转的话题，在进入之前，你知道，你确实是 AI 参与生物学领域的先驱。

And then I want to get into age reversal, and before we get to that, you know, you've really been a pioneer in this field of AI being involved in biology.

你知道，你跟我提过你的博客。我不知道，是 30 年前吗？

You know, you were talking to me about your blog. I don't know, was it 30 years ago?

生物奇点。是的。25 年前。

Bios singularity. Yeah. 25 years ago.

所以你有这个博客，生物奇点，做预测。你能谈谈它吗？

So you have this blog, Bios Singularity, predicting. Can you talk a little bit about it?

是的，当然。事实上，我在 90 年代初就对 AI 产生了兴趣，那是在我从医学院毕业之后。我十几岁的时候对计算机非常感兴趣。当时第一批计算机刚刚问世，我试着写代码，我热爱它。那真是太棒了。但我选择了医学，因为我觉得生物学要复杂得多，所以我应该先尝试弄清楚它。但很快我就意识到，我相信你作为科学家也有同感，生物学是如此复杂得令人难以置信。我说，好吧，我们根本没有机会弄清楚它，因为会有太多的数据集。所以那是我第一次对 AI 产生兴趣。当然，当时 AI 还非常原始。但快进一下，影响我的书之一是雷·库兹韦尔写的。我相信很多关注科技的人都知道他。他写了这本书《奇点临近》。他称之为奇点，在这个点上，计算或技术呈指数级增长，以至于你甚至无法预测第二天会发生什么，因为它有点像自我训练的 AI 模型。他有一些图表，绘制了 AI 的进步，说到 2029 年它将达到人脑水平，我们将实现 AGI。那真是令人难以置信，大多数人认为他是在胡说八道或者科幻小说。他们不相信。那怎么可能发生？但我非常兴奋。事实上，我有一本雷签名的书。受到那本书的启发，我创办了这个名为“生物奇点”的博客。我说，好吧，计算在呈指数级增长，但生物学在某种程度上也是一种计算。它基于信息，但只是复杂得多。所以它也应该呈指数级扩展。如果你绘制那条曲线，那就意味着，根据我 25 年前的计算，事实上我把它写在博客的“关于”页面上，到 2035 年左右我们应该能够治疗所有疾病，到 2045 年左右我们应该能够完全逆转衰老。事实上，到 2050 年代我们将达到我称之为“人类 2.0”的地步，因为那时我们对生物学有了完整的理解。然后我们可以真正地工程化它。我们可以创造新的生物有机体。我们可以改变我们的生物学、我们的基因组，重新编程它。

Yeah, sure. So in fact, I got interested in AI in the early 90s, after I graduated medical school. I was very interested in computers when I was a teenager. The first computers had come out at the time, and I was trying to code, and I loved it. It was just so wonderful. But I went into medicine because I figured biology is much more complex, so I should first try to figure that out. But then immediately I realized, and I'm sure you did too, as a scientist, that biology is so incredibly complex. I said, well, we don't have any chance of figuring this out, because there are going to be so many data sets. So that's when I first got interested in AI. Of course, at the time, AI was very primitive. But fast forward, one of the books that influenced me was from Ray Kurzweil. I'm sure a lot of people who follow technology know him. He wrote this book, The Singularity Is Near. He called it a point of singularity where computation or technology advances exponentially so much that you cannot even predict what will happen next day, because it's sort of like self-training AI models. And he had these figures where he would plot the advances of AI, saying by 2029 it will be at the human brain level, and we'll reach AGI. It was just unbelievable, and most people thought he was talking crap or science fiction. They didn't believe it. How could that happen? But I got very excited. In fact, I have a signed copy from Ray for the book. So being inspired from that, I started this blog called Bios Singularity. I said, okay, computation is going exponential, but biology is sort of a computation as well. It's based on information, but it's just much more complex. So it should also expand exponentially. And if you plot that curve, that means that, based on my calculations 25 years ago, in fact I wrote it on the about page of the blog, by year 2035 or so we should be able to treat all diseases, and by 2045 or so we should be able to completely reverse aging. In fact, by the 2050s we will get to a point I call Human 2.0, because at that point we have a complete understanding of biology. Then we can truly engineer it. We can create new biological organisms. We can change our biology, our genome, reprogram it.

重写我们的免疫系统。

Rewrite our immune system.

是的，完全正确。在许多可能的方式上，因为如果你仔细想想，它有点混乱。我们认为生物学是个奇迹，但它是一种糟糕的遗留工程。这不是糟糕的工程；而是遗留，因为生物系统找到了某种东西，它们无法摆脱它，它们不能从白板开始，所以它们在它之上构建。所以你得到一层又一层的调控。然后，当然，在我研究的免疫系统中，你会得到很多自身免疫性疾病。免疫系统杀死很多人，即使在流行病期间也是如此，或者它不能识别癌细胞。那么为什么我们不能设计一个免疫系统 2.0，一个白板，真正精心设计的免疫系统呢？我说到 2045-50 年我们会达到那个点。事实上，当时人们觉得这听起来很疯狂。但现在我觉得我太保守了。我们可能会达到。但关键是，我特别在“关于”页面写了我们将因为人工智能而做到这一点。我只是采用了雷绘制的图表。我说，好吧，到 2029 年 AI 将达到那个点。它将足够好，可以应用于生物学，这将使我们能够解决疾病，然后衰老。事实上，时机相当不错，甚至有点保守，我对此感觉很好。这就是为什么我全力投入 AI。哇，它正在发生。它真的在发生。

Yeah, exactly. In many possible ways, because it's kind of messed up if you think about it. Biology, we think, is a miracle, but it's a bad kind of legacy engineering. It's not bad engineering; it's legacy, because biological systems find something, they can't get rid of it, they can't start from a clean slate, so they build on top of it. So you get regulation over regulation over regulation. And then, of course, with the immune system that I study, you get lots of autoimmune diseases. The immune system kills a lot of people, even during pandemics and things like that, or it doesn't recognize cancer cells. So why shouldn't we be able to design an immune system 2.0, a clean slate, really greatly engineered immune system? And I said by 2045-50 we'll get to that point. And actually, again, at the time it sounded really crazy to people. But now I feel that I was too conservative. We'll probably get there. But the key point is that I wrote specifically in the about page that we will do this because of artificial intelligence. I was just taking the plot that Ray plotted. I said, okay, by 2029 AI is going to be at that point. It will be good enough to apply to biology, and that will allow us to solve diseases, and then aging. The fact that the timing was pretty good, again even a bit conservative, I feel great about it. That's why I'm all in on AI. Wow, it's happening. It's really happening.

所以衰老非常复杂，而且如你所知，它不是单一过程。我们有这 12 个生物学标志。我们现在有 12 个。基因组不稳定、线粒体功能障碍、细胞衰老，等等。我们知道器官以不同的速率衰老。它们以不同的速率达到峰值，以不同的速率衰老，一切都在以非常复杂的方式相互作用。你认为理解衰老过程和逆转它的瓶颈是什么？

So aging is very complex, and as you know, it's not one process. We've got these 12 hallmarks of biology. We now have 12. Genomic instability, mitochondrial dysfunction, cellular senescence, on and on. And we know organs are aging at different rates. They reach their peak at different rates, and they age at different rates, and everything is interacting in a very complex way. What do you see as the bottleneck for understanding the aging process and also reversing it?

我的意思是，与其说是瓶颈，不如说这是我们必须思考衰老的方式。生物学实际上是被编程来防止衰老的。对吧？所以它在某种程度上不像汽车，因为一旦你制造了一辆汽车，你必须不断把它送到修理厂或重新喷漆。生物学在内部做到这一点。如果不是这样，我们会立即衰老。有一种疾病叫做早衰症。这些孩子在 7 或 8 岁时就衰老了，他们变得像 80 或 90 岁，因为他们的一个基因发生了单点突变，因为他们失去了修复能力，无论是 DNA 修复，无论是清除旧细胞，还是清理组织，然后像干细胞一样再生，创造新细胞。所以这个程序有时持续几十年，否则我们无法生存。对于一些动物，一些生物体，只有几年。对于我们，大约是 50 到 100 年。对于一些鲸鱼，是几百年。所以同样的生物学，只是其中一个决定我可以养一条鲸鱼，或者不管怎样，有些动物活得更长，因为它们不被捕猎，或者它们可以更晚繁殖，等等。所以在生物系统中发生的事情是，这个程序不知何故崩溃了，你开始失去所谓的韧性。对吧？所以当你 30 或 40 岁时，你有韧性。你能比 70 或 80 岁的人承受更多的损伤，因为你的系统是这样的，即使你受伤或生病，你也能更容易恢复。

I mean, more so than the bottleneck, this is the way we have to think of aging. Biology actually is programmed to prevent aging. Right? So it's not like a car in a way, because once you make a car, you have to constantly bring it to a repair shop or repaint it. Biology does that internally. If it didn't, we would age immediately. There is a disease called progeria. These children get aged by the age of 7 or 8, they become like an 80 or 90 year old, because of a single point mutation in one of their genes, because they lose their ability to repair, whether it's DNA repair, whether it's getting rid of old cells, or cleaning up the tissues and then regenerating like stem cells creating new cells. So this program continues for sometimes decades, otherwise we wouldn't survive. For some animals, for some organisms, it's only a couple of years. For us, it's about maybe 50-100 years. For some whales, it's hundreds of years. So same biology, it's just that one of them decided that I can keep a whale, or whatever, some animals live longer because they're not getting hunted, or they can reproduce later, and so on. So what happens in the biological system is that somehow this program breaks down, and you start to lose what's called resilience. Right? So when you are age 30 or 40, you're resilient. You can tolerate much more damage than someone who's 70 or 80 years old, because your systems are such that even if you get wounded or get sick, you can recover easier.

但这种韧性会丧失，其原因是某种信息丢失，因为生物系统拥有特定信息：它知道某些基因何时该开启、何时该再生，比如你的皮肤。你知道为什么会长皱纹吗？因为你的细胞停止产生胶原蛋白，然后各种垃圾堆积在皮肤下，那些本该清理的巨噬细胞之类的没有尽职。信息或通讯出现了故障，而细胞内通讯正是衰老的标志之一。至于为什么会这样，十二大衰老标志给出了很多原因，比如肠道细菌。这些细菌产生各种代谢物，帮助免疫系统不断再生、保持最佳状态。如果这发生变化，你的新陈代谢、血糖水平、线粒体突变等都会改变。所有这些都会累积——表观遗传变化、DNA 突变——而生物体不知怎的就忘了自己该做什么。此外，因为损伤一旦发生，修复比预防更难，对吧？如果你持续保养你的车或房子，它出故障的可能性远低于你等到什么都不行了再处理。是的，你可以逆转，但需要付出更多努力。所以我认为，对于年轻一代，在未来十年左右，将不是逆转而是预防衰老过程——将这种韧性再维持几十年。我们会达到这样一个阶段：如果你二三十岁，你不会再衰老，因为会持续逆转。但已经衰老的人，比如你 80 岁或 90 岁，那我们就得逆转这个过程。那更困难，但我们能做到。绝对能。不过，这需要大量的工程方法，因为你需要修复大多数衰老标志。如果你年轻，你就预防这些标志发生；你维持信息的时间要长得多。这两种情况都会发生。我们只需要弄清楚丢失的是什么信息，然后把它放回去。

But that resilience is lost, and the reason why it's low is that there's a sort of information loss because the biological system has certain information: it knows when certain genes should be turned on, when things should be regenerated, like your skin. You know why you get wrinkles? Because your cells stop making collagen, and then all kinds of crap accumulates under your skin, and the macrophages or whatever were supposed to clean there don't do their job. There's a breakdown in information or communication, and intracellular communication is one of the hallmarks of aging. And then why that happens is the 12 hallmarks—many reasons, for example, the bacteria in your gut. These bacteria produce all kinds of metabolites that help your immune system constantly regenerate and keep it in optimal shape. If that changes, then your metabolism changes, your glucose levels, your mitochondrial mutations, and so on. All these things accumulate—epigenetic changes, DNA mutations—and somehow the biology forgets what it's supposed to do. Also, because when damage happens, it's harder to fix than to prevent, right? If you continuously take care of your car or your house, the likelihood of breakdown is much less than if you wait until nothing works. Yes, you can reverse it, but it's going to take a lot more effort. So I think what will happen is that for younger individuals in the next decade or so, it's not going to be reversal but prevention of the aging process—maintaining that resilience for decades more. We'll come to a point where if you're 20 or 30 years old, you won't age anymore because it's going to be constant reversal. But people who have already aged—say you're 80 or 90—then we'll have to reverse that process. That's more difficult, but we'll be able to do it. Definitely. However, it will require lots of engineering approaches because you need to fix most of those hallmarks. If you're younger, you prevent those hallmarks from happening; you maintain the information much longer. Both of those will happen. We just need to figure out what information is being lost and put it back.

