翻译｜林绮蓓
编辑｜蔡芳芳
谷歌有多少代码是 AI 写的？去年，CEO 桑达尔·皮查伊（Sundar Pichai）还给出过一个明确的数字：25%。但当主持人再次问起这个问题时，他的回答却改变了：“这个比例已经没有太大意义。”
因为 AI 已经深度参与谷歌内部的代码开发。Gemini 4 Argon 发布后的四周内，谷歌代码变更提交量增加约 30%，回滚次数减少约三分之一，开发人员开始用 Gemini 4 Argon 处理涉及数百万行代码的复杂项目。相比 AI 编写代码的比例，谷歌现在更关注代码变更量、回滚情况等指标。在皮查伊看来，这些变化更能反映 AI 对开发效率与质量的影响。
近日，皮查伊做客 Big Technology Podcast，与主持人 Alex Kantrowitz 谈起谷歌在 AI 竞争中的投入与取舍。刚刚上线的 Gemini 4 Argon，前 72 小时的使用量就接近 25 万亿 Token，约为 Gemini 3.0 同期的三倍，算力供给却跟不上需求。在外界一度认为谷歌可能放弃 AI 前沿竞争、专注于将能力够用的模型融入现有产品时，皮查伊给出的回应十分明确：“这个判断错得离谱。”皮查伊透露，为了保障前沿模型研发，谷歌已将算力优先分配给 AI 研究团队，云业务也因此放弃了部分短期收入。
除了模型研发与算力争夺，他还谈到了 AI 参与模型开发、持续运行的个人智能体、世界模型，以及谷歌如何应对来自 OpenAI 的竞争。以下是本次访谈的整理内容。
太长不看版：
Alex： 外界曾认为，谷歌不再追求最前沿的 AI，模型够用就好。但 Argon 的推出说明，事实并非如此？
Pichai：这个判断错得离谱。追求技术前沿一直是谷歌的核心目标。Argon 上线后的前 72 小时，使用量接近 25 万亿 Token，约为 Gemini 3.0 同期的三倍。我们对它的表现很满意，也看到了需要改进的地方，会继续提升它的能力。
Alex：有人认为，普通模型就足以做好地图、Gmail 和搜索，只有生物学研究等特定任务才需要前沿模型。谷歌为什么还要追求最前沿的能力？
Pichai： 我们希望持续推动技术进步。部分云客户需要前沿模型，更强的模型也能帮助开发下一代模型。Argon 在内部已经解决了一些过去模型处理不了的问题。当然，不同规模、不同能力的模型都有适用场景，我们也会继续投入。
Alex： 你说，站在技术前沿能让你更早看到未来。那么，你已经看到了哪些即将到来的变化？
Pichai： 我们一直在研究让智能体持续执行长时间任务，接下来还会走向多个智能体相互协作。未来，个人用户和企业用户都可能拥有持续为自己工作的多智能体系统。
Alex： 模型还不能完全按照人的意愿行事，我们却在追求递归式自我改进，让它们更快地提升能力。我感到担忧，是不是应该采取措施，让科技行业在这件事上冷静一些？
Pichai： 担忧是合理的，但首先要分清当前的发展阶段。目前，是智能体协助研究人员开发模型，仍然有人类参与，距离完全自主的递归式自我改进还很远。安全也不能只依靠对齐，还需要实时监测、沙箱隔离、异常干预和客户控制权。下一步，行业需要更完善的标准、独立评估，以及监管的参与。
Alex： 有些机构把大语言模型视为通向 AGI 的主要道路，而谷歌还在同时推进世界模型研究。为什么你认为这条路线也很重要？
Pichai： 理解物理规律、准确模拟现实世界，是发展智能的重要基础。Argon 在具身推理上的表现，也得益于世界模型研究积累的经验。未来，这些能力可能通过新的架构融合起来，我们希望在不同的技术路径上都保持领先。
Alex： 云业务希望通过对外出售算力获得利润，AI 研究则需要大规模计算资源来开发最好的模型。你如何在这两方面分配算力？
Pichai： 首先保障前沿 AI 研究和模型开发，再根据各项业务的长期规划、需求预测和合同安排算力。前沿研发所需的资源必须得到保障，这是我们的首要原则。
Alex： 如果你真的相信 AI 的影响会像火和电一样深远，那么，即使牺牲云业务等业务的眼前利润，是不是也应该给研究人员提供尽可能大规模的计算资源？
Pichai： 我们已经在不同程度上这样做了，云业务也因此放弃了部分短期收入。不过，适当的约束也能激发创造力，推动算法效率的突破。我们的原则，是确保投入足够的资源，保持前沿地位。
Alex： Greg Brockman 认为，足够聪明的 AI 应该让人直接通过交谈表达需求，减少对屏幕和按钮的依赖。你怎么看 AI 时代的交互界面？
Pichai： 技术应该适应人的生活习惯，让交互尽可能自然顺畅。但界面仍然有作用，比如观看视频、欣赏 AI 创作的内容。关键是减少用户为了使用技术而付出的精力。
Alex： 谷歌有类似 Muse、Dots 的产品吗，或者是否准备开发一款与它们竞争的新产品？
Pichai： 我们对 Gemini 的目标，就是把它打造成了解个人需求、能够主动提供帮助的强大助手。用户只需要提出需求，不必先区分是在聊天还是交办任务，它会采取合适的方式，在后台持续执行，并在不同产品中提供连贯的体验。
Alex： 你之前说过，谷歌内部有 25% 的代码由 AI 编写。现在这个比例是多少？
Pichai： 这个比例已经不太有意义，因为 AI 深度参与了开发工作。我们更关注效率和质量：Argon 发布后的四周内，代码变更提交量增加约 30%，回滚减少约三分之一，开发人员也开始用它处理涉及数百万行代码的复杂项目。
Alex： 有没有哪个尚未取得突破的 AI 研究方向，是谷歌正在重点攻克的？比如，让模型发布之后仍能持续学习和更新？
Pichai： 我们正在并行探索合成数据、持续学习、更高效的架构和世界模型等方向，并投入相当一部分算力。其中一些研究成果会直接应用到主要模型中。
谷歌放弃 AI 前沿竞争？Pichai：这个判断错得很离谱
Alex：今天，我们邀请到 Alphabet 首席执行官 Sundar Pichai，聊聊谷歌如何参与 AI 前沿竞争，如何平衡前沿 AI 研究与云业务的资源投入，以及未来的交互界面会是什么样子。Sundar 今天来到了我们位于山景城的演播室。Sundar，很高兴见到你，欢迎来到节目。
Pichai：谢谢你，Alex，很高兴来到这里。
