第91集：迈克尔·莱文谈生物智能的新前沿

准备好与主持人Vasant Dhar一起进入一个勇敢的新世界。

Get ready to enter a brave new world with your host, Vasant Dhar.

《勇敢新世界》由纽约大学数据科学中心（CDS）支持。若贵机构有意通过学生项目与CDS合作，请发送邮件至cdsindustry@nyu.edu。想获取更多节目花絮与评论，请订阅我们的通讯vasantdar.substack.com。新年快乐，欢迎收听新一季《勇敢新世界》。今天的嘉宾是塔夫茨大学生物学教授Michael Levin。

Brave New World is supported by the Center for Data Science or CDS at NYU. If your organization is interested in engaging with CDS through student projects, please email cdsindustry@nyu.edu. For more color on the podcast and additional commentary, please subscribe to our newsletter at vasantdar.substack.com. Happy New Year, and welcome to another season of Brave New World. My guest today is Michael Levin, who's a professor of biology at Tufts University.

Michael运用生物学、神经科学、计算机科学和哲学的独特概念与方法研究认知与智能。他的研究将细胞层面的生物活动与有机体特性、人工智能及意识联系起来，形成了引人入胜的学术成果。Michael，欢迎来到《勇敢新世界》，非常高兴你能参与节目。

Michael studies cognition and intelligence using a unique blend of concepts and methods from biology, neuroscience, computer science and philosophy. It makes for some fascinating reading that relates what's going on in biology at the level of the cell to properties of organisms and AI and consciousness. Michael, welcome to Brave New World. I am delighted to have you on the show.

非常感谢。很高兴认识你，能来这里我也很开心。

Well, thank you so much. Yeah. And nice to meet you. I'm happy to be here.

在开始前，Michael请简单介绍一下自己。比如成长过程中哪些经历影响了你，最终让你投身于这个令人着迷的研究领域？

So before we get started, just tell us a little bit about yourself, Michael. Like, what, what influenced you growing up, and how did you, end up studying what you're studying, which I find fascinating?

简短来说，我从小就对工程学——特别是电子电气方面——很着迷，喜欢拆解东西。同时我也常在户外观察昆虫、虫卵、毛毛虫等。我一直对复杂性、能动性，以及人造工程产物与自然造物之间的区别充满好奇，比如偏好从何而来、智能如何具现化这类问题。本科时我最初主修计算机科学，想从事AI研究，后来获得了生物学和计算机科学双学士学位。

Yeah. Well, the short, I guess the short version is that, as as a pretty young kid, I was always interested in engineering in particular electrical kinds of things, electronics. I would take things apart and at the same time, I spent a lot of time outside looking at bugs and insects and eggs and caterpillars and whatnot And, always just always interested in in complexity, in agency, and the difference between engineering artifacts that we make, and the kinds of things that exist in the natural world, asking where do preferences come from, where do you know, how how intelligence is embodied, those kinds of things. So I had that interest from a very early age. Then I went to undergraduate initially to study computer science.

接着我攻读遗传学博士学位，完成博士后研究后，现在塔夫茨大学运营实验室，并在艾伦探索中心以非常规方式研究具身智能相关课题。同时我还兼任哈佛大学Wiese研究所教职，从事生物工程研究。

I wanted to work in AI and I got two BS degrees, one in biology and one in computer science. After that, I went to grad school and I got a PhD in genetics and yeah, and then a postdoc and now, I run a lab at Tufts University and, the Allen Discovery Center where we, address all all of the things around, embodied intelligence in very unconventional kinds of ways. I also have a have a, a faculty appointment at the Wiese Institute at Harvard where we do a lot of bioengineering as well.

这确实是条不寻常的道路。我是说，你同时涉足生物学和计算机领域，这很特别。但事实上，这正是当下最该深耕的领域。

So that's a sort of an unusual, path. Right? I mean, you've sort of been immersed in biology and computing. That's unusual. I mean, it's it's exactly the right space to be in at the moment.

是什么让你转向生物学领域的？你刚说对人工智能感兴趣，那又是什么促使你研究生物学的？

What sort of led you towards the biology side of things? Right? You said you were interested in AI. What what led you towards biology?

说来有趣——最初转行时，如今大家都明白这种跨学科结合很棒，生物与计算相得益彰。但当年我从计算机科学转向生物学时，情况艰难得多。很多人根本不认为这是个好主意，我读研时甚至遭遇了不少阻力。

Well, at first and and it's and it's funny. And and that nowadays, I think everybody understands that these kinds of interdisciplinary things are are are great and that biology and computation work together nicely. Back in the day when I was trying to make the switch from computer science to biology, it was it was much harder. It did not seem to a lot of people that this was this was a good idea at all. There was a lot of resistance to it actually when I went to grad school and things like that.

驱使我转向的原因很简单：我们当时并不清楚实现真正人工智能所需的知识。我们缺乏工具，但随处可见的昆虫卵和蠕虫卵——这些能从物质跃升为心智的复杂自组装系统——让我明白：一个静默的卵母细胞通过生化物质自组织，最终形成具有目标、偏好、问题解决能力的生命体。这使我确信发育生物学才是理解"心智如何寓于生命物质"的途径，在创造人工智能前，我们必须先理解眼前现成的范例。

But what led me to it was simply the realization that we did not know what we needed to know in order to make real artificial intelligence. So you know, we we just, we we just didn't have the tools and yet in under every rock there are eggs of various insects and worms and so on and these are incredibly complex self assembling systems that make the complete journey from matter to mind. So you have there a little oocyte, a bunch of biochemicals sitting there quiescently, and that system self assembles into a creature with goals, preferences, problem solving capacities, you know, self self driven behavior and inner perspective. And it became clear to me that developmental biology was the way to understand what I really wanted to understand, which is how mind becomes embodied in living matter. That before we could build such things, we really needed to understand the one example of it that we have right in front of us.

确实。听你讲述时我在想，研究这类现象的工具和手段实在太关键了。七八十年代我进入AI领域时，受赫伯特·西蒙影响，认知模型还停留在长时记忆/短时记忆的抽象计算框架——那时我们根本没有现在研究大脑和生物学的工具。所以很大程度上，选择进入某个领域的时机，取决于相关技术是否已发展成熟。

Yeah. And, you know, and and as you were talking, I was thinking that it's really sort of important to have the tools and the instruments to be able to study this kind of phenomenon. Right? I mean, at one point, we just didn't have the tools to study it. You know, when I got into AI, you know, late seventies, early eighties, you know, I was influenced by Herb Simon who, you know, had this sort of model of cognition as consisting of, you know, long term memory, short term memory, and there was a mechanism that sort of, you know, evoked the two and, you know, made things happen.

这种抽象计算模型曾是AI的重要组成部分，但当时我们确实缺乏现今研究大脑与生物学的工具。所以说，很多时候成功在于天时地利——当某个学科的技术工具发展到足够成熟时，恰好能把握住机遇。

Right? I mean, that was sort of a abstract model of computation that was, you know, a fairly big part of AI at the time, and we just didn't have the tools to study the brain and biology, like we do now. So a lot of it is just, I guess, being in the right place at the right time and sort of knowing when a discipline is right to go into because the instrumentation has been developed sufficiently.

仪器设备固然关键，但概念认知同样重要。比如过度关注大脑虽然看似合理，实则可能妨碍我们更广泛地理解智能。当然大脑具有独特功能，但我认为智能远比大脑古老——最基础的生命系统就已具备。要真正理解智能、识别非常规心智、构建新智能并与之伦理相处，我们必须摆脱对大脑和神经系统的过度聚焦。

Yeah. That's that's part of it. Certainly, the instrumentation is critical, but also it's conceptually it's really important because, for example, I think that the focus on the brain, while in some ways obvious, but in other ways I think is really misleading in in this whole effort to understand intelligence more broadly. I mean obviously brains do the interesting things that that are unique, but can get into it here if you want, but but I actually think that intelligence is way older than brains. It's extremely basal systems have it and that in order to really understand what's going on and to recognize unconventional minds around us and to build new minds and to relate to them ethically, we need to, loosen our hyper focus on the brain and the nervous system.

因此部分原因在于拥有工具——实验工具，但另一部分在于拥有能让你提出更广泛问题的概念框架。我认为一些经典的发育生物学家，其中许多人在分子生物学兴起后某种程度上被边缘化了，实际上早已触及这个理念并理解到，身体的自我构建与心智的自我构建从根本上说是同一个问题。顺便提一下，图灵就是其中之一。我认为图灵有过这种洞见，尽管他没有明确谈论太多。

And so part of it is having the tools, the experimental tools, but part of it is having a conceptual framework that allows you to ask broader questions and I think some of the classic developmental biologists, many of whom were sort of sidelined in some ways once molecular biology got started and so on, actually were on to this idea and understanding that, you know, the self construction of the body and the self construction of the mind are really the the same problem in a in a fundamental sense. Turing by the way being one of them I think. I think Turing had this insight, although he didn't explicitly talk much about this.

那么让我们深入探讨你理论的核心。我最近通过阅读关于虚构记忆的内容才相对熟悉你的研究。记忆并非被动存储的客观存在，而是一个我们主动虚构的过程——这个观点让我觉得很有趣。

So let's get deeper into this, right, the essence of your theory. And, you know, I became sort of familiar with your work relatively recently when I read, something about confabulation. Right? That that memory isn't this sort of passive thing there that's corded, you know, in some objective sense, but we actually, you know, confabulate. You know, that that memory seems to be an active process, which I thought was interesting.

一方面我们都知道这点对吧？我们明白自己会编造事物，或者对情境采取不同视角。但作为长期研究AI的人，我从未认真思考过记忆除了被对象化之外的其他可能。

And I you know, on on the one hand, we all know that. Right? We we we know that we sort of make things up or or we can bring different perspectives on a situation. You know? But I've been an AI for a long time, and I never sort of seriously thought about memory as anything but sort of objected.

你懂吗？记忆被记录后我们提取它，但提取过程未必忠实。我们可能在过程中进行虚构，但记忆本身是稳固的，对吧？

You know? It's it's it's recorded, and then we retrieve it. And we may retrieve it, you you know, not faithfully. We may sort of make things up along the way, but but memory itself is is solid. You know?

它总是以某种方式被记录的。但阅读你的著作后，我开始质疑这点——研究表明事实或许并非如此，记忆可能完全是另一回事。这让我思考其在司法系统、社会不公等领域的深远影响，毕竟我们总假设存在某种底层真相，而你的理论正在动摇这个基础。

It it was recorded in some way. And when I read your work, it was like, you know, it it it sort of questions that and demonstrates that that, you know, perhaps isn't the case, and then memory just seems to be something else altogether. And so that really got me thinking about the larger implications of it in in the legal system, in, you know, injustice, like, all kinds of things, right, where we sort of assume that there's some sort of underlying truth. Right? And and your theory seems to sort of, you know, throw that into question.

让我们回归主题。假设我们是初次见面的聚会宾客，当我问及你的工作时，你会如何向普通人描述你的理论和思想？

So let's just go back. I mean, how would you sort of summarize your your your thinking and a theory, you know, you know, pretending that we're sort of we're a party and we've just met, and I'm asking you what you do. You know, how how would you describe that to, you know, regular people like myself?

