诺奖得主、DeepMind创始人：AI离彻底改变人类社会的能力只有10年
Nobel laureate and DeepMind founder: AI is only 10 years away from having the ability to completely change human society

文|opencat Text|opencat

本文转自：AI寒武纪 This article is reprinted from: AI Cambrian

11.22

知识分子 Intellectuals

The Intellectual

● ● ●

谷歌DeepMind CEO Demis Hassabis说，将人工智能视为普通技术是错误的，人工智能将具有“划时代的意义”，很快将治愈所有疾病、解决气候和能源问题并丰富我们的生活，AGI大概需要 10 年时间，因为还需要 2 到 3 项重大创新，下一项就是基于代理的系统，能够完成你给它的特定目标或任务。
Google DeepMind CEO Demis Hassabis said that it is wrong to view artificial intelligence as ordinary technology. AI will have "epoch-making significance" and will soon cure all diseases, solve climate and energy issues, and enrich our lives. AGI will probably take 10 years because 2 to 3 major innovations are still needed, the next being agent-based systems capable of completing specific goals or tasks given to them.

谷歌DeepMind创始人、CEO Demis Hassabis 最近在《泰晤士报》和《泰晤士报商业版》的主办的科技峰会上发表演讲，Hassabis回顾了DeepMind的创立，谈了 AGI、AlphaFold 和 AI 的未来。
Google DeepMind founder and CEO Demis Hassabis recently gave a speech at the technology summit hosted by The Times and The Sunday Times Business, where he reviewed the founding of DeepMind and discussed AGI, AlphaFold, and the future of AI.

照例先给大家划个重点（访谈全文附在文后）： As usual, let's highlight the key points for everyone (the full interview is attached at the end of the article):

Demis Hassabis看见了什么？ Hassabis已经在游戏的微观世界中看到了一点，并且理解得很清楚：从一个随机的系统AlphaZero开始8小时就可以训练出超越最顶尖人类的国际象棋实体，虽然这只是游戏狭窄领域，但一定会扩展出世界模型。
What did Demis Hassabis see? Hassabis has already seen a bit in the microcosm of games and understands it very clearly: starting from a random system, AlphaZero can be trained in 8 hours to surpass the top human chess players. Although this is just a narrow field of games, it will certainly expand into a world model.

DeepMind 的初心： Hassabis 30 年前就开始研究 AI 了！从游戏 AI 到神经科学，他一直坚信 AI 的潜力。2010 年，他创立 DeepMind，因为他看到了深度学习和强化学习的巨大潜力，以及 GPU 等硬件的快速发展。他想打造一个通用的、能自我学习的 AI 系统，这正是 DeepMind 的初心！
The original intention of DeepMind: Hassabis started researching AI 30 years ago! From game AI to neuroscience, he has always believed in the potential of AI. In 2010, he founded DeepMind because he saw the great potential of deep learning and reinforcement learning, as well as the rapid development of hardware like GPUs. He wanted to create a general AI system capable of self-learning, which is exactly DeepMind's original intention!

游戏 AI，AGI 的“练兵场”： DeepMind 早期专注于游戏 AI，是因为游戏可以快速验证算法的有效性，而且容易进行基准测试。但他们的目标不仅仅是赢得游戏，而是开发通用的 AI 技术，并将其应用于其他领域，例如科学和商业。
Game AI, the "training ground" for AGI: DeepMind initially focused on game AI because games can quickly verify the effectiveness of algorithms and are easy to benchmark. But their goal is not just to win games; it is to develop general AI technology and apply it to other fields, such as science and business.

AlphaFold：AI for Science 的典范： Hassabis 一直对用 AI 解决科学难题充满热情，而蛋白质折叠问题是他最想攻克的目标之一。AlphaFold 的成功（Hassabis因AlphaFold 获得2024诺贝尔化学奖），证明了AI 在科学领域的巨大潜力！
AlphaFold: A model of AI for Science: Hassabis has always been passionate about using AI to solve scientific problems, and the protein folding problem is one of his main targets. The success of AlphaFold (Hassabis won the 2024 Nobel Prize in Chemistry for AlphaFold) demonstrates the great potential of AI in the scientific field!

多模态模型，AGI 的关键： Hassabis 认为，多模态模型是 AGI 系统的关键组成部分，例如 DeepMind 的 Gemini 模型，它可以处理文本、图像、音频、视频和代码等多种输入。
Multimodal models, the key to AGI: Hassabis believes that multimodal models are a key component of AGI systems, such as DeepMind's Gemini model, which can handle various inputs like text, images, audio, video, and code.

通往 AGI 的道路：更强大的 Agent： 现在的聊天机器人大多是被动的问答系统，而未来的 AI 系统需要更主动、更智能，能够像 AlphaGo 一样进行规划和推理，并在现实世界中采取行动。
The Road to AGI: More Powerful Agents: Most current chatbots are passive Q&A systems, but future AI systems need to be more proactive and intelligent, capable of planning and reasoning like AlphaGo, and taking action in the real world.

AGI 时代，还有多远？ Hassabis 预计，我们距离 AGI 还有大约 10 年的时间。
How Far is the AGI Era? Hassabis estimates that we are about 10 years away from AGI.

DeepMind 的未来： DeepMind 将继续以研究为导向，同时也会加大产品研发的投入，与谷歌的其他部门合作，将 AI 技术应用于更多产品和服务中。
The Future of DeepMind: DeepMind will continue to be research-oriented while also increasing investment in product development, collaborating with other Google departments to apply AI technology to more products and services.

