文 | 划重点KeyPoints
上线72小时，吞吐近25万亿Token，消耗量是上一代Gemini 3.0同期的3倍。
这是Google最新旗舰模型Gemini 4 Argon交出的首秀成绩单。哪怕是家底丰厚的Google，面对瞬间涌入的调用需求，也不得不再次实施算力限流。
在科技播客《Sources》的最新访谈中，Google CEO桑达尔·皮查伊复盘了Gemini 4 Argon的发布细节，首次公开解释了Gemini 3.5 Pro为何神秘消失，以及DeepMind核心管理层变动的幕后原因。
1.砍掉3.5 Pro直接冲刺Argon，内部研发已进入RSI阶段
过去几个月，外界一直好奇曾在I/O大会上提及的Gemini 3.5 Pro去哪了。皮查伊在访谈中直接给出了答案：为了抢夺Agent编程能力的高地，被团队主动舍弃了。
今年年中，Google在训练3.5系列时发现，模型在Agentic Coding上存在明显缺口。团队判断，利用体积更小、效率更高的Flash版本进行高频迭代，是补齐短板的最快路径。于是Google在约3个月内密集推进了从3.5、3.6、3.7到3.8 Flash的连续迭代。
当这套技术方案跑通并准备应用到Pro版本时，Gemini 4.0 Argon的预训练已经完成。考虑到此时再发布3.5 Pro生命周期极短，在算力受限的环境下，团队决定将Flash验证成功的所有技术积累，直接注入Argon的后训练阶段。
这一决策直接促成了Argon在长程任务上的跃升。在外部测试中，Argon耗时一天多、花费不到2000美元算力便自主通关了《我的世界》。而在Google内部，它已经被指向拥有数百万行代码的Chromium代码库。
至于外界传闻的Carbon模型，皮查伊澄清那只是内部前沿模型迭代的代号。目前团队正基于Argon推进模型蒸馏，研发能力接近但成本更低的轻量版本，以向更多用户开放。
随着新模型融入内部开发环境，皮查伊认为2027年是生产力的黄金时代。他透露，Google内部已经进入研究人员与模型深度协同的递归自我改进（RSI）阶段：模型正直接参与提出训练配方建议、自动评估下一个检查点。皮查伊本人每天也在使用内部Agent汇总预训练、中训练、后训练和数据群组的24小时研发进展，每天节省约1小时的信息同步时间，并通过接入遥测数据的Agent实时监控各部门的算力使用效率。
2.首谈DeepMind高层变阵，自由现金流转负背后的博弈
针对近期备受关注的Google DeepMind人事震荡：传奇工程师Jeff Dean离职创业、Demis Hassabis退居首席科学家与主席、Koray Kavukcuoglu接棒掌管DeepMind，皮查伊首次给出了官方解释。
除了效力27年的Jeff Dean希望开启新探索创立Discovery Loop外，内部调整的核心原因在于：前沿模型研发已经演变成一项吞噬全部精力的超大规模工程。
皮查伊指出，从数据筹备、算力调度、算法演进到统筹人类团队与模型Agent协同，前沿模型的迭代速度越来越快，需要一位专职负责人投入全部精力统筹日常运转。Koray在DeepMind任职超十年，且早在18个月前就已搬到湾区推进跨部门整合，由他接管模型研发管线水到渠成。而Demis则得以抽身，将精力集中到以Isomorphic Labs为代表的AI制药、生命科学、机器人等落地方向。
模型竞赛的加速也直接反映在财务账面上。今年夏天，Google出现了自IPO以来的首次自由现金流转负，同时继续上调2027年资本开支指引。
在解释为何敢于如此激进投入时，皮查伊算了一笔时间账：自研芯片和数据中心的前置周期极长，今天市场拿着完美信息所评估的算力水位，其实是两年前做出的决策，而去年这个时候，行业甚至还没有迎来Agent编程的爆发。如今，Google已经在规划2028至2029年的土地与电力储备。
他强调，今天的AI基建并非当年闲置的暗光纤泡沫，所有算力都在被真实消耗。除了C端和B端的需求信号以及云业务的长期合同积压外，底层研发突破也给了他足够的确定性。皮查伊在访谈中直接剧透：根据目前的内部基准和演进曲线，下一代Gemini 5.0将比现在的4.0 Argon强得多。
3.底层共享微虚拟机，太空TPU与量子计算时间表浮出水面
在产品架构上，Google正在抹平企业端与消费端Agent的界限。
皮查伊透露，无论是面向B端还是C端的Gemini Agent，底层跑的都是同一个模型、同一套运行框架和同一个运行时，唯一的区别在于注入的是企业上下文还是个人上下文。Google在企业端为每个任务自动分配独立微虚拟机、实现7×24小时安全运行的设计，正在同步下放到消费端。在交互界面上，Gemini现有的“对话”和“任务”将合并为单一输入框。
针对外界关心的Google搜索与Gemini何时合并，皮查伊表示不需要强行合并两款产品，云端24小时在线的Agent层会共享用户的个人上下文，自然接管跨端任务，让用户在搜索、Gemini应用与Gmail之间无缝切换。而在Web交互底层，他认为像人类一样操作界面的Browser Use与MCP、UCP等原生协议将长期共存。
以下是访谈原文：
1.Argon发布反响与算力受限下的访问策略
主持人：Gemini 4 Argon广泛发布已经过去一周左右了。我很想听听你对新模型早期市场反响的看法。
皮查伊：我们对此感到非常兴奋。在此之前我们在内部已经使用了一段时间，它绝对是一个世界级的尖端模型。现在的尖端模型在智能广度和深度上都非常全面。如果把智能水平在代理任务、知识工作和编程等维度上做一张热力图，我认为它在很多维度上都非常强大，而且竞争力极强。从Argon发布后的前72小时来看，我们处理了大约25万亿个Token，大约是3.0版本的3倍。由于正处于发布初期，我们在算力上确实受到了一些限制。来自客户的反馈非常棒，特别是在代码和代理能力方面。当然，我也清楚还有需要改进的地方。在尖端模型领域，你不断发布产品、收集反馈、然后推进下一步，这是我见过的最具动态变化的时期。
主持人：你觉得目前还有哪些领域是你们希望继续做好的？
皮查伊：我们希望在编程能力上继续精进。这是一个非常优秀的模型，我们的工程师在内部非常喜欢使用它，但我们深知在任何时刻都有可以做得更好的地方。发布之后，我们内部一直在持续推进，根据用户反馈进行行为层面的优化。
主持人：你刚才提到了算力限制。像谷歌这样的公司在面对这样的新模型时也会遇到算力瓶颈，这很有意思。Argon是一个大模型，运营成本显然很高。你能讲讲是如何决定对其进行访问限制的吗？目前它只面向API以及AI Pro和Ultra订阅用户开放，显然还没有覆盖到最广泛的用户群，对于这种规模的模型来说这确实是个挑战。你们是如何思考这种限制与发布策略的？
皮查伊：我们在3.0模型时期就积累了丰富的蒸馏经验，能够提炼出能力几乎同样强大但效率高得多的模型。我们目前正在开发这些版本，这能帮助我们尽可能让更多人广泛使用。同时，面对这个尖端模型，安全至关重要。我们在外部测试上投入了大量精力，例如通过美国政府的AI安全研究所项目，以及向网络安全研究人员提供早期访问权限，同时也在内部进行了广泛测试，随后才将其推向外部。正如你所说，这是一个大模型，但考虑到其规模，我们的推理速度相当快。从现在开始，我会持续扩大访问权限。我们在Gemini应用中首先面向所有Ultra和Pro层级的用户推出，随后又在Anti-Gravity应用中推出了Pro版本，并且我们正在开发可以服务更多用户的4.0系列蒸馏版本。