当然。关于记忆的研究只是我们工作的一个切面，但让我们先聚焦于此。从神经科学角度看，这绝非新理论——人们早就认识到记忆具有可塑性，回忆过程会改变记忆内容，而且神经科学界对记忆载体本质仍存争议，因为突触是极其动态的结构。虽然神经科学已有这些认知，但我将其与发育和进化生物学进行更广泛的类比，这确实为计算方式提供了全新的思路。

Sure, sure. Well, okay, I mean, and the memory thing is just it's one sort of slice of what we do, but let's just focus on that for a moment. I should say that from the perspective of neuroscience this is in no way a new theory in the sense that people have known for a long time now that memories are labile, that recall is not non destructive, that you cannot simply recall memories without changing them. And also in neuroscience there's an ongoing controversy about what the substrate of memory actually is because synapses are incredibly dynamic kinds of structures. So some of this already exists in neuroscience, but I I, draw a much wider parallel here to developmental and evolutionary biology in a way that does suggest some very, different ways to do computing.

在图灵的理论框架中，一个被明确确立的观点是数据与机器之间存在非常清晰的界限。即存在一个机器，它作用于被动数据上——这一点我们稍后还会再讨论，因为我想要打破这种区分。这里的核心思想是：机器作用于这些可靠的数据上。换句话说，就像大多数计算架构所承诺的那样，数据（如纸带）具有稳定性和保真度。通过多层抽象，当你在某个层级操作时，无需关心底层的硬件实现。

And and the thing that, was sort of nailed down in in Turing's formulation is a very sharp distinction between data and machine. So you have this machine and it operates on passive data, and that's that's something that, we'll we'll come back to towards the end of what I'm about to say, because I I I'd like to blow up that distinction as well. So so the idea here is that you have you have a machine that's going to act on this passive data and the data is reliable. In other words, the tape is reliable in the sense that most of our computing architectures are committed to the the stability and the fidelity of the data. You have these layers of abstraction where if you're operating at a certain level, you don't need to worry about the the metal underneath.

你不需要担心总线变形、晶体管数量变化或比特位突然损坏这类问题。当然，不可靠计算确实存在，但计算的核心目的正是将这些底层问题屏蔽在高层之外，使你能在抽象层级编程而无需关注底层实现。而生物学采取了完全相反的路径——这正是我认为双方都值得思考的原因。生物学从一开始就承认其基质是不可靠的。

You don't need to worry about, you know, your your bus changing shape or your the number of transistors changing or, bits of, you know, just suddenly, degrading. And I mean, of course, there's unreliable computing and things like that. But but typically, the whole point of it is to shield all that stuff from the higher levels so that you can program at a level of abstraction where you don't need to really worry about how it's done underneath. Biology takes the exact opposite approach, and this is this is why I think I think it's interesting for for both sides. What happens in biology is that you know from the start that your substrate is unreliable.

蛋白质会代谢，分子会流动，你永远无法精确掌握任何物质的数量。损伤会发生，突变会出现，环境在变化，你自身也在改变。进化面临一个两难困境：如果物种拒绝改变，必将因环境变化而灭绝；但如果改变，就不再是原来的自己。这个古老的哲学命题——如何在保持自我的同时学习与适应——正是生物学的优化方向：它不追求保真度或维持不变，而是优化显著性（salience），优化意义。接下来我会具体解释。

Proteins are gonna come and go, molecules are gonna come and go, you never quite know how many of anything you have. There will be damage, there will be mutations, the environment will be different, you will be different. Evolution faces a dilemma in that if you're a species that that doesn't change, you will surely die out when other things change, But if you do change, you are no longer yourself, so again you're gone. And so this idea, this ancient philosophical kind of question about how do I persist and yet learn and change and adapt is this idea that what biology optimizes is not for fidelity and keeping something the same, it optimizes salience. It optimizes for for meaning, and I'll talk about what what that is.

让我们暂时回归实际记忆问题。作为生命体（无论人类或其他生物），你无法直接接触过去，只能通过身体内生物物理介质留下的记忆痕迹（engrams）来获取记忆。可能是突触可塑性，也可能是格兰斯曼所说的RNA。

So let's just go back to actual memory for a second. So as a living being, whether you're a human or anything else, you don't have access to your past. What you have access to is some engrams that have been left in your body in some biophysical medium. Maybe it's synaptic plasticity. Maybe it's RNA as Glansman says.

可能性很多。我个人认为是多种机制共同作用。你所拥有的是...

It could be many things. I actually think it's many things. What you have at

关于这点——你说‘记忆痕迹’具体指什么？

Just just on that. So when you say, it's engrams, what do you mean by that?

没错，这个概念并非我首创，它由来已久。记忆痕迹就是身体因过往经历产生的生物物理变化。无论是RAM还是生物基质，要形成记忆就必须发生某种物理改变。

Yeah. So so I certainly didn't come up with the notion of engrams. It's it's very old. The idea of an engram is simply some biophysical change in your body that is left over from a prior experience. So in order for you to have memory, whether that's RAM or anything else or some kind of biological substrate, something has happened.

或许这是某种训练，或许是联想性条件反射，无论如何，那段记忆必然在你的大脑或身体某处留下痕迹。从生物物理层面看，存在某种介质，一种记忆载体。在RAM（随机存取存储器）中，你可能经历了状态翻转——但无论如何，必须存在物理层面的改变。因此在每个当下时刻，你我都必须动态重构关于我们是谁、我们在做什么的叙事，这些叙事源自我们拥有的记忆痕迹。因为你无法直接触及过去，只能通过过去在身体和大脑中留下的痕迹来建构历史。

Maybe it was some kind of training, maybe it was, you know, associative conditioning, whatever it was, some that that memory has to leave a trace in your brain or in your body somewhere. There's something biophysically, there's a medium, there's a memory medium. In in in RAM, you might have flipped, you know, flipped a flip flop state or or or whatever it is, but somewhere, there has to be a physical change. And so at any given time, at this now moment, you and I, have to reconstruct dynamically on the fly a story of who we are, what we are, what we're doing from the engrams that we have. Because because again, you don't have access to the past, have access to the history, to the to the traces that the past has left in your body and in your brain.

这意味着可以将记忆视为来自过去自我的讯息。你真正拥有的是一系列存储于某种介质中的信息，这些信息由过去的你所留存。此刻你必须即时构建一个关于世界模型与自我模型的故事，并持续调整这个故事以适应现状。这里有两个有趣之处：首先，优质学习会产生高度压缩的记忆痕迹——你不会记录见过的每个像素，而是将千万细节压缩为规律性认知，最终记住的是规律本身。

Now this means that one way to think about it as memories as messages from your past self. What you really have are a bunch of messages in some medium that have been left for you by the actions of your past self and you you must now construct on the fly a story about a model, a model of the world, a model of yourself, and continuously manage that story about about to to adaptively to make it fit current circumstances. Now there's a couple of interesting things there. First, in any good kind of learning creates very compressed engrams, meaning you don't record every pixel you've ever seen. What you record are compressed kinds of generalizations of okay I saw a thousand particulars and from this I infer a rule and what I remember is the rule.

你丢弃了相关性细节从而实现压缩，这就是我们所说的有效学习——你不是机械记忆细节，而是推断出了模式。这意味着记忆痕迹中存储的并非事件原貌，而是需要解压缩的简化表征。想象一个自动编码器：海量信息输入后，中间形成如领结般的紧缩节点，这就是压缩表征。领结左侧代表学习与概括过程，右侧则是创造性解读——为了将经验应用于新情境，你必须对已舍弃细节的信息进行非演绎性诠释。

And you throw away the correlations, that's how you get the compression and you learn, and this is what we call, you know, effective learning is you didn't just overtrain on the particulars, you actually inferred a pattern. And that means that what you get in these engrams is not the exact details of what happened, you get a compressed representation that has to be reinflated. So think about an autoencoder where you've got all this all this large stuff coming in, you've got a tight, node in the middle, it's like a bowtie node in the middle, that is your compressed representation. So the so the left part of the of the bowtie is your learning and your generalization. The right part is creative interpretation because having thrown away a bunch of details in order for you to take those lessons and apply them to your current scenario, you would have to interpret them and that interpretation is not, it is it is, it's it's not deductive.

你无法通过算法精确理解其含义。这需要某种程度的创造力，举个极端例子（人类情况没那么极端）：想象毛毛虫化蝶的转变。已知若训练毛毛虫根据特定颜色刺激觅食，化蝶后仍保留原始记忆。这有两处惊人之处：第一（已足够惊人）是在变态过程中大脑几乎重构——多数细胞死亡，突触连接断裂，全新大脑被重建，因为从二维空间爬行的软体生物，变成了需要三维空间飞行控制的硬翅生物。

You can't just, you know, algorithmically know exactly what it means. You don't know what it means. You have to apply some some degree of creativity, and I'll give you an extreme example, which is in, you know, in humans it's it's less extreme, but but here's a biological example. Imagine the butterfly the the caterpillar to butterfly transition. Okay?

记忆能在大脑重构后存续本就惊人（现有计算介质都无法做到），但更惊人的是：毛毛虫的原始记忆对蝴蝶毫无用处——蝴蝶不食叶而吸花蜜，不爬行而飞翔。因此信息处理的关键不在于保真度，设计挑战也不在于如何在脑重构中保存信息，而在于如何将信息映射到全新躯体上。蝴蝶继承的不是具体记忆，而是可适配新形态的通用经验。

It's it's now known that if you are a caterpillar and we train you to find leaves to eat on a given a certain color stimulus, that caterpillar will then become a a moth or a butterfly that still remembers the original information. Now there's two wild things about that. One, the the less amazing thing, which is amazing enough, but the less amazing thing is that during that process, the brain is basically refactored. So most of the cells die, all the connections are broken, you get a completely different brain rebuilt because now instead of a two a soft bodied robot that moves in two dimensional space, you need a controller that runs a hard bodied thing that flies in three d space. Okay.

我们都某种程度上处于这种状态——继承着来自过去的讯息。神经科学表明，人类大脑极其擅长根据各种证据（包括自身记忆）编织合理故事。这些故事不必拘泥于事实保真度，而是服务于当下的适应价值。

So so the fact that the memories survive brain refactoring is of course amazing and something that we, you know, none of our computational media can really do. But here's the more amazing fact. Notice that the actual raw memories of the caterpillar are of absolutely no use to the butterfly because because knowing how to crawl to leaves is not something that the butterfly can use. It doesn't want leaves. It drinks nectar and it doesn't crawl, it flies.

（接前文）这个极端例子说明，我们继承的所有过往信息都需要在新的生命阶段进行创造性重译。就像蝴蝶必须将毛毛虫时期的觅食经验转化为飞行导航的参考框架，人类也在不断将历史记忆重新语境化为当下可用的认知模块——这种转化过程本质上是一种生存智慧的体现。

So what has to happen to that information is not merely the fidelity, it's not the the design challenge isn't how do I keep the information even though my brain is gonna be completely refactored, It's how do I remap the information onto a totally new body? What the cat what the butterfly inherits from its former life is not the exact memories. It inherits some sort of a general, lessons that can be remapped onto a new body and onto a new structure. So I think so that's that's kind of an extreme example, but we are all to some extent in that in that boat because all of us inherit these messages from our past, and we know from neuroscience that our brains are incredibly good at concocting reasonable stories given various forms of evidence, including your own memories. And those stories are not tied to necessarily to fidelity of what actually happened, they're tied to current adaptive value.

