书童按：本篇是Guillaume Verdon接受Lex Fridman播客采访实录的第二部分。延续上篇对有效加速主义（e/acc）哲学根基的探讨，本篇深入人与AI共生的未来图景、末日概率（p(doom)）的合理性辨析、量子机器学习的前沿探索等议题。Verdon以物理学家的视角，从热力学第二定律出发论证生命与文明的增长本性，主张人类应拥抱AI增强而非恐惧替代，批判末日论者对未来的偏见式采样，并分享量子计算与量子深度学习的技术洞见。访谈纵横于哲学思辨与技术前沿，既有对人类中心主义的解构，亦有对资本主义市场机制的坚守，视野开阔，发人深省。初稿采用Claude API机器翻译及排版，书童仅做简单校对及批注，以飨诸君。
Lex Fridman (00:50:13) 那么，如果事实证明，宇宙中意识之美的载体不止人类，AI也能将同样的火焰传承下去——这让你害怕吗？你担心AI会取代人类吗？
LEX FRIDMAN (00:50:13) So if it turns out that the beauty that is consciousness in the universe is bigger than just humans, the AI can carry that same flame forward. Does it scare you, are you concerned that AI will replace humans?
Guillaume Verdon (00:50:32) 在我的职业生涯中，有一个时刻让我意识到：也许我们需要把任务交给机器，才能真正理解我们周围的宇宙——而不是仅靠人类拿着纸笔把一切算出来。对我来说，这种放手一部分主动权的过程，反而给了我们理解世界的巨大杠杆。量子计算机在理解纳米尺度的物质方面，远胜过人类。类似地，我认为人类面临一个选择：我们是否接受AI将解锁的智力和操作杠杆，从而确保我们能够沿着文明规模与范围不断增长的道路前进？我们可能会被稀释——也许会有大量AI工作者——但总体而言，出于自身利益，通过与AI结合并增强自己，我们将实现更高的增长和更大的繁荣。
GUILLAUME VERDON (00:50:32) So during my career, I had a moment where I realized that maybe we need to offload to machines to truly understand the universe around us, right, instead of just having humans with pen and paper solve it all. And to me that sort of process of letting go of a bit of agency gave us way more leverage to understand the world around us. A quantum computer is much better than a human to understand matter at the Nanoscale. Similarly, I think that humanity has a choice, do we accept the opportunity to have intellectual and operational leverage that AI will unlock and thus ensure that we’re taken along this path of growth in the scope and scale of civilization? We may dilute ourselves, right? There might be a lot of workers that are AI, but overall out of our own self-interest, by combining and augmenting ourselves with AI, we’re going to achieve much higher growth and much more prosperity, right.
Guillaume Verdon (00:51:49) 对我而言，我认为最可能的未来是人类用AI增强自己。我认为我们已经走在这条增强之路上了——我们有手机用于通信，随时带在身上。我们有可穿戴设备，很快就会拥有与我们共享感知的设备，比如Humane AI Pin，或者说，从技术上讲，你的特斯拉汽车就具有共享感知能力。如果你们有共享的体验、共享的上下文、彼此通信并且有某种输入输出接口，那它本质上就是你自己的延伸。对我来说，人类用AI增强自己，以及那些不锚定于任何生物基质的AI，二者将会共存。而让各方利益对齐的方式——我们其实已经有了让由人类和技术组成的超级智能体对齐的机制。公司本质上是大型的混合专家模型，我们在公司内部有任务的神经路由机制，也有经济交换的方式来对齐这些庞然大物。
GUILLAUME VERDON (00:51:49) To me, I think that the most likely future is one where humans augment themselves with AI. I think we’re already on this path to augmentation, we have phones we use for communication, we have on ourselves at all times. We have wearables, soon that have shared perception with us, right, like the Humane AI Pin or I mean, technically your Tesla car has shared perception. And so if you have shared experience, shared context, you communicate with one another and you have some sort of IO, really it’s an extension of yourself.And to me, I think that humanity augmenting itself with AI and having AI that is not anchored to anything biological, both will coexist. And the way to align the parties, we already have a sort of mechanism to align super intelligences that are made of humans and technology, right? Companies are sort of large mixture of expert models, where we have neural routing of tasks within a company and we have ways of economic exchange to align these behemoths.
Guillaume Verdon (00:53:10) 对我来说，我认为资本主义就是那条路。我确实认为，无论是什么样的物质或信息配置，只要能带来最大化的增长，我们就会收敛到那里——这纯粹是物理原理使然。所以我们要么让自己与这个现实对齐，加入文明规模与范围加速扩张的进程；要么被甩在后面，试图减速，退回森林，放弃技术，回到原始状态。至少在我看来，这就是摆在面前的两条路。
GUILLAUME VERDON (00:53:10) And to me, I think capitalism is the way, and I do think that whatever configuration of matter or information leads to maximal growth, will be where we converge, just from like physical principles. And so we can either align ourselves to that reality and join the acceleration up in scope and scale of civilization or we can get left behind and try to decelerate and move back in the forest, let go of technology and return to our primitive state. And those are the two paths forward, at least to me.
Lex Fridman (00:53:54) 但有个哲学问题是：人类对齐能力是否存在极限？让我以一种论证的形式提出来。有个叫Dan Hendrycks的人写道，他同意你的观点，即AI的发展可以被视为一个进化过程，但对他——对Dan来说——这并不是件好事，因为他认为自然选择会偏好AI而非人类，这可能导致人类灭绝。你怎么看？如果这真是一个进化过程，而AI系统可能不需要人类呢？
LEX FRIDMAN (00:53:54) But there’s a philosophical question whether there’s a limit to the human capacity to align. So let me bring it up as a form of argument, this guy named Dan Hendrycks and he wrote that he agrees with you that AI development could be viewed as an evolutionary process, but to him, to Dan, this is not a good thing, as he argues that natural selection favors AIs over humans and this could lead to human extinction. What do you think, if it is an evolutionary process and AI systems may have no need for humans?
Guillaume Verdon (00:54:36) 我确实认为，我们实际上正在通过市场对AI的空间施加进化压力。现在我们运行那些对人类有正效用的AI，这就产生了选择压力——如果你认为当一个神经网络的API实例在GPU上运行时，它就”活着”的话。
GUILLAUME VERDON (00:54:36) I do think that we’re actually inducing an evolutionary process on the space of AIs through the market, right. Right now we run AIs that have positive utility to humans and that induces a selective pressure, if you consider a neural net being alive when there’s an API running instances of it on GPUs.
Lex Fridman (00:55:01) 对。
LEX FRIDMAN (00:55:01) Yeah.
Guillaume Verdon (00:55:01) 哪些API会被运行？那些对我们有高效用的。这就像我们驯化狼并把它们变成狗——狗的表达非常清晰，非常对齐。我认为我们有机会引导AI并实现高度对齐的AI。而且我认为人类加AI是一个非常强大的组合，我不确定纯粹的AI会淘汰这种组合。
GUILLAUME VERDON (00:55:01) Right. And which APIs get run? The ones that have high utility to us, right. So similar to how we domesticated wolves and turned them into dogs that are very clear in their expression, they’re very aligned, right. I think there’s going to be an opportunity to steer AI and achieve highly aligned AI. And I think that humans plus AI is a very powerful combination and it’s not clear to me that pure AI would select out that combination.
Lex Fridman (00:55:40) 所以人类现在正在创造选择压力，以创造与人类对齐的AI。但考虑到AI的发展方式以及它能多快地增长和扩展，对我来说，一个担忧是意外后果——人类无法预见这个过程的所有后果。AI系统可能造成的意外后果的破坏规模非常大。
LEX FRIDMAN (00:55:40) So the humans are creating the selection pressure right now to create AIs that are aligned to humans, but given how AI develops and how quickly it can grow and scale, to me, one of the concerns is unintended consequences, like humans are not able to anticipate all the consequences of this process. The scale of damage that could be done through unintended consequences with AI systems is very large.
