Demis Hassabis：从单一梦想走向集体行动

我觉得《无限机器》和你讲述的 DeepMind 与德米斯·哈萨比斯的故事，吸引了很多 AI 圈的人。我知道我读这本书时，大概 24 小时就看完了，故事非常引人入胜，而且你对德米斯的接触机会非常难得。我想，你们在酒吧里一坐就是好几个小时，聊各种各样的事情，这真是一个绝佳的机会，能深入了解当今 AI 竞赛中最有趣的人物和公司之一。感谢你写了这本书，也感谢你来播客聊聊。我们今天的目标是利用你的报道，来理解塑造我们正在经历的 AI 竞赛的人、决策和动态。所以，也许从这一点开始：我觉得你书中一个有趣的主题是，如今实验室之间的这场竞赛是否不可避免，或者如果不同的参与者或决策不同，结果会不会不一样。我很好奇，做完这些工作后，你认为有没有另一条路，还是说这基本上是不可避免的？

I feel like The Infinity Machine and your story of DeepMind and Demis Hassabis has captivated a bunch of folks in the AI world. I know that I inhaled this book. I think I finished it in like 24 hours. It's just a gripping story and you got this incredible access to Demis. I think you know, sitting in this pub for multiple hours at a time, talking about all sorts of things, just a fascinating opportunity and window into definitely one of the most interesting people and companies in this AI race today. So appreciate you writing the book and coming on the podcast to talk about it. Our goal today is to use your reporting to understand the people, decisions, and dynamics that shape the AI race we're living through. So maybe to start, I feel like one of the interesting themes in your book is whether this race we have today between the labs was inevitable or could have gone differently if different players or decisions had been made. I'm curious after doing this work, do you think there was another path or was this kind of inevitable?

嗯，我认为这是不可避免的。当你有这样一种极其强大的技术时，多个国家的多个实验室都会拼命去尝试构建它。我们知道中国的技术栈相当强大。所以尽管缺乏半导体，他们还是会去尝试，而且实际上做得相当不错。然后显然在美国，还有其他几个国家，比如法国的 Mistral、加拿大的 Cohere。肯定会有很多参与者，因为这项技术太诱人了，不可能只有一个团队感兴趣。但整个辩论中奇怪的是，当时人们并不这么看。你知道，当德米斯创办 DeepMind 时，他真的希望避免竞赛动态。回想起来这似乎很天真，但他当时确实这么希望。

You know, I think it was inevitable. I think when you have this sort of supremely strong technology, there's going to be multiple labs in multiple countries that are just desperate to try and build it. And we know the China stack is pretty strong. And so despite the lack of semiconductors, they were going to have a go at it and they are actually doing pretty well. And then clearly in the US, and then in a few other countries, you've got Mistral in France, Cohere in Canada. There was bound to be many players, just as the technology is too sweet for it to be only interesting to one team. What's strange about this whole debate though is that's not how people saw it back then. You know, when Demis was starting DeepMind, he really hoped that he could avoid the race dynamic. It seemed naive in retrospect, but that's what he hoped.

你书里有一个非常引人注目的轶事：我相信德米斯在面试时会说，'看，某个时候我们可能非常接近 AGI，我们会飞到某个地堡，像一个团队一样解决所有问题。你愿意登上那架飞机去做吗？' 我认为 Anthropic 的创立初衷也部分源于这种信念：嘿，我们需要成为那个在 AGI 悬崖边上解决问题团队。这显然是创办所有这些公司的关键部分。你觉得这些人，尤其是德米斯，他们还相信这一点吗？还是说他们的想法随着过去几年事态的发展而改变了？

One really compelling anecdote you had was, I believe Demis in his interviews would say, you know, he was interviewing candidates and he'd say, 'Look, at some point we may be really close to AGI and we'll go fly to a bunker and figure this all out as kind of like one team. Would you be willing to get on that flight and do that?' I think certainly part of the founding origin of Anthropic is also this belief like, hey, we need to be the team that, as we get to the precipice of AGI, we're the ones figuring this out. Clearly a key part of starting all these companies. Do you think these folks, I mean particularly Demis, do they still believe this or how has their thinking evolved having seen the way things have played out these past years?

不，我认为德米斯从一个极端摆到了另一个极端。他一开始认为可能存在单一实验室的场景，而且他指的就是 DeepMind 自己。现在他走到了相反的极端，认为这个领域非常拥挤，因此一个实验室单独追求安全几乎毫无意义，因为如果一个实验室安全而其他实验室不安全，这并不会让世界更安全。所以他真的转变了看法，认为这是一个只有政府才能解决的集体行动问题。

No, I think Demis has swung from one extreme to the other. You know, he began by thinking there could be a singleton scenario, just one lab, and by that he really meant DeepMind himself. To the opposite extreme where he now sees that there's a very crowded field and therefore it's almost pointless for one lab to pursue safety by itself because if one lab is safe and the other ones aren't, it doesn't make the world safer. So he really has shifted to seeing this as a collective action problem that only a government can solve.

我觉得你书中非常引人深思的一点是，显然早期有一次 AI 安全峰会，德米斯和团队组织起来，分享了他们取得的进展以及为什么对此感到担忧。而房间里的一些人，比如里德·霍夫曼、埃隆·马斯克，得知这些信息后，他们就想，'哦，这项技术真的开始奏效了。我们自己可能也应该在这方面做点什么。' 你认为，显然要让安全在未来发挥作用，需要这些不同的公司同意分享和协作。我想这给德米斯留下了不好的印象。你觉得他和其他关键参与者，鉴于过去的经历，会如何看待分享他们正在做的事情？

One thing I thought was really compelling from your book is obviously there was this early AI safety summit that Demis and the team put together and they're sharing the progress they've made and why they're worried about it. And some people in the room like Reid Hoffman, Elon Musk take that information and then they're like, 'Oh this technology really is starting to work. We should ourselves potentially do something around that.' Do you think, obviously to get safety to work going forward you need these different companies to agree to share things and collaborate. I imagine that left a bad taste in Demis's mouth. How do you think he and other key players in this space think about potentially sharing what they're doing, given that past experience?

是的，你说得完全对。那是在 2015 年。2015 年夏天，他们在 SpaceX 举行了这次峰会。埃隆·马斯克是主办方。想法是 DeepMind 会把他拉进自己的阵营，让他参与他们的努力，让他主持这个安全监督委员会。这样他就不会建立竞争对手。当然，2015 年底，他确实建立了一个竞争对手。

Yeah, I mean, you're totally right. That was in 2015. Summer of 2015, they have this summit at SpaceX. Elon Musk is hosting it. And the idea is he's going to be brought into the tent by DeepMind. He's going to be part of their efforts. He's going to be chairing this sort of safety oversight board. And therefore, he wouldn't set up a competitor. And of course, at the end of 2015, he did set up a competitor.

OpenAI 的出现真正让人意识到我们将面临一场竞赛。如果你现在问及未来的合作，从戴密斯或任何实验室领导的角度来看，坦率地说，你无法信任其他人。获得信任的唯一途径是有一个政府执法者站出来说：‘这是给所有人的规则。将会有一个公平的竞争环境。你们都必须遵守某种安全减速，在发布模型前进行预测试，所有这一切，你们都必须做。’然后反应会是：‘那中国那边呢？’所以最终我认为这必须成为中美之间的合作，尽管这个前景对许多听众来说可能很遥远。

Open AI and so that really drove home the reality that we're going to have a race. I think if you ask now about future collaboration, from Demis's point of view or any of the lab leaders, frankly, you can't trust the other guys. The only way you get trust is if you have a government enforcer that comes along and says, 'Here are the rules for everybody. There's going to be a level playing field. You're all going to have to abide by some sort of safety slow-down, pre-testing of models before you release them, all that stuff, and you all have to do it.' Then the reaction will be, 'Yeah, but what about the guys in China?' That's why ultimately I think this has to be a US-China collaboration, remote though that prospect may seem to many listeners right now.

你认为戴密斯和其他人认为这真的现实吗？显然这是一个紧迫的问题，但似乎人们对政府没有太大信心，更不用说全球范围内的政府间合作了。从理性上讲，这是解决这个问题的明确途径，但你认为他们真的认为这有可能成功吗？

Do you think Demis and others think that's actually realistic? Obviously it's such a pressing problem, but it doesn't seem like folks have the most confidence in governments, and certainly inter-government collaboration across the world. Intellectually it makes sense that that is the clear way to solve this problem, but do you think they think there's actually a probability this happens successfully?

我认为他们应该相信这是可能发生的。我们有食品药品监督管理局，人们抱怨它慢，但它确实有专家评审员审查临床试验，判断‘这种药安全吗？可以发布吗？’如果你能对药品这样做，那么对可能造成更大规模破坏的 AI 模型也应该这样做。在英国，有一个政府 AI 安全研究所的例子，相当有效，拥有高素质的科学家。那里的首席科学家是杰弗里·欧文，他最初在 DeepMind，对早期模型做了很多后训练，是一位非常严肃的研究人员，也在 OpenAI 工作过，是达里奥的密切合作者。他们曾发现系统中的漏洞，并悄悄与未能发现相同漏洞的私人实验室分享。这表明你可以在公共机构内部聚集技术专长，因为有些人有足够的动力和公德心去为政府工作。我认为我们不应该放弃这一点。

Well, I think they ought to believe that it could happen. We do have the Food and Drug Administration, and people complain it's slow, but it does have expert reviewers who look at clinical trials and say, 'Is this drug safe or not? Can it be released?' If you can do that for pharmaceuticals, you should do it for AI models that can be more destructive on a grander scale than a drug. In Britain, you have an example of a government AI safety institute that's pretty effective, with high-caliber scientists. The chief scientist there is Jeffrey Irving, who was at DeepMind originally, did a lot of the post-training on the early models, a really serious researcher, also at OpenAI, and a close collaborator of Dario. They have found vulnerabilities in systems and quietly shared them with private labs that failed to find the same vulnerabilities. So it shows you can aggregate technical expertise inside public institutions because some folks are motivated and public-spirited enough to go work for the government. I don't think we should give up on that.