主持人：这就是传闻中的Carbon模型吗？
皮查伊：在任何特定时刻，我们都会有尖端模型的内部版本，其中一些只是代码代号，这与我们推出更高效版本的Argon并不矛盾。
2.安全测试、竞争格局与长期前沿定位
主持人：所以更高效的版本即将推出。关于安全性测试的过程，我很好奇你是怎么看待的。比较像Anthropic这样的私有尖端实验室与谷歌、Meta或大型云服务商在安全和隐私上的做法很有意思。听起来你们对这个模型的安全测试非常严谨，耗时似乎也比之前的模型更长，可以这么说吗？
皮查伊：绝对是这样的。鉴于其强大的代理能力和极其出色的编程水平，特别是在将其大规模部署到企业客户时，你需要确保万无一失。例如，除了模型对齐工作之外，更重要的是构建算力和系统来对运行中的模型进行实时监控。这种实时监控对于部署后的模型至关重要，能让你尽早发现任何不符合预期的行为并及时进行干预。为此，我们花时间构建了相应的系统、沙盒环境以及实时监控机制，并为客户提供所需的保护措施，确保他们清楚模型运行的位置、权限以及所需的遥测数据。
主持人：你有多关注新模型之间的这场拉力赛？我的意思是，我知道你肯定关注各种动态，看谁领先、谁落后。大家去年看到Gemini 3的时候都惊呼谷歌重回尖端行列了。现在看基准测试，Argon也确实名列前茅。但感觉大家的注意力总是在转移，看谁处于领跑位置。你对此有多关注？你有多鼓励团队去关注这些？很想听听你现在的想法。
皮查伊：这是一个非常好的问题。这是一个变化极快的时期，外界的声音非常多，我当然会关注X上的动态。获取外部的声音非常重要，因为这能让你了解哪些地方做得好，哪些地方需要改进。虽然我会观察所有这些外部反馈，但更重要的是，我们内部建立了一套非常严密的评估体系来评估模型。我们可以清晰地看到模型所处的性能曲线，从而在内部获得高质量的信号。我对团队的要求是，我们坚决致力于在极长期内保持在AI技术的最前沿。我们并非最近才涉足这一领域，而是长期专注于此，希望在十年后依然处于智能的最前沿。这意味着我们需要做好所有必要的工作，包括安全保障，这对于开发尖端智能至关重要。今年对我们而言，最重要的事情之一就是加快产品发布节奏。行业变化极快，尖端模型也在飞速演进。因此，我们今年所做的一大调整，就是在代理编程上取得更深层次的突破，并以此来优化我们的发布周期。从3.5、3.6、3.7到3.8 Flash，我们在大约3个月内完成了这一系列迭代，因此在4.0版本中，我们也期望保持并提升这种发布节奏。
3.代理能力跃升与2027生产力/RSI展望
主持人：Argon有一些方面让我很感兴趣。比如我看到它在大约一天多的时间里独立通关了《我的世界》，算力消耗据说还不到2000美元，这非常令人印象深刻。我很好奇这种级别的代理能力意味着什么。游戏是一回事，但你可以将其推演到很多潜在的事情上，比如高级决策和长周期任务。当你看到《我的世界》这个案例时，你是怎么想的？这如何影响你的长远规划？
皮查伊：在内部，我们已经见证了Argon带来的阶跃式生产力提升。自内部部署以来，它被应用到了许多大型代码库中。例如，在Chromium中，如果你想创建一个体验，比如拖拽标签页并将其放入AI体验中，同时保持其实时运行而无需重新加载，过去团队需要打通Chromium的所有底层架构。但现在，你可以直接让Argon去处理，它能够持久耐心地完成整个流程。我们在各个地方都看到了这样的例子。这让我感到兴奋的是，我认为2027年将是我们的生产力黄金时代，这得益于新一代模型及其与我们内部环境的深度融合。你还可以将这些模型深度应用到模型自身的研发中，也就是行业内所说的递归自我改进（RSI）。这些都是我们能够看到持续进展的实际领域，非常令人兴奋。《我的世界》中展示的能力表明，只要有足够的算力和资源，我们在许多领域的推进速度将会超越以往。
主持人：任何可验证的领域其实都可以，对吧？
皮查伊：确实如此，而且在其他领域，始终可以让人类参与到反馈循环中，不断推动进展。这感觉就像有无限的可能。这是一个只要你能想到、就能投入并取得进展的时刻，最近在数学领域已经看到了这一点。你本以为需要很长时间才能解决的事情，只要投入精力和资源，就能展现出戏剧性的进展，就像OpenAI在其最新模型中在数学领域所做的那样。因此，我认为我们在2027年将会看到更多这样的成果。
主持人：是的，我最近也听到谢尔盖·布林·（Sergey Brin）在谈论RSI并推动团队在这方面发力。谷歌目前在内部的RSI进展到了什么阶段？现在的研发有多少是在模型辅助下进行的？
皮查伊：目前的进展是，AI研究人员正在与模型进行深度的协同工作。模型会参与进来，自动推荐技术路线、评估下一个检查点等，因此模型和研究人员之间存在着健康的协作关系。虽然流水线中的某些环节可能离递归自我改进还有一定距离，但整个连续体正在发生位移。这就是为什么我认为思考我们如何开发尖端模型如此重要，并且我预计在未来12个月内，由于递归自我改进的应用，我们将会看到持续的加速发展。
4.Gemini 3.5 Pro的去向与编程能力待攻克方向
主持人：在模型话题结束前顺便问一句，我在上次年度开发者大会（Google I/O）上听你们提到了Gemini 3.5 Pro，但外界一直有疑问，那个模型后来怎么了？是直接变成了Argon，还是发生了什么？
皮查伊：当我们开发3.5系列时，我们同时在开发Flash和Pro。到了今年年中，我们内部认识到，模型在代理能力（特别是编程相关能力）以及它对整体进展的基础性作用上存在差距。因此我们希望在该领域加速攀登，而团队认为最快的攀登方式是使用Flash。这是一个更小、更高效的模型。这样你就能在其中获得更多、更密集的迭代循环。正是这种思路引导我们推出了3.5到3.8 Flash系列。等我们完成这些时，我们有机会将其应用到Pro版本中，但当时Argon的预训练已经完成了。所以是因为意识到3.5 Pro的生命周期会非常短，所以决定把所有的积累应用到Argon的后训练中。在算力和时间等资源受限的环境下，你只能做出这样的取舍。但这也给我们提供了一个机会，去向未来真正需要的地方提前布局。所以在4.0版本中，你们会看到不同的模型版本。Argon是我们的最大模型，也是目前我们对外发布的版本，但我们后续还会推出更多模型。同时，在此期间我们也在将Argon提供给Pro层级的用户。
主持人：目前在编程方面还存在哪些困难？你们仍在努力攻克的难关是什么？
皮查伊：倒谈不上是什么巨大的困难，只是尖端技术在不断演进。如果放在两个月前来看，Argon在编程的各个方面都堪称最前沿。所以这并不是说遇到了什么过不去的坎，而是你不断地给它设定更复杂的端到端代理任务，并且希望模型能够做得极其出色——就像一个世界级的优秀程序员团队那样，以极佳的品味、出色的代码可审查性以及高度的安全性来完成任务。因此这是一个不断发展的技术前沿。比如说幻觉问题，这永远是一个需要不断取得进展的前沿领域，它是当前最先进技术固有的一部分，但我们一直在努力将其减少。这既是一个机遇，也是一个挑战。