你这么认为吗？我们先谈谈预防衰老，如果你是年轻人，因为预防总是更容易。如果有一个 20 或 30 岁的人，你认为方法会是找到——首先，我们是否知道所有的修复过程？我们有已知的部分，对吧？

Do you think so? Let's first talk about preventing aging if you're a younger person, because it's easier to prevent. If you have a person who's 20 or 30 years old, do you think the approach would be finding—first of all, do we even know all the repair processes? We have what we know, right?

但我们还有很多要发现的。

But we still have a lot to discover.

我们可能还有很多要发现的。那么你认为会不会有一个发现，让我们弄清楚——我们知道自噬、干细胞耗竭、应激反应基因、抗氧化剂、DNA 修复、线粒体修复本身，对吧？我们是要增强或调校这些机制，让它们持续以最佳状态运转吗？还是说我们又会回到这个信息问题——为什么这些会衰退？我们是要去研究它的信息层面，也许是表观遗传学？而且会更针对那些基因，还是会有更多这种——我们会谈到细胞重编程和部分重编程——但我很好奇你认为 AI 会如何参与这个过程。我想我们不知道，这是问题的一部分，但然后我们必须弄清楚如何把这些疗法给到人们，对吧？这是方程式的另一部分。

We probably have a lot to discover. So do you think there's going to be a discovery where we figure out—we know things like autophagy, stem cell depletion, stress response genes, antioxidants, DNA repair, mitochondrial repair itself, right? Are we going to be enhancing or tuning these up so they keep working at their prime continually? Or do you think we're going to have again this information—why are those things going down? Are we going to go to the information of it, the epigenetics perhaps? And is it going to be more targeted towards those genes, or are we going to have more of this—we'll get into cellular reprogramming and partial reprogramming—but I'm curious how you see AI coming into that process. I guess we don't know, that's part of the problem, but then we have to figure out how to give these treatments to people, right? That's another part of the equation.

你提到的那些生活方式改变——它们很有帮助，但只能减缓衰老过程。我不认为有什么能逆转这个过程。可能会有某种局部逆转，暂时性的，但仍然是希望事情不会更快变糟。比如，有些人能活到 100 岁，有些人只能活到 60 岁，对吧？所以长寿者身上有某种优势。事实上，有些超级百岁老人能活到 110 岁——非常少——但我认为这主要是基因。他们的生活方式可能有一点帮助，但他们的生物学特性能够更长时间地维持那种信息。所以我们必须触及核心：是什么在破坏信息、导致信息丢失？当然，你必须关注基因组，因为那是蓝图。但不仅仅是基因组，还有之后影响你的东西——你的微生物组、你的代谢物，这些东西如何变化，是加速还是逆转。而且这也必须是一种工程方法。皮肤衰老和免疫衰老、大脑衰老是非常不同的问题。你的皮肤细胞不断更新，所以你只需要让编程好的干细胞进入那里，清理环境，清除衰老细胞，让它再生，产生胶原蛋白等等。但大脑不是这样——你不想再生你的神经元，因为你会失去你的身份。所以它们必须以不同的方式处理。我认为，对于年轻人群，重新设计某些生物学——听起来很激进，但会更可靠。如果我们能通过基因工程改变基因组，添加或改变某些基因，使 DNA 损伤被更长时间地检查呢？有些动物有更好的 DNA 损伤蛋白，它们进化出了这种能力。大象很少得癌症，因为它们有多个 P53 基因拷贝，P53 就像基因组的守护者——它防止突变和癌症。所以大象很少得癌症。裸鼹鼠，你可能很了解——它们就像老鼠一样。

So the ones you mentioned about lifestyle changes—they help a lot, but they only slow down the aging process. I don't think there's anything that reverses that process. There might be some local reversal for a temporary period of time, maybe, but it's still kind of hoping that things won't go bad a little bit longer. For example, some people live to 100, others only to 60, right? So there's something good about those who live long. In fact, there are supercentenarians who make it to 110 years old—very few people—but I think it's mostly genetics. Their lifestyle might have helped a little bit, but something about their biology is able to maintain that information much longer. So we have to get to the core: what are the things that are disrupting that information loss? And of course, you have to focus on the genome because that's the blueprint. But it's not just that; it's what affects you afterwards—your microbiome, your metabolites, how those things are changing, whether accelerating or reversing. And it has to be an engineering approach as well. Skin aging is a very different problem than immune aging or brain aging. Your skin cells are constantly renewing, so all you have to do is have programmed stem cells go in there, clean the environment, remove senescent cells, get it regenerated, produce collagen, and so on. But the brain is not like that—you don't want to regenerate your neurons because you'd lose your identity. So they have to be dealt with differently. Some of it, I think, for the younger population, seems like redesigning certain biology—sounds radical, but it would be more foolproof. What if we could change the genome through genetic engineering, adding or changing certain genes so that DNA damage is checked much longer? There are certain animals with better DNA damage proteins; they evolved to do that. Elephants rarely get cancer because they have multiple copies of the P53 gene, which is like the guardian of the genome—it prevents mutations and cancer. So elephants rarely get cancer. Naked mole rats, you probably know that very well—they're like rats.

它们生活在地下，但普通老鼠只能活几年，而它们能活 30 到 40 年。结果发现，它们在某个免疫基因上发生了突变，这个基因叫 seag gas，也参与免疫优化和 DNA 修复。就像你知道的，一两个基因就能带来巨大差异。那么我们能否改造人类，阻止这种信息退化？对于那些已经受损的人，我们就得考虑修复、逆转，然后维持。那会更有挑战性，但我们也会做到的。

They live underground but normal rats live a couple of years and these guys live 30-40 years. So it turns out they have some mutation in some immune gene called seag gas that's also involved in immune optimization and DNA repair. Just like you know, one or two genes make a huge difference. So can we engineer humans to block that degradation of information? For those who have already had the damage, then we're going to have to think about repairing that, reversing it, and then maintaining it. That's going to be a bit more challenging, but we'll get to that too.

你怎么看待基因疗法？显然有基因编辑、基因疗法，而现在我们只知道我们知道的，对吧？就像我们知道的这些长寿基因，但你认为 AI 能帮助我们分析人类基因组吗？我不知道它还需要什么其他数据集，但我们会给它一切，帮助我们发现这些基因之间的相互作用，以及所有这些组合。你认为这会发生吗？我们真的会发现这个方程里比我们原先知道的要多得多。

What do you think about gene therapy? There's obviously gene editing, gene therapy, and right now we only know what we know, right? Again, like with these longevity genes we know about, but do you think that AI is going to be able to help us analyze the human genome? And I don't know what other data sets it will need but we'll give it everything and help us figure out well actually there's interaction of these genes together and when you know like all these combinations. Is that something that you think is going to happen? We'll actually figure out there's a lot more to this equation than we originally knew.

是的。这是关键问题，因为我们知道基因组里所有的基因。我们已经完全解码了，而且我们基本知道它们的功能，大部分都知道，即使你不知道每一个与衰老相关的基因，我们也知道很多。问题是，不同的基因，首先可以产生不同的蛋白质，你知道，有各种剪接等等。但即使在不同的背景下，我们也怀疑，同样的蛋白质可以杀死细胞或导致存活。比如在免疫系统中，我们有这些受体，叫 TNF 受体之类的，它们可以根据细胞的背景发出存活信号或死亡信号、自杀信号。所以非常关键的是，正如你指出的，这些基因和蛋白质以网络方式、拓扑网络的方式，它们做什么？比如如果我干扰，像这些叫 Yamanaka 因子的东西，你可以从正常细胞生成干细胞，对吧？就像完全再生，但问题是它们也可能导致癌症，因为它们只需要在特定时间活跃。如果它们一直活跃，就会导致畸胎瘤之类的。所以那部分太复杂了，我们绝对需要 AI 来为我们模拟。如果我在某个年龄，在所有其他事物的背景下有这个基因，加上这些表观遗传程序以及所有代谢物等等，因为这些不断向细胞发出信号，让蛋白质做某些事情等等。如果我干扰那个特定基因会发生什么，或者我如何改进它？因为你还必须考虑其他基因组。你的基因疗法可能和别人非常不同，因为你可能有一些很好的基因具有协同作用，但另一个人可能没有这么好的基因，所以即使你试图改进，那实际上会起作用，或者不会帮助。所以这只是复杂性的问题。有太多信息，AI 不仅要整合这些，还要有几乎时间上的模型模拟。这是非常重要的一点，因为现在的模型是静态的。它们有很好的理解，但它们不知道如果一个细胞在 2 分钟前靠近肿瘤会发生什么，旁边的细胞，那个背景有什么影响。这是一个行为问题。这和机器人技术是同样的问题，对吧？所以，物理智能或生物智能，一旦这些模型用大量数据进化，我认为我们将能够模拟这个，AI 将能够决定你应该接受哪种基因疗法。所以，你需要一个新的免疫系统副本，但让我为你设计。

Yeah. That's the critical problem because we know what all the genes are in the genome. We have it decoded completely, and then we pretty much know their functions, most of them, even if you don't know every single gene involved in aging, we know a lot of them. The problem is that different genes, first of all, can create different proteins, you know, there's all that splicing that happens and so on. But even we doubt that in a different context, so the same protein can kill a cell or cause survival. Like in immune system, we have these receptors called TNF receptors or whatever, they can have a survival signal or a death signal, suicide signal, depending on the context of the cell. So that is very critical, how, as you pointed out, how these genes and proteins in a network fashion, in a sort of a topological network, what do they do? Like if I interfere, like these things called Yamanaka factors where you can generate a stem cell from a normal cell, right? So like complete regeneration, but the problem is that they can also cause cancer because they only need to be active at certain times. If they're active all the time, they can cause teratomas and things like that. So that part is so complex that we absolutely are going to need AI to simulate that for us. If I have this gene in the context of all the other things at a certain age, with these epigenetic programs plus all the metabolites and so on, because those are constantly signaling the cell and letting the proteins do something and so on. What would happen if I interfere with that particular gene or how can I improve that? Because you have to consider the other genome too. Your gene therapy might be very different than somebody else's because you might have some great genes that are synergistic, but that other person might have not so great genes, so even if you try to improve it, that would actually work or it wouldn't help. So it's just a matter of complexity. There's so much information that the AI has to not only put that together but have sort of almost a temporal simulation of the model. That's a very important point because right now the models are kind of static. They have a good understanding but they don't know what would happen if a cell comes next to a tumor just 2 minutes earlier, the cell next to it, what that context affects. There's a behavioral issue. It's the same problem with the robotics, right? So, kind of the physical intelligence or the biological intelligence, once those models are evolved with a lot of data, I think we will be able to simulate this and AI will be able to decide this is the gene therapy you should get. So, you need a new copy of immune system, but let me design it for you.

这太令人兴奋了，因为我们不仅在谈论延长寿命和治愈疾病，还在谈论以某种方式消除副作用。我的意思是，你知道，人们对不同的食物、治疗和一切的反应都不同，对吧？这就是为什么有些人对疫苗可能有严重反应，而其他人没有。所以，想到这一点真的很令人兴奋。

It's so exciting because not only are we talking about extending our lifespan and curing disease, but we're talking about getting rid of side effects in a way. I mean, you know, people all respond to different foods and treatments and everything differently, right? That's why some people have a terrible response to perhaps maybe a vaccine, and others don't. And so, it's really exciting to think about that.