从外部很难判断。这几乎是一个令人沮丧的故事：戴密斯曾经理想主义，然后意识到这种动态不会实现。所以很难说，从你的书中字里行间看，他是已经认命了，还是在政府间合作的实际可能性中找到了新的希望，因为你给出的例子甚至只是在一个国家内部，当然没有达到跨国协调的规模。

From the outside it's hard to tell. It's almost like such a dispiriting story of Demis being idealistic and then realizing this dynamic isn't going to play out. So it's hard to tell, reading between the lines in your book, whether he's resigned to this dynamic or has found newfound hope in the actual possibility of inter-government collaboration, because even the examples you give are just within a single country, and certainly not at the scale of coordination across countries.

是的。他说的话——我书的最后一行是：‘我仍然乐观。’这是戴密斯试图说服我的引述。我在采访中追问：‘你真的乐观吗？因为我们似乎陷入了竞赛动态。’2025 年 1 月，DeepSeek 出现，标志着中国实验室进入这个领域。我写到 2025 年底，问他：‘你还乐观吗？’他说：‘是的，我仍然乐观。’我认为他坚持的原因，除了他还能说什么之外，是当戏剧性的事情发生时，世界会凝聚起来采取行动。看看新冠：各国政府确实做了人们之前没预料到的事情。在危机中，人们会反应。像贸易这样的问题，保护主义对经济表现的影响缓慢而微妙，你不会得到政治反应去修复它。但如果有一个急性事件——比如 Anthropic 的模型出现，人们突然担心——你会看到美国政府从自由放任立场 180 度大转弯，变成‘我们最好控制这个’。这表明冲击效应在决定政府是否采取行动方面极其重要。

Yeah. What he says—the very last line of my book is, 'I'm optimistic still.' That's a quote from Demis trying to persuade me. I was pushing him in my interviews: 'Are you really optimistic? Because it seems like we've got a race dynamic.' In January 2025, DeepSeek came out, signaling the advent of Chinese labs into this space. I'm writing through 2025 into the end of 2025, and I'm saying, 'Are you still optimistic?' And he says, 'Yes, I'm optimistic still.' I think the reason for that insistence, apart from what else he would say, is that when something dramatic happens, the world does coalesce and take action. Look at COVID: governments on a national basis really did stuff people didn't expect beforehand. In a crisis, people react. When you have something like trade, protectionism has a slow and subtle effect on economic performance, and you don't get a political reaction to fix that. But if you have an acute thing—like an Anthropic model comes out and people suddenly get worried—you see the US government do a 180 from a laissez-faire position to 'we better control this.' That shows the shock effect is extremely important in determining whether you get government action.

换个话题，显然你工作的一个关键部分是 DeepMind 的故事以及它最终与谷歌的关系。这是一个不可思议的故事。我记得你是在 ChatGPT 之前开始这个项目，或者说让戴密斯同意这个项目的。显然在 AlphaFold 之后有一些有趣的事情，但从你项目开始以来，世界以许多有趣的方式演变。人们一直在谈论戴密斯和 DeepMind，但你做了研究。你认为流行的叙述对戴密斯和 DeepMind 最大的误解是什么？

Switching gears, obviously a key part of your work is the story of DeepMind and its ultimate relationship with Google. It's an incredible story. I recall you started this, or got Demis to agree to this, right before ChatGPT. Obviously post-AlphaFold there were some interesting things, but the world evolved in a bunch of interesting ways from the start of your project. People talk about Demis and DeepMind all the time, but you've done the research. What do you think the popular narrative got most wrong about Demis and DeepMind?

对我来说，最不寻常的是人们基本上低估了戴密斯。我的书几周前出版后，我接受了很多采访，人们经常念错他的名字。他们说‘Deise’，说‘Hassabis’而不是‘Habis’。他们几乎不知道他是谁。我的出版商企鹅出版社在设计封面时，想用这张照片，但同时他们认为人们不会认出他。所以你不可能靠一个没人认识的人来卖书。

The extraordinary thing to me is how people basically discounted Demis. I've been interviewed a bunch of times since my book came out a few weeks ago, and often people can't pronounce his name. They say 'Deise,' they say 'Hassabis' instead of 'Habis.' They barely know who he is. My publisher, Penguin Press, was figuring out the cover design, and they wanted to use this picture, but at the same time, they didn't think people would recognize him. So you couldn't sell books based on some person nobody recognized.

在这一点上，我猜山姆、达里奥，他们上封面没问题。

At this point, I assume Sam, Dario, they'd be fine for a book cover.

是的，完全同意。埃隆肯定没问题。而戴密斯是在 2010 年创立了最初的实验室，远早于其他人，真正创造了后来 OpenAI 复制的模型。

Yeah, totally. Elon, for sure. And Demis is the guy who founded the OG lab in 2010, way before anybody else, really created the model that OpenAI then copied later.

我记得做研究时去见达里奥，他说没错，德米斯是原创人物，整个 AI for science 领域都是他的，当时确实如此，现在可能没那么绝对了。所以我觉得人们低估了他的重要性，也低估了谷歌这家公司，因为他们觉得创新者困境会让谷歌太慢，OpenAI 先推出了模型和聊天机器人，现在有了巨大的品牌效应，没人能追上。结果证明错了。到 2025 年底，谷歌 DeepMind 的 Gemini 3.0 在排行榜上已经优于对手。当然之后 Anthropic 又大幅追赶。但关键是，我认为人们太早把 OpenAI 和山姆·奥特曼捧为赢家，低估了德米斯本人和谷歌 DeepMind 这家公司。

I remember going to see Dario when I was doing my research and you know he said yes Demis was the original figure and he's got the whole AI for science space to himself which was true at the time I think it's less true now but so I think people just underestimated how important he is they underestimated Google as a company because they thought innovator's dilemma they were too slow OpenAI went first with a model with a chatbot and now they've got this massive brand effect OpenAI does and so nobody can catch up. Well that proved wrong. And as of late 2025 the Google DeepMind models Gemini 3.0 were better on the leaderboards than the adversaries. Of course since then we've seen a big Anthropic surge. But the point is I think people were too quick to crown OpenAI and Sam Altman as the winner and underestimated both Demis as a person and Google DeepMind as a company.

有趣的是，从你的报道来看，到 2026 年，编码智能体崛起，Anthropic 和 Claude Code 率先推出，OpenAI 的 Codex 紧随其后。感觉谷歌一直在做惊人的科学，排行榜上表现也很好，但不知为何在时代精神和产品实际使用上却挣扎。如果两个里程碑事件是 ChatGPT 和消费产品，然后是 Claude Code 和编码智能体，谷歌在这两个领域都没有特别有竞争力的产品——从使用量看。产品本身不错，模型也好，但就是没搞定后面那部分。你怎么看？

It is interesting obviously since your reporting I guess in 2026 you've had these rise of coding agents and Anthropic and Claude Code were the first there and OpenAI with Codex was right behind. It does feel like Google continues to do amazing science and they always do well on the leaderboards but for whatever reason it struggles in the zeitgeist and actual usage on the product side. If these two monumental moments were ChatGPT and consumer products and then Claude Code and coding agents, Google doesn't really have a super competitive product in either of those spaces from a usage perspective. The products themselves are good, the models are good, but they haven't really figured out the latter part. I'm wondering what you make of that.

人们使用 Gemini 比我们意识到的要多，它已经整合到谷歌搜索的 AI 模式里了。但我理解你的大观点：这两个证明消费体验的标志性时刻——消费端的 ChatGPT 和最近的编码——都不是来自谷歌 DeepMind。我认为这或许表明，部分原因是德米斯的个性和知识背景——神经科学博士，对智能本质的广泛研究，因此构建 AI 的方法也非常广泛。有种什么都试试的态度。每当有两条不同路径，他们会说两条都走，如果能找到第三条，可能也会走。他们非常对冲。而 Anthropic 之所以在编码上成功，是因为它愿意下更集中的赌注。它从未涉足生成式视频，没有 Sora 的竞品。OpenAI 作为初创公司没有谷歌那样的声誉包袱，所以愿意推出一个初期经常幻觉的聊天机器人。这其实和 Transformer 本身有相似之处。OpenAI 押注全力扩展 Transformer——这发明于谷歌——而谷歌同时探索多条路径。后来 OpenAI 也开始做很多事，Anthropic 则决定聚焦编码。鉴于这些发展，你看到德米斯和 DeepMind 的焦点有转变吗？作为一个对 AI 多方面都感兴趣的博学者，要同时关注很多事和当前热点之间存在张力。OpenAI famously 收缩了，说我们做得太多，现在要集中精力在编码模型上追赶 Anthropic。你觉得德米斯和 DeepMind 的人怎么想？

People use Gemini more than we realize, it's bundled into Google search with the AI mode. But I take the broader point that both of these seminal moments in proving out the consumer experience — ChatGPT on the consumer side and more recently coding — neither came from Google DeepMind. I think that perhaps shows us that partly because of Demis' personality and his intellectual formation — a PhD in neuroscience, this very broad study of what intelligence might be, a very broad approach to building AI. There's this let's try everything approach. Whenever there are two different paths, they say we'll do both, and if we can find a third path we'll probably do that too. They're very hedged. Whereas I think Anthropic got to coding because it was willing to take a more concentrated bet. It never went into generative video, never had a Sora equivalent. And OpenAI being a startup didn't have the reputational baggage that Google has, so was willing to put out a chatbot that hallucinated a lot at the beginning. That honestly has echoes of the transformer itself. OpenAI made the bet of going all in on scaling transformers, which were invented at Google, while Google pursued many paths. Then OpenAI started doing many things and Anthropic decided to focus on coding. Have you seen a shift in Demis and DeepMind's focus given what's played out? It's such a tension as a polymath interested in many facets of AI to focus on many things versus what's happening at the moment. OpenAI famously consolidated and said we were doing too much, now we've got to hone in and catch up to Anthropic on coding models. How do you think Demis and the DeepMind folks think about this?