5.DeepMind领导层变动及其与谷歌的融合
主持人：举一个具体的例子。自从Google DeepMind发生一些领导层变动以来，我们还没聊过，我想向你请教一下。CEO Demis Hassabis退居二线成为首席科学家，同时也把更多精力投入到Isomorphic Labs中，对此我非常好奇；同时Corey被提拔为DeepMind的负责人；Google传奇人物Jeff Dean离开并创立了Discovery Loop。当时这一消息传出时，引发了大量媒体报道和对幕后情况的种种猜想。自那以后，我没有听你或任何相关人士谈及此事。我很想听听你的看法，或许你可以借此平息一些谣言：当时的考量是什么？是什么促成了这些变动？为什么会发生这样的变化？
皮查伊：当你看到人们在某个岗位上投入了一段时间后，自然会经历一些演变。
主持人：我记得Jeff在谷歌工作了27年。
皮查伊：整整27年。我感到非常自豪，我和Jeff共事了很长时间。他在谷歌的贡献具有开创性，我们也非常期待看到他在Discovery Loop能够做出什么成绩。首先我想说的是，他当时确实希望在人生某个阶段去做一些让自己感到兴奋、并且能和志同道合的人一起做的事情。至于内部的变动，我想说的是，我们正处于这样一个行业时期：尖端模型的开发是一项会吞噬一切的工作。过去我们有广泛的AI研究、AGI研究等等，但现在尖端模型的开发是全方位占据精力的。它正在快速推进，这是一项规模庞大的系统工程，从数据、算力、算法工作，到管理团队以及利用模型代理来完成任务，无所不包。因此我认为这本身就已经变成了一项至关重要的核心工作。
皮查伊：Demis其实已经思考这个时刻很久了。他特别关注两个方面：一方面是如何安全地度过这个尖端技术爆发的时期，另一方面是如何将其应用到解决其他问题上。他在AlphaFold中做的工作，以及我们现在在Isomorphic Labs所做的事情，都是如何将AGI应用于解决其他问题（比如医学、药物发现）的例证，还有机器人技术等等，这些最终会应用到所有领域。所以我想他希望把更多时间花在这些事情上。他一直对这些充满热情，而且他很早就涉足其中。
皮查伊：一方面，我们看到模型开发已经成为一个相当成熟的流程；另一方面，Corey在DeepMind也已经工作了十多年之久。因此我们看到了让Corey挺身而出的好时机。Demis仍然作为DeepMind的主席参与其中，但他也花更多时间参与AI安全领域的对话，同时思考如何更实际地将AI应用于科学和医疗等领域，未来还会涉足其他领域。
主持人：这是否意味着DeepMind与谷歌其他部门的关系发生了某种转变？它是在变得更加紧密融合吗？这一变动自然而然会产生哪些后续影响？
皮查伊：多年来，包括Demis掌舵时期在内，双方经历了一个自然的融合过程。例如，Gemini应用本身就是在DeepMind内部构建的。Corey大约在18个月前就已经搬到了湾区。双方一直有着紧密的集成，这对于DeepMind的算力支撑以及对谷歌其他业务的影响来说都是顺理成章的。因此，这其实是一个已经持续了很久的深度协作过程，可能外界只是把这次变动当作了一个被关注的节点；但对我们而言，这是一种早已在进行的深度耦合。
6.算力分配、前瞻规划与激进投资的底气
主持人：你们是如何管理和分配算力来支持长期研究投入的？DeepMind历史上以长期的前瞻性研究闻名，而我们现在正处于由近期前沿模型、Gemini应用、各类产品以及商业化主导的AI浪潮中。过去一两年里，算力的分配发生过变化吗？
皮查伊：问题在于，现在的赌注非常高，而且我们很幸运，能够看到前沿智能带来的非凡机遇——我们在许多产品和服务中拥有非常成功的消费者业务，同时又是全球顶级的云服务商之一，拥有提供这项技术的广阔机遇。因此，机遇空间非常巨大，我们需要提前做好规划。我们有一套相当严谨的方法来进行前瞻性规划。这些团队都在独立地进行规划，而所有这一切的基础是，我们长期以来一直是一家“AI优先”的公司。伴随这一战略而来的，是对投资算力的坚定承诺，无论是为了基础研究还是为了开发模型，这都是我们规划未来的基石。在此基础之上，各个业务部门（比如Thomas和云团队）会根据自身的长期规划、营收预测和市场机遇来制定需求，我们以此为他们规划算力。回过头来看，你总能找出一些可以做得更好的地方。放眼全球，没有任何一家公司能像我们一样深切感受到技术前沿推进的速度有多快。就在去年这个时候，代理编程的爆发甚至还没有发生。而且别忘了我们自己制造算力芯片，这需要很长的交货周期。所以在任何特定时刻，你都在承受着两年前做出的算力决策所带来的结果。但大家评估你时，依据的却是他们今天所掌握的完美信息。
主持人：然而你两年前就必须做出决断。
皮查伊：是的，必须在两年前做出决定。基于当时的信号做出判断，保持自信，抱持长期的乐观态度，并在投资上采取进取姿态。我认为当今世界上的任何一家公司，都不会有人拍着胸脯说他们在两年前的决策就百分之百完美。你必须去经历和处理这些时刻，这本就是这一行的一部分。我们计划在算力投资上持续保持最先进水平，以此来推动基础研究和前沿模型的开发。除此之外，我们还将有机会把技术应用到更多领域，我们计划尽最大努力去做好这一切。
主持人：也就是说，你们现在已经在为2028年和2029年做规划了？
皮查伊：是的，这取决于……比如如果你在规划土地、电力供应等资源，那确实已经是在看未来更远的时间了。
主持人：今年夏天发生了一个重大节点，谷歌自公司IPO以来首次出现了自由现金流为负的情况。华尔街的很多人都注意到了这一点，大家看到后都在惊叹：“谷歌正在继续加大激进投资的力度。”我记得你们还上调了2027年的资本支出预期，现在又在考虑2028和2029年。是什么支撑着你们做出这样的决心？特别是在今年夏天谷歌自由现金流转为负数的那种时刻，这绝非一个小决定，你们是如何思考这一点的？是什么给了你们这份笃定？
皮查伊：我们所看到的强劲且持久的需求信号。要做出这样的举措，你需要在多个层面具备信心。首先，我们很幸运，能够从极其多元化的机遇组合中获得早期信号——从消费端到企业端，我们都看到了强劲的需求信号。
主持人：你们掌握的市场信号可能比地球上任何一家公司都要多。
皮查伊：我以前没从这个角度想过，但你这么说确实有道理。因此我们不仅拥有强劲的外部需求信号，还拥有来自谷歌自身的内部需求——作为一个处于前沿的企业，我们自身如何使用和采用AI。你可以亲眼看到能力前沿的进展，能够在脱离需求信号的情况下，以客观的方式去评估能力前沿的发展状况，看看性能曲线看起来如何，比如缩放定律（scaling laws）、新模型与旧模型的对比、我们能够解决的任务复杂度、以及内部的评测基准。
皮查伊：因此你可以获得第一手的直观感受，也能从研究突破中感受到：比如Gemini 5.0会不会比目前的Argon好得多？答案是肯定的。所以你在技术栈的各个层面都能获得这些信心信号。此外，我们还在签署拥有长期承诺收入的耐久合同。这绝不仅仅是盲目的希望，我们有实际的业务订单作为支撑。把所有这些因素综合起来，你就会以一种负责任且纪律严明的方式全力投入，而这正是我们正在做的。