顺便说一下，我称之为人类 2.0。也许我们会达到人类 3.0，那将发生在生物奇点时刻。这意味着我们重新改造自己。我总是想，大多数科学家或医生会想，这个人或病人有什么问题。我总是想相反。有些人，我说，他们有什么好的？比如这个人吸烟 50 年，从未得肺癌，或者饮食很糟糕等等。这个人活到了 110 岁，不管什么原因。那么，这些人有什么好的？为什么我们不能把所有这些人的优点都拿来，然后重新改造那些不幸没有天生拥有这些优点的人，甚至让它变得更好。这就是人类 2.0。

Which, by the way, I call human 2.0. And maybe we'll get to human 3.0, which will happen at this biosingularity moment. What that means is that we kind of re-engineer ourselves. I always think about, like, most scientists or most doctors think like, what's wrong with this person or patient. I always think the opposite. There are certain people, I'm saying, what's right about them? Like this person has smoked for 50 years, never got lung cancer, or had a terrible diet or whatever. This one lived to be 110 for whatever reason. And so, what is good about those people? Why can't we take what's good about all of those people and then re-engineer those that are not so lucky to be born with what's so good, and then even make it better. So that's the human 2.0.

对吧？我的意思是，那对我来说也很令人兴奋，对吧？我的意思是，我们确实知道，就像你说的，我们可以活，人类现在有能力活到，我想最老的大概是 121 岁，也许是 123 岁的法国女性。我的意思是，现在在 2026 年，我们知道人类至少可以活到 115 岁，我是说 115、116 岁，那被认为是目前的极限，但你知道，全世界只有 300 人活到 110 岁及以上。为什么？为什么不是其余的 80 亿人？

Right? I mean, that's exciting to me as well, right? I mean, we do know, like you said, we can live, humans are capable right now of living to be, I think the oldest was like 121 maybe, 123 French woman. I mean, the fact that right now in 2026 we know that humans can at least live to be 115, I mean at 115, 116, that's considered sort of the current limit, but you know, only 300 people in the world are 110 and older. Why is that? Why not the rest of the 8 billion?

对，是的，这很迷人。我非常兴奋拥有这种超级计算能力来帮助我们弄清楚。当山中伸弥发现 Yamanaka 因子，突然你可以把这个老细胞完全逆转成诱导多能干细胞时，你是怎么想的？你记得吗，那时你有没有想到衰老，你当时想，这几乎是你能得到的最年轻状态？

Right, yeah, it's fascinating. And I'm so excited for having this supercomputing power to help us figure that out. What did you think when the Yamanaka factors were discovered by Shinya Yamanaka and all of a sudden you could take this old cell and completely revert it to an induced pluripotent stem cell? Do you remember, was that something, did aging come into your mind at that point where you were thinking, well, that's the youngest almost you could get?

是的，当然。事实上，当时我是某个衰老研究小组的成员。我想论文发表一个小时后，我就在打字，你知道，就是它，这太神奇了。所以我应该说，对我来说有两个时刻，我认为衰老将是可逆的或可治愈的，不管你怎么称呼它。

Yeah, of course. In fact, at the time I was part of some aging groups. I think like an hour after the paper was published, I was typing, you know, this is it, this is amazing. So I should say that there were two moments for me that I thought that aging was going to be reversible or curable, however you call it.

这有点像生物学里的“chachipit”时刻。第一个时刻是那只叫多莉的羊。你可能知道，那是第一只克隆羊。大概是 1996 或 1997 年，具体日期我记不清了，但那是 90 年代。基本上，科学家从一只羊身上取了一个细胞，然后通过克隆复制出一只一模一样的羊。那是在胚胎层面，但算是精确复制。所以这意味着，只要有足够的信息，你就可以一次又一次地复制出同一个人，对吧？然后第二个时刻，当然，就是山中因子（Yamanaka factors），我记得是 2006 年。那一刻我们知道了，我们可以在细胞层面完全抹去细胞的年龄，然后把它带回多能干细胞状态，再用它来重建整个生物体。所以这意味着我们有了无限的再生能力——没有上限。事实上，我们已经知道，我们的 DNA 已经延续了数十亿年。而你能在实验室里做到这一点，并且能生成它，这太神奇了。但当然，问题在于，那要怎么应用呢？事实上，我记得最近日本有一项研究，开始用山中因子做临床试验，因为那不是一个非常可控的系统——你不知道那些细胞会不会发展成肿瘤。在小鼠身上，有些确实长了肿瘤。不管你能不能控制它们，或者更重要的是，我想大卫·辛克莱（David Sinclair）很快会启动一项试验——我们能不能做部分重编程？因为大多数时候你并不想要多能干细胞。你只是想让你的皮肤细胞回到更早期的、更像干细胞的状态。比如，我研究免疫系统，对我们来说，我可以把免疫细胞分为初始型、记忆型、效应型和分化型。初始型细胞有点像年轻细胞，它们有巨大的扩增潜力，能产生记忆和效应细胞群。其他的则不断死亡、变老。我们能不能真的把细胞逆转回初始状态？我实际上花了很长时间试图做到这一点。所以也许这种部分重编程能实现这一点，那将是革命性的，因为如果你还能把这些细胞输送回去，你就能让你大部分老化的皮肤细胞变成更年轻的版本。我想这项试验会和大卫·辛克莱合作。是的。所以，这些事向我们展示了我们可以逆转衰老。但当人们说“哦，那不可能，你不能逆转衰老，这是熵增什么的”，我们在实验室里一直在做。为什么不在整个生物体层面做呢？

It's kind of like the chachipit moment of biology. The first moment was the sheep called Dolly. You probably know it was the first cloned sheep. It was 1996 or 1997, something like that. I can't remember the exact date, but it was in the '90s. So basically, the scientists took a cell from one sheep and then recreated an exact copy of that sheep by cloning it. It was at the embryo level, but it was sort of an exact copy. So that means there was enough information that you could just recreate the same person again and again. Right? And then the second, of course, the Yamanaka factors in 2006, I think. And that was the moment that we knew we could completely erase the age of the cell on a cellular level and then bring it back to a pluripotent stem cell level and then use that to recreate the whole biological organism. So it means we have unlimited supply of regenerative capacity—there's no limit to it. In fact, we already know that our DNA just keeps going for billions of years. And the fact that you could do that in the lab and generate it was amazing. But of course, the problem was, okay, so then how do you apply that? In fact, I think there was just a recent study that started in Japan using the Yamanaka factors in clinical trials, because it was not a very controlled system—you didn't know if those cells would develop tumors. In mice, some of them did develop tumors. Whether you can control them, or importantly, I think there's going to be a trial started by David Sinclair soon—can we do partial reprogramming? Because most of the time you don't want the pluripotent cell. You just want your skin cells to go early enough to their more stem-like level. Like, I work in the immune system, and for us, I can divide immune cells into naive, memory, effector, and differentiated. So the naive cells are kind of the young guys—they have huge potential to expand and make memory and effector populations. The other ones constantly die and get older. Can we actually revert the cells towards the naive? And I actually spent a long time trying to do that. So maybe this partial programming will enable that, and that will be revolutionary, because then if you can also deliver those, you can make most of your old skin cells turn into a younger version. I think the trial is going to be with David Sinclair. Yeah. So again, these things showed us that we can reverse aging. But when people say, 'Oh, that's impossible, you can't reverse aging, this entropy whatever,' we do it in the lab all the time. Why not do it at a total organism level?

所以，关于这种部分细胞重编程，正如你提到的，你基本上是拿一个老细胞，把这些不同的蛋白质——我想现在可以用更少的蛋白质了——但把它们放在细胞上较短的时间，这会改变表观遗传程序，但细胞仍然保持其身份。它不会变成干细胞，但似乎更年轻了。我知道有一些研究，自从最初的一些研究发表以来，我没有跟进所有文献，但我想是胡安·卡洛斯·伊斯皮苏亚（Juan Carlos Izpisua）——他现在我想在 Altos Labs，但当时他在索尔克研究所——他在小鼠身上做了这个。我想他们甚至可能用的是早衰症小鼠或某种加速衰老模型，并且有一些逆转——某些器官似乎在一定程度上恢复了活力，这些动物的寿命也延长了。但有趣的是，并不是所有 12 个衰老标志都消失了。

So with this partial cellular reprogramming, as you mentioned, you're basically taking an old cell and putting these four different proteins—I think they can do it with fewer now—but putting them on for a shorter period of time on the cell, and that changes the epigenetic program in a way that the cell still keeps its identity. It doesn't become a stem cell, but it seems to be more youthful. I know there's been some work, and I haven't followed all this literature since the first studies that came out, but I think it was like Juan Carlos Izpisua—he's now, I think, at Altos Labs, but at the time he was at the Salk Institute—and he had done this in mice. I think they were even perhaps progeria mice or some sort of accelerated aging model, and there was some reversal—certain organs seemed to be rejuvenated in a sense, and the life expectancy was extended in those animals. But what's interesting is that not all of the 12 hallmarks of aging go away.

嗯。

Yeah.

对。所以你会希望完全逆转衰老，但基因组——体细胞突变仍然存在，我想端粒不会重置，线粒体也是……

Right. And so you would hope that you would reverse aging totally, but there's genomic—somatic mutations are still there, I think telomeres don't get reset, mitochondria so...

你觉得——首先，我很想理解为什么会这样。那是什么原因？如果你基本上抹去了当前的表观遗传程序并把它逆转回去，为什么不是所有东西都改变？我不知道你有没有什么想法。但你觉得 AI 会帮助我们理解这一点吗？

Do you think—first of all, I'd love to understand why that is. So what is it? If you're essentially wiping out the current epigenetic program and reverting it back, why does not everything change? I don't know if you have any ideas. But do you think AI is going to help us understand that?

当然。我是说，我们还应该指出，我们确实需要生成大量数据。所以我想每当我谈到 AI 时，人们会说：“好吧，那为什么 AI 现在做不到呢？”有两个原因。一是我们没有足够的数据。我们可能只了解所有生物学的 10% 到 20%。我们还有很多数据要生成。第二个是……

Definitely. I mean, we should also point out that we do need to generate lots of data. So I think whenever I talk about AI, people say, 'Okay, well, why can't AI do it now?' For two reasons. One is that we don't have enough data. We probably know maybe 10-20% of all the biology. We still have lots of data to generate. The second is the...

你说的是科学家。

You're talking about scientists.

是的。科学家或者自动化实验室，不管是什么。所以现在我们能在一次实验中生成数百万个数据点，但即使那样也不够——我们需要生成数十亿个数据点，等等。但当然，要处理这些，我们还需要超级智能和超级计算。所以我们必须拥有比现有算力强大数千倍的算力。人们会说：“好吧，那他们为什么要建所有这些数据中心？这还不够吗？”嗯，我们会需要的。如果你想治愈所有疾病并逆转衰老，我们可能需要太空中的数据中心，以及更多，因为有太多数据需要实时模拟。而且随着我们学习算法，我们可能会变得更高效。所以这是一个问题。另一个是，正如你指出的非常重要的一点，这种部分重编程或完全重编程非常令人兴奋，但它并不能完全解决衰老问题。它会让你 80 岁时眼睛在一段时间内看得更清楚，或者皮肤变得更好。但它能对你的心肌或脑细胞、神经元起作用吗？这才是关键点。因为如果你有一个完美的身体，但你的大脑在衰老，那就完了。

Yeah. Scientists or automated labs, whatever it is. So right now we're able to generate millions of data points in one experiment, but even that's not enough—we need to generate billions of data points and so on. But of course, to handle that, we also need superintelligence and supercompute. So we have to have compute that's thousands of times more than what's available. And people say, 'Okay, well, why are they building all these data centers? Isn't this enough?' Well, we're going to need it. If you want to cure all diseases and reverse aging, we're going to need probably data centers in space and a lot more, because so much data has to be simulated in real time. And we might get much more efficient doing that as we learn algorithms. So that's one issue. The other is that, as you pointed out something very important, this partial reprogramming or total reprogramming is super exciting, but it doesn't completely solve the aging problem. It will make your eyes see better for a certain period if you're 80 years old, or your skin gets better. But will it work on your heart muscle or on your brain cells, neurons, which is the critical point? Because if you can have a perfect body but your brain is aging, then that's it.