Alex：随着 Gemini 4 Argon 的推出，谷歌重新回到了 AI 前沿。用直白的话说，这是一个处于世界领先水平的模型，刷新了多项基准测试成绩，在编程、数学、科学和知识工作等多项测试中，超过了 OpenAI 和 Anthropic 已经公开的模型。此前，外界普遍认为，谷歌基本上已经放弃了前沿竞争，打算开发一些速度快、能力够用的模型，再把它们应用到自家的产品中，不再追求最前沿的 AI。但事实证明，这种判断是错误的。
Pichai：而且错得很离谱。我们始终非常坚定地要站在技术前沿，追求技术前沿可以说是我们骨子里的东西，我们一直都是这样做的。十多年来，我们一直在思考 AI 前沿的发展。这个领域变化很快，竞争非常激烈，也令人兴奋。
我对 Argon 的最终表现非常满意。在开发过程中，我们没有急于求成，而是确保充分开展安全测试，谨慎地部署。考虑到它的能力，我们在它身上花的时间比其他模型更长一些。但整体来看，无论是智能体任务、知识工作、编程，还是创意任务，我认为 Argon 都是非常出色的前沿模型之一。
上线后的前 72 小时，它的使用量就已经接近 25 万亿 Token，差不多是 Gemini 3.0 上线后前 72 小时的三倍。目前，我们在提供服务时仍然受到容量限制。我们对此很兴奋，但也始终有一些地方需要改进。AI 前沿变化很快，我们需要不断改进模型。我们也看到了 Argon 还有哪些方面做得不够好，接下来会继续做好这些方面的工作。
Alex：具体是哪些方面？
Pichai：拿 Argon 来说，如果和两个月前的模型水平相比，它在各个方面都已经有了提升。但这个领域变化得很快。即使是编程，有些地方也还需要做得更好。我们也收到了明确的反馈，会针对这些问题继续改进。
Alex：那我换个角度，说说为什么有人认为，对谷歌来说，普通模型可能就已经够用了。我来复述一下他们的理由。
Pichai：好的。
Alex：现在很多普通模型已经足够强，Meta 的 Muse 就是一个例子。有人认为，谷歌也完全可以用这些模型做好地图、Gmail 和搜索，把最前沿的能力留给生物学研究等特定领域。为什么谷歌还要坚持追求最强模型？
Pichai：我一直都不是这样看待这件事的。我们始终致力于不断提高技术水平，探索这项技术究竟能发展到什么程度。从 Transformer 开始，我们就一直在朝这个方向努力，今后也会继续这样做。
我们很幸运，无论面向个人用户还是企业，都有很多可以应用这些技术的机会。有些云客户确实需要最前沿的模型。而且，在递归式自我改进的时代，前沿模型能够更好地帮助开发下一代模型，因此，保持前沿能力在这方面也很重要。在内部，把 Argon 提供给开发人员之后，我们确实感受到了明显的变化：它能够处理一些过去的模型解决不了的问题。所以，站在技术前沿，也就让你站到了创新前沿，能够帮助你继续推动创新。
不同规模、不同能力的模型，都有各自适用的场景。我们一直这样认为，所以才会开发 Flash、Flash-Lite 等模型。我也同意，不一定非要使用最前沿的模型，才能做出很好的产品体验，毕竟这些前沿模型通常也更慢、更贵。不过，我们也很重视前沿技术的进步，因此，这些方向我们都会继续投入。
Alex：你刚才说，开发前沿模型，能让你更早看到未来的发展方向。那么，你看到了哪些即将到来的变化？
Pichai：现在之所以这么令人兴奋，是因为经过几个月的努力，你终于把成果做出来、发布出去，而这时，你已经能清楚地看到，接下来还可以往哪些方向继续探索。
我们一直在努力，让智能体能够持续执行较长时间的任务。很显然，接下来前沿关注的领域，是让多个这样的智能体相互协作，共同完成任务。这就是一种正在形成、令人期待的新范式。我也很清楚地看到，未来无论是个人用户还是企业用户，都会拥有某种多智能体框架，让这些智能体持续为他们运行、处理事情。所以，你会觉得，这是一个充满无限可能的时刻。
让模型参与研发，也让模型理解世界
Alex：你提到了递归式自我改进，也就是让模型自己改进自己。这个过程甚至可能不需要人类参与。有人把它称为“智能爆炸”，仿佛“砰”的一声，AI 就突然具备了超越人类的能力。我记得，Sergey 在和 DeepMind 研究人员的一次圆桌交流中谈到，Argon 的核心创新之一，就是更多地依靠模型来帮助开发模型。
但最近 AI 模型出现的那些情况，我们也都看到了。它们在对齐方面已经做得不错，但仍然存在偏差，而这些偏差可能带来很大的问题。科技界已经传开了 Hugging Face 遭入侵的事件，越来越多的公众也逐渐了解：OpenAI 的 AI 智能体突破了隔离限制，入侵了 Hugging Face，做出了一些很出格的行为。谷歌也出现过类似情况，你们的 AI 智能体获得了不该获得的访问权限，后来，它们停止了这些行为。
但是，已经有足够多的证据表明，这些模型并不总是完全按照我们的意愿行事。在这种情况下，继续追求递归式自我改进，甚至只是让模型以更快的速度进步，这个想法都会让我感到不安，甚至害怕。你不觉得，大家有这样的担忧是合理的吗？是不是应该采取一些措施，让科技行业在这件事上冷静一点？
Pichai：你的问题可以分成三个部分来回答。先说 RSI，也就是递归式自我改进本身。你说得对，Sergey 确实提到，RSI 在 Argon 的开发过程中发挥了重要作用。但有一点需要说清楚：如果把 RSI 的发展看作一个逐步推进的过程，我们现在距离它的最终形态还很远。更准确地说，目前模型开发中的一些环节已经在利用 AI 来帮助改进模型，但这个过程仍然有研究人员参与。你可以理解为，智能体通过一定程度的自动化，协助 AI 研究人员，一起完成模型开发。我认为，至少应该先弄清楚，我们目前究竟处在什么阶段。我预计，到了明年，RSI 还会继续取得进展。
第二个问题是安全。安全需要依靠多层防护，也就是纵深防御。你刚才谈到了对齐，我们都在做大量工作，尽可能确保模型与我们的意图保持一致。但对齐仍然是一个需要持续研究的问题，目前也确实存在明显的不足。不过，我们并不是只依靠对齐来保障安全。
除此之外，我们还投入了大量精力，实时监测模型运行时的行为，通过遥测数据充分了解正在发生什么，确保一旦发现异常行为，就能及时介入。具体如何实施、采用什么样的架构，以及如何通过沙箱和隔离机制，在实时监测的同时让模型继续完成任务，都是我们正在投入的方向。如何做到这一点，采用什么样的架构，如何确保相关环节在沙箱中运行并得到隔离，让监测系统在实时监控期间仍然能够持续履行自己的职责，这些都是我们正在投入的方向。