AGI 时代，人类将进入富足时代！Hassabis 认为，AGI 将彻底改变经济和社会，消除能源和资源的稀缺性，让人类进入一个物质极大丰富的时代。我们需要提前思考如何分配这些财富，例如，是否应该实行全民基本收入制度。
The AGI Era: Humanity Will Enter an Era of Abundance! Hassabis believes that AGI will fundamentally change the economy and society, eliminating the scarcity of energy and resources, and ushering humanity into an era of material abundance. We need to think ahead about how to distribute this wealth, for example, whether a universal basic income system should be implemented.

访谈全文：强烈推荐 Full interview: Highly recommended

注意:这是Demis Hassabis10月1日的访谈，此时距离10月9日他获得2024年诺贝尔化学奖还有几天时间，但是访谈视频今天才放出来。
Note: This is an interview with Demis Hassabis on October 1st, a few days before he won the 2024 Nobel Prize in Chemistry on October 9th, but the interview video was only released today.

主持人： 我想，在座的各位几乎都知道DeepMind，也知道它现在在做什么。让我们先简单回顾一下您的故事，因为您在2010年左右创立了DeepMind，而在此之前，人工智能经历了40年的寒冬，作为一名科学记者，我当时并没有关注人工智能。DeepMind为何在那个时候出现？是有什么有利因素吗？
Host: I think almost everyone here knows about DeepMind and what it is doing now. Let's briefly review your story, because you founded DeepMind around 2010, and before that, artificial intelligence had gone through a 40-year winter. As a science journalist, I wasn't paying attention to AI at the time. Why did DeepMind emerge at that time? Were there any favorable factors?

Demis Hassabis： 嗯，我研究人工智能实际上已经超过30年了，最初是做游戏，为游戏设计人工智能，以及模拟游戏。后来我学习了计算机科学和神经科学，并且一直在观察人工智能领域的发展。在您提到的90年代的人工智能寒冬时期，都是逻辑系统，也就是所谓的专家系统。你们很多人可能还记得深蓝在国际象棋比赛中击败了加里·卡斯帕罗夫（俄罗斯国际象棋棋手，国际象棋特级大师，前国际象棋世界冠军），这些都是预编程系统，实际上是程序员和系统设计者解决了问题，并将其封装成规则。计算机、人工智能系统实际上根本不智能，它只是在执行这些启发式方法。这样做的问题是，最终会得到脆弱的系统，它们无法学习新东西，当然也无法发现新东西，因为它们显然天生就受到设计者或程序员已知能力的限制。
Demis Hassabis: Well, I've actually been studying artificial intelligence for over 30 years, initially working on games, designing AI for games, and simulating games. Later, I studied computer science and neuroscience and have been observing the development of the AI field. During the AI winter of the 90s you mentioned, it was all about logic systems, the so-called expert systems. Many of you might remember Deep Blue defeating Garry Kasparov (a Russian chess player, chess grandmaster, and former world chess champion) in chess matches. These were pre-programmed systems, where programmers and system designers solved the problems and encapsulated them into rules. Computers and AI systems were not truly intelligent; they were just executing these heuristic methods. The problem with this approach is that you end up with fragile systems that cannot learn new things, and certainly cannot discover new things, because they are inherently limited by the known capabilities of the designers or programmers.

所以对我来说，很明显，在整个90年代，我在剑桥和麻省理工学院学习期间，这仍然是主流观点，尤其是在那些地方，逻辑系统才是正道。我认为这就是出现很多人工智能寒冬的原因，因为它们天生就脆弱且局限。所以在2010年，DeepMind的想法是，我们可以看到深度学习刚刚在学术界被发明出来。
So for me, it was clear throughout the 90s, during my studies at Cambridge and MIT, that this was still the mainstream view, especially in those places, that logic systems were the right path. I think this is why there were many AI winters, because they were inherently fragile and limited. So in 2010, the idea of DeepMind was that we could see deep learning had just been invented in academia.

强化学习是我们发现的东西，大脑中的多巴胺系统，动物和包括人类在内都使用强化学习来学习。
Reinforcement learning is something we discovered, the dopamine system in the brain, used by animals and humans alike to learn.

因此，对我来说，显而易见的是，我们需要构建的是一个能够自学且通用的学习系统，这就是DeepMind的起源。然后我们也看到了GPU和硬件加速等技术的进步。所以我使用了第一代GPU，它是用于计算机图形、计算机游戏的，但它们是非常通用的，事实证明，世界上的一切都是矩阵乘法。我们很早就开始了，我们觉得这就像一个阿波罗计划，需要付出巨大的努力才能将所有这些新奇的想法和成分整合在一起可以取得非常快的进展，结果也确实如此。
Therefore, it is obvious to me that what we need to build is a self-learning and general learning system, which is the origin of DeepMind. Then we also saw advancements in technologies like GPUs and hardware acceleration. So I used the first generation of GPUs, which were for computer graphics and computer games, but they were very versatile, and it turned out that everything in the world is matrix multiplication. We started early, and we felt it was like an Apollo program, requiring tremendous effort to integrate all these novel ideas and components to make very rapid progress, and indeed, that was the case.