主持人：听你满怀信心地谈论这些信号令人印象深刻，我相信这些判断。但与此同时，人们也在试图应对各种担忧——比如Corey的变动，比如对新云（NeoClouds）的担忧，对这些基础设施融资的担忧，以及它们的脆弱性。放眼整个市场（不仅仅是谷歌，尽管谷歌参与其中许多事务），你有没有感到任何让人担忧的迹象，比如市场是否出现了过度透支？
皮查伊：我刚才回答上一个问题时，显然是从谷歌自身的角度出发的。至于更广阔的市场，每当你经历这样一个机遇巨大、回报丰厚且有大量玩家涌入的时刻，市场中自然会出现活力与波动，我认为这是这种时期不可分割的一部分。正因如此，我极力将精力集中在我能够掌控、能够看清的信号上——比如需求信号、以及我们在业务合同和订单积压等方面清晰的业务前景，并围绕这些来进行规划。你需要以这样一种方式来设计公司：即便市场出现某些波动，其长期发展弧线对公司而言依然是向好的。这就是为什么保持长远眼光如此重要。如果你回顾过去的五年，疫情曾推动了向数字化的大规模转变，后来其中一部分需求有所回调。
皮查伊：但对于大多数这样做的公司而言，顺应并推动这一转变的投资是被证明是明智的，它们也因此取得了更好的成绩。所以你必须理解长期的长期趋势，并进行投资以抓住转型机遇。事实上，我们过去几年中很大一部分员工规模的增长都是为了谷歌云，而我们在云业务规模化之后正享受到随之而来的红利。我们进入云市场的时间相对较晚，但在一段长达6到8年的时间里，我们采取了非常激进的投资姿态，而这正是帮助我们度过当前这一关的关键所在。我认为我对当前AI浪潮的看法也是如此。
7.算力受限何时缓解、AI辅助决策与个人Agent战略
主持人：你能否在未来的某个时间节点看到，我们不再受困于算力限制？有没有这样一个时刻是你能够展望到的？
皮查伊：我有点犹豫要不要这么说，因为——
主持人：是的，你得小心点，因为如果这么说大家会吓坏的。
皮查伊：不，我的意思是，我们生活在一个全球经济以极其复杂的方式交织在一起的时代，任何一个孤立的事件都可能产生无处不在的“蝴蝶效应”。这就是我们在两件事之间进行区分的原因。我并不是说从今以后一切都会一帆风顺——生活从来都不是那样的。经济会增长，也会有波折，我们需要经历业务周期。所有这些都是世界运转中不可避免的组成部分。
主持人：而且我认为，没有人会认为这就像过去的泡沫——这绝不是当年的光纤泡沫。我们在切切实实地使用所有这些技术。
皮查伊：我们目前处于严重的算力受限状态，但更重要的是，我们今天已经在以切实的、具备实际价值的方式在使用这些技术。我们看到了明确的前景，知道未来这一切都将运行得更好。但这并不意味整个行业或经济体就不会出现过载，或者遭遇不可控的系统性冲击。所以这是两码事。正因如此，我将精力集中在我们可以掌控的事情上：利用我们长期的视角、多年的经验去更好地预测、建模和投资，并以相当有纪律的方式将其应用到我们所拥有的广阔机遇中。我把精力和关注点放在这些事情上，而不是去纠结市场是否存在过度狂热。
主持人：我听说过你曾表示，希望AI能在谷歌内部的算力分配上为你提供更多帮助。我很想知道，在过去几个月里，AI有参与进来吗？你是否觉得它帮助你做出了更好的决策？或者说，在如此重大且关键的事情上，你会把某些决定交由模型来做吗？
皮查伊：我从两个方面来回答这个问题。第一，鉴于这件事事关重大，所有决策都严格由人参与（Human-in-the-loop）来做出，包括我自己也深度参与其中。但第二，对于我个人而言，我也部署了AI Agent来协助我。比如有一个并不复杂但非常实用的场景：在我们这样的业务规模下，仅仅是了解模型研发的最新进展，每天就有预训练、中训练、后训练数据等各种沟通群，如果纯靠自己去跟进过去24小时发生的事情，每天轻松就会花掉一小时。但现在我运行着Agent，它会自动帮我汇总并呈现这些信息，如果需要，我依然可以随时深入了解。同样地，也有Agent能够实时监控我们各个产品线的算力使用情况、效率和利用率，并为我提供决策输入和洞察。我们正是通过这些方式将AI融入到日常工作中。当然，也有极其优秀的人员在处理这些工作。AI让我们每个人都能更好地完成本职工作，这就是我目前使用它的方式。
主持人：所以近期还不会出现由AI完全决定算力分配的情况？
皮查伊：早在五年前，我们就已经在用AI来优化数据中心的能耗了。因此在做出算力分配决策时，AI所提供的输入本身就是核心要素之一。
主持人：我本来还想帮你甩个锅，比如说是模型说的，不是我说的。
皮查伊：这是一个渐进的过程。等模型发展到能够完全承担那项决策时，你自身就已经开始投身于下一个层面的问题了。我就是这样看待它的。
主持人：行业目前正在展开一场关于个人Agent（Personal Agent）的竞赛，比如Muse和Instinct都做得很好。我记得在上次I/O大会上，你们展示了Gemini Spark。你认为谷歌在面向消费者的Agent领域是否需要采取更多行动？我知道你们最近推出了Gemini Workspace，让Agent具备邮箱和Workspace身份等，但在消费者赛道上，尽管现在还处于早期，但我猜你一定希望谷歌能占据更大的份额。
皮查伊：这是一个极其重要的焦点领域。事实上，我们为Gemini设定的目标，就是让它成为一个主动、强大且个性化的助手。这是我们打造Gemini的核心方向。你在I/O大会上看到的Spark正是朝着这个方向迈进。当然，我们在近期的企业级活动中，也在Gemini Agent的语境下讨论了更多。需要明确的是，企业级和消费级，它们底层使用的是同一个模型、同一套架构和运行环境，区别仅在于处理的是企业数据还是个人上下文。我们在发布会上宣布的所有企业级功能，都具备世界级的安全治理能力——每一个任务本质上都会配置一个微型虚拟机（microVM）来安全地24小时不间断运行，并提供实时监控，从而为企业提供完善的治理、控制和审计能力。我们正在把所有这些能力带给普通消费者。你现在已经可以在Gemini中看到聊天（Chat）和任务（Task）功能。我们近期的愿景是将其整合为一个统一的入口，让聊天与任务之间的界限自然衔接——哪些任务需要持久化运行，哪些需要定时调度，以及如何将结果反馈给你。从我们观察到的Gemini用户使用数据来看，这种趋势已经非常明显：越来越高比例的查询呈现出明确的代理（Agentic）特征。在接下来的几个月里，我们将把这些能力更广泛地提供给用户。此前我们主要面向Ultra和Pro用户推出了这些功能，接下来我们会进一步扩大覆盖面。全新的Argon模型带来了阶跃式的提升，作为消费者，你未来不仅能在Gemini应用里与它对话，还能在Gmail等各种上下文环境中调用它。我们在搜索等各个层面都在对此进行深度投入。
8.Agent时代的协议、界面与搜索/Gemini融合
主持人：我一直想问你一个问题，特别是考虑到你过去掌管Chrome的经验：在我们当前所处的时代，所有这些Agent都在像人类一样使用浏览器来完成任务，比如 Muse 就是
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。
KeyPoints
Over 72 hours, nearly 25 trillion tokens were swallowed up, three times more than the previous generation of Gemini 3.0 in the same period.
