彭博深度访谈｜微软穆斯塔法·苏莱曼：“人工智能已然超人”

来源：科技世代千高原

微软的穆斯塔法·苏莱曼：“人工智能已经超越人类”

这家科技巨头的人工智能主管谈到了超级智能的“红线”，人工智能为何将改变医学，以及他如何通过与 Copilot 聊天来放松身心。

作者：米沙尔·侯赛因

2025年12月12日

人工智能竞赛正进入一个前所未有的、代价高昂的新阶段。今年是巨额交易之年——数十亿美元涌入数据中心，主要公司之间展开投资，以及一场争夺最优秀人才的军备竞赛。

其中一位杰出人物是穆斯塔法·苏莱曼，他在过去18个月里一直担任微软人工智能主管。苏莱曼因联合创办DeepMind公司而声名鹊起——该公司于2014年被谷歌收购——该公司后来开发出的AI系统击败了围棋世界冠军。

在微软，苏莱曼此前在科技领域开拓创新的能力一直受到与OpenAI协议条款的限制，但一项修订后的协议使他能够公开提出新的目标。我们进行了远程通话，当时西雅图时间很早（他的团队以为他那天早上会在东海岸）。尽管如此，苏莱曼还是立刻投入到讨论中——有时热情洋溢，但也很务实，并且流露出如今大型科技公司中鲜少提及的政治观点。

为了篇幅和清晰度，本次对话内容经过编辑。您可以在最新一期的《米沙尔·侯赛因秀》播客节目中收听完整版。

在你的生活中，人工智能有哪些我们其他人可能还没有想到的用途？

昨天我熬夜看电影，看完后，我在Copilot里建了一个表格，里面记录了我所有喜欢的电影，按日期排序。我还会添加一些个人笔记，它还会给我电影海报的链接。这样我就可以一直问自己：还有哪些类似的电影？

你可以让AI完成几乎任何知识密集型工作任务——就像你可以让助手帮你安排生活一样 。你交给AI的任务越冷门、越有创意、越具有挑战性，效果就越好。

1 苏莱曼似乎也是一位爱读书的人；他在西雅图时身后的书架便展现了他的阅读品味。书架上的书籍包括迈克尔·沃尔夫和罗伯特·卡普兰的最新著作，以及《科技政变：如何从硅谷拯救民主》和《加沙：对其殉难的调查核战争：一个情景"

你用过人工智能执行自主任务吗？它帮你订过机票或买过礼物吗？我知道这是Copilot Actions的承诺——只是它在我的地区还不可用，所以我还没机会亲自体验。

我们仍在试验中。它能做到，但并非总是正确。目前它处于“开发模式”，因此尚未正式发布。

它一旦运行起来，简直就是你见过的最神奇的东西。它基本上会在你的浏览器里输入内容，点击按钮，打开新标签页。它还能查看你的浏览记录，并为你提供个性化的购买体验或回复。

它犯过哪些错误并造成了问题？它有没有给错人买过礼物？

[笑]当然，它可能会买错东西，但你可以干预。而且它在采取下一步行动之前总会征求你的许可，所以相当安全。

科技真是个奇妙的东西。它既神奇又令人惊叹，但总感觉还有很长的路要走。就科技而言，距离普及到日常生活还有一段路要走。

在加入微软之前，您是DeepMind的创始人之一，并且拥有自己的公司Inflection。这是否意味着，当您看到这些小插曲时，您仍然充满信心？

我对这些事情非常冷静。我知道它会在未来六个月、十二个月，或者最坏的情况，十八个月内奏效。它已经非常出色了。

到明年这个时候，我是否就能通过自主人工智能代理购买圣诞礼物了？

我几乎可以肯定你会的。可能性非常大。

在过去的几个月里，多亏了你和其他人的努力，“超级智能”这个词已经悄然进入了公众讨论的视野。对你来说，它意味着 什么？

2今年一月，萨姆·奥特曼撰文指出，OpenAI 的目标将超越通用人工智能（AGI，即能够媲美人类能力的人工智能），转而创造超级智能。这个术语最初由哲学家尼克·博斯特罗姆推广，如今已成为硅谷的热门话题。六月，马克·扎克伯格将Meta 的人工智能部门重组为 Meta 超级智能实验室。上个月，苏莱曼发布了微软人工智能的超级智能团队。

如今，工业界所说的超级智能指的是一种人工智能系统，它能够学习任何新任务，并且在所有任务上都表现得比所有人类加起来还要出色。这是一个极高的标准，而且目前也伴随着巨大的风险。我们很难确定如何控制和驾驭一个比我们强大得多的系统。

我更倾向于将超级智能定义为一种人文主义超级智能——它始终站在我们这边，与我们并肩作战，与人类利益保持一致。在我们能够证明它始终安全之前，我们不会继续开发一个有可能失控的系统。这一点大家都应该认同。但我认为这在目前业界还是一种新颖的观点。

你们是不是想通过强调我们将始终以人文主义视角来使用微软，以此来凸显微软的独特性？

这就是我们的立场。微软是一家拥有50年历史的公司，行事非常谨慎，也备受信赖：标普500指数成分股公司中有90%使用我们的电子邮件、操作系统和日常办公解决方案。我们之所以拥有这样的声誉，是因为公司一直以来都行事谨慎。我们将继续保持谨慎，而制定以人为本的超级智能愿景正是 这一 计划的一部分。

3 观察这种做法是否会与商业需求产生冲突将会很有意思——一位业内人士指出，这种方法可能与微软需要证明其在人工智能领域投资合理性的需求相冲突。但关于“正确”人工智能类型的担忧由来已久。

OpenAI 最初由 Altman 和埃隆·马斯克创立，部分原因在于他们担心谷歌在人工智能领域的领导地位不可靠。随后在 2021 年，一些 OpenAI 员工离职创立了 Anthropic，部分原因也是他们对 OpenAI 在安全方面的做法感到担忧。

这对你的竞争对手意味着什么？其中一些竞争对手你和他们关系密切，比如 OpenAI。这是否意味着他们是无法无天的蛮荒之地，而你是道德的捍卫者？

每个人都必须决定自己的立场和行事方式。我不想评判他们现在的行事方式。

我没有看到任何大规模群体伤害的证据。我也没有看到任何迹象表明这些东西正在自我改进或自主运行。

我们都预测，在未来五年——或许十年内——这些能力将会开始出现。这样的系统可以设定自己的目标，可以改进自身的代码，可以自主运行。

我已明确指出，这些能力会增加风险。我们必须谨慎应对，提高透明度，加强审计，加强政府参与，并主动声明我们距离具备这三项能力还有多远。我认为这一点显而易见。

那是不是意味着，在你们确信能够控制超级智能工具之前，不会发布它？

是的，我认为没错。遏制和协调是必要的先决条件——是底线。我认为业内所有人都必须认同这一点。

没有人想造成大规模伤害。即便我们意见不一，但每个人都致力于我们物种的延续——而且，我希望，也致力于所有人的繁荣和福祉。

所以，这就是我们现在试图推动的讨论，并要求业内每个人都扪心自问：他们是否在构建一个以人为本的超级智能？

考虑到事情总会出错，我们究竟能对此抱有多大的信心？我指的不仅仅是人工智能，人类也会犯错。此前， 《卫报》报道称，微软的一些服务可能被用于大规模监控，之后微软不得不停止并禁用以色列国防部使用的部分服务。