Guillaume Verdon (00:56:10) 但上行空间的规模——
GUILLAUME VERDON (00:56:10) The scale of the upside.
Lex Fridman (00:56:12) 是的。
LEX FRIDMAN (00:56:12) Yes.
Guillaume Verdon (00:56:13) 对吧？
GUILLAUME VERDON (00:56:13) Right?
Lex Fridman (00:56:13) 我猜这是——
LEX FRIDMAN (00:56:13) Guess it’s-
Guillaume Verdon (00:56:14) 通过用AI增强我们自己，现在无法想象的上行空间。机会成本——我们正处在一个岔路口，对吧？我们要么走创造这些技术的道路，增强自己，在AI的帮助下攀登卡尔达肖夫等级（书童注：Kardashev Scale，衡量文明技术发展水平的量表，以能源利用能力为标准），成为多行星物种；要么我们完全不孕育这些技术，把所有潜在的上行空间都留在桌面上。
GUILLAUME VERDON (00:56:14) By augmenting ourselves with AI is unimaginable right now. The opportunity cost, we’re at a fork in the road, right? Whether we take the path of creating these technologies, augment ourselves and get to climb up the Kardashev Scale, become multi-planetary with the aid of AI, or we have a hard cutoff of like we don’t birth these technologies at all and then we leave all the potential upside on the table.
Lex Fridman (00:56:42) 对。
LEX FRIDMAN (00:56:42) Yeah.
Guillaume Verdon (00:56:42) 对我而言，出于对未来人类的责任——通过扩大文明规模，我们可以承载更多的人口——出于对这些未来人类的责任，我认为我们必须让那个更伟大、更宏大的未来成为现实。
GUILLAUME VERDON (00:56:42) Right. And to me, out of responsibility to the future humans we could carry, with higher carrying capacity by scaling up civilization. Out of responsibility to those humans, I think we have to make the greater grander future happen.
Lex Fridman (00:56:58) 在硬切断和全速前进之间，有中间地带吗？谨慎有任何论据吗？
LEX FRIDMAN (00:56:58) Is there a middle ground between cutoff and all systems go? Is there some argument for caution?
Guillaume Verdon (00:57:06) 我认为，正如我所说，市场会表现出谨慎。每个有机体、每家公司、每个消费者都在为自身利益行事，他们不会把资本分配给对他们有负效用的东西。
GUILLAUME VERDON (00:57:06) I think, like I said, the market will exhibit caution. Every organism, company, consumer is acting out of self-interest and they won’t assign capital to things that have negative utility to them.
Lex Fridman (00:57:21) 问题在于市场并不总是有完美信息，存在操纵，存在恶意行为者搅乱系统。它并不总是一个理性和诚实的系统。
LEX FRIDMAN (00:57:21) The problem is with the market is, there’s not always perfect information, there’s manipulation, there’s bad faith actors that mess with the system. It’s not always a rational and honest system.
Guillaume Verdon (00:57:41) 嗯，这正是为什么我们需要信息自由、言论自由和思想自由，以便能够收敛到对我们所有人都有正效用的技术子空间。
GUILLAUME VERDON (00:57:41) Well, that’s why we need freedom of information, freedom of speech and freedom of thought in order to be able to converge on the subspace of technologies that have positive utility for us all, right.
Lex Fridman (00:57:56) 那让我问你关于p(doom)的问题——末日概率。这个词说起来挺有意思，但经历起来可不有趣。在你看来，AI最终杀死全部或大部分人类的概率是多少——也就是所谓的末日概率？
LEX FRIDMAN (00:57:56) Well let me ask you about p(doom), probability of doom. That’s just fun to say, but not fun to experience. What is to you the probability that AI eventually kills all or most humans, also known as probability of doom?
Guillaume Verdon (00:58:16) 我不喜欢那种计算方式。我认为人们只是随便抛出数字，这是非常草率的计算。要计算概率，比方说你把世界建模为某种马尔可夫过程，如果你有足够多的变量，或者隐马尔可夫过程——你需要对所有可能的未来空间做随机路径积分，而不仅仅是你的大脑自然倾向的那些未来。我认为p(doom)的估算者是有偏见的，因为我们的生物本性。我们进化出了对负面的、可怕的未来的偏见采样，因为那是进化的最优解。所以那些神经质程度较高的人，整天每天都在想一切都会出错的负面未来，并声称他们在做无偏采样。某种意义上，他们没有对所有可能性的空间做归一化处理，而所有可能性的空间是超指数级庞大的，很难有这样的估计。
GUILLAUME VERDON (00:58:16) I’m not a fan of that calculation, I think people just throw numbers out there and it’s a very sloppy calculation, right? To calculate a probability, let’s say you model the world as some sort of Markov process, if you have enough variables or hidden Markov process. You need to do a stochastic path integral through the space of all possible futures, not just the futures that your brain naturally steers towards, right. I think that the estimators of p(doom) are biased because of our biology, right? We’ve evolved to have bias sampling towards negative futures that are scary, because that was an evolutionary optimum, right. And so people that are of, let’s say higher neuroticism will just think of negative futures where everything goes wrong all day every day and claim that they’re doing unbiased sampling. And in a sense they’re not normalizing for the space of all possibilities and the space of all possibilities is super exponentially large and it’s very hard to have this estimate.
Guillaume Verdon (00:59:40) 总的来说，我认为我们无法以那样的粒度预测未来，因为混沌。如果你有一个复杂系统，你在几个变量上有一些不确定性，如果你让时间演化，你就有了李雅普诺夫指数（Lyapunov exponent）这个概念。一点点模糊会在我们的估计中呈指数级地变成大量模糊，随着时间推移。我认为我们需要表现出一些谦逊，承认我们实际上无法预测未来。我们拥有的唯一先验是物理定律，这正是我们所主张的。物理定律说，系统会想要增长，而为增长和复制而优化的子系统在未来更有可能出现。所以我们应该力求最大化我们当前与未来的互信息，而通往那条路的方式是加速而非减速。
GUILLAUME VERDON (00:59:40) And in general, I don’t think that we can predict the future with that much granularity because of chaos, right? If you have a complex system, you have some uncertainty and a couple of variables, if you let time evolve, you have this concept of a Lyapunov exponent, right. A bit of fuzz becomes a lot of fuzz in our estimate, exponentially so, over time. And I think we need to show some humility that we can’t actually predict the future, the only prior we have is the laws of physics, and that’s what we’re arguing for. The laws of physics say the system will want to grow and subsystems that are optimized for growth and replication are more likely in the future. And so we should aim to maximize our current mutual information with the future and the path towards that is for us to accelerate rather than decelerate.
Guillaume Verdon (01:00:40) 所以我没有p(doom)，因为我认为，类似于谷歌的量子霸权实验——我当时就在他们运行模拟的房间里——那是一个量子混沌系统的例子，你甚至无法用世界上最大的超级计算机估算某些结果的概率。那就是混沌的一个例子，而我认为这个系统对任何人来说都过于混沌，无法对某些未来的可能性有准确的估计。如果他们真有那么厉害，我想他们在股市交易上会非常富有。
GUILLAUME VERDON (01:00:40) So I don’t have a p(doom), because I think that similar to the quantum supremacy experiment at Google, I was in the room when they were running the simulations for that. That was an example of a quantum chaotic system where you cannot even estimate probabilities of certain outcomes with even the biggest supercomputer in the world, right. So that’s an example of chaos and I think the system is far too chaotic for anybody to have an accurate estimate of the likelihood of certain futures. If they were that good, I think they would be very rich trading on the stock market.