This is Gemini 4 Argon's first report card from Google's latest flagship model. Even the rich Google had to re-introduce rate limit in the face of a sudden influx of calls.
In a recent interview with the tech podcast Sources, Google CEO Sanddar Pichay took back details of Gemini 4 Argon ' s publish, explaining publicly for the first time why Gemini 3.5 Pro disappeared mysteriously and the reasons behind the changes in DeepMind ' s core management.
1. Cut off 3.5 Pro's direct sprinting Argon, internal RSI stage of development
Over the past few months, the outside world has been curious about where Gemini 3.5 Pro, mentioned in the I/O Assembly, went. In an interview, Pichai gave the answer: the team voluntarily abandoned Agent in order to rob him of his programming capacity.
In mid-year, Google, in training the 3.5 series, found that there was a clear gap in the model on the Agenic Coding. The team judged the use of smaller and more efficient Flash versions for high frequency iteratives to be the fastest route to complete the panels. Thus, Google pushed intensively from 3.5, 3.6, 3.7 to 3.8 Flash over a period of about three months.
Gemini 4.0 Argon pre-training was completed when the technology package ran and was ready for application to Pro. Taking into account the extremely short life cycle of the publish3.5 Pro at this time, and in an environment of limited computing, the team decided to validate all the successful technology accumulation of Flash directly into the post-training phase of Argon.
This decision directly contributed to Argon's leap forward on long-range tasks. In an external test, Argon cleared My World on its own after more than a day, at less than $2,000. Within Google, it has been pointed at the Chromium Code Library, which has millions of line codes.
As for the anecdotal Carbon model, Pichai clarified that it was only an iterative code for the internal front model. The team is currently building on the distillation of the Argon propulsion model and developing a light version with nearer but lower cost to make it accessible to more users.
As the new model is integrated into the internal development environment, Pichai sees 2027 as the golden age of productivity. He revealed that Google had entered the step-by-step self-improvement (RSI) phase, where researchers were working with the depth of the model: the model was directly involved in the development of training formulas and the automatic assessment of the next inspection point. Pichai himself is using the in-house Agent 24-hour development of pre-, mid-, post- and data sets to save approximately one hour of synchronization of information per day and to monitor real-time the efficiency of the use of the units by accessing the telemetry data by Agent.
2. Start with DeepMind, the game behind free cash flow.
For the first time, Pichay gave an official explanation for the recent popular DeepMind personnel shock: the departure of legendary engineer Jeff Dean, Demis Hassabis, chief scientist and chairman of Deepmind, Koray Kavukcuoglu, who had taken charge of DeepMind.
In addition to the 27-year-old Jeff Dean’s desire to launch a new exploration to create Discovery Loop, the central reason for internal adjustments is that front-line model development has evolved into a massive project that consumes all of its energy.
Pichai noted that from data preparation, computing and algorithmic evolution to the integration of human teams with model Agent, the frontier model is becoming more and more dynamic, requiring a dedicated head of staff to devote all his or her energies to the integration of day-to-day operations. Koray worked in DeepMind for more than 10 years and moved to the Bay Area 18 months ago to promote cross-sectoral integration, where he took over the development of a model pipeline to the canal. Demis, on the other hand, was able to pull out and concentrate on AI pharmaceuticals, life sciences, robotics, etc., represented by Isomorphic Labs.
The acceleration of the model competition is also directly reflected in the financial books. This summer, Google showed its first free cash flow negative since IPO, while continuing to adjust its 2027 capital expenditure guidelines upwards.
In explaining why they are so aggressive, Pichai has set aside a time line: the pre-manufacturing cycle of self-study chips and data centres, the calculus level assessed today by the market with the perfect information, was the decision taken two years ago, when the industry was not even ready for the outbreak of Agent’s programming. Today, Google is already planning land and electricity reserves from 2028 to 2029.
He stressed that today ' s AI infrastructure was not an idle dark fibre-optic foam, and that all the computing power was being consumed. In addition to the demand signals at the C and B sides and the backlog of long-term contracts for cloud operations, he was given sufficient certainty by the bottom R&D breakthroughs. Pichai was directly involved in the interviews: the next generation of Gemini 5.0 would be much better than the current 4.0 Argon based on current internal benchmarks and evolution curves.
3. Substrate shared microviral machines, space TPU and quantum schedule surfaces
In the product architecture, Google is levelling the boundaries between the enterprise and the consumer-end Agent.
Pichai disclosed that Gemini Agent, whether for B or C, ran the same model, the same operating framework and the same running time, the only difference being the injection of the enterprise context or the individual context. Google is simultaneously decentralizing to the consumption end the design of auto-allocation of an independent micro-virtual machine for each mission at the enterprise level to achieve 7x24 hours of safe operation. On the interface, Gemini's existing “dialogue” and “task” will be merged into a single input box.
In response to the question of when Google searches of outside interest were merged with Gemini, Pichai stated that there was no need for a forced merger of the two products, and that the 24-hour on-line Agent layer would share the user ' s personal context, which naturally would take over the cross-cutting task and allow the user to switch seamlessly between search, Gemini applications and Gmail. On the other hand, at the bottom of the Web interface, he believed that primary agreements such as Browner Use, which operated interfaces like humans, would coexist with MCP, UCP and others.
Here are some of the interviews:
1. Argonpublish Reaction and Calculator-limited access strategy
Moderator: Gemini 4 Argon has been a week or so since extensive publish. I'd like to hear your views on the early market repercussions of the new model.