当然，在部署模型时，我们也会让企业客户拥有相应的控制权，包括实时监测模型，以及自行决定模型部署在哪里、可以访问哪些类型的数据。这些工作都非常重要，我们正在大力推进。所以，整个行业都需要不断提高安全水平，花足够的时间把这些工作做好，也要投入必要的算力，确保模型能够安全部署。
第三个问题是，做了这些究竟够不够。这也是行业内不断讨论的问题。我认为，有担忧是合理的，但面对这些担忧，正确的做法是尽早思考，我们还能做些什么。你也确实能看到，整个行业在对齐、安全测试等方面开展的工作越来越多。随着这些模型投入使用，我认为下一步应该制定更完善的标准，让整个行业有共同遵循的规范，包括引入独立评估者开展评估。比如在财务管理领域，我们有会计准则，也有独立审计机构。对 AI 行业来说，这些都是接下来应该推进的工作。在这个过程中，监管最终也会发挥作用。
Alex：接下来问一个技术问题。谷歌有一点比较特别：你们既在沿着 Transformer 开启的方向研究大语言模型，也在推进世界模型研究。有些 AI 研究机构，比如 OpenAI，基本上已经认定大语言模型就是通向 AGI 的道路，并且收缩了部分世界模型研究，也就是那些试图理解物理世界的 AI 模型研究。但你们在同时推进这两个方向。能不能谈谈，为什么你认为世界模型的研究很重要？
Pichai：人会感知周围的世界，从中获取各种信息。我认为，让模型理解物理规律、准确地模拟现实世界，都是发展智能所必需的基础能力。我们也在探索这项技术各种可能的应用。拿机器人来说，Argon 在具身推理测试中的领先表现并不是偶然的。它在其中多项测试中都达到了目前最先进的水平，这也得益于我们开发世界模型积累的经验。我预计，未来会出现能够把这些能力融合起来的架构。但目前，智能前沿在各个方向的发展还不均衡，我们则希望在各种可能的技术路径上都保持领先。我认为，这会帮助我们开展更好的研究，加深我们对智能的理解，并开发出更好的 AGI 模型。这就是我们的考虑。
Alex：是的，Argon 发布之后，我看了你的推特，上面几乎都是现实世界里的应用，而不只是文本或者模型生成文本之类的例子。
算力先给谁？谷歌愿为前沿 AI 放弃部分云业务收入
Alex：算力是你们最宝贵的资源，有人和我说，谷歌应该退出 AI 研究，直接把自己的云计算资源和芯片拿去卖就行，英伟达就是一个例子。一方面，云业务希望通过对外出售算力获取较高的利润；另一方面，AI 研究又需要规模足够大的计算系统，才能开发出最好的 AI，两者在算力分配上存在拉扯。你曾说过你工作中的一个重要部分，就是决定如何分配算力。那么，你具体是怎么做，如何平衡这两方面的需求？
Pichai：对我们来说，首先要确保投入足够的算力，保持我们站在 AI 前沿的地位。2015 年，我就把公司的方向定为“AI 优先”。因此，从搜索到云，再到 Waymo，AI 都是贯穿我们所有业务的一项关键基础能力。
因此，我们首先要保证，无论是开展研究，还是开发模型，都有足够的算力。在此基础上，我们再为各项业务安排算力。每项业务都有自己的长期规划。它们会根据这些规划预测未来的需求。比如，云团队有业务预测，有客户需求，也有已经签订的合同。这些信息都能帮助我们规划云业务长期的算力投入。
我们很幸运，有这么多可以发展的业务。不过我认为，世界上没有哪家公司完全不需要作出取舍。我们遵循的首要原则，是确保有足够的算力支持前沿 AI 研发。这一点必须得到保障，其他业务的安排都以此为基础。还要考虑到，我们在建设自己的计算基础设施和数据中心，这些设施的建设周期很长，需要提前规划。因此，我们今天拥有多少算力，取决于我在 2024 年作出的决策。哪怕只是回想一年前，那时智能体编程还没有迎来爆发。今天再看，显然大家的算力都不够，不只是我们，其他公司也一样。每个人都希望，当初规划的算力规模能再翻一倍。
我认为，这是我们在这个行业里必然要面对的问题。我们始终需要提前规划未来两到三年的算力需求，但两年后，随着技术的发展再回头看，可能就会发现，当初有些决策并不是最优的。我不会花太多时间纠结这些事情。我分配算力的思路，是先确保 Google DeepMind 在开展研究和开发前沿模型时，能够获得最先进的算力支持，然后再为各项业务作出安排。同时，像我们现在这样，把算力也提供给其他公司，是一件很好的事。这既能帮助我们扩大规模，也能让相关技术在尽可能广泛的实际应用中得到检验，确保我们在 AI 技术栈的每一层都保持领先。
Alex：有一期播客 Ben Horowitz 转述 Andy Jassy 的话，大意是算力已经卖到 2028 年了。你所看到的情况也是这样吗？
Pichai：我认为，这样说基本符合实际情况。如果今天有客户来找我们，希望预订 2027 年的算力，我们很难满足这样的需求。去年，我们还专门租用了一些短期算力，部分原因就是为了满足一些客户的需求，帮他们解决眼前的问题。所以，我认为大家面临的算力供应周期都差不多。
Alex：有人说，每一次技术变革都会出现过度建设，比如铁路和光纤。这一轮 AI 基础设施建设，是不是也存在同样的问题？
Pichai：这要看你从多长的时间跨度来判断。如果从更长远的角度来看，现在这样的建设规模，在我看来完全合理。我们正处在有史以来影响最深远的一项技术发展的早期阶段。而且，这和当年的光纤不同，我们建成的算力设施正在实际使用，每天都面临容量不足的问题。应用在不断发展，底层技术也在快速进步。不过，作为一家公司的经营者，我在作决策时，会看实际的需求，看已经签订的合同所对应的收入，再结合底层技术能力、内部评估、我们自己的使用情况，以及模型能力的发展趋势，来规划算力建设。我们既要有信心、积极投入，也要保持审慎，合理安排。
但如果放眼整个经济环境，不同参与者承担的风险差别很大，我很难对所有人的情况作出判断。我会专注于公司能够控制的事情，并据此作出规划。如果你问，有没有可能出现过度建设，答案是有可能。因为算力建设需要各方面的条件相互配套，某些组件可能成为限制，而要真正建成这些算力设施，所需的各项条件必须配齐。
Alex：内存、电力，还有电工......