Lex Fridman (01:01:23) 但话虽如此，人类确实有偏见，根植于我们的进化生物学，害怕一切能杀死我们的东西，但我们仍然可以想象能杀死我们的不同轨迹。我们不知道所有其他不一定会的轨迹，但我认为，结合一些基于人类历史的基本直觉来推理，仍然是有用的——比如看看地缘政治，看看人性的基本面，强大的技术如何能伤害很多人？这似乎基于此，看看核武器，你可以开始估算p(doom)，也许是在更哲学的意义上，而非数学意义上。哲学意义上是指：有这种可能性吗？人性倾向于那个方向吗？
LEX FRIDMAN (01:01:23) But nevertheless, it’s true that humans are biased, grounded in our evolutionary biology, scared of everything that can kill us, but we can still imagine different trajectories that can kill us. We don’t know all the other ones that don’t necessarily, but it’s still I think, useful combined with some basic intuition grounded in human history, to reason about like what… Like looking at geopolitics, looking at basics of human nature, how can powerful technology hurt a lot of people? It just seems grounded in that, looking at nuclear weapons, you can start to estimate p(doom) maybe in a more philosophical sense, not a mathematical one. Philosophical meaning like is there a chance? Does human nature tend towards that or not?
Guillaume Verdon (01:02:25) 我认为，对我来说，最大的存在风险之一是AI的权力集中在极少数人手中，尤其是如果这是控制信息流的公司和政府的混合体。因为这可能为一种反乌托邦的未来铺平道路——只有极少数人和政府中的寡头拥有AI，他们甚至可以说服公众AI从未存在过。这就开启了威权集中控制的场景，对我来说，这是最黑暗的时间线。而现实是，我们有这些事情发生的数据驱动先验。当你给予太多权力，当你过度集中权力时，人类会做可怕的事情。
GUILLAUME VERDON (01:02:25) I think to me, one of the biggest existential risks would be the concentration of the power of AI in the hands of the very few, especially if it’s a mix between the companies that control the flow of information and the government. Because that could set things up for a sort of dystopian future where only a very few and an oligopoly in the government have AI and they could even convince the public that AI never existed. And that opens up sort of these scenarios for authoritarian centralized control, which to me is the darkest timeline. And the reality is that we have a data-driven prior of these things happening, right. When you give too much power, when you centralize power too much, humans do horrible things, right.
Guillaume Verdon (01:03:23) 对我来说，在我的贝叶斯推断中，这比基于科幻的先验有更高的可能性——比如”我的先验来自《终结者》电影”。所以当我和这些AI末日论者交谈时，我只是要求他们追溯一条通过马尔可夫链事件的路径，这条路径会导致我们的末日，并实际给我每次转换的良好概率。而很多时候，那条链中会有一个非物理的或极不可能的转换。但当然，我们天生就会害怕事物，我们天生会对危险做出反应，我们天生会认为未知是危险的，因为这是生存的好启发式方法。但出于恐惧，我们有更多的损失。我们有太多要失去的，太多的上行空间会因为出于恐惧而预先阻止正面未来的发生而失去。所以我认为我们不应该屈服于恐惧。恐惧是心智的杀手，我认为它也是文明的杀手。
GUILLAUME VERDON (01:03:23) And to me, that has a much higher likelihood in my Bayesian inference than Sci-Fi based priors, right, like, “My prior came from the Terminator movie.” And so when I talked to these AI doomers, I just ask them to trace a path through this Markov chain of events that would lead to our doom and to actually give me a good probability for each transition. And very often there’s a unphysical or highly unlikely transition in that chain, right. But of course, we’re wired to fear things and we’re wired to respond to danger, and we’re wired to deem the unknown to be dangerous, because that’s a good heuristic for survival, right. But there’s much more to lose out of fear. We have so much to lose, so much upside to lose by preemptively stopping the positive futures from happening out of fear. And so I think that we shouldn’t give into fear, fear is the mind killer, I think it’s also the civilization killer.
Lex Fridman (01:04:43) 我们仍然可以思考事情出错的各种方式。比如，美国的开国元勋们思考了人性，这就是为什么会有关于必要自由的讨论。他们真正深入地审议了这一点，我认为同样的事情可能也可以为AGI做。人类历史确实表明我们倾向于集中化，或者至少当我们实现集中化时，很多坏事会发生。当有独裁者时，很多黑暗、糟糕的事情会发生。问题是，AGI能成为那个独裁者吗？AGI在发展时，能否因为其权力而成为集中化者？也许是因为人类的对齐，也许是同样的倾向，同样的斯大林式集中化和集中管理资源分配的倾向？
LEX FRIDMAN (01:04:43) We can still think about the various ways things go wrong, for example, the founding fathers of the United States thought about human nature and that’s why there’s a discussion about the freedoms that are necessary. They really deeply deliberated about that and I think the same could possibly be done for AGI. It is true that human history shows that we tend towards centralization, or at least when we achieve centralization, a lot of bad stuff happens. When there’s a dictator, a lot of dark, bad things happen. The question is, can AGI become that dictator? Can AGI when develop, become the centralizer, because of its power? Maybe because of the alignment of humans, perhaps, the same tendencies, the same Stalin like tendencies to centralize and manage centrally the allocation of resources?
Lex Fridman (01:05:45) 你甚至可以看到这在表面上是一个令人信服的论点：”嗯，AGI如此聪明，如此高效，如此擅长分配资源，我们为什么不把它外包给AGI呢？”然后最终，无论什么力量用权力腐蚀人类的心智，都可能对AGI做同样的事。它只会说：”好吧，人类是可有可无的，我们会摆脱他们。”就像乔纳森·斯威夫特（Jonathan Swift）几个世纪前——我想是1700年代——的《一个温和的建议》（A Modest Proposal），他讽刺性地建议，我想是在爱尔兰，穷人的孩子被作为食物喂给富人，这将是个好主意，因为它减少了穷人的数量，并给穷人带来额外收入。所以从几个方面减少了穷人的数量，因此更多的人变得富有。当然，它漏掉了一个很难放入数学方程的基本部分——人类生命的基本价值。所以，这一切都是在说，你担心AGI成为你刚才谈到的权力集中者吗？
LEX FRIDMAN (01:05:45) And you can even see that as a compelling argument on the surface level. “Well, AGI is so much smarter, so much more efficient, so much better at allocating resources, why don’t we outsource it to the AGI?” And then eventually whatever forces that corrupt the human mind with power could do the same for AGI. It’ll just say, “Well, humans are dispensable, we’ll get rid of them.” Do the Jonathan Swift, Modest Proposal from a few centuries ago, I think the 1700s, when he satirically suggested that, I think it’s in Ireland, that the children of poor people are fed as food to the rich people and that would be a good idea, because it decreases the amount of poor people and gives extra income to the poor people. So on several accounts decreases the amount of poor people, therefore more people become rich. Of course, it misses a fundamental piece here that’s hard to put into a mathematical equation of the basic value of human life. So all of that to say, are you concerned about AGI being the very centralizer of power that you just talked about?
