人工智能与未来：从国际象棋到 DeepMind

欢迎来到第二届皇家电视学会与工程技术学会公众讲座。我是今晚的主持人蒂姆·戴维。当我们构思这个系列讲座时，目标是聆听一些世界上最杰出的思想家的声音，他们工作在尖端科技与人类创造力及媒体交汇的领域。去年有幸目睹迈克·林奇精彩演讲的我们，都感到这方向对了。今晚现场的兴奋感证明了公众讲座在激发想象力方面的持久魅力。我深信，我们正幸运地享受另一个需要公众参与和讨论的黄金进步时代。200 多年前，早熟的汉弗里·戴维就曾以关于人类进步和科学知识的售罄讲座让观众惊叹，甚至造成马车交通堵塞。值得注意的是，作为柯勒律治的朋友，他对主观与技术的交汇及其带来的无限可能性着迷。所以今晚，在这个美妙的机构，我们延续这一传统，邀请到特别嘉宾德米斯·哈萨比斯。德米斯将演讲约 35 分钟，然后进行问答环节。我将主持，我们非常期待与大家进行一场精彩的讨论。如各位所知，德米斯被公认为人工智能领域的世界领先思想家，而 AI 近年来已成为主导媒体和公众想象的热门话题。在我的领域，我们最近进行了一项 BBC 调查，探讨哪些工作面临计算机化风险，吸引了 230 万页面浏览量。德米斯无需过多介绍，但他的成就让最杰出者的履历也黯然失色：13 岁成为国际象棋大师，五次赢得世界游戏锦标赛冠军，成功的扑克玩家，剑桥大学计算机科学双一等荣誉学位，开创性的视频游戏开发者，多项重要神经科学研究的主导者——特别是关于记忆塑造与未来想象相似性的里程碑式论文——当然，还有 2011 年创立 DeepMind，2014 年出售给谷歌，德米斯担任工程副总裁，专门负责 AI。DeepMind 的目标是解决智能问题——一个谦逊的目标。具体来说，德米斯说他参与构建能优雅地应对意外的事物。我认为这对公众讲座的听众来说是一个很好的简介。德米斯，请开始。

Welcome to the second Royal Television Society and Institution of Engineering and Technology public lecture. I'm Tim Davy, chairing tonight. When we conceived this lecture series, our aim was to hear from some of the world's finest minds working at the intersection of cutting-edge science and technology with human creativity and media. The amazing response, and those of us who saw Mike Lynch in full flow last year, suggested we were onto something. The excitement in the room tonight speaks to the enduring appeal of public lectures in sparking our imagination. I really believe we are lucky enough to be enjoying another golden age of advancement that demands public engagement and debate. Over 200 years ago, a rather precocious Humphry Davy wowed crowds and caused carriage traffic jams with sellout lectures on human progress and scientific knowledge. Notably, as a friend of Coleridge, he was fascinated by the intersection of the subjective and the technical, and the endless possibilities this throws up. So tonight, at this wonderful institution, we continue that tradition with our very special speaker, Demis Hassabis. Demis will talk for about 35 minutes, then open up for Q&A. I'll be conducting, and we're really up for a good discussion with you all. As you know, Demis is acknowledged as a world-leading thinker in AI, which has become one of the hottest topics dominating media and public imagination. In my world, we ran a recent BBC survey on which jobs are at risk of computerization, attracting 2.3 million page views. Demis needs little introduction, but his achievements put even the highest achievers' CVs deeply in the shade: chess master at 13, five-time World Games champion, successful poker player, double first in computer science at Cambridge, a pioneering video games developer, leader of numerous important neuroscience research pieces—notably his landmark paper on the similarities of how we shape memory and imagine the future—and then, of course, founder of DeepMind in 2011, sold to Google in 2014, with Demis becoming VP of Engineering with special responsibility for AI. DeepMind has set out to solve intelligence—a humble objective. Specifically, Demis says he is involved in building something that can 'expect the unexpected gracefully.' I think that's a great brief for a public lecture audience. Demis, the floor is yours.

非常感谢蒂姆的慷慨介绍。能在这里做讲座真是非常愉快。我将谈论人工智能及其对未来的影响——实际上，可能是相对近期的未来。AI 是让机器变聪明的科学。我最初通过游戏进入 AI 领域。游戏对我来说始于国际象棋，正如蒂姆所说。我四岁就开始下棋。我认为，如果你从这么小就认真下棋，而且像我一样是个内向的孩子，你会开始思考你的大脑是如何想出这些走法、这些想法，让你下棋并获胜的。进入青少年时期后，我开始大量思考这个问题。与此相关，我也接触了计算机。我八岁时用国际象棋比赛赢得的奖金买了第一台电脑——ZX Spectrum 48k，并开始自学编程。我想很早的时候——在座的工程师们会有共鸣——我就直觉地意识到计算机是一种特殊的机器。大多数机器，比如汽车和飞机，扩展了我们的物理能力：汽车让我们比跑步更快，飞机让我们飞翔。但计算机在思维领域做到了这一点——它们真正扩展了大脑的能力。当我编写第一个程序并做基本数学计算时，这一点变得清晰。我惊讶地发现，你可以让程序运行一整夜，然后去睡觉，第二天早上醒来发现电脑在你睡觉时解决了一些问题。这感觉像是一个非常强大、在某种程度上神奇的工具。所以我对计算机和游戏的热爱自然而然地结合在了视频游戏设计中。实际上，我接受这次讲座的一个原因是，正如蒂姆所说，我喜欢将 RTS、创意艺术、IET 和工程学融合在一起的想法。这就是我进入商业视频游戏领域的原因——因为在 90 年代早期和中期，电脑游戏确实在推动工程学的尖端发展，甚至包括为运行这些游戏而制造的机器。我记得 90 年代关于英特尔推出新奔腾处理器的争论，人们说：“我们还需要多少算力来运行文字处理器和电子表格？我们不是已经有足够的算力了吗？”一个答案是，如果我们想要更逼真、更复杂的游戏，就需要更强大的计算机，配备更大的内存和图形芯片。所以很长一段时间，游戏实际上在推动尖端硬件的发展。此外，我设计和编程的游戏都以 AI 作为核心游戏机制。我最著名的游戏叫《主题公园》，这里有一些截图。它于 94 年发布，非常成功。它实际上是同类游戏中的第一款。其理念是设计你自己的迪士尼乐园，成千上万的小人会进来玩你的游乐设施。他们对你的主题公园的喜爱程度会影响他们的情绪和幸福感，进而影响经济模型——你可以为汉堡和气球收取多少费用。你的设计越好，赚的钱就越多，从而可以进一步扩建。这款游戏和另一款叫《模拟城市》的游戏是首批将 AI 作为核心游戏机制的游戏，并真正催生了一个完整的类型。

Thanks very much, Tim, for that very generous introduction. It's a real pleasure to be here giving this lecture. I'm going to talk about artificial intelligence and its impact on the future—in fact, it could be the relatively near-term future. AI is really the science of making machines smart. I got into AI first through the medium of games. Games started for me with chess, as Tim said. I started playing chess when I was very young, at age four. I think if you play chess seriously from such a young age and you're quite an introspective kid, which I was then, you start thinking a lot about how your brain comes up with these moves, these ideas that allow you to play and win. I started thinking a lot about this as I got into my teenage years. Allied with that, I got into computing. I actually bought my first computer, a ZX Spectrum 48k, with some winnings from a chess tournament when I was about eight years old, and I started teaching myself how to program. I think very early on—and the engineers in the audience will resonate with this—I realized on an intuitive level that computers are a special type of machine. Most machines, like cars and planes, extend our physical capabilities: cars let us move faster than we can run, planes let us fly. But computers do that in the realm of our minds—they really extend the capabilities of the brain. This became clear to me when I wrote my first programs and did basic math calculations. It struck me that you could set something running overnight, go to sleep, and wake up the next morning to find your computer had solved some problem for you while you slept. This felt like a really powerful, in some ways magical, tool. So my love of computers and my love of games came together in an obvious way in designing video games. Actually, one reason I accepted this lecture is I love the idea, as Tim said, of the confluence of bringing together the RTS, the creative arts, the IET, and engineering. That's why I got into commercial video games—because at the time, in the early and mid-90s, computer games were really pushing the cutting edge of engineering, and even the machines being built to run these games. I remember debates in the '90s about Intel bringing out their new Pentium processors, people saying, 'How much more power do we need to run our word processors and spreadsheets? Haven't we got all the computing power we need?' One answer was that if we wanted more realistic and complex games, we would need more powerful computers with larger memory and graphics chips. So for a long time, games were actually driving the development of cutting-edge hardware. Furthermore, the games I used to design and program all involved AI as a core gameplay mechanic. My best-known game was called Theme Park, with some screenshots here. It came out in '94 and was very successful. It was actually the first game of its type. The idea was that you designed your own Disney World, and thousands of little people would come in and play on your rides. How enjoyable they thought your theme park was would impact their emotions and happiness, which fed into the economics model of how much you could charge for hamburgers and balloons. The better your design, the more money it made, allowing you to expand further. This game, and another called SimCity, were the first games with AI as a core gameplay component and really spawned a whole genre.

所谓的经营模拟游戏。这些游戏之所以如此受欢迎，原因之一在于 AI 会适应玩家的游戏方式。这意味着每个玩这款游戏的人都有独特的体验。我记得人们会寄信给游戏杂志，向我们展示他们主题公园的最终状态。人们创造出了许多令人惊叹的设计，连我们这些游戏发明者都没想到能做到。这在我 16、17 岁编写这款游戏时深深触动了我。我想，如果我将职业生涯奉献给推进 AI，那将是多么不可思议的技术。

Management simulation games, as they're called. One of the reasons these games were so popular is that the AI adapted to the way the player played the game. That meant every single person who played this game had a unique experience. People used to send in, I remember, to magazines, game magazines, and write to us, showing what end state they got their theme park into. There were all these amazing designs that people had created that we had no idea could be done, even as the inventors of this game. That really struck a chord with me when I was around 16 or 17 years old when I wrote this game. I thought, maybe if I devoted my career to advancing AI, what an incredible technology that could be.

在游戏行业打拼并经营自己的游戏公司之后，我重返学术界攻读神经科学博士学位。我觉得这是我在发起像 DeepMind 这样的项目之前需要的另一块拼图。我想更深入地了解大脑如何解决想象和记忆等难题。我特意选择这些课题作为博士研究方向，因为至少在 2000 年代中期，我们在计算机算法方面还不太擅长这些。我想研究大脑如何解决这些我们尚不知道如何赋予机器的难题。我将在演讲的后半部分再回到这个话题。

After having a career in games and running my own games companies, I went back to academia to do a PhD in neuroscience. I felt that was another piece of the puzzle I needed before launching an effort like DeepMind. I wanted to understand a bit more about how the brain solved tough problems like imagination and memory. I specifically picked those topics for my PhD because those are things that, at least back in the mid-2000s, we were not very good at doing in computer algorithms. I wanted to look at the way the brain solved some of these very tough problems that we didn't yet know how to imbue our machines with. I'll come back to that in the second half of my talk.

所有这些不同的经历最终促成了我在 2010 年创立 DeepMind。我花了 20 多年的时间才走到这一步，积累了足够的我认为是基本要素的东西——无论是在算法层面，还是在组建创始科学团队和建立联系方面——从而真正打造出像 DeepMind 这样的机构，并有可信度地去追求像解决智能这样宏大的使命。

All of these different experiences culminated in finally setting up DeepMind in 2010. It's been a 20-year-plus journey for me to get to this point and have enough of what I thought were the basic ingredients, both on an algorithmic level and in terms of the founding scientific team and making those contacts, to actually put together something like DeepMind and plausibly go after as big a mission as solving intelligence.