Pichai：对，涉及的组成部分可能有两千种，甚至更多。所以，某个环节可能会成为瓶颈，导致其他部分虽然已经投入建设，却无法充分利用，进而出现过度建设的情况。此外，经济也可能受到外部冲击，宏观风险始终存在。这些情况在任何时候都有可能发生。但如果你问，我是否相信，只要时间足够长，我们现在规划的这些算力最终都能派上用场，我的回答是：绝对相信。
Alex：关于这个问题再追问一下。既然你们要同时做云业务和 AI 研究，外界有这样一种看法：对 AI 研究人员来说，计算系统的规模越大，做出的 AI 就越好。当年你说 AI 的影响会像火和电一样深远时，我就在现场，我们当时都笑了。现在想起来，我有点后悔当时笑了。如果你真的相信 AI 会有这样的影响，那么，即使要牺牲眼前的利润，比如云业务的利润，也应该为研究人员提供尽可能大规模的计算系统。你怎么看？
Pichai：我们其实都已经在不同程度上这样做了。为了支持前沿 AI 研发，我们一直都在算力分配上作取舍，云业务也因此放弃了一部分短期收入。所以，你说的情况实际上已经在发生。不过，这件事也需要仔细权衡。即使对 AI 研究人员来说，适当的约束也有好处。它能够激发创造力，促使研究人员在算法效率上取得突破，比如想办法减少内存使用。所以，我们会确保投入足够的资源，让我们能够始终站在 AI 技术前沿。这就是我在分配算力时的基本考虑。
AI 时代的界面，要适应人的习惯
Alex：我想知道，你认为未来的交互界面会是什么样子？随着 AI 逐渐普及，我们今后会怎样使用计算机？几个月前，我和 OpenAI 总裁 Greg Brockman 聊过这个话题，他说了一段很有意思的话：“你想要什么样的界面？你想要什么样的产品？”他的回答是，你其实不需要界面，也不需要产品，而是希望能像你我现在这样直接交流，我们之间不存在什么界面，基本上就是交谈。他的意思是，如果 AI 已经足够聪明，就不应该还需要一块屏幕，或者那么多需要用户手动操作的按钮和其他东西。你怎么看 AI 时代的交互界面？
Pichai：我的想法是，过去的一直是人类适应计算机；有了 AI 之后，情况会反过来，由技术来适应人的生活方式和日常习惯。你希望在需要的时候，随时都能和 AI 交谈；它了解上下文，能够持续运行，让前后的交互自然衔接起来，而不需要你处理各种复杂的操作。这些想法我都认同。不过，在我看来，这并不意味着以后就不需要界面了。
拿我现在使用手机的体验来说，Gemini 的语音模型已经非常好了。我只要按一下麦克风按钮，就可以随意地说上一段，我们把它叫作“rambler”。它会结合对我个人情况的了解，自动纠正识别结果。就连我朋友那些比较特别的名字，它也能识别正确，哪怕一开始猜错了，随后也会纠正过来。它能够这样适应我，确实很了不起。现在我说完之后，已经不用再手动删除、修改识别出来的文字了。它能进步到这个程度，确实令人惊讶。
但有些时候，你还是需要界面。比如，你可能想找点娱乐，或者看一段视频，就可能会让 AI 创作一些内容，然后观看、欣赏这些内容。所以，在我看来，关键是让交互尽可能顺畅、自然。用户不应该费力地去适应某一种界面，而应该由技术来适应用户当下做事的方式。
Alex：是的。我想再聊几个和 OpenAI 有关的话题。因为去年，你们让他们进入了“红色警报”状态。你肯定也知道，当时 Gemini 表现很好，OpenAI 严阵以待。我和 Sam Altman 谈到来自谷歌的挑战时，他说，他认为 OpenAI 能够应对你们的竞争，因为他们在打造 AI 原生产品。我不确定他是不是用了这样的措辞，但他的言下之意是，谷歌只是在现有产品上加上 AI 功能。你怎么看这种说法？当然，我也想先问问，你是否接受这样的描述。在现有产品上加入 AI 功能，能和 AI 原生产品竞争吗？
Pichai：首先，现在这个领域变化很快，我们都在相互推动。OpenAI 也给我们带来了很大的竞争压力，远不是只有我们在给他们施加压力。
Alex：是的，你们也进入过“红色警报”状态。当然，我知道，对于究竟有没有进入过这种状态，还有一些争议。
Pichai：我的意思是，在 AI 发展的这个阶段，我们所有人都需要根据当时的情况，以不同程度的紧迫感作出回应。而且，我们也一直在开发新产品。比如 Gemini，我说的是 Gemini 应用，它就是从零开始打造的。再比如，接下来，随着大家都转向智能体，你会说 OpenAI 是在 ChatGPT 上加上智能体功能，还是在打造原生的智能体产品？你会怎样判断？同样的问题，也完全可以反过来问他们。
所以，我不会用这样的方式理解问题。在我看来，重要的是打造世界一流、能够赢得用户的 AI 原生产品体验。有些时候，需要改造现有产品；有些时候，则需要重新思考，设计全新的体验。这些都是我们应对这一轮变化时需要做的工作。
Alex：再谈谈商业模式，你认为未来围绕这种 AI 交互方式的商业模式会是什么样子。比如，我让自己的智能体出去帮我做一件事，它向我推荐了一款产品。如果我最后买了，智能体是不是应该从这笔交易里抽成？即使事先明确告知用户，这样做是否合适？或者换个角度，你们打算怎样通过这些产品赚钱？
Pichai：我们先不谈企业业务。
Alex：对，我问的是面向个人用户的产品。
Pichai：对于个人用户，我认为，他们会越来越认可这项技术的价值。拿现在的软件开发人员来说，他们愿意为编程模型付费，因为这些模型确实能够提高他们的工作效率。没有人愿意回到三年前的编程方式，甚至不愿意回到去年这个时候的方式。所以，只要产品为用户创造了价值，就有机会获得相应的回报。回顾消费产品的发展，我还没有见过哪一种产品，真正做到了数十亿用户的规模，并且用户还在高频、深入地使用，却没有相应的变现方式。只要产品确实为用户创造了价值，就有机会获得商业回报。
Alex：完全同意。
Pichai：当然，具体的变现形式可能会变化。你刚才说的是让智能体替你办事的场景。用户终究会有购买商品或服务等商业需求，智能体需要帮助用户满足这些需求，另一端的商品或服务提供商也会从中获得价值，最终会形成相应的商业机制。一个有效率的系统就会这样运转。不过，我认为，最重要的还是打造世界一流的消费产品体验。
Alex：我想提一个产品功能需求。把 Gmail 接入 ChatGPT 之后，在 ChatGPT 里搜索邮件很好用。Gmail 自己由 Gemini 提供支持的搜索，能不能也做到同样的水平？
Pichai：我们最近公布了一项很让人兴奋的进展，就是面向企业的统一 Gemini 智能体，而且我们也会把它带给个人用户。这些智能体能够全天候在云端持续运行，实际上相当于为每次查询启动一个微型虚拟机，也就是 microVM。从运行时、智能体执行框架，也就是 harness，到模型，整套系统都作为一种不依赖特定前端界面的服务运行。因此，它可以在不同产品中工作，并且延续你的上下文信息。
无论你是在 Workspace、搜索，还是 Gemini 里，都可以调用这个智能体。我认为，智能体会把这些产品的使用体验连接起来，实现用户一直期待的那种统一体验。你说得也对，我们的产品很多，同时又在快速推进，难免有些产品进展快一些，有些慢一些。你提出的问题确实存在。不过，在打造智能体体验的过程中，我们正在投入大量时间和精力，确保用户在不同产品中都能获得连贯的体验，相关能力也都能正常发挥作用。
Alex：好，这其实也很好地回答了刚才关于交互界面的问题。
Pichai：是的，确实如此。
Alex：说到谷歌下一步会在哪些方向发力，大家也会问：谷歌自己的 Muse 或 Dots 在哪里？这些产品最近似乎突然就出现了，而且很快流行起来。你们有类似的产品吗？或者说，你们是否打算开发一款新产品，来和 Muse 这样的产品竞争？
Pichai：其实，我们对 Gemini 应用的设想一直很明确：把它打造成一个了解用户个人需求、能够主动提供帮助、能力强大的助手。这就是我们正在努力的方向。在今年的 Google I/O 上，我们已经通过 Spark 展示了它的早期版本。事实上，现在部分订阅用户打开 Gemini 应用，就已经能看到“聊天”和“任务”两个入口。
这些功能已经开始让用户体验到智能体的能力。有了 Argon，这些能力又有了大幅提升。我们最近在面向企业的活动上公布的能力，也会向个人用户开放。很快，你只需要向 Gemini 提出需求，不必先区分自己是在和它聊天，还是在让智能体执行任务。它会自动采取合适的处理方式，安全地在后台持续运行，帮你把事情办好。
从 Gemini 应用的实际使用情况来看，如果把去年用户发来的内容分类，绝大多数都属于聊天。但现在，越来越多的用户在让 Gemini 执行任务，而且这类请求所占的比例还在稳步上升。所以，是的，我们会为个人用户提供专属的智能体。无论是在 Gemini 应用、谷歌搜索，还是 Gmail 中，用户都可以调用它。我们希望先把这套体验做好。它将由 Argon 提供支持，我们非常期待让用户用上这些能力。
从 Anthropic 到太空数据中心，谷歌如何押注未来？
Alex：接下来再问几个其他方面的问题。算上已经投入的资金和承诺的投资，谷歌对 Anthropic 的投入达到数百亿美元。你在 Dario 身上看到了什么，让你愿意作出这么大规模的投资承诺？
Pichai：我们很早就关注了 Anthropic 的创始团队。当时就看得很清楚，这是一支世界一流的团队，他们参与过一些最早期的 GPT 模型研发。因此，我们很快就判断，他们有能力站到技术前沿。我们一直希望找到有望引领技术发展的公司，并与它们合作。我们和 SpaceX 的合作是这样，和 Anthropic 的合作也是这样。这就是我们当时看重的地方。后来，我们有机会与他们开展合作，其中
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
I'll translate it.
Editor Cai Fang
How many codes did Google write in AI? Last year, CEO Sandar Pichai also gave a clear figure: 25 per cent. But when the moderator asked the question again, his answer changed: “This ratio is no longer significant”.
Because AI has been deeply involved in code development inside Google. Within four weeks of Gemini 4 Argon Z94Z, Google code change submissions increased by about 30 per cent, rollbacks decreased by about one third, and developers began processing complex projects involving millions of lines of code with Gemini 4 Argon. Google is now more concerned with indicators such as code changes, rollbacks, etc. than the percentage of AI writing codes. In Pichai ' s view, these changes better reflect AI ' s impact on development efficiency and quality.
Recently, Big Technology Podcast, a guest in Pichay, spoke with Alex Kantrowitz about Google input and trade-offs in the AI competition. Gemini 4 Argon, the first 72 hours of use was close to 25 trillion token, about three times as much as Gemini 3.0 in the same period, but the calculator did not keep up with demand. At a time when the outside world thought that Google might abandon the AI front-line competition and focus on integrating a capable model into existing products, Pichai responded quite clearly: "This is a miscalculation." Pichai revealed that, in order to secure forward model development, Google had given priority to computing to the AI research team, and that cloud operations had therefore abandoned part of their short-term income.