Pichay: We are very excited about this. We have been using it internally for some time, and it is definitely a world-class cutting-edge model. The current cutting-edge models are very comprehensive in terms of intellectual breadth and depth. I think it is very powerful and highly competitive in many dimensions, if one tries hard at the level of intelligence in terms of proxy tasks, knowledge work and programming. From the first 72 hours of Argonpublish, we handled about 25 trillion Tokens, about three times the version of 3.0. Since we are in the early stages of publish, we do have some limitations on our ability to calculate. Feedback from clients was excellent, particularly in terms of code and agency capabilities. Of course, I also know that there is room for improvement. In the field of cutting-edge models, you keep publish products, collect feedback and then push next step, the most dynamic period I've ever seen.
Moderator: What areas do you think you want to continue to do?
Picay: We want to continue to improve on programming capabilities. This is a very good model, which our engineers very much like to use internally, but we know that there is room for doing better at all times. Since publish, we have been moving forward internally to optimize the behavioural level based on user feedback.
Moderator: You just mentioned the limitations of arithmetic. It's interesting that companies such as Google also encounter math bottlenecks when facing new models like this. Argon is a LLM and the operating costs are clearly high. Can you tell me how you decided to limit her access? It is currently open only to API and AI Pro and Ultra subscribers and apparently has not reached the widest range of users, which is a real challenge for models of this size. How did you think about this limitation and publish strategy?
Pichay: We have accumulated a wealth of distillation experience during the 3.0 model period, capable of extracting models that are almost equally powerful but far more efficient. We are currently developing these versions, which will help us to make them as widely available as possible. At the same time, security is crucial in the face of this cutting-edge model. We have invested a great deal of effort in external testing, such as through the United States Government ' s AI Security Research Institute project and the provision of early access to permission for cybersecurity researchers, as well as extensive in-house testing, which was subsequently pushed outside. As you said, this is a LLM, but given its size, our reasoning is fast. From now on, I will continue to expand my visit to permission. We started with all users at the Ultra and Pro levels in the Gemini application, followed by the Pro version in the Anti-Gavity application, and we are developing a distilled version of the 4.0 series that will serve more users.
Moderator: Is this the legendary Carbon model?
Picay: At any given point, we have internal versions of cutting-edge models, some of which are code codes, which are not in contradiction with the more efficient version of Argon.
2. Safety tests, competitive patterns and long-term frontier positioning
Moderator: A more efficient version is about to be launched. I'm curious about the security test process. It is interesting to compare private cutting-edge laboratories like Anthropic with Google, Meta or large cloud service providers in terms of safety and privacy. Sounds like you guys have a very rigorous safety test for this model, and it seems to take longer than the previous model.
Pichay: Absolutely. Given its strong agency capacity and excellent programming levels, especially when deploying it to corporate clients on a large scale, you need to make sure that everything is done. For example, in addition to model alignment, it is more important to construct algorithms and systems for real-time monitoring of operational models. Such real-time monitoring is essential for post-deployment models to enable early detection of any undesired behaviour and timely intervention. To that end, we have taken time to construct systems, sandbox environments and real-time monitoring mechanisms, and to provide clients with the necessary protection measures to ensure that they know where the models are running, permission and the required telemetry data.
Moderator: How much do you care about the pull race between the new models? I mean, I know you must be concerned with the dynamics of who leads and who lags behind. Last year, when you saw Gemini 3, you were shocked that Google was back on top. Argon is certainly at the top of the benchmark. But it feels like attention is always shifting to see who's running. How much do you care? How much do you encourage the team to look at this? I'd love to hear your thoughts now.
PJ: This is a very good question. This is a very fast time for change, with a lot of outside voices and, of course, I am following developments on X. Access to external voices is important because it gives you an idea of where it works well and where it needs to be improved. While I will observe all this external feedback, more importantly, we have a very rigorous internal assessment system in place to evaluate the model. We can clearly see the performance curve in which the model is located, thereby obtaining high-quality signals internally. My demands on the team are that we be firmly committed to staying at the forefront of AI technology in the very long term. We are not only recently involved in this area, but have focused on it for a long time, hoping to remain at the forefront of intelligence in a decade. This means that we need to do all the necessary work, including safety and security, which is essential for developing cutting-edge intelligence. One of the most important things for us this year is to speed up the publish pace of the product. Industry changes very rapidly and cutting-edge models are evolving at a fast pace. Thus, one of the major adjustments we made this year was to achieve a deeper breakthrough in the proxy programming and thereby optimize our publish cycle. From 3.5, 3.6, 3.7 to 3.8 Flash, we have completed this series within approximately three months, so in version 4.0, we also expect to maintain and enhance this publish rhythm.
3. Growth of agency capacity versus 2027 productivity/RSI outlook
Moderator: Argon has some aspects that interest me. It is very impressive to see, for example, that it cleared My World on its own for about a day or so, with a calculated consumption of less than $2,000. I wonder what this level of agency means. Game is one thing, but you can push it to a lot of potential things, like high-level decision-making and long-cycle tasks. What do you think when you see the case of My World? How does this affect your long-term planning?
Pichay: Internally, we have witnessed the step-up productivity that Argon has brought. Since its internal deployment, it has been applied to many large coding repositories. For example, in the Chromium, if you want to create an experience, such as dragging the tab and putting it into the AI experience, while keeping it running on time without reloading, the team in the past needed to access all of Chromium ' s sub-structures. But now, you can just let Argon handle it, and it can do the whole process with patience and patience. We have seen such examples everywhere. It excites me that 2027 will be the golden age of our productivity, thanks to a new generation of models and their deep integration with our internal environment. You can also apply these models to the development of the models themselves, that is, to self-improvement in industry. These are practical areas where we can see continued progress, and they are very exciting. The ability shown in My World shows that we will move faster in many areas than we have in the past, provided that we have sufficient capacity and resources.
Moderator: Any verifiable field actually works, right?
Pichai: Indeed, and in other areas, human participation in the feedback cycle can continue to drive progress. It feels like there's an infinite possibility. This is a time for you to think about, to invest in and to make progress, as has been seen recently in the field of mathematics. You thought it would take a long time to solve things, and with energy and resources, you could show dramatic progress, as Open AI did in his latest model in mathematics. I therefore believe that we will see more such results in 2027.
Moderator: Yes, and recently I heard Sergei Brin talking about RSI and pushing the team to work on it. What stage has Google made in the RSI? How much of this research is being done with modeling?
Pichai: The progress is that AI researchers are working in depth with the model. Models are involved, automatically recommend technical routes, assess the next checkpoint, etc., so there is a healthy collaborative relationship between models and researchers. While some links in the water flow line may be some distance away from retrogressive self-improvement, the entire continuum is shifting. That is why I think it is so important to think about how we develop cutting-edge models, and I expect that in the next 12 months we will see sustained accelerated development as a result of the application of incremental self-improvement.
4. Gemini 3.5 Pro ' s routing and programming capability to be targeted
Moderator: By the way, before concluding the issue of the model, I heard you mention Gemini 3.5 Pro at the last Annual Developer Conference, but there's always a question about what happened to the model. Did it just turn into Argon, or did it happen?