那么它会不会改变现在处于老年人环境中的微生物组呢？因为如果发生这种情况，如果你的新陈代谢是老年人的新陈代谢，微生物组也是老年人的新陈代谢，而且你的 DNA 积累了大量的突变，线粒体也有突变，你可以稍微逆转这一点，拥有一些再生能力，但它们很快就会再次变老，对吧？因为环境不好。所以，如果你住在一个糟糕的社区，你建了一栋漂亮的房子，但社区很糟糕，你的房子在那里不会持续很久。所以你的邻居也必须干净。所以我认为这是一件好事，这很可能会增加一定的寿命和生活质量，这是肯定的。但我们必须把它推得更远，真正了解它是否是 12 个标志。实际上，我最近问了 GPT，它又提出了四五个标志。

So will it modify the microbiome that has now the environment of an old person? Because if that happens, if your metabolism is an old person's metabolism and the microbiome is an old person's metabolism, and your DNA has accumulated a bunch of mutations and mitochondria have mutations, you can reverse that a bit, have some regenerative capacity, but they will quickly become old again, right? Because the environment is not great. So if you live in a bad neighborhood and you created this beautiful house, but it's a very bad neighborhood, your house is not going to last very long there. So your neighbors have to be clean as well. So I think it's a great thing and that's probably going to add certain years to lifespan and the quality of life for sure. But we have to push that much further and really understand whether it's 12 hallmarks. Actually, I asked GPT recently, it came up with another four or five hallmarks.

它们是什么？

What were they?

我记不太清了。其中一个是与免疫系统相关的。这是最近的事。但确实，这很有趣。试着回忆一下，有一个与新陈代谢有关。你知道，因为我们根据我们能测量和看到的东西来分类标志，而我认为 AI 能比我们看到更多一点。所以无论如何，这将是一个严肃的工程问题。如果我们有一种药丸，吃了之后突然变年轻，我会非常惊讶。这对我来说似乎非常不现实。

I can't remember exactly. One of them was related to the immune system. This was recently. But yeah, it was quite interesting. Trying to remember, one had to do with metabolism. You know, because we classify hallmarks based on what we can measure and see, and I think AI can see a little bit more than we can. So anyway, this is going to be a serious engineering problem. I would be very surprised if we have one pill you take and then you suddenly become young again. That seems very unrealistic to me.

是的。

Yeah.

我的意思是，你知道，然后另一个问题是，在实验室里，我们传递这些治疗的方式就像腺病毒，对吧？然后就像，好吧，那会导致癌症吗，因为病毒会进入正确的细胞？

I mean, you know, and then the other question is like in the lab, the way we're delivering these treatments is like an adeno virus, right? And then it's like, well, is that going to cause cancer because the virus goes to the right cell?

它会进入正确的细胞吗？正是如此。

Is it going to go to the right cell? Exactly.

我的意思是，我们肯定需要开发很多工程。所以我喜欢用 AI 模型做的事情之一就是开发一些新方法、新技术。它们有太多的护栏，所以不允许我深入其中。但你知道，因为我认为我们没有足够的工具。当然，我们现在有 CRISPR，但实际上 Dudana 的实验室刚刚推出了一种更好的细菌基因组编辑工具。所以想象一下，可能有各种各样的其他工具我们可以构建，使这种定位和编辑更加完美。而且它必须是可编程的。你必须真正创建电路。我们可以在培养中编程免疫细胞。我们可以给一种药物，它会关闭它们的反应，或者我们可以创建与门和非门。如果它们看到两个分子，它们就会反应；如果它们看到一个，它们就不会。你可以真正地编程生物学。所以我们必须开发这些比病毒更好的新工具，也许生成大量的数据集，然后操纵器官等等。可能对于一些器官，当它们太老时，修复它们可能太难了。所以你可能考虑直接换一个新的。

I mean, there's definitely a lot of engineering we have to develop. So one of the things that I like doing with AI models is to develop some new methods, new technologies. They have a bit too much guardrail, so they don't allow me to go too deep into it. But you know, because I don't think we have enough tools. Of course, we have CRISPR now, but actually Dudana's lab just came out with something even better for bacteria for genome editing. So imagine there are probably all kinds of other tools that we can build that will make this localization and editing much more perfect. And it has to be programmable. You have to literally create circuits. We can program immune cells in culture. We can give a drug that will shut down their response, or we can create AND gates and NOT gates. If they see two molecules, then they respond; if they see one, they don't. You can literally program the biology. So we have to develop these new tools that are better than viruses, maybe generate lots of data sets, and then manipulate the organs and so on. Could be that for some organs, when they're too old, it might be just too difficult to repair them. So you might consider just putting a new one.

你知道，就像可能是一个不可逆转的点，你的肾脏之类的。

You know, like it might be a point of no return, your kidneys or whatever.

然后你会有这些 3D 打印的器官工厂。实际上，我们和我的一个同事做了很多合作。他可以打印小组织、肺和类似的东西。所以其中一些将是移植新器官。其中一些将是预工程化。

Then you'll have these organ factories which are 3D printed. Actually, we did a lot of collaboration with a colleague of mine. He can print small tissues, lungs, and pieces like that. So some of them will be transplanting new organs. Some of them will be pre-engineering.

然后数字孪生、分析和模拟将能够判断你是否会排斥这个。

And then the digital twin, the analysis and simulation, will be able to figure out if you're going to reject this or not.

没错。

That's right.

对。

Right.

你对使用这个叫做 GPT-4b micro 的模型得出的新数据有什么看法？我猜这个模型被用来找出如何在四种不同的山中因子中制造某些突变，使它们更有效。所以他们基本上能够将诱导多能性的效率提高 50 倍。你如何解读这些数据？

What do you think of the new data that came out using this model called GPT-4b micro? Where I guess there's this model that was used to figure out how to make certain mutations in the four different Yamanaka factors to make them more effective. So they were able to basically 50-fold more effective or efficient at increasing this induced pluripotency. How do you interpret that data?

所以我不认为那个模型比我们现在拥有的更好。可能当前的模型要好得多。我认为可能有两个差异，我不知道所有细节。一个是他们可能移除了护栏，因为当前模型中有很多生物安全护栏。如果你问 GPT-5.5 同样的问题，它会拒绝做。它会说，“哦，这是生物危害，如果你突变并创造一种新病毒或新癌症，等等。”所以这可能是一个原因。另一个是如果你让这些模型思考更长时间。所以像 GPT-5.5 pro，思考模型是相同的预训练，但 pro 模型可以思考两个小时，思考可以花两分钟。所以它们能思考的时间越长，它们就能迭代得越多。它们可以一次又一次地运行这些场景。所以我的推测是，那个模型可能运行了很长时间。当然，你需要大量的算力和大量的 token，这对 OpenAI 来说不是问题。然后你可能会得出一个即使更智能的模型在更短时间内也无法得出的解决方案，因为那个特定案例实际上是在运行实验场景，比如，“好吧，如果我做这个突变，潜在的结果会是什么？”它运行所有的模拟。“哦，好吧，那么如果我把那个突变改到这里，然后如果我添加另一个突变呢？”然后一次又一次地运行实验。所以随着你思考的时间越长，你不断地让解决方案变得更好。而且这会变得更好。所以如果你有更多的算力，更多的智能，你说，“好吧，GPT-7 或 6，不管怎样，去思考一个月，找到能结合到这个受体并导致那个结果的完美分子，”它可能会找到。

So I don't think that model is any better than what we have right now. Probably the current models are much better. I think there might have been two differences, and I don't know all the details. One is that they probably removed the guardrails, because there are a lot of biosecurity guardrails in the current models. If you ask the same question to GPT-5.5, it will refuse to do it. It will say, 'Oh, this is a biohazard, what if you mutate and create a new virus or new cancer, whatever.' So that might be one reason. And then the other is if you let these models think longer. So like GPT-5.5 pro, the thinking model is the same pre-training, but the pro model can think for two hours, thinking can take two minutes. So the longer they can think, the more they can iterate. They can run these scenarios again and again and again. So my speculation is that that model probably ran for a long period of time. Of course, you need a lot of compute and a lot of tokens, not a problem for OpenAI. Then you will probably come up with the solution that even a more intelligent model couldn't come up with in a shorter period of time, because that particular case is really running experimental scenarios like, 'Okay, if I do this mutation, what would be the potential outcome?' It's running all the simulations. 'Oh, okay, so what if I change that mutation to here, and then what if I add another mutation?' And running the experiment again and again. So you're constantly making the solution better and better as you think longer. And this will get better. So if you have much more compute, much more intelligence, and you say, 'Okay, GPT-7 or 6, whatever, go and think for a month, find the perfect molecule that will bind to this receptor and cause that,' it'll probably figure that out.

那是什么？听起来我们将需要做很多这种类型的模拟，而“我们”我指的是研究人员和科学家。要移除那个环境中的一些护栏，让研究人员能够做出这些新发现，需要什么？以及什么样的，我猜，我们如何保护免受新的疯狂生物危害或生物安全问题？

What is it? It sounds like we're going to need to do a lot of this type of simulation, and by 'we' I mean researchers and scientists. What is it going to take to remove some of those guardrails in that environment for researchers to be able to make these new discoveries? And what sort of, I guess, how do we protect from a new crazy biohazard or biosafety issue?

嗯，我认为 OpenAI 正在与可信的人合作。所以你必须得到他们的批准。所以我认为然后，无论是公司还是类似的东西。

Well, I think OpenAI is partnering with trusted people. So you have to be approved by them. So I think then, whether it's a company or something like that.

这和网络安全是同一个问题，对吧？所以 Anthropic 有个新模型叫 Mitos，他们决定不发布它，因为他们说它对网络安全来说太危险了，因为这个模型可以破解任何东西，能找到别人看不到的所有东西。所以事实上，连政府都认为那很重要。我不知道他们是不是夸大其词，或者是不是真的，但如果你把它发布到世界上，你就必须加上那个护栏，因为有人可以用那个模型黑进你的银行账户，或者有人可以用它来制造新的病毒基因之类的。所以我认为这会在个人或机构的基础上决定。希望这些 AI 公司会分享它，因为他们可能决定不分享。他们可能会说，好吧，我们为什么不内部开发所有药物，不发布任何这些模型呢？有些人可能就在这么做。我认为这不是好事，因为你真正需要的是，正如我所说，你需要大量数据，需要大量科学家把所有这些数据输入模型，但不仅是数据，还有他们在野外、在世界中的经验。你实际上是在训练那些模型。即使它是超级智能，它也会对数据如饥似渴，以至于你必须合作或把它发布给其他人。另外，我认为这对民主化医疗保健很重要，因为每个人都问一个问题：如果你找到了衰老的治疗方法，这只会对超级富豪可用，我永远负担不起，或者癌症治疗。我说恰恰相反。多亏了 AI，它会超级便宜，因为如果你能像初创公司那样创造一种药物，比如说无法与大药企竞争，他们可以用 AI 以便宜 100 倍的价格找到一种药物，而且如果你能用数字孪生进行临床试验，那才是花钱的地方。你可以用几百万美元而不是几十亿美元开发一种药物。所以药物开发或治疗开发的成本将低几个数量级，这将让大量人获得它。但当然，AI 必须被分享。它是全人类的产品，应该归全人类所有。我是这么看的。

It's the same problem with cyber security, right? So Anthropic has this new model called Mitos and they decided not to release it because they said it's too dangerous for cyber security because this model can just crack into anything, can find all these things that others cannot see. So, in fact, even the governments thought that was important. I don't know if they're exaggerating or if it's true, but you have to put that guardrail if you release it to the world because somebody can use that model to hack into your bank account, or somebody can use it to create a new virus gene or something like that. So I think that will be made on an individual basis or institution basis. Hopefully these AI companies will share that, because they might decide not to share it. They might say, well, why don't we just develop all the drugs internally and not release any of these models? Some might be doing that, for example. I don't think that would be a good thing because what you really need is, as I said, you need a lot of data, you need a lot of scientists putting all that data into the models, but not only the data but their experience in the wild, in the world. You're actually training those models. Even if it's superintelligence, it's going to be so hungry for data that you're going to have to collaborate or release it to others. Also, I think this will be important to democratize healthcare because one question everybody asks is, well, if you find the treatment for aging, this is only going to be available for the super rich, I'm never going to be able to afford it, or treatment for cancer. I say the opposite actually. Thanks to AI, it will be super affordable because if you can create a drug like in a startup, let's say cannot compete with a big pharmaceutical company, they can find a drug using AI 100 times cheaper, and if you can do the clinical trial using digital twin, that's where all the money goes. You could develop a drug for a couple of million dollars rather than a couple of billion dollars. So the cost of drug development or treatment development will be magnitudes lower, and that will give a huge number of people access to that. But of course, AI has to be shared. It's a product of all humanity, and it should be the possession of all humanity. That's how I view it.