它们与我现在能讲的最佳故事相关联。

They're tied to what what is the best story I can tell right now.

我觉得这非常迷人，比如毛毛虫变蝴蝶的例子。在阅读你的著作时，另一个让我着迷的例子是，当你将气味分子注入青蛙卵后，受试者竟会在觅食时主动寻找那种气味。要知道，我过去几年一直在研究嗅觉，这个现象立刻抓住了我的注意力——这到底是怎么发生的？

I found that really fascinating, the example of caterpillar to the butterfly. The other example I found really fascinating in in reading your work was this example you cite about, you know, when a odorant molecule is injected into a frog egg, the subject actually seeks out, that smell in its search for food. And, you know, I mean, I've been studying smell for the last couple of years, and, you know, I that just sort of got my attention. I was like, well, how the hell is that happening? Right?

那你是否对这种现象有理论解释？比如当把气味注入青蛙卵后，它为何会在后续觅食中寻找这种气味？

I mean, do you have a a theory of, like, what's happening when you inject a, you know, an odor into an into a frog egg and it sort of looks for that in its food subsequently?

具体到这个案例我并没有明确理论，但我认为这体现了一个更普遍的生物学现象：信息总是能跨越尺度传递。想象一下——清晨醒来时你可能有某些高层次社交或职业目标，这些属于高级认知目标。

What's Yeah. Don't I don't have a specific theory about what's happening in that individual case, but I think it is a an example of a much more general phenomenon, which is that in biology, information always travels across massively across scales. So so just just think about about the following. When you in in the morning, you wake up and you have certain high level social career goals, whatever. These are very high level executive cognitive goals.

要实现这些目标，你必须先起床。这意味着你的身体需要把这些抽象认知转化为肌肉细胞膜上钾钙离子的跨膜运动。人体本质上是实现'意念操控物质'的机器，这种信息能贯穿所有层级才真正令人惊叹。如果我说仅凭意念就能让化学物质穿过细胞膜，你大概会觉得我疯了，或是在谈论某种罕见的身心疗法。但实际上，这就是你身体每时每刻都在做的事。

In order for you to execute those goals, you have to get up out of bed, which means that your body has to transduce these very abstract cognitive kinds of things into the movement of potassium and calcium ions across cell membranes in your muscle. Your body is a machine for doing, you know, mind to matter kinds of controls. It's, it's it's a it's remarkable that, that kind of information propagates all of those levels. If I said to you that just with the power of my mind alone I can, make chemicals cross my cell membranes, you would either think I'm crazy or that this is some sort of very rare mind body medicine, you know, some sort of yoga, you know, some kind of crazy, like like unusual mind body medicine. But actually, this is you know, all day every day is what your body is doing.

这就是最普通的自主运动——高层行为与细胞内物理事件之间的联结。催眠皮肤学之所以有效，各种安慰剂效应研究之所以成立（正如Benedetti所说'语言与药物具有相同作用机制'），原因就在于此。我在青蛙案例中提到这些是因为：胚胎细胞内分子事件竟能转导为由大脑控制的大型动物行为级联，这种跨尺度信息传递才真正非凡。被注入的化合物必须被细胞解析，并最终影响宏观行为。

This is completely conventional voluntary motion. It's the linkage between these very high level behavioral kinds of things and physical events that happen inside your cells. This is why hypno dermatology works, this is why all kinds of placebo effect studies, you know, as Benedetti says, words and drugs have the same mechanism of action. So the reason I bring all that up in the frog case is this: what's remarkable about that case is that you have a molecular event inside the cells of an embryo which gets transduced into a behavioral cascade of a large scale animal that is presumably run by the brain. So again that information, it has to be analyzed by the cells, the compound that you're injecting has to be analyzed by the cells and it has to make its way up.

这种信息必须抵达完全不同的描述尺度——行为层面。虽然具体机制尚不明确，但本质就是信息在层级间的跃迁。关于之前讨论的记忆问题，最后补充一点：你也可以从发育信息和进化等更大尺度来思考这个机制。

That information has to make its way to a completely different scale of description which is behavior. So I I don't know exactly how it works in that case, but I think that's what we're looking at. I think all of this is about information, bouncing across levels. And, one one one last thing I wanna add to the to the previous business about the memory is simply this. You can you can also think about how this works on a much larger scale with, developmental information and evolution.

这里有个简单的例子，然后我会告诉你它如何与你询问的记忆问题相关联。想象你是一只蝾螈，我们观察肾小管的横截面，这是一条通向肾脏的小管。通常肾小管由大约8到10个细胞协同构成，中间留有空腔。对这些动物可以做的操作是，你可以在卵子中复制染色体或遗传物质，这会使细胞变大。这样做的结果是，蝾螈的实际体型完全相同，但横截面上的肾小管细胞数量从8-10个减少到4-5个，因为细胞变大了。

Here here's a simple example and then and then I'll I'll tell you how how it fits to the whole memory thing you're asking about. If imagine you're a salamander, a newt, and we look at the cross section of the kidney tubule, it's a little tube that goes to the kidney. That tubule is normally made up of let's say eight to 10 cells cooperating with each other to make this tube and leave a hole the middle. Now one thing you can do with these animals is you can create extra copies of the chromosomes, of the genetic material in the egg, and that makes the cells bigger. So if you do this what you find is that the actual salamander is exactly the same size, but the tubules now in cross section have instead of eight to 10 cells they have four to five cells because the cells are bigger.

更令人惊叹的是，如果你让细胞变得极其巨大——现在说的是六倍体蝾螈，对吧？这些蝾螈的遗传物质有大量副本。细胞会大到肾小管周围只能容纳一个细胞。它会怎么做？

What's even more amazing is that if you make the cells really gigantic, so now this is I think six n newts. Right? Newts that have many many copies of their, of their genetic material. The cells get so big that only one cell will fit around that tubule. What does it do?

它会自我弯曲并在中间留出空腔。这完全是一种不同的分子机制——不再是细胞间通讯，而是某种细胞骨架的弯曲。想想这意味着什么：作为一只初生的蝾螈，你不能预设任何条件。你无法预知自己会有多少DNA，无法预知细胞的尺寸或数量，必须拥有一种算法来构建连贯的个体（理想状态下是蝾螈），即使身体部件不断变化。

It bends around itself leaving a hole in the middle. So so this is a completely different molecular mechanism instead of cell to cell communication, now you've got cytoskeletal bending of some sort. So think about what this means. You're a newt coming into the world, you cannot assume that you you can what what can you count on? Well, you can't count on knowing how much DNA you're gonna have, you can't count on the size or the number of your cells, you have to have an algorithm that can make something coherent, ideally a newt, even with your own parts change.

更不用说环境会变化，连你自身的部件也会改变。我可以举些惊人的例子——这在涡虫实验中体现得淋漓尽致。研究最终表明，算法能完全无视极其混乱的硬件条件。我认为生物学从最初就倾向于这样的观点：硬件不可信赖，而来自前代的信息只是建议。

Never mind the environment's gonna change, but your own parts are gonna change. And so and and I could tell you some amazing examples. This is in in planaria. This really went all the way to the end with an algorithm that just basically completely ignores a very junky hardware. What happens there is that I think biology leaned in from from the earliest moments to the idea that the hardware is not to be trusted and that the information that you have from past generations is a suggestion.

这是讯息，而非具体的身体构造，不是制造特定机器的蓝图。这是你应当根据当下情境自由解读的数据，因为一切终将不同。我认为生物不会过度依赖进化先例。进化创造的是问题解决机器，这又回到记忆作为可自由解读讯息的观点。

It's a message. It is not any kind of a body, you know, a plan to create a specific machine. It is data that you should be able to interpret in whatever way makes sense at the moment knowing that everything will be different. I think living things do not overtrain on their evolutionary priors. They what evolution makes are problem solving machines and and that, you know, that that kind of goes goes back to this issue of memories as as messages that you are now free to interpret.

正因如此，我们团队一直在构思基于这个理念的新计算平台——不追求信息保真度，而是强调显著性和实时解读能力。

And this is why, I think, you know, we can we've been thinking in my group about a new, computational platform that is built around this idea. Not not not the fidelity of information, but the saliency and the on the fly interpretation.

这实在太迷人了。我一直在想为什么这两个目标（显著性与保真度）会相互冲突——是因为容量限制还是其他原因？因为我联想到你提到的持续悖论：我们的细胞在变化，今天的我们与昨天已不同，一年后更是天差地别。

That is absolutely fascinating. So, you know, I I I wondered what why those are sort of competing objectives, you know, the salience versus fidelity, and and whether that's just whether they're competing because of limited capacity or or something else. You know? Because I was thinking, you know, you talk about sort of this persistence paradox where, you know, our cells change. You know, we are physically different today than we were yesterday and and a lot different in the year from now.

然而这些记忆似乎挥之不去。你懂我意思吗？我写了一份与这个播客配套的通讯稿，之前还有一篇题为'去他妈的莎士比亚'——请原谅我的粗话——我在里面讲了个故事，笑点就是'去他妈的莎士比亚'。这个故事我讲过无数遍，因为它实在太搞笑了。我总怀疑每次讲述方式是否相同，但核心内容始终如一。

And yet these memories seem to persist. You know what mean? You know, I write a newsletter that accompanies this podcast and then a previous newsletter called fuck Shakespeare, excuse the French, where I recited a story where the punchline was fuck Shakespeare. And I've told that story so many times, you know, because it's it's just hilarious. And I wonder whether I tell it the same way every time, but the essence of it is the same.

你知道吗？我记不清具体细节了。但核心内容没变。我想表达的是，我确实可能在脑中存储了那段经历的压缩版本。由于那件事本身足够有趣且相关，所以我记住了大量细节。

You know? I I don't remember the details. The essence of it is the same. But I guess what I'm getting at is that, yeah, I can buy that I stored some compressed version of that episode in my head. And because that episode was so interesting and relevant, I remember a lot of the details of it.

但我觉得也可以这样看：我对那件事的记录并不完美，毕竟认知能力有限——记忆力有限、注意力有限。不过有些人确实能比别人更精确地记录事件。所以我想说的是，或许有些人天生记忆力更好，因为我们拥有更优越的'硬件'，因此记忆精度也更高。要知道，有些人什么都记不住。

But I guess I I would I would I I could also take the view that my recording of that episode was, you know, imperfect because I'm I'm limited cognitively. I've got a limited amount of memory, limited attention. But that you know, some people would have recorded that thing in a lot more fidelity than others. So so I I guess what I'm saying is that maybe some of us are just blessed with better memories than others because we've got better hardware, and so we have higher fidelity as well than others. You know, some people can't remember anything.

所以我在想，为什么这些特质会相互竞争。

And so I wonder why those are competing.

确实。我认为这两者并不互斥，都很重要。但显著性这个概念远不止人类记忆和对事件片段的回忆这么简单。让我们回到蝾螈的故事——当你刚来到这个世界时，完全不知道自己的身体会是什么样。

Yeah. I mean, I think I think they're not mutually exclusive, both are important. But but the thing about the salience is that it's much broader than the human memory and and recall of episodic events. I mean, the the the let's go back to the, to the to the newt story. You you coming into the world, you have no idea what your embodiment is going to be like.