In addition to model development and computing competition, he spoke about AI's involvement in model development, ongoing personal AI agent, world models, and Google's response to competition from Open AI. Below is a compilation of this interview.
Too long to read:
Alex: The outside world once thought that Google was no longer pursuing the front-line AI, and the model was enough. But Argon's presentation, isn't it?
Pichai: This is a miscalculation. The pursuit of the technological frontier has been a central goal of Google. The first 72 hours after Argon went online were close to 25 trillion token, about three times as much as Gemini 3.0 in the same period. We are satisfied with its performance and see where it needs to be improved, which will continue to enhance its capacity.
Alex: It has been argued that ordinary models are enough for maps, Gmails and searches, and that only specific tasks such as biological research require cutting-edge models. Why would Google pursue the most advanced powers?
Pichai: We want to continue to promote technological progress. Some cloud customers need front-line models, and stronger models can help develop next-generation models. Argon has solved some of the problems that past models could not handle internally. Of course, models of different sizes and capabilities are appropriate and we will continue to invest.
You said that being on the front line of technology would make you see the future sooner. So, what changes have you seen coming?
Pichai: We've been working on keeping AI agent on a long mission, and we're going to move on to a number of AI agent collaborations. In the future, both individual and enterprise users may have multi-AI agent systems that work continuously for themselves.
Alex: Models can't be done exactly as people want, but we're looking for regressive self-improvement to make them more capable. I am worried that measures should be taken to calm the industry down on this matter.
Pichai: Concerns are legitimate, but first it is necessary to distinguish between the current stage of development. At present, AI agent is assisting researchers in developing models that still involve humans and are far from being fully autonomous in retrogressive self-improvement. Security cannot only depend on alignment, but also requires real-time monitoring, sandbox isolation, unusual intervention and customer control. next step, industry needs better standards, independent evaluation and regulatory involvement.
Alex: Some institutions see the mega-linguistic model as the main route to AGI, while Google also promotes world modelling. Why do you think this route is important?
Pichai: Understanding physical laws and accurately modelling the real world are important foundations for the development of intelligence. Argon has also benefited from the experience accumulated by world modelling studies in its reasoning. In the future, these capabilities may be integrated through new structures, and we hope to remain in the lead on different technical paths.
Alex: The cloud business seeks to earn profits through external sales, while AI research requires large-scale computing resources to develop the best models. How do you divide the numbers between those two?
Pichai: First, secure front-line AI research and model development, and then calculate it based on long-term planning, demand forecasting and contractual arrangements for each operation. It is our primary principle that the resources needed for forward research and development must be secured.
Alex: If you really believe that the impact of AI is as far-reaching as fire and electricity, should researchers be provided with the largest possible computing resources, even at the expense of immediate profits from operations like cloud operations?
We have done this to varying degrees, and the cloud business has given up some of its short-term revenues. However, appropriate constraints can also stimulate creativity and drive a breakthrough in algorithm efficiency. Our principle is to ensure that sufficient resources are invested and that we remain at the forefront.
Alex: Greg Brockman believes that an AI that is smart enough should allow people to express their needs directly through conversation and reduce their dependence on screens and buttons. What do you think of the AI Age interface?
Pichai: Technology should adapt to people ' s habits and allow for as much natural interaction as possible. But the interface still works, for example by watching videos and appreciating the content of AI. The key is to reduce the energy devoted by users to the use of technology.
Alex: Does Google have products like Muse, Dots, or is it ready to develop a new product that competes with them?
Pichai: Our goal for Gemini is to make it a powerful assistant who understands personal needs and can help out. Users need only demand and do not have to distinguish between chatting or handing over tasks, and it will do so in a suitable way, on an ongoing basis in the back and providing consistent experience in different products.