主持人： 这是您在普林斯顿时期预想的结果吗？您是否想过15年后，我会在这里与您对话，人工智能会成为热门话题，并且蛋白质折叠问题会被解决？
Host: Is this the result you envisioned during your time at Princeton? Did you ever think that 15 years later, I would be here talking to you, that artificial intelligence would become a hot topic, and that the protein folding problem would be solved?

Demis Hassabis： 实际上，它大致沿着我们计划的路线发展，当然，过程中也有一些小插曲和意想不到的事情，但当我们在2010年开始时，我们认为要达到通用人工智能大约需要20年的时间。我认为我们可能距离这个目标还有10年左右的时间。从现在开始，大致是那个时间线，用人工智能系统进行科学研究，在通往人工智能的道路上解决科学问题一直是我的主要热情所在。蛋白质折叠一直是我最想解决的科学难题之一，如果我们能够取得突破，它将带来变革。
Demis Hassabis: Actually, it has developed roughly along the lines we planned. Of course, there have been some minor detours and unexpected events along the way, but when we started in 2010, we thought it would take about 20 years to achieve general artificial intelligence. I think we might be about 10 years away from that goal. From now on, that's roughly the timeline, and conducting scientific research with AI systems has always been my main passion on the road to AI. Protein folding has been one of the scientific challenges I most wanted to solve, and if we can make a breakthrough, it will be transformative.

主持人： 好的，让我们回到这一点，我认为我们也应该谈谈人工智能，因为有趣的是，自从ChatGPT出现以来，我们作为一个社会一直在非常深入地讨论人工智能，它与您一直在做的人工智能是截然不同的，作为一名观察者，您的人工智能一直都非常具体，观察它有点奇怪，你知道，它开始做一些毫无意义的事情。它非常擅长电脑游戏。
Host: Okay, let's get back to that point. I think we should also talk about artificial intelligence because, interestingly, since the emergence of ChatGPT, we as a society have been discussing AI very deeply. It is quite different from the AI you have been working on. As an observer, your AI has always been very specific, and it's a bit strange to watch, you know, it starts doing some meaningless things. It is very good at computer games.

Demis Hassabis： 我不会说它们毫无意义，但它们更多的是为了好玩，也许你可以这么说。我们从游戏入手，部分原因是我的游戏背景以及认真下棋等等。但我可以看到，游戏与人工智能一直有着悠久的历史。从图灵和香农在人工智能领域的早期开始，所有这些伟大的，他们都是从象棋程序开始的。几乎每个AI先驱都这样做过。而且，它一直是我们的试验场。你能用你的算法思想快速取得进展吗？然后很容易衡量你的水平，如果你能击败世界冠军或最好的计算机，那么你就知道你做得很好。但关键是，它们始终是达到目的的手段，而不是目的本身。所以我们的想法是:
Demis Hassabis: I wouldn't say they are meaningless, but they are more for fun, maybe you could say that. We started with games partly because of my background in gaming and playing chess seriously, etc. But I can see that games have always had a long history with AI. From the early days of Turing and Shannon in the field of AI, all these greats, they all started with chess programs. Almost every AI pioneer has done so. And it has always been our testing ground. Can you make rapid progress with your algorithmic ideas? Then it's easy to measure your level, and if you can beat the world champion or the best computer, then you know you're doing well. But the key is, they have always been a means to an end, not the end itself. So our idea is:

不要仅仅为了击败围棋或国际象棋的冠军，而是要以一种能够推广到其他领域的方式来做到这一点，包括科学和商业应用。这就是我们用深度强化学习和AlphaGo所做的，所有这些都是非常通用的系统，我们至今仍在使用。
Instead of just aiming to defeat champions of Go or chess, it should be done in a way that can be generalized to other fields, including scientific and commercial applications. This is what we did with deep reinforcement learning and AlphaGo, all of which are very general systems that we still use today.

现在，当你谈到像AlphaFold或我们的科学程序，它们解决了蛋白质折叠等问题时，你真正感兴趣的是解决方案本身。如果你找到了治疗癌症的方法，你不会在乎它是如何做到的。你只想要治疗癌症的方法。所以你真的想全力以赴。所以你首先要做的就是把你所有的通用技术作为基线。然后你再看领域本身，如果这个领域对社会或商业足够有价值，那么你就在上面添加定制的东西。这就是你如何得到像AlphaFold这样的突破性程序。但最终，DeepMind的目标，从我们创立之初到现在，仍然是实现通用人工智能，这意味着一个通用的系统，它能够开箱即用地完成任何你能完成的认知任务。完全通用，就像阿兰·图灵在50年代所定义的那样，能够计算任何可计算的东西。这是人工智能作为一个领域的最初目标，也是DeepMind的目标。
Now, when you talk about programs like AlphaFold or our scientific programs that solve problems like protein folding, what you're really interested in is the solution itself. If you find a way to cure cancer, you don't care how it was done. You just want the method to cure cancer. So you really want to go all out. So the first thing you need to do is use all your general technologies as a baseline. Then you look at the field itself, and if this field is valuable enough to society or business, you add customized elements on top of it. This is how you get breakthrough programs like AlphaFold. But ultimately, DeepMind's goal, from our inception to now, remains to achieve general artificial intelligence, which means a general system capable of performing any cognitive task you can do right out of the box. Fully general, as defined by Alan Turing in the 1950s, capable of computing anything computable. This was the original goal of AI as a field and is also DeepMind's goal.