也许和你一直以来的认知相同，但这种惊人可塑性存在的原因是——虽然有些例外（比如线虫），但大多数生物，通俗地说，都带着'初心'降临世间。它们必须从零开始构建自我模型和世界模型，不能做任何预设——这种特质赋予了它们惊人的能力。我再举个例子来说明新颖性和可塑性的起源：我们创造了一些从未存在过的合成生物，有趣的是无法用长期选择历史来解释它们的特性。

Now now maybe it's the same as it always has been, but but but the reason that these these this incredible plasticity exists is because most creatures, and there are some exceptions, I mean I think nematodes actually might be an exception to this, but but most creatures are, you know, a kind of a colloquial way to say it is they have a kind of beginner's mind to it, and they they come into the world and they have to build a self model and a world model from scratch. They cannot assume what and and that gives them incredible power. I'll I'll give you another example, to address the origins of novelty and the plasticity and these kinds of things. We build some synthetic creatures that have never been here before. And the interesting thing about these synthetic creatures is that you can't use a long history of selection to explain their features.

明白吗？你必须思考：为什么完全相同的标准细胞组（即相同硬件）能做出与常规功能完全不同的事？比如我们可以从成人气管上皮提取细胞，在简单条件下它们会重启多细胞性，自组装成我们称为'Anthrobot'的结构。这是什么？这是一种仅由气管上皮细胞（不含神经元）自组装成的小型结构，能够自主移动、游动，并具有包括修复遇到的神经损伤在内的非凡特性。

Okay? You have to you have to ask why a perfectly standard set of of cells so the same hardware can do something completely different than what it normally does. So we can take cells from the tracheal epithelium of an adult human and under some pretty simple conditions they reboot their multicellularity, they self assemble into something we call an anthrobot. What's an anthrobot? It's a self assembling little structure, it has only the tracheal epithelial cells, it has no neurons, it starts to move around, it swims around and it has some pretty remarkable properties including the ability to heal neural wounds that it comes across.

目前，我们正在进行各种研究，探索它们的记忆、偏好等方方面面，但关键在于——进化过程中从未出现过类人机器人。为何一位80岁患者的细胞能欣然重构出不同的身体结构、原始生物形态，并表现出不同行为？我认为这正是我们所见到的根本可塑性。生命并非不愿存储准确信息，有时确实需要且这无可厚非，但更需要随时准备以新方式重新诠释并复用自身硬件。我在多篇论文中列举了许多例证，表明形态发生与身体的自我构建并非固定不变的过程。

Now, we're doing all kinds of work studying their memories and their preferences and all this kind of stuff, but the point is this, there has never been any anthrobots in evolution. Why is it that an 80 year old patient's cells are perfectly happy to reconstruct a different body plan, a different kind of proto organism, and have different kind of behaviors? This is this is I think what we're seeing is this really fundamental plasticity. It isn't it isn't that life doesn't want to store accurate information, sometimes you need that and and that's fine, but you need to be ready to reinterpret your and reuse your own hardware in new ways. And there are many examples that I go through in various papers where I show morphogenesis and the self construction of bodies is not a hardwired process.

这无法用复杂性的涌现来合理解释。像细胞自动机这类流行范式认为简单规则能涌现复杂现象——复杂性本身很廉价，涌现复杂性很容易，但生物学运作机制并非如此。它不仅是复杂性的涌现，更是认知的涌现。这些系统实际上是在解剖空间、基因表达空间和生理状态空间中解决问题。

It is not well explained by the emergence of complexity. It's a very popular paradigm like cellular automaton, so on, where you have simple rules and then something complex emerges. Complexity is cheap, it's easy to have emergent complexity, but that isn't, what biology is doing. It's not just emergent complexity, it's emergent cognition. These things are actually solving problems in anatomical space, in gene expression space and physiological state space.

这些都是不遵循固定规则的问题解决主体。它们有偏好，能通过不同手段复用现有硬件达成相同目标。我认为进化本质就是寻找方法构建这类问题解决机器，这些机器能索引物理定律、计算法则、数学规律等赋予我们的可能性空间来解决问题，即便周遭一切（包括自身组成部分）都在剧变时，仍能构建出具有生命力的连贯系统。

All of these are problem solving agents that are not following some rote set of rules. They have preferences, they have ways to, meet the same goal by different means, by reusing the hardware at their disposal. And I think this was what most of evolution is about is finding ways to construct these kind of problem solving machines that index into the space of affordances given to us by the laws of physics, the laws of computation, mathematics and so on to be able to solve problems and, construct something, coherent that's going to have a life even even when absolutely everything around you is changing, including your own your own parts.

肯定存在某种更高层次的目标函数在驱动，即便是生物学领域也不例外。对吧？那么这个目标会是什么？仅仅是保存吗？

There's got to be some sort higher level objective function just driving, you know, even biology. Right? So what would that be? Would that be just, like, preservation?

是的。我无法给出确切答案，但可以提供几个假说。我不认为是保存——虽然这是新达尔文主义的经典论述：最可能被观察到的就是会被观察到的，这本质是保存的故事。但根据我刚阐述的框架，所有证据都表明：指望保持不变是徒劳的。

Yeah. I so so I can't give you a firm answer, but I'll give you some some hypotheses. I don't think it's preservation because if you stick with preservation, and I know that's the standard, you know, neo Darwinian story is that the things that, that are are, most likely to be observed will be observed and that's the kind of that's that's a story of preservation. I don't think it's preservation because the whole, everything about the frameworks that I just told you is about the idea that you cannot hope to remain the same. It's futile.

唯一可能的保存形式是引导你走向改变——适应性改变。这里有几个假说：最初我认为生命在做的是解决问题，即寻找特定问题的解决方案；后来认为更准确的说法是它在主动寻找问题。

You need to, the only kind of preservation is the kind that leads you to change, to to adaptive change. And so so here are here are a couple of hypotheses. At first, I thought that what life was doing was problem solving. It was it was, it was solving, looking for solutions to a specific problem. I then thought that probably better than that would be it's act what it's actually doing is looking for problems.

我认为它不止在寻找特定问题的解决方案，而是在寻找自己能解决的问题。但现在我有了更深入的认识：它真正在做的是拓展到新的问题空间，那里存在它将最终寻求解决方案的新问题。不妨想象一下

Instead of looking for solutions to a specific problem, I think it's actually doing more than that. I think it's looking for problems that it can solve. But I actually now I actually think we can do even better than that. I think what it's really doing is looking to expand into new problem spaces in which there are new problems to which it will then eventually look for solutions. So just imagine

什么意思？这具体是指什么？

Meaning what? Like what does that mean?

是的。想象一下，你是一个单细胞生物，你的所有行为都围绕着在生理和转录空间中导航。这个空间包含所有可能的基因表达和生理状态，你的目标是适应性导航。外界会有各种因素试图将你拉离稳态吸引子——那个生命最舒适的区域，而你的任务就是努力留在那里。所以你实际上是在管理自己在这个空间中的航行轨迹。

Yeah. So so imagine, you're you're a single cell and all of your actions are around navigating physiological and transcriptional space. So the space of all possible gene expressions, the space of all possible physiological states, Your goal is to navigate that space adaptively. There are things going on that try to pull you out of a homeostatic attractor where life is good, and your job is to try to stay there. And so what you're doing is you're managing your, navigation into that in in that space.

你的认知光锥（即你可能追求的目标范围）极其微小，因为你只关心当下——那些你试图调控的设定点，无论时间还是空间维度都非常局限。这些目标大约只有一个细胞大小。如果你是一个细菌，唯一关心的就是周围各种营养物质的浓度。你的记忆可能回溯二十分钟，预测能力或许能展望五分钟，仅此而已。你的认知范围非常有限。

Your cognitive light cone, which is the size of the goals you could possibly pursue, is extremely small because all you care about is what's going on, meaning the set points, the things you are trying to manage are all very small both in space and time. They're about the size of a single cell. If you're a bacterium, only thing you care about is your local concentration of various goodies. You've got maybe a memory going back twenty minutes, maybe predictive capacity going forward maybe five minutes, that's it. You've got a very small cognitive glycol.

你原本只在生理空间、代谢空间这类领域活动。现在想象你突然获得了与同类细胞连接的能力——比如通过被称为间隙连接的电突触。当进化出这种连接后，细胞开始共享记忆。

And you're operating in physiological space, metabolic space, those kinds of things. Now imagine, suddenly you acquire the ability to, connect yourself to other cells that are like you. So one way to do that is through these electrical synapses known as gap junctions. So evolution discovers these gap junctions, cells are now connected. When cells are directly connected through these gap junctions they start sharing memories.

这形成了一种意识融合：记忆分子扩散后，你无法区分彼此的记忆，于是我们拥有了共同记忆。由此获得更强的计算能力（虽然背后有更长的演化故事），关键是现在能存储关于解剖形态的大型目标状态。你体内的单个细胞不知道什么是手指，也不清楚该长几根手指，它们只关注局部目标。但细胞集体却明确知晓这些——实验证明：若截断蝾螈胚胎的肢体，细胞会立即识别偏差并再生，直到精确恢复原有解剖结构，因为它们存有空间定位记忆。

It becomes a kind of mind meld because the memory molecules spread and that means you can't tell your memory from my memory so we now share, it's our memory. You have more computational capacity and because you are now, there's a much longer story to be told, the bottom line is that you now have the ability to store large goal states as memories that are about anatomical shape. So individual cells in your body don't know what a finger is or how many fingers you're supposed to have or anything like that. All they know is their own local, goals that they can pursue. But a large cellular collective absolutely knows those things, and we know that's true because if you try to deviate it, so if you have an early embryo or in fact a salamander and you cut off some fingers of the whole limb, those cells will immediately recognize that they've been deviated and they will build back and they will stop when it's done, when the exact, correct anatomy has been restored because they have a memory of of where in anatomical space they need to be.

因此，多细胞化的本质是将自身从原先的生理/代谢空间，投射到单细胞无法触及的全新领域——解剖形态空间。在这里，生命创造出各种奇妙的动植物形态。随着肌肉和神经的出现，生命又跃入三维空间：能移动、奔跑、实现立体空间目标。再后来大脑进一步发展，我们进入了语言空间——这完全是另一个维度的突破。我认为生命的核心（如果非要定义目标函数的话，虽然这种框架本身可能有问题）...

So what you've done in in acquiring multicellularity is be is project yourself from the spaces you used to operate in, is physiological and and metabolic, into an entirely new space that individual cells couldn't have found, couldn't have, operated in, which is anatomical morphospace. So so you do this and you're in anatomical morphospace and you solve all kinds of problems in making all kinds of, amazing looking, you know, plants and animals, And then eventually you discover muscle and nerve, and suddenly you're projecting yourself into three-dimensional space because now you can move, and now you can run around and and do go and, work towards goals in three-dimensional space. So now because of this new kind of system, you're able to your brain and your musculature is now able to project you through another kind of space. And then eventually, your brain develops further and now you're in linguistic space and suddenly suddenly you can you can operate into in a completely different space. So I think I think what life is really doing is the the overall objective function to the extent that there is any objective function, I mean, think there's a problem with that framing to begin with.

但就我们所能观察到的近似目标函数而言，我认为生命始终在尝试将自己投射到能更好繁衍的新空间中去。

But but but to the extent that there that that we can see something that looks like an objective function, what I think it's doing is trying to project itself into new spaces in which it can thrive.