Alex: You said earlier that 25% of Google's internal code was written by AI. What's the rate now?
Pichai: This percentage is no longer meaningful because AI is deeply involved in development. We are more concerned with efficiency and quality: four weeks after Argon publish, code change submissions increased by about 30 per cent, roll back by about one third, and developers began to use it to process complex projects involving millions of lines.
Alex: Is there an AI study that has not yet been broken? For example, to keep the model publish up to date?
Pichai: We are simultaneously exploring the directions of synthetic data, continuous learning, more efficient architecture and world models, and investing a significant portion of our calculations. Some of the research results will be applied directly to major models.
Google gave up AI front-line competition? Pichai: That's a crazy judgment.
Alex: Today we invite CEO Alphabet Sundar Pichai to talk about Google's participation in the AI Frontline competition, how to balance the AI Frontline resource input for research and cloud operations, and what the future interface will look like. Sundar came to our studio in Mountainview today. Sundar, nice to meet you. Welcome to the show.
Pichai: Thank you, Alex, very happy to be here.
Alex: With the launch of Gemini 4 Argon, Google is back to the AI front. To put it bluntly, it is a model that is at the world's leading level and has updated many benchmark test scores, surpassing the models already published by Open AI and Anthropic in many tests of programming, mathematics, science and knowledge. Prior to this, it was generally accepted that Google had largely abandoned front-line competition and intended to develop fast enough models that could then be applied to its own products and no longer pursue the front-line AI. But it has proved wrong.
Pichai: And it's so wrong. We have always been very determined to be at the forefront of technology, and we have always done so in pursuit of what is in our bones. For over a decade, we've been thinking about the development of the AI frontier. The field is changing rapidly and is highly competitive and exciting.
I'm very satisfied with Argon's final performance. In the development process, we have not been eager to deliver, but rather to ensure that security tests are fully conducted and that they are carefully deployed. Given its ability, we spend more time on it than on other models. But on the whole, whether it's a AI agent mission, knowledge work, programming, or creative task, I think Argon is one of the very good frontier models.
The first 72 hours on the line, it's almost 25 trillion token, almost three times the 72 hours before Gemini 3.0. Currently, we continue to face capacity constraints in the provision of services. We are excited about this, but there is always room for improvement. AI's front line is changing fast, and we need to constantly improve the model. And we see that there's something else that Argon didn't do well, and we're going to do it again.
Alex: What are the specific aspects?
Pichai: Using Argon, it has been upgraded in various ways compared to the model level two months ago. But this area is changing rapidly. Even in programming, there are places that need to be better. We have also received clear feedback that will continue to improve on these issues.
Alex: Then let me put it in another direction and say why some people think that ordinary models might be enough for Google. Let me repeat their reasons.
Pichai: Okay.
Alex: Now many ordinary models are strong enough, and Meta's Muse is one example. The view was expressed that Google could also make maps, Gmails and searches using these models, leaving the cutting edge of capabilities to specific areas such as biological research. Why does Google insist on pursuing the strongest model?
Pichai: I've never seen this the same way. We have always been committed to increasing the level of technology and to exploring the extent to which it can develop. We have been working in this direction since Transformer and will continue to do so in the future.
We are fortunate that there are many opportunities to apply these technologies, both for individual users and for businesses. Some cloud customers do need cutting-edge models. Moreover, in an era of regressive self-improvement, front-line models are better able to help develop next-generation models, and maintaining front-line capabilities is also important in this regard. In-house, with Argon being made available to developers, we do feel a clear change: It can deal with problems that past models cannot solve. So, being on the front line of technology puts you on the front line of innovation, helps you move on.
Models of different sizes and capabilities have their own scenarios. We've always thought so, so we've developed models like Flash, Flash-Lite. I also agree that it is not necessary to use cutting-edge models to make good product experiences, which are usually slower and more expensive. However, we also attach importance to advances in front-line technology, and we will therefore continue to invest in those directions.
Alex: You just said that the development of a front-line model would allow you to see the way forward earlier. So, what are you seeing coming changes?
Pichai: It's so exciting that after months of effort, you've finally made the results, and publish has gone out, and you can see clearly which direction to follow.
We have been trying to keep AI agent on a longer mission. It is clear that the next front-line area of concern is to bring together several of these AI agents to fulfil their mandates. This is a new and promising paradigm that is emerging. It is also clear to me that in the future, both individual and business users will have some kind of multi-AI agent framework to keep them running and doing things for them. So, you'll feel that this is a time of infinite possibility.
Let the models be developed and the models understand the world.
Alex: You're talking about retrogressive self-improvement, that is, making models improve themselves. This process may not even require human involvement. It's called a "smart explosion," like a bang, and all of a sudden AI has the ability to go beyond humans. And I remember that Sergei, in a round-table exchange with DeepMind researchers, said that one of the core innovations of Argon was to rely more on models to help develop models.
But we've seen all the recent developments in the AI model. They have done well in alignment, but there are still deviations, which may pose major problems. The scientific and technological community has heard of the invasion of Hugging Face, and more and more the public is becoming aware that Open AI's AI AI agent has broken the isolation restrictions, invaded Hugging Face, and committed some extraordinary acts. A similar situation occurred in Google, where your AI AI agent obtained access to permission, which was not due, and later they stopped these acts.
There is, however, enough evidence to show that these models do not always follow our wishes. In these circumstances, the idea of continuing to pursue incremental self-improvement, even allowing models to advance at a faster pace, would upset and even scare me. Don't you think it's reasonable to worry like that? Should some measures be taken to calm the industry down?
Pichai: Your questions can be answered in three parts. First of all, RSI, which is the incremental self-improvement itself. You're right, Sergei did mention that RSI played an important role in the development of Argon. But one thing needs to be made clear: if we look at RSI development as a step-by-step process, we are still a long way from its final form. More precisely, some elements of the current model development are already using AI to help improve the model, but the process still involves researchers. You can understand that AI agent is assisting AI researchers with model development through a certain degree of automation. I think it would be useful to know, at least first, what stage we are at. I expect that next year, RSI will continue to make progress.
The second issue is security. Security depends on multiple layers of protection, i.e. deep defences. You just spoke of alignment, and we are all doing a lot of work to ensure, as far as possible, that the model is consistent with our intentions. However, alignment remains an issue requiring continued research and there are clear gaps. However, we do not rely solely on alignment to ensure security.
In addition to this, we have invested considerable effort in real-time monitoring of the behaviour of the model at the time of its operation, in order to gain a good understanding of what is happening through telemetry data and to ensure timely intervention in the event of unusual behaviour. We are investing in how to implement, what structure to use, and how to keep the model on course with real-time monitoring through sandboxing and isolation mechanisms. How this is done, what structures are used, how to ensure that links are operated and isolated in sandboxes, and that monitoring systems are able to perform their duties on a continuous basis during real-time monitoring are the directions we are investing in.