Pichay: When we developed the 3.5 series, we were also developing Flash and Pro. By mid-year, we recognized that there was a gap in the model ' s proxy capacity (especially programming-related capacity) and its underlying role in overall progress. That is why we want to accelerate the climb in this area, which the team considers the fastest way to climb is to use Flash. This is a smaller and more efficient model. So you can get more and more incisive in it. It's this idea that led us to the 3.5 to 3.8 Flash series. By the time we're done, we have an opportunity to apply it to the Pro version, but Argon's pre-training was completed. Therefore, it was decided to apply all accumulations to Argon ' s post-training because of the realization that 3.5 Pro ' s life cycle would be very short. You can only make such trade-offs in resource-constrained environments like calculus and time. But it also provides us with an opportunity to move ahead where we really need to go. So in version 4.0, you'll see different models. Argon is our top LLM and now our external version of publish, but we'll have more models later. Meanwhile, in the meantime, we are also making Argon available to users at the Pro level.
Moderator: What are the current programming difficulties? What's the stakes that you're still working on?
Picay: Not to mention the enormous difficulties, but the cutting edge of technology is evolving. If seen two months ago, Argon was at the forefront in all aspects of programming. So that is not to say that there has been some sort of trouble, but that you constantly set it up with more complex end-to-end proxy tasks, and you hope that the model will do an excellent job — like a world-class team of good programmers, with excellent taste, good code reviewability and high security. This is therefore an evolving technological frontier. For example, the problem of hallucinations, which is always a frontier area requiring continuous progress, is an inherent part of the most advanced technology of the time, but we have been trying to reduce it. This is both an opportunity and a challenge.
5. DeepMind leadership change and integration with Google
Moderator: Give a specific example. We haven't talked since Google DeepMind had some leadership changes, and I would like to ask you something. The CEO Demis Hassabis retreated to the second line as the leading scientist, and I was curious to know that more energy was devoted to Isomorpic Labs; Corey was promoted to head of DeepMind; and Google legend Jeff Dean left and created Discovery Loop. At that time, the news had triggered a large number of media reports and speculations about what was behind the scenes. Since then, I have not heard from you or from anyone concerned. I'd like to hear your opinion, and perhaps you could use it to calm down some rumours: What were the considerations at the time? What contributed to these changes? Why is this happening?
Picay: When you see people in a position for some time, there's a certain evolution.
Moderator: I remember Jeff working at Google for 27 years.
Pichai: For 27 years. I'm very proud that I worked with Jeff for a long time. His contribution to Google is groundbreaking, and we very much look forward to seeing what he can do in Discovery Loop. First of all, I would like to say that he did want to do something at some stage of his life that he was excited about and able to do with people who were like-minded. As for internal changes, I would like to say that we are at an industry stage in which the development of cutting-edge models is a work that will devour everything. We used to have a wide range of AI, AGI, and so on, but now the development of cutting-edge models is all-out. It is advancing rapidly and is a large-scale system engineering effort, ranging from data, computing, algorithmic work to management teams and the use of model agents to accomplish tasks. That is why I believe that this in itself has become a vital core.
Pichay Demis has been thinking about this for a long time. He paid particular attention to two aspects: how to safely survive this period of sophisticated technology and how to apply it to other issues. What he did in AlphaFold, and what we do now in Isomorphic Labs, is an example of how AGI can be applied to other problems (such as medicine, drug discovery) and robotics, and so on, and so on. So I think he wants to spend more time on these things. He's been passionate about these things, and he's been involved early.
Picay: On the one hand, we see that model development has become a fairly mature process; on the other hand, Corey has been working in DeepMind for over a decade. So we saw a good time for Corey to step up. Demis is still involved as President of DeepMind, but he has also spent more time in dialogue in the area of AI security, while reflecting on how to apply AI more practically in areas such as science and medicine, and in other areas in the future.
Moderator: Does this mean that there has been some change in the relationship between DeepMind and the rest of Google? Is it getting closer? What follow-up effects would this change naturally have?
Pichai: Over the years, including the Demis period, the parties have undergone a natural process of integration. For example, Gemini applications are themselves built within DeepMind. Corey moved to the Bay Zone about 18 months ago. Both sides have been closely integrated, which is logical for DeepMind ' s computing support and for its impact on other Google operations. Thus, this is a deep collaborative process that has been going on for a long time, and perhaps the outside world is simply looking at the change as a node of concern; but for us, it is an already ongoing in-depth convergence.
6. Capital allocation, forward planning and the bottom line for radical investment
Moderator: How did you manage and allocate the math to support long-term research inputs? DeepMind has a long history of forward-looking research, and we are now in an AI wave dominated by recent frontier models, Gemini applications, products and commercialization. Has there been any change in the distribution of arithmetic in the last year or two?
Picay: The problem is that the stakes are very high and that we are fortunate to see the extraordinary opportunities that frontier intelligence offers — that we have very successful consumer operations in many products and services, and that we are one of the top cloud service providers in the world, with great opportunities to provide this technology. So, the opportunities are enormous, and we need to plan ahead. We have a fairly rigorous approach to forward planning. These teams are planning independently, all on the basis that we have long been an "AI first" company. This strategy is accompanied by a strong commitment to the calculus of investment, whether for basic research or for modelling, which is the cornerstone of our planning for the future. On this basis, each business unit (e.g. Thomas and the Clouds Team) develops its needs based on its own long-term planning, revenue forecasting and market opportunities, by which we plan for them. Looking back, you can always find something better. Looking at the world, no company can feel as deeply as we do how fast the technology frontier is advancing. Just this time last year, the surrogate programming outbreak has not even happened. And don't forget we make our own calculus chips, which takes a long delivery cycle. So at any given moment, you are bearing the fruit of the calculus that you made two years ago. But you are assessed on the basis of the perfect information they have today.
Moderator: But you had to make a decision two years ago.
Picay: Yes, a decision had to be made two years ago. It was judged on the basis of the signals of the time, remained confident, had a long-term sense of optimism and had a proactive attitude towards investment. I believe that no company in today ' s world will have anyone who says that they made 100 per cent perfect decisions two years ago. You have to go through and deal with these moments, which are part of this. We plan to sustain the state-of-the-art level of investment in computing as a means of promoting basic research and frontier model development. Beyond that, we will have the opportunity to apply technology to more areas, and we plan to do our best to do all that.
Moderator: So now you're planning for 2028 and 2029?
Picay: Yes, that depends on... for example, if you're planning resources like land, electricity, and so on, it's really going to be a long way to go.
Moderator: There was a major node this summer, and Google has had a negative free cash flow for the first time since IPO. A lot of people on Wall Street have noticed this, and what you're looking at is amazing: “Google is continuing to scale up radical investment.” I recall that you also increased your expectations of capital expenditure in 2027 and are now considering 2028 and 2029. What sustains your determination? Especially at a time when Google's free cash flow turned negative this summer, it's not a small decision. How do you think about it? What gives you this certainty?
Pichay: The strong and sustained demand signals we see. You need confidence at many levels to do this. First of all, we are fortunate to be able to get early signals from an extremely diverse combination of opportunities — from the consumer to the business end — and we all see strong demand signals.