当然，你最近看到的是像这些语言模型之类的东西。实际上是ChatGPT进入了大众市场，进入了公众的视野。但实际上，所有顶级实验室，包括谷歌和DeepMind，都在研究语言模型。我们有自己的内部模型，叫做Chinchilla，谷歌也有他们的模型。当然，它们都是基于Transformer架构的，这是谷歌研究院发明的，所有当前的模型都是基于它的。所以这是一个激动人心的时刻，因为语言显然是一种通用能力。这就是为什么每个人都对聊天机器人感到非常兴奋的原因。而且非常有趣，而且有点出乎意料的是，这项技术能够扩展到如此程度。我认为我们比以往任何时候都更接近构建这些类型的通用系统。但目前你仍然需要专门的系统来在特定领域做到最高水平。
Of course, what you've seen recently are things like these language models. In fact, ChatGPT has entered the mass market and the public eye. But actually, all the top labs, including Google and DeepMind, are researching language models. We have our own internal model called Chinchilla, and Google has their models too. Of course, they are all based on the Transformer architecture, which was invented by Google Research, and all current models are based on it. So it's an exciting time because language is clearly a universal capability. That's why everyone is so excited about chatbots. And it's very interesting, and somewhat unexpected, that this technology can scale to such an extent. I think we are closer than ever to building these types of general systems. But currently, you still need specialized systems to achieve the highest level in specific fields.

主持人： 大型语言模型更接近AGI吗？我的意思是，它感觉更像是在与人互动，而这感觉就像AGI。但它真的是吗？
Host: Are large language models closer to AGI? I mean, it feels more like interacting with a human, and that feels like AGI. But is it really?

Demis Hassabis： 我认为，多模态，现在甚至不应该说是大型语言模型，因为它们不仅仅是大型语言模型。它们也是多模态的。例如，我们的Gemini模型从一开始就是多模态的。它们可以处理任何输入。视觉、音频、视频、代码，所有这些东西，以及文本。所以我认为我的观点是，这将成为AGI系统的一个关键组成部分，但可能仅凭它本身还不够。
Demis Hassabis: I think, multimodal, now we shouldn't even say large language models, because they are not just large language models. They are also multimodal. For example, our Gemini model is multimodal from the start. They can handle any input. Visual, audio, video, code, all these things, as well as text. So I think my point is that this will become a key component of AGI systems, but it may not be enough on its own.

我认为从现在到我们实现AGI，还需要两到三个重大创新。这就是为什么我给出的时间表是超过10年的原因。其他人，我的一些同事和同行，以及我们的一些竞争对手，他们的时间表比这要短得多。但我认为10年左右是比较合适的。
I believe that from now until we achieve AGI, we need two to three major innovations. That's why I give a timeline of more than 10 years. Others, some of my colleagues and peers, as well as some of our competitors, have much shorter timelines. But I think around 10 years is more appropriate.

主持人： 这与DeepMind内部的，我猜是内部的紧张关系是如何协调的呢？因为我感觉，尤其是在早期，你们就像世界上资金最雄厚的大学实验室之一。就像贝尔实验室之类，一个伟大的商业研究机构。但现在你们正在做一些非常有用的事情，你们有一系列的，我的意思是，你提到了蛋白质折叠，但你们还有天气预报。你们刚刚在国际数学奥林匹克竞赛中获得了一枚奖牌。对不起，我相信如果你自己去参加的话，你也能获得金牌，但你们的系统获得了一枚银牌。你们正在做所有这些其他的事情。你们在背后也在做吗？你们有团队在思考吗？但是现在我们需要继续前进，制造AGI？
Host: How is this coordinated with the internal, I guess internal tensions at DeepMind? Because I feel, especially in the early days, you were like one of the best-funded university labs in the world. Like Bell Labs, a great commercial research institution. But now you're doing some very useful things, you have a range of, I mean, you mentioned protein folding, but you also have weather forecasting. You just won a medal at the International Mathematical Olympiad. Sorry, I believe if you participated yourself, you could win a gold medal, but your system won a silver medal. You're doing all these other things. Are you doing it behind the scenes as well? Do you have teams thinking about it? But now we need to keep moving forward and create AGI?

Demis Hassabis： 是的，我们有一个很大的组织。正如你所说，我们最初的DeepMind模式有点像贝尔实验室，它是世界上最好的工业实验室之一，能够发明未来，能够长远思考。我们真正展示了这种模式能够做什么。我认为它在为我们今天看到的各种技术奠定基础方面非常有效。所以我认为任何类型的深度技术初创公司都需要时间来发展其成熟的技术。
Demis Hassabis: Yes, we have a large organization. As you said, our initial DeepMind model was a bit like Bell Labs, one of the best industrial labs in the world, capable of inventing the future and thinking long-term. We've really shown what this model can do. I think it has been very effective in laying the foundation for the various technologies we see today. So I think any type of deep tech startup needs time to develop its mature technology.

在过去的两三年里，我们已经到了一个非常激动人心的时刻，这项技术已经相当成熟。它已经准备好应用于各种领域。很明显，有科学、数学、医学，以及所有这些领域的进步。可以算是应用科学。如果你愿意，但也有生产力和商业应用，比如聊天机器人或重新构想工作流程和电子邮件等等，这还处于萌芽阶段，以及帮助编码等等。我们显然也致力于所有这些工作，我们是谷歌的引擎。谷歌拥有令人难以置信的，我想是150亿用户，以及平台和产品，而人工智能是所有这些的核心，新的功能不断涌现，来自我们DeepMind开发的一些技术。
In the past two or three years, we've reached a very exciting moment where this technology has become quite mature. It's ready to be applied in various fields. Obviously, there are advancements in science, mathematics, medicine, and all these fields. It can be considered applied science, if you will, but there are also productivity and commercial applications, like chatbots or reimagining workflows and emails, etc., which are still in their infancy, as well as helping with coding, etc. We are obviously committed to all this work, and we are the engine of Google. Google has an incredible, I think, 15 billion users, and platforms and products, with AI at the core of all this, with new features constantly emerging from some of the technologies developed by DeepMind.