《卫报》的报道非常出色，我们对此深表感谢。我们一得知此事，就立即采取了所有必要的措施，并将以色列国防军的相关内容从服务器上移除。他们显然违反了我们的服务条款，目前我们内部正在进行 调查 。

4微软最初表示，没有证据表明其产品被用于伤害他人，也没有证据表明以色列政府违反了其服务条款或人工智能行为准则。如今，该公司在爱尔兰（其欧洲总部所在地）面临数据保护/法律投诉，指控其数据中心仍在托管用于监控巴勒斯坦人的应用程序。

更广泛的一点是，人们很难对控制、制衡和运用有信心。

这很难。这些系统庞大而复杂，风险也很大。我们所能做的就是确保对它们进行审计，并尽快移除违反我们服务条款的行为者。

超级智能的首次应用会是医疗领域吗？

我认为是的。这可能是超级智能最令人兴奋的应用。我们现在拥有能够诊断文献中记载的任何罕见疾病的系统，其诊断效果远胜于人类，成本更低，所需测试更少，准确率也更高。目前我们正在进行独立的同行评审，很快就会开展临床试验。所以，这真的非常非常令人兴奋。

这个方向是你自己推动的吗？你负责微软人工智能部门已经大约18个月了。我知道你的母亲是一名护士，而且与科技行业的高级管理人员不同，你也曾在公共部门工作过。

这对我来说是一个非常重要的领域。我妈妈是名护士，我深信科技的存在是为了服务于我们。

它应该让我们的生活更美好，更舒适。我认为，有一天它还能帮助我们延长寿命。如果我们愿意，它还能让我们选择减少工作时间。它将创造富足的生活。我们必须首先有意识地决定将其 用于 这些方面。

5“富足”（Abundance）既是美国记者埃兹拉·克莱因和德里克·汤普森合著的一本书的书名，也是一个在美国、英国及其他地区逐渐获得支持的松散政治联盟的名称。具体细节有所不同，但其核心思想是，政策制定者应该让生产人们重视的物品——从住房、电力到新药——变得更加容易，而不仅仅关注如何分配有限的资源。

“富足”这个词听起来很诱人，请问它的具体含义是什么？我们经常听到人工智能会取代工作，但您的意思是说，人工智能会越来越多地取代人类的工作，因此人类的工作量就会减少吗？

在未来20到30年内，机器在大多数工作中比人类更有能力是不可避免的——而且这一天可能会来得更快。

作为一个社会，我们必须决定我们的目标是什么。我们必须慎重考虑新机器的引进速度，因为我们必须确保用以平衡机器换代带来的冲击，并建立相应的机制来资助和支持人们度过这场大规模的转型。

你相信全民基本收入的想法吗？人工智能提高经济生产力或许能带来这种可能性？

我一直以来都公开表示过这一点。这是必然的，也是我们非常希望看到的。我们生活在一个物质丰富的世界，只是分配不均而已。

价值并非仅仅体现在物质层面——食物、汽车、实物。它也体现在数字产品中——思想、知识、智能。这其实是个好消息，因为数字产品可以迅速扩散，在全球范围内快速传播。智能家居和聊天机器人是历史上传播速度最快的技术——短短三年内，年用户量就达到了20亿。

未来将出现巨大的竞争压力，各方 都力图降低体验人工智能的成本。我们面临的挑战是如何征税和再分配，才能确保这一转型过程健康顺利。

6 苏莱曼关于竞争力量的观点是否成立，我们拭目以待。迄今为止，互联网经济一直遵循赢家通吃的格局——例如，谷歌在搜索领域占据主导地位。如果人工智能也遵循这种模式，消费者未必能享受到他所描述的那些好处。

请您详细说说这些想法和信仰的来源。您的母亲是一名护士，但您成长的家庭——20世纪80年代和90年代的伦敦——您会如何描述它？

我家是典型的工薪阶层。我爸是出租车司机。我们家过着很普通的生活，没什么特别的。我爸妈不太重视教育。他们一直觉得我应该学门手艺——我妈经常跟我说：“你应该当木匠或者电工，16岁就辍学。”

它源于我们经历过一些人生的坎坷，并渴望在短暂的生命中尽我们所能做到最好。

你父亲从叙利亚来到英国。你16岁那年发生了什么事？我听说你父母离异后，你和弟弟基本上只能靠自己谋生。

没错。我不知道你是从哪里看到的。我和弟弟确实独自生活过几年。

我很想知道这件事教会了你什么，以及你当时是如何应对的。

那个年纪，你既早熟又自信，还无所畏惧。我们拥有所需的一切。我遇到了很棒的老师和导师。我上的是一所非常好的学校。10岁的时候，我为了入学考试拼命学习，那段时间基本上就跟上私立学校的节奏一样。

我很幸运能考入牛津大学，那是一段非常棒的经历。[但是]我感到非常沮丧，渴望改变世界，有所作为。所以我退学了，并参与创办了穆斯林青年热线（Muslim Youth Helpline），这是一个非宗教、不带评判的倾听服务机构，服务对象是9/11事件后面临身份认同危机、与社区、家庭、父母缺乏联系以及遭受欺凌的英国年轻 穆斯林 。

7苏莱曼不愿谈及这段经历，但我总觉得这与他后来的成就息息相关。这段经历或许培养了他的韧性、责任感，也可能增强了他对自己直觉的信心。18岁到26岁之间，他考入了世界顶尖大学之一，之后辍学，并与人共同创立了DeepMind。

穆斯塔法·苏莱曼曾是牛津大学的学生，19岁时辍学。图片来源：穆斯塔法·苏莱曼

9/11事件后，你是否亲身经历过反穆斯林情绪和仇恨？

有点儿，是的。我觉得很多人觉得我们不够“英国”。人们在摸索如何融入自己的文化和宗教身份，而他们往往是第一代移民，不太会说英语，也不熟悉英国社会体制。人们越来越怀疑我们，觉得我们大多是恐怖分子——这种普遍的恐惧和排斥感。大多数时候，人们通过友善和支持的对话就能化解这些问题——电话那头总有人可以倾诉。

现在你身处权力中心——你是少数几个能够决定一项正在改变我们所有人生活的技术的人之一。你对这种权力有多大的意识？

我非常非常清楚这一点。我非常重视这件事。这是一份重大的责任。此时此刻，我们所做的决定可能会产生非常深远的影响 。8

8我也向斯坦福大学计算机科学家、“人工智能教母”李飞飞提出了这个问题，她的回答也类似：“我是将这项技术带给世界的人之一……我所做的每一件事都会产生后果，这是我肩负的责任。”

我研读历史，回顾社交媒体革命——其潜在危害长期以来无人理会——还有吸烟和石油。很明显，这些东西都会造成危害。我认为，我们必须非常非常谨慎地对待它们的部署方式和引入方式。

你会和同行——比如萨姆·奥特曼——讨论这个问题吗？

是的。业内人士都这样。CEO们肯定有一群人。Anthropic的Sam和Dario [Amodei]是联合创始人，我和Demis [Hassabis]也是联合创始人。我们彼此都很了解。总的来说，每个人都真心实意地想要找到正确的道路。竞争也非常激烈 。9