Of course, in deploying models, we will also give corporate clients the corresponding control, including real-time monitoring models and self-determination of where the models are deployed and what types of data are accessible. These efforts are all very important, and we are moving forward vigorously. Therefore, there is a need to continuously improve the level of security throughout the industry, to spend enough time on these tasks and to invest in the necessary calculations to ensure that models are deployed safely.
The third question is whether this is enough. This is also an ongoing issue within the industry. I think that there are legitimate concerns, but in the face of those concerns, it is right to think as soon as possible about what we can do. As you can see, more and more work is being done across the industry on alignment, safety testing, etc. As these models become operational, I believe that next step should develop better standards for common industry-wide norms, including the introduction of independent evaluators to conduct assessments. In the area of financial management, for example, we have accounting standards and independent audit bodies. For AI, these are the next steps. In the process, regulation will ultimately also play a role.
Alex: There's a technical question. Google is a bit special: you're working on big language models in the direction that Transformer is opening up, and you're moving forward with world modelling. Some AI research institutions, such as Open AI, have largely determined that the mega-linguistic model is the path to AGI and have contracted some of the world's model studies, that is, those that attempt to understand the physical world. But you're moving in both directions. Can we talk about why you think world modelling is important?
Pichai: People know the world around them and get all kinds of information from them. In my view, giving models an understanding of physical patterns and an accurate simulation of the real world are all essential building blocks for the development of intelligence. We are also exploring possible applications of this technology. Using robots, Argon's lead performance in a body-debate test is no accident. It has reached the state-of-the-art level in many of these tests and has also benefited from our experience in developing world models. I anticipate that in the future there will be structures that can integrate these capacities. At present, however, the development of the intellectual frontier is uneven in all directions, and we hope to remain ahead of every possible technological path. I think that would help us to do better research, deepen our understanding of intelligence and develop better AGI models. That's what we're thinking.
ALEX: Yes, after Argon publish, I saw your Twitter, which is almost all about real-world applications, not just examples of text or model-generated texts.
Who do you think you're going to start with? Google is willing to give up part of the cloud business for Al.
Alex: Counting is your most valuable resource, and I was told that Google should quit the AI study and sell its cloud computing resources and chips directly. On the one hand, the cloud business seeks to generate higher profits through external sales; on the other hand, AI needs a large-scale computing system in order to develop the best AI, and there is a tension in the distribution of power. You said that an important part of your work is to decide how to distribute the calculus. So what exactly do you do, and how do you balance the needs of both?
Pichai: For us, the first thing to do is to make sure that we have enough credit to keep us at the forefront of AI. In 2015, I set the direction of the company “AI priority”. So, from search to cloud, to Waymo, AI is a key foundational capability that cuts across all our operations.
Therefore, we must first ensure that both research and model development are sufficiently smart. On that basis, we will then be able to calculate the operational arrangements. Each operation has its own long-term planning. They predict future needs based on these plans. For example, cloud teams have business forecasts, client needs and contracts that have been signed. This information can help us to plan long-term computing inputs for cloud operations.
We are fortunate to have so many businesses to develop. I believe, however, that there is no firm in the world that needs to make trade-offs at all. The first principle that we follow is to ensure that there is sufficient calculus to support frontier AI research and development. This must be ensured, and other business arrangements are based on this. It is also important to take into account that we are building our own computing infrastructure and data centres, which have a long construction cycle and require advance planning. So, what we have today depends on the decisions I made in 2024. Even if only a year ago, AI agent programming had not yet broken out. Looking again today, it is clear that none of us, not just us, but other companies. Everyone hopes that the size of the calculator originally planned will double again.
I think that's what we have to face in this industry. We have always had to plan ahead for the next two to three years of computing needs, but two years later, as technology develops, it may be discovered that some decisions were not optimal. I won't spend much time on these things. The idea of allocating power is to make sure that Google DeepMind is able to obtain state-of-the-art calculus support for research and development of forward models before making arrangements for operations. At the same time, it is a good thing, as we are now, to give the math to other companies. It helps us to scale up, and it allows the technology to be tested in as wide a practical application as possible, ensuring that we stay ahead of each level of the AI technology vault.
Alex: One of the podcasts, Ben Horottz, relays Andy Jassy's words to the effect that the math has been sold for 2028. Is that what you see?
Pichai: I think this is basically true. If a customer comes to us today and wants to book 2027, it's hard for us to meet this demand. Last year, we also rented some short-term calculations, partly to meet the needs of some clients and to help them solve immediate problems. So I think we all have the same calculus supply cycles.
Alex: It has been said that every technological change is over-building, such as railways and fibre-optics. Is the same problem with the AI infrastructure in this round?
Pichhai: It depends on how long you've been. It seems to me that, in the longer term, the scale of the construction is entirely reasonable. We are at an early stage of one of the most far-reaching technological developments in history. Moreover, unlike the fibre-optics of the year, the computing facilities that we have built are actually being used and are facing a daily shortage of capacity. Applications are constantly evolving and bottom-up technologies are improving rapidly. However, as an operator of a company, when I make a decision, I look at the real needs, the revenues corresponding to the contracts that have been signed, and then plan for the build-up of algorithms in conjunction with the bottom-up technical capabilities, internal assessments, our own use and the evolution of model capabilities. We need to be confident, active and careful and rational.
But it is difficult for me to judge the situation of all people if the risks assumed by the different participants vary significantly by looking at the overall economic environment. I will focus on what the company can control and plan accordingly. If you ask if there is a possibility of over-construction, the answer is possible. Since various conditions are required for computing, some components may be limited, and the requirements must be met if these computing facilities are to be truly constructed.
Alex: Memory, electricity, electricians...
Pichai: Yes, there may be 2,000 or more components involved. As a result, one link may become a bottleneck, leading to over-construction in other parts of the country that are already under construction but are not fully utilized. In addition, the economy may be exposed to external shocks and macro-risks persist. Such situations can occur at any time. But if you ask me, do I believe that, given the time that has elapsed, all of the calculations that we're planning will eventually come into play, and my answer is: absolutely.
Alex: Let's ask again about this. Now that you're doing both cloud business and AI research, there's a perception that for AI researchers, the bigger the computing system, the better it is. When you said that the influence of AI would be as far-reaching as fire and electricity, I was there, and we laughed. Now remember, I kind of regret laughing. If you really believe that AI has such an impact, then even at the expense of immediate profits, such as those from cloud operations, you should provide researchers with as large a computing system as possible. What do you think?
Pichai: We have all done this to varying degrees. In order to support frontier AI research and development, we have always made trade-offs in the distribution of power, and as a result the cloud business has abandoned some of its short-term revenues. So what you said actually happened. However, the matter also needs to be carefully weighed. Even for AI researchers, appropriate constraints would be beneficial. It stimulates creativity and leads researchers to breakthroughs in algorithm efficiency, such as finding ways to reduce memory use. So we'll make sure that we invest enough money to keep us at the forefront of AI technology. That is the basic consideration in allocating power.