这在某种程度上是很好的，因为产品所需的技术类型实际上与你无论如何都会进行的AGI研究类型有90%的相似之处。这些东西已经有很多融合了，而在五年前或十年前，如果你想将人工智能融入产品中，你必须这样做，因为通用系统和学习系统还不够好，你必须回到逻辑网络，专家系统。像早期那一代的助手，例如，都仍然建立在那种旧的技术之上，这就是为什么它们很脆弱，它们不能泛化，最终它们也没那么有用，而建立在这些学习系统上的新一代助手将更加强大。而且，实际上，非常令人兴奋，我实际上认为像Gemini以及我们自己对未来多模式助手的设想，目前称为Astra，它们是通往AGI系统的关键路径，因为它们实际上推动了朝着这个方向的研究。所以，这里有一段Astra工作的视频。
This is good in a way because the type of technology needed for products is actually 90% similar to the type of AGI research you would be doing anyway. These things have already converged a lot, whereas five or ten years ago, if you wanted to integrate AI into products, you had to do so because general systems and learning systems weren't good enough, and you had to go back to logic networks, expert systems. Like the early generation of assistants, for example, were still built on that old technology, which is why they were fragile, they couldn't generalize, and ultimately they weren't that useful, whereas the new generation of assistants built on these learning systems will be much more powerful. And, in fact, very exciting, I actually think that things like Gemini and our own vision for future multimodal assistants, currently called Astra, are the key path to AGI systems because they actually drive research in this direction. So, here's a video of some Astra work.

以下视频来源于 The following video is sourced from

谷歌黑板报 Google Blog

视频来源：“谷歌黑板报” Video source: "Google Blog"

Demis Hassabis： 这只是一个能够在日常生活中帮助你的通用助手的开始。我听到要把它做成员工。也会有不同的形式。你可以在手机上看到它，你可以在眼镜上看到它。我无法形容这会有多么神奇。如果我们回到五年前，你告诉我我们会走到今天这一步，你只需用相机指向某个东西，它就能完全理解空间环境。这真是太不可思议了。就好像它有概念，它理解什么是物体，甚至能认出我们所在的街区。仅仅是通过窗户看到的周围景色。像记住你把东西放在哪里的记忆，这可能非常有用，就像一个助手一样。个性化和所有这些东西都在这个我称之为下一代助手的产品中出现。我称之为通用助手，因为我想象你把它随身携带在不同的设备上。它是同一个助手，无论它是和你玩游戏，还是在你的桌面上帮助你工作，或者是在移动设备上陪你旅行。
Demis Hassabis: This is just the beginning of a general assistant that can help you in everyday life. I hear it will be made into an employee. It will also come in different forms. You can see it on your phone, you can see it on your glasses. I can't describe how amazing this will be. If we go back five years and you told me we would reach this point today, where you just point a camera at something and it fully understands the spatial environment, it would be incredible. It's as if it has concepts, it understands what objects are, and can even recognize the neighborhood we're in. Just by seeing the surrounding scenery through a window. Memories like remembering where you put things could be very useful, just like an assistant. Personalization and all these things appear in what I call the next-generation assistant. I call it a general assistant because I imagine you carry it with you on different devices. It's the same assistant, whether it's playing games with you, helping you work on your desktop, or accompanying you on your mobile device.

主持人： 是这样吗？我认为我理解对了。有些人会认为这是通往通用智能的一步。如果没有我们还没有的秘密武器，那么它本质上就是这个目标。它与我们目前使用的方法之间没有无法弥合的差距。它只是你达到了70%，你达到了80%，你达到了90%吗？还是有什么其他的东西需要解决？
Host: Is that so? I think I understand correctly. Some people would consider this a step towards general intelligence. If there isn't a secret weapon we don't have yet, then essentially this is the goal. There's no unbridgeable gap between it and the methods we currently use. It's just a matter of whether you've reached 70%, 80%, 90%, or is there something else that needs to be addressed?

Demis Hassabis： 嗯，我们肯定需要这些系统，我相信你们所有人都已经体验过各种最先进的聊天机器人了。这些系统非常被动，它们是问答系统。它们对于回答问题，也许做一些研究，总结一些文本之类的事情很有用。
Demis Hassabis: Hmm, we definitely need these systems, and I'm sure all of you have already experienced various state-of-the-art chatbots. These systems are very passive; they are question-and-answer systems. They are useful for answering questions, maybe doing some research, summarizing texts, and things like that.