迈克尔，你完全让我震惊了。所以，是的，我猜，也许说目标函数有点夸张，因为也许生命的目的就是没有目的。它只是存在。但是，你知道吗，当你说话时，我想起了这部叫《我的章鱼老师》的电影。不知道你有没有看过，是佩帕·埃尔利希拍的。

Michael, you're completely blowing my mind. So, yeah, I, I guess, you know, maybe objective function is putting it strongly because maybe the purpose of life is just to there is no purpose. It just exists. But, you know, but as you were talking, I was reminded of this, film called My Octopus Teacher. I don't know if you've It's, seen by Peppa Ehrlich.

她去年是这个播客的嘉宾。你知道，电影里有这样一个场景，鲨鱼撕掉了章鱼的一条触手，但它又长回来了。对吧？所以在某些生物体内，肢体似乎能再生。对吧？

She was a guest, last year on this podcast. And, you know, there's the scene where, the shark, like, tears off one of the octopus' limbs, and it grows back. Right? But so, you know, so it seems like in some organisms, it it grows back. Right?

那些生物知道受伤后该做什么，但其他生物不行。这怎么解释呢？对吧？因为人类失去肢体就无法再生。

That that that the organism knows what it needs to do when it's done, but in others, it doesn't. So what explains that? Right? Because in humans, you lose a limb, you can't do it

不，还不是时候。我们正在研究。我不认为这是根本性的，我觉得这只是暂时的限制，将来会被突破。

back. Mean, not not not yet. We're working on it. I don't actually think that's fundamental. I think that's, that's a that's a temporary limitation that will be lifted.

让我们...好吧。更普遍的观点是，如果我们的细胞没有形成能记住整体模式的集体，我们活不过一个月。换句话说，你体内的物质和细胞时刻都在更新，我们就像忒修斯之船——船的延续性在于替换机制（即其他细胞）的记忆中。这些模式记忆储存在你体内的生物电状态、生化状态、生物力学状态等等中。如果没有这些记忆，我们就无法抵抗退化、衰老、癌症等混乱。它时刻运作以维持我们的生命。

Let's let's okay. There's a there's a the more the more general point here is that if your cells didn't emerge into collectives that remember the overall pattern, we would not survive past about a month. In other words, you know, material and cells are coming and going in your body all the time, and we we are a ship of Theseus, the the ship is in the memory of the rip replacement machinery, which is the other cells, and it is these pattern memories present in your body, they're stored in bioelectric states, they're stored in biochemical states, biomechanical states and so on. If those memories were not there, we would not be able to resist disorder of degeneration, aging, cancer. It is working constantly to keep us going.

有些动物比如蝾螈很擅长再生。这方面的冠军是涡虫，它们在这方面表现惊人，我们可以详细聊聊。我觉得值得一提。

Now some animals like salamanders are very good at it. The champion of all this is the planaria, which are amazingly good at it and we can talk about that. I think that's worth mentioning.

详细说说这个‘冠军是涡虫’是什么意思。

Say more about that when you say the champion's lies then.

好的，那我们来说说涡虫。涡虫属于扁形动物，不同于蚯蚓，它们与我们的直系祖先相似。它们拥有真正的大脑，具备许多与你我相同的神经递质。以下是涡虫的一些惊人特性：首先，你可以将它们切成碎片——据记录最多可切276块——每一块都能再生出完整的、完美的小涡虫，包括大脑等所有器官。

Well, okay, so let's talk about planaria. So planaria are a flatworm, they're not like earthworms, they are our, they're similar to our direct ancestor. They have a true brain, many of the same neurotransmitters that you and I have. Here are some amazing properties of planaria. First of all, you can, chop them up into pieces, the record I believe is 276 pieces and every piece will regenerate a complete, a perfect little worm including brain and everything else.

因此它们对大规模损伤具有惊人的抵抗力。第二，它们对小规模损伤同样极具抵抗力，这意味着它们不会得癌症。它们之所以高度抗癌，是因为对自身形态有着惊人的控制力。第三，它们是永生的，不会衰老。这些无性繁殖的涡虫已经存在了约四亿年，在此期间它们持续繁衍。

So they're incredibly resistant to large scale injury. Number two, they are also incredibly resistant to small scale injury meaning they do not get cancer. So they are very cancer resistant because they have this incredible control over their morphology. The third thing is that, they're immortal, they do not age. So the asexual lines of these planaria have been around for, I don't know, four hundred million years and during this time, they they go on.

这些动物身上没有衰老迹象，因为它们能按自身蓝图持续更新。最后一点听起来像是个悖论，让我困惑了二十多年，但我想我们终于弄清了真相：涡虫的基因组非常杂乱。我的意思是，想象一下当像你我这样的动物生育时，后代不会继承我们体内的突变，对吧？

There is no evidence of of aging in these animals because they continuously replace according to their plan. Now the last thing which which, sounds like a well, it is it is a paradox and it drove me crazy for over twenty years, but I think we finally got a handle on what's going on. The last thing about planaria is this, they have a really junky genome. And what I mean by that is just imagine when when animals like you and I have children, the kids do not inherit our our the mutations in our bodies. Right?

如果你的手臂发生突变，孩子不会继承这一点，因为我们的生殖细胞来自精卵结合。而涡虫——至少无性繁殖的品种——并非如此。它们将自己撕裂成两半后再生，这就是它们的繁殖方式。因此任何未杀死新生细胞的突变，都会在分裂后重构半个身体时传递给下一代。

If you have a mutation in your arm, your kids don't inherit that because we set aside our germline and it comes from the egg and the sperm. Planaria, at least the asexual forms, are not like that. They tear themselves in half and then they regenerate. That's how they reproduce. So any mutation that doesn't kill the the neoblast cell that's there will be, propagated into the next generation as the thing has to recreate half the body, right, as it when after it divides.

因此它们持续积累突变——数亿年来这些生物一直在积累突变。它们的细胞是所谓的混倍体，染色体数量各不相同，简直一团糟。然而，拥有永生、再生和抗癌能力的生物却有着最糟糕的基因组，这难道不奇怪吗？我在所有生物课上都从未听过这种说法，这太不合常理了。

So they accumulate, so for for hundreds of millions of years these things have been accumulating mutations. Their cells are what's called myxaploid, meaning they have different numbers of chromosomes. They're it's an incredible mess. And yet, I mean does it not strike you as weird and and you know, and I certainly never heard this in any biology class that I ever took, that the animal with immortality, regeneration, cancer resistance is also the one with the worst genome. That that doesn't sound right.

是的。我们被灌输的基本认知——基因组决定你的本质，遗传信息保真度的重要性——在这里被打破了。这让我纠结多年，最终通过计算模型和一些研究，我想我们终于弄清了其中的机制。

Yeah. But by the the basic story that we've been that we've been all told that that the genome tells you what what you are and and it's, you know, and and and the importance of your of the of the fidelity of your genetic information is is broken here. And so this this drove me nuts for for for for years, and and I think we finally, through computational modeling and and some research, we finally, I think, figured out what's what's going on there.

所以你的意思是，既然基因组这么杂乱，是不是每个个体的基因组都不同？基本上...基本上是否存在一个统一的...呃...

So is there, like so, you know, you talk about having a junky genome. I mean, does that mean that every individual has a different genome? Basically. Basically Is there, like, a single Well,

首先，基因组绝不是身体的蓝图，绝对不是。是的，个体涡虫之间存在巨大差异，因为它们会不断积累突变。

first of all, the genome is is in no case blueprint of of the of of of the body at all ever. Yeah. There are there are tremendous differences between individual planaria because they accumulate they accumulate mutations.

为什么你说基因组不是身体的蓝图？

Why do you say the genome is not a blueprint for the body?

是的，绝对不是。这是一个重大误解。如果你观察基因组，它实际上编码了什么？基因组只编码蛋白质，仅此而已。你无法直接从基因组中看到任何关于身体结构、形态、眼睛数量或是否有眼睛的信息——这些内容都不在基因组里。

Yeah, absolutely not. This is this is a major misconception. If you look into the if you look at the genome, what what does the genome actually encode? The genome encodes proteins and that's it. There's you you cannot see directly in the genome anything about your body structure, your shape, how many eyes, do you have eyes, you we have we have none none of that is in there.

基因组里包含的是纳米级硬件的描述，每个细胞都拥有的微小信号蛋白。这就是基因组的内容。之后发生的一切都是生理软件在这些基因决定的硬件基础上运行的结果。这相当于拥有计算机中锗、硅、铜等所有组件的详细描述，但基因组不包含架构，也不包含现有软件的功能。

What's in there is the description of the the nano level hardware that every cell gets to have, the little signaling proteins. That's what's in the genome. Everything else that happens after that is the result of the physiological software playing out on top of this, on top of this genetically determined hardware. So it's the equivalence to the equivalence to, having a very good description of the, germanium, the silicon, the copper, everything else in your computer. What you are not getting in the genome is the architecture and the and the and the, you know, the the capabilities of the of the software that's there.

你说的软件是指什么？是指环境吗？还是说也是你...你...嗯。软件是从哪里来的？

And when you say the software, what do you mean? Is that the environment, or is that also part of the you you you yeah. Where's the software coming from?

好的。我先告诉你它是什么，然后我们可以探讨更深层次的问题——它来自哪里，因为这会让我们进入一些哲学领域。但让我们先谈谈我所说的软件是什么意思。我们知道，如果你的硬件具有可重编程的特性，就意味着你无法仅通过硬件本身来完整描述它的功能或未来行为。你必须能够理解它的动态特性。

What's Okay. I'll I'll tell you what it is and then we can tackle the much deeper question of where is it coming from because that that gets us into some philosophical waters. But but but let's talk about what what I mean by the software. So we know that if your if your hardware is has this property of being reprogrammable, it means that you cannot, tell the full story of what it's doing or what it's going to do just by an account of the hardware alone. You have to, you have to be able to understand its its dynamical properties.

以涡虫为例，基因组编码的是一组离子通道。这些离子通道是位于细胞表面的特定蛋白质，它们让离子来回流动——这就是基因组编码的全部内容。现在想象一下，当你有一片细胞群（比如一个含有这些离子通道的胚胎）时，它们就形成了一个可电激发的介质。明白吗？

So so in in this case, imagine, in the planarian case, what, what the what the genome encodes is a bunch of ion channels. These ion channels are specific proteins that sit in the surface of the cell and they let ions go back and forth. That's it, that's what the genome encodes. So now imagine once you have a field of cells, so let's say an embryo with a bunch of these ion channels, they form an electrically excitable medium. Okay?

这类似于一种化学可激发介质，图灵模式可以在其中形成，还有许多其他类型的介质。进化塑造这些通道的方式是，硬件默认会遵循物理和计算的法则，通过这种电介质运作。它会默认采用特定配置。就像你从工厂订购计算器时，一开机，所有计算器都从零开始。就是这样。

And it's it's akin to a chemically excitable medium in which Turing patterns can be formed, and and many other kinds of kinds of media. Now the way that evolution has has shaped these channels is that that hardware by default will adopt through the laws of physics and computation that operate through this electrical medium. It will adopt by default a particular configuration. It's like when you when you order a calculator, you know, from the factory, you turn it on, every calculator starts out at zero. That's it.