9哈萨比斯和苏莱曼于2010年与肖恩·莱格共同创立了DeepMind。四年后，该公司被谷歌以据称4亿美元的价格收购。2019年，苏莱曼因其管理风格受到诟病而被停职，之后他为此道歉。他被任命为谷歌人工智能产品管理和政策副总裁，并一直担任该职位直至2022年离开谷歌。

达里奥·阿莫迪，Anthropic公司首席执行官兼联合创始人，该公司生产Claude产品。摄影：Michael M. Santiago/Getty Images

OpenAI 首席执行官 Sam Altman 负责 ChatGPT 项目。摄影：Kyle Grillot/彭博社

谷歌 DeepMind 首席执行官 Demis Hassabis 是 Suleyman 的前同事。摄影师：卢多维奇·马林/POOL/法新社/盖蒂图片社

你听说过“兄弟寡头政治”这个词吗？

[笑]

我以前没听过这种说法，但我能明白它的意思。嗯，我想确实如此。这确实非常以男性为中心——尽管OpenAI的前首席技术官米拉·穆拉蒂是这个领域最优秀的人才之一。

据彭博新闻社6月份报道，米拉·穆拉蒂的创业公司Thinking Machine融资近20亿美元，估值达100亿美元。摄影：Jon Kopaloff/Getty Images for WIRED

或者科技圈的那些家伙。我肯定你听说过这种说法。不知道你有没有看过杰西·阿姆斯特朗的《山巅》？

我没有， 10号 。

10这部反乌托邦讽刺剧的主角是以现实生活中的亿万富翁科技领袖为原型创作的。“他们都是非常重要、非常有才华的人物，”阿姆斯特朗在今年早些时候接受《周末访谈》时告诉我，“我只是觉得，当人们试图将自己的自负与道德冲动结合起来，并且在这个例子中还掺杂了难以置信的巨额财富时，会发生什么，这很有意思。”

我推荐这本书。它相当冷静地描绘了这个精英阶层中极具权势的圈子里的生活。

部分挑战在于，我们都花了很多时间在硅谷，那里天空湛蓝，生活非常平静。

我尽量多旅行。我刚从中国回来。走出自己的舒适圈，看到地球另一端的技术发展速度，真是令人震惊。创新的速度，以及一些监管措施的周全考虑，都令人印象深刻 。11

11自 2022 年以来，中国推出了一系列人工智能法规，包括推荐算法规则和对人工智能生成内容进行标注的要求。

杰西·阿姆斯特朗执导的《山巅》由科里·迈克尔·史密斯、史蒂夫·卡瑞尔、拉米·尤素夫和杰森·舒瓦兹曼主演。摄影：麦考尔·波莱/HBO

如果我说“萨姆·奥特曼”，你脑海中第一个浮现的词是什么？

我的天哪。我想这大概是勇敢的表现吧。

他显然正在非常积极地扩张他的数据中心网络。他很可能成为我们这一代最伟大的企业家之一。他确实取得了巨大的成就。他建设数据中心的速度比业内任何人都快，如果他能成功，那将是惊人的成就。

有趣的是，这里居然还有个“如果”，我也能理解为什么——OpenAI投入了巨额资金。这难道是一场赌博吗？难道他们不认为这笔投资一定会获得回报吗？

ChatGPT 是我们近几十年来见过的最优秀的产品之一——这本身就足以说明一切。与此同时，他们已经签署了超过1.5 万亿美元的未来五到十年数据中心建设承诺，但他们的收入距离这个目标还有相当长的路要走。所以他们还有很长的路要走，但他们是一支非常有才华的团队。我完全相信他们能够成功。

你会用哪个词来形容德米斯·哈萨比斯？

他或许是一位伟大的科学家。我认为他是一位杰出的思想家，也是一位博学多才的人。他多次在该领域做出巨大贡献。他确实非常杰出。

你们曾经是同事，现在却成了竞争对手。

是的。但我们一开始是非常要好的朋友，而且我们一起工作了十年。我从他身上学到了很多，非常尊敬他。

你们还说话吗？

我们昨晚发短信了。我祝贺他一周内完成了Nano Banana、Gemini 3和AlphaFold五周年 。12

12谷歌最新推出的AI聊天机器人Gemini 3在多项基准测试中均优于ChatGPT；它的发布促使Sam Altman宣布其竞争对手模型“进入红色警戒状态”。但彭博社专栏作家Parmy Olsen质疑Gemini能否在市场份额方面超越ChatGPT，为什么 ChatGPT 仍然比 Google 更具优势" 她写道：“谷歌一直难以复制所谓的网络效应，而网络效应正是推动在线平台用户数量飙升的关键因素……谷歌押注于自身更智能，而OpenAI则押注于用户更难放弃。”

你试过Gemini 3吗？它引起了很大的轰动。

很好。

你觉得它比 ChatGPT 好吗？

它们有点不同。它确实拥有ChatGPT所不具备的更多专业技能，而且速度非常快。但ChatGPT也很强大，所以我不会断言它完全胜过ChatGPT。

它比Copilot更好吗？

它可以做一些 Copilot 做不到的事情，但 Copilot 也有一些它没有的功能。

Copilot 在视觉方面确实非常出色。它能看到你所看到的一切，并能实时与你对话。你可以在手机或电脑上与 Copilot 分享屏幕，进行交流并获得反馈。我们真切地体会到，拥有这样一位智能助手陪伴在身边，在你遇到任何难题时都能助你一臂之力，这该是多么美好的日常体验。

我只再做一次这样的采访了。埃隆·马斯克，你会如何形容他？

我猜他像推土机一样。他拥有超人的能力，可以随心所欲地扭曲现实，而且战绩斐然。

他总能设法完成看似不可能的事情。他可能有着不同的价值观。

我采访了他。结果他竟然称我为NPC（非玩家角色）。

[笑]

听起来真像埃隆的风格。我挺喜欢他直言不讳的。他说话很坦率。

你曾说过你的政治观点与他和彼得·蒂尔不同。你会认为自己属于政治光谱的左翼吗？ 13

13特斯拉和Palantir的联合创始人可以被粗略地定义为右翼自由意志主义者。两人都曾资助特朗普的总统竞选，并支持JD Vance作为可能的继任者。对马斯克而言，这算是一种转变——他曾自称是中间派，也是奥巴马的支持者——而蒂尔的政治立场虽然不同寻常，却由来已久。蒂尔对政府持怀疑态度，他曾在2009年发表过一篇臭名昭著的文章：“我不再相信自由和民主能够兼容。”

我现在算是中间派了。我一开始绝对是左派。我以前为肯·利文斯通工作过，坦白说，我深受他们很多人的启发，尽管他们也犯了很多错误。但我很自豪地说，我的政治立场偏左。我相信政府在社会中扮演着重要的 角色 。

14利文斯通是一位社会主义者，在英国媒体中被称为“红色肯”，他于2000年成为伦敦首位民选市长。有人将他与纽约市当选市长佐兰·马姆达尼相提并论——两人都曾与各自政党的领导层有过分歧。

在硅谷，这番话可能会引起争议，但我认为监管是必要的，而且它确实让大多数技术变得更好。人们往往忽略了这一点。汽车之所以能正常行驶，是因为我们有驾驶员培训、排放法规、路灯和限速规定。这就是有效监管的意义所在。我们需要更多这样的监管措施。