我认为他们确实有这种什么都试试的倾向，而且没改。上周我和那家机构的一位年轻科学家聊天，他说的基调就是：我们总是下多个赌注，从不真正全力投入一条路。从公司角度看，这也许有道理。想想消费端的情况，通用聊天机器人模型，他们在 2017 年 Transformer 论文后开发就晚了，产品化又落后于 ChatGPT，但到 2025 年追上了。所以如果你有谷歌那样雄厚的资金、深厚的技术储备、大量人才和惊人算力，你负担得起。他们不像苹果那样。苹果是极端，说我们不做 AI 这玩意儿，就收钱把它装到 iPhone 上。那是极其放手的态度。谷歌参与得更深，但不介意落后几年，因为它觉得自己能追上。我好奇某个时点这是否会改变。显然，预测市场中出现了价值 2 万亿美元以上的公司，占据了地盘。看看这种分散还是专注的方式会持续下去，会很有趣。

I think they do have this tendency to try and do everything. I don't think that's been cured. I was chatting last week with one of the young scientists from that shop and that was very much the tenor: we always make multiple bets, we never really go hard down one avenue. From a corporate point of view, maybe this makes sense. If we think about what happened with the consumer side, the general chatbot model, they were late both in trying to develop it after the transformer thing dropped in 2017 and then late again in productizing behind ChatGPT, but they caught up by 2025. So maybe if you have very deep pockets like Google and a very deep technical bench and tons of human talent and amazing amounts of compute, you can afford it. They're not doing Apple, right? Apple is the extreme where they say we're not going to do this AI thing, we'll just charge people to put it on our iPhone. That's extremely hands-off. Google is much more in it, but it doesn't mind being a couple of years behind because it figures it can catch up. I wonder at some point whether that will shift. Obviously you've seen companies worth $2 trillion plus emerge in prediction markets, gaining territory. It'll be interesting to see whether that focused or diverse approach continues.

你在书中谈了很多 DeepMind 和谷歌的治理，以及这两个组织如何协作。最有趣的部分之一是 DeepMind 考虑从谷歌分拆出来，背后还有一些融资。他们最终选择不拆。你觉得为什么他们最终决定不分拆？回顾来看，这个决定对吗？

You talk a lot in the book about the governance of DeepMind and Google and how these two organizations work together. One of the most interesting parts was DeepMind considering spinning out of Google. There was some financing behind that. They ultimately chose not to. Why don't you think they ended up deciding to spin out? And in retrospect, was that the right decision?

是的。有一个秘密计划叫“马里奥计划”，目的是分拆。当我从其他渠道发现时，德米斯不太高兴。我一度不得不和谷歌 DeepMind 的总法律顾问谈话，他试图说服我不能写这件事。其他人给我泄露了文件，我拿到了，信息确实可靠。

Yeah. So there was this secret plan called Project Mario to spin out and Demis was not so happy when I kind of discovered about it from other sources. I had to speak to the general counsel of Google DeepMind at one point who tried to persuade me I couldn't write about it. Other people leaked me the documents and I had them and definitely had good information.

这都是真的，我写了下来。所以随便吧。这个故事的重点部分回到了你的安全问题上：戴密斯非常希望获得对谷歌 DeepMind 模型的安全监督，或者当时只是 DeepMind。而谷歌总部在 Mountain View 并没有这么做，所以他必须有一个可信的剥离威胁。于是他去找了里德·霍夫曼。里德·霍夫曼承诺提供十亿美元来资助剥离，戴密斯用这个来向谷歌施压。虽然我不认为他提过里德·霍夫曼已经这么做了，但他知道自己有这个后备选项，所以向谷歌施压。我和很多当时在场的顾问谈过，当时 DeepMind 正在决定是否采用里德·霍夫曼的剥离方案。有些人说：‘你应该剥离，因为那样你会成为一家独立创业公司，拥有所有的激励、警觉性、灵活性和敏捷性。你可以给团队提供与业绩挂钩的高强度财务激励，那会很棒。你应该剥离。’而戴密斯最终的观点是：‘这在法律上可能会有一场法庭大战，看我们是否有权这么做。我只想做科学。我不想被法律斗争分心。我想要大量的算力。我留下。’这就是他做的决定。是对是错？嗯，这确实让他在 2020 年获得了诺贝尔奖。在他放弃与母公司斗争一年后，他推出了 AlphaFold，蛋白质折叠预测模型，这为他赢得了诺贝尔奖。

It was all true and I wrote it. So whatever. And the point of this story is partly it goes back to your safety thing: Demis really wanted to get safety oversight over the Google DeepMind models, or just DeepMind as it was then. And Google corporate in Mountain View wasn't doing that, so he had to have a credible threat of spinning out. So he went to Reid Hoffman. Reid Hoffman pledged a billion dollars to finance a spin-out, and Demis used that to kind of pressure Google. Although I don't think he ever mentioned that Reid Hoffman had done that, but he pushed Google knowing that he had this back-pocket option. I spoke to a lot of the advisers who were in the room when DeepMind was figuring out whether to go with a Reid Hoffman spin-out option. There were those who said, 'Look, you should spin out because then you'll be an independent startup and you'll have all the incentives, alertness, flexibility, and agility that go with that. You can give your team very high-powered financial incentives linked to your performance, and it'll be awesome. You should spin out.' And Demis' view in the end was, 'It's legally going to be a court battle over whether we have the right to do that. I just want to do science. I want to not be distracted by legal fights. I want access to tons of compute. I'm staying in.' So that was the call he made. Was it right? Was it wrong? Well, it did lead to him getting a Nobel Prize in 2020. A year after he gave up that fight with his parent company, he ships AlphaFold, the protein folding prediction model, and that gets him the Nobel Prize.

你刚才描述的方式更像是，嘿，这是谈判筹码的一部分，用来争取一些安全方面的东西。我是说，你认为当时 DeepMind 团队有多认真地考虑这个？

The way you just framed it was more like, hey, this was part of negotiating leverage to get some of the safety stuff. I mean, how seriously do you think the DeepMind team was considering this at the time?

嗯，我认为他们自己从未真正到达那个决策点。他们想要 B 计划剥离选项，因为为什么不呢？但他们不确定如何在与谷歌的谈判中使用它。我认为他们决定从不明确告诉谷歌他们有这个选项，而是暗示可能有什么背景，如果你们不按我们要求的做安全监督，我们可能会做一些你们意想不到的事情。但最终，他们从未明确挥舞过那个威胁，也从未执行过，尽管谷歌没有给他们想要的安全监督。所以从某种意义上说，这是一个关于创始人天真的故事，包括戴密斯本人和他的联合创始人穆斯塔法·苏莱曼，他现在在微软。他们进入这个局面时并不确定自己的最终目标，最终也从未真正使用那个剥离选项。所以从某种意义上说，花三年时间在 Mountain View 来回奔波是浪费。

Well, I think they never quite got to that decision point in their own heads. They wanted the plan B spin-out option because why not have that? And they weren't sure how to then use it in the negotiation with Google. I think they determined never to tell Google explicitly that they had that, but to kind of hint that there might be something in the background, and if you don't do what we want with safety oversight, we may do something you don't expect. But ultimately, they never really waved that threat explicitly, and they never exercised it, even though Google didn't give them the safety oversight they wanted. So in some sense, it's a story of the naivety of the founders, both Demis himself and Mustafa Suleyman, his co-founder who's now at Microsoft. They went into this not quite sure what their endgame was, and in the end, they never really used that spin-out option. So in some sense, spending three years going back and forth to Mountain View with this was a waste.

我想知道，在你与戴密斯的对话中，显然我们已经谈到了 AI 安全峰会，并意识到那并不理想。他是否还有其他关于过去十年如何展开的遗憾？

I'm wondering, in the course of the conversations you had with Demis, obviously we talked about the AI Safety Summit already and realizing that was not ideal. Are there any other things that he regrets about how the whole past decade has unfolded?

他后悔什么？我想他肯定不会后悔把那么多精力花在 AI 用于科学上。顺便说一句，我喜欢你书里的那个轶事：他们赢得 AlphaGo 的那天，AlphaGo 击败了世界最佳棋手，他已经在考虑生物了，有人用麦克风捕捉到了。那正说明了一个雄心勃勃、不断前进的领导者。他享受荣誉大约 10 秒钟，然后说：‘我们接下来要做的是解决蛋白质折叠。’那太棒了。但我认为严肃的一点是，他不仅因此获得了诺贝尔奖。他还认为，我认为正确，这对人工智能在整个社会中的可接受性至关重要。如果 AI 不能为人类带来明确的好处，而只是大量扰乱就业，这可能对生产力有好处，但对受影响的人来说是痛苦的，我不确定 AI 是否会在没有巨大反弹的情况下真正推广。所以我认为，从推进科学的角度看，这很好，我喜欢赢得诺贝尔奖，但 AI 没有这类科学成就也无法成功。我认为他对此毫不后悔。他后悔没有更快地进入聊天机器人领域吗？是的，当然。如果能像 OpenAI 的伊利亚那样理解就好了。Transformer 论文一出来，伊利亚就从椅子上跳起来，跑下走廊去找亚历克·拉德福德说：‘嘿，我们要基于这个 Transformer 架构构建一个语言模型。’论文发表当天，因为他从博士起就有准备，一直在思考如何处理像文本这样的序列数据。所以当 Transformer 论文出来时，他立刻看到了其重要性。如果 DeepMind 也能那么快有同样的认识就好了，但他们却花了两年或三年才赶上。坦率地说，戴密斯对此仍有点盲点。当我对他说：‘你晚了大约三年。’他回答：‘不，不，不，不。我们有 Chinchilla。我们有这些其他模型。’我说：‘是啊，是啊，但你们直到 2020 年底才发布那些。’而伊利亚那时已经研究了三年半。所以你晚了。但他非常抗拒这一点，我认为那是因为他确实后悔了。

What does he regret? I'd say that he would certainly not regret spending all that energy on AI for science. I love that anecdote in your book, by the way, that the day they won AlphaGo, AlphaGo beat the best player in the world, he was already on to bio and someone picked it up on a mic. That was just talking about an ambitious and always moving leader. He rests on his laurels for about 10 seconds and then says, 'The next thing we're going to do is we're going to solve protein folding.' That was amazing. But I think the serious point here is that he not only got a Nobel Prize out of it. He also views this, I think correctly, as absolutely central to the whole acceptability of artificial intelligence in society at large. If AI doesn't deliver clear benefits for humans, and it's just lots of job disruption which may be good for productivity but painful for people on the receiving end, I'm not sure that AI will really be rolled out without some huge backlash. So I think both from it's good to advance science, I like winning a Nobel Prize, but also AI can't succeed without this kind of science stuff. I think he doesn't regret that for a second. Does he regret not being faster onto chatbots? Yeah, of course. It would have been better to understand, like Ilya did at OpenAI. The moment the transformer dropped, Ilya was jumping out of his chair, running down the corridor to find Alec Radford saying, 'Hey, we're going to build a language model based on this transformer architecture.' On the day the paper dropped, because he had this prepared mind ever since his PhD, he had been thinking about how to deal with sequential data like text. So when the transformer paper came out, he immediately saw the significance. Of course it would have been great for DeepMind if they had had that same perception that quickly, and it took them instead two or three years to get there. Demis, frankly, still has a bit of a blind spot about this. When I say to him, 'You were sort of three years late,' he goes, 'No, no, no, no. We had Chinchilla. We had these other models.' And I say, 'Yeah, yeah, but you didn't release those till the end of 2020.' And Ilya by then had been working on it for three and a half years. So you were late. But he really resists that, and I think that is because he does regret that.