我们看待公司的另一种方式是将其视为 AI 的阿波罗计划，一个真正专注于雄心勃勃的长期目标的登月项目。我们已经聚集了超过 100 位，实际上现在接近 150 位，这个领域全球顶尖的研究科学家。我认为 DeepMind 是迄今为止世界上最大的机器学习专家聚集地。除了尝试构建 AI，我们还在实验组织科学事业的新方式。我们试图将硅谷初创公司的精华与麻省理工学院、伦敦大学学院、剑桥等顶尖学术机构的优势结合起来，看看能否融合成一种新的混合科学工作方式，既更高效、极具效率，又允许极致的创造力。

Another way we look at the company is as an Apollo program for AI, a sort of moonshot project that really focuses on very ambitious long-term goals. We've collected together more than 100, actually nearly 150 now, of the world's top research scientists in this area. I think DeepMind is by far the biggest collection of machine learning experts anywhere in the world. Another thing we're experimenting with, apart from trying to build AI, is new ways to organize scientific endeavor. What we tried to do with DeepMind is really combine the best from Silicon Valley startups together with the best parts you find in the best academic institutes like MIT, UCL, Cambridge, and so on, and see if we can fuse that into a new hybrid way of doing science that is more productive and extremely efficient but still allows for extreme creativity.

我们的使命，正如 Tim 所说，我们用一种两步走的方式来阐述。首先，我们谈论解决智能。我们使用'解决'这个词，它有点模糊，因为实际上我们的意思是理解自然智能，即人类心智，同时也人工地重现这种智能。然后第二步，我们想利用这项技术来帮助我们解决其他一切问题。这对你们中的一些人来说可能有点牵强，甚至有点异想天开，但我们真的相信，如果你能解决智能，第二步自然会从第一步中产生。我希望在演讲结束时，你们会同意这个猜想。

Our mission, as Tim said, we articulate it in a kind of two-step way. Firstly, we talk about solving intelligence. We use the word 'solve', which is a kind of ambiguous word there, because actually what we mean is understanding natural intelligence, so the human mind, but also recreating that intelligence artificially. Then step two, we want to use that technology to help us solve everything else. That might seem a little bit far-fetched, possibly a little bit fanciful to some of you, but we really believe that step two naturally follows on from step one if you can solve intelligence. I hope by the end of this talk you'll agree with this conjecture.

更实际地说，我们打算如何解决智能？我们在 DeepMind 试图做的是构建世界上第一个通用学习机器。关键方面是'通用'和'学习'这两个词。在 DeepMind，我们只对能够自我学习的算法感兴趣，即它们从原始经验或原始数据中自动学习。它们不以任何方式预编程。我们在这里谈论的是自主学习系统。第二点是通用性这个概念。我们感兴趣的是同一个系统能够开箱即用地在广泛的任务和环境中运行，无需重新配置。当然，我们有一个这样的通用学习系统的例子：人类心智，我们能够将心智应用于几乎无数的不同任务。

More practically, how are we going to solve intelligence? What we're trying to do at DeepMind is to construct the world's first general-purpose learning machine. The key aspects are the word 'general' and 'learning'. At DeepMind, we're only interested in algorithms that learn for themselves, so they learn automatically from raw experience or raw data. They are not pre-programmed in any way. What we're talking about here is autonomous learning systems. The second thing is this idea of generality. What we're interested in is the same system actually being able to operate across a wide range of tasks and environments out of the box with no reconfiguration. Of course, we have an example of such a general learning system: the human mind, where we're able to apply our minds to almost endless number of different tasks.

我应该说，当今大多数 AI，尽管现在是一个巨大的流行词且非常时髦，但大多数并非这种类型的技术。我们在 DeepMind 内部将大多数 AI 称为'狭义 AI'。我们的意思是，它是为特定任务以定制方式构建的预编程 AI。实际上，我们每天互动的大多数 AI，从手机上的 Siri 到自动驾驶汽车，都属于这种预编程类型的 AI。我们感兴趣的是我们所谓的通用人工智能，即通用学习系统的概念。

I should say most of AI today, although it's a huge buzzword right now and is very fashionable, most of it is not of this type of technology. We call most AI internally at DeepMind 'narrow AI'. What we mean by that is pre-programmed AI that has been built in a bespoke way for one specific task. Actually, most of the AI we interact with every day, from Siri on your phone to self-driving cars, is of this pre-programmed type of AI. What we're interested in is what we call artificial general intelligence, this idea of a general learning system.

也许我能给出的最著名、最清晰的例子是著名的深蓝对阵加里·卡斯帕罗夫的比赛。这是 AI 的一个分水岭时刻，在 1990 年代末，IBM 的深蓝在一场六局国际象棋比赛中击败了卡斯帕罗夫。但有趣的是，我从那场比赛中走出来，实际上对加里·卡斯帕罗夫的心智印象更深刻，而不是深蓝机器。当然，这是一项令人印象深刻的工程壮举，但深蓝是由一个出色的程序员团队以及一群国际象棋特级大师编程的，他们试图将国际象棋知识提炼成算法结构。那些程序员直接将想法和解决方案编程到机器中。这意味着深蓝虽然非常擅长国际象棋，但对其他任何事情都毫无用处，包括严格更简单的事情，比如玩井字棋，任何国际象棋特级大师都能轻松学会。显然，深蓝所知道或代码中包含的任何东西都无法帮助它处理甚至像那样严格简单的事情，更不用说其他领域，比如说话或开车，而这些加里·卡斯帕罗夫都能毫不费力地做到。

Perhaps the most famous and clearest example I can give is the famous Deep Blue match against Gary Kasparov. This was a watershed moment in AI when in the late 1990s IBM's Deep Blue beat Kasparov in a six-game chess match. But the interesting thing is I came away from that match actually more impressed by Gary Kasparov's mind than the Deep Blue machine. Of course it was an impressive engineering feat, but Deep Blue was programmed by an amazing team of programmers along with a bunch of chess grandmasters trying to distill chess knowledge into an algorithmic construct. Those programmers were directly programming the ideas and solutions into the machine. What that meant is that Deep Blue, although it was very good at chess, was no use for absolutely anything else, including strictly simpler things like playing noughts and crosses, which any chess grandmaster could trivially be taught. Obviously nothing Deep Blue knew or had in its code would help it with even something strictly simple like that, let alone other kinds of domains like speaking languages or driving cars, which of course Gary Kasparov could do effortlessly.

取而代之，我们在所谓的强化学习框架中思考智能。我将用这个小卡通图来说明其主要基本部分，因为这对于我接下来要展示的算法工作视频很重要。你从你的智能体系统开始，由这个小人物角色代表，该智能体发现自己处于一个环境中，可以是虚拟世界或现实世界。如果是现实世界，智能体很可能是一个机器人。

Instead of that, we think about intelligence in the framework of what's called reinforcement learning. I'm just going to illustrate what the main basic parts of that are in this little cartoon diagram because it's important for what I'm going to show next in terms of the videos of the algorithms working. You start off with your agent system, represented by this little humanoid character, and that agent finds itself in an environment, which could be virtual or real world. If it's the real world, the agent will probably be a robot.

智能体将是一个化身，它被赋予某种目标，并试图在该环境中实现该目标。智能体与环境互动只有两种方式：一是通过其感官装置获取关于世界的观察。目前我们主要使用视觉，但很快也会探索其他感官模态。这些观察总是不完整且有噪声的，所以你永远无法获得关于世界的全部信息，不像下棋那样是完美状态信息，你能看到游戏世界的一切。在现实世界中，当然，你无法看到所有信息。智能体系统的任务之一就是仅基于这些有噪声、不完整的观察，尽可能准确地构建外部环境的模型。智能体实时进行这一过程：每个时间步都会收到观察，并基于新证据不断更新其世界模型。智能体的第二个任务是选择它应该采取什么行动，即在当前情况下，什么行动最能帮助它实现目标。一旦决定了行动，它就会输出该行动，行动被执行，然后可能导致环境变化，进而产生新的观察，如此循环往复。

The agent will be an avatar and the agent has some kind of goal that it's been given to that it's trying to achieve in that environment. The agent only interacts with the environment in two ways: one is that it gets observations through its sensory apparatus, observations about the world. We mostly use vision at the moment, but we're also looking to use other sensory modalities soon. Those observations are always incomplete and noisy, so you never get full information about the world, unlike say a game of chess where it's perfect state information, you see everything in the game world. In the real world, of course, you don't get to see all the information. One of the jobs of the agent system is to build as accurate a model as possible of the environment out there based solely on these noisy, incomplete observations. The agent is doing this in real time: these observations are coming in every time step, and the agent is continually updating its model of the world based on this new evidence that it gets. The second job of the agent is to then pick what action it should take, what's the best action it can take in that particular moment in time that will best get it towards its goal from the current situation that it finds itself in. Once it's decided what that action should be, it outputs that action, that action gets executed, and that then may drive a change in the environment, which will then drive a new observation, and this goes around in an endless sort of cycle.

尽管这个图解释起来很简单，但它实际上隐藏了巨大的复杂性。我们知道，如果你能解决这个图背后所有的问题，那么这就足以实现真正的人工智能。我们之所以知道这一点，是因为生物系统（包括人类和大多数哺乳动物）就是这样学习的。事实上，在人类中，是多巴胺系统实现了一种强化学习的形式。

Although this diagram is very simple to explain, it actually hides an incredible amount of complexity. We know that if you could solve all the problems that underlie this diagram, this representation, then that would be enough for true artificial intelligence. We know that because this is the way that biological systems learn, including humans and most mammals. In fact, in humans, it's the dopamine system that implements a form of reinforcement learning.

我们接着测试了这类系统，实际上我们选择在电脑游戏上测试我们系统的智能。我们相信，一个真正的思维机器必须嵌入到感觉运动数据流中。除非你拥有影响你所处世界的能力和感知这个世界的能力，否则你无法拥有真正的智能和思维。这被称为具身认知。通常，当人们认同这种 AI 观点时，他们通常会开始研究基于真实世界环境的真实机器人。但机器人非常棘手：它们非常昂贵、非常慢，而且容易出故障。如果你和任何使用过或尝试过开发机器人的人交谈，你会听到很多工作实际上是在修理机器人的机械结构、电机、传感器等等。我们不想被这些分心；我们想专注于智能算法本身。所以我们决定首先使用视频游戏。这和我的一些背景有关，在这里派上了用场，我们重新利用游戏作为测试 AI 算法智能和能力的平台。

We went on to test these kinds of systems, and actually we chose to test the intelligence of our systems on computer games. A true thinking machine, we believe, would have to be embedded in a sensory motor data stream. You can't have true intelligence and true thinking unless you have the ability to affect the world that you're in and the ability to sense that world. This is called embodied cognition. Usually when people subscribe to this view of AI, they normally start working on real robots based in real world environments. But robots are very tricky to use: they're very expensive, very slow, and they break down. If you talk to anyone who has used or tried to develop on robots, you'll hear a lot of the work actually goes into fixing the mechanics of the robots, the motors, the sensors, and so on. We didn't want to be distracted by that; we wanted to focus on the intelligence algorithms themselves. So what we decided to use was video games in the first instance. It's a little bit to do with my background where it came in useful here in video games, and we repurposed the games as a platform for testing the intelligence and capabilities of our AI algorithms.