这就是默认状态。电子设备就是这样运作的。计算器的制造配方里根本不会提到零。它只是告诉你，如果你以特定方式连接晶体管，默认行为就会如此，因为我们知道电学定律决定了它会这样做。所以默认的涡虫——可以说是‘开箱即用’的涡虫——其电路系统默认会形成一个头部。

That's the default. That's what the elect electronics do. Now the recipe for how to build a calculator doesn't say anything about zero. What it says is here, if you if you connect the the transistors in a certain way, the default behavior is is it will be this because we know the laws of the laws of electricity that that's what it's gonna do. So the default planarian, you know, sort of out of the box planarian, has a an electrical system that defaults to saying one head.

实际上它包含许多信息，但我们只关注头部数量。默认情况下，这套电学硬件会稳定在吸引子状态，电路会稳定在‘形成一个头部’的吸引子状态。传统理论认为这是复杂性的前馈涌现。就像图灵模式——图灵研究化学可激发介质中的这些模式时，真正关注的是秩序的起源。对吧？

It actually has says many things, but let's just focus on head number. So so by default, that electrical hardware settles into an attractor state, that circuit settles into an attractor state that says build one head. Now the traditional story is this is feed forward emergence of complexity. So just like a Turing pattern, which is, you know, what Turing was actually onto when when he was studying these these patterns in chemically excitable media, He was interested in the origins of order. Right?

身体如何构建自身？我认为他明白这本质上与心智构建自身的过程具有对称性。我们知道，通过对称性破缺和理解自组织法则，最终会形成默认模式。在这里，这个模式就是‘一个头部’。这就是传统观点。而我们研究后发现，某些电路是可重编程的，因为它们具有记忆功能。如果将其重置为其他状态，它们会保持这种状态。

How how bodies construct themselves because I think he understood that this is basically a symmetry to how minds construct construct themselves. And what we know is that through symmetry breaking and and understanding the laws of self organization, you end up with a default pattern, and in this case the pattern says one head. And and that's and that's the conventional story. Now we looked at this and we said, well, some electrical circuits are reprogrammable in the sense that they have a memory. If you reset them to something else, they will keep.

对吧？就像触发器电路，你不需要更换硬件。给它一个不同信号，它就会保持新状态。我们首先开发了可视化该电路的工具，然后追问：‘它怎么知道要形成一个头部？’我们观察发现——

Right? So so like a like a flip flop, you don't need to change out the hardware. You give it a different signal, and then and then it holds. So what we were able to do is first, we developed the tools to visualize that electric circuit, and we said how does it know to have one head? And we said, look.

我们能看到。这就是电信号分布图。这意味着‘一个头部’。现在让我们重写它，将其改写为‘两个头部’的形态。

We can see it. Here's the electrical distribution. That that means one head. So now let's rewrite it. Let's rewrite it to a to a form that says two heads.

猜猜细胞会构建出什么？这样做之后，细胞会构建出双头涡虫。更惊人的是，如果你持续切割这只双头涡虫，这些片段将永远持续生成双头生物，因为电路已被永久重置为‘两个头部’。对这些细胞集体而言，关于正常涡虫形态的记忆现在就是‘双头’。我称之为软件的原因是：这些动物在基因上没有任何差异。

Well, guess what the cells build? If you do that, the cells will build a two headed planarian. And furthermore, if you keep cutting that two headed planarian, they will those pieces will in perpetuity continue to generate two headed animals because the circuit has now permanently been reset to two heads, and as far as that cellular collective is concerned, the memory of what a proper planarian should look like is now two headed. Now the reason I call it software is this. There is nothing genetically different about those those animals.

假如我要释放它们——虽然我不会这么做——但如果我把它们放入波士顿的查尔斯河，五十年后，科学家们可能会来采集样本，他们会发现一些单头蠕虫和双头蠕虫，然后说：‘哇，酷！物种形成事件。让我们测序基因组看看发生了什么。’而他们在基因组中会发现完全相同的序列。基因组没有任何差异。信息并不存储在那里。

If I were to release them, I'm I'm not going to, but if I were to release them into the Charles River here in Boston, fifty years from now, some scientists could come along, they would scoop up some samples, they would see some one headed worms and some two headed worms, and they would say, oh, cool. A speciation event. Let's sequence the genome and see what happened, and they would see absolutely nothing different in the genome. There's nothing different with the genome. That is not where the information is.

所以在我看来，这标志着——顺便说一句，我并非主张我们冯·诺依曼架构和狭隘的编程概念适合描述生命系统。我的意思是，这些概念必须被彻底改造。但‘软件’这个概念的价值在于：我们拥有完全相同的硬件（基因编码的蛋白质），却能支持截然不同的解剖结构发展路径——无论是趋向单头还是双头的吸引子。这种重编程能力不是通过重新布线、更换硬件、修改基因组或物理部件重组实现的，而是通过高层次的生理信号体验完成的。

So this is to me, this is the mark of by the way, I am not claiming that our von Neumann architecture and our very narrow notions of programming are suitable for describing living systems. That's not what I'm saying. I think those things will have to be radically altered. But what's what's useful about the notion of software is that we have the exact same hardware, genetically encoded proteins, that is now able to support completely different visions of where they should go in anatomical space, to the one headed attractor or the two headed attractor. So the ability to reprogram this thing, not by rewiring, not by changing the hardware, not by changing the genome, not by moving around the connection between the physical parts, but by an experience, a physiological signal done at a very high level.

我们不需要指定哪些基因该激活或抑制来构建头部——这涉及成千上万个基因，我们完全没碰那些。我们是在极高层次上进行编程，改变的是关于‘应该长几个头’的记忆指令，系统就会自动完成其余工作。这些特性正是我们想从计算范式中借鉴的：理解到硬件并非与系统交互的唯一层面。

We didn't have to say what genes to turn on and off to build ahead. There are thousands of genes that have to be activated. We didn't touch any of that. We are programming at a very high level here, and we are we are changing the memory that says how many heads you're supposed to have, and then the system does the rest. So those properties are exactly the kinds of things that we want to borrow from the computational paradigm of reprogramming where you you you understand that your your hardware is not the only level at which you can interact with the system.

就像我告诉学生们的：当你从Photoshop切换到Word时，不会拿出焊枪开始改电路。因为计算机科学早已证明，某些硬件支持原位重编程——而生物学绝对具备这种原位重编程能力。

You know, as I tell my students, this is why when you wanna go from Photoshop to Microsoft Word, you don't get out your soldering iron and start rewiring. It's because we've understood in computer science that some hardware re is is is reprogrammable in place, and biology is is absolutely reprogrammable in place.

太奇妙了。那么软件其实是由环境驱动的？

So fascinating. So the software is actually then being driven by the environment?

既是也不是。就刚才描述的案例而言，确实是我们人为改变了电压状态。对蠕虫来说，是环境重编程了它。但从细胞视角看，这本质上就是身体24/7自我调控的方式。绝大多数情况下（没有医生干预时），正是你的细胞在互相重编程。

Yes and no. I mean, it's environment in the sense that we came along and in this particular case that I've just described, we came along and we changed the voltage state to be different. So as far as the worm is concerned, it's the environment that reprogrammed it. So from the perspective of those cells, but fundamentally this is how the body controls itself 20 fourseven. So typically speaking in the vast majority of cases where you're not being reprogrammed by a doctor, it is your cells that are reprogramming each other.

它们都通过这个接口互相‘黑客入侵’——这构成了我和Josh Baumgarten提出的多元计算框架：生物学所有尺度中，每个子单元都在‘黑客’周围一切及自身。不是破坏性的黑客，而是Josh Tenenbaum所说的‘孩童式黑客’——不囿于所谓‘正确’使用方式，以最具适应性的方式调动周围资源。器官的形态、某些动物的再生能力、所有生物通过细胞更替维持形态的奥秘，都源于这种软件级的身体自控机制。虽然我们能从‘环境’进行黑客干预，但这本就是生命常态的运行方式。

They are all hacking each other through this interface and this gives rise to the polycomputing framework that Josh Baumgarten and I came up with, which is this idea that that in biology at all scales, every subunit is hacking everything around it and also itself. Not hacking in the negative sense of trying to ruin it, hacking in in Josh Tenenbaum's sense of the child is a hacker, meaning you you don't know or care what the quote unquote correct way of using or interpreting something is, you will do it in whatever you you will use the affordances around you and also your own parts in whatever way is most adaptive. So the reason that organs have the shapes that they do and that, you know, in some animals they regenerate and in all animals they persist over time despite cellular replacement is precisely because the software is how the body controls itself. It can be hacked by us from quote unquote the environment, but this is naturally how this works all the time.

那么让我们回到更高层次的讨论，关于是什么让我们具有思考和感知能力，以及你在多处提到的‘自我’这个概念。我记得你曾将自我描述为某种现实的低维投影。但在日常生活中这究竟意味着什么？当我们不断在细胞层面变化，却又能保留所谓的‘记忆’时，自我在这个情境中真正的含义是什么？

So let's go back to sort of the, the higher level of, you know, what makes us sort of thinking and sentient and and and this notion of the self that you actually talk about in a in a few places. You know, you talk about the self as being sort of, if I remember correctly, something like a low dimensional projection of some reality. But what does that mean at a just at sort of an everyday level? Like, what does the self really mean in this situation where we are constantly changing, right, at a cellular level, and yet we're preserving, quote, unquote, memory. Right?

我的意思是，我们不断变化却又保持着某种本质。无论其机制是信息传递还是其他形式，在我们生命历程中总有些抽象的东西被维系着。那么在这种情况下，自我真正的含义究竟是什么？

I mean, we we change, but yet we preserve this thing. So whatever the mechanism is, you know, whether you call it message passing or or something else, there's something abstract that's being maintained as we move through life. Right? So what does what does the self really mean here?

这个问题可以从多个角度来理解。我最确信的一点是：任何我们认为客观正确的概念都不应该只有单一的定义。这些概念在不同语境下可以有多种有用的定义方式。不过现在我们先讨论其中一种——需要说明的是，由于我的理论框架尚未完善，此刻讨论的并不涉及意识问题，即不涉及作为自我存在的主观体验。

There are a number of, ways to look at it. I'm I'm one of the things I'm most, sure about is that there there should be no single definition of anything that that we think is the is the sort of objective correct one. Think there there are multiple useful ways to to define these concepts that are useful in different different contexts. But let's let's just let's just talk about about one, and I should I should say that what I'm not talking about right now, because my my framework on this is not really ready yet, is I'm not talking about consciousness. So I'm not talking about the inner perspective of what it feels like to be a self.

我并非说这不重要，也并非否认意识的存在或回避这个难题。我认为这些都至关重要，也正在做相关研究，只是目前还不成熟。意识领域确实非常复杂。但从功能角度而言，我认为可以从几个不同维度来定义自我。

I'm not saying that's not important. I'm not saying consciousness doesn't exist. I'm not saying that there isn't a heart problem. I think all of these things are important, I'm working on some things on that, but but it's not ready, so I'm I'm not gonna say anything yet about consciousness. It's a really difficult field, but more more functionally, what I think we can can define selves in a couple of different ways.

定义自我的首要特征是认知光锥——即你所能构想的最大目标的规模与形态。比如细菌的目标范围，狗的认知光锥显然更大，但狗永远不会关心三个月后邻镇发生的事，它的认知系统无法维系那种层级的思维。

The first thing that that defines a self is a cognitive cone. It is the size and, and shape of the biggest goals you can possibly muster. And you know, you can think about the size of the goals of a bacterium, a dog will have a bigger cognitive light cone of course, but for example, a dog is never going to care about what happens three months from now in in the next town. It just can't. That that size, right, that cognitive system cannot maintain the size of of of that level of cognitive.