我们接下来想要的是更多基于代理的系统，能够完成你给它的特定目标或任务。这当然是我使用助手，数字助手需要做的。你计划一个假期，你在城市里旅行，你告诉它帮你订票。所以它们需要能够在世界上行动，执行动作并进行规划。所以我们需要规划、推理、行动，我们需要更好的记忆，我们需要个性化，所以它需要了解你的喜好，记住你告诉它的内容和你喜欢的东西。所以所有这些技术都是需要的。
What we want next are more agent-based systems that can accomplish specific goals or tasks you give them. This is certainly what I use assistants, digital assistants, for. You plan a vacation, you travel in a city, you tell it to help you book tickets. So they need to be able to act in the world, execute actions, and do planning. So we need planning, reasoning, action, we need better memory, we need personalization, so it needs to understand your preferences, remember what you tell it and what you like. So all these technologies are needed.

现在，我给出的简略说法是: Now, the brief statement I give is:

我们的一些游戏程序，比如击败了围棋世界冠军的AlphaGo，它有规划和推理能力，尽管是在这个狭窄的游戏领域。我们必须引入这些技术，并将它们应用于像Gemini这样的多模态模型，它们基本上是世界模型，正如你刚才看到的，它理解它周围的世界。但是如何在杂乱的现实世界中进行规划，而不是像游戏这样的干净的环境呢？所以我认为这是下一个需要取得的重大突破。
Some of our game programs, like AlphaGo, which defeated the world champion in Go, have planning and reasoning capabilities, although in this narrow game domain. We have to introduce these technologies and apply them to multimodal models like Gemini, which are essentially world models, as you just saw, it understands the world around it. But how to plan in the messy real world, rather than in a clean environment like a game? So I think this is the next major breakthrough that needs to be achieved.

主持人： 那这个系统也能应用AlphaGo级别的游戏和拉动某种方法吗？
Host: So can this system also apply AlphaGo-level games and pull some kind of method?

Demis Hassabis： 是的，没错。有两种方式可以实现，这是我们内部以及学术界目前正在进行的非常有趣的辩论。你可以想象，你希望你的通用代理系统能够做的一件重要的事情就是使用工具。这些工具可以是硬件，比如机器人或物理世界中的东西，但它们当然也可以是其他软件，比如计算器，诸如此类。但它们也可以是其他人工智能系统。所以你可以拥有，你可以想象一个像大脑一样的通用人工智能系统，然后调用像AlphaFold或AlphaGo这样的东西来下围棋或折叠蛋白质，或者因为它是全数字化的，你可以想象将这种能力折叠到通用大脑中，折叠到Gemini中。但这需要权衡取舍，因为这样你是否会用专门的信息超载它，比如太多的棋局，然后这会让它在语言方面变得更差。
Demis Hassabis: Yes, that's right. There are two ways to achieve this, and this is a very interesting debate currently happening internally and in academia. You can imagine that one important thing you want your general agent system to be able to do is use tools. These tools can be hardware, like robots or things in the physical world, but they can also certainly be other software, like calculators, and so on. But they can also be other AI systems. So you can have, you can imagine a general AI system like a brain, and then call upon something like AlphaFold or AlphaGo to play Go or fold proteins, or because it is fully digital, you can imagine folding this capability into the general brain, folding it into Gemini. But this requires trade-offs, because would you overload it with specialized information, like too many chess games, and then this would make it worse in terms of language.

这是一个开放的研究问题，你是想把它分离成一个工具，即使是一个通用人工智能可以在特定情况下使用的AI工具，还是你想把它上游到主系统中。有些东西你想上游到主系统中，比如编码和数学，因为事实证明，如果你把它放在主系统中，它实际上会让一切变得更好。所以有点像你在学习理论和儿童发展理论等等，实际上是为了思考哪些东西可能是通用的，并且最好放在主系统中而不是外围工具中。
This is an open research question, whether you want to separate it into a tool, even an AI tool that a general AI can use in specific situations, or whether you want to upstream it into the main system. Some things you want to upstream into the main system, like coding and mathematics, because it turns out that if you put it in the main system, it actually makes everything better. So it's a bit like learning theory and child development theory, and so on, actually thinking about which things might be general and are best placed in the main system rather than as peripheral tools.

主持人： 你现在组织还有多少比例是一个科学组织？你还有多少比例在努力成为贝尔实验室？
Host: What proportion of your organization is still a scientific organization? How much are you still striving to be like Bell Labs?

Demis Hassabis： 我们永远都会是一个以研究为主导的组织。这就是我们现在在Google DeepMind所做的。但我们越来越多地拥有一个越来越大的产品应用团队，与谷歌的其他部门进行互动。但我们仍然试图让我们的基础研究稍微不受其影响，这样它就可以根据我们自己的研究路线图进行更长远的、更具蓝天意义的思考，而不仅仅是由产品路线图所主导。
Demis Hassabis: We will always be a research-led organization. That's what we are doing now at Google DeepMind. But we increasingly have a growing product application team interacting with other parts of Google. But we still try to keep our fundamental research slightly insulated from that, so it can think more long-term and blue-sky according to our own research roadmap, rather than just being led by the product roadmap.

主持人： 您个人是如何跟上这一切的呢？
Host: How do you personally keep up with all of this?

Demis Hassabis： 嗯，我努力，我的意思是，我曾经在18个月前说过，我会保留我的晚上时间，而且我是一个很自律的人。所以我把午夜到凌晨3点的时间留给思考、阅读论文和提出想法，我仍然在伦敦。但现在我在加州有很多团队。所以很多黄金思考时间都被与美国的电话会议占据了。想想如何腾出这些时间。
Demis Hassabis: Hmm, I try, I mean, I said 18 months ago that I would keep my evening time, and I am a very disciplined person. So I reserve the time from midnight to 3 a.m. for thinking, reading papers, and coming up with ideas, and I am still in London. But now I have many teams in California. So a lot of my prime thinking time is taken up by conference calls with the U.S. Thinking about how to free up this time.