游戏非常好，因为显然你可以在云端运行它们，运行速度比实时快得多，你可以并行运行数百万次实验，而且很容易衡量进展，因为大多数游戏都有分数。所以你可以很方便地看到你的算法调整是否带来了优势，以及你是否朝着正确的方向前进。这对于像我们这样长期且雄心勃勃的使命非常重要：能够将一个雄心勃勃的使命分解成更小的、易于衡量进展的部分。游戏的另一个关键点是，它们显然是由其他人和工程团队设计的，并非专门为 AI 测试而设计。这意味着你必须处理各种有趣的问题，如果你自己是 AI 设计师，你永远不会设计这些问题。我认为这实际上确保了你在应用 AI 时不会对问题类型产生任何偏见。过去几十年 AI 研究的一个大问题是，通常 AI 设计师也设计问题，无论你是否喜欢，潜意识里你最终会设计出你知道你的 AI 算法擅长的问题。

Games are really good because obviously you can run them in the cloud, you can run them much faster than real time, you can run millions of experiments in parallel, and it's very easy to measure progress because most games fortunately have game scores. So you can see very conveniently if your algorithmic tweaks are gaining you an advantage and whether you're heading in the right direction based on the performance in those environments. That's something very important, especially for a very long-term mission like we have, and a very ambitious mission: to be able to break down an ambitious mission into smaller chunks that are very easy to measure progress on. The other key thing about games is that obviously they were designed by other people and other engineering teams, and they weren't designed specifically for AI testing. So what that means is that you have to deal with all kinds of interesting problems that you would never have dealt with if you designed yourself as an AI designer. I think that actually makes sure that there isn't any bias in the types of problems that you apply your AI to. One of the big problems in AI research over the last few decades is that generally speaking, it's the AI designers that also design the problems, and subconsciously, whether you like it or not, you end up designing problems that you know your AI algorithms are well suited to.

我们最初从 80 年代的 Atari 游戏开始，这确实是第一个拥有许多流行且具有挑战性游戏的标志性平台。我们决定从那里开始。我们使用了一个开源的 Atari 游戏模拟器，对其进行了优化，使其更稳定、运行更快，然后将我们的 AI 算法接入这个系统。现在我要给你们看一些 AI 系统工作的视频。但在那之前，我想解释一下你们将看到什么。这里的 AI 系统只接收原始像素作为输入，就好像我们设置了一个摄像头观察屏幕。它获得的唯一信息就是原始像素。它不知道自己在控制什么，不知道游戏的目标是什么，也不知道如何得分。它被告知的只是需要最大化分数。其他一切都是从零开始学习的。这里又涉及到了通用性，我们要求一个单一系统能够直接玩所有不同的游戏。显然有几十种非常不同的 Atari 游戏。

What we started off with was actually Atari games from the 80s, which are really the first iconic platform that had a lot of very popular challenging games on it. We decided to start with that. What we did is we started with an open source emulator for Atari games, souped it up, made it more robust, made it run faster, and then we plugged our AI algorithms into this system. Now I'm going to show you a couple of videos of the AI system working. But before I do that, I wanted to explain to you what it is you're going to see. The AI system here only gets the raw pixels as inputs, so it's almost as if we'd set up a video camera observing the screen. The only information that it gets is the raw pixels. It doesn't know anything about what it's controlling, it doesn't know what the aim of the game is, it doesn't know how to get points. All it's being told is that it needs to maximize the score. Everything else is learned from scratch. There's the sort of generality component that comes in again, where we require a single system to play all the different games out of the box. There are obviously dozens and dozens of very different Atari games.

我要展示的第一个视频是《太空侵略者》，这大概是有史以来最具标志性的游戏。我将展示这个视频的两个部分。一开始，当我播放视频时，你会看到智能体第一次遇到这个环境时的样子。它控制着屏幕底部的火箭，显然它在随机尝试动作，因为它不知道应该做什么，几乎立刻失去了三条命。如果你让机器训练一整夜，第二天回来，机器在游戏上已经达到了超人水平。它每次射击都能命中目标，不再会被杀死，它发现屏幕顶部的粉色母舰值很多分，它做出了惊人准确的射击。

The first video I'm going to show you is Space Invaders, probably the most iconic game there has ever been. I'm going to show you two parts of this video. In the beginning, as I roll the video now, you'll see what the agent looks like when it first encounters this environment. It's controlling the rocket at the bottom of the screen, obviously it's trying actions randomly because it has no idea what it's supposed to be doing, and it loses its three lives almost immediately. Now if you leave the machine training overnight and you come back the next day, the machine is now superhuman at the game. Every single shot it fires hits something, it can't be killed anymore, it's worked out that the pink Mothership at the top of the screen is worth a lot of points, it does these amazingly accurate shots.

你可以看到，它构建的世界模型非常精确。那些在 80 年代玩过《太空侵略者》的人会记得，敌人越少，它们移动得越快。如果你看最后一枪，你会发现它预测了那个太空侵略者会落在哪里，并提前开火。现在我要展示第二个视频，是《打砖块》游戏。这是我最喜欢的视频，我会展示智能体逐渐变得更好、更有能力的几个阶段。在这个游戏中，智能体控制粉色的球拍和球，目标是一块一块地击穿这面彩虹色的砖墙。一开始，经过 100 局游戏后，AI 系统还不太好。你可以看到它大部分时间都接不到球，但也许你能说服自己它开始明白应该把球拍移向球。经过 300 局游戏后，你可以看到系统已经变得和人类玩家一样好了，几乎总能接住球，即使球以非常垂直的角度弹回来。我们觉得这很酷，但我们想，如果让智能体运行更长时间，再玩 200 局会怎样？然后意想不到的事情发生了：它发现了最优策略——在侧面挖一条隧道，让球绕到墙后面。它再次以超高的精度完成了这个动作。有趣的是，尽管研究人员都是出色的程序员和工程师，但他们并不擅长玩雅达利游戏，所以实际上他们并不知道这个策略。所以我认为这是一个例子，你创造的系统实际上教会了你一些东西，这对我们来说是一个分水岭时刻。如果你们对这项工作感兴趣，它今年早些时候已经发表在《自然》杂志上，并登上了封面，我们甚至发布了代码，所以你们可以自己查看并尝试。

You can see that the model it's built of the world is extremely accurate. Those of you who played Space Invaders back in the 80s will remember that as there are less of them, they get faster. If you just watch the last shot, you'll see that it's sort of predicted where that space invader is going to end up and fires the shot ahead of time. So I'm going to show you our second video now, which is the game of Breakout. It's my favorite video, where I'm going to show a few more gradations of the agent getting better and more capable. In this game, the agent is controlling the pink bat and ball, and the aim of the game is to break through this rainbow-colored brick wall, brick by brick. To start off with, after 100 games, the AI system is not very good. You can see it's missing the ball most of the time, but you can see maybe you can convince yourself it's starting to get the hang of the idea that it should be moving the bat towards the ball. Now after 300 games, you can see that the system has now got pretty much as good as any human can play this, and almost always gets the ball back, even when it's coming back at very vertical angles. So we thought that was pretty cool, but we thought, what would happen if we left the agent running for longer and played another 200 games? And then this unexpected thing happened: it discovered the optimal strategy was to dig a tunnel around the side and send the ball around the back of the wall. It's doing that again with sort of super accuracy in terms of the motor control there. And one of the funny things is that although the researchers on that are amazing programmers and engineers, they're not so good at playing Atari games, so they didn't actually know about that strategy. So I think it's an example of a system that you've created actually teaching you something, which is quite a watershed moment for us. If you're interested in that work, this was then fully published in Nature and the front cover earlier this year, and we actually even released the code as well, so you can have a look at that and play with that yourselves.

现在我们转向 3D 游戏、机器人模拟器，当然我们对机器人感兴趣，但作为开发平台而非开发平台。现在只展示 3D 方面的一个内容。明年我们会有更多新消息宣布，但我会播放这个小视频：你们之前看到玩雅达利游戏的同一个智能体，现在在 3D 游戏中驾驶赛车绕赛道行驶。同样，这里的输入只有原始像素和方向盘控制，它仅通过原始驾驶经验学会了如何驾驶，甚至能以 200 公里/小时的速度超车。同样，仅从原始像素数据学习。所以我们正在向更高级的 3D 环境迈进，研究迷宫问题以及各种更复杂的路径规划问题。

Now we're moving on to 3D games, robot simulators, and of course we are interested in robots, but as a development platform rather than as a development platform. Now just show you one thing on the 3D stuff. We'll have a lot more announcements to make in the next year, but I'll show this little video of the same agent that you saw playing the Atari games actually now driving a racing car around a track in a 3D game. Again, the only inputs here are the raw pixels and the steering wheel controls, and it's learned just from raw experience driving the car around how to drive, and even do things like overtaking other cars at 200 km/h. Again, just from the raw pixel data. So we're now moving towards much more advanced 3D environments, where we're looking at maze problems and all kinds of much more complex pathfinding problems.

我在演讲开始时提到了神经科学，现在想再回到这个话题。我们经常说 DeepMind 构建的 AI 受神经科学启发，事实上，我们现在研究的许多领域都密切借鉴神经科学，以寻找关于大脑如何工作的新型算法。所以我们正在研究记忆、注意力概念、规划、导航和想象力。由于每个领域可能都需要一整场演讲来深入，我将只聚焦于想象力，因为我认为这与在座各位最相关，也是我博士期间的研究方向。事实证明，想象力相当依赖于大脑中一个叫做海马体的区域，在这张人脑图中，就是这里粉色的部分，位于大脑中心。海马体非常重要；它虽然很小，但非常关键，而且 50 多年来人们已经知道，如果海马体受损，就会导致失忆。所以海马体对情景记忆至关重要，但人们不知道海马体还有什么其他用途，例如，它是否参与想象力？我怀疑它可能参与，因为当我开始读博士、阅读关于记忆和海马体的文献时，我遇到了这些文献，它们将记忆描述为一个重构过程，而不是像录像带一样。记忆实际上就是这样工作的：如果你明天还记得这场演讲，它不会像存储在脑海中的录像带一样；实际上，你会从你之前拥有的各种事物和经历中组合它，比如其他讲座、也许之前参观大英博物馆的经历，以及今晚特定的内容。你的大脑所做的，以及海马体参与的，就是重构，将所有部分整合成一个连贯的整体，然后被大脑的其他部分识别为情景记忆。所以我在想：如果记忆是一个重构过程，那么如果我们把想象力看作一个类似的过程，但在这里它是一个构建过程——如果我们把记忆看作试图以大脑认为熟悉的方式组合你拥有的组件，那么创造力可能就是它的反面：你仍然在组合这些组件，但现在你试图创造一些新颖的、大脑判断为不熟悉的东西。所以我在想：如果记忆严重依赖海马体，那么想象力可能也严重依赖相同的大脑结构和相同的过程。我们决定测试这一点的方法，实际上是走遍全国，寻找海马体受损但仅限于海马体的患者。有非常罕见的疾病会导致这种情况，尽管像阿尔茨海默病确实会攻击海马体，但也会攻击其他大脑结构。我们需要的是只有这一大脑区域特定损伤的患者，然后我们测试了这些患者的想象力能力，而不是他们的情景记忆。我们给他们的是一项相当简单的想象任务：我们会给他们如下词语提示：想象你躺在白色沙滩上。