Guillaume Verdon (01:07:09) 我确实认为，现在AI有向集中化的偏见，因为计算密度和数据的集中化以及我们训练模型的方式。我认为随着时间推移，我们将耗尽可以从互联网上抓取的数据，而且我正在研究提高计算密度，以便计算可以无处不在，以分布式方式在环境中获取信息并测试假设。我认为从根本上说，集中式控制论控制——也就是拥有一个庞大的智能体，融合许多传感器，试图准确感知世界、准确预测它、预测许多许多变量并控制它、对世界施加其意志——我认为这从来就不是最优解。比方说你有一家公司，如果你有一家公司，我不知道，有10000人，他们都向CEO汇报。即使那个CEO是AI，我认为它也会努力融合所有传来的信息，然后预测整个系统，然后施行其意志。
GUILLAUME VERDON (01:07:09) I do think that right now there’s a bias over a centralization of AI, because of a compute density and centralization of data and how we’re training models. I think over time we’re going to run out of data to scrape over the internet, and I think that, well, actually I’m working on, increasing the compute density so that compute can be everywhere and acquire information and test hypotheses in the environment in a distributed fashion. I think that fundamentally, centralized cybernetic control, so having one intelligence that is massive that fuses many sensors and is trying to perceive the world accurately, predict it accurately, predict many, many variables and control it, enact its will upon the world, I think that’s just never been the optimum, right? Like let’s say you have a company, if you have a company, I don’t know, of 10,000 people, they all report to the CEO. Even if that CEO is an AI, I think it would struggle to fuse all of the information that is coming to it and then predict the whole system and then to enact its will.
Guillaume Verdon (01:08:28) 在自然界、在公司以及各种系统中出现的，是一种分层控制论控制的概念。在公司里，你有个人贡献者，他们为自己的利益行事，试图完成他们的任务，他们有一个精细的——就时间和空间而言——控制回路和感知领域。比如说你在一家软件公司，他们有自己的代码库，他们在一天内迭代它。然后管理层可能会检查，它有更广的范围，比方说有五个直接汇报对象。然后它每周对每个人的更新采样一次，然后你可以沿着链条向上，你有更大的时间尺度和更大的范围。而这似乎已经成为控制系统的最佳方式。
GUILLAUME VERDON (01:08:28) What has emerged in nature and in corporations and all sorts of systems is a notion of sort of hierarchical cybernetic control, right. In a company it would be, you have like the individual contributors, they are self-interested and they’re trying to achieve their tasks and they have a fine, in terms of time and space if you will, control loop and field of perception, right. They have their code base, let’s say you’re in a software company, they have their code base, they iterate it on it intraday, right. And then the management maybe checks in, it has a wider scope, it has, let’s say five reports, right. And then it samples each person’s update once per week, and then you can go up the chain and you have larger timescale and greater scope. And that seems to have emerged as sort of the optimal way to control systems.
Guillaume Verdon (01:09:25) 而这正是资本主义给我们的。你有这些层级结构，你甚至可以有母公司等等。这样容错性要强得多。在量子计算中——这是我的领域出身——我们有量子纠错中的容错概念。量子纠错是检测来自噪声的故障，预测它如何在系统中传播，然后纠正它——这是一个控制论回路。事实证明，分层的解码器，并且在每个层级都是局部的——
GUILLAUME VERDON (01:09:25) And really that’s what capitalism gives us, right? You have these hierarchies and you can even have like parent companies and so on. And so that is far more fault tolerant, in quantum computing, that’s my feel that came from, we have a concept of this fault tolerance in quantum air correction, right? Quantum air correction is detecting a fault that came from noise, predicting how it’s propagated through the system and then correcting it, right, so it’s a cybernetic loop. And it turns out that decoders that are hierarchical and in each level, the hierarchy are local-
Guillaume Verdon (01:10:00) ——分层的，并且每个层级都是局部的，表现要好得多，而且容错性要强得多。原因是，如果你有一个非局部的解码器，那么你在这个控制节点上有一个故障,整个系统就会崩溃。类似地，如果你有一个每个人都向其汇报的CEO，而那个CEO去度假了，整个公司就会陷入停滞。对我来说，我认为是的，我们看到AI有集中化的趋势，但我认为随着时间推移会有修正，智能会更接近感知。我们将把AI分解成更小的子系统，彼此通信并形成一个元系统。
GUILLAUME VERDON (01:10:00) … that are hierarchical. And at each level, the hierarchy are local, perform the best by far, and are far more fault-tolerant. The reason is, if you have a non-local decoder, then you have one fault at this control node and the whole system crashes. Similarly to if you have one CEO that everybody reports to and that CEO goes on vacation, the whole company comes to a crawl. To me, I think that yes, we’re seeing a tendency towards centralization of AI, but I think there’s going to be a correction over time, where intelligence is going to go closer to the perception. And we’re going to break up AI into smaller subsystems that communicate with one another and form a meta system.
Lex Fridman (01:10:56) 如果你看看今天世界上的层级结构，有国家，那些都是层级的。但相对于彼此，国家是无政府的，所以这是一种无政府状态。
LEX FRIDMAN (01:10:56) If you look at the hierarchies that are in the world today, there’s nations and those all hierarchical. But in relation to each other, nations are anarchic, so it’s an anarchy.
Guillaume Verdon (01:11:06) 嗯。
GUILLAUME VERDON (01:11:06) Mm-hmm.
Lex Fridman (01:11:08) 你预见这样一个世界吗，在那里没有一个总体的……你怎么称呼它？集中式控制论控制？
LEX FRIDMAN (01:11:08) Do you foresee a world like this, where there’s not a over… What’d you call it? A centralized cybernetic control?
Guillaume Verdon (01:11:17) 集中式控制中心。对。
GUILLAUME VERDON (01:11:17) Centralized locus of control. Yeah.
Lex Fridman (01:11:21) 你说那是次优的？
LEX FRIDMAN (01:11:21) That’s suboptimal, you’re saying?
Guillaume Verdon (01:11:22) 对。
GUILLAUME VERDON (01:11:22) Yeah.
Lex Fridman (01:11:23) 所以，在最顶层总会有竞争状态？
LEX FRIDMAN (01:11:23) So, it would be always a state of competition at the very top level?
Guillaume Verdon (01:11:27) 对。就像在公司里，你可能有两个部门在做类似的技术并相互竞争，然后你剪掉表现不佳的那个。这是一个树的选择过程，或者一个产品被砍掉，然后整个组织被解雇。这个尝试新事物和淘汰不奏效的旧事物的过程，正是给我们适应性的东西，帮助我们收敛到最好的技术和最该做的事情。
GUILLAUME VERDON (01:11:27) Yeah. Yeah. Just like in a company, you may have two units working on similar technology and competing with one another, and you prune the one that performs not as well. That’s a selection process for a tree, or a product gets killed and then a whole org gets fired. This process of trying new things and shedding old things that didn’t work, it’s what gives us adaptability and helps us converge on the technologies and things to do that are most good.
Lex Fridman (01:12:04) 我只是希望没有一种对AGI独特而对人类不独特的失败模式，因为你现在主要描述的是人类系统。
LEX FRIDMAN (01:12:04) I just hope there’s not a failure mode that’s unique to AGI versus humans, because you’re describing human systems mostly right now.
Guillaume Verdon (01:12:11) 对。
GUILLAUME VERDON (01:12:11) Right.
Lex Fridman (01:12:11) 我只是希望当一家公司垄断AGI时，我们会看到与人类相同的情况，也就是另一家公司会涌现出来并开始有效竞争。
LEX FRIDMAN (01:12:11) I just hope when there’s a monopoly on AGI in one company, that we’ll see the same thing we see with humans, which is, another company will spring up and start competing effectively.