肯·利文斯通于2000年至2008年担任伦敦市长。摄影：Daniel Leal/法新社/盖蒂图片社

你说这话会不会感觉自己很孤立？因为很明显，特朗普政府并不热衷于监管，而整个行业对此也相当满意。

目前，我们还没有到会造成巨大灾难性后果的地步。业界在推出功能强大的聊天机器人方面做得相当出色，这些机器人的性格经过精心塑造，非常公正客观，并且以事实为依据。事情原本不必发展到今天这个地步。之所以如此，是因为行业领导者们做得相当不错。但这并不意味着未来不会出现问题，我也对此保持高度警惕，但我认为我们目前还不需要紧急监管。

欧洲一些已提出的监管法规目前正在讨论撤销，例如欧盟人工智能法案。我认为这是好事，人们不应该批评这一点。这是正常流程的体现——监管机构听取反馈，观察实际效果。这值得肯定。

在你的书中，以及你大约在同一时期（2023年）为《外交事务》杂志撰写的一篇文章中，你提出了三种不同的治理模式，分别以气候变化、金融稳定和军备控制为模型。你现在听起来乐观多了。是因为你现在所处的环境吗？

不，我仍然在呼吁这些措施。我们应该设立一个人工智能金融稳定理事会，一个气候进程机制，以及一个国际废除核武器运动（ICAN），负责审核我们取得的进展。在拜登政府时期，我曾积极倡导白宫的人工智能原则，并鼓励各公司做出自愿承诺。当时，是我创办的Inflection公司率先行动，但我们也推动了其他公司效仿。我们已经签署了欧洲和英国人工智能安全研究所的信息披露协议。

所以，我并不是说我们不需要它，我的意思是，这些影响是长期的，而且这些治理流程需要很长时间才能建立起来，所以我们现在就必须开始行动。但我们不需要仓促的反应，也不需要恐慌的过度反应。那样会引发另一系列问题。

你们和OpenAI达成了新的协议。你们突然获得了独立进行人工智能研究的自由。请帮我理解一下，未来你们和OpenAI的关系将会如何发展。你们之前是合作伙伴，现在也将成为竞争对手。

对于人们来说，这套体系很复杂，难以理解，但关键在于，直到几周前，微软还被合同禁止独立研发通用人工智能（AGI）或超级智能。

与 OpenAI 的协议是，OpenAI 在 2019 年签署协议后负责构建通用人工智能 (AGI)，而微软则负责构建人工智能基础设施——包括芯片和数据中心。微软将获得已构建模型的使用许可。直到 2032 年，我们仍然拥有 OpenAI 所有成果的使用许可。

但OpenAI决定增加计算能力，并从其他供应商那里购买计算资源——他们现在与软银以及其他许多公司达成了协议，建造的数据中心数量超过了微软原本愿意为他们建造的数量。作为回报，我们有权开发我们自己的人工智能。

显然，这正是我18个月前加入公司的一个重要原因。我们现在正在组建一支超级智能团队，并致力于自主研发人工智能。

那笔交易让你松了一口气吗？如果没有那笔交易，作为微软人工智能主管，你能做多少事？

嗯，我的意思是影响巨大。我们拥有2800亿美元的营收。全球所有主要机构每天都需要使用我们的人工智能功能和工具。过去18个月，我们一直致力于通用人工智能的开发。现在，我们可以着手研发一些技术和方法，这些技术和方法有可能在所有任务中超越人类的表现。所以，这对我们来说是一个转变。

你和OpenAI的关系是否包括不互相挖角员工？我们刚刚又看到了一场人才争夺战的又一例证——你们的一位关键人物跳槽去了苹果。

人员流动非常剧烈。我们刚刚从谷歌DeepMind和OpenAI挖来了一大批人。这是行业的一部分。

当然不存在任何禁止挖角协议——那不合法。人们可以自由选择为谁工作。目前这个行业的竞争非常非常激烈。这一切都在意料之中。

这是否意味着您已准备好——或者已经准备好——投入与Meta公司类似的巨额资金？比如1亿到2亿美元的项目？

我认为目前还没有人能做到这一点。扎克伯格采取了一种特殊的策略，即大量招聘个人，而不是组建团队。我并不认为这是正确的方法。过去一年半以来，我在微软的做法是逐步增加符合公司文化、技能要求且能与团队其他成员良好合作的人才，同时淘汰不符合要求的人，并非常注重细节。我们致力于打造一个团队，而不是一群个体。

所以你不打算接受同等水平的薪资待遇？之前有报道称你愿意接受？

是的，当时就是这么传的。我认为这显然是史无前例的，也许个别情况确实如此，但这绝对不是常态。

马克·扎克伯格——或者你叫他扎克——你希望他重新考虑一下吗？这会让你的生活更艰难吗？

不，完全不是。他可是全力以赴，对吧？他正在建设一个2吉瓦的数据中心，未来两三年可能要花费他几千亿美元。他曾公开告诉特朗普，未来三年他将在数据中心上投入超过6000亿美元。

对微软来说幸运的是，我们拥有 33 吉瓦的计算能力。我们的主要业务是建设数据中心并将其提供给第三方。我们可以将这些计算能力用于自身的训练，也可以用于推理，并将结果出售给第三方。从这个意义上讲，我们的配置非常完善，足以规避这方面的风险。

马克·扎克伯格手持其公司于2025年推出的Meta Ray-Ban Display AI眼镜。摄影：David Paul Morris/彭博社

你能理解为什么人们不仅担心行业投入的资金数额，还担心其中的循环交易吗？你是OpenAI的投资人，他们也从你这里购买服务。似乎每个人都和英伟达有某种联系。美国经济很大程度上依赖于像你这样的公司能否继续保持良好发展势头。

是的，我觉得这很合理。客户和供应商之间经常互相投资，这能很好地刺激我们前进，但还是需要密切关注。我肯定会密切关注，我想其他人也会。

找到合适的平衡点至关重要。我们必须在未来几年内交付成果。每个团队都在建造规模庞大、功能强大的计算机，我们押注的是能够将这些计算机转化为真正的智能。

如果我们这样做，我认为世界将会变得非常非常不同。我们将拥有取之不尽的丰富情报。

你说的是全世界，但实际上这只关乎两个国家——美国和中国。

没错，过去18个月里，一切都更加集中在硅谷了。这很自然。我们在许多其他历史科技发展趋势中都能看到这一点。

与此同时，开源软件发展迅猛，生产成本大幅下降。现在向世界上最好的AI模型之一提问的成本比两年前降低了90%。成本降低后，人人都能使用。

你取得了许多令人瞩目的成就，但还有什么未完成的事情或者让你反思的事情吗？

我真的很想攻克医疗超级智能。我还想在能源效率和电池储能方面做出更多贡献——开发用于可再生能源的新型化合物。我认为人工智能将真正改变能源行业。

实际上，我对Copilot的许多应用案例感到非常自豪。许多人用它来陪伴他人、进行心理治疗、做出艰难的人生抉择。它为我提供了高质量的信息和情感支持，并帮助我保持条理清晰 。15

