DeepMind's Kilpatrick Says AI Models Will Absorb Coding Agent Harnesses

DeepMind 的 Kilpatrick 称 AI 模型将吸收编程 Agent 框架 Google DeepMind 的 Logan Kilpatrick 认为，专用 AI Agent 脚手架仅有 12 个月的寿命，此后基础模型将原生吸收这些能力，竞争价值将转向垂直领域专业知识与用户成果指标。 Logan Kilpatrick 认为，当前初创公司的 Agent 框架只有约 12 个月的生命周期，此后基础模型将吸收这些脚手架并原生执行工作流。他介绍了 Antigravity——Google 覆盖 Search、Cloud、AI Studio 及 Gemini 应用的统一 Agent 框架，同时重点强调了 Gemini 3.5 Flash 通过后训练获得的编程能力提升。Kilpatrick 还描绘了在通用 AGI 出现之前，垂直领域的“参差不齐”超级智能将率先降临于数学和科学等高可验证性领域的愿景，并讨论了用于取代独立传统架构的单一原生多模态模型 Google Omni。 Google 正在将其 AI 产品统一到名为 Antigravity 的单一 Agent harness 上，从以 Gemini 模型为核心转向贯穿 Search、Cloud、AI Studio 和 Gemini 应用的基础架构。 (0:00) Gemini 3.5 Flash 仅凭后训练（post-training）优化即在编程能力上超越了以往所有的 Pro 模型，展现出 Agent 编程作为狭义超级智能（narrow superintelligence）的巨大潜力。 (15:00) 超级智能预计将以“参差不齐”（jagged）的垂直领域形态逐步到来，数学、金融和科学发现等具备高可验证性的领域将率先取得突破。 (20:00) Google 推出原生多模态模型 Omni，取代过去针对文本、音频、音乐和视频分别训练的分散模型，并逐步将其应用于精准实时视频编辑等场景。 (30:00) Logan Kilpatrick 指出初创公司构建外置 Agent harness 的红利期大约只有 12 个月，因为基础模型未来将原生吸收这些脚手架（scaffolding）功能。 (40:00)

Logan Kilpatrick, lead for Google AI Studio and the Gemini API, argues that autonomous agent scaffolding is rapidly collapsing directly into foundation models, shifting Google's product architecture from isolated services toward a unified agent harness. Unified Agent Infrastructure Across Google's Ecosystem (0:00) Historically, Google lacked a single connective thread across its broad portfolio. Following the acquisition of the team behind Windsurf, Google established Antigravity—an agent harness comprising an IDE, a web-based agent-first interface, a CLI, an SDK, and Gemini API access. Kilpatrick describes Antigravity as the shared infrastructure linking Search, Google Cloud, AI Studio, and the consumer Gemini app. This shifts products from passive question-answering systems to autonomous agents capable of executing tasks on behalf of users. Kilpatrick rejects concerns that autonomous workflows will cannibalize Google's core surfaces by decreasing user engagement hours. Instead, he views search dynamics as positive-sum, noting that autonomous software agents represent a new category of search consumers. Success under Google DeepMind's leadership is framed around maximizing positive user outcomes rather than optimizing for human screen time across products serving over one billion users. Coding Agents, Post-Training, and Narrow Superintelligence (15:00) Kilpatrick views agentic coding as the most mature implementation of agentic AI and an early form of narrow superintelligence. Highlighting Gemini 3.5 Flash, he notes that its coding capabilities surpass earlier Gemini Pro models entirely through post-training optimization rather than fresh pre-training runs. Within Google, internal teams used Antigravity to build the native Gemini macOS desktop application faster than any previous desktop app in company history. Kilpatrick anticipates that artificial general intelligence will arrive in a 'jagged' fashion across specific domains rather than as a uniform leap. Domains with high inherent verifiability—such as formal mathematics, algorithmic finance, and scientific discovery—are positioned to achieve superhuman capabilities first. In creative coding, analytics from Google AI Studio showed that roughly 20% of user-generated applications were games, though full game generation remains constrained by orchestration scaffolding and asset pipelines rather than raw model intelligence. Native Multimodal Models Replacing Modular Routing (30:00) Google has adjusted its architecture away from chaining separate models—such as the Gemini text baseline, Lyria for music, and Veo for video—toward natively unified systems. Google Omni is built as a single multimodal model handling diverse inputs and outputs simultaneously. DeepMind CEO Demis Hassabis views this approach as a foundational step toward functional world models that grasp physical reality, avoiding the high compute costs associated with historical action-conditioned video generators. The initial public rollout, Gemini Omni Flash, focuses on fine-grained video editing, demonstrating capabilities such as seamlessly integrating dynamic elements into live video streams. In developer applications, Kilpatrick notes that AI Studio users generated 350,000 Android applications within a week of releasing Android vibe-coding capabilities, primarily targeting personalized utility tools with native OS-level integrations. The 12-Month Harness Horizon and Lab Dynamics (40:00) Kilpatrick predicts that dedicated agent harnesses built by external startups face an approximate 12-month shelf life before base foundation models natively absorb containerized execution, hosted tool suites, and agentic scaffolding. To evaluate model flexibility across varying