除了回到你在本播客开头提到的那件事，那就是人类落入坏人之手，那才是问题，而且那是需要非常认真对待的事情。

Except for going back to the thing that you mentioned at the beginning of this podcast, which is that humans in the wrong hands, that is the problem, and that is something that needs to be taken very seriously.

但解决方案也是 AI。所以现在，我听说 Mitos 基本上找到了网络安全问题中所有那些人们几十年都搞不清楚的漏洞。他们甚至不知道它们存在。所以它只是在修补所有这些安全漏洞。所以它将创造几乎完美的安全系统，比如它将无法被黑客攻击，因为 Mitos 实际上在防止这种情况。所以为了防止这种情况发生，你仍然需要 AI。你可能仍然有一些坏演员试图开发一种会导致大流行的病毒。为了防止这种情况，你还需要 AI。所以 AI 应该能够预测它，并已经准备好疫苗。它会说，好吧，有人可能会制造这种病毒，所以让我们做好准备。所以 AI 是解决所有这一切的办法。

But the solution to that is also AI. So right now, I hear that Mitos basically finds all these loopholes in cyber security issues that people couldn't figure out for decades. They didn't even know they existed. So it's just patching all those security bugs. So it will create almost perfect secure systems, like it will be unhackable because Mitos is actually preventing that. So to prevent that from happening, you still need AI. You might still have some bad actor trying to develop a virus that will cause a pandemic. To prevent that, you also need AI. So the AI should be able to predict it and already create the vaccine ready. It will say, well, somebody might make this virus, so let's get ready for it. So AI is the solution to all that.

有趣的观点。你似乎总是很乐观。我想问你另一个问题，你知道，我们在谈论这些模拟，以及我们如何用 AI 来基本上更便宜地进行临床试验，因为我们要做这些模拟，有生物标志物数据，而且会更短、更便宜、更容易。问题总是你测量什么，对吧？什么是生物标志物？终点是什么，对吧？在衰老方面，你现在可以看到，我的意思是，几乎每天都有新研究出来看这些表观遗传衰老时钟。正如大多数收听本播客的人所知，我请过 Steve Horvath 几次，他算是这些表观遗传衰老时钟的先驱，它们在过去的十年左右已经发展起来，变得更像年龄的生物标志物，比如你的生物年龄，而不仅仅是能预测你的实际时间年龄。所以你会发现现在有研究在看在治疗，以及它是否能逆转，所谓的，逆转生物衰老或表观遗传衰老，但还不清楚那是否真的是逆转衰老，对吧？那么，从你的角度来看，你认为我们应该关注哪些功能输出？

Interesting perspective. You always seem to have a positive outlook. I wanted to ask you another question about, you know, we're talking about these simulations and how we're going to use AI to essentially run these clinical trials cheaper because we're going to do these simulations and have biomarker data and it'll just be shorter and cheaper and easier. The question is always what do you measure, right? What is the biomarker? What's the endpoint, right? And in aging, you can now see, I mean, almost a new study every day coming out looking at these epigenetic aging clocks. And as most people listening to this podcast know, I've had Steve Horvath on a couple of times, and he's sort of the pioneer in these epigenetic aging clocks, and they've now developed over the last decade or so and become much more of a biological marker of age, like your biological age, not just to be able to predict your actual chronological age. And so you'll find now studies that are looking at treatments and whether or not it can reverse, quote unquote, reverse biological aging or epigenetic aging, but it's not clear that that's necessarily, you know, if that's really reversing aging, right? So how, what do you think from your perspective, what should we be looking at in terms of some of these functional outputs?

是的，我的意思是，那些表观遗传标记非常有用，但我不认为它们在预测真正衰老方面非常有用。我的意思是，这些标记有一个非常显著的问题。通常它们是通过血液分析完成的，但在血液中，你知道，我和 T 细胞打交道，所以你有这些我们称为效应细胞的细胞，它们有很多表观遗传变化，因为它们分化了，而且它们在老年时继续积累，然后你有这些更原始的初始细胞。所以这是一个组合，所以取决于那个组合是什么，它会影响输出。所以你可以直接看分化 T 细胞的比例，你可能会得到同样的信息。而且它不会告诉你皮肤、大脑或心脏发生了什么。它并不意味着如果免疫细胞变年轻了，或者年轻的细胞在扩张而老的细胞在死亡，那并不意味着你的皮肤变年轻了或你的肝脏变年轻了。所以在我看来，它的用途非常有限。但我们真的不需要那个，因为衰老可能是最容易测量的。我们确切地知道老年时哪里出了问题，对吧？所以你呼吸不太好，心脏工作不太好，肌肉不行，你只能举那么多，因为你肌肉减弱了，或者你的最大摄氧量更低了。这些都是表型。你可能甚至不需要抽血，只需测量老年人的能力。

Yeah, I mean, those epigenetic markers are very useful, but I don't believe that they are terribly useful as predicting true aging. I mean, there's a very significant problem with those markers. Usually they're done through blood analysis, but in the blood you have, like, you know, I work with T-cells, so you have these cells that we call effector cells that have lots of epigenetic change because they differentiate, and they continue to accumulate in old age, and then you have these naive cells that have more pristine kind. So it's a combination, so depending on what that combination is, it's going to affect the output. So you can actually just look at the proportion of your differentiated T-cells, you'll probably get the same kind of information. And it doesn't tell you what's happening in the skin or the brain or the heart. It doesn't mean that if the immune cells are getting younger, or the young ones are expanding and the old ones are dying, that doesn't mean that your skin is getting younger or your liver is getting younger. So it has a very limited use in my opinion. But we really don't need that because aging is probably the easiest way to measure. We know exactly what goes wrong in old age, right? So you can't breathe that well, your heart doesn't work that well, your muscles don't work, you can only raise so much because your weakened muscles, or your VO2 max is lower. These are all phenotypic. You don't even have to probably withdraw blood, just measuring the ability of an elderly person.

呃，他们走路是不是比以前好，你知道，100 米比过去走得更好？因为那是在看整体生物学，比如你的细胞、你的新陈代谢之类的，或者肌肉，对我来说那是或者你的认知能力。

Uh, can they walk uh better, you know, 100 meters than they used to? like because that's looking at the total biology like you know your cells your metabolism or whatever uh muscle to me that's or or your cognitive abilities

但这些无法被模拟，我是说。

but those can't be simulated I mean

>> 但这些无法被模拟，我是说 >> 它们最终可以被模拟，但现在还不能，它们无法被模拟，呃，因为正如我提到的，AI 缺少现实世界中的行为物理智能，因为那是大多数事情发生在现实生活中的地方。

>> but those can't be simulated I mean >> the they eventually they can be right now they can't they can't be simulated uh um because as I mentioned the AI is missing that behavioral physical intelligence in the real world because that's a that's most things are happening in real life.

呃，但我认为它们可以被模拟，但更重要的是，呃，我认为最终你必须尝试，无论 AI 得出什么，你都需要在人类身上测试，对吧。所以我的观点是，你不必做任何太花哨的事情，也不必等几十年才能看到效果。如果我给一个，呃，我不知道，80 岁的人这种治疗，他们突然能呼吸了，嗯，你知道，他们的最大摄氧量上升了。嗯，他们更敏锐了，他们能更好地思考，呃，他们能更好地记忆。嗯，你可以看看他们的免疫系统，我们可以看到细胞，我们知道哪些细胞更年轻或更差。或者你可以看看他们的皮肤，比如，哇，皮肤变年轻了。你看到了，你甚至不需要做任何事情。嗯，所以有很多衰老的表型特征可以被客观测量，实际上，而不仅仅是主观上，你会很快看到效果，比如他们正在做的这个部分重编程试验，它是为青光眼患者做的，我猜，因为那发生在老年，对吧，所以你的细胞在衰老。我的意思是，如果这些人开始看到它有效，对吧，他们的细胞得到了再生。嗯，你不需要看表观遗传学。嗯，所以我认为这将是那些测量的组合，可能我们会提出，AI 可能会提出这组生物标志物。我不认为我们知道，因为它将是一组生物标志物，比如，你知道，你的血糖、你的胆固醇可能在 30 岁时高，在 80 岁时可能高或低。我的意思是，没有一个非常具体的标志物能仅凭看它就能告诉你你的年龄。但组合效应，AI 可能能够通过查看各种数据集来预测你的年龄，并说，哦，这家伙一定是，呃，你知道，嗯，52 岁，基于这个。

uh but I think I think they can be simulated but more importantly uh I think eventually you have to tr whatever the AI comes out with you need to try it on on the humans right so uh my point is that you don't have to uh do anything too fancy or wait decades to see the effect if I give this treatment to um I don't know 80 year old and they're suddenly able to breathe Well, you know, their VMX went up. Um, they're sharper, they can think better, uh, they can remember better. Um, you you can look at their immune system and we can see that the cells are we know which cells are younger or worse. Or you can look at their skin like, oh wow, the skin is getting young. Like you see it, you don't even have to do anything. Um so so there are so many features phenotypic features of aging that could be um objectively measured actually and not just subjectively you will see the effect very very quickly like this partial reprogramming trial they're doing it's it's done for glaucoma patients I I guess uh because that happens in old age right so your your cells are aging so I mean if these people start to see it works right their their cells cells got regenerated. Um you don't need to look at the epigenetic. Um so I think uh it will be a combination of those um measurements probably we will come up with and AI will probably come up with this set of biomarkers. I don't think we know because it's going to be a set of biomarkers like um you know your glucose your cholesterol might be high when you're 30 and it will be high or low when you're 80. I mean there's not a very specific marker that will tell you your your age for just looking at that. But the combinatorial effect will will AI probably will be able to predict your age looking at all kinds of data sets and say oh this guy must be uh you know um 52 years old based on this.