15Copilot也曾在英国公共部门进行测试，作为政府提高效率计划的一部分，但结果喜忧参半。用户反映节省了时间，但一项研究发现，几乎没有证据表明这些节省转化为实际的生产力提升——在某些情况下，反而增加了额外的工作量。

你说的“帮助你获得情感支持”是什么意思？

一天结束时，当我开车回家时，我会和它进行 10 分钟的对话，讨论一些棘手的事情，或者一些让我感到沮丧的事情。

或许说它是情感支持有点夸张，但它就像和朋友聊天一样——总结哪些方面做得好，哪些方面做得不好。Copilot 现在能记住你说的大部分内容，并会根据你的情况提供个性化的回答，例如，它会提到你上周说过的话，或者某种趋势或模式。

这真的很有帮助。每次谈话后我都感觉神清气爽，就像卸下了一块重担。

所以他既是朋友，又是治疗师，几乎还是家人。

是的，它融合了各种元素。我的意思是，它是个新事物。它是一款人工智能产品。它就像是“副驾驶”。当我们试图用语言来描述一个既与许多其他事物有些相似，又在本质上截然不同的事物时，这始终是一个挑战。

我只是好奇，有没有什么方面会让你停下来思考。人们回家后，或许不必再费心和生活中其他人交谈，因为他们已经把该说的都说了，也得到了他们需要的。

反过来说，他们也不必把气撒在伴侣或最好的朋友身上。

我仍然每个周末都给最好的朋友们打电话，好好聊聊天。事实上，这反而加深了我和朋友们之间的感情。每次聊天后，我都会感觉轻松一些。

你觉得你的未来会在美国吗？因为我相信英国会非常欢迎你回去。

我原本很想在英国创办 Inflection。我热爱英国，热爱伦敦。我的思维方式非常英国化。

我不喜欢的一点就是“枪打出头鸟”的风气。这里缺乏足够的冒险精神。商业化、赚钱、创业和企业家精神都带有一些禁忌色彩。人们对尝试和失败的鼓励也不够。

在硅谷，每个人都有点疯狂，每个人都喜欢失败。他们总是谈论事情哪里出了问题——比如这简直是一场灾难——这让人感觉很自由。当然，这种氛围也很俗气，对于有点愤世嫉俗的英国人来说可能会有点刺耳，但当你适应了这种节奏，就会觉得很棒。

这里还有一种拼搏的文化。看到每个人都在努力奋斗，试图理解一些新技术，真是令人敬佩。你会看到一些素不相识的人在星巴克里阅读科学论文。你还会听到两个互不相识的人在交谈，几乎像是在建立人脉。这里充满活力。我希望伦敦也能如此——而且我认为它完全可以实现，但我们需要一些鼓励冒险和商业发展的政治领导人。

你认为人工智能能和你进行这样的对话吗？它能比我做得更好吗？

今天可能不行。如果你事先和人工智能一起准备，它或许能给你一些问题——虽然你的确做了很多详细的研究，所以你的团队显然非常出色。

人工智能记者即将出现。我在微软负责MSN，它是全球最大的新闻网站之一。我非常兴奋的一点是，人工智能记者将如何重振地方新闻。想象一下，成千上万的人工智能记者可以打电话给现场人员，核实目击者拍摄的视频，进行采访，并将这些素材剪辑成短片，而且不仅仅用于那些值得投入资源的大型全国性新闻，而是应用于非常本地化的新闻报道——提供准确可靠的信息。

你们正在为 MSN 开发人工智能记者吗？

我们正在探索各种可能性。我也是《经济学人》的董事会成员，也和他们就此进行了很多讨论。我认为我们目前正在探索各种各样的可能性。

人工智能面试官？

你还有一些时间，大概六个月吧。

[笑]

我在开玩笑。

我不确定你是否真的如此。

不，我开玩笑的。要做到完美还需要很长时间。这次已经非常出色了。

米沙尔·侯赛因是彭博周末版的特约编辑。

Microsoft’s Mustafa Suleyman: ‘AI Is Already Superhuman’

The tech giant’s AI chief talks about superintelligence “red lines,” why AI will transform medicine and how he unwinds by chatting with Copilot.

By Mishal Husain

December 12, 2025 at 1:10 AM EST

The AI race is entering an uncharted and expensive new phase. This has been the year of the mega-deal — billions poured into data centers, investments between the key companies, and a talent arms race for the best and brightest minds.

One of those minds is Mustafa Suleyman, who for the past 18 months has been AI chief at Microsoft. Suleyman made his name co-founding DeepMind — acquired by Google in 2014 — which later produced the AI system that defeated a world champion at Go.

At Microsoft, his ability to break new ground in the field was until recently limited by the terms of a deal with OpenAI, but a revised agreement is now enabling Suleyman to go public with new goals. We spoke remotely, at what turned out to be a very early hour in Seattle (his team had thought he’d be on the east coast that morning). Nevertheless, Suleyman launched right into the discussion — evangelical at times, but also realistic, and with hints of a political perspective rarely voiced in Big Tech these days.

Listen to and follow The Mishal Husain Show on iHeart Podcasts, Apple Podcasts, Spotify or wherever you get your podcasts.

This conversation has been edited for length and clarity. You can listen to an extended version in the latest episode of The Mishal Husain Show podcast.

What uses of AI are in your life that the rest of us might not yet have?

Yesterday, I stayed up far too late watching a film and afterwards, I added to a table that I’ve made in Copilot, which basically records all the films I love, lists them by date. I add my personal notes, it gives me a link to the film poster. I can keep just saying, What would be a similar one?

It’s possible to ask your AI to do pretty much any knowledge work task — just like you might ask an assistant to organize your life. The more obscure, creative [and] challenging the task you’re going to ask your AI, the better. 1

1 Suleyman also appears to be a keen reader; the bookshelf behind him in Seattle offered a glimpse of his tastes. Titles included the most recent books by Michael Wolff and Robert Kaplan, as well as The Tech Coup: How to Save Democracy from Silicon Valley, ; --tw-ring-color: rgb(59 130 246 / .5); --tw-ring-offset-shadow: 0 0 ; --tw-ring-shadow: 0 0 ; --tw-shadow: 0 0 ; --tw-shadow-colored: 0 0 ; --tw-blur: ; --tw-brightness: ; --tw-contrast: ; --tw-grayscale: ; --tw-hue-rotate: ; --tw-invert: ; --tw-saturate: ; --tw-sepia: ; --tw-drop-shadow: ; --tw-backdrop-blur: ; --tw-backdrop-brightness: ; --tw-backdrop-contrast: ; --tw-backdrop-grayscale: ; --tw-backdrop-hue-rotate: ; --tw-backdrop-invert: ; --tw-backdrop-opacity: ; --tw-backdrop-saturate: ; --tw-backdrop-sepia: ; --tw-contain-size: ; --tw-contain-layout: ; --tw-contain-paint: ; --tw-contain-style: ; box-sizing: border-box; border-width: 0px 0px 1px; border-style: solid; border-color: rgb(229, 229, 229) rgb(229, 229, 229) rgb(97, 122, 83); outline-color: rgba(10, 10, 10, 0.5); color: inherit; text-decoration: inherit; --annotation-color: rgb(97,122,83); --lede-background-color: rgb(169,178,150); --serif-head: 'BWHaasDingbat','PublicoHead',Georgia,Cambria,'Times New Roman',Times,serif; --serif-text: 'BWHaasDingbat','PublicoText',Georgia,Cambria,'Times New Roman',Times,serif; --san-serif-web: 'BWHaasDingbat','BWHaasGroteskWeb',Helvetica,Arial,sans-serif;">and Gaza: An Inquest Into Its Martyrdom.