runtime environments, he proposes an industry benchmark called 'HarnessBench'. Despite this consolidation, he maintains that independent startups hold an advantage in domain-specific execution speed and risk tolerance over large companies balancing massive user bases. Kilpatrick characterizes DeepMind's culture as retaining an academic, scientific mission under Demis Hassabis, distinguishing it from OpenAI's commercial focus and Anthropic's esoteric style. DeepMind functions as the 'engine room of Google,' balancing foundational research aimed at complex challenges with the operational deployment demands of Google's 13 products that each exceed one billion users. While foundation models are poised to assimilate the external runtime harnesses currently built by startups, the transition relies on models becoming universally adaptable across unstandardized environments, leaving open the question of how smoothly general models will manage domain-specific execution layers.

Google DeepMind 产品主管 Logan Kilpatrick 指出，AI 系统正由静态权重模型向自主 Agent 基础设施转型，而外部工程支架预计将在 12 个月内被底座模型原生吞噬；依托后训练的编程 Agent 已显现狭义超智能特征，推动多模态架构与垂直场景的深度重构。 统一 Agent 支架与模型吞噬假说 (40:00) Kilpatrick 提出“模型吞噬脚手架”的核心论点：两年前模型仅被视作输入输出 Token 的静态权重，而在 Gemini 3.5 时代，代码执行容器、托管搜索与 Agent 支架正直接围绕模型原生构建。Kilpatrick 预测，初创公司自建 Agent 支架的红利期仅剩约 12 个月，通用底座模型将逐步原生吸收外围工程脚手架，未来行业需要如 HarnessBench 这般的标准化基准来评测模型对各类运行时环境的自适应能力。 在 Google 内部，整合 Windsurf 团队后打造的 Antigravity 正成为贯穿 Search、Cloud、AI Studio 与 Gemini 应用的统一基础设施。该支架将既有产品从被动问答升级为代表用户自主执行任务的 Agent。配合内部上百名顶级工程师的密集测试与反馈飞轮，Antigravity 推动了内部 Token 消耗量的爆发式增长。 编程智能体的后训练跃升与超智能路径 (15:00) Kilpatrick 认为 AI 编程已展现出“狭义超智能”的特征，其现实价值在于拓展开发者的野心与交付效率，而非单纯的人力替代。例如，Google 团队借助 Antigravity 以历史最快速度上线了原生 Gemini macOS 桌面端应用。其中，Gemini 3.5 Flash 成为关键里程碑：该模型完全基于后训练（Post-Training）优化，在未进行新大规模预训练的情况下，编程性能全面超越了此前的所有 Gemini Pro 模型。 对于通用人工智能（AGI）的实现路径，Kilpatrick 预测其不会以全能统一的形式突然降临，而是呈现垂直领域分步达成的“锯齿状超智能”。他认为高可验证性与具备快速反馈回路的领域将率先突破，包括数学定理的形式化验证（如求解埃尔德什猜想）、金融算法与交易逻辑建模，以及特定垂直学科的科学假说生成与数据分析。 原生统一多模态与世界模型范式演化 (30:00) 在多模态与物理世界建模层面，Google 正在放弃将文本、音乐（Lyria）、音频及视频（Veo）等独立模型进行路由拼接的旧技术栈。新亮相的 Google Omni 采用跨文本、音频、图像和视频全模态原生统合训练架构，其首发版本 Gemini Omni Flash 聚焦高精度的实时视频编辑，可动态将外部实体无缝融入实时视频流并呈现符合物理交互规律的自然反应。 Kilpatrick 指出，“世界模型”的定义正从过去类似 Genie 那样昂贵且难以扩展的动作条件视频生成器，演进为对物理现实具备深层潜在理解的综合系统。尽管在 3D 游戏生成领域，编程 Agent 仍面临资产编排、渲染引擎（如 Three.js）对接及设计脚手架的瓶颈，但原生统合感知的全模态模型正成为通向通用物理理解的关键路径。 专用界面价值与初创企业专注优势 (10:00) 面对全能通用提示界面的诱惑，Kilpatrick 坚持专用界面能有效降低用户认知负荷，用户更偏好布局稳定可预测的专用交互，而非频繁动态重构界面。在 Google AI Studio 推出 Android 端 Vibe Coding 功能后，用户仅在一周内就构建了超过 35 万款移动应用，且约 20% 为游戏，反映出普通用户借助操作系统级深度权限构建个性化专属软件的庞大需求。 在竞争格局上，Kilpatrick 指出 Google 拥有 13 款用户超 10 亿的产品，面对数十亿受众的安全与合规托底责任，大厂在全局 Agent 推进上整体处于谨慎起步阶段。相反，独立初创团队无需维系庞大多元的产品线，依靠极度的垂直专注力与高风险偏好，能借助 AI 编程工具迅速填补工程鸿沟，在特定场景跑赢大企业。 科学导向的研发文化与十亿级工程落地 (45:00) Kilpatrick 剖析了主要 AI 实验室的文化差异，指出 DeepMind 始终烙印着诺奖得主 Demis Hassabis 的科学探索基因，聚焦于攻坚疾病等真实世界重大命题，而非单纯在 SWE-bench 等榜单上进行零和博弈；OpenAI 与 Anthropic 则分别反映出 Sam Altman 的商业执行力和 Dario Amodei 的独特研究哲学。随着 Hassabis、Sergey Brin 与 Noam Shazeer 等领导者坐镇一线，DeepMind 呈现出高强度的研发生态。 被定义为“Google 发动机室”的 DeepMind，肩负着将前沿技术落地到 Android、Cloud、Gmail 和 Workspace 等超级场景的重任。在 13 款月活突破 10 亿的旗舰产品中实现安全可控的模型调用，并探索多天级别的长程自主任务执行，构成了 Google 独有的工业级工程壁垒。 访谈表明，AI Agent 竞争正加速迈向深度系统集成与长程自主执行，但底座模型快速吸收工程支架的趋势对中间层工具构成了明确的时效压力。核心悬念在于：全模态与代码智能能否在 12 个月内彻底替代复杂脚手架实现全自动闭环，以及超大型平台如何在数十亿用户的安全责任底线与 Agent 的激进迭代之间取得稳固平衡。

## Background 本期节目由红杉资本（Sequoia Capital）的 Sonya Huang 主持，于 2026 年 6 月 11 日发布于播客及网络平台。节目主要围绕 Google AI Studio 和 Gemini API 负责人 Logan Kilpatrick 的深度对谈展开，探讨了初创生态狂热构建的 Agent Harness 仅有约 12 个月保质期、Google 内部统一底座 Antigravity 的演进、窄域超级智能与参差不齐的超级智能涌现，以及全模态单体模型 Omni 和以用户成效为导向的商业哲学。 ## TL;DR Logan Kilpatrick 深入剖析了从单一模型走向 Agent 架构的技术演进，指出外部脚手架（Harness）正在被大模型原生吸收，而 Google 自身正通过 Antigravity 框架重构全线核心产品。他认为编程领域已展现出“窄域超级智能”特征，且具备严密验证闭环的数学、科学等垂直赛道将率先迎来超级智能的“参差涌现”。Google 的终极目标并非争夺用户的眼球停留时长，而是最大化交付成果并解放人类生产力。同时，以 Omni 为代表的单体多模态世界模型，正在将离散的专用系统整合成具备物理与社交理解能力的新一代底座。 ## Atmosphere & style 整场对话在专业极客的深度洞察与轻松幽默的谈笑风生之间自如切换，散发着硅谷前沿科技圈特有的敏锐与真实感。Logan 既有大厂技术操盘手的严谨与清醒反思，又带有开发者布道师般的真诚热情，与 Sonya 针锋相对却又默契十足的提问碰撞出密集的信息增量。从开场关于舞台即兴生成小狗的调侃，到对大模型吞噬周边生态的犀利预判，听感上既高屋建瓴又极富人性温度，真实还原了顶尖实验室在技术爆发奇点下的思考脉搏。 ## Bulletin - `01:54` Google 正在全面将源于 Windsurf 团队的 Antigravity 作为贯穿 Search、Gemini App、Cloud 和 AI Studio 的统一 Agent 运行架构，而 Logan Kilpatrick 预测外部 Agent Harness 的独立壁垒仅剩约 12 个月保质期，模型将原生吸收外部脚手架。 - `30:00` Google 推出原生单体架构的 Omni 模型，彻底打破以往针对文本、音频、音乐和视频分别训练独立系统的路由分发模式，实现了对物理世界常识与社交互动的端到端理解与高保真视频编辑。 - `35:10` 在 Google AI Studio 支持通过“氛围编码”（Vibe Coding）生成原生 Android 应用后，开发者在上线仅一周内便构建了超过 35 万个解决长尾具体需求的个人定制软件。 - `22:15` 编程已表现出“窄域超级智能”特征，Logan 认为在通用人工智能（AGI）到来之前，世界将率先迎来在数学、金融和基础科学等高可验证性垂直领域非对称爆发的“参差不齐的超级智能”（Jagged Superintelligence）。 ## Key takeaways - `01:54` Antigravity 接替单一基座模型成为 Google 全线业务的统一 Agent 架构骨干，全栈产品共享约 80% 的调度与工具调用逻辑，其余 20% 针对特定场景进行调优。 - `05:00` Google 产品评估指标从最大化用户的“眼球停留时长”转向“最大化帮助客户达成最终目标”，且 Agent 的广泛部署正在推动搜索业务进入正和增长生态。 - `07:11` 生成式引擎优化（GEO）正在承接传统 SEO 的生态位，商业聚合平台的流量入口与消费决策权正逐步转移至具备自主采购与选型能力的 Agent 端。 - `08:35` 面对十余款十亿级用户的超级应用，Google 整体 Agent 化仍处于审慎的“爬行”阶段，但在 Gemini App 与支持超长程自主编程重构的 Antigravity 等产品上已迈入“行走”阶段。 - `10:00` 相比于万能生成式界面的认知过载，专精工具具备天然的心智确定性，生态未来将分化出数以万计解决具体细分痛点的垂直产品。 - `12:00` 评估智能体演进的核心硬指标正在从单纯的 Token 消耗总量，转向“智能体单次任务自主运行的平均持续时长”（Average Agent Run Duration）。 - `14:40` 得益于后训练（Post-training）红利以及 Antigravity 提供的真实工程实验场，Gemini 3.5 Flash 的编程表现已全面超越上一代 Pro 级模型。 - `18:00` Google 内部十万名工程师在日常研发中使用 Gemini 并提供高密度 A/B 测试反馈，构成了研究机构与初创团队难以复制的工程飞轮。 - `20:00` 编程代码智能体已具备窄域超级智能特征，它并非人类开发者的替代品，而是通过显著降低交付成本重置并放大了人类个体的野心水平与能动性。 - `22:15` 超级智能将以锯齿状（Jagged）分批涌现，具备严密逻辑验证闭环的数学证明、量化金融与科学模型赛道将最先实现质的突破。 - `27:10` 在游戏制作落地路径上，短期内“代码智能体 + 成熟游戏引擎”相比纯粹的世界模型能提供更大确定性，当前瓶颈主要在于游戏资产与状态流编排脚手架的缺失。 - `30:00` Omni 模型采用单体原生融合架构，避免了多模态路由拼接带来的延迟与失真，能够精准理解现实社交动态并在视频编辑中保留人类主体的真实性。 - `37:46` “模型”的概念已从纯粹的静态权重矩阵演变为集成搜索、沙盒代码执行与工具链调用的复合系统工程，模型正在持续内化并吞噬外围脚手架。 - `40:00` 第三方 Agent Harness 仅有约 12 个月的窗口期，初创公司的核心壁垒在于对垂直领域客户痛点的极致专注，以及在边缘创新场景中承担更高风险的敏捷度。 - `45:00` Google DeepMind 兼具 Demis Hassabis 科学求索的学者基因与敏捷攻关的“战役”文化，同时扮演着向 Google 旗下 13 款十亿级用户产品工程化部署模型的底座角色。 ## Detailed walkthrough ### `00:00` 现场即时编辑与“理解世界”的 Omni 模型 节目开场伴随着轻松的幕后闲聊与调侃，Logan Kilpatrick 与 Sonya Huang 谈到是否应该像刚才在场下聊天那样，直接把镜头外的即兴互动剪进正片里作为开场，因为这种无缝衔接展现出了 AI 技术的全新潜质。Logan 分享了一段令他极为震撼的亲身经历：在一次演讲活动中，他与负责模型团队的同事 Tulsi 同台。Logan 当时在台上开玩笑让台下观众现场试试编辑视频，结果有人立刻拍下现场照片，并用 Omni 进行了实时端到端编辑。令人瞠目结舌的是，在编辑后的视频画面中，一只逼真的小狗突然溜达上了舞台；台上的其他嘉宾见状低下头看狗并会心微笑，而台上的 Logan 则一边继续侃侃而谈关于 AI 的各种宏大话题，一边自然地伸手去摸跳上膝盖的小狗。 Logan 笑称台下观众和台上嘉宾当时的笑声“绝对不是因为我的笑话好笑”，而是被小狗毫无违和感地闯入画面给逗乐了。模型在捕捉这种现实世界的细微动态与社交默契（Social Subtlety）上表现得无懈可击，不仅准确渲染了交互细节，还让整个虚实结合的过程显得自然流畅。Logan 感慨，这种对现实物理世界与人类行为深刻理解的“世界模型”雏形，正迫使所有人重新审视并消化其对未来内容创作乃至整个数字生态的颠覆性意义。 - **Omni 实时交互演练**：通过单张现场照片与实时指令，直接生成符合物理规律与人物交互动态的高保真视频片段。 - **世界理解的微妙细节**：模型精准捕捉到了旁人视线下移、低头莞尔微笑，以及演讲者一边继续讲话一边抚摸跳上大腿的小狗等连贯微表情与肢体反应。 --- ### `01:54` 告别单模型时代：Antigravity 成为贯穿 Google 的新“连接线” Sonya 正式引出 Logan 的身份背景——他不仅掌舵 Google AI Studio 和 Gemini API，更是 Google 为下一代开发者构建工具的核心人物。刚刚过去的 Google I/O 大会上，Google CEO Sundar Pichai 正式宣布全面迈入“Agentic Gemini 时代”。Logan 坦言，早在 Gemini 2.0 阶段 Google 就曾提及 Agent 愿景，但当时稍微有些操之过急；直到如今的 Gemini 3.5 时代，Agent 驱动的软件开发和原生 Agent 产品才真正迎来成熟与爆发。 Logan 抛出了一个核心观点：长期以来，Google 旗下的数十款核心产品（约有数十至上百款产品）在技术底座上缺乏统一的系统架构。Gemini 大模型的诞生曾首次充当了这根“主连接线”（through line），让所有 Google 产品全线接入统一的大模型能力；然而如今，这一角色正全面被 Antigravity（源自 Windsurf 团队，现升级为 Google 的核心 Agent Harness）所接替。Google 的全部业务线正在全面重构，转型为由 Agent 驱动的原生产品。 ```mermaid flowchart TD accTitle: Google 核心连接层的技术演进 accDescr: 展示从最初分散的底层，到 Gemini 单一模型作为连接线，再到 Antigravity 作为通用 Agent Harness 驱动所有产品的进化历程。 subgraph Past["早期架构"] P1["分散的单点产品架构"] end subgraph EraGemini["Gemini 阶段"] M1["统一的基础模型: Gemini API"] P2["驱动搜索、Cloud、应用"] M1 --> P2 end subgraph EraAgentic["Agentic 时代 (当前)"] direction TB A1["通用 Agent Harness: Antigravity"] subgraph Products["接入产品矩阵"] direction LR G1["Google Search"] G2["Gemini App"] G3["Google Cloud"] G4["AI Studio"] end A1 --> Products end Past -.-> EraGemini EraGemini -.-> EraAgentic ``` 面对 Sonya 关于“Antigravity 究竟是一个 IDE 还是非 IDE 工具”的提问，Logan 解释称 Antigravity 是一个完整的生态系统：它既包含核心的本地 IDE，也拥有面向 Web 端的 Agent-First 原生交互界面，同时还提供 CLI 和 SDK。开发者既可以通过 Gemini API 直接调用这一免运维的基础设施级托管 Agent，而更深远的意义在于：这个通用的 Harness 框架正直接驱动着 Google 几乎所有核心产品的 Agent 化能力——包括 Google Search、Gemini App、Google Cloud 以及 Google AI Studio。 - **核心定位演进**：从“Gemini 大模型作为唯一技术连接线”，演变为“Antigravity Agent Harness 作为整个产品矩阵的通用架构骨干”。 - **全栈生态覆盖**：提供桌面 IDE、Web 端 Agent、命令行 CLI、开发者 SDK，以及云端全托管 API。 --- ### `04:09` 编码 Harness 即通用 Harness：专精与通用的分化 在探讨 Agent Harness（智能体框架/线束）与 Coding Harness（代码生成线束）的关系时，Logan 指出，业内关于二者是否是一回事一直存在微妙的辩论。事实证明，编写代码不仅是 Agent 极具商业价值的垂直落地场景，更是通往“通用 Agent Harness”的最佳试金石。换句话说，针对代码环境磨砺出的感知、调用工具、纠错和执行循环，已经证明了它本身就是极其强大的通用 Agent Harness。 当然，在实际产品落地中依然存在针对特定场景的深度裁剪与优化。Logan 透露，驱动各业务的底层核心框架大约有 80% 的代码与逻辑是完全共享的，但剩下的 20% 会根据场景进行专精定制。例如，Google AI Studio 中的 Harness 针对快速原型探索与“氛围编码”（Vibe Coding）进行了特别调优；而面向 C 端用户的 Gemini App 则针对 7×24 小时全天候运行、低延迟响应和后台持续处理进行了专向优化。 - **80/20 架构法则**：80% 为通用调度、工具执行与记忆线束，20% 针对不同任务形态进行超参数与交互调优。 - **分化形态**：AI Studio 偏向交互式“氛围编码”与原型生成；Gemini App 则偏向始终在线的消费级自主助理。 --- ### `05:00` 自我蚕食危机？衡量标准从“用户停留时长”转向“任务达成结果” Sonya 提出了一个尖锐的商业化难题：随着 Agent 能力从单纯的检索和信息摘要深入到能够代读邮件、代写代发甚至自主决策，用户的屏幕视线与使用时长（Eyeball Hours）必然会急剧减少。当用户不再亲自打开邮箱或翻看网页，Google 是否会面临现有商业帝国的剧烈自我蚕食（Cannibalization）？ Logan 回顾了当前 AI 时代初期外界对搜索业务的普遍恐慌：当时业内几乎一致认为“大模型能直接给出答案”必定会导致搜索业务沦为负和博弈（Negative-Sum Game）；但实际结果却是强烈的正和效应（Positive-Sum）——不仅人类的探索欲望被激发而搜索得更多，蓬勃兴起的各路自主 Agent 本身也成为了高频搜索调用的新用户群体。 更关键的思维转变在于 Google 对产品成功终极标准的重新定义。Logan 引用了他与 Google DeepMind 负责人 Demis Hassabis 的多次深入讨论指出：技术的终极意义在于替人类完成繁杂的具体事务。“对 Google 而言，未来的成功绝不是最大化用户在产品前消耗的眼球停留时长，而是最大化帮助客户达成最终目标（Maximizing outcome for customers）”，让用户能更快解决问题并回归真实的现实生活。 - **搜索的正和生态**：人类的主动搜索行为不仅没有萎缩，反而因更深维度的探索而增加；同时以 Agent 为主体的程序化搜索量呈现爆发式增长。 - **价值理念重塑**：从传统的广告流量逻辑（以停留时间和展示曝光量为衡量标准）全面转向以交付成果、解放人类生产力为核心的成效导向。 --- ### `07:11` 从 SEO 到 GEO：Agent 驱动增长对商业模式的重塑 针对 Sonya 提出的“Agent 驱动增长”（Agent-Led Growth, ALG）概念——在开发者已习惯全权交由 Agent 替自己挑选云数据库与技术栈的当下，未来普通消费者的购物决策也将完全托管给 Agent——商业聚合平台的流量变现与广告竞价规则势必迎来剧烈震荡。 Logan 认为，商业逻辑的发展呈现出很强的延续性与自适应性，未来的演变可能并不会像外界想象的那样断崖式剧变。传统的搜索引擎优化（SEO）与如今兴起的生成式引擎优化（GEO，Generative Engine Optimization）在本质逻辑上高度互为镜像，彼此之间存在着极强的承袭与映射关系。技术演进是层层叠加、螺旋上升的，商业生态也会在底层机制上自然完成与自主决策 Agent 的对接。 - **技术镜像效应**：GEO（生成式引擎优化）正在无缝承接传统 SEO 的生态位，成为引导 Agent 作出推荐与采购决策的核心手段。 - **决策权转移**：用户正主动将工具选型、底层云架构与消费选择权让渡给 Agent，商业变现的接触点正在向模型决策端漂移。 --- ### `08:35` 进化度评估：走在“爬行”阶段的 Google 与步入“行走”的先锋产品 当被要求用“爬行、行走、奔跑”（Crawl, Walk, Run）的三阶段模型来评估 Google 现有产品矩阵的 Agent 化程度时，Logan 坦率地将其整体定义为“爬行”（Crawl）阶段。他强调，这种克制主要源自 Google 面临的产品体量压力——旗下拥有超过十多款用户量破十亿级的超级应用。针对这种量级的全人类公共级基础设施，技术落地必须承担起应有的谨慎守护职责（Stewardship Responsibility）。长尾大众用户在隐私保护、行为可控性等方面尚未做好让 AI 全面放飞“奔跑”的准备，搜索团队尤为需要以稳健的节奏引导用户习惯渐进演变。 ```mermaid flowchart LR accTitle: Google 产品 Agent 化的三阶段成熟度分布 accDescr: 展现从 Crawl 到 Walk 再到 Run 阶段的产品分布与对应的产品形态特征。 subgraph Stage1["爬行阶段 (Crawl)"] direction TB S1["主力消费级产品线"] S2["Google Search, 邮件及核心办公套件"] S3["核心考量: 10亿+用户规模与稳健监管责任"] end subgraph Stage2["行走阶段 (Walk)"] direction TB W1["前沿探索产品"] W2["Gemini App (7x24h 伴生助手)"] W3["Antigravity (自主代码重构)"] end subgraph Stage3["奔跑阶段 (Run)"] direction TB R1["前沿全自主实验室阶段"] R2["消耗海量 Token、自主重构 OS 的完全自动化系统"] end Stage1 --> Stage2 --> Stage3 ``` 不过，在一些特定的先锋产品线上，Google 已经迈入了“行走”（Walk）阶段。Logan 点名了 Gemini App 与 Antigravity：前者正致力于打造 7×24 小时随时在线、能主动代表用户在现实世界执行系列复杂动作的智能伴侣；而后者作为自主编程平台，已经能够支持 Agent 在无需人类干预的情况下连续消耗数以十亿计的 Token，花费上千美元的云端算力，自主完成重构操作系统等超高复杂度的工程任务。这些先锋实验正揭示出下一代自主软件构建的未来雏形。 - **多产品用户量级壁垒**：坐拥十余款 10 亿级以上用户的超级产品，要求系统必须在权限可控、隐私合规与渐进式用户体验上保持审慎。 - **先行“迈步”的产品矩阵**： - **Gemini App**：探索 7×24 小时始终在线、代表用户在互联网环境执行多步骤真实世界任务的超级 Agent 助理。 - **Antigravity**：支持长程自主编程，具备消耗数十亿 Token 规模进行系统级底层软件与操作系统自主重构的前沿工程能力。 ### `10:00` 专精产品与万能界面的认知负荷博弈 关于未来用户究竟会面对少数几个整合式 AI 入口还是数万个分散的产品形态，Logan Kilpatrick 认为这本质上取决于人类消费产品时的心理机制。人类天生习惯产品的模块化和专精化（compartmentalization and specialization）；如果一个产品试图包揽所有任务，用户为了让它达成特定目标，反而需要耗费更多心智与时间去引导和交互。反之，单一功能产品具备即开即用的确定性：“我点开日历应用，它就单纯展示日历，我完全不需要应对任何其他干扰。”Logan 甚至抛出了一个关于传统办公工具的尖锐洞察（hot take）：幻灯片（Slide decks）之所以经久不衰，正是因为人们希望关键信息始终固定在预期的位置。虽然动态生成式界面（Generative Interface）概念听起来极其炫酷，但在很多场景下它给大脑带来的认知过载远大于便利。 对于整体 AI 生态而言，Logan 坚信未来一定会涌现出数以万计去解决各类细分需求的产品，这种长尾繁荣令人兴奋；至于 Google 自身最终是向用户提供三款核心产品还是庞大的产品矩阵，则将取决于公司的顶层战略决策。 ```mermaid flowchart TD accTitle: 通用界面与专精工具的心智成本对比 accDescr: 对比万能生成式界面与独立垂直工具在用户心智负担与确定性上的差异 subgraph GeneralInterface ["万能生成式 AI 界面"] A["单一入口承载全部需求"] --> B["需要大量提示词与交互引导"] B --> C["认知负荷高 / 输出结果不确定"] end subgraph SpecializedTools ["垂直专精产品 (如日历、幻灯片)"] D["单一专精功能界面"] --> E["信息结构固定且可预测"] E --> F["即开即用 / 零额外心智成本"] end ``` ### `12:00` 智能体落地的分水岭：长程任务（Long-Horizon）与评估指标 针对许多企业认为“目前只有编程智能体（Coding Agents）真正跑得通”的观点，Logan 认为这取决于企业界对“可用（working）”的具体评判门槛。在模型质量尚未跨越特定阈值的复杂垂直领域，想要完全甩手托管任务确实还不现实。但他指出，整个行业正在经历关键的转折点。业内通常习惯用 OpenRouter 等平台的 Token 消耗量来衡量全球智能总量的爆发，但 Logan 更关心一个目前各家尚未公开的关键指标——智能体单次运行或单个任务的平均持续时长（average agent run duration）。 各大前沿实验室已经开始宣传模型能够自主执行长达三天的任务，尽管这仍是极少数极端案例，但长程任务（Long-Horizon Tasks）的时长正在快速爬升。Logan 预计，即使企业目前在编程之外尚未深切体会到智能体的价值，随着底层能力的涌现，其他垂直场景也将在今年迅速跨过可用门槛。在 Google DeepMind 的众多前沿赌注中，长程智能体与编程模型被视为主导级别的能力，因为一款卓越的编程模型本身就是全公司各项业务的“终极加速器”。 - **Token 消耗量 vs. 任务生命周期**：行业常用 Token 总消耗衡量智能普及，但“Agent 单次自主运行能够维持多久”才是反映智能体推理深度的硬指标。 - **长程任务外溢效应**：自研顶级代码模型不仅服务开发者，更直接赋能内部科研与系统工程，成为推动全盘业务运转的底层杠杆。 ### `14:40` 编程模型战局：Windsurf 渊源、Antigravity 与后训练红利 谈及开发者生态的实际体感，主持人 Sonya Huang 提出了一个尖锐问题：许多开发者在 Claude 与 OpenAI Codex 之间平分秋色，却鲜少将 Gemini 作为主力编程工具。Logan 坦言这种观察并非空穴来风，并回顾了瞬息万变的技术叙事：在落地 Gemini 3 时，外界曾一度盛赞 Google 赢下了模型能力的阶段性胜利；然而随着圣诞假期至一月份编程智能体浪潮的席卷，行业风向几乎在一夜之间以“曲速（warp speed）”切换。这给团队敲响了警钟：想要做出真正适应极长程软件工程（SWE）任务的代码模型，前提是必须拥有一个能够承载这种交互的真实产品环境。 这也正是 Google 促成与 Windsurf 团队合作的根本原因。该团队加入后构建了 Antigravity，成为连接产品与模型的关键纽带；Sundar Pichai 在 Google I/O 上展示的内部 Token 消耗飙升曲线，正是这一飞轮加速运转的直接证据。Logan 将 DeepMind 内部的代码团队比作“AI 领域的复仇者联盟（Avengers of AI）”，汇聚了公司内顶尖的人才推石上山。他特别指出，算力集群可用性带来的预训练窗口期（Pre-training Windows）往往存在滞后性，外界往往会因此误判研发进度。以最新推出的 Flash 模型为例，其编程能力已经全面超越了此前发布的所有 Pro 级模型，而这一惊人的跨代超越完全是由后训练（Post-training）收益驱动的。 - **Antigravity 的战略定位**：单纯在评测集上刷分无法磨练出真正优秀的工程师智能体，Antigravity 为代码模型提供了真实、长时间运行的软件工程实验场。 - **预训练窗口与后训练突破**：外界容易忽视万卡集群大规模预训练的排期周期；而 3.5 Flash 仅凭后训练就击败了上一代 Pro 模型的编程表现，展现了对齐与强化学习的巨大杠杆。 ### `18:00` 内部 Dogfooding 哲学与十万人工程飞轮 在模型内部试用（Dogfooding）的机制上，Google DeepMind 并没有推行盲目排外的教条主义。Logan 明确表示，团队内部完全允许甚至鼓励工程师体验外部竞争对手的模型：“如果不去使用外部产品，你根本无法真正理解生态系统里正在发生什么。”Logan 自己就是全生态各类模型的深度用户。 但与此同时，将 Gemini 作为日常主力开发工具（Daily Driver）构成了不可替代的竞争壁垒。Google 拥有超过十万名顶尖工程师，这支庞大的人才队伍每天在实际工作中调用 Gemini、进行实时 A/B 测试并持续回传结构化反馈。这种超大规模的真实工业级反馈飞轮（Feedback Flywheel），是任何初创公司或缺少生态纵深的研究机构都无法复制的核心优势。 ```mermaid flowchart LR accTitle: Google 内部 Dogfooding 增强飞轮 accDescr: 展现十万工程师使用 Gemini 产生反馈、通过 A/B 测试改进模型的循环过程 A["十万级内部工程师日常使用"] --> B["真实工程场景与长程编码"] B --> C["高密度交互数据与 A/B 测试反馈"] C --> D["后训练与模型能力迭代"] D --> A ``` ### `19:15` “软着陆”式自我强化循环与万卡 TPU 资源闸门 面对业界热议的“软着陆/软飞跃（Soft Takeoff）”假说——即代码智能体足够强之后，将自主加速模型架构与训练算法的研发，从而触发自我强化的智能爆炸——Logan 确认，内部确实已经观察到了一些早期苗头，代码模型正在切实辅助研究员编写实验代码并加速研究流程。 然而，距离完全由 AI 主导的自动化迭代依然存在现实鸿沟，其中最大的制约因素在于超大规模训练的**物理资源配置壁垒**。在当前阶段，发起一次前沿模型训练任务需要调动数以万计的算力芯片，耗资极其高昂。人类研究员必须牢牢坐在驾驶位上掌控战略决策，正如 Logan 形象的总结：“你绝不可能任由一个智能体随随便便就调用一万张 TPU，去启动一个根本没什么实际意义的巨型训练任务。”真正的自举闭环仍然受制于物理世界的严苛资源边界。 ### `20:00` 编程领域的窄域超级智能与人类野心的重置 Logan Kilpatrick 指出，从产品交付的角度来看，智能体编程（agentic coding）带来的提速已经在团队内部得到了切实验证。以内部使用的 Antigravity 为例，团队借助它构建移动应用的端到端上线速度，比 Google 历史上任何移动端团队都要快；Josh 的团队利用它打造 Gemini macOS 客户端时，同样创下了 Google 交付 Mac 应用的最快纪录。对于 Sonya Huang 提出的“如果系统能用代码构建一切，人类在同等水平上无法竞争，这是否已属于‘窄域超级智能’（narrow superintelligence）”这一问题，Logan 坦言，现在的编程体验确实给人这种感觉。尽管如何量化定义见仁见智，但核心事实在于代码生成极其有效。 Logan 强调，AI 社区在追求 AGI 的同时，绝不能忽视甚至贬低当前技术已经具备的非凡现实价值。对他而言，当下的 AI 并没有剥夺人类开发者的价值，反而是人类能力的“加速器”（accelerant）。他分享了个人的心态转变：过去脑海里冒出新想法时，常常因为实现成本太高而只能感叹“要是能做出来该多好”，随后作罢；而现在却面临相反的“幸福的烦恼”——哪怕只想做个最小可行产品（MVP），只要有技术加持，他就会反问自己为什么不再多推进 10 步、做得更宏大。这种“野心水平的重置”（resetting my level of ambition）虽然带来了一层更重的责任与负担，但也极大扩展了人类个体的能动性（agency）。 ```mermaid flowchart LR accTitle: 从窄域超级智能到人类野心重置 accDescr: 展示编程智能体如何带来窄域超级智能体验，并重置开发者的野心水平。 subgraph S1["效率跃迁"] A["Antigravity / 智能体编程"] --> B["超高速产品交付<br/>(Gemini Mac App / 移动端)"] end subgraph S2["认知重塑"] B --> C["窄域超级智能体验<br/>(Narrow Superintelligence)"] C --> D["开发者能动性增强<br/>(野心重置：从想放弃到多迈10步)"] end ``` - **内部交付标杆**：Josh 带领的团队通过智能体编码，创造了 Google 内部端到端上线 Mac 原生应用的最快速度记录。 - **追求平衡的视角**：必须同时兼顾“构建通用技术（AGI）”与“充分利用当下极具冲击力的专业能力”，避免因仰望未来的 AGI 而掩盖了当前技术的惊人突破。 --- ### `22:15` 参差不齐的超级智能：下一个突破的垂直领域 在 AGI 全面降临之前，Logan 认为世界将率先迎来一系列“参差不齐的超级智能”（jagged superintelligence）——即在不同垂直领域呈非对称分布的高阶智能。当被问及除代码之外下一个最快见证超级智能的领域时，Logan 指出，拥有“高可验证性”（verifiability）的学科必然跑在最前列。数学、金融以及科学研究都是最具潜力的候选者。在这些领域，AI 的输出可以通过形式化证明、数据对账或实验逻辑进行检验，从而形成快速的正向迭代飞轮。 Logan 从宏观叙事层面分析，这类高正面价值的突破越早到来，对整个人类社会就越有利。如果高影响力、推动人类福祉的重大科学突破能在技术演进早期落地，大众就能直观建立起对 AI 潜力的正面认知。谈及当前数学证明领域的火热进展，Sonya 幽默地援引推特上的热梗“为什么埃尔德什（Paul Erdős）有那么多留给后人的未解问题”，Logan 大笑表示赞同，称这完全可以印在文化衫上，这种高难度但闭环可验证的课题正是探索智能边界的绝佳试金石。 - **参差不齐的超级智能（Jagged Superintelligence）**：超级智能不会以平滑整齐的方式瞬间覆盖所有认知领域，而是在不同垂直赛道分批次、锯齿状地逐个涌现。 - **突破前置条件**：具备严密判定闭环的领域（可验证性强的数学、量化金融、科学模型）最具爆发潜力，且对重塑公众的 AI 信任至关重要。 --- ### `24:25` “Vibe Coding”制作独立游戏：从预测到工程脚手架挑战 话题转向 Logan 曾在推特上的公开预测：“到 2025 年底，每个人都将能够凭借直觉写代码（vibe code）来制作视频游戏。”Logan 承认这个目标目前虽然还没完全达到能做《使命召唤》（Call of Duty）或《侠盗猎车手》（GTA）等 3A 大作的水平，但距离可玩作品的门槛已经无限接近。现实开发中最大的难点往往不在代码本身，而在于边缘工程支撑：像 Three.js 这样的框架虽然降低了门槛，但原生的编码智能体仍然搞不定游戏专属的资产管线，例如精灵图（sprite）生成、音效整合以及长线状态编排。 Logan 强调，要打造可重玩（replayable）、有深度且具备良好审美体验的游戏，必须依赖精心设计的产品脚手架（product scaffolding）。从 Google AI Studio 早期的后台使用数据来看，当时有近 20% 的应用构建尝试都是游戏，足见大众对于自制游戏的热情极高；不过随着生态与受众扩展，目前金融相关的工具（尤其是加密货币）以及利用 Google 媒体能力的内容生成占据了更大的比例。 | 探索维度 | 当前现状与瓶颈 | 破局关键与未来演进 | |:--- |:--- |:--- | | **模型编码能力** | 基础逻辑与 Three.js 代码生成基本就绪 | 解决原生多模态资产（如 Sprite 生成、音效编排）的统合 | | **游戏开发脚手架** | 缺乏针对复杂交互与状态流的原生粘合层 | 构建专业 Orchestration Layer，实现深度且具复玩性的产品体验 | | **GDM 历史渊源** | Demis Hassabis 以游戏研究开启 AI 生涯 | Kaggle 联合 DeepMind 推出 Game Arena 作为检验 AGI 的代理标尺 | --- ### `27:10` 游戏智能体的实现路径：游戏引擎与世界模型之辩 探讨街头普通人何时能凭灵感直接搓出一款好玩的游戏时，Logan 给出了激进的判断：“我想说就在今年。”他认为目前模型能力已经不再是短板，真正的差距在于那些深谙优秀游戏制作之道的人，能否搭建出适配的脚手架去弥合最后一公里。对于普通用户来说，这既是一个认知与发现机制（discoverability）的普及过程，也是等待最后几个细分模型缺陷在数周或数月内被攻克的质变前夜。 关于实现路径究竟是“游戏引擎 + 编程智能体”还是直接依赖“世界模型”（world models），Logan 认为在短期内（如 1-2 年），**“代码智能体 + 成熟游戏引擎”**依然能提供最大的超额收益（alpha）。现阶段的世界模型虽然想象空间极大，但其内在架构的高度开放性（open-endedness）使其缺乏约束，若无极其精密的外部脚手架进行锚定，很难直接用于制作具备确定逻辑的游戏。与此同时，世界模型本身的定义正在迅速泛化与模糊化。正如 Google I/O 大会上发布的 Omni 模型，它能原生处理任意模态输入并生成任意模态输出，DeepMind 联合创始人 Demis Hassabis 正式将其定义为一种“世界模型”，这标志着从传统物理模拟架构向全模态统一架构的认知跃迁。 ### `30:00` 从动作条件视频到统一的单体 Omni 模型 过去所谓的传统世界模型（如 Genie）本质上更接近一种受动作条件约束的视频模型（action-conditioned video model），其运行成本高昂且在架构上极难扩展。如今业界重新定义“世界模型”时，所指的实际上是“对现实物理世界具有真正认知与理解能力”的模型。它虽然目前尚未完全实现毫秒级的实时交互，但已经能够完成诸多过去专属世界模型的生成与视觉推演任务，模糊了纯视频生成与物理世界建模之间的界限。 针对外界关于该技术底层是否仅是“Gemini 串联专用视频模型”的疑问，答案是它完全由一个单一体素（single model）驱动。过去谷歌必须训练多达八种独立的专用模型分别处理不同模态：基线 Gemini 处理文本、Lyria 负责音乐创作、Veo 专攻视频生成，外加一系列独立的音频模型。而全新的 Omni 模型打破了“通过路由器（Router）分发请求给不同下游小模型”的伪全模态做法，是原生构建的真正 Omni 单体架构。当前公开的 Omni Flash 是该技术架构的第一次迭代，首先落地于效果最成熟的精准视频编辑场景，未来更高阶的版本将进一步解锁多模态混合生成能力。 ```mermaid flowchart TD accTitle: 谷歌模型架构从多模型路由向单体 Omni 的演进 accDescr: 展示了从过去分散的独立多模态专用模型，走向单一原生 Omni 统一架构的演变历程。 subgraph Past["过去：离散的多模型碎片化系统"] direction TB R["路由分发层"] M1["文本：Gemini"] M2["视频：Veo"] M3["音乐：Lyria"] M4["音频套件"] R --> M1 R --> M2 R --> M3 R --> M4 end subgraph Present["现在：原生单体架构"] direction TB OMNI["统一单体 Omni 模型 (如 Omni Flash)"] OMNI --> Cap1["物理世界常识与语境理解"] OMNI --> Cap2["端到端原生多模态视频/音频编辑"] end Past -->|架构重构与能力内化| Present ``` - 过去针对不同模态分别独立训练模型：文本（Gemini）、音乐（Lyria）、视频（Veo）以及成套的独立音频系统。 - 真正的 Omni 架构并非在外层做路由调度分发，而是从底层将全模态融合进单一模型体系中。 - Omni Flash 作为首轮迭代，优先切入质量已达商用水平的原生视频编辑场景，后续将陆续放出更高性能的大尺寸版本。 ### `32:38` 细致入微的视频编辑与拒绝“水果岛”的真实性哲学 这种对世界常识的深入理解在实际视频修改中体现得淋漓尽致。在一次演讲活动中，模型团队负责人 Tulsi 与同台交流时，台下观众利用 Omni 进行了现场实时视频编辑测试：画面中凭空生成了一只走上舞台的小狗，舞台上的其他嘉宾自然地低头看向小狗并会心微笑，随后小狗一跃跳上膝头被抚摸，而台上的演讲者依然能够神态自若地继续高谈阔论。这种微表情、视线焦点以及肢体动作与新增物体之间的天衣无缝衔接，证明模型已经掌握了复杂的场景物理互动与社交逻辑，远超生硬的视觉贴图。 面对生成式媒体的爆发，内容创作者最核心的壁垒始终在于“真实性带来的超额收益”（alpha and authenticity）。许多创作者坚持不使用 AI 替身或数字人来制造内容，因为核心的言论、嗓音与面容代表着个人身份。Omni 带来的范式转变在于它“不改变人物主体本身”，而是允许用户自由替换布景、咖啡桌或周边杂物等非核心背景细节。这种理念完全颠覆了市面上充斥着 AI 化身或“水果岛”（Fruit Island）式的廉价拼接视频，让 AI 真正回归到放大人类真实表达、而非取代人类特质的定位。 - 现场编辑演示细节：真实观众利用 Omni 将虚拟小狗注入舞台视频，模型自动补全了其他嘉宾看狗、低头微笑、小狗跳跃上膝以及主体抚摸的动态互动细节。 - 视频内容的核心吸引力在于真实性（Authenticity），这也是创作者抵御同质化 AI 生成内容的核心壁垒。 - Omni 的价值主张在于保留人的真实言行，仅针对非个人化的场景细节（如演播室布景、背景陈设）进行丰富与风格化改造。 ### `35:10` 在 AI Studio 中“氛围编码”：35 万个原生 Android 应用的诞生 Google AI Studio 引入了直接“氛围编码”（vibe coding）生成原生 Android 应用的功能。从产品战略而言，开发者在利用谷歌技术构建初创产品时往往要在复杂的生态与数十个不同的管理后台中来回穿梭；AI Studio 的首要设计原则就是充当大一统入口，使用户无需跳出界面就能调动整个谷歌生态系统的能力。通过降低原生移动开发的门槛，甚至连此前从未写过移动端代码的开发者也能在几分钟内构思并运行属于自己的 Android 应用。 以解决身边实际问题为导向的个人定制软件（Personal Software）正迎来爆发式增长。从该功能上线到录制播客当天上午，开发者仅用一周时间就在 AI Studio 中构建了超过 35 万个原生 Android 应用。虽然目前尚未观察到这些“氛围编码”生成的应用在应用商店掀起商业巨浪，但它们正在切实解决具体的个人长尾需求——例如为了在自家后院种植树木而临时搭建的花园养护小工具。相比于能力日益强大的 Web 网页端，原生移动操作系统所独有的底层丰富度（Native richness）、硬件上下文以及如原生短信般的无缝交互体验，依然是网页端无法轻易替代的。 - 跨生态整合：AI Studio 致力于将分散在谷歌各处的开发者工具链整合进单一工作流，避免开发者在多个管理后台间跳转。 - 爆发式增长：自功能上线仅仅一周左右，AI Studio 中生成的原生 Android 应用总量已突破 350,000 个。 - 个人定制软件崛起：绝大多数生成的应用并非奔着 App Store 商业变现而去，而是用户针对自己后院种树、日常记录等具体私人问题制作的定制工具。 - 原生系统优势：相较于网页端，移动操作系统在原生上下文调度、系统级短信集成与交互质感上具备不可替代的优势。 ### `37:46` 核心命题：模型正在吞噬脚手架与外壳 在探讨“模型正在吞噬脚手架”（The Model Eats the Harness / Scaffolding）时，Logan 明确指出：如今大家口中所谓的“模型”，早已不再是两年前狭义上的那一组静态权重（weights）。过去大语言模型走红之初，模型仅仅意味着一套最纯粹的输入 Token 并输出 Token 的矩阵运算；而如今无论是 Gemini 3.5、GPT 还是 Claude 系列，背后都是一套围绕基础权重极速向外扩张的庞大系统级架构。 在这个新型系统中，Agent 工具调用（Tool Calling）、托管工具链、Google 搜索增强、沙盒代码执行乃至容器化运行环境，都在逐步变成原生系统的一部分。工程实践中，外部脚手架（Scaffolding）往往因为灵活性走在模型迭代前面一两步，但随着模型能力的演进，底层系统会像吸收养分一样迅速将这层外部脚手架吞噬，并转化为原生模型能力的组成部分。虽然在多搜索源切换等特殊异构场景中，外部脚手架依然拥有其垂直生态位，但核心交互架构的原生化整合已经成为不可逆转的技术趋势。 - “模型”定义的演变：两年前模型仅仅指代纯粹的权重矩阵（纯 Token 吞吐），如今则是深度融合了代码执行、搜索与代理工具调用的复杂系统工程。 - 吞噬脚手架的机制：外部 Harness/脚手架通常先行探索一到两步，随后模型系统便会将其标准化并完全内化为原生功能。 - 外部脚手架的剩余价值：在需要接入差异化搜索引擎或非标准化外部工具链等特定定制化场景下，外部封装层仍会阶段性保留其独立价值。 ### `40:00` Agent Harness 的 12 个月保质期与初创公司的“专注”护城河 当前整个创业生态圈掀起了一股疯狂打造 Agent Harness（智能体运行底座/套件）的浪潮，许多团队认为 Harness 就是超额收益（Alpha）的聚集地。Logan Kilpatrick 提出了截然不同的看法：这种局面可能只有大约 12 个月的保质期。正如模型曾经原生吞噬了搜索能力与代码执行一样，未来的模型会直接将这层外部脚手架（Scaffolding）“消化并向上吸收”（Upstreamed），让模型原生具备调度与交互能力，真正的超额收益将转移到其他维度。Sonya Huang 随即追问：很多应用层公司自建 Harness 是为了避免被单一模型供应商绑定、保持跨模型切换的灵活性，模型供应商是否真能通吃？Logan 回应称，通用能力足够强的大模型天然应当能够自适应各种不同的 Harness，如果一个模型换了个 Harness 就无法工作，那它就称不上是通用模型。因此，Logan 提议整个生态应当建立一个像 **HarnessBench** 这样的基准测试，专门衡量各大模型适配不同 Harness 的泛化能力。 在“模型吞噬套件与周边工具”的大趋势下，独立应用层初创公司究竟靠什么生存？Logan 认为，现在反而是创业公司拥有前所未有历史机遇的黄金期。模型大厂追求的是解决高度通用的大问题，而真正深厚的商业价值往往蕴藏在垂直细分赛道中。初创公司最大的核心竞争力就是“极致的专注”（Focus is the superpower of startups）。相比于像 Google 这样背负着数十亿用户和众多产品的巨头，初创公司没有沉重的历史包袱，能够在一个具体垂直领域深入理解客户痛点与生态格局，从而在专业纵深上“套圈”（Run laps around）甚至领先最顶尖的模型实验室。 同时，两年前人们曾担忧 AI 可能会压缩独立开发者的生存空间，如今事实证明恰恰相反。代码生成等 Agent 原生基元（Agentic Primitives）大幅抹平了初创公司与大厂之间长期积累的代码库差距，让小团队能够以不可思议的敏捷度快速产出软件。此外，大企业在面对智能体的不确定性时受制于风险偏好，而愿意承担更多边界风险的初创公司，能够迅速赢得那些同样渴望尝试激进创新产品的先锋用户群体。 - **模型吞噬脚手架周期**：外部 Harness 的独占壁垒预计仅有约 12 个月，随后能力会被模型原生吸收。 - **倡议建立 HarnessBench**：行业需要客观评测不同基础模型对各类第三方 Harness 的兼容与适配效率。 - **初创公司的三重复合优势**： - **垂直专注度**：在大模型专注通用能力时，深耕垂直行业的领域知识与客户洞察； - **研发速度平权**：AI 编码极大缩短了与大公司存量代码库和人力规模的差距； - **风险偏好差异**：小团队更敢于在创新边缘承载风险，抢占前沿用户生态。 ```mermaid flowchart TD accTitle: 智能体技术栈价值演进与初创公司定位 accDescr: 展示从外部脚手架到原生吸收的演进，以及初创公司的核心竞争力。 subgraph Evolution["技术栈演化 (约12个月周期)"] direction TB H1["当前阶段：外部 Harness / 脚手架<br/>（各家争抢自建套件以防锁定）"] H2["未来演进：模型原生吸收能力<br/>（搜索、代码执行、调度被 Upstream）"] H1 -->|模型能力泛化| H2 end subgraph StartupEdge["初创公司的 Alpha 护城河"] direction TB S1["垂直领域沉淀<br/>（深度理解客户与场景）"] S2["极致专注与敏捷<br/>（AI 编码追平存量差距）"] S3["高风险承担能力<br/>（抢先验证 Agentic 新范式）"] end H2 -.->|基础能力通用化| StartupEdge ``` --- ### `45:00` Google DeepMind 的工程师文化与大厂叙事的真实感 谈及 Google DeepMind（GDM）内部的研发氛围，以及 Demis Hassabis、回归的联合创始人 Sergey Brin 与 Noam Shazeer 汇聚一堂的场景，Logan 用“不可思议”（Incredible）来形容这一历史时刻。他总结了 GDM 文化的三个关键切面：首先是深厚的“战役”（Strikes）文化——如同纪录片《思考的游戏》（The Thinking Game）中所记录的那样，当发现某个在投资组合中此前未被充分重视的赛道被友商暂时领先时，DeepMind 能够迅速召集全球最顶尖的极客集中攻关，在极短时间内抹平差距并反超；其次，实验室文化高度映射了创始人特质：OpenAI 带有 Sam Altman 那样无与伦比的商业与运作嗅觉，Anthropic 带有 Dario Amodei 某种深奥独特的学者哲学，而 GDM 则深深烙印着诺贝尔奖得主 Demis Hassabis 纯粹以科学求解世界的底色——始终铭记做 AI 的初心是为了攻坚疾病、科学探索等人类真正的难题，而不是沉迷于在 SWE-bench 上为了单纯刷分而进行零和博弈。 Logan 幽默地借用美剧《硅谷》（Silicon Valley）中 Gavin Belson 的经典名言来讽刺行业内过度的恶性攀比：“我们绝不能让别人比我们更能把世界变得更美好！”（"We can't let other people make the world a better place more than we can."）他强调，推动人类进步从来不是零和游戏。