>> 我知道我们有那个模型时钟基础，现在正在研究各种可能逆转表观遗传衰老的小分子。现在有一些数据集显示，如果你逆转表观遗传衰老，会有一些功能相关性，与一些功能改善相关，比如前衰弱之类的东西，你知道，比如改善，但嗯，归根结底，你知道，我认为有趣的是，看看是否会有公司出来试图销售某种药物，声称它能逆转衰老，而他们实际上只是看一个 >> 生物标志物，即逆转。

>> I know we have uh that model clock base that's looking now at a variety of small molecules that might reverse epigenetic aging. Now there are some data sets showing that you if you reverse epigenetic aging there is some functional correlation with some functional improvements like pre-frailty things like that you know like improve but um it at the end of the day you know I think it it'll be interesting to see if there's going to be companies that come out trying to sell some sort of drug to claiming it reverses aging when they're really just looking at one >> biioarker which is reversing

>> 正如我所说，他们主要看的是免疫衰老，或者可能是清除终末分化的免疫细胞，比如在老年时，你会积累这些 CMV 特异性 T 细胞，嗯，CMV 是一种你无法真正清除的病毒，所以免疫系统必须不断控制它，嗯，那些免疫细胞有点像传教士，它们应该退休，但它们不断扩张，有些个体可能有 20% 到 30% 的 T 细胞专门针对这个 CMV 的一个肽段，它们没有帮助，但变得有害，因为那些家伙老了，它们应该退休，但它们没有，它们引起炎症，因为它们活跃，嗯，它们不给年轻细胞让位，而且它们在表观遗传上是关闭的，因为它们是分化的，它们的端粒更短。所以，呃，你知道，你可能通过某些治疗清除一些这样的细胞，这很好。嗯，但然后你有间接效应，对吧？所以，如果你能控制免疫系统和炎症，那将对全身产生巨大影响，那并不意味着你的皮肤刚刚再生，但它会帮助清理 >> 衰老。

>> it's most as I say it's mostly the immune aging that they're looking at or or sort of maybe getting rid of the terminally differentiated immune cells like for example in old age you you accumulate these CME specific tea cells um CMV is a virus that you can't really get rid of so the immune system constantly have to keep it under check um and those immune cells they kind of become like missionaries they should retire but they keep on expanding and some indiv individuals might have like 20 30% of all their tea cells just dedicated to like one peptide of this this CMV and they're they're not helpful but they become uh harmful because those guys are are old they should retire they don't and they cause inflammation because they're they're active and um and they don't give place for the young guys to come in uh and they're they are epigenetically you know closed because um they're differentiated their telomeres are shorter So, uh, you know, you might be getting rid of some of those cells with certain treatments, which is great. Um, but then you have the indirect effects, right? So, if you can if you can control the immune system and inflammation, that's going to have huge effect all over your that doesn't mean your skin got just regenerated, but it it will it will help clean up >> aging.

是的。是的。没错。嗯，我还在想的另一件事是，你知道，你提到了最大摄氧量，你知道，肌肉力量、肌肉质量。我们有所有这些标志物，它们随着年龄下降，但我们不一定知道它们以某种方式导致衰老。所以问题是，AI 能否利用所有这些相关性数据，比如我们有所有这些不同的功能终点，我们观察它们，并能够将其与个性化的数据集区分开来，而不是像实际上如何治愈衰老，你改变什么来驱动逆转衰老？我的意思是，有很多问题。

Yeah. Yeah. Exactly. Um, also the other thing I was thinking about is like, you know, we you're mentioning V2 max and, you know, muscle strength, muscle mass. We have all these markers that sort of like decrease with age and yet we don't know necessarily that they cause aging in a way. So the question is like will AI be able to take all this correlational data like we have all this you know all these different functional out you know endpoints that we look at and and be able to differentiate it from like personalized you know this personalized um data set versus like actually like how do you cure aging like what do you change that's going to drive you know reverse the aging I mean there's there's a lot of questions

嗯，你提到了一个与大脑相关的有趣点，这也是我一直在思考的，因为你知道，我们身体里有很多修复过程，对吧，我们可以修复很多 DNA 损伤，你知道线粒体功能，你知道所有这些事情，但在大脑中，我们可以生长新细胞来替换旧细胞，在大脑中，它不那么强健，对吧，大脑的某些部分可以，嗯，你可以生长新神经元，神经发生，有神经可塑性。这在某种程度上是修复过程的重要组成部分，但它不像一个大的，你不是完全替换大脑，你也不想，你知道，正如你提到的，因为那样记忆会消失，你的身份也会消失，你知道，事情变得非常复杂。

Um, you mentioned something interesting that had to do with the brain and that is something that I've been thinking about as well because you know we we have a lot of repair processes in our body right we can repair a lot of DNA damage and you know mitochondrial function and you know all these things but in the brain we can grow new cells replace the old cells in the brain it's not as robust right there's some parts of the brain that can um you can grow new neurons neurogenesis there's neurop plasticity. That's a big part of of the repair process in a way, but it's not like a big you you're not you're not totally replacing the brain and you don't want to, you know, as you mentioned because then memories go away and your identity and you know, it gets very complicated.

>> 嗯，你怎么看待 AI >> 介入其中？如果我们可以逆转心脏衰老等等，那一切都很好，但大脑如此重要 >> 现在。它是否只会是，你知道，延缓与年龄相关的疾病、神经炎症等等？我们可以修复这些，但像我们真的能逆转大脑衰老吗？

>> Um, how do you see AI >> intervening in that? Like everything's great if we can reverse our heart aging and all this, but our brains that's so important >> now. Is it just going to be a, you know, delay age related disease, neuroinflammation, all that stuff? We can we can fix that, but like are we going to be able to really reverse brain aging?

>> 嗯，你知道，我得问 AI 才能弄清楚。但你知道，我能想到几种可能发生的情景。呃，首先，你知道，神经元，或者整个大脑，必须有某种非常好的维护策略，对吧，所以有些神经元可以存活几十年，也许 70、80 年，而且不仅仅是神经元，还有其他细胞类型可以存活很长时间，它们分裂不多，有一些再生，呃，它不是零，这非常重要，因为这意味着如果你让我们做一个完整的实验，假设你每个月或每年替换 0.01% 的神经元，类似这样。

>> Um, you know, I I would have to ask AI to to to figure that out. But, you know, I can I can think of several scenarios how that might happen. uh first of all you know neurons um or the brain overall must have some very good maintenance policy right so so there are neurons that live for decades maybe 70 80 years and not just neurons but there are other cell types that can live for very long they don't divide very much there is some regeneration uh it's not like zero and that's very important because that means that if you let's just do a total experiment let's just say that you replace 0.01% of your neurons uh every month or every year something like that.

我不认为那会对你的大脑结构产生巨大影响，因为他们所做的可能是，你知道，某个地方的神经元与其他神经元通过突触相互作用，然后它被替换，新神经元可能有一些其他突触，但无论如何那会替换那个网络，因为他们有那种能力。所以如果你慢慢做，我想你不会失去太多。事实上，我们仍然会失去记忆，对吧？所以我们不能记住一切，或者我们一直在产生幻觉。说到幻觉，对吧？想象一下这发生在我身上。不，不，不，那没发生。不，不，我记得。所以那就像大脑，也许部分是新神经元，他们就是不知道。所以他们编造了，对吧？

I don't think that's going to make a huge difference in your brain structure because what they're doing is that they're probably, you know, there's some neurons somewhere interacting with a bunch of other neurons via synapses, and then it gets replaced, and the new neurons might have a few other synapses other than that, but that's going to replace that network anyway because they have that capability. So if you do this slowly, I think you won't lose a lot. In fact, we still lose memories, right? So we can't remember everything, or we hallucinate all the time. Talk about hallucination, right? Imagine that this happened to me. No, no, no, it didn't happen. No, no, I remember that. So that's like brain, maybe part of it is new neurons that they just didn't know. So they just made it up, right?

另一件事是，这些神经元可能有一些内部再生能力。我的意思是，细胞可以自我维持，如果它有，你知道，一种很好的内部清理方式，比如自噬，这是一个非常重要的机制，如你所知，或者它有某种特殊的 DNA 损伤修复能力，就像干细胞那样。所以原始的干细胞不会变老，你知道，即使 100 岁，它们仍然像年轻人一样。然后你还有所有这些其他细胞，比如胶质细胞等等，它们在那里防止其他事情发生，比如炎症。你知道，胶质细胞当然在某种程度上是免疫系统的一部分，但免疫系统在极少数情况下不被允许进入大脑。那是一个受保护的区域，因为免疫系统造成太多损伤，如果你不能快速替换，那是个大问题，但它们有自己的清理网络，它们可能有某种淋巴系统等等。

The other thing is that these neurons probably have some internal abilities to regenerate. What I mean by that is that the cell can maintain itself if it has, you know, sort of a great way to clean up internally like autophagy, which is a very important mechanism, as you know, or it has some really special DNA damage correction ability, like stem cells have that right. So pristine stem cells they don't get old, you know, even in 100 years old they're still like a young person. And then you have all these other cells like glial cells and so on that are there to prevent all the other stuff that happens, the inflammation. You know, glial cells of course are part of the immune system in a way, but they are like the immune system is not allowed into the brain in very rare cases. It's like a protected area because the immune system causes too much damage, and if you can't replace it quickly, that's a huge problem, but they have their own network of cleaning up and they probably have some sort of like a lymphatic system and so on.

所以如果我们能弄清楚，或者如果我能弄清楚，我们可能真的能够，也许不是完全再生，但显著延长。也许再 10 年、20 年、30 年，随便多少。然后我们可能到达一个点，这现在有点科幻了，你知道，假设 50 年后 AI 可能能够弄清楚你大脑中的所有突触连接，比如每一个神经网络和神经递质等等。所以最终你可能能够真正模拟你的大脑，你进入矩阵级别。所以那可能让 AI 说，好吧，我要替换所有这些神经元，但我要确保它们重新连接所有这些突触，这样你就不会失去你的身份。或者，我可以在这里保留一个副本，然后我们可以创建一个新大脑，然后转移到那个新大脑，那个我发现的精确状态。我不是说现在这是可能的。那真的是科幻时代，但你可以想象在某个时候我们可能达到那个水平。

So if we can figure that out, or if I can figure that out, we might be able to really, maybe not completely regenerate, but extend it quite significantly. Maybe another 10 years, 20 years, 30 years, for whatever. And then we might come to a point, and this goes into a little bit of science fiction now, you know, let's say in 50 years time AI might be able to figure out all of the synaptic connections in your brain, like every single neural network and the neurotransmitters and everything else. So eventually you might be able to literally simulate your brain, you go into the Matrix level. So that might allow AI to say, okay, I'm going to replace all these neurons but I'm going to make sure that they reconnect all these synapses so that you don't lose your identity. Or alternately, I can keep a copy here and then we can create a new brain and then transfer to that new brain that exact state that I found. I'm not saying that this is possible right now. That's really science fiction era, but you can imagine that at some point we might get to that level.

所以我不太担心。我认为如果能度过这几十年，然后保持大脑健康和自我保存，也许，你知道，到 120 岁、130 岁。事实上，你知道，那些活到 100 岁的人，他们思维非常敏锐，对吧？

So I'm not too worried. I think if it can pass this couple of decades and then keep the brain healthy and self-preserving for maybe, you know, age 120, 130. And in fact, you know, people who actually live to age 100, they have very sharp minds, right?

因为如果你没有敏锐的思维，你就不会活得很老。所以那是高度相关的。

Because if you don't have a sharp mind, you don't live very old. So that's super correlated.

所以如果我们能再保持几十年，我们可能会找到其他解决方案。所以如果我们能保持神经炎症低，如果我们能增加脑源性神经营养因子，这些我们知道在改善神经可塑性和生长新神经元方面起作用的东西，并至少以某种可预测的方式做我们能做的一切，而且我们可以有像芯片这样的东西用于记忆部分，你知道，我们总是可以补充。

So if we can keep it for a couple of more decades, we'll probably find some other solutions. So if we can keep the neuroinflammation low, if we can increase brain-derived neurotrophic factors, some of these things that we know do play a role in improving neuroplasticity and growing new neurons and to do all the things that we can at least in some predictable way, and we can have like chips for the memory part, you know, we could always supplement that.

增加容量。

Increase the capacity.

并希望 AI 能帮助我们弄清楚如何将这些疗法输送到大脑。

And hopefully AI will help us figure out how to deliver these therapies to the brain.

是的，输送总是最大的问题，对吧？

Yeah, delivery is always the biggest problem, right?

嗯，这真是一次迷人而激动人心的对话。Derya，我还有几个问题，是收尾问题。我真的很想知道，如果你能访问，假设没有护栏，你能访问衰老生物学中的所有数据，你知道，T 细胞库、纵向队列、百岁老人数据，就像一切，任何你能想象的。你拥有这一切，你有一个惊人的模型，你可以……

Well, this has been such a fascinating and exciting conversation. Derya, I have a couple more questions, closing questions for you. And I really kind of was just wanting to know if you had access, let's say there were no guardrails and you had access to all this data in aging biology, you know, the T-cell repertoire, longitudinal cohorts, centenarian data, like everything, just anything you can imagine. You had it all and you had this model that was amazing that you could...