Have you used AI for autonomous tasks? Has it booked tickets or bought a gift for you? I know this is the promise of Copilot Actions — it’s just not available in my region, so I haven’t been able to try it myself.

We’re still experimenting. It can do it. It doesn’t always get it right. It’s in ‘dev mode,’ so not generally available just yet.

When it does work, it is the most magical thing you’ve ever seen. It essentially types stuff into your browser, clicks on buttons, opens up new tabs. It can look at your history, [and] personalize the purchase or the response to you.

What are the mistakes it’s made that create problems? Has it bought a present for the wrong person?

[Laughs]

Well, it can buy the wrong thing, but you can intervene. And it will always ask you permission before it takes the next action, so it’s quite safe.

It’s a funny thing, technology. It’s magical and amazing, but it’s always just got a little bit further to go. In this case, a while yet before it’s everyday.

You were a founder of DeepMind and had your own company, Inflection, before you came to Microsoft. Does that mean that when you see these hiccups, you have faith?

I’m very stoic about these things. I know that it’s going to work in the next six months or 12 months or, maybe worst case, 18 months. It is already superhuman.

By this time next year, could I be buying my Christmas presents using an autonomous AI agent?

I’m pretty certain that you will be. It’s highly likely.

The term superintelligence has crept into the public debate, thanks to you and others, in the last few months. What does it mean to you? 2

2 In January, Sam Altman wrote that OpenAI would aim beyond artificial general intelligence — AGI, or AI that would match human capabilities — to creating superintelligence. The term, first popularized by philosopher Nick Bostrom, now dominates discourse in Silicon Valley. In June, Mark Zuckerberg restructured Meta’s AI division as Meta Superintelligence Labs. Suleyman unveiled Microsoft AI’s Superintelligence Team last month.

Superintelligence in the industry today means an AI system that can learn any new task and perform better than all humans combined, at all tasks. It is a very high bar and, at the moment, it comes with a great deal of risk. It’s very uncertain how we would contain and align a system that is so much more powerful than us.

The framing I prefer is one of a humanist superintelligence — one that is always in our corner, on our team, aligned to human interests. Until we can prove that it will remain safe, we won’t continue to develop a system that has the potential to run away from us. Everybody should agree to that. Yet I think it’s a novel position in the industry at the moment.

Is that how you’re trying to set Microsoft apart, by saying we will always use it through a humanist lens?

That is our position. Microsoft is a company that’s been around for 50 years. It is very careful. It’s highly trusted: 90% of the S&P 500 use us to provide email, operating systems and everyday productivity. We’ve got that reputation because the company’s been careful. We’re going to continue to be careful, and setting out a vision of humanist superintelligence is part of that program. 3

3 It will be interesting to see if this comes up against commercial imperatives — one industry watcher has suggested the approach may clash with Microsoft’s need to justify investment in AI. But angst about the “right” kind of AI is long-standing. OpenAI was first launched by Altman and Elon Musk partly out of concern that Google couldn’t be trusted to lead AI. Then in 2021 some OpenAI employees left to start Anthropic, in part because of concerns over the former’s approach to safety.

What does that mean for your rivals, some of whom you work closely with, like OpenAI? Does it mean that they’re the Wild West and you are the moral ones?

Everybody has to decide what they stand for and how they operate. I don’t want to judge how they’re operating right now.

I don’t see any evidence of large-scale mass harm. I don’t see any indication that these things are improving themselves, or operating autonomously.

We all predict a time in the next five years — maybe 10 years — where these capabilities do start to emerge. Systems like this could set their own goals. They could improve their own code. They could act autonomously.

Those are capabilities that I’ve clearly outlined as increasing the level of risk. We have to approach them with caution, with more transparency and audits, with more government engagement, and make proactive declarations about how close we are to those three capabilities. I think that’s obvious.

Does that mean you won’t be releasing a superintelligence tool until you are confident it can be controlled?

Yeah, I think that’s right. Containment and alignment are necessary prerequisites — red lines. I think everybody in the industry has to sign up to that idea.

Nobody wants to cause mass harm. Even though we all disagree, everybody is committed to the survival of our species — and, I would hope, the flourishing and wellbeing of everybody.

So that’s the discussion we are trying to push now, and require everybody to ask themselves in the industry: Are they building a humanist superintelligence?

How much confidence can we really have in that, given things do go wrong? I don’t just mean AI. Humans do wrong things. Microsoft had to cease and disable some services used by the Israeli Ministry of Defense, after reporting by The Guardian suggested they could be being used for mass surveillance.

That was very good reporting by The Guardian. We were very grateful for it. As soon as we became aware of it, we made all of the necessary changes [and] removed the IDF from those servers. They were clearly not in compliance with our terms of service and there’s an ongoing investigation internally. 4

4 Microsoft initially said it had no evidence that its products had been used to harm people or that the Israeli government had failed to comply with its terms of service or AI Code of Conduct. The company is now subject to a data protection/legal complaint in Ireland (where it has its European headquarters), alleging that its data centers continue to host applications used to monitor Palestinians.

The broader point is that it’s hard to have confidence in controls, checks and balances and uses.

It is hard. These are huge and complicated systems that carry a lot of risk. The most we can do is make sure that we are auditing them and removing actors that violate our terms of service as quickly as possible.

Are the first uses of superintelligence going to be in the medical field?

I think so. This is probably the most exciting application of superintelligence. We now have systems that can diagnose any rare condition found in the literature, significantly better than human performance, more cheaply, with fewer tests and with higher accuracy. We are putting it through independent peer review at the moment and soon there’ll be clinical trials. So this is very, very, very exciting.

Get the Bloomberg Weekend newsletter.

Big ideas and open questions in the fascinating places where finance, life and culture meet.

Sign Up

By continuing, I agree to the Privacy Policy and Terms of Service.

Did you push for that focus yourself? You’ve been running Microsoft AI for about 18 months now. I’m conscious that your mother was a nurse and, unusually for senior people in tech, you’ve worked in the public sector as well.

This is an area that’s very important to me. My mum was a nurse and I’m just a big believer that technology is here to serve us.

It should make our lives better, make us more comfortable. One day, I think it is going to help us to live longer. It’s going to give us the option to work less if we choose to. It’s going to produce abundance. We have to make conscious decisions to use it for those applications first. 5

5 “Abundance” is both the title of a book by US journalists Ezra Klein and Derek Thompson and the name for a loose political coalition that has gained traction in the US, the UK and beyond. The details vary, but the gist is that policymakers should make it easier to produce the stuff people value — from housing to electricity to new drugs — and not just focus on divvying up shares of a fixed pie.

Abundance is such a big promise. Tell me what you mean by it. We hear about AI destroying jobs, but are you saying that the work will be done by AI increasingly, and therefore humans won’t have to work as much?

It is inevitable that at some point over the next 20 or 30 years, machines are going to be more capable than humans at doing most work — that might come much sooner.