最后，GDM 目前的定位是整个 Google 的“动力舱”（Engine Room）——不仅包含前沿科研，更需要把 Gemini 等模型能力部署并赋能到 Android、Google Cloud、Gmail 和 Workspace 等产品线中。在全世界范围内，只有两家公司面临着在多达十几个十亿级用户（Billion-user）产品中部署最新大模型的工程挑战，而 Google 独占 13 个这样的产品，这种在极致规模下落地前沿技术的经验在外部几乎无法复制。 在被问及高频活跃在社交媒体上是否曾让公关部门提心吊胆时，Logan 对 Google 市场与公关团队展现出的信任与开放表示衷心感激。大公司往往因为层级繁复、参与人数众多，把原本震撼人心的技术魔法经过层层公关流程稀释得平淡无奇，丧失了故事的真实感。他选择以一种直接、真诚的开发者语言与社区对话，既维护了建立在彼此信任之上的底线，又让开发者能够真切感受到 Google 正在探索世界上最前沿的技术并攻克最严峻的挑战。 - **三家前沿实验室的文化画像**： - **Google DeepMind**：诺贝尔奖科学家 Demis Hassabis 奠定的科研立足点，以攻克疾病与基础科学为终极追求； - **OpenAI**：由顶级商业运作大师 Sam Altman 驱动的高强度市场攻势； - **Anthropic**：带有 Dario Amodei 风格的深度学者思考与特立独行路线。 - **十亿级落地挑战**：Google 拥有 13 个十亿级用户产品，GDM 作为“动力舱”承担着将科研成果大规模工程化落地的独特使命。 --- ### `50:00` 告别与尾声：赋予前沿技术真实的人性温度 访谈进入尾声，Sonya Huang 高度赞扬了 Logan Kilpatrick 以及团队中的 Josh 等人，称赞他们成功在当今时代最重大的技术变革中注入了真诚且富有温度的人性色彩（"A really sincere human touch on the most important problem of our time"），让庞大的科技叙事变得生动可感。双方回顾了整场对话涵盖的广泛议题：从智能体架构、编码范式、世界模型，到 Harness 的演进生命周期以及 DeepMind 的组织文化。 在轻松愉快的欢笑声中，Logan 对受邀参与深感尽兴，并开玩笑打趣道，非常期待看到后期制作团队会把他们俩交谈的背景合成出什么天马行空的场景，甚至逗趣地期待画面前会不会凭空生成一只小狗来“圆自己的小狗梦”。Sonya 笑着回应表示一定会帮他实现心愿，在亲切融洽的氛围中正式为整场对谈画上圆满句号。 ## Notable quotes > `05:00` 对 Google 而言，未来的成功绝不是最大化用户在产品前消耗的眼球停留时长，而是最大化帮助客户达成最终目标（Maximizing outcome for customers） — Logan Kilpatrick > `10:00` 我点开日历应用，它就单纯展示日历，我完全不需要应对任何其他干扰。 — Logan Kilpatrick > `18:00` 如果不去使用外部产品，你根本无法真正理解生态系统里正在发生什么。 — Logan Kilpatrick > `19:15` 你绝不可能任由一个智能体随随便便就调用一万张 TPU，去启动一个根本没什么实际意义的巨型训练任务。 — Logan Kilpatrick > `24:25` 到 2025 年底，每个人都将能够凭借直觉写代码（vibe code）来制作视频游戏。 — Logan Kilpatrick > `27:10` 我想说就在今年。 — Logan Kilpatrick > `40:00` Focus is the superpower of startups — Logan Kilpatrick > `45:00` We can't let other people make the world a better place more than we can. — Logan Kilpatrick > `50:00` A really sincere human touch on the most important problem of our time — Sonya Huang ## Mentioned Logan Kilpatrick, Sonya Huang, Sundar Pichai, Demis Hassabis, Sergey Brin, Noam Shazeer, Tulsi, Josh, Sam Altman, Dario Amodei, Paul Erdős, Gavin Belson, Google, Google DeepMind, Google AI Studio, Gemini API, Gemini, Gemini 2.0, Gemini 3, Gemini 3.5, Gemini 3.5 Flash, Antigravity, Windsurf, Omni, Omni Flash, Lyria, Veo, Genie, Google Search, Google Cloud, Android, Gmail, Workspace, Three.js, OpenRouter, OpenAI, Codex, Claude, Anthropic, Kaggle, Game Arena, SWE-bench, HarnessBench, The Thinking Game, Silicon Valley, Call of Duty, GTA, Sequoia Capital ## Video description The entire startup ecosystem is racing to build agent harnesses. Logan Kilpatrick, who leads Google AI Studio and the Gemini API, argues that scramble has a roughly 12-month shelf life. Models will absorb the scaffolding and run it natively, so the edge moves elsewhere. Google's own bet runs in parallel: a single agent harness, born from the Windsurf team and now called Antigravity, has become the connective tissue across search, the Gemini app, Cloud, and AI Studio — the role Gemini-the-model used to play. Logan makes the case that coding already feels like narrow superintelligence, and that "jagged" vertical superintelligence (in math, finance, and science) will arrive well before AGI. He argues Google's real goal is maximizing outcomes for users, not eyeball time. He unpacks Omni, the single model built to replace multiple separate systems Google once trained for text, audio, music, image, and video. His throughline: AI is an accelerant for human ambition, not a substitute for it. Hosted by Sonya Huang, Sequoia Capital [Source](https://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI6013959399.mp3)

[5:00] Logan Kilpatrick: 对 Google 而言，未来的成功绝不是最大化用户在产品前消耗的眼球停留时长，而是最大化帮助客户达成最终目标（Maximizing outcome for customers） [10:00] Logan Kilpatrick: 我点开日历应用，它就单纯展示日历，我完全不需要应对任何其他干扰。 [18:00] Logan Kilpatrick: 如果不去使用外部产品，你根本无法真正理解生态系统里正在发生什么。 [19:15] Logan Kilpatrick: 你绝不可能任由一个智能体随随便便就调用一万张 TPU，去启动一个根本没什么实际意义的巨型训练任务。 [24:25] Logan Kilpatrick: 到 2025 年底，每个人都将能够凭借直觉写代码（vibe code）来制作视频游戏。 [27:10] Logan Kilpatrick: 我想说就在今年。 [40:00] Logan Kilpatrick: Focus is the superpower of startups [45:00] Logan Kilpatrick: We can't let other people make the world a better place more than we can. [50:00] Sonya Huang: A really sincere human touch on the most important problem of our time

[15:00] Logan Kilpatrick: I describe it as like the Avengers of AI internally... some of the best people inside of Google trying to push the rock up the hill on this stuff. [20:00] Logan Kilpatrick: I used to kick around ideas and they were like slightly out of reach... Now I have the opposite problem. I'm kicking around an idea and I'm like, I could probably make this even more ambitious. [40:00] Logan Kilpatrick: the model eats the harness [45:00] Gavin Belson (quoted by Logan Kilpatrick): We can't let other people make the world a better place more than we can.