主持人： 我们可以把计时器放在这里吗，如果可以的话。否则，我们会让你错过你的饮料，而且我不确定我们还有多少时间。未来会怎样？我认为你是其中的一员，你是签署了其中一封公开信的人之一，警告说，你知道，真正的生存风险，但这并没有特别明确的定义。你对希望、炒作和末日持什么立场？
Host: Can we put a timer here, if possible? Otherwise, we'll let you miss your drink, and I'm not sure how much time we have left. What does the future hold? I think you are one of them, you are one of the people who signed one of the open letters, warning that, you know, there is a real existential risk, but it is not particularly well-defined. What is your stance on hope, hype, and doomsday?

Demis Hassabis： 我认为这个等式的两边都有很多疯狂的炒作。有一种现在被称为末日阵营的人，他们认为肯定会出错。然后还有那些波莉安娜阵营的人，他们认为这只是另一种技术,以前在移动互联网时代就见过这种情况，它会像那样发展壮大，但是我们作为一个社会和作为人类适应性很强，没什么特别的。
Demis Hassabis: I think there is a lot of crazy hype on both sides of this equation. There is a group now called the doomsday camp, who think things will definitely go wrong. Then there are those in the Pollyanna camp, who think this is just another technology, seen before in the mobile internet era, and it will grow and develop like that, but we as a society and as humans are very adaptable, nothing special.

显然这是错的。我认为，这比互联网或移动设备之类的东西要重要得多。我认为这是一个划时代的定义。我一直这样认为，我想对更多人来说，这一点正变得越来越清晰，但我从我还是个孩子的时候就一直这样认为，这就是为什么我毕生致力于此的原因。
Obviously, this is wrong. I think this is much more important than things like the internet or mobile devices. I think this is an era-defining moment. I've always thought so, and I think it's becoming clearer to more people, but I've thought this way since I was a kid, which is why I've dedicated my life to it.

我认为它可以做到。它将产生难以置信的影响。当然，我做这一切的原因是因为我认为人工智能对世界将产生难以置信的积极影响。我认为我们距离用人工智能治愈所有疾病的目标已经不远了，通过材料科学和新能源以及我认为人工智能可以发明的其他东西来帮助应对气候变化，以及在我们的日常生活中，只是提高生产力，丰富我们的日常生活，平凡的管理工作。我认为，平凡的管理工作可以自动处理。我认为这些都很棒，而且很快就会实现。
I think it can be done. It will have an incredible impact. Of course, the reason I do all this is because I believe AI will have an incredibly positive impact on the world. I think we are not far from the goal of curing all diseases with AI, helping to tackle climate change through materials science and new energy, and other things I think AI can invent, and in our daily lives, just improving productivity, enriching our daily lives, mundane administrative tasks. I think mundane administrative tasks can be handled automatically. I think these are all great and will be realized soon.

但是这些系统存在风险，它们是新技术的新系统，它们非常强大。
However, these systems carry risks. They are new systems with new technologies, and they are very powerful.

我已经在游戏的微观世界中看到了这一点，我理解得很清楚，比如下棋，你从一个随机的系统AlphaZero开始，到了早上喝咖啡休息的时候，它可以打败我，然后到了午餐时间，它已经比世界冠军更厉害了，然后到了下午，在大约8个小时内，它就比最好的国际象棋，高端的、硬编码的国际象棋计算机更厉害了。
I have already seen this in the microcosm of games, and I understand it very clearly. For example, in chess, you start with a random system like AlphaZero, and by the time you take a coffee break in the morning, it can beat me. Then by lunchtime, it is better than the world champion, and by the afternoon, in about 8 hours, it is better than the best chess, high-end, hard-coded chess computers.

所以它是世界上有史以来最伟大的国际象棋实体，在8个小时内从随机变成了这样。我实际上观察了这个过程超过8个小时，这真是太不可思议了，当然，那只是一个游戏，而且范围很窄。
So it became the greatest chess entity ever in the world, transforming from random to this in 8 hours. I actually observed this process for over 8 hours, and it was truly incredible. Of course, that was just a game, and the scope was very narrow.

但我认为没有理由认为这种能力不能推广到这些更通用的语言和世界模型系统中。所以它将非常强大，但必须谨慎处理。而且我认为我们根本不知道。所以我签署那封信的原因是我只是想给那些认为这里没什么可看的，实际上有一些未知风险的波莉安娜主义者一些压力。
But I see no reason to think that this capability cannot be extended to these more general language and world model systems. So it will be very powerful, but it must be handled with care. And I think we simply don't know. So the reason I signed that letter is that I just wanted to put some pressure on those Pollyannas who think there's nothing to see here, when in fact there are some unknown risks.