嗯，我也觉得很有趣，你在书里提到：对 Transformer 架构的部分抵制源于深厚的神经科学背景，以及一种信念——通往 AGI 的道路必须经过更接近人类学习和大脑的方式，也就是强化学习的世界，而不是仅仅通过消费互联网来理解人类。而你的观点是，事实证明，通过 Scaling 这些模型，你能理解的东西远超预期。

Well, I thought it was interesting too, the way you put it in your book: part of the resistance to the Transformer architecture came from a deep neuroscience background and the belief that the path to AGI must flow through something that looks more like how humans learn and the human brain, which was the RL world, rather than just understanding what it means to be human by consuming the internet. And I think your point was that it turns out you can understand it way more than one might have anticipated from scaling these models.

是的。顺便说一句，我认为机器人领域可能也存在类似的两难困境。为了在模拟中训练机器人，你需要真实世界的模拟，而视频可能是构建这些模拟的重要工具。所以如果 DeepMind 在通往更好机器人的正确路径上押对了注，它具备所需的条件。但我们会看到，语言和 Transformer 模型的问题是否会在机器人领域重演——我们低估了语言对超级智能的重要性——那么要搞定机器人，是不是要通过建造实体机器人，让它们在物理实验室里摔断胳膊？

Yeah. And by the way, I think there may be some version of that same dilemma with robotics. In order to train robotics in simulation, you need real-world simulations, and probably video is one important tool for building those simulations. So potentially if DeepMind gets the right bet on the right pathway to much better robotics, it has what it takes. But we'll see whether there's some version of what went wrong with language and the Transformer model—the gem underestimated the importance of language to superintelligence—and then maybe to get robotics right, is it something you do by building physical robots and having them crash and break their arms in some physical lab?

但这是一个有趣的观点，因为我在机器人领域看到了许多相似之处，人们正在尝试 10 到 15 种不同的方法。我想如果你是谷歌，考虑到 DeepMind 处理其他事情的方式，你会涉足其中很多方法。与此同时，初创公司会全力押注某一种方法。看看结果如何会很有趣，因为当配方不明确时，DeepMind 保持开放选择并非不合理。但这也会让你容易受到那些全力押注 Transformer 或代码或机器人领域下一个等价物的人的冲击。

But it's an interesting point because I see many parallels in the robotics space where there are 10 or 15 different approaches people are trying. And I imagine if you're Google and given the way DeepMind approached other things, you'll put your hand in a bunch of these approaches. Meanwhile, startups will be all-in on one approach or another. It'll be interesting to see how that plays out because you can certainly empathize that when the recipe is not clear, it's not particularly irrational for DeepMind to keep its options open. But it also leaves you vulnerable to someone who really goes all in on the Transformer or on code or whatever the next equivalent is in the robotics space.

100%。而且我认为这里有一个更大的有趣问题：风险投资支持的创新是否胜过那种超大规模科技巨头的 AI 方法。

100%. And I think here there's a larger fascinating question about whether venture-backed innovation beats sort of hyperscaler tech behemoth AI approaches.

我对这个问题有偏见。

I have a biased opinion on that one.

是的，这很公平。但我没那么有偏见。我现在写了一本关于超大规模公司的书，但我之前的书是关于风险投资的。我两种都做过。但我认为这是一个有趣的平衡，因为从某种意义上说，这是一个非常资本密集型的项目。你需要非常雄厚的财力。如果你看看 OpenAI 在生成式 AI 上与谷歌长期对抗的能力，我今年一月在《纽约时报》上写道，我认为 OpenAI 有 50%的几率在明年夏天之前破产。

Yeah, that's fair. But I'm less biased. I've written about a hyperscaler now, but my previous book was about venture. I've kind of done both. But I think it's an interestingly balanced thing because in one sense this is a very capital-intensive project. You need very deep pockets. If you look at OpenAI's ability to fight Google over the long haul on generative AI, I wrote in the New York Times in January that I thought OpenAI had a 50% chance of going bust by next summer.

顺便问一下，现在还是 50%吗？

Is it still 50% by the way?

是的。我的意思是，我说过会失败。我的意思是，它以一个折扣价卖身给某个黑客来扩大规模。我对他们削减开支的一些举措印象深刻。取消 Sora 是一个明智之举。他们做了一些令人印象深刻的成本削减。所以我对此表示赞赏。另一方面，关于 Sam 被 Bret Taylor 取代的传闻相当普遍，而且普遍感觉领导层已经受损。我认为总体而言，我仍然认为他们有 50%的几率在明年夏天之前失败。这不是因为他们没有好的技术——他们有。技术很棒。只是商业模式有问题。而且问题在于你面对的是谷歌，它拥有无限的现金可以把你耗死。所以从某种意义上说，这看起来像是一次有利于现有企业的平台转移。另一方面，正如我们一直在讨论的，当你拥有无限的人才、算力和资金时，你往往不会做出这些战略性的集中押注。而如果你面对的是许多不同的风险投资支持的初创公司，每家公司都做出不同的高度集中的押注，那么也许其中某个押注会成功。

Yeah. I mean, I said fail. What I mean is it sells itself at a discount to some hack to scale. And I've been impressed by some of the expenditure cuts they've done. Canceling Sora was a smart move. They've done some impressive cost cutting. So I give them credit for that. On the other hand, the rumors around Sam being replaced by Bret Taylor are pretty widespread, and there's just a general sense that the leadership is damaged goods. I think on net I'm still around 50/50 that they fail by next summer. And that's not because they don't have great tech—they do. The tech is great. It's just the business model is problematic. And it's problematic because you're up against Google, which just has unlimited amounts of cash to spend you into the ground. So in one sense, this looks like a platform shift that benefits the incumbents. On the other hand, as we've been discussing, when you have the luxury of unlimited amounts of talent and compute and money, you tend not to make these strategic concentrated bets. And if you're up against a bunch of different venture-backed startups that each make different highly concentrated bets, then maybe one of those bets works out.

但这是一个利润非常微薄的游戏，对吧？即使以 Anthropic 为例，一直存在一个问题：他们能否筹集足够的资金来留在这场竞赛中，或者基本上必须出售或与超大规模公司深度绑定，他们在某种程度上已经这样做了。感觉它可能正朝着与谷歌或亚马逊合作的轨道发展。然后就在你可能怀疑的时候，嘿，他们不断提高留在游戏中的所需资金，编码模型爆发了，而且爆发得非常厉害。现在 Anthropic 是湾区最热门的公司。但如果这些东西晚六个月或十二个月才爆发，那将是一个非常有趣的反事实。它当时开始真正起作用并非必然。而且很多这样的事情，事后看来你可以写出一个非常清晰的故事，但事情的发展并非必然。

But it is a game of very thin margins, right? Even if you take Anthropic, there was always a question about whether they would be able to raise enough capital to stay in this race, or basically have to sell or be deeply embedded with a hyperscaler, which they are to some extent. And it felt like it might be headed on that track with either Google or Amazon. Then right at the moment where you might have wondered, hey, they keep raising the amount of capital required to stay in this game, the coding models hit, and they hit in a huge way. And now Anthropic is the hottest company in the Bay. But it's a really interesting counterfactual if that stuff had hit six months later or twelve months later. It certainly wasn't inevitable that it was really going to start working then. And a lot of this stuff, it seems you can write a very clean narrative in retrospect, but it's not an inevitability that things play out that way.

我想回到你刚才说的关于 Demis 的事情，他显然专注于科学，并相信我们必须证明 AI 可以对社会产生有益的结果，否则世界会反对它。有一件事让我特别印象深刻，你在开头就提到了：你不能把 Demis 放在封面上，因为人们认不出他是谁。在很多方面，今天世界与 AI 联系在一起的人是 Sam 和 Dario，他们占据了公众关于如何思考 AI 的讨论中的大量氧气。我想知道 Demis 对此有何感受，以及随着时间的推移，你认为他是否觉得自己必须比现在更成为一个公众人物，以便以更符合他感受的方式塑造这场对话？

I want to go back to something you said about Demis, which was him obviously focusing on science and this belief that we have to show that AI can have beneficial outcomes for society, or else the world will turn against it. And one thing I'm struck by in particular, you kind of alluded to it right at the beginning: that you couldn't put Demis on the cover because people wouldn't recognize who he was. And in many ways, the people the world associates with AI today are Sam and Dario, and they're taking a lot of the oxygen in the public conversation about how people think about AI. I'm wondering how Demis feels about that, and over time, do you think he feels like he will have to become more of a public figure than he is today to shape that conversation in a way that's more analogous with how he feels?

是的。首先，要说明的是，我的出版商达成的妥协是让他上封面，但把他弄得模糊一点，这样他看起来像一个神秘的书呆子，如果你认不出他，它仍然是一张很酷的照片。

Yeah. First of all, to be clear, the compromise my publisher struck was to have him on the cover, but to kind of fuzz it up so that he looks like a kind of enigmatic geek, and if you don't recognize him, it's still a cool picture.