你在为我描述天堂。

You're describing heaven for me.

是的。是的。那提示会是什么？你会问它什么问题？我的意思是，会有不止一个，但第一个是什么？

Yes. Yes. What would be the prompt? What would be the question you would ask it? I mean, there'd be more than one, but what would be the first?

是的。希望 AI 不会回答 42 作为答案。所以我可能问的第一件事是，不是说只是去弄清楚衰老或什么的，因为我认为必须有一个顺序。所以想象你拥有所有这些数据。什么是开发一种干预措施的最实用、最快的方法，针对老年人，比如 70 岁、80 岁，能立即增加五年寿命？所以对我来说，那是最关键的直接问题，因为那个群体没有太多时间，所以我们必须极其快速地开发这些技术，我们应该有，你知道，甚至两年、三年的延长，这样我可以在那之后提出下一个提示。所以我想那是我会问的第一个提示。

Yeah. Hoping that the AI won't answer 42 as an answer. So the first thing I would probably ask is, not saying that just go figure out aging or whatever, because I think there has to be a certain sequence. So imagine that you have all this data. What would be the most practical, quickest way you can develop an intervention to an elderly person, say age 70, 80 years old, that will immediately add five years to their lifespan? So to me, that would be the most critical immediate question to ask because that population doesn't have a lot of time, and so we have to develop these technologies extremely quickly, and we should have, you know, even two years, three years extend so that I can come up with the next prompt after that. So I guess that would be the first prompt I would ask.

太好了。好的，还有一个问题。所以，这个是没有钱。钱不是问题。好的。没有像你有完全访问权限。

That's great. Okay, there's another question. So, this one is there's no money. Money is no object. Okay. There's no like you have complete access.

你现在在描述这么多天堂。

You're describing so many heavens now.

我知道。我只是好奇你的答案。你打算亲自构建你自己的数字孪生，我现在就计划这样做。你会优先考虑哪些测试？比如你会优先考虑哪些数据集，一个人如何获得它们，你会多久做一次这些测试，你会如何将这些信息组织到 AI 中，以真正获得最大的收益，你知道，从它将给你的信息中受益，对吧？

I know. I'm just curious what your answer is. You're going to personally build your own digital twin, which I plan to, right now. What tests would you prioritize? Like what data sets would you prioritize, how can a person get them, how often would you take these tests, how would you organize this information into the AI to really get the biggest bang, you know, benefit from the information it's going to give you, right?

但你说钱不是问题，对吧？

But you said money is not an issue, right?

钱不是问题。

Money's not an issue.

所以我会把它分成两部分。

So I would divide it into two parts.

一部分是我们必须建立一个庞大的实验室，部分自动化，在那里我会生成海量的细胞和组织数据。因为我们必须从第一性原理出发，去理解比如单个 T 细胞内部发生了什么——那成千上万的蛋白质、代谢物，它们在做什么？然后这将使我能够创建所谓的虚拟细胞，最终是虚拟组织，以及细胞在时空上是如何行为的。那可能是最昂贵的一部分，我需要很多钱。你说过没有限制，对吧？那好。第二部分就是我们之前谈到的：来自人类的行为数据。这些数据不仅仅是血浆中的蛋白质、代谢物水平，你的完整微生物组、全基因组测序——顺便说一句，这些都是可能的。如果成本不是问题，你可以轻松做到英国生物银行对 50 万人做的事情，但针对 100 万人。而且我认为如果对 100 万人做，那几乎就能覆盖全人类。并不是每个人都是完全不同的；我们有很多共同点。所以从人类身上收集大量生物数据，但非常重要的是行为数据。我认为这是数字孪生中完全缺失的东西。就像我们之前谈到的，一个人能走多远、能举多重——这些不会出现在任何生物标志物组合中，但它们可能极其重要。或者思考能力、认知水平，这可能直接与大脑衰老相关。还有很多东西，以及人类在特定环境中会发生什么。即使你生活方式非常健康，那也可能帮不了你太多。比如，我在纽约市住了十年。我的压力水平非常高，而那种压力是有害的，因为免疫系统不断认为有威胁，导致炎症。事实上，我认为住在纽约的人心脏病发作风险是两倍。你的环境、情绪状态、与他人的互动——所有这些都会影响你的衰老过程、你的韧性、你的乐观程度。顺便说一句，乐观是对抗衰老最好的事情之一，一项又一项研究证明了这一点。所以能够吸收坏事并继续前进——韧性。但这些是生物数据集中没有的行为数据。所以我会对全世界不同地方的 100 万人做这件事。然后在实验室里，解码我能找到的每一种细胞类型，把它们全部整合到超级智能中，然后瞧，我们就有了数字孪生。

One part is that we have to set up a huge lab, partially automated, where I would generate enormous amounts of data on cells and tissues. Because we have to go by first principles to understand what's going on in, say, an individual T-cell—all those thousands of proteins, metabolites, what are they doing? Then that will enable me to create what's called the virtual cells, and eventually virtual tissues, and how cells behave in a spatiotemporal manner. That would probably be the most expensive part, and I'll need a lot of money. You said no limit, right? So okay. The second part would be what we talked about earlier: behavioral data from humans. That data isn't just plasma levels of proteins, metabolites, your full microbiome, your full genome sequencing—all these things are possible, by the way. If cost is not an issue, you can easily do what UK Biobank did for 500,000 people, but for a million people. And I think if you did it for a million people, that would pretty much cover all of humanity. It's not like everybody is perfectly different; we share a lot of things. So from humans, collect lots of biological data, but very importantly, behavioral data. I think this is something totally missing in a digital twin. Like we were talking earlier, the ability to walk a certain distance, to lift some weights—these don't show up in any biomarker sets, but they could be extremely important. Or the ability to think, their cognitive level, which could be directly related to brain aging. And lots of things, and what happens when humans are in certain environments. Even if you have a very healthy lifestyle, that may not help you much. For example, I lived in New York City for a decade. My stress level was so high, and that stress is harmful because the immune system constantly thinks there's a threat, causing inflammation. In fact, I think people living in New York have twice the heart attack risk. Your environment, your emotional states, how you interact with others—all these impact your aging process, your resilience, your optimism. By the way, being optimistic is one of the best things for aging, study after study shows that. So being able to absorb bad things and keep going—resilience. But these are behavioral data not available in biological sets. So I would do that for a million people all over the world. And then in the lab, decode every single cell type I can find, put them all together into the superintelligence, and voila, we have a digital twin.

好的，Derya。那么假设有人现在就想用我们今天能用的模型来构建他们的小型数字孪生。我们今天在消费级层面能汇总的数据——我们可以输入的生物特征数据。你今天会如何构建那个迷你数字孪生？

Okay, Derya. So let's say someone wants to build their little mini digital twin right now using the models we have access to today. The type of data that we can aggregate at the consumer level today—biometric data that we can put in. How would you build that mini digital twin today?

是的，好问题。事实上，构建一个迷你数字孪生是可能的，它不必像我描述的那样复杂，因为那个更偏向临床试验和开发治疗方法。但回到英国生物银行的例子，他们没有数万亿的数据集；他们只用了每个人的几百个数据点，就能预测很多疾病。所以这意味着我们利用今天收集的数据就能拥有很大的预测能力。另一个例子是我有的这个血糖仪——每五分钟显示我的血糖水平，然后我把这些数据输入 ChatGPT。一旦你有了额外的数据集，那就变得非常有价值，因为假设你有你的化验值、胆固醇、血糖、每天的步数、睡眠等等。这些本身其实是非常丰富的数据，因为它们积累了大量的底层生物学信息，而且它们把 AI 放入了情境中——你的迷你数字孪生。所以我的建议是，尽可能每天给 AI 提供大量数据，并保持在同一上下文中，同一个窗口，这样模型就能记住。实际上，也有一些技巧可以做到这一点。你可以把它当作一个数据库，告诉模型：‘去查我的数据库，看看根据我的新数据，事情发生了什么变化，你能给出什么建议。’比如，我提供我服用的所有补充剂、我吃的食物类型——所有这些都会产生很大影响。所以模型开始真正个性化它的风格。它会了解你的风格，并为你提供建议，而不是给出像‘你应该走 10000 步’这样笼统的说法。它知道 Derya 不能每天走 10000 步，但 3000 步对他来说就够了。

Yeah, great question. In fact, it is possible to build a sort of mini digital twin that doesn't have to be as sophisticated as I described, because that one is more for clinical trials and developing treatments. But going back to the UK Biobank example, they didn't have trillions of data sets; they only used a few hundred data points from each person, and they were able to predict a lot of diseases. So that means we can have a lot of predictive power with the data we're collecting today. Another example is this glucose meter I have—every five minutes it shows my glucose level, and then I take that data and put it into ChatGPT. Once you have additional data sets, that becomes very valuable because, let's say you have your lab values, your cholesterol, your glucose, your daily steps, your sleep, and so on. These are actually very rich data on their own because they accumulate lots of underlying biology, but also they put AI into a context—your mini digital twin. So my suggestion would be to provide the AI as much data as you can on a daily basis, and keep it in the same context, the same window, so the model can remember. Actually, there are some tricks to do that as well. You can keep it as a database and tell the model, 'Go check my database and see, based on my new data, how things have changed and what suggestions you could give.' For example, I provide all the supplements I take, the type of food I eat—all these things will make a big difference. So the model starts to really personalize its style. It will know your style and make suggestions for you, rather than giving blanket statements like 'You should walk 10,000 steps.' Well, it knows that Derya cannot walk 10,000 steps every day, but 3,000 would be enough for him.

那我们说的是哪种模型？你会用 GPT-5.5 Pro 吗？还有那些智能体和编解码器，它们如何帮助分析你正在创建的那个数据库？

And what kind of model are we talking about? Would you be using the GPT-5.5 Pro? And then what about these agents and Codex, and how does that come into helping analyze that database that you're creating?

是的，我认为这些模型正变得越来越智能体化。我知道 OpenAI 例如，他们把智能体集成到了他们的 Codex 模型，也就是编码模型中，很快我相信它会成为所有 ChatGPT 的一部分。你不需要非常复杂的模型来做这件事。重要的是真正保持那个上下文。所以希望模型能有更大的记忆，并且能记住。ChatGPT 可以保留关于你的某些记忆，但它仍然有限。不仅仅是 ChatGPT；你可以用 Gemini，例如，它有更长的上下文窗口，或者 Claude 也可以。我认为大多数模型都能处理这些信息，而且它们处理大数据集没有问题。正如我提到的，我可以放入数百万个数据集，它们能够分析这些。

Yeah, I think these models are becoming more agentic all the time. I know OpenAI, for example, integrated agents into their Codex model, the coding model, and soon I'm sure it will be part of all of ChatGPT. You don't need very sophisticated models for that. What is important is really maintaining that context. So hopefully the models will have a larger memory and they can remember. ChatGPT can keep certain memories about you, but it's still kind of limited. It's not just ChatGPT; you can use Gemini, for example, which has a longer context window, or Claude for that matter. I think most of the models can handle that information, and they don't have problems dealing with large data sets. As I mentioned, I can put millions of data sets and they're able to analyze that.

它们需要的是记住一个月前的情况，因为那是“之前”和“之后”。之前和之后的对比极其有价值。所以模型会知道他开始服用维生素 D3，哦，那之后这些指标变了，你可能没注意到，或者血糖看起来更好了，因为那个变化发生了。于是它开始建立这些联系，而我认为这是关键点，因为你需要把所有这些上下文放进 AI 模型里，才能更好地预测该用什么、不该用什么。好吧，你之前在用那个，也许那并不是个好主意，那就换掉它，或者调整剂量之类的。

What they need is that they need to remember how things were a month ago because that's before and after. Before and after is extremely valuable. So the model will know he started taking vitamin D3. Oh, these things changed after that that you may not notice or glucose looks better because of that change. So it starts to make those links, and that's the critical point because you need all of that context in the AI model to give you a better prediction on what to use and what not to use. Okay, you were using that, well maybe that was not a great idea, so change it or change the dose or whatnot.