I spoke about neuroscience at the beginning of the talk and just want to come back and touch on that now. We talk a lot about the AI we build at DeepMind as being neuroscience-inspired, and in fact many of the research areas we're looking at now, we are looking to neuroscience very closely for inspiration about new types of algorithms as to how the brain works. So we're looking at memory, attention concepts, planning, navigation, and imagination. Now because each of those areas would probably need a whole talk to get into, I'm just going to focus on imagination because I think that's most relevant for the audience here, and is also what I did for my PhD. Now it turns out that imagination is quite dependent on an area of the brain called the hippocampus, which is actually here in pink at the center of your brain in this diagram of the human brain. The hippocampus is very important; it's quite a small brain region but it's very critical, and it's been known for more than 50 years that if you damage the hippocampus, then you become amnesic. So it's well known that the hippocampus is vital for episodic memory, but what wasn't known was what else the hippocampus was useful for, for example, was it involved with imagination? Now I suspected that it might be because when I started reading the literature on memory and hippocampus when I started my PhD, I came across this literature that was talking about memory as being a reconstructive process rather than like a videotape. That's actually the way that memory works: if you remember this lecture tomorrow, it won't be like a stored videotape somewhere in your mind; actually, you'll be combining it from all sorts of components of things and experiences that you've had before, other lectures, perhaps other visits to the British Museum, as well as specific pieces of content that are to do with this evening specifically. What your brain does, and the hippocampus is involved with, is reconstructing that, pulling all those parts together into a coherent whole which then is recognized by the rest of your brain as an episodic memory. So I was thinking: if memory works as a reconstructive process, then if we think about imagination as being a similar process but in this case it's a constructive process—if we think of memory as trying to put components that you have together in a way that your brain thinks looks and judges as familiar, perhaps creativity is the converse of that: you're still bringing together those components, but now you're trying to create something novel that your brain judges as unfamiliar. So I was thinking: if memory is heavily dependent on the hippocampus, then maybe imagination is also very heavily dependent on the same brain structure and the same processes. The way we decided to test this was actually by going around the country to patients who had damage to the hippocampus, but only the hippocampus. There are very rare sorts of diseases that cause that, although things like Alzheimer's actually do attack the hippocampus but also other brain structures. What we needed were patients that had only specific damage to this one brain region, and we tested those patients on their imaginative abilities rather than their episodic memory. What we did was give them a fairly simple imaginative task: we would give them word cues like the following: imagine you are lying on a white sandy beach.

想象一个美丽的热带海湾，尽可能详细地描述你周围能看到、听到和体验到的一切。当我们比较并分析他们的描述时，我们用一个非常复杂的评分系统来评估描述的丰富程度，然后将其与年龄、教育程度和智商匹配的对照组进行比较。这些对照组除了海马体完好健康外，其他方面都匹配。我们发现，在我们的丰富度指标（或指数指标）上，海马体受损患者所想象的场景比健康对照组要贫乏得多。你可以在这两个柱状图中看到这一点。左边是五名患者，右边是十名匹配的对照组。这是描述场景的丰富度指数，你可以看到患者相比对照组有巨大缺陷。所以，海马体确实对想象未来非常重要。实际上，《新科学家》报道了这项研究，它成为了一项重要研究，并被《科学》杂志列为 2007 年重大突破之一。《新科学家》有一个很好的标题，说海马体患者被困在当下。所以，我们知道他们无法很好地回忆过去，现在发现他们也无法想象未来。但如果你和他们交谈几分钟，他们会显得完全正常，你可以正常对话，因为他们处理当下的方式和健康人一样。

Beach in a beautiful tropical bay, describing as much detail as you can what you can see and hear and experience around you. And what happened was, when we compared and broke down their descriptions, we scored it in a very complex scoring system to break down how rich that description was, and we compared it to age- and education- and IQ-matched control subjects. So they were matched in every way except obviously they had intact and healthy hippocampi. We found that on our richness measure, or exponential index measure, the imagined scenes that the hippocampal patients were describing were hugely impoverished compared to the healthy controls. And you can see that in these two bar charts. On the left-hand side here is the patients, the five patients, and then here are the ten matched controls to these patients. And you can see this is the richness index of the described scene, and you can see that the patients are massively deficient compared to the controls. So it seems indeed that the hippocampus is very important for imagining the future. And actually, New Scientist reported on this work, and it became quite a big study that was also listed in Science as one of the big breakthroughs of 2007. And New Scientist had quite a nice headline talking about hippocampal patients being stuck in the present. So the idea was, we know they can't remember the past very well, and now it turns out they can't imagine the future either. But if you were to talk to them, for a few minutes they would seem completely normal to you, and you'd be able to converse with them completely normally because they're processing the present in the same way that a healthy person would.

随后我们用功能磁共振成像进行了跟进，发现海马体并非单独支持想象。它是一个更大脑网络的关键核心部分，包括从后压部皮层到内侧前额叶皮层、内侧颞叶、外侧颞叶皮层和顶叶皮层等各种脑区。当你扫描一个人并让他们想象场景时，所有这些区域都会可靠地激活。这个网络就是所谓的想象网络。

We then followed this up in fMRI, and we found, of course, the hippocampus doesn't support imagination on its own. It's a critical core part of a much larger brain network that includes all sorts of other brain regions, from the retrosplenial to the medial frontal cortex, to the medial temporal lobe, lateral temporal cortex, and parietal cortex. So all these regions reliably come on when you scan somebody in the brain scanner and get them to imagine scenes. And this network is the imagination network, if you like.

最近我在想，人类是这样想象的，那动物呢？它们也能想象吗？比如，老鼠能想象吗？我们知道老鼠有很好的记忆，但它们能想象未来吗？在我展示相关研究之前，我需要先介绍一些关于老鼠的知识。首先要说的是位置细胞。位置细胞可以看作是老鼠大脑中关于它们位置的 GPS 坐标。想象一个盒子，老鼠可能在里面，从俯视图看，老鼠在盒子里四处游荡。在这些实验中，你直接记录老鼠大脑的活动，你会发现海马体中的细胞在环境的特定位置放电。例如，细胞 A 只在老鼠穿过盒子的某个特定区域时放电，而细胞 B 只在另一个区域放电。发现这一点的杰出人物是约翰·奥基夫，他在 70 年代发现了这些位置细胞，就在 UCL 附近，去年他因此获得了诺贝尔奖。我很幸运，他是我博士答辩委员会的成员，所以我对他很了解，也对整个老鼠文献非常熟悉。

So much more recently, I was thinking, okay, so this is how humans imagine, but what about animals? Can they imagine as well? For example, can a rat imagine? We know that rats have very good memory, but can they do things like imagine the future? Before I show you the study we did to investigate that, I need to take you through a couple of things we know about rats. The first thing to tell you about is place cells. Place cells can be thought of as the GPS coordinates in a rat's brain about where they are. So if we imagine a box that a rat might be in, looking top-down on this environment, the rat might be roaming around this box. In these experiments, you record directly from the rat's brain while they're roaming around, and what you find is that cells in the hippocampus fire in specific places in the environment. For example, cell A might fire only when the rat is traversing through this particular part of the box environment. Conversely, cell B might only fire in another part of the environment. The amazing person who discovered this, John O'Keefe, who discovered these place cells in the 70s, just around the corner at UCL, won the Nobel Prize for this last year. I was lucky enough to have him on my vivA committee for my PhD, so I got to know him well and the entire rat literature very well.

我开始思考的是，自从发现位置细胞以来，人们发现当老鼠穿过环境时，位置细胞会按顺序放电。例如，考虑一个新环境，一个线性轨道。这是一个线性盒子，老鼠从左向右移动。如果你记录老鼠的大脑，你会发现位置细胞会按 A、B、C、D 的顺序放电，模拟老鼠穿过环境的方式。然后在 90 年代，其他人发现，当老鼠在迷宫环境中行走和导航后，在它们睡觉时记录大脑，老鼠会重放它们清醒时经历的轨迹。你可以把这看作是老鼠在做梦。它们会重放 A、B、C、D 的轨迹，但现在它们正在熟睡，只是大脑在重放。有趣的是，老鼠的大脑实际上以比真实体验快一个数量级的速度重放这些轨迹。所以，如果你认为做梦有助于老鼠了解它们所处的环境，那么它们实际上比清醒时更有效地学习。这表明老鼠有记忆，也许它们会做梦，但这并不表明它们清醒时能想象新的体验。

What I started thinking about is, since that discovery of place cells, people have found that sequences of place cells fire in sequence when a rat moves through an environment. For example, take a new environment here, a linear track. This is like a linear box, and the rat is moving from left to right. What you find if you record from the brains of these rats is that place cells will fire in order A, B, C, and D, mimicking the way the rat is moving through that environment. Then in the 90s, other people found that when they recorded from the brains of these rats while they were asleep after they had walked around and navigated in one of these maze-like environments, the rats would replay the trajectories they had experienced in their awake session before. So you could think about this as rats dreaming. They would replay this trajectory of A, B, C, D, but now they would be sleeping soundly, and it would be just their brains replaying this. What's interesting is that the brains of these rats actually replay these trajectories an order of magnitude faster than they actually experienced it in real life. So if you think about dreaming as helping the rats learn about the environments they're in, then they are actually learning from this much more efficiently than they can experience it when they're awake. So that shows that rats have memory and perhaps they dream, but it's not showing that they actually imagine new experiences while they're awake.

我们想明确地证明这一点，最近我和 UCL 的一些同事在《eLife》上发表了一项研究，我认为它明确地表明老鼠确实会想象。我们设计了一个简单而优雅的实验来验证这个假设。这次我们用的是 T 迷宫。同样，我们从俯视图看老鼠的环境。T 迷宫有一个屏障，这个屏障阻止了老鼠。老鼠从 T 迷宫的竖道开始，屏障阻止它进入横臂，但屏障是透明的，所以老鼠可以看到横臂上的东西。在最初的阶段，老鼠在竖道中来回跑，为了让老鼠对横臂产生兴趣，我们在其中一个横臂上放了一些食物，一个米粒，用黄点表示在右侧横臂上。老鼠到达屏障时能看到米粒，但够不到。所以它非常有动力去思考如何得到米粒，但无法越过屏障。在它体验了这个环境一段时间后，我们让老鼠睡觉，同时记录它的大脑活动。然后我们再次唤醒老鼠，把它放回环境中，但这次我们……

We wanted to show that unequivocally, and recently we published a study with some colleagues of mine at UCL in eLife that I think unequivocally shows that rats do imagine. So we designed this simple but elegant design to test this hypothesis. What we had here is a T-maze this time. Again, we're looking top-down on the rat environment. The T-maze has a barrier, and this barrier stops the rat. The rat starts off in the stem of the T-maze, and it stops the rat moving to the arms of the T-maze, but it's see-through, so the rat can see past to the arms and see what's on the arms. So the rat is initially in the first session running up and down the stem of the T-maze, and what we do to make the rat really interested in the arms is we put some food, a rice pellet, on one of the arms here depicted by this yellow dot on the right-hand arm. The rat can see this when it gets to the barrier, but it can't reach the rice pellet. So obviously it's very motivated to think about trying to get the rice pellet, but it can't get past the barrier. So after it has experienced that environment for a while, we let the rat go to sleep, and we are of course recording from the rat's brain while this is going on. Then we wake the rat up again and now we put it back in the environment, but this time we...