Moderator: You may have more market signals than any company on Earth.
PJ: I didn't think about it that way before, but that makes sense. We therefore have not only strong external demand signals, but also internal needs from Google itself — how we use and adopt AI ourselves as a front-line enterprise. You can see for yourself the progress of the capacity frontier, and you can assess the development of the capacity frontier in an objective manner, away from demand signals, and see how the performance curve looks, such as the scaling laws, the comparison of new models with old models, the complexity of tasks we can solve, and the benchmarking of internal assessments.
Picay: So you can get a first-hand visual feeling, and you can feel a breakthrough in research: would Gemini 5.0 be much better than Argon? The answer is yes. So you get these signs of confidence at all levels of technology. In addition, we are in the process of signing durable contracts with long-term commitments. This is not just blind hope; we are backed by actual business orders. Combining all of these elements, you will be fully engaged in a responsible and disciplined manner, which is exactly what we are doing.
Moderator: I believe that you are impressed to talk about these signals with confidence. At the same time, however, attempts are being made to address concerns — such as the changes in Corey — about NeoClauds, the financing of these infrastructure, and their vulnerability. Look at the whole market (not just Google, although Google is involved in many of these matters), do you have any worrying signs, such as whether the market is overstretched?
Picay: The answer I just gave to the previous question was clearly from the perspective of Google itself. As for the wider market, every time you experience such a moment of great opportunity, high returns and a large influx of players, there will naturally be dynamism and volatility in the market, which I think is an integral part of this period. That is why I have tried to focus on and plan around signals that I can control and that I can see clearly — for example, demand signals, and our clear business prospects in terms of business contracts and backlogs of orders. You need to design the company in such a way that, even if there are certain fluctuations in the market, its long-term development arc remains good for the company. That is why long-term vision is so important. If you look back over the past five years, the epidemic has contributed to a large-scale shift towards digitalization, and some of its demands have since been reversed.
Pichai: But for most companies that do so, investments that adapt to and drive this shift have proved wise and have led to better results. So you have to understand the long-term trends and invest to seize the opportunities for transformation. In fact, much of the increase in the size of our staff over the past few years has been for Google Clouds, and we are enjoying the dividends that follow when the cloud is scaled up. We entered the cloud market relatively late, but for a period of six to eight years, we took a very aggressive investment stance, which is the key to helping us get through this. I think that's what I think about the current wave of AI.
7. Timing of constraints, AI support decision-making and individual Agent strategy
Moderator: Can you see, at some point in the future, that we are no longer constrained by arithmetic? Is there a moment you can see?
PJ: I'm a little hesitant to say that because...
Moderator: Yes, you have to be careful, because if you say so, you'll freak out.
Picay: No, I mean, we live in an era when the global economy is interwoven in extremely complex ways, and any isolated event can have an pervasive “butterfly effect”. That is why we distinguish between two things. I am not saying that everything will be fine from now on — life is never like that. There will be economic growth and setbacks, and we need to go through business cycles. All of these are inevitable components of the functioning of the world.
Moderator: And I don't think anyone would think it's like the past bubble — it's not the fiber-optic bubble of that year. We are using all of these technologies on the ground.
Picay: We are currently in a state of severe computational constraint, but more importantly, we are using these technologies today in a practical and meaningful way. We see clear prospects and know that all of this will work better in the future. This does not mean, however, that an entire industry or economy will not be overloaded or subject to uncontrolled systemic shocks. So it's two different things. That is why I have focused my efforts on what we can do: use our long-term perspective, years of experience to better predict, model and invest, and apply it in a rather disciplined way to the vast opportunities we have. I have focused my energy and attention on these things, not on the existence of excessive fanaticism in the market.
Moderator: I have heard you say that I hope that AI will help you more in the distribution of power within Google. I wonder if AI has been involved in this in the last few months? Do you think it helped you make better decisions? Or are you going to leave some decisions to models for something so big and crucial?
Pichay: I answer this question in two ways. First, given the importance of the matter, all decisions are strictly made with the participation of people (Human-in-the-loop), including myself. But secondly, for me personally, I have deployed AI Agent to assist me. For example, there is an uncomplicated but very practical scenario: On a business scale like ours, simply keeping abreast of the latest developments in model development, there are a variety of communication groups each day, such as pre-training, middle- and post-training data, and it takes an hour to follow up on what has happened in the past 24 hours on its own. But now I'm running Agent, which will automatically help me aggregate and present this information, and I'll still be able to understand it if I need it. Similarly, Agent is able to monitor in real time the computing, efficiency and utilization of our product lines and to provide me with input and insight into decision-making. It is in these ways that we have integrated AI into our daily work. Of course, there are also very good people dealing with these tasks. AI allows us all to do our job better, which is how I use it now.
Moderator: So there will be no complete distribution of power by AI in the near future?
Picay: Five years ago, we were already using AI to optimize energy consumption in data centres. Thus, the input provided by AI is itself one of the core elements in the decision-making on the allocation of power.
Moderator: I was going to throw a pot for you, like a model, not me.
Pichai: This is a gradual process. And when the model develops to be able to take full responsibility for that decision, you yourself are starting to get into the next level of the problem. That's how I see it.
Moderator: The industry is currently running a competition on individual Agent, such as Muse and Instinct. I remember the last time you showed Gemini Spark. Do you think Google needs to do more in the consumer-oriented Agent area? I know you've recently launched Gemini Workspace to give Agen the e-mail and workspace identities, but on the consumer track, although it's still early, I'm sure you want Google to take a bigger share.
Pichai: This is an extremely important focus area. In fact, the goal we set for Gemini was to make it a proactive, powerful and personal assistant. This is the central direction we've been building Gemini. It is in this direction that you see Spark at the I/O Assembly. We have, of course, discussed more in the context of Gemini Agent in our recent corporate activities. It needs to be made clear that at the enterprise and consumer levels, they use the same model, the same set of structures and operating environments at the bottom, with the difference between processing business data and individual context. All the enterprise-level functions that we announced at the publish conference have world-class security governance capabilities — each mission is essentially equipped with a microviral (microVM) that operates safely around the clock and provides real-time monitoring, thus providing companies with sound governance, control and auditing capabilities. We are bringing all these capabilities to ordinary consumers. You can now see Chat and Task in Gemini. Our vision for the near future is to integrate it into a single portal, allowing for a natural link between chat and tasks — which tasks need to be sustained, which need to be scheduled, and how to feed back to you. From the data we have observed on the use of Gemini users, this trend is already clear: the increasing proportion of queries shows a clear proxy. In the coming months, we will make these capabilities more widely available to users. These features were introduced primarily for Ultra and Pro users, and we will further expand coverage. The brand-new Argon model has brought a leap forward, and as a consumer, you will be able to talk to it not only in Gemini applications, but also in various context environments such as Gmail. We are deeply involved in this at all levels, including search.
8. Agent-era protocols, interfaces and search/gemini integration
Moderator: I've been meaning to ask you a question, especially in view of your past experience in running Chrome: In our current era, all these Agents are using browsers to accomplish their tasks like humans, like Muse.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.