而有些人类能设定超越生命长度的目标，甚至包含星球尺度的规划，人类的认知光锥就非常宏大。因此我认为，可以将自我理解为一个持续变化的系统，具有特定的认知光锥，这个系统始终处于意义建构和故事叙述的过程中（对人类而言是语言性的，但并非必须）。

Well, some humans are able to have goals that, you know, they're working towards things that may or may not happen long after they're dead. They have these planetary scale, you know, goals. So so the human cognitive icon is huge. So I think one thing that, one way to understand a self is as a continuously changing system with a defined cognitive lichone and that system is constantly in the process of, another way I say it is it's in the process of sense making and storytelling. In our case it's linguistic, but it doesn't have to be.

我认为即使简单系统也可能拥有非语言形式的自我——这是对其认知光锥边界的持续演化建模。你试图界定'我'与外部世界的分界，对外部世界建立模型，评估其中的能动性，判断自己处于学习还是被训练状态，揣测互动的另一端是谁。这是你不断维护更新的动态叙事：关于你的本质、关切、目标、承诺，也是选择特定认知视角的承诺——任何认知系统都必须抉择关注焦点、信息粒度及追踪状态，毕竟全知是不可能的。

I think very simple systems can have selves as well that are just not, couched in terms of language that are, it's a continuously evolving model of the boundary of your cognitive light cone. So what you're trying to say is this is where I end and where the outside world begins, you are making models of that outside world, you are trying to estimate how much agency is in that outside world, you might want to know am I learning or am I being trained, you know, who's on the other end of this interaction we're having. It's continuously developing story or model that you are trying to maintain and update and refine about what you are, what you care about, what your goals are, what your commitments are. It's also a commitment to certain perspectives, so any cognitive system has to choose what am I going to pay attention to, how much am I going to coarse grain, which states am I going to track. You can't track everything, it's impossible.

因此，自我是一系列对自身视角的承诺，是你处理信息的方式，以及你为实现目标所能调用的智力程度，并以当下合理的方式重构记忆。我们都熟悉这种感觉——虽然人类不是蝴蝶幼虫，但我们都经历过童年，拥有过随着荷尔蒙重塑大脑而彻底改变的视角、目标和偏好。你会回首感叹：哇，我曾在意那些事，如今已无关紧要。还有些记忆可能垂直也可能不垂直，谁知道呢？但你的目标是在当下讲述最合理的故事。我认为这就是本质。

So a self is a set of commitments to your perspective and to how you're going to process that information and the degree of intelligence that you have to, available to you to pursue those goals and to remap whatever memories you have in a way that makes sense now. And we're all familiar, even humans, I mean, we're not butterfly caterpillars, but but even, you know, even humans have the experience of having been children, having a set of perspectives and goals and preferences that change radically once the hormones start to remodel our brain, where you look back and you say, wow, I cared about that stuff, well that's irrelevant at this point, and then you have other memories which may or may not be vertical. Who knows? But your goal is to tell the best story you can at this at the current time. I I think that's what is

你提到意识，虽然它并非你研究的核心，但与你讨论的内容相关。我记得读过你关于意识可能源自或关联于记忆过程的观点——既然记忆是主动而非被动的，那么编造故事、虚构情节这类行为或许都是意识的组成部分？甚至可能是意识的必要条件？

So, you know, you you mentioned consciousness, and I I realize you're not that consciousness is sort of not sort of down the middle of your plate, but it is related to what you're talking about. Right? And and at some point, I remember reading something that that you were talking about that that maybe maybe consciousness derives from or is related to a process of memory actually, you know, since memory is active as opposed to passive, maybe making up stories and confabulation and things like that are all part of consciousness. Right? That maybe they are actually even necessary conditions for consciousness.

请继续展开

So say more

好的，关于意识我谈几点。一种理解方式是：意识是作为记忆不确定性的存在所体验到的感受。这个过程大多是潜意识的（颇具讽刺），我们无需刻意努力。正如马克·索姆斯所说，意识是对外部世界可感知的不确定性。

about Yeah. Okay. So I'll say a couple things about consciousness. One way I think to think about consciousness is as what it feels like to be a being uncertain of their memories. The process of constantly, and of course much of this is is is unconscious ironically enough because we don't we don't have conscious effort in doing it, but but being so so Mark Soames says that consciousness is palpated uncertainty about the outside world.

这种感受迫使你为未知的下一步做出优化决策，这是生命体最根本的特征。我进一步认为：你不仅对外界不确定，对自我也不确定。你无法随时知晓体内编码记忆的分子结构及其含义，只能猜测而非确知。

It's the it's the it's the feeling of having to make decisions to optimize for certain needs based on not actually knowing what's going to happen next. So you know, very fundamental living thing kind of thing. I like it and I extend that to the idea that not only are you not certain about the outside world, you're also not certain about yourself. You do not know out of the box, at any given moment what the molecular structures in your body that encode these memories, what they mean. You can take a guess, but you actually don't know for sure.

因此，意识的核心在于不断努力构建最合理的故事——既要满足各种心理驱力，又要符合生物本能。此外...

So the the idea the the with the hard work of constantly trying to tell the best, the most reasonable story that preserves various psychological drives, and of course, various biological imperatives, that is I think a fundamental aspect of consciousness. And the other thing that

没有不确定性就没有意识。如果你对某事始终确信无疑，只需套用决策规则而无需思考，那就不算真正的意识。

So there's no uncertainty. There's no consciousness. So you're if you're always sure about something, that you don't even need to think about it, you you just sort of apply a decision rule that would not be necessarily be consciousness. Yeah.

我是说，在某种程度上，我认为你刚才描述的——虽然我讨厌这个词——但人们提到‘机器’时想到的就是这种情形。当人们说‘那不是意识，那只是机器’时，他们脑海中浮现的是某种算法化的存在：按部就班地执行步骤，所有数据和指令都毫无疑义，一切都被预设好，每个环节都有明确定义，就这样机械运转。我同意这种基质不太可能孕育意识，但我要补充个古怪观点——这与我们几个月前预印的研究有关，正式论文这几天就会发表。

I mean, to to the extent to the extent that I mean, I think what you've just described, and and I hate this word, but but I think what you've just described is what people have in mind when they use the word machine. When people say, well, that's not the conscious, that's just a machine. I think what they have in mind is something that is algorithmic in the sense that it cranks through a set of steps where all the data and all the instructions are not under any kind of doubt. Everything is prescribed, everybody knows what everything means and you just sort of roll along. I agree that I think that kind of thing is not the sort of substrate where you would expect to find consciousness, but I will I will put a a weird twist on it, which is, and this this relates to some work that we preprinted some months ago, and the real paper is gonna be out, I think, this week or in a couple of days.

在生命科学领域存在两大阵营：机械论者与有机论者。机械论者认为万物皆自下而上运作，所有重要事实都存在于化学层面。他们不算真正的还原主义者，因为不愿讨论量子泡沫，但对化学情有独钟。

In in the life sciences, there are two basic camps. One one is the mechanist and one is the organist. So so the mechanists think, look, all everything proceeds bottom up. All the important facts are to be found at the level of chemistry. They're not really reductionist because they don't wanna talk about quantum foam, but they like chemistry.

严格来说不算还原主义，但他们痴迷化学，认为最佳叙事要用化学语言讲述，万物皆由此衍生，我们都是化学机器——仅此而已。有机论阵营则主张：化学故事与自下而上的规则无法囊括生命奇迹的全部要义，因为存在涌现现象——不仅是不可预测性，也不仅是复杂性，更是涌现的认知与内在视角，这些都无法用硬件规则解释。然后他们需要具体阐明其本质——这非常困难，但学者们正在研究。

So so not really reductionism, but for some reason they love chemistry, and and the idea is that the best stories are told in the language of chemistry, and everything else proceeds from that, and we are all chemical machines, and and that's all you get. The, organicist camp says, look, the story of chemistry and the bottom up rules does not capture everything there is of importance about the majesty of living things because there are emergent not just emergent unpredictability, just emergent complexity, but actually emergent cognition and emergent inner perspective that is not captured by the rules of the hardware. Right? And then and then it's up to them to specify what what that is, and then that's that's really hard, but people are people work on it.

顺便问下，这是否与自由意志vs决定论有关？

By the way, would this be related to sort of free will versus determinism?

有关联。

It's related.

前者会说超越...

That is the former would say beyond

确实有关联。虽然这是个稍有不同的议题——我们需要探讨你对自由意志概念的预期作用，这可以深入讨论。但关键在于，我要提出个疯狂主张——作为计算机科学家你可能会觉得难以接受，但我还是决定说出来。

It's it's related. Yeah. It's it's related, although although that's a slightly separate question that where we would have to get into what what work you expect the concept of free will to do, and we can talk about that. That's a that's a slightly more, you know, it's an in-depth discussion. But but the thing is that so so here's the crazy claim that that I'll make and and and it probably, you know, will will probably not sound good to you as a computer scientist, but I'll I'll do it anyway.

我认为，出于与有机论者相同的理由——他们发现生命体中有重要方面无法被化学定律所涵盖，尽管化学定律并未被明确违反——基于同样的原因，我们应当对算法系统中涌现的、极其基础（微小到不可思议）但确实存在的初级认知乃至意识持开放态度。我要提出的主张是：算法作为一种形式模型，即便由你编写创造，也无法完全捕捉系统的全部奥秘——就像生命体不违反化学定律那样，系统也不会违背算法步骤。然而，即便是极其简单的算法，也能做出算法本身无法直观解释的行为。举个简单例子（如果你没看过这篇论文我可以发你）：我们研究了冒泡排序、选择排序等六行代码的确定性排序算法。这些算法透明直白，被研究了几十年，人们自以为完全了解其行为——但它们实际所做的，恰恰就是算法步骤本身所规定的。

I I think for the exact same reasons that the organicists find important aspects of living things that are not captured in the laws of chemistry, even though the laws of chemistry are not violated explicitly. For that for that same reason they should be open to emergent, very basal, so so incredibly small, but emergent basal cognition and possibly consciousness in algorithmic systems because, and the claim that I'll make, is that algorithms are a formal model that does not capture everything there is to know about a system even though you wrote the algorithm, you made it, the system is not violating the, the individual steps of the algorithm any more than living things violate chemistry, and yet even extremely simple algorithms can do things that that are not obvious from the algorithm. And the very simple example, I I can send you a copy of this paper if you haven't seen it, is we looked at sorting algorithms like bubble sort, selection sort, you know, sorting algorithms. Six lines of code, deterministic, transparent, people have been studying them for decades, where people think they know what they do and they do what do they do? They do exactly what the algorithm says they do.