是的，这很有意思。这让我想起一个我确实想问你的问题，你知道，这些 AI 模型和未来的 AI 进展，当你思考这些特质时，比如持久记忆、扩展的上下文处理，似乎这些变得更加重要。

Yeah, that's interesting. It kind of reminded me of a question that I did want to ask you about, you know, these AI models and future AI advances when you think about these qualities. So, like persistent memory, expanded context handling, it seems like those seem to be more important.

绝对如此。对我来说，记忆带来了上下文，所以模型现在能够思考相当长的时间，而它们不会，因为以前模型即使在同一上下文窗口内，即使你有一百万的上下文窗口，过一会儿它们就会掉线，因为它们会忘记自己在想什么。现在它们有了这种不断去检查的能力。所以我认为在接下来的几个月里这将会发生。那将产生巨大的影响。是的，记忆就是一切。

Absolutely. For me, memory which brings the context, so the models are now able to think for quite a long time, and they don't, because previously the models would just, even in the same context window, if you had a million context windows, after a while they would just fall off because they would forget even what they were thinking about. Now they have this ability to constantly go and check on it. So I think in the next few months this is going to happen. So that will have a tremendous impact. Well, memory is everything.

所以你认为，那我们说的是多长时间？比如假设你 6 个月前开始服用维生素 D 补充剂。假设你有同一个窗口，你从那个窗口开始，你有那个入口点，那个日期，然后你不断添加你的数据。它现在已经有了所有数据。它能回溯那么远吗，或者它能回溯多远？

So you think, so how long are we talking? Like let's say you started a vitamin D supplement 6 months ago. Put that you have the same window and you start in that window, you have that entry point that the date, and then you keep adding your data. It has got all the data right now. Can it go back that far or how far can it go back?

如果你的数据在数据库里的某个地方，比如，我采用了一种由著名 AI 研究员 Karpathy 描述的技术。你可以创建自己的维基，有点像维基百科那种东西，个人化的。你拿，你知道，如果你所有的数据都在某个地方，你可以让 AI 把所有数据拉出来，放进一个类似维基百科的东西里。你可以每天或每周做一次，取决于环境。所以现在你在构建自己的健康数据库，AI 可以帮助你更新它。如果你有那些数据，它可以回溯多年，没关系。你可以有 10 年的数据，它会分析所有这些。

If you have that data somewhere in your database, for example, I adapted a technique that Karpathy, who's a famous AI researcher, described. So you can create your own wiki, sort of Wikipedia kind of a thing, like personal. You take, you know, if you have all your data somewhere, you can ask AI to pull all that and put it into a Wikipedia-like thing. You can do it daily or weekly depending on the environment. And so now you're building your own health database, which AI can help you update. If you have that data, it can go years, doesn't matter. You can have 10 years of data, it will analyze all of that.

它有那种记忆，它可以……

It has that memory, it can like...

是的，所以在同一个上下文里，如果你提供所有这些，我的意思是它仍然受限于可能一百万个 token 之类的，但没有人会有一百万 token 的数据集，即使你计算 10 年。所以那不是问题。问题在于如果你想要它持续，比如你只给 AI，好的，这是今天的数据，它应该能记住昨天是什么，两个月前是什么，这样你就不必保留自己的数据库，然后一次又一次地提供所有数据，因为你每次都得这么做，对吧？那会消耗很多 token 之类的。但我认为这将会被解决。

Yeah, so in the same context, if you provide all of that, I mean it's still limited with maybe a million tokens or something, but no one's going to have a million token data set, even if you calculate 10 years. So that's not a problem. The problem is if you want this to be continuous, like you just give AI, okay here's the data today, that it should be able to remember what was yesterday, what was 2 months ago, so you don't have to keep your own database and give all that again and again, because you have to do that every time, right? And that will spend a lot of tokens and stuff like that. But I think this is going to be solved.

你如何不产生偏见？你如何降低自己对你所知 GPT 5.5 Pro 将要反馈给你的内容的偏见能力，对吧？比如基于你问它的内容，我的意思是我发现有时候我可能能稍微引导它一点。你知道我在说什么吗？

How do you not bias? How do you lower the ability of yourself to bias what you know GPT 5.5 Pro is going to feed you back, right? Like based on what you're asking it, and I mean I find sometimes I might be able to bias it a little bit. Do you know what I'm talking about?

是的，当然。我的意思是，这就是为什么我认为我们在某种程度上处于实验阶段。每个人都需要自己做某种验证，因为模型在变得更好。我这么说的意思是，当然不要尝试有害的事情，也不要冒险，但对于日常使用，你可能在服用维生素 D，然后你停止服用维生素 D，这样你只是在做一个前后对比的实验，然后你收集前后的数据，然后 AI 给你一个解决方案，说，嗯，你知道，我认为服用这个剂量的维生素 D 很重要。所以你可以重新开始那个剂量，然后看看会发生什么。如果你达到了和以前一样的水平，那就意味着 AI 做出了一个好的预测。你需要看到，之后你必须要有前后的记录，这样你才是评判者。好吧，这是个好主意，所以我很高兴我听了 Jupy 的话。好吧，如果这不是个好主意，它也没要你的命，没让你生病。所以那也没关系。

Yeah, sure. I mean that's why I think we are in sort of the experimental phase, in a way. Everyone has to do their own kind of validation as the models are getting better. What I mean by that is that again, of course don't try harmful things and don't go into risk, but for daily use, you might be taking vitamin D and then you stop taking vitamin D so that you're just doing an experiment like before and after, and then you collect that data before and after, and then AI gives you one solution, says well, you know, taking this dose of vitamin D I think is important. So then you can start that dose again and then see what happens. If you reach the same level as before, it means that AI made a good prediction. You need to see, after you have to have that record before and after, so that you are the judge. Well, this was a good idea, so I'm glad that I listened to Jupy. Well, if it wasn't a good idea, it didn't kill you. It didn't make you sick. So that's also fine.

是的。我猜对于已经在服用很多补充剂的人来说，比如，他们不会有那种前后对比。然后你还得知道要等多久，你知道，比如清除期，还有……

Yeah. I guess for someone that's already taking a lot of supplements for example, they're not going to have that before and after. Then also you have to know like how long do you wait, you know, for example for the wash out period, and...

是的，等等。希望在于如果你非常频繁地提供数据，事实上，我可以提一件事。比如实验室数值，你去测量胆固醇、血糖、钠，不管什么，他们总是给你一个范围，对吧？所以如果在这个范围内，就是正常的。嗯，你怎么知道呢？因为你可能处于范围的上限，那可能是你的异常，别人的正常。有人可能稍微超出正常范围，但仍然没事，或者反之亦然，因为我们不知道个性化水平。所以我们计算的是基于人群的。所以好吧，这个范围对这个人群是好的。所以在某种程度上，如果你有三四次测量，比如每隔几个月，你就可以发展出你自己的设定点正常值。AI 会知道你的血糖正常值是 90，不是 70，不是 100，也不是 105，别人可能是 102。所以它基于那些测量知道这一点。然后它开始基于你的数据集、你的设定点给你建议，因为如果你的血糖是 100，突然降到 70，也许那不是好事。我只是举个例子。所以这就是为什么持续数据收集如此重要。用血糖仪，我每 5 分钟收集一次。数据越多越好。

Yeah, and whatnot. The hope is that if you provide that very frequently, in fact, I can mention one thing. For example, the lab values, like you go and measure your cholesterol, glucose, sodium, whatever, they always give you a range, right? So if it's within this range, it's normal. Well, how do you know that? Because you can be at the top of the range, that might be your abnormal, somebody else's normal. Somebody might be a little bit over the normal and might still be okay, or vice versa, because we don't know the level on a personalized level. So we calculate population base. So okay, so this range is good for this population. So in a way, if you have three or four measurements, let's say every few months, you can develop your own set point normal. The AI will know your normal for glucose is 90, not 70, not 100, or not 105, somebody else might be 102. So it knows that based on those measurements. So then it starts to give you advice based on your data set, your set points, because if yours is 100 and suddenly dropped to 70, maybe that's not a good thing. I'm just giving an example. So that's why continuous data collection is so important. With glucose meter, I collected every 5 minutes. The more data, the better.

好的，Derya，非常感谢你今天坐下来和我聊这个激动人心的前沿领域，你知道，治愈疾病、延长人类预期寿命，显然还有健康寿命、逆转衰老，也许达到人类 2.0，你知道，我们还在增强基因特征。这是一个非常激动人心的时代，如果我们能在未来 10 到 15 年内不死……

Well, Derya, thank you so much for sitting down with me today and talking about this exciting, I mean, frontier that we're exploring, you know, curing disease, extending human life expectancy, obviously health span, reversing aging, perhaps getting to human 2.0, you know, where we're enhancing genetic features as well. Very exciting time to be in, and if we cannot die in the next 10 to 15 years...

那可能会更激动人心。是的，绝对如此。因为，你知道，我要说的最后一件事是，这在人类历史上是独一无二的。

It may be even more exciting. Yes, absolutely. Because, you know, the last thing I will say, this is so unique in human history.

因为十年前，如果你告诉某人：“你应该保持健康，做这个做那个，”他们可能会说：“好吧，那也只能延长我两三年寿命。我只想好好生活，不在乎多活几年老年的时光。”那时候这么说完全合理。但现在情况不同了。多活一年，可能就能让你达到那个临界点——届时将有能力治疗许多疾病、逆转衰老，再给你十年、二十年。而一旦你达到那个点，你又能再获得十年、二十年。所以在我看来，现在每一天都很重要。这就是为什么：别死。

Because a decade ago, if you told someone, 'You should be very healthy, do this, do that,' they could say, 'Well, it's only going to extend my life by maybe two or three years. I just want to live my life, and I don't care about living a few more years in old age.' And that was perfectly relevant. That's not the case now. Living an extra year could make you reach that threshold where there will be the ability to treat many diseases, reverse your aging, and give you another decade, another 20 years. And once you reach that, you get another 10 years, another 20. So even every day counts now, in my opinion. So that's why: don't die.

人们可以在哪里了解更多你的研究并关注你？我在 X 上关注了你。也许你可以告诉大家如何关注你，你的 X 用户名是什么，以及还能在哪里找到你。

Where can people find out more about your research and follow you? I follow you on X. Maybe you can tell people how to follow you, what your X user handle is, and where else they can find you.

是的，我的主要账号在 X 上，是 @DeryaTR（带下划线）。如果他们搜“Derya Nutmas”，我想应该能找到我。那是我进行大部分交流的地方。我有一个 LinkedIn 账号，但不常发帖。我一直在计划开一个 YouTube 频道，但我觉得永远不会去做，因为我永远没时间。你做视频真的很了不起，因为视频制作很费精力。所以对我来说，X 是最快的方式。事实上，我甚至有过一个 Substack 账号，但就是没时间写长文。所以 X 是最好的方式。

Yeah, my main account is on X. It's @DeryaTR under dash. If they write 'Derya Nutmas,' I think I'll show up. That's where I do most of my communication. I have a LinkedIn account, but I don't post that often there. I've been planning to start a YouTube channel, but I don't think I'll ever do that because I'll never have the time. It's really amazing what you're doing, because video takes a lot of effort. So for me, X is the fastest way. In fact, I even had a Substack account, but just couldn't find the time to write long messages. So X is the best way.

嗯，我真的很鼓励大家去 X 上关注你。我是说，你每天都会在 X 上发一些有趣的内容，所以我强烈推荐大家关注你。

Well, I really encourage people to follow you on X. I mean, every day there's something interesting that you're posting on X, so I highly recommend that people do follow you.

已经有很多人关注了。

As many already do.

再次感谢你所做的研究，我很期待看到未来几个月会发生什么。

So, thanks again for the research you're doing, and I'm excited to see what's going to happen in the next couple of months.

期待那一天的到来。非常乐观。谢谢，非常感谢。这次访谈很棒。

Looking forward to it. Very optimistic. Thank you. Thank you very much. It was great.