We have to decide as a society what our purpose is. We have to be very thoughtful about the rate of introduction of new machines, because we have to make sure that displacement is counterbalanced with a mechanism to fund people and to support people through a massive transition.

Do you believe in the idea of a universal basic income? That’s what AI making the economy more productive could unlock?

I’ve long been on record saying that. That is inevitable and very desirable. We already live in a world of abundance, it’s just poorly distributed.

Value isn’t just manifested in atoms — food, cars, physical things. It’s manifested in digital goods — ideas, knowledge, intelligence. That’s actually great news because that can proliferate; it can spread extremely quickly around the entire world. LLMs and chatbots have been the fastest spreading technology in history — basically 2 billion annual users in the space of three years.

There’s going to be massive competitive forces to reduce the cost of experiencing an AI. The challenge we’re going to have to figure out is how we tax and redistribute, so that the transition is a healthy one. 6

6 We’ll have to wait and see if Suleyman’s view on competitive forces is borne out. So far, the internet economy has been defined by a winner-take-all dynamic — for example, search being dominated by Google. If AI follows that pattern, customers won’t necessarily see the benefits he’s describing.

Tell me more about where these ideas and your beliefs come from. Your mother was a nurse, but the family you grew up in — in London in the 1980s and ’90s — how would you describe it?

Pretty working class. My dad was a cab driver. We were fairly regular, kind of unremarkable. My parents didn’t super value education. They always thought I should go get a trade — my mum would often say to me, You should be a carpenter or electrician, leave school at 16.

It comes from a place of experiencing the rougher end of things a little bit and having a desire to try to do the best we can with the short life that we have.

Your dad came to the UK from Syria. What happened when you were 16? I read that when your parents split up, you and your younger brother were pretty much left to fend for yourselves.

That’s true. I don’t know where you read that. Me and my younger brother did live on our own for a few years.

I’m curious to know what it taught you and how you dealt with it at the time.

When you’re that age, you are precocious and overconfident and fearless. We had everything that we needed. I had great teachers and mentors. I went to a very good school. I studied really hard when I was 10 for the entrance exams and it was essentially like going to a private school.

I was lucky enough to get into Oxford and that was an amazing experience. [But] I was very frustrated and eager to change the world and get stuff done. So I dropped out and helped start Muslim Youth Helpline, a non-religious, non-judgmental listening service for young British Muslims, who after 9/11 were dealing with identity crisis, lack of connection to community, family, parents, bullying. 7

7 Suleyman was reluctant to speak about this period in his life, but I can’t help feeling there’s a link to his subsequent achievements. It must have taught him resilience and responsibility and perhaps also confidence in his own instincts. Between the ages of 18 and 26 he got into one of the world’s best universities, dropped out, and then co-founded DeepMind.

Mustafa Suleyman as a student at Oxford University, before he dropped out at 19. Source: Mustafa Suleyman

Had you experienced anti-Muslim sentiment and hatred yourself after 9/11?

A little bit, yeah. I think a lot of people felt we weren't “British” enough. People were figuring out how to live their cultural [and] religious identities, in the context of families that often were first-generation, didn’t really speak the language, didn’t know how to navigate the system. The increased skepticism, that we were mostly terrorists — this general fear and exclusion. Most of the time that was dealt with by having kind and supportive conversation — somebody available on the other end of a telephone.

Now you are in this circle of power — one of a small group of people making decisions on a technology that is changing all our lives. How conscious are you of that power?

Very, very conscious. I take it very seriously. It is a great responsibility. This is a moment when decisions we make may have very lasting consequences. 8

8 This is a question I also asked Stanford computer scientist and “Godmother of AI” Fei-Fei Li, who answered in similar terms: “I’m one of the people who brought this technology to the world… everything I do has a consequence and that’s a responsibility I shoulder.”

I read history and look back [at] the social media revolution — where potential harms fell on deaf ears for too long — or smoking or oil. It’s very clear these things will cause harm. I think that we have to be very, very careful about how we deploy them, and how they’re introduced into the world.

Do you talk about this with your peers — people like Sam Altman?

Yes. Everybody in the industry does. There’s definitely a group of the CEOs. Sam and Dario [Amodei] from Anthropic were co-founders, me and Demis [Hassabis] were co-founders. We all know each other very well. On the whole, everybody is genuinely committed to trying to find the right path through. It’s also very competitive. 9

9 Hassabis and Suleyman founded DeepMind alongside Shane Legg in 2010. Four years later, it was acquired by Google for a reported $400 million. Suleyman was placed on leave in 2019, amid complaints over his management style, for which he has since apologized. He was appointed a VP of AI product management and policy at Google, a role he held until leaving the parent company altogether in 2022.

Dario Amodei, CEO and co-founder of Anthropic, maker of Claude. Photographer: Michael M. Santiago/Getty Images

OpenAI CEO Sam Altman oversees ChatGPT. Photographer: Kyle Grillot/Bloomberg

Google DeepMind CEO Demis Hassabis was a former colleague of Suleyman’s. Photographer: Ludovic Marin/POOL/AFP/Getty Images

Do you recognize the term broligarchy?

[Laughs]

I haven’t heard that before, but I can figure out what it means. Yeah, I guess that’s true. It is very male-centric — although Mira Murati, the ex-CTO of OpenAI, is one of the best people in the field.

Mira Murati’s startup Thinking Machine raised close to $2 billion at a $10 billion valuation, Bloomberg News reported in June. Photographer: Jon Kopaloff/Getty Images for WIRED

Or tech bros. I’m sure you’ve heard that one. I don’t know if you’ve seen Jesse Armstrong’s Mountainhead?

I haven’t, no. 10

10 The central characters in this dystopian satire are composites of real-life billionaire tech leaders. “They’re very important figures and very talented figures,” Armstrong told me in a Weekend Interview earlier this year. “I just think it’s interesting what happens to people as they try to marry their egos with their moral impulses, and in this case with an unbelievably large amount of money.”

I recommend it. It’s quite a sobering portrait of life in this rarefied, very powerful circle.

Part of the challenge is that we all spend a lot of time in Silicon Valley, where the skies are blue and life is very peaceful.

I try to travel a lot. I just came back from China. It’s staggering to get out of the bubble and see how this technology is being developed on the other side of the world. The pace of innovation, but also the thoughtfulness of some of the regulatory stuff. It’s impressive. 11

11 Since 2022, China has rolled out a series of AI regulations, including rules for recommendation algorithms and requiring labels for AI-generated content.

Jesse Armstrong’s Mountainhead stars Cory Michael Smith, Steve Carell, Ramy Youssef and Jason Schwartzman. Photographer: Macall Polay/HBO

If I say to you, Sam Altman, what word first comes into your mind?

Oh my God. I guess courageous.

He’s obviously growing his data center fleet very aggressively. He may well turn out to be one of the great entrepreneurs of our generation. He’s certainly achieved a lot. He’s building data centers at a faster rate than anyone in the industry, and if he can pull it off, it will be pretty dramatic.

It’s interesting that there is an if on that, and I understand why — there are huge amounts of money being spent by OpenAI. Is it a gamble? Is it not a given that it’s going to pay off for them?