Guillaume Verdon (01:12:24) 到目前为止一直是这样。我们有OpenAI。我们有Anthropic。现在，我们有xAI。我们有Meta，甚至是开源的，现在我们有Mistral，它非常有竞争力。这就是资本主义的美妙之处。你不必过于信任任何一方，因为我们总是在每个层面对冲我们的赌注。总有竞争，这对我来说至少是最美好的事情，就是整个系统总是在转变，总是在适应。
GUILLAUME VERDON (01:12:24) That’s been the case so far. We have OpenAI. We have Anthropic. Now, we have xAI. We have Meta even for open source, and now we have Mistral, which is highly competitive. That’s the beauty of capitalism. You don’t have to trust any one party too much because we’re always hedging our bets at every level. There’s always competition and that’s the most beautiful thing to me, at least, is that the whole system is always shifting and always adapting.
Guillaume Verdon (01:12:54) 维持这种活力就是我们避免暴政的方式。确保每个人都能访问这些工具、这些模型，并能为研究做出贡献，就能避免智能暴政——极少数人控制世界的AI并用它来压迫周围的人。
GUILLAUME VERDON (01:12:54) Maintaining that dynamism is how we avoid tyranny. Making sure that everyone has access to these tools, to these models, and can contribute to the research, avoids a neural tyranny where very few people have control over AI for the world and use it to oppress those around them.
Lex Fridman (01:13:23) 当你谈论智能时,你提到了多体量子纠缠。
LEX FRIDMAN (01:13:23) When you were talking about intelligence, you mentioned multipartite quantum entanglement.
Guillaume Verdon (01:13:28) 嗯。
GUILLAUME VERDON (01:13:28) Mm-hmm.
Lex Fridman (01:13:29) 先问一个高层次的问题：你认为什么是智能？当你思考量子力学系统并观察其中发生的某种计算时，你认为宇宙能够进行的那种计算有什么智能之处？而人类大脑能够进行的计算只是其中的一小部分？
LEX FRIDMAN (01:13:29) High-level question first is, what do you think is intelligence? When you think about quantum mechanical systems and you observe some kind of computation happening in them, what do you think is intelligent about the kind of computation the universe is able to do; a small, small inkling of which is the kind of computation a human brain is able to do?
Guillaume Verdon (01:13:52) 我会说智能和计算并不完全是一回事。我认为宇宙确实在进行量子计算。如果你能访问所有自由度和一台非常非常非常大的量子计算机，有很多很多量子比特，比方说，每个普朗克体积有几个量子比特——这差不多是我们拥有的像素——那么你就能在一台足够大的量子计算机上模拟整个宇宙，当然，假设你看的是宇宙的有限体积。我认为至少对我来说，智能是——我回到控制论——感知、预测和控制我们世界的能力。
GUILLAUME VERDON (01:13:52) I would say intelligence and computation aren’t quite the same thing. I think that the universe is very much doing a quantum computation. If you had access to all the degrees of freedom and a very, very, very large quantum computer with many, many, many qubits, let’s say, a few qubits per Planck volume, which is more or less the pixels we have, then you’d be able to simulate the whole universe on a sufficiently large quantum computer, assuming you’re looking at a finite volume, of course, of the universe. I think that at least to me, intelligence is, I go back to cybernetics, the ability to perceive, predict, and control our world.
Guillaume Verdon (01:14:46) 但实际上，现在看来，我们使用的很多智能更多是关于压缩。它是关于操作化信息论。在信息论中，你有分布或系统的熵的概念，熵告诉你，如果你有最优代码，你需要这么多比特来编码这个分布或这个子系统。AI，至少我们今天为LLM和量子所做的方式，非常像试图最小化我们的世界模型与世界之间、与来自世界的分布之间的相对熵。我们在学习，我们在计算空间中搜索以处理世界，以找到那个已经提炼出所有方差、噪声和熵的压缩表示。
GUILLAUME VERDON (01:14:46) But really, nowadays, it seems like a lot of intelligence we use is more about compression. It’s about operationalizing information theory. In information theory, you have the notion of entropy of a distribution or a system, and entropy tells you that you need this many bits to encode this distribution or this subsystem, if you have the most optimal code. AI, at least the way we do it today for LLMs and for quantum, is very much trying to minimize relative entropy between our models of the world and the world, distributions from the world. We’re learning, we’re searching over the space of computations to process the world, to find that compressed representation that has distilled all the variance in noise and entropy.
Guillaume Verdon (01:15:58) 最初，我从黑洞研究进入量子机器学习，因为黑洞的熵非常有趣。某种意义上，它们在物理上是宇宙中密度最高的物体。你无法在空间上比黑洞更密集地打包更多信息。所以我在想，黑洞实际上是如何编码信息的？它们的压缩代码是什么？这让我进入了算法空间，搜索量子代码空间。它也让我实际进入了，你如何从世界获取量子信息？我做过的一些工作，现在是公开的，是量子模数转换。
GUILLAUME VERDON (01:15:58) Originally, I came to quantum machine learning from the study of black holes because the entropy of black holes is very interesting. In a sense, they’re physically the most dense objects in the universe. You can’t pack more information spatially any more densely than in a black hole. And so, I was wondering, how do black holes actually encode information? What is their compression code? That got me into the space of algorithms, to search over space of quantum codes. It got me actually into also, how do you acquire quantum information from the world? Something I’ve worked on, this is public now, is quantum analog digital conversion.
Guillaume Verdon (01:16:50) 你如何从真实世界以叠加态捕获信息而不破坏叠加态，而是为量子计算机数字化来自真实世界的信息？如果你有能力捕获量子信息并学习它的表示，现在你就可以学习可能在其潜在表示中有一些有用信息的压缩表示。我认为我们文明面临的许多问题实际上都超越了这个复杂性障碍。温室效应是一种量子力学效应。化学是量子力学的。核物理是量子力学的。
GUILLAUME VERDON (01:16:50) How do you capture information from the real world in superposition and not destroy the superposition, but digitize for a quantum mechanical computer information from the real world? If you have an ability to capture quantum information and learn representation representations of it, now you can learn compressed representations that may have some useful information in their latent representation. I think that many of the problems facing our civilization are actually beyond this complexity barrier. The greenhouse effect is a quantum mechanical effect. Chemistry is quantum mechanical. Nuclear physics is quantum mechanical.
Guillaume Verdon (01:17:43) 很多生物学、蛋白质折叠等都受量子力学影响。所以，解锁用量子计算机和量子AI增强人类智力的能力，对我来说似乎是文明需要发展的基本能力。我花了几年时间做这个，但随着时间推移，我对开始看起来像核聚变的时间线感到厌倦。
GUILLAUME VERDON (01:17:43) A lot of biology and protein folding and so on is affected by quantum mechanics. And so, unlocking an ability to augment human intellect with quantum mechanical computers and quantum mechanical AI seemed to me like a fundamental capability for civilization that we needed to develop. I spent several years doing that, but over time, I grew weary of the timelines that were starting to look like nuclear fusion.
Lex Fridman (01:18:17) 我可以问一个高层次的问题，也许通过定义的方式，通过解释的方式：什么是量子计算机，什么是量子机器学习？
LEX FRIDMAN (01:18:17) One high-level question I can ask is maybe by way of definition, by way of explanation, what is a quantum computer and what is quantum machine learning?