事实证明，如果你不预设结论而是探究其行为，将这些算法视为在可能的有序字符串空间中导航的智能体，就会发现非常有趣的现象：会出现延迟满足效应、具备错误规避能力、还会优化出我们称之为'聚类'的行为——这与算法设计初衷毫无关联。这种在基础层级就显现的意外特性前所未见。我的核心观点是：虽然意识本身不应与绝对确定的系统挂钩，但那些看似算法化的事物，我认为本质上未必如此。我们必须警惕不要将形式模型与实体本身混淆——当你面对六行排序代码时，其实已置身于存在你未曾预料的、具有目标导向性的涌现智能的领域。

Turns out that if you actually, if you don't assume that that's the case and you actually probe their behavior, if you if you pretend that they are agents navigating the space of possible sort possible ordered strings, you find some very interesting behavior. You find delayed gratification, you find the ability to navigate around errors, you find that they are actually optimizing something we found called clustering, which has nothing to do with with what you've said in the algorithm, and extremely basal simple thing, which is surprising and and had never been seen before, you can already find at that level. So so my my point simply is that, yes, consciousness by itself should not be associated with extremely, determined systems where nothing is under, question. However, even the things that look to us like they are algorithmic, I don't think are necessarily that at all. I think we have to be really careful not to confuse our formal models with the thing itself, and even already by the time you get to a six line sorting algorithm, you are already I think in the land where you have some emergent goal directed intelligence that, you did not anticipate.

迈克尔，你的研究让我思考一个更宏观的问题：当前AI范式（我们现有的AI技术）与生物学现实之间存在多大程度的脱节？生物学对智能的影响似乎远超我们的理解。这对AI未来意味着什么？要创造下一代智能机器，真正需要突破什么？或者从另一个角度：你认为当前AI存在哪些局限性？在实现真正智能的道路上，它会在哪里碰壁？

So, you know, Michael, the you know, one of the other sort of larger questions that your work made me think about was, how much at odds the current paradigm of AI, like where we are with AI, is with the reality of biology. Biology really seems to matter, you know, more than we might understand in intelligence. So what does this tell you about the future of AI and what's really needed to create sort of the next generation of intelligent machines? And you can think of this question also, I guess, from the standpoint of, like, know, what are the current limitations of AI that that that you see? Where where will it run into a wall, you know, when it comes to intelligence?

首先，现有AI（即便是所谓'类脑'系统）与生物运作方式完全不同，这点毋庸置疑。但即便如此，也不能推论说这就不是智能，或会遭遇特定瓶颈。我不认为智能存在特权载体——盲目进化并非创造心智的唯一途径。

Well, a few things. First of all, I I think it's absolutely true that the current AI, even though even though with systems that are supposed to be neuromorphic and stuff like that, are not at all the, not at all the way that biology does things, completely different. However, having said that, I don't think you can conclude from this that that it is not intelligence or that it's going to run into some particular wall. I do not believe in any kind of privileged substrate for intelligence. I don't think that blind evolution has any monopoly on creating minds.

生物的物质载体对智能而言并非关键要素。我们完全可以用其他基质构建智能体。虽然我原本准备写篇论文列举当前AI与生物的五六点差异，但最终决定搁笔。这部分是因为我意识到（当然可能完全错误）：如果我的观点成立，将导致不仅能创造擅长解决问题的智能体，更会催生具有内在视角、需要道德考量的存在。我不愿为制造数万亿个我们无法保障其待遇的此类存在推波助澜。

I don't think there's anything about the material substrate of biology that is critical here. I don't see any reason why we couldn't build intelligent beings out of completely other substrates. Now having said that, again, I I actually started I I started writing a paper on here are the, I don't know, five or six ways in which current AI is is not biological and if and and and then I stopped and I'm I'm not gonna write it. It's it's somebody else will, so it's so it's not gonna be, know, help anything, but but but I'm not gonna be the one to do it because it occurred to me that to the extent, to whatever extent I'm right, and I could be completely wrong, but to to the to whatever extent I'm right, it's gonna lead to the creation of not just intelligent beings in the sense of, performing well on problem solving tasks, but actually beings with an inner perspective who matter in a moral sense. And I'm not interested in you know contributing to to trillions of those that we can't control how how they're gonna be treated.

总会有人去做这件事，这是无法阻止的。但关于AI我想强调的是：当前方法虽与生物学无关，却不能因此否定其真实性或智能本质。几个月前有位AI领域的重要人物对我说：'这些语言模型没什么神奇的，我亲手构建它们，完全清楚它们的运作。'

So somebody else will do it, you know, there's no way to stop it. But my point is my point about about AI is this, The current approach is absolutely not biological, but I don't think you can say from this that it isn't real or that it isn't really intelligence or that you know what it's doing. Somebody said to me a few months ago, I this is somebody who, you know, was a significant figure in in making these these things and he said, oh, these language models, there's nothing surprising going on there. I build them. I know I know what they do.

我回应道：'你连冒泡排序的真实行为都说不清，怎么可能宣称了解这些复杂系统？没错，代码是你写的，但历史早已证明：创造者未必真正理解自己的造物。'

I I I write I write this code. I I build them. And I I said, you you don't even know what bubble sort does. There's no way you can say that you know what these things do. Fine, you build them, but we already know that just because you build something doesn't mean that you understand what's really going on.

我们必须保持极大的谦逊，承认生物学家至今仍未理解生物基质以及生命意识从何而来。我们确实不理解，神经科学家也同样不理解。同样地，我们对计算基质的理解也不足以断言它能做什么或不能做什么。因此我认为必须对此保持谦卑——虽然我不认为当前的语言模型之类的东西与生物大脑有任何相似之处，但这并不意味着那里不存在意识，也不意味着意识不会在意想不到的地方涌现，毕竟我们对自己在做什么都还不清楚。

And we are, we really, there needs to be a lot of humility around the idea that the biologists do not understand the biological substrate and where the consciousness of living things come from. We do not, we absolutely do not, nor do the neuroscientists. And similarly, we do not understand the computational substrate enough to be able to say what it does and doesn't do. So I think we have to have a lot of humility about this. I don't think that certainly current language models and things like this, I do not believe they're anything like a biological mind, but that does not mean that there is no mind there or that mind isn't going to crop up in a surprising place because we don't know what we're doing yet.

我认为我们必须对这些事情保持非常开放的态度，对所有不确定性保持开放。那么界限在哪里？我...我真的不知道。过去几年AI发展最令人惊奇之处在于，它如何彻底将语言问题解决能力与曾经密不可分的生物进化史分离开来。过去，如果有东西对你说话，来源是显而易见的。

And I think we have to be very open about these things, about the uncertainty of all of these things. And so, you know, where is the limitation? I I I don't know. I think I think what's really interesting about and surprising about the the advances of AIs over the last few years is how far it's dissociated linguistic, problem solving from the typical biological history that always used to go with it. In the past, if something was talking to you, it was very obvious.

你过去能明确指出它的来源，知道它有生物进化背景。但现在我们已与之分离。那么问题解决能力强的环境下究竟会孕育出怎样的真实意识？我们根本不知道。任何声称自己确定知道AI模型在做什么的人——无论是否参与构建——我认为都是错误的。

You could say exactly where it came from, but you knew it had a biological history and so on. We now have dissociated with that. And, you know, to what extent does actual do do real minds emerge under circumstances that are good for problem solving? We simply do not know. Anybody who says they're certain about what these AI models are doing I think is wrong, whether you built them or not.

我们实际上并不知道答案，因为我们不理解基础认知的构成要素。这与大脑无关，与模仿或不模仿大脑无关。涌现的认知会出现在非常令人惊讶、非常微小的角落，而我们对此一无所知。

We actually do not know because we do not understand the ingredients of basal cognition. It has nothing to do with the brain, it has nothing to do with imitating or failing to imitate the brain. Emergent cognition shows up in very surprising, very minimal places and and we just don't know.

在我们结束前，这可能要转向科幻领域了——你如何看待AI与人类的未来？你预见到会出现混合体、赛博格这类形态吗？多种不同类型的...

You know, this might be going into science fiction as we as we wrap up, but what do you see as sort of the the the future of AI humans? Do do you see hybrids, cyborgs, that kind of stuff emerging, like multiple different kinds of

物种？我对此著述颇多。我认为'机器'这个概念、机器与有机体之间的区分终将被扫进历史垃圾堆。如果我们能活得足够久，发展顺利的话，我认为成熟物种将拥有我所谓的'具身自由'——你不必终生受限于进化偶然赐予的躯体，不必受那些宇宙射线是否击中关键DNA片段等随机事件导致的健康与智力限制。当我们真正掌握技术时（这本质上是再生医学的未来范式），每个人都能获得自己想要的任何身体形态。

species? I've I've written on this on this quite a bit. I think that the distinction between, I I think the word machine and the distinction between machines and organisms is gonna go in the dustbin of history. I think that we are ultimately assuming we live you know long enough, all goes well the mature species I think is going to have what I call freedom of embodiment, which is this idea that you don't just live your life in whatever body you happen to have been born into that was driven by the vagaries of evolution, the cosmic rays that hit, that did or did not hit some important DNA element or something else causing you to, you know, to have limitations of health and IQ. Once we really understand what we're doing, which is basically the future of the regenerative medicine, kind of, you know, deep regenerative medicine paradigm, everybody can have whatever embodiment they want.

这意味着各种感觉运动架构、进化材料与设计材料的任意组合——各种混合体、赛博格、嵌合体——我们将被难以置信的多样性生命形式包围，它们的身心远超达尔文所说的'最美丽的无尽形态'。自然界现有的一切多样性，在这个向我们敞开的可能性空间里不过是个小点——这才是我所见的未来图景。

This means whatever sensory motor architectures, whatever combination of evolved material, design material, you know, hybrids, cyborgs, chimeras of of every kind, going to be surrounded by an incredible variety of creatures with diverse bodies and minds that, go so far beyond, you know, Darwin's phrase of endless forms, the most beautiful. I mean, know, that all of all of the variety of nature is a tiny dot in this option space of what's open to us, and and that's that's what I see for the future.

迈克尔，我不知道时间都去哪儿了，一小时甚至更久就这么过去了。但我觉得我们只是触及了表面。有很多地方我本可以深入探讨，再聊个十到二十分钟。但我会先消化今天的对话内容，并考虑安排后续讨论，因为这个领域的深度和丰富性远超人们想象。不过真的非常感谢你抽时间与我们分享你的见解。

Michael, I don't know where the hour went, or hour plus went, you know, when the time's gone. You know, but I you know, part of me feels like, you know, we barely scratched the surface. There's so many places where I could have gone down and, you know, for another ten or twenty minutes. But I'm just gonna reflect on what we've talked about and, you know, think about a follow-up conversation to this because the space is just so much deeper and richer than people realize. But I really wanna thank you for your time, you know, and sharing your thinking with us.

这次对话非常愉快，拜读你的著作也让我受益匪浅。太精彩了。

Really enjoyed the conversation, enjoyed reading your your work. Fascinating.

非常感谢。很荣幸受邀。我很乐意再来交流。我认为你的观众群体，特别是从事计算机和信息技术工作的人，真的需要了解生物学领域正在发生的变革。

Thank you so much. Thank you for having me. Yeah. I'm happy to come back again. I think I think, you know, your audience and and especially people who work on, in computer computing and information technology really need to understand, the op what what's going on in biology and there's yeah.

未来蕴藏着不可思议的机遇，这些机遇将创造更美好的明天。谢谢。

The opportunities for the future are are and for for a better future are are incredible. So thank you.

确实如此。没错。我越来越深刻地认识到这一点。

Indeed. Indeed. Yeah. No. I'm I'm realizing that increasingly.

再次感谢你，迈克尔，我们保持联系。好的，非常感谢。

Well, thanks again, Michael, and we'll be in touch. Cool. Thanks very much.

嘿。

Hey.