所以，他们就是这么做的。但我觉得，针对你的问题，戴米斯确实明白他需要更多地露面。他在某种程度上很擅长自我叙事，也就是说，他对自己从过去到现在的旅程回顾性故事讲得很好。他跟我聊了超过 30 个小时，当然这其中有原因。此外，还有纪录片《思维游戏》，很多人都看过；在此之前，还有一部关于 AlphaGo 的纪录片，那可不是偶然制作的——他们当时去首尔与世界围棋冠军比赛时，真的带了一个纪录片团队。所以，他擅长这种回顾性的叙事。但他不擅长前瞻性的叙事。他不像山姆那样，仅仅通过发几条推文就能抢走人们对 DeepMind 新发布的关注。

So, that's what they did. But yeah, I think to your question, Demis does understand that he needs to be out there more. He's good at kind of self-narrativizing in one way, which is that the retrospective story of his journey up until where he is now is something that he's been good at communicating. He spent more than 30 hours talking to me, and of course there is a reason for that. But also there was the documentary 'The Thinking Game' which a lot of people have seen, and before that there was a documentary about AlphaGo which didn't get made by mistake—they actually brought a documentary team with them to Seoul when they played that game against the world Go champion. So there is this retrospective storytelling which he's good at. What he doesn't do is prospective. He doesn't, like Sam, have the ability to kill the attention on a new DeepMind release simply by putting out a couple of tweets.

他们总是提前一天行动，好像知道要发生什么似的。

They would always go one day before, like they know something was going to happen.

是的，当我们说提前一天行动，意思是在 X 上预告一下，发几条帖子，甚至不发布任何东西。但那是因为山姆在 X 上的粉丝数是戴米斯的五六倍，回音室效应强得多，即使别人在发布产品，他也能主导叙事。这对 DeepMind 的产品采用和人才招聘都是个问题。控制叙事确实很重要，我觉得 Google DeepMind 的人多少明白这一点，但他们是否有能力应对，我们拭目以待。我的意思是，他们不会像达里奥那样做——与五角大楼公开争吵，发布未公开的神话故事等等，这些事能立刻让你成为超级名人。而戴米斯在某些方面太理智了，不会与政府公开对抗。所以这可能让他不那么出名。也许这是战略谨慎的副产品，使他在媒体上不那么引人注目。

Yeah, and when we say go one day before, we mean just trail something on X, just put out a couple of posts and not even release something. But that was because Sam's following on X is like five or six times Demis's, the echo chamber is so much stronger that he could really dominate the narrative even when the other guys were releasing the product. And that's a problem both in terms of product adoption for DeepMind and also talent recruitment. Controlling the narrative does matter, and I think they kind of get that over at Google DeepMind, whether they quite have the ability to fight it, we'll see. I mean, what they don't do is what Dario does, which is pick a public fight with the Pentagon, unreleased mythos, all these things which just immediately turn you into a massive celebrity. And Demis in some ways is too sensible to pick a public fight with the government. So maybe that makes him less notorious. And maybe it's a kind of byproduct of strategic caution that he's a bit less front of mind in the media.

我很好奇，你最终了解到这在多大程度上能吸引研究人员。显然，这场战斗很大一部分是人才争夺战。谷歌有非常出色的研究人员，比如你在书中提到的杰克·雷，他离开又回来又离开。你怎么看，或者在与人们交谈时，那些被 DeepMind 吸引的人是什么类型，他们在多大程度上需要改变招聘品牌才能与其他公司保持同步？

I'm curious what you ended up learning about the extent to which that does or doesn't attract researchers. Obviously a huge part of the battle here is the battle for talent. You've had, I mean, Google has incredible researchers, you've had folks like Jack Rae who you talk about in the book who left then came back then left again. How do you think about, or in talking with folks, the types of people that are attracted to DeepMind, the extent to which they do or don't need to change their hiring brand to stay on par with the others?

是的，这是个好问题，我认为这和我们刚才说的资金雄厚的超大规模企业与风险投资模式之间的竞争有关。因为我认为杰克·雷离开 DeepMind 去 OpenAI 时，他向我解释说，这是因为 OpenAI 当时集中押注语言模型，他们只关心这个，而杰克做的就是这件事，所以他想去那里。我认为今天在机器人领域也是如此。有些非常优秀的研究人员，他们想去一个他们自己相信的集中押注的地方，然后全力以赴成为团队的一部分，实现他们认为正确的道路。而如果你把一个机器人研究人员放在一个庞大、谨慎、多元化的实验室里，说我们非常相信你，但我们也同时相信其他五个押注，他就不会感觉那么好。缺少了一种激情。

Yeah, it's a great question and I think it sort of overlaps with what we were saying about the rivalry between the deep-pocketed hyperscaler and the venture model. Because I think for Jack Rae when he left DeepMind and went to OpenAI, he explained to me that this was because there was a concentrated bet on language models at OpenAI and that's all they cared about at the time and that's what Jack was doing, so he wanted to be there. And I think that would be true today in robotics. There'll be certain researchers who are really great and they want to go somewhere where there's one concentrated bet which they believe in themselves, and then they will just work all out to be part of that team and to realize what they think is the right path. Whereas if you put a robotics researcher in this big careful diversified lab and say we really believe in you but we also believe in these other five bets we're making at the same time, you don't feel so good about it. There's a kind of passion piece that's missing.

你怎么看这个决定？显然 AlphaFold 的工作最终被剥离或分拆成了 Isomorphic，对吧，成为一家不同的公司，Alphabet 仍然拥有很大一部分，但这是一个独立的努力。这合理吗？还是你认为这条路会被重复？

What do you make of the decision? Obviously the AlphaFold work eventually got folded out or spun out into Isomorphic, right, in a different company that Alphabet still owns a bunch, but is a separate effort. Does that make sense or is that a path forward you think will be repeated?

这是个好问题。我认为 Isomorphic 确实有这种效果。我认为分拆也是为了能够与制药合作伙伴建立合作关系，也许在分拆的形式下更容易一些。

That's a great question. I mean, I think Isomorphic does have that effect. I think the spin-out was also about trying to be able to do partnerships with pharma partners, and maybe that was a bit easier in a spun-out format.

只是因为他们不想直接与谷歌合作吗？

Just because they wouldn't want to work with Google directly or something?

也许吧。而且可能还有一种感觉，就是这件事本身作为一个独立实体可以变得非常庞大，而且谷歌和 DeepMind 之前各自在 AI 健康领域有过尝试，健康领域有自己的政治和非常长的周期。他们觉得它需要一个不同的归宿。所以他们这么做了。但我认为你提出了一个很好的观点，可能还有一个额外的理由，那就是你给人们提供了一个选择，让他们可以在一个他们自己就是押注、他们的事情就是核心的地方工作。我认为这是一个非常重要的招聘工具。我还要指出，在硅谷可能众所周知，Anthropic 的人员流失率相对其他公司非常低，我认为这表明，如果你的领导者身份像达里奥那样，公开且不加过滤地表达对安全的极度担忧以及对责任和社会影响的极度关注，写那些长文，这可能意味着有些人永远不会加入你，因为他们觉得你有点古怪，但加入的人绝对爱你、相信你，这种惊人的忠诚度非常特别。

Maybe, yeah. And maybe also just the sense was that this by itself as a freestanding thing could become so big, and there'd been this history of Google and DeepMind separately doing AI for health, and health has got its own politics, its own kind of very long lead times. They felt it just needed to be in a different home. So they did that. But I think you're raising a good point that there might have been an extra reason to do it, which is that you give people the option of working somewhere where they are the bet, their thing is the thing. And I think that's a super important recruiting tool. I also would note, it's probably fairly well known in the valley, but the Anthropic churn is very low relative to everybody else, and I think that shows you that if your identity as a leader is like Dario where you're kind of out there and unfiltered about your extreme concern for safety and your extreme concern for responsibility and social impact, you write these long essays, it probably means that some people would never join you because they think you're a bit wacky, but the ones who do join absolutely love you and believe in you, and there's this kind of amazing loyalty, which is kind of special.

你作品中的另一个关键人物，我们还没谈到的，是大卫·西尔弗，他显然在强化学习工作中发挥了关键作用，并且是早期与他们合作很多事情的人之一。显然，自从你的书出版后，他最近离开了 DeepMind，创办了另一家公司。我很好奇你怎么看这件事。

Another kind of key character in your work that we haven't talked about is David Silver, who obviously played a key role in a bunch of the reinforcement learning work and was one of the early collaborators with them on a bunch of things. He obviously, I think since publishing your book, he recently left DeepMind to start another company. I'm curious what you made of that.

我认为这实际上符合你的叙事。我的意思是，他是一个极度坚定的强化学习信徒，以至于我认为 DeepMind 的大多数同事最终都觉得他有点过头了。

I think it kind of fits your narrative here actually. I mean, in the sense that he was this intensely determined believer in reinforcement learning to a point where I think most of his colleagues at DeepMind ended up thinking it was just too much.