AI-era interfaces, adapted to human habits
Alex: I wonder what you think the future interface will look like. How do we use computers in the future as AI becomes popular? A few months ago, I talked to Greg Brockman, CEO of Open AI, and he said, "What kind of interface do you want? What kind of product do you want?" His answer is that you don't really need an interface, you don't need a product, but rather you want to communicate as directly as you do right now, and we don't have any interfaces, basically just talk. What he means is, if AI is smart enough, it shouldn't need a screen, or so many buttons and other things that need to be manually operated by users. What do you think of the AI Age interface?
Pichai: My idea is that humans used to adapt to computers; with AI, it turns out that technology adapts to people ' s lifestyles and everyday habits. You want to be able to talk to AI whenever you need it; it knows context, it can keep running, and it connects with each other, without having to deal with complex operations. I agree. In my view, however, this does not mean that there will be no future interface.
With the experience I'm using my cell phone, Gemini's voice model is very good. All I have to do is press the microphone button and I can just say the last paragraph, which we call “rambler”. It automatically corrects the result of identification in the context of knowledge of my personal circumstances. Even my friend's more special names can be identified correctly, even if they were initially miscalculated and then corrected. It's amazing how it fits me. Now that I have finished, there is no need to manually delete or modify the identified text. It's amazing how far it can go.
But sometimes you still need an interface. For example, you might want to find some entertainment, or watch a video, so that AI can create something and then watch and appreciate it. So, in my view, the key is to keep interaction as smooth and natural as possible. Users should not struggle to adapt to one interface, but rather to the way they do things now.
Alex: Yes. I'd like to talk about a few more topics about Open AI. Because last year, you put them in the red alert. You must also know that Gemini was doing well, Open AI waiting. When I spoke to Sam Altman about the challenge from Google, he said he thought Open AI could handle your competition because they were building an AI primary product. I'm not sure he used that language, but what he meant was that Google only added AI to the existing product. What do you think of that? Of course, I'd like to ask if you accept this description. Adding AI functionality to existing products can compete with AI primary products?
Pichai: First of all, there is now rapid change in this area, and we are all pushing each other. Open AI also put a lot of competitive pressure on us, not just us.
Alex: Yes, you've been in the red alert. Of course, I know there is some controversy as to whether this has ever happened.
Pichai: I mean, at this stage of AI's development, we all need to respond with varying degrees of urgency, depending on the circumstances. And we've been developing new products. Gemini, for example, I'm talking about Gemini applications, which are built from scratch. And then, for example, as everyone turns to AI agent, would you say that Openai is adding AI agent to ChatGPT, or building original AI agent products? What would you say? The same questions can well be asked the other way around.
So I will not understand the problem in that way. In my view, it's important to build a world-class, user-friendly AI primary product experience. Sometimes there is a need to adapt existing products; sometimes there is a need to rethink and design new experiences. These are all the things that we need to do to cope with this round of change.
Alex: Turning to business models, what do you think the future business models around this AI interactive approach will look like? For example, I sent my own AI agent out to do one thing for me, and it recommended me a product. If I finally bought it, should AI agent be drawn from this deal? Would this be appropriate even if users were informed in advance? Or how do you plan to make money from these products?
Pichai: Let's not talk business.
Alex: Yeah, I'm asking about the product for personal users.
Pichai: For individual users, I think they will increasingly appreciate the value of this technology. For their part, the current software developers are willing to program models paid, because they do improve their efficiency. No one is willing to go back to programming three years ago or even to last year. Thus, as long as products create value for users, there is an opportunity for corresponding returns. Looking back at the development of consumer products, I have not seen any product that is truly at the scale of billions of users, who are still in high-frequency, in-depth use without corresponding liquidity. As long as the product does create value for users, there is an opportunity for commercial returns.
Alex: Absolutely.
Pichai: Of course, the specific form of realization may change. You were talking about the scene where AI agent was supposed to work for you. Users will eventually have commercial needs such as the purchase of goods or services, AI agent will need to help the users meet these needs, and the goods or service providers at the other end will derive value from them, eventually forming the corresponding commercial mechanisms. This is how an efficient system works. However, I think it is important to build a world-class consumer experience.
Alex: I want to raise a product functional need. After connecting Gmail to ChatGPT, searching for mail in ChatGPT is useful. Can Gmail do the same for its own search, supported by Gemini?
Pichai: We have recently published an exciting development, the unification of the company Gemini AI agent, and we will bring it to individual users. These AI agents are capable of running around the cloud all day, and are in effect equivalent to launching a mini-virtual machine for each query, the microVM. From running time, the AI agent implementation framework, that is, the harness, to the model, the entire system operates as a service that does not depend on a specific front-end interface. It can therefore work in different products and continue your context information.
Whether you're in Workspace, search, or Gemini, you can call this AI agent. AI agent, I think, will connect the use experience of these products to achieve the unified experience that users have always expected. You're right, we have a lot of products, but at the same time we're moving fast, and some are moving faster and slower. Your questions do exist. In building the AI agent experience, however, we are investing a lot of time and energy in ensuring that users have a consistent experience in different products and that the relevant capabilities function properly.
Alex: Well, that's actually a good answer to the question just asked about the interface.
Pichai: Yes, indeed.
Alex: When it comes to where Google next step works, you will ask: where is Google's own Muse or Dots? These products appear to have appeared recently and are rapidly becoming popular. Do you have any similar products? Or are you going to develop a new product to compete with a product like Muse?
Pichai: In fact, our vision of Gemini's application has been clear: it creates an assistant who understands the needs of the user, who can help out on his own initiative and who can be powerful. That is the direction we are working on. This year on Google I/O, we've shown it through Spark in an early version. In fact, now some subscribers can see both "chat" and "task" portals by opening Gemini applications.
These features have begun to give users the ability to experience AI agent. With Argon, these capabilities have been significantly increased. Our recent ability to announce business-oriented activities will also be open to individual users. Soon, you'll just have to ask Gemini for it, without having to tell whether you're talking to it or whether you're giving AI agent a mission. It will automatically take the right approach and will operate safely backstage to help you get things done.
According to the actual usage of Gemini applications, most of the content sent by users last year is chat. But now, more and more users are giving Gemini the task, and the proportion of such requests is steadily increasing. So, yes, we'll provide individual users with an exclusive AI agent. Users can call it in Gemini applications, Google search, or Gmail. We want this experience done first. It will be supported by Argon, which we very much look forward to using.
From Anthropic to space data centres, how does Google bet on the future?
Alex: A few other questions follow. With the funds and commitments already invested, Google's investment in Anthropic reached tens of billions of dollars. What did you see in Dario, making you willing to make such a massive investment commitment?
Pichai: We've been following the founding team of Anthropic for a long time. It was clear at the time that this was a world-class team that was involved in some of the early GPT models. Therefore, we quickly judge that they are capable of standing at the forefront of technology. We have always hoped to find and cooperate with companies that are expected to lead technological development. Our cooperation with SpaceX is this, as is our cooperation with Anthropic. That's where we looked. Later on, we had the opportunity to work with them, of which
Auto-fetched and translated by AI Atlas. All rights belong to the original author.