我们需要定义，我认为我们有时间，但十年时间对于即将到来的如此重大的事情来说并不长。所以我们需要对可控性等方面进行更多研究，从理论层面理解这些系统的作用。你知道，非常重要的事情，比如我们如何为系统定义目标和价值观，以及我们如何确保它们坚持这些目标和价值观。这些都是当前新兴技术的未知数。所以我想说我是一个谨慎的乐观主义者。所以我认为如果我们齐心协力，我们就能解决这个问题。我们在国际范围内这样做，让所有最优秀的人才都参与进来。我们现在就开始行动。我很高兴看到在英国和美国成立的人工智能安全研究所，我们将倡导这种情况的发生，并测试最新的模型。但我们需要更多这样的机构。我只是在鼓励这种情况的发生。而且我认为，如果有足够的时间，有足够的脑力,人类的聪明才智,我们会做好的。但风险是存在的，我们不能抄近路，我们需要认真对待它。我认为应该怀着敬畏之心。我认为这项技术值得我们去努力。
We need to define, and I think we have time, but ten years is not long for something so significant that is coming. So we need to do more research on controllability and other aspects, to understand what these systems do from a theoretical level. You know, very important things like how we define goals and values for the systems, and how we ensure they adhere to these goals and values. These are the unknowns of current emerging technologies. So I would say I am a cautious optimist. So I think if we work together, we can solve this problem. We do this internationally, involving all the best talents. We start acting now. I am pleased to see AI safety institutes being established in the UK and the US, and we will advocate for this situation to happen and test the latest models. But we need more such institutions. I am just encouraging this to happen. And I think, given enough time, enough brainpower, human ingenuity, we will do well. But the risks are there, and we cannot take shortcuts, we need to take it seriously. I think it should be approached with awe. I think this technology is worth our effort.

主持人： 你所说的有点吓人。我的意思是，如果你从国际象棋中发生的事情进行概括，国际象棋还好。但假设我每天去办公室工作，靠下棋谋生，不是靠打败其他人，而是因为它对下棋有一些功利价值。你所说的系统可能会消除几乎所有的人类价值。
Host: What you said is a bit scary. I mean, if you generalize from what happens in chess, chess is fine. But suppose I go to the office every day to work, making a living by playing chess, not by defeating others, but because it has some utilitarian value in playing chess. The system you mentioned might eliminate almost all human value.

Demis Hassabis： 我认为即将出现一些重要的哲学讨论。它们很快就会出现。我们如何分配？我们应该生活在一个AGI运作的零和世界中。所以激进的富足。像能源之类的东西不应该短缺，因此资源和其他东西也不应该短缺。我认为这确实改变了经济的动态，我指的是长远来看。所以我们现在需要开始思考这个问题，为它做好准备。我们希望如何分配这些额外的富足和财富。无论是某种普遍基本供应还是其他类似的东西。我们需要现在就开始思考这个问题，经济学家和像他们这样的人。我觉得他们需要现在就开始研究这个问题。
Demis Hassabis: I think some important philosophical discussions are about to emerge. They will appear soon. How do we allocate? Should we live in a zero-sum world operated by AGI? So radical abundance. Things like energy shouldn't be scarce, and therefore resources and other things shouldn't be scarce either. I think this really changes the dynamics of the economy, I mean in the long run. So we need to start thinking about this now and prepare for it. How do we want to allocate this extra abundance and wealth? Whether it's some kind of universal basic provision or something similar. We need to start thinking about this now, economists and people like them. I feel they need to start studying this issue now.

主持人： 最后一个问题英国。我们在英国，美国正走向人工智能的中心。我们正在讨论英国是否需要更多计算能力，政府是否需要支持它。这是必需的吗？我们有哪些可能会落后的方面？
Host: The last question is about the UK. We in the UK and the US are moving towards the center of artificial intelligence. We are discussing whether the UK needs more computing power and whether the government needs to support it. Is this necessary? In what areas might we fall behind?

Demis Hassabis： 我认为这是一个巨大的增长领域，我认为英国政府应该鼓励它。我想说的是，更重要的是鼓励国内投资，地方投资，规划许可等等。我认为这是一个建立新世界的绝佳机会。我认为这是一个巨大的机遇。许多大公司都认为英国是一个吸引人的研究和开展业务的地方。我们在这里有一个很棒的生态系统。我们拥有一流的大学。这就是我们DeepMind在这里的原因之一。我们在这里有助于建立一个由英国优秀初创公司组成的生态系统，这些公司基于人工智能或与人工智能相邻。所以我认为政府只需要释放这种潜力，让公司很容易在这里投资和建设，包括大型数据中心。
Demis Hassabis: I think this is a huge growth area, and I think the UK government should encourage it. What I want to say is that it is more important to encourage domestic investment, local investment, planning permissions, and so on. I think this is a great opportunity to build a new world. I think this is a huge opportunity. Many big companies see the UK as an attractive place for research and business. We have a great ecosystem here. We have world-class universities. This is one of the reasons why DeepMind is here. We are here to help build an ecosystem of excellent UK startups, companies based on or adjacent to artificial intelligence. So I think the government just needs to unleash this potential, making it easy for companies to invest and build here, including large data centers.

⭐星标AI寒武纪，好内容不错过⭐
⭐Star AI Cambrian, don't miss good content⭐

个人公众号：AI寒武纪 Personal WeChat Official Account: AI Cambrian

亲爱的读者们，不星标《知识分子》公众号，会错过每日科学新知！星标《知识分子》，紧跟前沿科学，一起探索科学的奥秘吧！
Dear readers, if you don't star the "Intellectuals" WeChat Official Account, you will miss out on daily scientific knowledge! Star the "Intellectuals" to keep up with cutting-edge science and explore the mysteries of science together!

请戳上图卡片添加星标 Please click the card above to add a star

关注《知识分子》视频号
Follow the "Intellectuals" video account

get更多有趣、有料的科普内容 Get more interesting and informative popular science content

END