嗯，你知道，他是英雄，首先，DeepMind 在 2012-2013 年推出的 Atari 游戏系统取得了惊人的突破，那几乎就像 ImageNet 但针对早期智能体，而 David Silver 是其中的关键人物。然后是 AlphaGo，接着是 AlphaZero，它几乎完全从强化学习中学习，没有深度学习，但严重偏向强化学习。然后 David Silver 和他的博士导师 Rich Sutton 发表了一篇论文，叫《经验就够了》之类的。基本上，那篇论文的信息是“从经验中学习，从强化学习中学习，不要从数据中学习”。David 非常坚持这个观点，认为从数据中学习是次等的，因为数据包含错误。以围棋为例，如果你用人类围棋选手过去的棋局训练，即使是专家棋局，那些选手对围棋的理解也不完美，你需要超越它才能获得超级智能。所以机器需要从自己的经验中学习，而不是依赖人类通过文本或其他数据传递的固化知识。他对此深信不疑，对他来说一切都是智能体，只有智能体，而且它们必须从自身学习。我想 Demis 告诉过我一次，实际上不止一次，那种方法可能在未来的某个时候最终获胜，当你已经拥有 AGI，现在只是在将其完善为更强大的超级智能时。因为最终，如果机器从自己的数据中学习，那更纯粹，会带你走得更远，但要达到 AGI，你需要用现有数据引导自己，这就是整个语言模型革命向我们展示的。很长一段时间几乎没有强化学习，除非你算上 RLHF，而且如果它们的基础模型不够强，它就行不通，除了这些特定领域。我认为这个时刻的迷人之处在于大型预训练语言模型和这些可验证领域的强化学习的结合。但有趣的是他这个时候离开了，因为感觉强化学习又流行起来，成为改进这些模型的方法。所以几年前当强化学习更不受欢迎时离开可能更有意义。我觉得这个时机很有意思。

Um you know he was the hero when, first of all, the Atari game playing system that DeepMind rolled out in 2012-2013 made these incredible breakthroughs, and that was almost like ImageNet but for early agents, and David Silver was a key person in that. And then there was AlphaGo, and then there was AlphaZero, which was pretty much all learning from reinforcement learning, no deep learning, but heavily skewed towards reinforcement learning. And then came this paper from David Silver and Rich Sutton, his PhD supervisor, called 'Experience is Enough' or something like that. And basically 'learn from experience, learn from reinforcement learning, don't learn from data' was the message of that paper. And David is just very very hard over on that vision that learning from data is inferior because the data includes mistakes. If you take the Go analogy, if you train on human Go players' past games, even expert games, those players don't have perfect understanding of Go, and you need to get beyond that to have superintelligence. So the machine needs to learn from its own experience, not rely on the crystallized knowledge of humans passed on through text or other data. And he is such a believer in that that it's all agents for him, only agents, and they have to learn from themselves. And I think Demis told me once, or actually more than once, that that kind of approach may ultimately win in some future when you already have AGI and now you're just perfecting it into even greater superintelligence. Because ultimately, if the machine learns from its own data, that is purer and will get you further, but to get to that AGI you need to bootstrap yourself with existing data, and that's what the whole language model revolution showed us. For a long time there was almost no reinforcement learning unless you count RLHF, and it didn't work if their base models weren't strong enough, except for these specific domains. And I think what's fascinating about this moment in time is the combination of large pre-training LMs and then reinforcement learning on these verifiable domains. But it's interesting that he left at this time because it feels like reinforcement learning is back in vogue as a way to improve these models. So it almost would have made more sense a few years ago when reinforcement learning was more out of favor. I thought it was interesting timing that it happened now.

是的。我的意思是，我认为你说得对，但这反映了人类在感觉和行动之间的滞后，对吧？我认为他一段时间以来一直觉得，他被困在一个大组织中，这个组织根本不想把超过一小部分筹码放在强化学习上。即使在强化学习内部，他对应该如何做的愿景也与其他人不同。我认为他是一个典型的创业型人物，想在一个小组织中，让他的愿景成为组织的愿景，因为他在强化学习方面非常有远见，所以他去做自己的事情是有道理的。

Yeah. I mean, I think you're right, but it sort of reflects the human lag between feeling something and acting on it, right? I think he'd been feeling for a while, frankly, that he was swamped in a big organization which didn't fundamentally want to put more than a small amount of its chips on the reinforcement learning table. And even within reinforcement learning, his vision of how it should be done differed from some other people's. I think he is a classic startup sort of person who wants to be in a small organization where his vision is the vision, because he's extremely visionary on reinforcement learning, and so it makes sense for him to go do his own thing.

我们在这里讨论了很多的一件事是，归根结底，有几个人运营这些实验室，他们有着强烈的个人历史和关系。我认为 Sam 和 Dario 的关系已经被充分记录。Elon、Sam，都在当前的法庭案件中浮出水面。我认为 Demis 与每个人的关系可能不太为公众所了解。我想知道你是否能谈谈这个，以及他对目前这个领域的另外两个主要人物的看法。

One thing that we've kind of talked about a bunch here is just like in the end of the day there are a few people who run these labs and have intensely personal histories and relationships. I think the Sam-Dario relationship has been incredibly well documented. Elon, Sam, all coming out in this current court case. I think Demis' relationship with each of them is probably less well understood by the general public. And I'm wondering if you could just talk a little bit about that, and his feelings on the two main other protagonists of the space right now.

所以，Demis 和 Elon 的关系非常有趣。他试图买下他，对吧？

So, Demis's relationship with Elon is very interesting. He tried to buy him, right?

是的。但回顾一下，这一切的起源是他们都是由同一个风投机构 Founders Fund 资助的。Elon 是 SpaceX 的人，Demis 是 DeepMind 的人。在 2012 年左右，他们都被邀请参加一个 LP 场外会议，他们都做了演示，然后开始交谈。Elon 总是很有竞争意识。那本应是一个值得关注的场外会议。但 Elon 说，‘嗯，我拥有世界上最重要的技术，因为即使你的 AI 搞砸了世界，我们都可以搬到火星，成为一个多行星物种。所以我拥有最重要的东西。’这时，Demis 说，‘是的，但如果你认为你在火星上会安全，记住我的 AI 将能够征服太空飞行，它会跟着你去火星。所以你还是不安全。’然后一阵沉默，Elon 说，‘嗯。’接着他说，‘嗯，我想投资你的 B 轮，’他给 B 轮开了一张 500 万美元的支票。然后，如你所说，他在 2014 年初想买下 DeepMind，以防止它被卖给 Google。Demis 只是挥手拒绝。有一个疯狂的故事，Founders Fund 的合伙人 Luke Nosek，他也是 SpaceX 的董事会成员，在洛杉矶的一个派对上看到 Elon，他们一致认为必须阻止 DeepMind 被卖给 Larry Page。‘她是个超人类主义者。你不能信任他。’所以他们上楼到一个壁橱里，在半夜给伦敦的 Demis 打 Skype，说，‘你必须卖给我们，而不是他们。卖给 SpaceX。卖给 Tesla。做点什么，但不要卖给 Google，因为他们是邪恶的。’Demis 说，‘你知道，不，不，Google 有计算机。我要卖给他们。再见。’

Yeah. But going back, the origin of this whole thing is that they were both funded by the same VC shop, Founders Fund. Elon was the SpaceX guy and Demis was the DeepMind guy. In 2012 or thereabouts, they both get invited to an LP offsite and they both present, and then they get talking to each other. Elon's always very competitive. That would have been a pretty valuable offsite to pay attention to. But so Elon is like, 'Well, I've got the most important technology in the world because even if the world is screwed up by your AI, we can all move to Mars, be a multiplanetary species. So I've got the most important thing.' At which point, Demis says, 'Yeah, but if you think you're going to be safe on Mars, remember that my AI will be able to conquer space flight, and it will just follow you to Mars. So then you won't be safe after all.' And then there's a silence, and then Elon goes, 'Hm.' And then the next thing he says, 'Well, I'd like to invest in your Series B,' and he writes a $5 million check into the Series B. And then, as you say, he wants to buy DeepMind at the start of 2014 to prevent it from being sold to Google. And Demis just waves him off. There's this crazy story where Luke Nosek, the Founders Fund partner who is sitting on SpaceX's board, sees Elon at a party in LA and they agree that they've just got to stop DeepMind from being sold to Larry Page. 'She's a transhumanist. You can't trust him.' So they go up to some closet upstairs in this party and they Skype Demis in the middle of the night in London and say, 'You got to sell to us, not to them. Sell it to SpaceX. Sell it to Tesla. Do something, but just don't sell to Google because they're evil.' And Demis is like, 'You know, no, no, Google's got the computer. I'm selling to them. Goodbye.'

晚安。然后他挂了电话，埃隆就发疯了，对吧？开始称戴米斯是邪恶天才，这指的是戴米斯当年做游戏设计师时参与的一款游戏。你知道，他诋毁戴米斯，对此执念很深。我觉得这一点在我们最近进行的审判中又有所体现，就是埃隆对萨姆的审判，以及埃隆对戴米斯作为必须被制衡的邪恶天才的那种执念。这就是那段历史。戴米斯非常想强调，现在他们相处得很好。我没有直接从埃隆那里听到过，但我感觉可能已经翻篇了。埃隆已经转向和萨姆争斗了。很可能他现在对戴米斯已经没什么了。但你知道，这就是历史。然后和萨姆相比，如果你对比他们两个，性格和背景简直天差地别，对吧？戴米斯有诺贝尔奖，萨姆连本科学位都没读完。因此，戴米斯不太把萨姆当回事，对吧？他没有大学学位，更别说博士或诺贝尔奖了。对戴米斯来说，萨姆——而且我觉得这确实有些道理——萨姆是硅谷网络那种熟练的终极化身，他知道如何假装直到成功，如何提前说出不成熟的真相，他非常擅长筹钱，利用他在硅谷的人脉。你知道，这都很棒，但这和做一个严肃的科学家不是一回事。你不能信任那样的人。而且我觉得，你看，我认为人们在某种程度上已经认同了戴米斯对萨姆的看法，但戴米斯一直这么觉得。所以我不认为他喜欢过萨姆。

Good night. And he puts the phone down and then Elon goes nuts, right? And starts calling Demis an evil genius, which is a reference to a game that Demis worked on when he was a video game designer. And you know, vilifies Demis and is obsessed with this. And I think this has come out again a bit more in the trial that we've just been having recently, you know, the Elon versus Sam trial, and the kind of obsession that Elon had with Demis as the evil genius that had to be counteracted. So that's the history there. Demis is kind of very keen to say that these days they get on fine. And I haven't had that directly from Elon, but I kind of feel that probably water under the bridge. Elon has moved on to fighting with Sam. It probably is true that he's okay with Demis now. But, you know, that's the history there. And then with Sam, it's just such a complete difference in personality and background if you compare the two of them, right? So Demis has a Nobel Prize; Sam didn't finish his first degree. Therefore, Demis doesn't take Sam very seriously, right? He doesn't have a college degree, let alone a PhD or a Nobel Prize. And to Demis, Sam—and I think there's some truth in this frankly—that Sam is the sort of skillful ultimate embodiment of the Silicon Valley network who knows how to fake it till you make it, to tell the premature truth, who is just a master at raising money, at leveraging his connections in the valley. And you know, that's all great, but it's not the same as being a serious scientist. And you can't trust somebody like that. And I think, you know, look, I think people have come around to Demis' view to quite some extent with Sam, but Demis always felt that. And so I don't think he ever liked Sam.