Guillaume Verdon (01:18:27) 量子计算机实际上就是一个量子力学系统，我们对它有足够的控制，它可以保持其量子力学状态。量子力学是自然界在非常小的尺度上的行为方式，当事物非常小或非常冷时，它实际上比概率论更基础。我们习惯于事物是这个或那个，但我们不习惯用叠加态思考，因为，嗯，我们的大脑做不到。所以，我们必须把量子力学世界翻译成，比如说，线性代数来理解它。不幸的是，这种翻译平均而言是指数级低效的。你必须用非常大的矩阵来表示事物。但实际上，你可以用很多东西制造量子计算机，我们已经看到各种各样的玩家，从中性原子、囚禁离子、超导金属光子，在不同频率上。
GUILLAUME VERDON (01:18:27) A quantum computer really is a quantum mechanical system, over which we have sufficient control, and it can maintain its quantum mechanical state. And quantum mechanics is how nature behaves at the very small scales, when things are very small or very cold, and it’s actually more fundamental than probability theory. We’re used to things being this or that, but we’re not used to thinking in superpositions because, well, our brains can’t do that. So, we have to translate the quantum mechanical world to, say, linear algebra to grok it. Unfortunately, that translation is exponentially inefficient on average. You have to represent things with very large matrices. But really, you can make a quantum computer out of many things, and we’ve seen all sorts of players, from neutral atoms, trapped ions, superconducting metal photons at different frequencies.
Guillaume Verdon (01:19:38) 我认为你可以用很多东西制造量子计算机。但对我来说，真正有趣的是量子机器学习既是关于用量子计算机理解量子力学世界，所以把物理世界嵌入AI表示，也是量子计算机工程是把AI算法嵌入物理世界。把物理世界嵌入AI、把AI嵌入物理世界的这种双向性，物理和AI之间的这种共生关系，实际上这就是我追求的核心，即使到今天，在量子计算之后。它仍然在这个将物理和AI真正融合的旅程中。
GUILLAUME VERDON (01:19:38) I think you could make a quantum computer out of many things. But to me, the thing that was really interesting was both quantum machine learning was about understanding the quantum mechanical world with quantum computers, so embedding the physical world into AI representations, and quantum computer engineering was embedding AI algorithms into the physical world. This bi-directionality of embedding physical world into AI, AI into the physical world, this symbiosis between physics and AI, really that’s the core of my quest really, even to this day, after quantum computing. It’s still in this journey to merge really physics and AI.
Lex Fridman (01:20:29) 量子机器学习是一种在保持自然的量子力学方面真实的自然表示上进行机器学习的方式？
LEX FRIDMAN (01:20:29) Quantum machine learning is a way to do machine learning on a representation of nature that stays true to the quantum mechanical aspect of nature?
Guillaume Verdon (01:20:43) 对，它是学习量子力学表示。那将是量子深度学习。或者，你可以尝试在量子计算机上做经典机器学习。我不建议这样做，因为你可能会有一些加速，但很多时候，加速伴随着巨大的成本。使用量子计算机非常昂贵。
GUILLAUME VERDON (01:20:43) Yeah, it’s learning quantum mechanical representations. That would be quantum deep learning. Alternatively, you can try to do classical machine learning on a quantum computer. I wouldn’t advise it because you may have some speed-ups, but very often, the speed-ups come with huge costs. Using a quantum computer is very expensive.
Guillaume Verdon (01:21:08) 为什么？因为你假设计算机在绝对零度下运行,而宇宙中没有物理系统能达到那个温度。你必须做的是我一直提到的，这个量子纠错过程，它实际上是一个算法冰箱。它试图把熵从系统中抽出来，试图让它更接近0K。当你计算在量子计算机上做深度学习需要多少资源时，比如说，做经典深度学习，会有如此巨大的开销，不值得。这就像考虑用火箭穿越城市，进入轨道再返回来运送东西。没有意义。就用送货卡车。
GUILLAUME VERDON (01:21:08) Why is that? Because you assume the computer is operating at zero temperature, which no physical system in the universe can achieve that temperature. What you have to do is what I’ve been mentioning, this quantum error correction process, which is really an algorithmic fridge. It’s trying to pump entropy out of the system, trying to get it closer to zero temperature. When you do the calculations of how many resources it would take to, say, do deep learning on a quantum computer, classical deep learning, there’s such a huge overhead, it’s not worth it. It’s like thinking about shipping something across a city using a rocket and going to orbit and back. It doesn’t make sense. Just use a delivery truck.
Lex Fridman (01:21:53) 你能用量子深度学习弄清楚、预测、理解什么样的东西,而用深度学习做不到？所以，将量子力学系统纳入学习过程？
LEX FRIDMAN (01:21:53) What kind of stuff can you figure out, can you predict, can you understand with quantum deep learning that you can’t with deep learning? So, incorporating quantum mechanical systems into the learning process?
Guillaume Verdon (01:22:05) 我认为这是一个很好的问题。从根本上说，任何具有足够量子力学关联、对经典表示来说很难捕获的系统，量子力学表示应该比纯经典表示有优势。问题是，哪些系统有足够的、非常量子的关联？但这也是，哪些系统仍然与工业相关？这是一个大问题。人们倾向于化学、核物理。我实际上从事过处理来自量子传感器的输入。如果你有一个量子传感器网络，它们捕获了世界的量子力学图像，以及如何后处理，那就成为一种量子形式的机器感知。例如，费米实验室有一个项目探索用这些量子传感器探测暗物质。对我来说，这与我从小就想理解宇宙的追求是一致的。所以，有一天，我希望我们能有非常大的量子传感器网络，帮助我们窥视宇宙的最早期部分。例如，LIGO是一个量子传感器。它只是一个非常大的。所以，是的，我会说量子机器感知、模拟、理解量子模拟，类似于AlphaFold。AlphaFold理解了蛋白质配置的概率分布。你可以用量子机器学习更有效地理解电子配置的量子分布。
GUILLAUME VERDON (01:22:05) I think that’s a great question. Fundamentally, it’s any system that has sufficient quantum mechanical correlations that are very hard to capture for classical representations. Then, there should be an advantage for a quantum mechanical representation over a purely classical one. The question is, which systems have sufficient correlations that are very quantum? But it’s also, which systems are still relevant to industry? That’s a big question. People are leaning towards chemistry, nuclear physics. I’ve worked on actually processing inputs from quantum sensors. If you have a network of quantum sensors, they’ve captured a quantum mechanical image of the world and how to post-process that, that becomes a quantum form of machine perception. For example, Fermilab has a project exploring detecting dark matter with these quantum sensors. To me, that’s in alignment with my quest to understand the universe ever since I was a child. And so, someday, I hope that we can have very large networks of quantum sensors that help us peer into the earliest parts of the universe. For example, the LIGO is a quantum sensor. It’s just a very large one. So, yeah, I would say quantum machine perception, simulations, grokking quantum simulations, similar to AlphaFold. AlphaFold understood the probability distribution over configurations of proteins. You can understand quantum distributions over configurations of electrons more efficiently with quantum machine learning.
Lex Fridman (01:23:53) 你合著了一篇题为《量子深度学习的通用训练算法》的论文。那涉及Baqprop，带Q。做得很好，先生。做得很好。它是如何工作的？你能提一些有趣的方面吗，Baqprop以及我们为经典机器学习所知的一些东西如何转移到量子机器学习？
LEX FRIDMAN (01:23:53) You co-authored a paper titled A Universal Training Algorithm for Quantum Deep Learning. That involves Baqprop, with a Q. Very well done, sir. Very well done. How does it work? Is there some interesting aspects you can just mention on how Baqprop and some of these things we know for classical machine learning transfer over to the quantum machine learning?
Guillaume Verdon (01:24:19) 是的。那是一篇古怪的论文。那是我在量子深度学习领域的第一批论文之一。每个人都在说：”哦，我认为深度学习会被量子计算机加速。”我说：”好吧，预测未来的最好方法就是发明它。所以，这里有一篇100页的论文，祝你愉快。”本质上，量子计算通常是，你把可逆操作嵌入量子计算。
GUILLAUME VERDON (01:24:19) Yeah. That was a funky paper. That was one of my first papers in quantum deep learning. Everybody was saying, “Oh, I think deep learning is going to be sped up by quantum computers.” I was like, “ Well, the best way to predict the future is to invent it. So, here’s a 100-page paper, have fun.” Essentially, quantum computing is usually, you embed reversible operations into a quantum computation.