移除那个障碍，现在它可以在整个迷宫里自由跑动了。它很开心地移动着，在主干道和左右臂之间穿梭，完全探索了整个环境。显然，正如我刚才提到的位置细胞，我们发现有些位置细胞，比如细胞 A 和细胞 D，在迷宫的手臂区域放电。现在我们可以回头查看老鼠睡觉时收集的数据，看看老鼠是否在它实际体验之前就已经想象过通往米粒的轨迹。别忘了，老鼠睡觉时从未在这条手臂上走过，它只是看到过那条手臂。结果我们发现我们的猜想基本被证实了。实际上，如果我们回头分析睡眠数据，会发现这种回放，或者说是预演，轨迹 A-B-C-D。而且这不仅仅是随机的预演，向右臂的预演显著多于向左臂的预演，这正好符合行为上的意义，对吧？所以，虽然这是拟人化地描述老鼠，但你可以想象老鼠想要得到那颗米粒，它在想象如何到达那里，几乎是在想象自己走到米粒前吃掉它。然后它就在梦到这些想象的经验。我认为这就是实际情况。当然，我们还会进一步研究，计划在更复杂的环境中进行后续研究。所以，人类会想象，老鼠会想象，那么机器呢？这是关键：想象力对于规划未来、制定好的计划至关重要。这显然也是我们希望机器能够做到的。所以我把这张幻灯片命名为《仿生人会梦见电子羊吗？》，这当然是指菲利普·K·迪克的名著和我最喜欢的电影之一《银翼杀手》。这里我只给大家看一个很短的视频片段，让你一窥之前玩《太空侵略者》的机器的内心。我之前说过，智能体系统的目的之一是构建世界模型，以便预测游戏世界中将要发生的事情。在这个大约 10 秒的视频中，机器从《太空侵略者》得到初始输入——这是游戏画面——然后自由想象或梦见接下来 10 秒可能发生的事情。你可以看到它在梦见移动火箭、得分和射击一些太空侵略者。画面有点模糊，因为世界中的不确定性，一切都是概率性的，所以不像看到实际屏幕那样确定，但我认为这是基于想象的规划的雏形。

Remove that barrier so now it's free to run around the whole maze. It does that happily. It moves around the stem and the arms, both the left arm and the right arm, and it fully explores this whole environment. Obviously, as I've just told you with the place cells, what we can do is we find that there are place cells, let's say cell A and cell D, that fire on the arm section of the maze. Now what we can do is then go back to look at the data we collected when the rat was asleep and see if the rat was imagining about those trajectories towards the rice pellet before it ever had experienced it in reality. So don't forget, when it was sleeping, the rat had never experienced walking on this arm; it only seen that arm. And what happens is we find our conjecture was sort of proven. That actually, we find if we go back and analyze the sleep data, you get this replay, or pre-play if you like, of this trajectory A-B-C-D. And what's more, this is not just random pre-play; you actually get significantly more pre-play to the right-hand arm than to the left-hand arm, which is exactly what you would expect if it's behaviorally consequential. Right? So, I mean, this is anthropomorphizing the rat, but you could imagine the rat wanting to get to that rice pellet and is imagining plans of how could it get there, right? Almost imagining itself walking to the rice pellet and then eating it. And so then it's dreaming about its imagined experiences. So that's really what's going on here, I think. And of course, we're now going to look into this further. We have a number of plans to look at follow-on studies from this with more complex environments. So you know, humans imagine, rats imagine, so what about machines? This is something that's key: imagination to planning for the future, making plans, good plans about the future. So this is obviously something that we also want our machines to be able to do. So I've entitled this slide 'Do Androids Dream of Electric Sheep?', which is of course a reference to Philip K. Dick's famous book and one of my favorite films, Blade Runner. And this is really just going to give you a very short excerpt here with this video, which is a little bit of an insight into the mind of the machine that you saw earlier playing Space Invaders. So I told you that one of the purposes of the agent system is to build a model of the world so it can predict the future of what's going to happen in that game world. And here, what I'm going to show in this sort of 10-second video is the machine getting an initial input from Space Invaders like this—this is the game position—and then freely imagining or dreaming about what might happen over the next 10 seconds. So you can see it's dreaming about moving its rocket and it's dreaming about getting points and shooting some of the Space Invaders. Now it's quite fuzzy because there's uncertainty about what might happen in the world; it's all probabilistic. So it's not as certain as seeing an actual screen, but it's the beginnings, I think, of imagination-based planning.

最后我想谈谈想象力。一旦你开始思考，机器能有想象力吗？创造力呢？这是我经常被问到的问题，当然也与在座许多人的工作密切相关。我认为机器距离真正的创造力还很远，但我不认为这是不可能的。而且我认为，当我们开始理解创造力这个神秘过程时，如何用算法实现它会变得更加明显。所以我想展示几个可能会让你惊讶的小例子。我们拿一张大英博物馆正面的照片。然后我们对机器说——这是一种新型算法，最初由德国马克斯·普朗克研究所最近发明，我们在内部实现了自己的版本——你可以给它一张像梵高画作这样的艺术图片，然后要求将那张照片以梵高的风格重新绘制。结果你会得到这样的输出。它还没有达到可以挂在卢浮宫的水平，但你可以开始思考，当我看到这些东西时，输出的连贯性令人惊讶。然后我们可以看其他例子。实际上，这是一张正在硅谷建造的新谷歌园区的概念艺术图，我们给它一幅修拉的画作，然后要求它输出。这是我最喜欢的例子之一，它生成了相当不错的原始版本，但以修拉的风格呈现。这不是真正的创造力，对吧？从某种意义上说，这是算法性的，因为我们所做的是解构原始照片和画作的特征，然后交换这些特征，用画作的特征覆盖照片的特征，从而得到这类输出。但令人惊讶的是，这里并没有太多你认为是创造力的东西，却得到了非常有趣的输出。所以，当我们最终发现创造力是什么时，它可能并不像我们想象的那么神秘。

Now I'm just going to end by talking a little bit about so that's imagination. So once you start thinking, well, could machines have imagination? What about creativity? Now this is something I get asked about all the time, and of course is very relevant to a lot of the work that people in this room do. And I think we're a long way away from machines being truly creative, but I don't think it's impossible. And I think that when we start to understand what this process is, this mysterious process of creativity is, I think it will become actually more obvious how to implement that in an algorithm. So I just want to show a couple of little hints at things that might surprise you. So let's take a picture of the British Museum, the front of this building. And if we then say to the machine—and this is a new type of algorithm that was actually first very recently invented at Max Planck Institute in Germany, and then we've implemented our own version of this internally—and what you can do is you can give it an artistic picture like this Van Gogh picture and say you want that photo redrawn in the style of Van Gogh. And actually, you end up with outputs that are like this. It's not ready yet to be hung in the Louvre, but you can sort of start thinking it's pretty surprising when I saw these things like how actually coherent the output can be. And then we can look at other examples. Actually, this is a piece of concept art for a new Google campus that's being built in Silicon Valley, and we give it a Seurat painting and then we ask it to output it. This is one of my favorite ones, and it produces a pretty good version of the original but in the style of Seurat. And this isn't true creativity, right? In some sense, this is algorithmic because what we're doing here is deconstructing the features of both the original photo and the painting, and then we're swapping those features, overwriting the photo features with the painting features, and that gives these kinds of outputs. But the surprising thing here is that there isn't much sort of what you would regard as creativity here, and yet you get these very interesting outputs. So it may be that creativity isn't as mysterious as it seems to us when we ultimately find out what it is.

最后我想谈谈更大的图景，以及人工智能未来可能产生的影响。我认为我们社会面临的一些重大问题包括信息过载和系统复杂性。现在我们无论走到哪里，都被信息淹没。比如基因组学、大数据，但在电视娱乐领域，现在有那么多电视频道和观看方式，你如何真正找到自己感兴趣的内容？个性化是一种可能有所帮助的技术，但它目前基于相当原始的群体智慧协同过滤技术，无法提供真正符合你兴趣长尾的独特推荐。而在系统复杂性方面，我们想要掌握的系统——气候、疾病、能源、宏观经济学，甚至粒子物理学——现在变得如此复杂，以至于即使是最优秀、最聪明的人类专家团队也难以理解这些系统的含义并做出有用的预测。所以我认为解决智能问题、解决人工智能问题，可能是所有这些问题的元解决方案。如果我们能解决智能问题，那么也许我们就能……

Now I'm just going to end by talking a little bit about the bigger picture and sort of the impact that AI might have in the future. And so I think some of the big problems that are facing us as a society are information overload and system complexity. So everywhere we go now, we're deluged by information. So things like genomics, big data in general, but in the world of TV entertainment, I mean there are so many TV channels now and modes of watching things, how can you really find what it is that you're interested in? And personalization is one kind of technology that might help, but it doesn't really work because it's really based at the moment on quite primitive sort of wisdom of the crowds collaborative filtering technology, and that doesn't give you unique recommendations that are unique to your, what I would call, long tail of interests. And then in terms of system complexity, the kinds of systems we would like to master—climate, disease, energy, macroeconomics, even particle physics—are becoming so complex now that even teams of the best and brightest human experts are having difficulty comprehending the implications of these systems and actually making useful predictions about them. So I think solving intelligence, solving AI, is potentially a kind of meta-solution to all these problems. If we can solve intelligence, then maybe we can...

我们可以用它来帮助人类专家更好地掌握所有这些其他系统。我的梦想，我想用通用 AI 做的事情，就是构建 AI 科学家，或者让 AI 辅助科学成为可能。当然，如果我们拥有如此强大的东西，就需要考虑其使用的伦理问题。与所有强大的新技术一样，AI 与过去的许多技术并无不同。我们必须非常注意以道德和负责任的方式使用这些技术。虽然人类级别的 AI，我认为通用 AI，还需要几十年，但我们现在就应该开始辩论。我们正在通过内部伦理委员会以及支持学术工作和会议来做到这一点。最后，向神经科学致敬，以这种受神经科学启发的方式构建 AI，有助于我们更好地理解自己大脑的奥秘和运作。未来，当我们把智能提炼成算法结构，并与人类心智的能力进行比较时，我们将更好地理解自己心智的独特之处，比如做梦、创造力，也许还有伟大的意识问题。正如我心目中的科学英雄之一费曼所说：‘我无法创造的东西，我就无法真正理解。’

We can use it to help us, as human experts, get a better handle on all these other systems. My dream, the thing I'd like to use general AI for, is to build AI scientists or to make AI-assisted science possible. Of course, if we have something this powerful, we need to think about the ethics of its use. As with all new powerful technologies, AI is no different from many other technologies in the past. We have to be very cognizant about using these technologies ethically and responsibly. Although human-level AI, I think general AI, is many decades away, we should start the debate now. That's what we're doing with our internal ethics committees and by supporting academic work and conferences on these topics. Finally, with a nod to neuroscience, building AI in this neuroscience-inspired way helps us better understand the mysteries and workings of our own minds. In the future, as we distill intelligence into an algorithmic construct and compare it with the capabilities of the human mind, we'll better understand what's unique about our own minds, like dreaming, creativity, and perhaps the great consciousness question. As Feynman said, one of my all-time scientific heroes, 'What I cannot build, I do not truly understand.'

太棒了。暂停一下。刚才说得太好了。我们把灯打开，我要问一个问题，但不会独占话筒，然后交给观众。我们有 20 分钟时间，所以稍微停顿一下，想想你想问 Demis 什么。我有一个很简单的问题：你详细描述的那个 20 年计划——搞定你的国际象棋技能，虚拟环境中的游戏，然后你需要新的东西——但那只是前 20 年，而你刚才又说还需要几十年。人们对此非常着迷。从原始数据到构建环境图景——我们经历了 Atari 游戏，而另一端是 AI 科学家处理原始数据。在你所处的这一代，用你的话说，一个现实的登月计划是什么？

Brilliant. Take a pause. That was brilliant. If we get the lights up, I'm going to ask one question but not hog the limelight, and I'll hand over to the audience. We're going to have 20 minutes, so have a little pause, think about what you want to ask Demis. One question I've got, very simple actually: I was amazed by what you detailed, your 20-year plan—get your chess skills sorted, virtual environments in gaming, then you needed the new. But that was the first 20 years, and you just talked about decades away. People are pretty obsessed with that. Taking raw data and building the picture of the environment—we were at Atari games, and then the other end of the spectrum is the AI scientist taking raw data. Give us a sense, in the generation you're in, what's a realistic moon landing in your terms?