我记得《经济学人》在谈论你的书时说，这简直是对历史伟人理论的一次检验。但我很好奇——我们聊了很多关于这四位主角的事。我在想，做完所有这些工作之后，这一切是不是某种必然？或者说，由戴米斯和可能其他这些人来掌舵公司，到底有多重要？我很好奇你的想法。

I think it was The Economist that, in talking about your book, said it's really like a test of this great man theory of history. But I'm curious—we talked a lot about these four protagonists. And I'm wondering, after having done all this work, was all of this kind of inevitable? Or to what extent did it matter that it was Demis and maybe some of these other folks at the helm of the companies? I'd be curious your thoughts on that.

我曾写过一本关于艾伦·格林斯潘的书，他是全球经济中极其强大的角色，但最终无法阻止 2008 年的金融泡沫。我把那本书叫做《知道的人》，因为他知道泡沫是危险的。事实上，他读博士时研究的就是这个。他对泡沫破裂非常着迷，但他无法阻止这件事发生。我觉得几乎可以用同样的书名来写戴米斯的书，对吧？我把它叫做《无限机器》，但它也可以叫《知道的人》，因为戴米斯从一开始就知道这东西是危险的。但作为一个实验室的领导者，即使是一个非常有钱的实验室，即使他有诺贝尔奖得主的声望，他明白这是危险的。但他能做什么呢？因为如果他把自己的实验室搞安全了，并不能阻止其他人不安全。所以我认为竞赛动态中有某种必然性。而谁在领导这些实验室确实很重要。我的意思是，显然，达里奥首先通过公开与五角大楼对抗，其次更重要的是通过他发布 Mythos 的方式，改变了 AI 安全的叙事。这是一个明显的例子。另一方面，萨姆决定发布 ChatGPT，尽管它当时在胡编乱造。那是一个选择。他本可以不那么做。而这完全改变了 AI 竞赛的走向。而戴米斯，以一种更低调的方式，告诉当时的英国首相里希·苏纳克，嘿，我们应该举办一个全球 AI 安全峰会。你知道，后来就发生了。然后在布莱切利公园举办了一次，中国也来了，这算是国际 AI 安全对话的开端。所以我认为戴米斯做事更在幕后，但没错，我认为这些领导人的个性很重要，但也许这不是唯一重要的因素——还有潜在的力量在起作用。

I once wrote a book about Alan Greenspan, super powerful player in global economics, but ultimately unable to stop the financial bubble of 2008. And I called the book The Man Who Knew because he understood that bubbles were dangerous. That's in fact what he read his PhD about. He was obsessed with bubbles blowing up, but he couldn't stop this thing from happening. And I feel that one could have almost used the same title for the Demis book, right? I called it The Infinity Machine, but it could have been called The Man Who Knew because Demis has known from the beginning that this thing is dangerous. But as the leader of one lab, even a very powerful rich lab, even he with his stature as a Nobel Prize winner, he understands that it's dangerous. But what can he do? Because if he makes his own lab safe, it doesn't stop the other guys from being unsafe. And so I think there is a sort of inevitability to the race dynamic. And it does matter who is leading these labs. I mean, clearly, Dario changed the narrative on AI safety first by fighting with the Pentagon in public and secondly and more importantly by the way that he did the Mythos release. So that's a clear example. On the other hand, Sam decided to release ChatGPT even though it was hallucinating. That was a choice. It didn't have to do that. And that completely colored the way the AI race played out. And Demis, in a more understated way, told Rishi Sunak, the UK prime minister at the time, hey, we should have a global AI safety summit. And you know, that's what happened. Then there was one in Bletchley Park and the Chinese came and that was kind of the beginnings of an international conversation on AI safety. So I think Demis is more behind the scenes in what he does, but yeah, I think it matters the personalities of these leaders, but perhaps it's not the only thing that matters—there are underlying forces as well.

好吧，在我们结束之前，鉴于你和戴米斯相处了那么长时间，我觉得有几个关于那个过程的问题。我真的很想深入了解一下——你大概花了三年时间，在同一家酒吧和他见面。我想知道，在那一系列对话中，你对他的看法在哪个时刻改变最大？

Well, before we wrap up, I definitely, given that you spent so much time with Demis, I think there are a few questions about that process. I was really curious to dig into—you spent, I guess, three years meeting him at the same pub. And I'm wondering across those series of conversations, what was the moment your view of him kind of shifted the most?

有很多惊喜，但其中之一是他对探索科学深层奥秘的信念之深。这实际上是一种精神信念，我之前毫无察觉。但你知道，他有时会在我们的谈话中爆发出来。每次谈话持续两个小时。所以我们能深入探讨。他会开始敲桌子说，看，这张桌子，它是一个谜。一个谜。我们不明白它。就像这些原子在跳动，它们之间有间隙，但桌子却是实心的。为什么你的笔记本电脑只是一堆沙子和铜，却能思考？你知道，这到底是怎么回事？为什么世界被设置成这样，让它能够运转？这是一个我们必须理解的谜。背后一定有某种智能。这不可能是巧合。也许就像上帝。也许如果我们以正确的方式接近科学，我们就能更了解自然。我们会更接近某种我们或许可以称为上帝的东西。我完全不知道他会这么想，但我认为他确实如此。这解释了他为什么仍然推动开发这种超级强大的 AI，尽管他知道这很危险。因为对他来说，这是一种准精神追求。

There were lots of surprises, but one of them is the depth of his conviction about discovering the deep mysteries of science. And this is actually a kind of spiritual conviction which I had no inkling of before. But you know, he would sometimes erupt in these conversations I was having. They went on for two hours each time. So we could already get deep on stuff. And he would start banging the table and saying, look, this table, it's a mystery. It's a mystery. We don't understand it. Like these atoms jumping around and there are gaps between them and yet the table is solid. And why is your laptop able to think when it's just a bunch of sand and copper? And you know, what's going on here? Why is the world set up like this so that it functions? It's a mystery we have to understand. There must be some sort of intelligence behind it. This can't just be coincidence that it's like this. Maybe it's like God. Maybe if we approach science the right way, we understand more about nature. We will be getting closer to something that we could perhaps call God. Now, I had no idea that he would feel that way, but I think he does. And it explains why he nonetheless pushes forward to develop this super strong AI, which he knows to be dangerous. It's because it's a kind of quasi-spiritual quest for him.

感觉对这个领域的很多人来说，实现 AGI 有一种宗教或精神元素。看到这一点真的很有趣。当然，我认为这在你的作品中表现得非常清楚。我想，你知道，感觉你得到了一本非常广泛和开放的书。有没有戴米斯不太愿意谈论的事情，或者你本来想写进书里但最终没有实现的部分？

It feels like for a lot of people in the space, there's a religious or spiritual element to getting to AGI. And it's really interesting to see play out. And certainly I thought that came through really clearly in your work. I guess, you know, it feels like you had such a wide-ranging and open book. Were there things that Demis wasn't really willing to talk about, or areas that you would have liked to put in the book but didn't end up coming through?

是的。他一开始就很明确，他会谈论很多关于他自己的事，但不会谈论他的家庭，所以我就没写。只有一些简短的提及——他娶了他在剑桥的女朋友。他们仍然在一起。有两个孩子。一切都很正常。

Yeah. He was very clear at the start that he would talk a lot about himself but he wouldn't talk about his family, and so I left that out. And there are passing references—he married his girlfriend from Cambridge. They're still married. They have two children. It's all very normal.

嗯，但我基本上不谈——我写格林斯潘时，采访他很多很多前女友是理解这个人的一种方式，而且确实很好地展现了他的人性面。对戴密斯，我没那么做。另一件事是，他不想谈自己和桑达尔·皮查伊以及山景城谷歌领导层之间的争执。我确实写了，因为别人告诉了我，所以我写了不少。所以他不希望被写进去的偏好没有被尊重。他不太想让我谈他解雇联合创始人穆斯塔法·苏莱曼的方式。他有点像是说：‘我并没有真的解雇他，你知道，有一个过程，总法律顾问对霸凌行为进行了某种调查，等等，那算是导火索，但不是穆斯塔法被赶走的深层原因。在 DeepMind，除非戴密斯想，否则什么都不会发生。’我认为他就是觉得是时候让穆斯塔法走了，因为他们在太多问题上意见不合。所以，有些事他不想被写进去，有时他的偏好被尊重了，有时没有。

Um, but I basically don't talk—I mean, most of what I wrote about Greenspan, going to speak to his many, many, many ex-girlfriends, was one way of understanding the guy. And it actually brought out the human side of him in a good way. And I didn't really go there with Demis. Another thing is he didn't want to talk about fights between himself and Sundar Pichai and the leadership of Google in Mountain View. I did write about that because other people told me about that, so I get into that quite a lot. So his preference that this should be left out was not honored. He didn't really want me to talk about the way that he fired his co-founder Mustafa Suleyman. He would sort of say, 'I didn't really fire him, you know, there was a process, some sort of inquiry into bullying done by my general counsel and so forth, and that was kind of the trigger, but it wasn't the deep reason why Mustafa was pushed out. Nothing happened at DeepMind unless Demis wanted it to happen.' I think he just decided it was time for Mustafa to go because they disagreed on too many issues. So yeah, there were some things that he didn't want included, and sometimes he won that preference, and other times he didn't.

关于你的书，公众讨论很多。我觉得你上了很多节目，很多人都在谈论它，而且我认为他们触及了我们今天讨论的很多主题。你觉得书中是否有某些部分或方面被讨论得不够，或者你希望人们更多地谈论？

There's been a ton of public discourse about your book. I feel like you've been on lots of folks have been talking about it, and I think have picked up on a lot of the themes we've discussed here today. Are there parts of the book or aspects of it that you think are underdiscussed or you wish people would talk about more?