Guillaume Verdon (01:24:47) 那里的技巧是做一个前馈操作并做我们所说的相位踢（phase kick）。但实际上，它只是一个力踢（force kick）。你只是用与你希望优化的损失函数成正比的某种力踢系统。然后，通过执行反计算，你从参数的叠加态开始，这相当古怪。现在，你不只是有参数的一个点，你有许多潜在参数的叠加态。我们的目标是——
GUILLAUME VERDON (01:24:47) The trick there was to do a feedforward operation and do what we call a phase kick. But really, it’s just a force kick. You just kick the system with a certain force that is proportional to your loss function that you wish to optimize. And then, by performing uncomputation, you start with a superposition over parameters, which is pretty funky. Now, you don’t have just a point for parameters, you have a superposition over many potential parameters. Our goal is-
Lex Fridman (01:25:24) 是用相位踢以某种方式调整参数吗？
LEX FRIDMAN (01:25:24) Is using phase kick somehow to adjust the parameters?
Guillaume Verdon (01:25:28) 对。因为相位踢模拟了让参数空间像n维中的粒子，你试图在神经网络的损失景观中获得薛定谔方程、薛定谔动力学。你做一个算法来诱导这个相位踢，这涉及一个前馈、一个踢。然后，当你反计算前馈时，那么所有这些相位踢和这些力的误差会反向传播并击中各层中的每一个参数。
GUILLAUME VERDON (01:25:28) Right. Because phase kicks emulate having the parameter space be like a particle in end dimensions, and you’re trying to get the Schrödinger equation, Schrödinger dynamics, in the lost landscape of the neural network. You do an algorithm to induce this phase kick, which involves a feedforward, a kick. And then, when you uncompute the feedforward, then all the errors in these phase kicks and these forces back- propagate and hit each one of the parameters throughout the layers.
Guillaume Verdon (01:26:04) 如果你把这个与动能的模拟交替进行，那么它就像一个在n维中移动的粒子，一个量子粒子。原则上的优势是它可以在景观中穿隧并找到对于随机优化器来说很难找到的新最优解。但同样，这是一个理论性的东西，在实践中，至少以我们目前计划的量子计算机架构，这样的算法运行起来会极其昂贵。
GUILLAUME VERDON (01:26:04) If you alternate this with an emulation of kinetic energy, then it’s like a particle moving in end dimensions, a quantum particle. The advantage in principle would be that it can tunnel through the landscape and find new optima that would’ve been difficult for stochastic optimizers. But again, this is a theoretical thing, and in practice with at least the current architectures for quantum computers that we have planned, such algorithms would be extremely expensive to run.
Lex Fridman (01:26:41) 也许这是一个问不同领域之间区别的好地方，你曾涉足的领域。所以，数学、物理、工程，还有创业，堆栈的不同层次。我认为你在这里谈论的很多东西在数学方面有一点，也许物理几乎在理论中工作。
LEX FRIDMAN (01:26:41) Maybe this is a good place to ask the difference between the different fields that you’ve had a toe in. So, mathematics, physics, engineering, and also entrepreneurship, the different layers of the stack. I think a lot of the stuff you’re talking about here is a little bit on the math side, maybe physics almost working in theory.
Guillaume Verdon (01:27:03) 嗯。
GUILLAUME VERDON (01:27:03) Mm-hmm.
Lex Fridman (01:27:03) 数学、物理、工程和为量子计算、量子机器学习制造产品之间有什么区别？
LEX FRIDMAN (01:27:03) What’s the difference between math, physics, engineering, and making a product for a quantum computing for quantum machine learning?
Guillaume Verdon (01:27:14) 是的。TensorFlow Quantum项目的一些原始团队成员，我们在学校开始的，在滑铁卢大学，有我自己。最初，我是一名物理学家、应用数学家。我们有一名计算机科学家，我们有一名机械工程师，然后我们有一名物理学家。那主要是实验性的。组建非常跨学科的团队并弄清楚如何沟通和分享知识，真的是做这种跨学科工程工作的关键。
GUILLAUME VERDON (01:27:14) Yeah. Some of the original team for the TensorFlow Quantum project, which we started in school, at University of Waterloo, there was myself. Initially, I was a physicist, applied mathematician. We had a computer scientist, we had a mechanical engineer, and then we had a physicist. That was experimental primarily. Putting together teams that are very cross-disciplinary and figuring out how to communicate and share knowledge is really the key to doing this interdisciplinary engineering work.
Guillaume Verdon (01:27:51) 有很大的区别。在数学中，你可以为了数学而探索数学。在物理学中，你是在应用数学来理解我们周围的世界。在工程中，你试图黑掉世界。你试图找到如何应用我知道的物理学，我对世界的知识，来做事情。
GUILLAUME VERDON (01:27:51) There is a big difference. In mathematics, you can explore mathematics for mathematics’ sake. In physics, you’re applying mathematics to understand the world around us. And in engineering, you’re trying to hack the world. You’re trying to find how to apply the physics that I know, my knowledge of the world, to do things.
Lex Fridman (01:28:11) 嗯，特别是在量子计算中，我认为工程上有很多限制。它似乎非常困难。
LEX FRIDMAN (01:28:11) Well, in quantum computing in particular, I think there’s just a lot of limits to engineering. It just seems to be extremely hard.
Guillaume Verdon (01:28:17) 是的。
GUILLAUME VERDON (01:28:17) Yeah.
Lex Fridman (01:28:18) 所以在理论上用数学探索量子计算、量子机器学习有很多价值。我想问一个问题是，为什么建造量子计算机如此困难？你对将这些想法付诸实践的时间线有什么看法？
LEX FRIDMAN (01:28:18) So, there’s a lot of value to be exploring quantum computing, quantum machine learning in theory with math. I guess one question is, why is it so hard to build a quantum computer? What’s your view of timelines in bringing these ideas to life?
Guillaume Verdon (01:28:43) 对。我认为我公司的一个总体主题是，我们有一些……有一种从量子计算的大规模流出，我们正在转向不是量子的更广泛的基于物理的AI。所以，这给了你一个提示。
GUILLAUME VERDON (01:28:43) Right. I think that an overall theme of my company is that we have folks that are… There’s a sort of exodus from quantum computing and we’re going to broader physics-based AI that is not quantum. So, that gives you a hint.
Lex Fridman (01:29:00) 我们应该说你的公司名字是Extropic？
LEX FRIDMAN (01:29:00) We should say the name of your company is Extropic?