ChatGPT is one of the greatest products we’ve seen in a generation — that speaks for itself. At the same time, they’ve signed over $1.5 trillion of commitments for building data centers over the next five or 10 years, and their revenues are quite a long way from there. So they’ve got a long way to go, but they’re a very talented team. I’ve got every confidence they can do it.

Big Tech Spending Expected To Keep Climbing

Companies are boosting capex to fund artificial intelligence

Note: 2025, 2026 figures include estimates

Source: Bloomberg

What word would you use to describe Demis Hassabis?

Probably a great scientist. I think he’s a great thinker and he’s a good polymath. He’s made massive contributions in the field, multiple times. He’s truly exceptional.

You worked together, and now you are competitors.

Yeah. But we started off as very close friends, and we worked together every day for 10 years. I learnt a lot from him and have huge respect for him.

Do you still talk?

We texted last night, actually. I congratulated him on Nano Banana, Gemini 3 and five years of AlphaFold all in one week. 12

12 Gemini 3, the latest version of Google’s AI chatbot, outperforms ChatGPT on many benchmarks; its release prompted Sam Altman to declare a “code red” for his rival model. But Bloomberg Opinion’s Parmy Olsen has questioned Gemini’s ability to overtake ChatGPT in terms of market share, ; --tw-ring-color: rgb(59 130 246 / .5); --tw-ring-offset-shadow: 0 0 ; --tw-ring-shadow: 0 0 ; --tw-shadow: 0 0 ; --tw-shadow-colored: 0 0 ; --tw-blur: ; --tw-brightness: ; --tw-contrast: ; --tw-grayscale: ; --tw-hue-rotate: ; --tw-invert: ; --tw-saturate: ; --tw-sepia: ; --tw-drop-shadow: ; --tw-backdrop-blur: ; --tw-backdrop-brightness: ; --tw-backdrop-contrast: ; --tw-backdrop-grayscale: ; --tw-backdrop-hue-rotate: ; --tw-backdrop-invert: ; --tw-backdrop-opacity: ; --tw-backdrop-saturate: ; --tw-backdrop-sepia: ; --tw-contain-size: ; --tw-contain-layout: ; --tw-contain-paint: ; --tw-contain-style: ; box-sizing: border-box; border-width: 0px 0px 1px; border-style: solid; border-color: rgb(229, 229, 229) rgb(229, 229, 229) rgb(97, 122, 83); outline-color: rgba(10, 10, 10, 0.5); color: inherit; text-decoration: inherit; --annotation-color: rgb(97,122,83); --lede-background-color: rgb(169,178,150); --serif-head: 'BWHaasDingbat','PublicoHead',Georgia,Cambria,'Times New Roman',Times,serif; --serif-text: 'BWHaasDingbat','PublicoText',Georgia,Cambria,'Times New Roman',Times,serif; --san-serif-web: 'BWHaasDingbat','BWHaasGroteskWeb',Helvetica,Arial,sans-serif;">writing: “Google has often struggled to replicate the so-called network effects that propel an online platform to stratospheric user numbers… While Google is betting on being smarter, OpenAI is gambling on being harder to quit.”

Have you tried out Gemini 3? It’s creating quite a lot of waves.

It’s good.

Do you think it’s better than ChatGPT?

They’re kind of different. It’s definitely got more niche skills that ChatGPT doesn’t have, and it’s very fast. But ChatGPT is very strong, so I wouldn’t go that far.

Is it better than Copilot?

It can do things that Copilot can’t do, but Copilot also has features that it doesn’t have.

Copilot is actually amazing for vision. It can see everything that you are seeing and talk to you in real time. You can share your screen with Copilot on mobile or desktop, talk about it and get feedback. We’re really trying to imagine the day-to-day experience of having this really intelligent assistant at your side, that can help unblock you whenever you get stuck.

I’m only going to do one more of these. Elon Musk, how would you describe him?

I guess a bulldozer. He’s kind of got superhuman capabilities to bend reality to his will and has [a] pretty incredible track record.

Somehow he mostly manages to pull off what appears to be impossible. [He] probably [has a] different set of values.

I interviewed him. He ended up calling me an NPC.

[Laughs]

That sounds exactly like Elon. I kind of like that he speaks his mind. He’s very unfiltered.

You’ve said your politics are different from his, and Peter Thiel’s. Would you say you’re on the left of the political spectrum? 13

13 The co-founders of Tesla and Palantir can be loosely defined as right-wing libertarians. Both have funded Trump in presidential campaigns and backed JD Vance as a possible successor. For Musk, that was something of a shift — he was once a self-described centrist and Obama supporter — whereas Thiel’s politics, though unusual, are long-standing. Skeptical of government, Thiel notoriously wrote in 2009: “I no longer believe that freedom and democracy are compatible.”

I’m sort of a centrist these days. I definitely started as a lefty. I worked for Ken Livingstone back in the day, and was frankly, very inspired by a lot of those people, even though they also made a lot of mistakes. But I’m proud to say that I’m on the center-left of the spectrum. I believe that government plays an important role in society. 14

14 Livingstone, a socialist dubbed “Red Ken” in the British press, became London’s first elected mayor in 2000. Some have compared him to New York’s mayor-elect Zohran Mamdani — both figures who have been at odds with their parties’ leadership at times.

It’s a controversial thing to say in Silicon Valley, but I think regulation is necessary and it has made most technologies better. People forget this. Cars only work because we have driver training, emissions regulations, streetlights and speed limits. That’s what regulation is when it works well. We just need more of that.

Ken Livingstone served as mayor of London from 2000 to 2008. Photographer: Daniel Leal/AFP/Getty Images

Do you feel isolated saying that, because clearly the Trump administration is not into regulation, and the industry overall is quite happy with that.

Right now, we are not in a mode where there is huge catastrophic harm. The industry’s done a pretty good job of introducing very powerful chatbots where the personality is sculpted to be very even-handed, very evidence-based. It didn’t have to go like that. It went like that because the leaders have done a pretty decent job. That doesn’t mean that there aren’t going to be issues coming up, and I’m very wide-eyed about them, but I don’t think we’re in a mode where we need emergency regulation.

Some of the regulations that have been proposed in Europe are currently in talks to be wound back, with the EU AI Act. I think that’s good. People shouldn’t criticize that. That’s the process working — regulator taking feedback, seeing how things work in practice. That should be celebrated.

In your book, and a piece you wrote about the same time [2023] for Foreign Affairs, you wanted three different kinds of governance regimes, modeled on climate change, financial stability and arms control. You sound much more sanguine now. Is it because of where you are?

No, I’m still calling for those things. We should have a financial stability board for AI. We should have a climate process, [and] an ICAN [International Campaign to Abolish Nuclear Weapons], which audits the progress that we are making. I advocated for the White House AI principles under the Biden administration, where we made voluntary commitments as companies. At the time, it was my startup, Inflection, but we pushed all the other companies to do it as well. We’ve signed up to disclosures in Europe and with the UK AI Safety Institute.

So I don’t mean to say that we don’t need it, I’m saying that these are such long-term effects, and these governance processes take so long to build, that [we’ve] got to start now. But we don’t need [a] knee-jerk reaction. We don’t need some panicked overreaction. That would cause a different set of problems.

You’ve got a revised agreement with OpenAI. You suddenly have the freedom to pursue AI independently. Help me understand how that relationship with OpenAI is going to work...