我在那张神经科学幻灯片中提到了其中一些事情。我认为除了 Atari 游戏之外，需要解决的重要大事——我们还在梯子的第一级。Atari 的事情意义重大，因为那是第一次有人构建了我们所谓的端到端智能体，它能够接收原始数据、做出决策，并在一个大的循环中完成这一切。我认为下一个重大突破将是：它能否真正学习抽象概念，超越单纯的感知输入，并对它所处的世界有真正的语义理解？那是我们的一个大目标。

I touched on some of those things in that neuroscience slide. The big things I think are important to solve beyond the Atari games—we're on the first rung of the ladder there. The Atari thing was significant because it was the first time anyone built what we call an end-to-end agent, something that took raw data, made decisions, and did that in one big cycle. I think the next big breakthroughs will be: can it really learn abstract concepts, go beyond just perceptual inputs, and have a real underlying understanding of the semantics of the world it finds itself in? That's our big kind of goal.

这会像计算机发展那样呈指数级增长吗？我们是否都会在 20 年内被进步的速度所震撼？我想知道 20 年内我们能达到什么程度。

Is this going to be exponential like the way computers developed? Are we all going to be bowled over in 20 years by the speed of progression? I want to know how much we can get in 20 years.

这很难预测，因为我们需要至少十几个真正重大的突破，而研究突破的时间尺度是出了名的难以预测。我认为未来几年我们会看到一些非常令人惊讶的事情，但至于我们能走多远，现在很难说。精确的时间尺度很难确定。

It's hard to predict because we need at least a dozen really huge breakthroughs, and research breakthroughs are notoriously hard to predict in terms of time scales. I think we'll have several very surprising things over the next few years, but as to how far we'll get all the way, it's too hard to say from here. Precise time scales for that are difficult.

蒂姆·马歇尔。几周前，我参加了一个会议，会上一个机器人正在执行前列腺手术——不仅仅是执行手术，它实际上能理解肿瘤并决定是否继续。今天我们在新闻中看到了提供医疗服务的挑战。你认为 AI 在健康领域的诊断和治疗中可以发挥什么作用？我们目前的医疗实践方式还是 19 世纪的范式。

Tim Marshall. A few weeks ago I was at a conference where a robot was performing a prostate operation—more than just performing the operation, it could actually understand the tumor and make a decision whether to proceed or not. Today we've seen in the news the challenges of providing healthcare. What role do you think AI can play in diagnosis and treatment in health? The way we practice medicine at the moment is a 19th century paradigm.

太好了。医疗保健实际上是我们首先关注的主要应用领域之一。我认为我们可以彻底改变护理质量和效率。正如你所说，我们还在使用 19 世纪的方法。以可消化、可操作的方式向外科医生和全科医生提供这类最新信息，将真正帮助整个医疗保健领域。这是我们未来几年打算大力投入的事情。

That's great. Healthcare is actually one of the main application areas we're focusing on first. I think we could revolutionize the quality of care and the efficiency of it. As you say, we're still using 19th century methods. Having this sort of latest information available in a digestible, actionable way to surgeons and GPs will really help the whole healthcare space. It's something we're looking to get heavily involved with in the next few years.

你为什么决定发布代码？从伦理角度，你是否担心代码会落入坏人之手？

Why did you decide to publish the code, and were you ever worried from an ethics point of view that the code might get into the wrong hands?

我们尽量对我们所做的事情保持开放。我们通常几乎发布我们所做的一切，并且在可能的情况下，我们也会开源一些东西。我们的神经网络库 Torch，我们在此基础上构建算法，就是开源的。《自然》杂志的审稿人和编辑要求我们发布代码，所以我们考虑了一下，并决定在这种情况下没问题。但这并不总是适用于我们所做的事情；我们必须逐案考虑。总的来说，在可能的情况下，我们喜欢参与并支持整个学术界。我们认为知识共享很重要，这是人类尽可能快速进步的方式。

We try to be as open as possible about what we're doing. We generally publish almost everything we do, and where we can, we open source things as well. Our neural network libraries called Torch, which we build our algorithms on top of, are open source. There was demand from Nature reviewers and editors to release our code, so we thought about it and decided in that case it was fine. But that may not always be the case for the stuff we do; we have to consider that on a case-by-case basis. In general, where we can, we like to engage and support the general academic community. We think it's important that knowledge is shared, and that's the way humanity can advance as quickly as possible.

你是地球上最聪明的人之一，现在谷歌收购了你们。我要问你一个你预料到的问题：史蒂芬·霍金对 AI 的担忧，一旦精灵从瓶子里出来，我们就完了。你们在做什么来试图遏制它？

You're one of the largest brains on the planet, and Google has now bought you. I'm going to ask you the question you're expecting: the Stephen Hawking thing, the concern about AI that once the genie is out of the bottle, we're all right. What are you doing to try to contain it?

这是……

This is the...

是的，当然。实际上，几个月前我和斯蒂芬·霍金就这个问题进行了一次长谈。我觉得他非常愉快。我们聊了一个小时，但他有太多问题。在我们讨论了我们具体如何应对之后，他感到相当放心。看，这里有一些关于自主学习系统的重大问题：我们应该给它们什么目标，给它们什么价值体系，如何确保这些正是我们想要的？有一些非常棘手的研究需要做。这方面的工作还不多，部分原因是还没有系统可以真正进行试验，所以一切都只是思想实验。大多数思考这个问题并为此担忧的人并不在人工智能领域；他们要么是哲学家，要么是其他著名科学家或实业家，但自己并不实际从事人工智能工作。如果他们从事人工智能工作，我认为他们会看到目前的问题要实际得多。很容易被那些几十年后的科幻场景冲昏头脑。我相信，随着我们构建更强大的系统，我们将对价值体系等问题有更好的答案，包括数学证明和实证工作，这将使我们更好地了解如何控制这些系统。

Yes, sure. I actually had a long chat with Stephen Hawking about this a few months ago. I think he was very enjoyable. We spent an hour together, but he had so many questions. I think he was quite reassured after we talked about how we were specifically approaching it. Look, there are big issues here about autonomous learning systems: what goal should we give them, what value system should we give them, how can we make sure those are exactly what we want? There are very tough pieces of research that need to be done. Not much work has been done in that area yet, partly because there have been no systems to really try this out on, so it's all been thought experiments. Most people thinking about this and worrying about it to that extent are not in the AI field; they're either philosophers or other famous scientists or industrialists, but not actually working in AI themselves. If they were working in AI, I think they would see that the problems are much more practical at the moment. It's easy to get carried away with science fiction scenarios that are many decades away. I have confidence that as we build more powerful systems, we'll have much better ideas about answers to these questions about value systems and so on, including mathematical proofs and empirical work that will allow us to have a much better idea of how to keep these systems under control.

我很高兴看到你们成立了伦理委员会，并且在思考这些问题，但你们拿了雅虎的钱，我对此很担心，因为你太聪明了。我希望你所做的一切实际上能改善社会，而不是把我们干掉。

Well, I'm pleased to see that you've got the Ethics Committee on board and you're thinking about these issues, but you have taken the Yahoo dollar and I am worried about this because you're so smart. I hope that everything you do actually improves society rather than kills us off.

绝对。但我认为，如果我们构建得当，人工智能可能成为人类最伟大的东西；我们将解决所有重大问题。关于企业责任与个人责任，我们花了很长时间对谷歌进行尽职调查。我们有其他选择，包括保持独立，我们决定联合，部分原因是谷歌高层同意伦理委员会等事项，并认为这是管理 DeepMind 技术使用的好主意。我们已经排除了军事或情报应用等明显用途。默认情况下，DeepMind 的东西不能用于这些目的。我们召开了伦理委员会的首次会议；委员会中有非常杰出的人物，其中许多是你提到的那些对此担忧的人，而不仅仅是持积极态度的人。其中很大一部分，因为我们还有几十年的时间，是开始教育每个人什么是真正的问题，并将事实与科幻分开。这是第一个起点。然后我们可以进入真正核心的技术难题。有一些难题，但我非常有信心，如果我们投入足够的脑力和时间，我们会解决这些问题。

Absolutely. But I think AI could be the greatest thing for humanity if we build it right; we'll solve all these big issues. Regarding corporate responsibility versus personal responsibility, we spent a long time doing due diligence on Google. We had other options, including staying independent, and we decided to join forces partly because the people high up at Google agreed with things like the Ethics Committee and thought it was a good idea to govern the use of DeepMind technology. We've already ruled out obvious things like military or intelligence applications. By default, DeepMind stuff cannot be used for those things. We've had our inaugural meeting of the Ethics Committee; there are very big luminaries on it, many of whom are some of the people you mentioned who are worried about this stuff, not just those who think positively about it. A big part of that, because we are decades away, is to start educating everyone on what the real issues are and separate fact from science fiction. That's the first starting point. Then we can get to the hub of the really core technical difficult questions. There are some, but I'm very confident that if we apply enough brainpower with enough time, we'll solve those problems.

太好了。你在谷歌已经大约一年了。你能给我们举几个例子或轶事，说明 DeepMind 如何改变了公司，接管了一些流程，改变了公司未来的运作方式吗？另外，关于伦理委员会，为什么你们没有公开或公布委员会成员是谁？

Great. You've been in Google now for about a year. Can you give us a couple of examples or anecdotes about how DeepMind has changed the company, taken over some processes, changed the way the company works going forward? And just to tag on to the Ethics Committee, why haven't you publicized or published who is on it?

第一个问题：我很高兴地说，几乎没有什么变化，而这正是关键。主要协议之一是我们的总部仍在英国，在国王十字附近。我们投资了那里的研究团队；整个 DeepMind 仍然在英国。我们作为一个半自主的单元运作。好处是我们获得的算力大大加速了我们的进展，还有谷歌的其他资源。DeepMind 如何影响了谷歌作为一家公司？这很难说。谷歌非常大，但我认为我们在某些方面影响了它。谷歌研究其他部门的工作方式发生了变化。谷歌研究有数千人，还有数千人从事机器学习，但我们有一个比其他地方更一致、更具体的使命。我们带来了一种更长期的研究重点，我认为谷歌现在想要更多这样的重点，这需要相当不同的组织架构和管理流程，其中一些现在正在山景城被采用。你的第二个问题关于为什么不公开：嗯，我们还在非常早期，这方面有很多关注。目前，这主要是关于教育人们，让大家了解这些问题。一旦你开始公开，那可能会改变辩论。我希望在给自己带来额外的公众审视之前，有一段安静、幕后、冷静、沉着的辩论期。在某个时候，我想我们会宣布这些人是谁，以及正在讨论的一些问题。话虽如此，我们已经做了很多公开的事情。在波多黎各有一个关于人工智能伦理和安全的大型会议，一月份在纽约大学还有一个我们赞助的会议，我是主题演讲者，我还与 Facebook、那里的人工智能负责人、微软和其他一些公司一起担任程序委员会成员。我认为明年的下一步将是创建一个跨行业小组，汇集所有大公司和学术实验室，除了我们自己的内部委员会之外。

First question: I'm pleased to say almost nothing has changed, and that's the whole point. One of the main agreements was that our headquarters is still in the UK, around King's Cross. We've invested in the research team there; the whole of DeepMind is still UK-based. We work as a semi-autonomous unit. The plus sides are the amount of compute power we have access to has really accelerated our progress, along with other resources from Google. How has DeepMind affected Google as a company? That's harder to say. Google is very big, but I think we have affected it in some senses. The way some other parts of Google Research do their work has changed. There are thousands of people in Google Research and thousands working on machine learning, but we have a more coherent, specific mission than the more applied machine learning elsewhere. We bring together a kind of longer-term research focus that I think Google wants more of now, and that requires quite different organizational structures and management processes, some of which are being adopted over in Mountain View now. Your second question about why we don't publicize: well, we're very early days, and there's a lot of scrutiny on this. At the moment, it's about simply educating people, getting everyone up to speed with the issues. Once you start making things public, that can change the debate. I wanted a period of quiet, behind-the-scenes, calm, collected debate before we brought on ourselves additional public scrutiny. At some point, I think we will announce who these people are and a little bit about the issues being discussed. Having said that, we already do lots of public things. There was a big conference in Puerto Rico about AI ethics and safety, another one at New York University in January that we're sponsoring and I'm keynoting, and I'm on the program committee along with Facebook, the heads of AI there, Microsoft, and some other companies. I think the next stage next year would be to create a cross-industry panel and bring together all the big companies and academic labs working on this, in addition to our own internal committee.