Guillaume Verdon (01:29:03) Extropic，没错。我们做基于物理的AI，主要基于热力学，而不是量子力学。但本质上，量子计算机非常难以建造，因为你必须诱导这个零开温度的信息子空间。做到这一点的方法是通过编码信息，你在代码中编码代码，在代码中编码代码，在代码中编码代码。需要大量冗余来做这个纠错，但最终，它是一种算法冰箱，真的。它只是把熵从虚拟的、去局域化的子系统中抽出来，该子系统代表你的”逻辑量子比特”，也就是你实际想运行量子力学程序的有效载荷量子比特。它非常困难，因为为了扩展你的量子计算机，你需要每个组件都具有足够的质量才值得。因为如果你试图做这个纠错，这个量子纠错过程，在每个量子比特和你对它们的控制中，如果它不够充分，就不值得扩展。你实际上添加的错误比你移除的更多。有一个阈值的概念，即如果你的量子比特在控制方面具有足够的质量，那么扩展实际上是值得的。实际上，近年来，人们一直在跨越阈值，它开始变得值得。
GUILLAUME VERDON (01:29:03) Extropic, that’s right. We do physics-based AI, primarily based on thermodynamics, rather than quantum mechanics. But essentially, a quantum computer is very difficult to build because you have to induce this zero temperature subspace of information. The way to do that is by encoding information, you encode a code within a code, within a code, within a code. There’s a lot of redundancy needed to do this error correction, but ultimately, it’s a sort of algorithmic refrigerator, really. It’s just pumping out entropy out of the subsystem that is virtual and delocalized that represents your “logical qubits”, aka the payload quantum bits in which you actually want to run your quantum mechanical program. It’s very difficult because in order to scale up your quantum computer, you need each component to be of sufficient quality for it to be worth it. Because if you try to do this error correction, this quantum error correction process, in each quantum bit and your control over them, if it’s insufficient, it’s not worth scaling up. You’re actually adding more errors than you remove. There’s this notion of a threshold where if your quantum bits are sufficient quality in terms of your control over them, it’s actually worth scaling up. Actually, in recent years, people have been crossing the threshold and it’s starting to be worth it.
Guillaume Verdon (01:30:38) 这只是一个非常漫长的工程跋涉，但最终，对我来说真正疯狂的是我们对这些系统有多么精致的控制水平。这实际上相当疯狂。人们正在跨越……他们正在实现里程碑。只是总的来说，媒体总是走在技术前面。炒作有点太多了。这对筹款有好处，但有时它会导致寒冬。这是炒作周期。我个人对10年、15年时间尺度上的量子计算持乐观态度，但我认为在此期间可以做其他的探索。我认为它现在掌握在好手中。
GUILLAUME VERDON (01:30:38) It’s just a very long slog of engineering, but ultimately, it’s really crazy to me how much exquisite level of control we have over these systems. It’s actually quite crazy. And people are crossing… They’re achieving milestones. It’s just in general, the media always gets ahead of where the technology is. There’s a bit too much hype. It’s good for fundraising, but sometimes it causes winters. It’s the hype cycle. I’m bullish on quantum computing on a 10, 15-year timescale personally, but I think there’s other quests that can be done in the meantime. I think it’s in good hands right now.
Lex Fridman (01:31:22) 嗯，让我探索一些不同的美丽想法，无论大小，在量子计算中可能从记忆中跳出来的，当你合著了一篇题为《通过Qudit探针实现渐近无限量子能量传送》的论文时。出于好奇，你能解释一下qudit与qubit相比是什么吗？
LEX FRIDMAN (01:31:22) Well, let me just explore different beautiful ideas, large or small, in quantum computing that might jump out at you from memory when you co-authored a paper titled Asymptotically Limitless Quantum Energy Teleportation via Qudit Probes. Just out of curiosity, can you explain what a qudit is versus a qubit?
Guillaume Verdon (01:31:45) 是的。它是一个D态量子比特。
GUILLAUME VERDON (01:31:45) Yeah. It’s a D-state qubit.
Lex Fridman (01:31:49) 它是多维的？
LEX FRIDMAN (01:31:49) It’s a multidimensional?
Guillaume Verdon (01:31:50) 多维的，对。它就像，嗯，你能有一个量子力学的整数浮点概念吗？这是我必须思考的东西。我认为那项研究是后来量子模数转换工作的前兆。那很有趣，因为在我硕士期间，我试图理解真空、空无的能量和纠缠。空无具有能量，这说起来非常奇怪。我们的宇宙学方程与我们对涨落中存在多少量子能量的计算不匹配。
GUILLAUME VERDON (01:31:50) Multidimensional, right. It’s like, well, can you have a notion of an integer floating point that is quantum mechanical? That’s something I’ve had to think about. I think that research was a precursor to later work on quantum analog digital conversion. There was interesting because during my masters, I was trying to understand the energy and entanglement of the vacuum of emptiness. Emptiness has energy, which is very weird to say. Our equations of cosmology don’t match our calculations for the amount of quantum energy there is in the fluctuations.
Guillaume Verdon (01:32:36) 我试图黑进真空的能量，而现实是你不能直接黑进它。它在技术上不是自由能。你对涨落的无知意味着你无法提取能量。但就像股市一样，如果你有一只随时间相关的股票，真空实际上是相关的。如果你在一个点测量了真空，你获得了信息。如果你把那个信息传达到另一个点，你可以推断真空处于什么配置，达到某种精度，并统计地平均提取一些能量。所以，你”传送了能量”。
GUILLAUME VERDON (01:32:36) I was trying to hack the energy of the vacuum, and the reality is that you can’t just directly hack it. It’s not technically free energy. Your lack of knowledge of the fluctuations means you can’t extract the energy. But just like the stock market, if you have a stock that’s correlated over time, the vacuum’s actually correlated. If you measured the vacuum at one point, you acquired information. If you communicated that information to another point, you can infer what configuration the vacuum is in to some precision and statistically extract, on average, some energy there. So, you’ve “teleported energy”.
Guillaume Verdon (01:33:18) 对我来说，这很有趣，因为你可以创造负能量密度的口袋，也就是低于真空的能量密度，这非常奇怪，因为我们不理解真空如何（传播？）引力。有一些理论认为真空或时空本身的画布实际上是由量子纠缠制成的画布。我在研究如何在局部降低真空的能量会增加量子纠缠，这非常古怪。
GUILLAUME VERDON (01:33:18) To me, that was interesting because you could create pockets of negative-energy density, which is energy density that is below the vacuum, which is very weird because we don’t understand how the vacuum gravitates. There are theories where the vacuum or the canvas of space-time itself is really a canvas made out of quantum entanglement. I was studying how decreasing energy of vacuum locally increases quantum entanglement, which is very funky.
Guillaume Verdon (01:33:58) 这里的事情是，如果你对UAP和诸如此类的奇怪理论感兴趣，你可以试着想象它们在周围。它们会如何推动自己？它们会如何超越光速？你需要一种负能量密度。对我来说，我尽了我的努力，试图黑进真空的能量，并达到物理定律允许的极限。但那里有各种警告，你显然不能提取比你投入的更多。
GUILLAUME VERDON (01:33:58) The thing there is that, if you’re into to weird theories about UAPs and whatnot, you could try to imagine that they’re around. And how would they propel themselves? How would they go faster than the speed of light? You would need a sort of negative energy density. To me, I gave it the old college try, trying to hack the energy of vacuum and hit the limits allowable by the laws of physics. But there’s all sorts of caveats there where you can’t extract more than you’ve put in, obviously.
Lex Fridman (01:34:41) 但你是说传送能量是可能的，因为你可以在一个地方提取信息，然后基于此，对另一个地方做出某种预测？
LEX FRIDMAN (01:34:41) But you’re saying it’s possible to teleport the energy because you can extract information one place and then make, based on that, some kind of prediction about another place?
Guillaume Verdon (01:34:56) 嗯。
GUILLAUME VERDON (01:34:56) Mm-hmm.
Lex Fridman (01:34:57) 我不确定该如何理解这个。
LEX FRIDMAN (01:34:57) I’m not sure what to make of that.
Guillaume Verdon (01:34:58) 是的，这是物理定律允许的。但现实是关联会随距离衰减。
GUILLAUME VERDON (01:34:58) Yeah, it’s allowable by the laws of physics. The reality though is that the correlations decay with distance.
Lex Fridman (01:35:06) 当然。
LEX FRIDMAN (01:35:06) Sure.
Guillaume Verdon (01:35:06) 所以，你将不得不在离你提取它的地方不太远的地方付出代价。
GUILLAUME VERDON (01:35:06) And so, you’re going to have to pay the price not too far away from where you extract it.