我认为对于科学创新有一些有趣的启示。如何做深度科技公司——我们谈到了 DeepMind 的一些缺点，比如缺乏集中押注等等。但从积极方面看，我觉得他发展出所谓的‘科学品味’非常有趣。我的意思是，当他创办第一家初创公司 Elixia（一家电子游戏制作公司）时，他基本上因为产品工程上过于雄心勃勃而搞砸了：图形能有多棒？能不能用早期的强化学习让角色更有趣？他把技术团队逼得太紧，结果他们无法按时交付产品。那是个糟糕的结果。他从第一家公司赚了些钱，但远没有他希望的那么成功。但快进到 DeepMind，2018 年，AlphaFold 研究进行了两年，他的 AlphaFold 团队已经做出了世界上最好的蛋白质预测系统。团队负责人 Andrew Senior 说：‘好了老板，我们已经做到了，让我们宣布胜利然后继续前进吧，因为我们不可能精确预测蛋白质——别开玩笑了，那是不可能的。我们是世界第一，这已经够好了。’戴密斯说：‘不，不，重点不是这个。我们想要预测蛋白质，这样研究生物学家和医学研究人员就能用我们的预测来制造药物和其他重大突破。仅仅比别的实验室好是没有意义的。我们要解决这个问题。’而团队负责人 Andrew Senior 说那不可能。于是戴密斯参加团队会议，这就是科学品味发挥作用的地方。他倾听他所谓的‘想法的流动性’。如果研究团队在互相碰撞各种可以探索的可能性，那就是流畅的，你应该继续前进，投入更多资源，更加努力。如果出现沉默，没人有好主意，那当然应该放弃。他做了这个测试。他倾听团队，他们想法交流很流畅。于是他就换掉了项目负责人。Andrew Senior 去做别的事。他换上了另一个人 John Jumper，后来 John Jumper 和他一起获得了诺贝尔奖。我认为从搞砸第一家初创公司到在蛋白质折叠上做出正确判断，这个演变过程对于如何在公司内部做前沿科学提供了一个教训。

I think there are sort of interesting takeaways for scientific innovation. How you do deep tech companies—we talked about some of the shortcomings at DeepMind, the lack of concentrated bets and so forth. But on the upside, I think it's very interesting how he developed what he calls scientific taste. What I mean here is that when he had his first startup, which was this video game production company called Elixia, he basically blew it up by being too ambitious in terms of product engineering, like how fantastic could the graphics be? Could you do kind of early reinforcement learning to make the characters more interesting? And he drove his technical team over the brink in terms of how ambitious they had to be, and they couldn't ship product on time as a result. So that was a bad outcome. He got some money out of that first company, but it wasn't as much of a success as he hoped. But then you fast forward and he's doing DeepMind, and he gets this moment in 2018 where he's two years into the AlphaFold research, and his AlphaFold team has produced the best protein prediction system in the world. The boss of that team, Andrew Senior, says, 'Okay boss, we've done this now, let's declare victory and move on, because we're not going to predict proteins precisely—give me a break, that's impossible. We're the best in the world. That's good enough.' And Demis said, 'No, no, that's not the point. We want to predict proteins so that research biologists and medical researchers can use our predictions to build medicines and other fantastic breakthroughs. There's no point just being the best better than the other labs. We want to solve this problem.' And the guy who's running the team, Andrew Senior, says that's impossible. So Demis sits in the meetings of the team, and this is where the scientific taste comes in. He listens to what he calls the fluidity of ideas. If the research team are bouncing possibilities of stuff they could look into off of each other, then that's fluent, and you should move forward, put more resources in, push harder. If there was kind of a silence and nobody had any good ideas, of course you should give up. So he does that test. He listens to the team; they are fluid in their exchange of ideas. And so then he basically switches out the project lead. Andrew Senior does something else. He puts in this other guy, John Jumper, who then goes on to become the co-winner of the Nobel Prize with him. And I think that evolution from blowing up his first startup to being right about protein folding holds a lesson for how you do frontier science inside a company.

我喜欢这个故事。我想，显然，这项工作的一部分确实有助于更广泛地提升和讲述戴密斯的故事，这可能是整个工作中不太为人所知的故事。或者总的来说，AI 生态系统中是否有其他你同样觉得值得更多曝光或传记的人，其他 AI 世界中的无名英雄？

I love that story. I guess, obviously, part of this work really helps elevate and tell the story of Demis more broadly, which is maybe an underknown story throughout this work. Or in general, are there other folks in the AI ecosystem that you feel similarly deserve more light shown on them or an exposé, some kind of biography, other unsung heroes in this AI world?

嗯，在我的书里，我确实有伊利亚·苏茨克沃和大卫·西尔弗这样的配对。一个代表深度学习传统，在多伦多师从杰弗里·辛顿获得博士学位；另一个代表我所谓的埃德蒙顿-阿尔伯塔传统，师从强化学习大师里奇·萨顿。我认为这两个人——我在戴密斯的故事里嵌入了他们俩的迷你传记，并列他们的两种方法——可以写一本精彩的双人传记。

Well, in my book, I do have this sort of pairing of Ilya Sutskever and David Silver. One of them representing the deep learning tradition, PhD under Geoffrey Hinton in Toronto, and on the other hand David Silver representing what I call the Edmonton Alberta tradition, PhD under Rich Sutton, who was the sort of guru of reinforcement learning. I think both of those individuals—and I have kind of embedded in my story about Demis a kind of mini biography of the two of them, juxtaposing their two approaches—and there could be a fantastic double biography of the two of them.

当然，我认为随着谷歌 AI 运势时好时坏，人们一直猜测也许有一天戴密斯会成为谷歌 CEO。你觉得这可能吗？

Certainly, I think at various times as Google's AI fortunes have seemed better or worse, folks have speculated that maybe one day Demis will become CEO of Google. You think that's possible?

是的，有可能。问题在于他是否想这么做，因为我认为他确实喜欢能思考科学，他每天做双班——意思是和家人一起吃晚饭，然后回到桌前工作到凌晨 4 点，那是他做科学的时间。现在，我觉得实际上这些天很多深夜时间都花在给山景城打电话，试图协调所有向他汇报的人。所以科学时间已经被侵蚀了，但如果他成为 CEO，显然就一点不剩了。我认为他对于是否想要这个职位确实很矛盾。但在我看来，这完全取决于——比如如果桑达尔决定离开。如果不是戴密斯，谁可能成为 CEO 继任者？如果是一个戴密斯觉得非常舒服的人，他可能完全乐意不去做 CEO。

Yeah, it's possible. There's a question about whether he would want to do that because I think he does love being a little bit able to think about the science, and he does this double shift every day—meaning he eats dinner with his family and then goes back to his desk until 4:00 a.m., and that's when he's doing science. Now, I think actually these days a lot of that late time is spent doing calls to Mountain View and trying to coordinate with all the many people who report to him out of Mountain View. So already that science time is being eaten into, but if he became CEO, obviously there'd be nothing left of it at all. And I think he's genuinely conflicted about whether he would want that. But it seems to me that it all depends—like if Sundar were to decide that he's leaving. Who would be the likely CEO replacement if it wasn't Demis? And if it was somebody that Demis felt very comfortable with, he'd probably be perfectly happy not to do that.

如果他不舒服，他可能会毛遂自荐，因为他不想为他不认同的人工作。嗯，塞巴斯蒂安，这真是一次有趣的对话。我总喜欢把最后一句话留给嘉宾。我想这次，你想把大家引向哪里已经很清楚了，但我还是把最后一句话留给你。我们的听众还能从这次对话中得到什么？另外，请宣传一下你的书。

If he was not comfortable with it, then he might kind of push his hat into the ring because he wouldn't want to work for somebody that he didn't agree with. Well, Sebastian, it's been such an interesting conversation. I always like to leave the last word to the guest. I think in this case, it's pretty clear where you might want to point people, but I'll leave the last word to you. Anything our listeners should take away further from this conversation? And also, please plug the book.

嗯，你看，我真的很喜欢写这本书。这是我的第六本书，和戴密斯·哈萨比斯在英国酒吧顶层待了 30 多个小时，真的很特别。他想让我指出，我们喝的是咖啡，不是啤酒。但仅仅是这个人的知识广度——他能即兴谈论神经科学、计算机科学、物理学、生物学、化学，还有电影史、小说、科幻——简直太疯狂了。我确实试图在书中传达那种能量。这是我第一次在书中使用第一人称作为手法。因为我想用我和他的对话。所以让戴密斯即兴发挥，然后我问他问题，或者在他说话时穿插我的想法，帮助读者理解。这是一次非凡的经历，促使我在自己的写作技巧上进行了创新。所以，是的，我很享受。我希望你们读了也会喜欢。

Well, look, I really enjoyed writing this book. It's my sixth book and spending more than 30 hours with Demis Hassabis in the top of a British pub was really quite special. He would like me to point out that we were drinking coffee, not pints of beer. But just simply the intellectual range of this guy who can riff about neuroscience, computer science, physics, biology, chemistry, the history of movies, novels, science fiction. It was just wild. And I do try to communicate that energy on the page. It's the first time I've ever used the first person as a device in a book. Because I wanted to use that dialogue I had with him. And so having Demis riff, then me kind of ask him a question or intersperse in what he's saying with what I'm thinking to help the reader understand it. This was such an extraordinary experience that I was driven to innovate the craft of writing in terms of my own craft. So yeah, I enjoyed it. I hope you do too if you read it, everybody out there.

太棒了。非常感谢你来播客谈论这本书。

Amazing. Well thanks so much for coming on the podcast to talk about it.

太棒了。谢谢你，雅各布。这很棒。

Fantastic. Thank you Jacob. It was great.

我是雅各布·埃夫隆，这里是《无监督学习》播客，在这里我可以和人工智能领域最聪明的人交谈，问他们大量关于模型发展以及这对世界上的企业意味着什么的问题。我希望大家清楚，我对此乐在其中。这是我除了在红点投资做投资人之外的夜间和周末项目。但我们能请到这些了不起的嘉宾，全靠像你这样的听众订阅播客、与朋友分享。这最终是让这一切运转起来的关键。所以，请考虑这样做。非常感谢你的支持和收听。我们下期再见。

I'm Jacob Efron and this has been Unsupervised Learning, a podcast where I get to talk to the smartest people in AI and ask them tons of questions about what's happening with models and what it means for businesses in the world. As I hope is clear, I have a ton of fun doing this. It's a nights and weekends project in addition to my day job as an investor at Redpoint. But our ability to get these incredible guests on really comes from folks like you subscribing to the podcast, sharing it with friends. It's really what ultimately makes this whole thing work. And so, please consider doing that. And thank you so much for your support and listening. We'll see you next episode.