我们已经讨论过通用视频游戏机，对此我非常感谢。但当你进入下一个阶段，即想象力时，必然有无数个随机的坐标点，因为我认为我们每个人的想象都不同，以至于你会花无穷无尽的时间去分析这些，才能让机器以有规划的方式想象。那么你如何处理这些坐标呢？

We've done general video gaming machines, so I appreciate that very much. But when you go the next stage up to imagination, there must be so many individual random coordinates, if you like, because I think we all imagine differently, that you'll spend an infinite amount of time trying to analyze these to actually make machines imagine in a planned fashion. So how do you cope with these coordinates?

你做不到，我也会构建那样的系统。因为这里很多电视观众，我能看到，我猜他们有点不寒而栗：‘他在开玩笑，他是搞创意的’，这有点笨拙，你承认，就像把诗句混在一起。当第一位导演出现，第一位作家出现，环境中的那些点，选择变得指数级增长，对比太空入侵者的屏幕。就像棋盘上的米粒。所以我认为这里大多数人的工作很长时间内都是安全的。我不认为会有导演能执导像雷德利·斯科特那样质量的作品，那可能是计算机最后才能做到的事情之一。所以我们谈论的是非常受限的事情。人类比计算机做得更好的一件非常困难的事是：他们拥有这种基本的蛮力想象力，但人类的一大优势是审美判断。他们知道并非所有路径都平等，有些路径可能比其他的更有成果。即使比较国际象棋大师和计算机下棋，计算机可能要看数百万步才能做出一个决定，而国际象棋大师只看几百步，但都是经过深思熟虑的。在某种意义上，我们的大脑甚至会过滤掉低层次的脑力活动，任何不会产生有用结果的走法或轨迹。

You can't, I'd build that as well. Because a lot of the audience for the television audience here, I can see, and I suspect there was a bit of a shudder around: 'He's kidding, he's the creative bit,' which is a bit clunky, you admit, you know, kind of merging the verses. When the first director comes, when the first writer, when those points in the environment, the choices become exponential versus a space invader screen. It's like the rice on the chessboard. So I think most people's jobs in here are safe for a long while. I don't think there's going to be any directors directing something of the quality of Ridley Scott or something for you know, that's probably going to be one of the last things that a computer will be able to do. So we are talking about very constrained things. That's a very difficult thing that humans do better than computers: they have this rudimentary kind of brute force imagination, but one of the big things that humans do is they have aesthetic judgment. They know that not all paths are equal, and some are likely to be more fruitful than others. Even if you compare a chess grandmaster playing chess to a computer, the computer might look at millions of moves to make that one decision, whereas a chess grandmaster only looks at a few hundred, but judicious ones. In some sense, our brains even filter out low-level power brain activity, any of the kind of moves or trajectories that are not going to yield anything useful.

你能做到这种过滤吗？你们在研究过滤这部分吗？

Can you do that filtration? Are you working on that filtration part?

部分原因在于你对所处世界的建模能力。如果你能更好地建模世界，那么你就能更好地预测哪些事情值得花费算力去想象或思考。目前，我们知道我们在这方面还处于非常早期的阶段。

Part of that is to do with how well you model the world that you're in. So if you're better at modeling the world, then that means you should make better predictions about what are going to be useful things to spend your compute time imagining or thinking about. And at the moment, we know we're still very early stages of that.

来自第四频道的 David Abraham。你提到了人们在娱乐中面临的选择数量这一挑战。这是我们行业正在花大量时间思考的问题。你们是否更具体地研究推荐引擎领域，并且会代表谷歌利用该算法的力量吗？

David Abraham from Channel 4. You touched on the challenge of how many choices people have in entertainment. Something that we in our industry are spending a lot of time thinking about. Are you working more specifically on the area of recommendation engines, and are you going to be capturing the power of that algorithm on behalf of Google?

是的，我们实际上在研究各种形式的推荐系统。我认为这是一个非常有趣的领域，我们的技术也相当适用。同样，这关乎你是否能对用户的旅程和轨迹进行建模，从而提供更有吸引力的内容或推荐。我认为我们目前的系统还不够好。我们正在对此进行实验，同样处于早期阶段，我们正在为谷歌内部和外部合作伙伴研究这个。

Yes, we are looking at recommendation systems in all forms actually, all sorts of forms. I think it's a very interesting area and it's something that our technology is quite applicable to. Again, it's about whether you can model user journeys and trajectories through things in a way that delivers much more compelling content or recommendations. I think the current systems we have are not good enough. We're experimenting with that, again quite early days, and we're looking at that for things both internally at Google and external partners.

这对在座许多从事媒体行业的人来说非常有趣，在一个选择爆炸的世界里如何提供内容。几年内会有普遍应用吗？你认为会宣布一些产品吗？系统能让第四频道、BBC、其他广播公司、独立公司提供他们的内容？你认为这并不遥远？

That would be deeply intriguing to a number of us in the room who are working in the media business about how to serve up stuff in a world where choices have exploded. The general application in a few years, you think it'll be announcing stuff? Products? Systems by which the likes of Channel 4, the BBC, other broadcasters, independent companies can serve their content up? You think that's not far away?

是的，我认为在未来几年内，你会开始看到底层算法帮助这些推荐系统。然后大概四五年后，会出现全新的系统，你可以用与现在不同的方式与之互动。

Yeah, I think in the next couple of years you'll start seeing under the hood algorithms helping these kinds of recommendation systems. And then maybe four or five years out, actual totally new systems that you might interact with in a different way to we do now.

来自 IBM 的 Daniel。我认为我们是 AI 领域的另一个大投资者。我喜欢你关于重大突破的评论。这可能需要不止一个玩家才能把我们带到未来。你的演讲中有几个问题让我很感兴趣。一个是围棋。几年前，我认为围棋是一个重大挑战，是计算机难以解决的问题。对这里的观众来说可能有点深奥。第二个问题，直接一点，对于 RTS，AI 如何被用来不是取代创造力，而是增强和支持，让我们作为人类能做更多有创意的事情，而不是让计算机来做？

Daniel from IBM. I think we're one of the other big investors in AI. I liked your comments around the big breakthroughs. This probably takes more than one player to get us to the future space. I had a couple of questions that intrigued me in your talk. One was Go. Years ago, I think there was a big effort around Go; it was the big one that was hard for computers to solve. It might be a bit esoteric for this audience. And then the second question, just to bring us to the point, is for RTS, how can AI be used not to supplant creativity but to enhance and support and allow us to do more creative things as humans, not as computers?

先快速回答第一个问题，如果确实如此，你可以之后再问。围棋，对于不了解的人来说，是一种东方棋盘游戏，可能是最复杂的游戏。中国和日本的人玩这个而不是国际象棋。计算机难以攻克的一个原因是分支因子，每一步的选择数量大约有 100 种，而国际象棋只有 20 种左右。所以当你开始规划时，分支因子会爆炸。如果你想用蛮力方式，宇宙中的原子都不足以描述围棋的棋局数量。你很擅长围棋，对吧？我围棋下得还不错。你是世界冠军吗？我确实是围棋世界冠军。但第二个问题是，在国际象棋中，因为这是一个非常物质化的游戏，皇后比车值钱等等，很容易手动编写一个评估函数来判断你的程序是赢是输。而在围棋中，所有棋子价值相同，所以你是否赢更多取决于棋盘的整体格局。从某种意义上说，这是一个更美的游戏，非常具有审美愉悦感，但手动编写评估函数要难得多。所以我们将在围棋上发布一些非常重要的公告。自从深蓝击败卡斯帕罗夫以来，这 20 年来一直是 AI 研究界的圣杯。最后一个问题是关于帮助创造力。对于这里的观众来说，这其实是同一回事。

Just do a quick one on the first one, you might want to do that afterwards if it's really so. Go, for those you don't know, is an oriental board game which is probably the most complex game there is. It's what they play in China and Japan instead of chess. One reason it's been so hard for computers to crack is that the branching factor, the number of choices you have in each move, is on the order of 100, whereas in chess it's more like 20. So as you start planning, that branching factor explodes. If you're going to do it in a brute force way, there aren't enough atoms in the universe to describe how many Go positions there are. You're good at Go, aren't you? I'm reasonably good at Go. Were you world champion? I was world champion at Go. But then the second problem is that in chess, because it's a very materialistic game, the queen is worth more than a rook and so on, it's quite easy to hand-program an evaluation function to tell you whether your program is winning or losing. In Go, all the pieces are worth the same, so whether you're winning or not is much more about the overall pattern of the board. It's a much more beautiful game in some sense, very aesthetically pleasing, but much harder to hand-code an evaluation function. So we have some very big announcements to make on Go. It's been the holy grail for the AI research community for the last 20 years since Deep Blue beat Kasparov. And then the last question was on helping creativity. For this audience, that's really the same thing.

在科学领域也是如此，对吧？还有医生，以及我真正在思考的：AI 以更易消化的方式为你呈现正确的信息，这样你就可以利用它做任何事。在媒体领域，我们谈到了推荐。你脑海中还有其他关于创意过程或媒体创作的想法吗？我认为那要难得多。所以推荐是显而易见的。我们正在研究音乐，这对计算机来说比视觉更受限，而视觉极其困难。在音乐创作、音乐分析方面有一些非常有趣的工作，我认为很有前景。所以我想那会是下一个领域。

Thinking about in science too, right? And with doctors, and with what I'm really thinking: AI surfacing the right information for you in a much more digestible way, so you can just leverage that for whatever it is. The world media, we talk about the recommendation area. Is there anything else in your head that might spring to mind in terms of the creative process, the creation of media? I think that's a lot tougher. So I think recommendation is the obvious one. We are looking at things like music, which is a more constrained domain for a computer than visuals, which is incredibly hard. There's some very interesting work being done in music composition, music analysis, which I think is pretty promising. So I would imagine that would be the next place.

很好。我再问三个问题，因为我们真的没时间了。抱歉，我们本来可以整晚继续的。请那位先生提问。有人有麦克风吗？没有的话我们就用话筒。那位先生，后面还有人想提问吗？我一直关注前排。我们请最后那排末尾的先生。

Very good. I'm going to take three more because we're really running out of time. I'm sorry, because we could just keep going all night here. Take the gentleman here. Has anyone got a microphone on them? Because no, we'll just get the mics. The gentleman here, anyone at the back want to have a... because I've been very front focused. We'll take the gentleman on the end of the row right at the back there.

你好，我是 IBM 的 Nara。在我们的文化历史中，我们一直被对痛苦和快乐的追求所塑造，这塑造了我们的思维和行为方式。你如何教会机器痛苦和快乐？

Hi, Nara from IBM. We've been shaped from the search of pain and pleasure during our cultural history, and this shapes the way we think and behave. How can you teach pain and pleasure to a machine?

嗯，一个问题是我们是否需要，以及是否应该。我们内部有一个概念叫内在动机。你知道，情绪和其他东西驱动人类行为，而不仅仅是外部奖励。目前，我们的机器没有类似的东西。但也许为了完成更复杂的任务，我们正在游戏世界中工作，那里大多数时候方便地有分数。但即使你开始进入更复杂的游戏，像《我的世界》这样的开放式游戏，现在没有分数了。那么你如何决定该做什么，什么是有用的，什么是好的，什么是进步？我这里说的是智能体系统。当然，这更像我们人类的现实世界，然而我们有自己的内在驱动力，可能是进化而来的，帮助影响我们的行为。所以我认为思考这个问题很有趣。我们有神经科学家或这些领域的专家作为顾问与我合作，我对这个非常着迷。但我们还没有明确的答案。谢谢。

Well, one question is whether we need to, but also whether we should. There is this idea that we look at internally of intrinsic motivation. You know, there are emotions and other things that drive human behavior, not just external rewards. At the moment, our machines don't have anything like that. But maybe to do more complex tasks, you know, we're working in game worlds where there is conveniently a score most of the time. But even if you start going to more complex games, open-ended things like Minecraft, now there isn't a score anymore. So how are you going to decide what you should do, what is useful, what is good, that you're making progress? I'm talking about the agent system here. Of course, that's more like the real world for us as humans, and yet somehow we have our own internal drives, probably evolved, that help influence our behavior. So I think it's interesting to think about. We have neuroscientists or experts in these areas who work with us as consultants, and it's something I'm very fascinated by. But we don't have a definite answer on that yet. Thank you.

好的，还有两个问题。我们看看后面那位先生。我想后面还有人想提问。我看到有人举手了。是的，就是他，最后面那位。然后我们请穿红衣服的先生作为最后一个问题。别紧张。

Right, two more. So we'll check the gentleman at the back. I think he's someone still want to a question at the back. I got a hand up there. Yes, there he is, right at the back there. I just feel and then we'll take the gentleman in red as the last question. No pressure.

我的问题实际上和上一个很相关。我在想，你有没有考虑过泛化目标？你谈到了如何有观察和行动，但大概你定义了目标并告诉系统如何衡量目标。如果我们认为 AI 是人类智能的仆人，那么你有没有考虑过 AI 可以从环境中推导出目标？AI 能否像人类一样，我们分割目标，可能有生活目标，但专注于子目标来实现它。而且，不仅仅是，我有点想象人类对系统说‘你能帮我做这个那个’，AI 能够从它听到的东西中推导出目标，但更进一步，能够在被明确指示之前推导出目标。所以能够预测人们可能希望从 AI 那里得到的目标。

My question leads on quite well from the last actually. I was thinking, have you thought about generalizing the goals? So you talked about how you have observations and actions you take, but presumably you define the goal and tell the system how to measure its goals. If we're thinking about AI as something which is a servant to human intelligence, then have you thought about AI which can derive its goals from the environment? Can AI also, as humans we segment our goals, we might have life goals but we focus on subgoals to get there. Also, not just you know, I'm sort of imagining a human saying to a system 'can you help me with this and that' and the AI being able to derive its goal from the things that it hears, but also going further than that, being able to derive goals before they're specifically instructed. So being able to anticipate goals that people might want out of AI.

没错。我的意思是，这是一个很好的问题，实际上这是一个迷人的研究领域：机器能否学习自己的目标，但通过观察你，例如通过观察你了解你喜欢什么，然后尝试甚至能够先发制人地猜测你需要什么，甚至在你提出要求之前。所以我认为这些系统非常有趣。我的意思是，即使那样，你仍然会有某种顶级目标，即满足用户，对吧？尽管它可能学习子目标是什么。这是另一个非常活跃的研究领域：如何自动将一个大目标分解为子目标？当然，这是我们大脑毫不费力就能做到的事情。你知道，如果你要计划从这里去巴黎旅行，你的大脑不会计划从这到巴黎的肌肉纤维运动，对吧？但这就是目前机器人学的工作方式：它们没有层次定义。你的大脑实际做的是，在高层，你需要到达机场航站楼，然后坐火车去那里，等等。只有当你从椅子上站起来时，你的大脑才会解包肌肉纤维运动来站起来走路。而目前，因为我们没有解决自动生成子目标的问题，一个试图完成那个任务的机器人必须规划从原始动作运动一直到巴黎，当然这是不可行的，因此也是不可处理的，因为它最终会变成，就像另一位先生的问题，你想象所有那些从这开始的肌肉纤维运动路径，基本上有无限多条。所以我们需要这个：我们需要解决的关键问题之一就是子目标问题。

That's right. I mean, that's a great question, and actually it's a fascinating research area: can the machines learn their own goals, but through observation of you, learn what it is you like through observing you, for example, and then try to maybe even be able to preemptively guess what it is that you need before you even ask for it. So I think those systems are very interesting. I mean, even there you would still have some kind of top-level goal which is to satisfy the user, right? Although it may learn what the subgoals are. And that's another very active area of research: how do you break down a large goal into subgoals automatically? Of course, that's something our minds do effortlessly. You know, if you're going to plan a trip to Paris from here, your brain is not going to plan over your muscle fiber movements all the way from here to Paris, right? But yet, that's how robotics works at the moment: they have no defining of hierarchy. What your brain actually does is, at a high level, you need to get to the airport terminal, then take a train there, and so on. Then only at the point where you get up off that chair does your brain then unpack the muscle fiber movements to get up off a chair and walk. Whereas at the moment, because we haven't solved this problem of automatically generating subgoals, a robot trying to do that task would have to plan over the primitive action movements all the way to Paris, and of course it's not feasible and therefore not tractable, because it ends up becoming, you know, speaks to the other gentleman's question about you imagine all those paths from here on muscle fiber movements, there's basically an infinite number of them. So we need that: one of the key things we need to solve is the subgoal problem.

最后一个问题。哦不，麦克风，抱歉，为了让大家听到。Malcolm，广播工程师，接着之前的一个问题。在你的记忆、导航、想象力等列表中，你没有提到情绪。既然我们要处理与人类的互动，这对你的伦理和参数冲突有多大的爆炸性？

The last question. Oh no, Mike, sorry, just to get that across. Malcolm, har broadcast engineer, following on from an earlier one. In your list of memory, navigation, imagination, etc., you didn't have emotions. As we're going to be dealing with interacting with humans, how explosive is that with your conflicts of ethics and parameters?

是的，我的意思是，这又和下面那位先生关于情绪的问题有关。我们目前系统中没有与之等价的东西。但如果你考虑情绪，可能这太简单了，但情绪的一部分是内在驱动力。如果它们给了我们内在驱动力，那么这就是我们需要探索并尝试找出可能需要哪些驱动力的事情。可能我们需要类似的驱动力，这样这些系统才能与人类共情。显然，这是一部很棒的第四频道剧集，我想在座有些人看过《人类》，我非常喜欢，那很有趣：他们试图与他们服务的人类共情。或者，你可能想要拥有非常不同类型的系统。

Yeah, I mean again, this relates to the question of the gentleman down here about emotions. We currently have no equivalent of that in our systems. But if you think of emotions, and probably this is too simplistic, but part of emotions are internal drives. If they give us internal drives, then that's something we do need to explore and try and work out what ones might be needed. And it might be we need similar ones so that these systems can empathize with humans. Obviously, this is a great Channel 4 series, I think some people in the audience of 'Humans' which I really love, and that's interesting: they're trying to empathize with the humans that they serve. Alternatively, you might want to have systems that have very different types.

所以我可以想象我们可能需要两种类型的 AI。情感……最后问一下，你认为 AI 必须将情感作为驱动力来应对吗？

So I could imagine we might need both types of AI. Emotion... just to finish on that, do you think it's essential that AI will have to cope with emotion as a driver of, if you like, the response?

我认为某些类型的情感。你可能需要它们有两个原因。一是这样这些系统能够共情，与人类更好地携手合作。另一个是，如果它们所处的环境没有很多外部奖励信号，它们就必须有一些内在驱动力来引导它们朝着正确的方向前进。

I think some types of emotions. There are two reasons you might want them. One is so that these systems can empathize and work better hand in hand with humans. The other thing is if it turns out the environments they're in don't have many external reward signals, they have to have some internal drive to get them going in the right direction.

太棒了。这是我的荣幸。我马上要把话筒交给 IET 主席 Naomi Kimer。但 Demis，非常高兴你能来。谢谢。

Brilliant. It was a privilege. I'm going to hand to Naomi Kimer from the IET, the president of the IET, in a second. But Demis, it's been wonderful to have you here. Thank you.

谢谢。作为 IET 主席，我们非常高兴能与皇家电视学会共同主办这次讲座，也很高兴能邀请到像您这样杰出的人士，Demis。像今天这样的活动实现了 IET 慈善使命的一部分，即激励、告知和影响人们。我不知道你们怎么想，但您确实激励、告知并影响了我，让我相信 AI 将改变世界。我喜欢您的 AI 之旅始于游戏，机器学习通过玩耍完成，这与人类非常相似。我很欣赏您让这一切听起来如此直截了当，从我们的海马体到想象老鼠，再到对机器创造力的追求。您使命的第二部分——用 AI 解决其他一切问题——听起来非常合乎逻辑，似乎很合理。您追求 AI 科学家真正应对一些重大挑战。所以听您演讲，AI 能为人类带来积极改变听起来非常真实和可行。即使它不会很快执导电影，但请再次和我一起感谢 DeepMind 创始人 Demis Hassabis，带来了一场绝对精彩的讲座。我还要感谢我们充满活力的主持人 Tim Davey，BBC Worldwide 的 CEO，他做得非常出色。我想我们都受到了激励、告知和影响。能参与其中我感到非常高兴。非常感谢你们的到来。你们会很高兴听到外面已经准备了饮品。谢谢。

Thank you. As the president of the IET, we are absolutely delighted to have co-hosted this lecture with the Royal Television Society, and delighted to have someone of your extraordinary caliber, Demis. Events like today fulfill part of the IET's charitable remit, which is to inspire, inform, and influence people. I don't know about you, but you sure as hell have inspired, informed, and influenced me on a topic that I believe is going to change the world. I love the idea that your AI journey started with games and that machine learning is done through play, pretty much the same as it is for humans. I've enjoyed the way that you've made it all sound pretty straightforward, actually, from our hippocampus to imagining rats, to the quest for machine creativity. It just sounded quite logical that your journey to the point two of your mission, which is to use AI to solve everything else, seems quite reasonable. Your quest for the AI scientist to really tackle some of those big important challenges. So listening to you, it all sounds incredibly real and feasible that AI can make a positive difference to humanity. And even if it's not going to be directing any movies anytime soon, so once again, please join me in thanking Demis Hassabis, founder of DeepMind, for an absolutely fantastic lecture. And I would also like to thank our feisty chair, who did an excellent job, Tim Davey, the CEO of BBC Worldwide. I think we've all been inspired, informed, and influenced. So it's been very nice for me to be part of this. Thank you very much for coming. You'll be delighted to hear that drinks are now served outside. Thank you.