[0:00] So we could edit this set so it looks like we're... We should, yes. Yeah, I want this where we were talking off camera. Like we should do that for the intro because I think it just like makes all this stuff more capable. I've seen these examples of like such subtle nuance that like make me appreciate that it's like the world understanding playing out. I was giving a talk and was on stage with my friend Tulsi who leads the model team. I had mentioned to someone in the crowd to like edit the video and they literally like took the picture, edited it with Omni in real time. And this like dog came on the stage in the edited version. The other guests sort of like look down and see the dog. They like chuckle a little bit. This is while I'm like opining about whatever AI nonsense. Maybe they're just laughing at your jokes. Yeah, it was not my jokes. They laugh at the dog coming up. It jumps onto my lap. I sort of like acknowledge the dog. I keep talking. I'm like petting it or whatever. And just like there's like so much subtle, subtlety in getting that right. And the model crushed it. And it's just it's very interesting and like still trying to like absorb and digest like what that means for, you know, the way we make content and all these other things. That's so interesting. I'm delighted to have Logan on the show. Logan runs Google AI Studio and the Gemini API. You spend a lot of your time thinking and building for the next generation of builders. Yes. So I'm excited to talk to you about everything from agentic AI to AI coding, world models, and more today. and right off the heels of Google I.O. So what better timing? Yeah, I'm super excited. Thank you for having me. Wonderful. Let's start with agentic AI. So Sundar opened I.O. by calling this the agentic Gemini era. What does agentic AI mean for Google? Yeah, it's a good question. I think, and we were, we sort of, if you followed closely, we did sort of mention some of these things back with like Gemini 2.0, which I think was like a little bit early. And so I think this era, this like Gemini 3.5 era, feels like it's actually becoming true now. And we're in the era of agentic coding or agentic products and everything agents, as far as Gemini goes. I think for us, this agentic layer, and I think we announced this actually at IO sort of being powered by the anti agent harness is this like additional through line for Google that sort of connects all of our products that they sort of like based on now And so historically like prior to Gemini there actually like wasn a through line for the you know, probably sub hundred number of Google products that we have, the 50 Google products we have. There wasn't a through line. We had Gemini. It became this through line. Everything is now sort of using Gemini in some way. That's now becoming true for anti-gravity as sort of all of the products rebase to become sort of like agentic native products and like actually taking action on behalf of users and helping them get things done. You see this like new through line emerging, which I think is actually really, really interesting. And sorry, help me with anti-gravity is the IDE, right? Or the non-IDE. Yeah. Anti-gravity is a lot of things, which I think is sort of, again, is an opportunity for us. You have sort of a core IDE. You have sort of like the agent first experience if you want it on the web you have a cli you have an sdk so i actually think and i don't know how much we've framed it this way but like it really is an ecosystem of stuff that we built and it's designed to sort of like meet developers wherever they are so you could use it through the gemini api if you want to and you want a managed agent that you don't have to do any of the sort of infrastructure work for um and then the most interesting bit is like it not just the ecosystem of anti stuff It also powering like literally it the same harness is actually powering all the other Google products So anti will be powering a bunch of agent stuff in search in the Gemini app across like cloud and Studio which is really exciting I see. So it used to be the Gemini API. So like the language model was a through line in terms of how I gets baked into every Google product. Yeah. And now it's not only the API, it's the coding harness. Exactly. that's being used in each of these products. And therefore, it's a coding agent itself that's driving more agentic properties. Yeah. The products. Yeah. Fair description. Fair description. I think more generically, too, it's just like it is the agent harness. I think like coding as sort of like a specialized use case of the agent harness, I think, is obviously powerful. But it is like coding has proved to be the general purpose agent harness in addition to also working really well for coding. Are agent harness and coding harness synonymous or not? There's definitely nuance. I think there's like optimization that you can squeeze out of like specializing and actually you see this where like the, you know, technically the agent harness that gets used for the way that AI studio uses it is like a little bit specialized for, you know, the vibe coding use case and the way that the Gemini app is using the agent harness is a little bit specialized for the sort of consumer always on 24 by seven agent. So I think. [5:00] You have that base harness that probably has like 80% of the same stuff, and then you specialize for coding or for whatever the use case is. How do you think about the cannibalization of the existing business, especially now that you are going much more aggressively into agentic properties? Because I could see, for example, if all you're doing is search or summarization, there's not as much of a cannibalization fear. Whereas if you're actually going through my emails, replying to them for me, like, am I even going through my email anymore? And so I could imagine that there's actually just fewer human eyeball hours on your products as a result of having more agentic capabilities. Is that fair or how do you think about the cannibalization? Yeah, it's interesting. I think one sort of observation I have is that like at the beginning, and I think Sundar has done a great job of sort of talking through this is at the beginning of the sort of current AI era, like everyone assumed that AI being able to answer questions for you was going to be like negative sum for search. And actually what's ended up happening is it's been incredibly positive sum for search. Like people are searching more, people are doing more. And agents are searching too Yeah and agents actually again there like this whole market that spawned at the same time that agents are doing more at the same time that humans are also searching more And so I think it will be interesting Obviously there a finite amount of like human time in the world. Um, but from, from like my early feelings of how a lot of this is playing out, it does feel like it's, it's very positive. Some from like an ecosystem value creation, like how the human behavior aspect of it turns out, I think is like somewhat clear in the next one to two years, much less clear, you know, three to five years from now when the technology has improved and the products probably look a little bit different than the way that they do. But ultimately, like that is the success of product. I think like we have a bunch of conversations with Demis all the time. And it's like the point of building the technology is so that it can go and do stuff for you. Like that point, like success for Google, like probably doesn't look like, you know, maximizing eyeball time in front of our products. It's like maximizing outcome for customers to like do the thing that they want to do so that they can go and live their life and do what they want and so I feel like you'll you'll probably see us go down the route of like maximizing outcomes for customers and like not maximizing eyeballs yeah I have this term stuck in my head agent-led growth like it seems to me so I'm using coding agents a lot in my personal time and you know I just let the agent make all the infrastructure choices for me I like I don care which database you tell me Yeah And and so and the reason I ask is you know it's true in coding today. I would imagine it's maybe going to be generally true for a lot of things, let's say shopping down the line. How do you think that's going to change how advertising works, how value capture works for for the aggregators? It feels like it's a very similar trend. This isn't perfectly true, but a lot of these things are just like proxies of each other. The way that SEO works, I think, is directly correlated with the way that... I forgot what the term now for it. It's like GEO is the generative engine optimization or whatever it's called. And so it does feel like there's a lot of correlation between the things. My guess is it looks like much less of a radical shift than I think maybe what we assume right now, just because these things compound on top of each other. If you were to grade the scale of agenticness in terms of crawl, walk, run, where are we in terms of how agentic the Google suite of products is? Yeah, that's a great question. It's definitely like crawl right now. And I think some of this is like all of the inherent product tension for Google is like you have what 13 billion plus user products And so like I actually think we have some more like labs like experiences where you probably closer to running or walking But I think like most of the product experience today is definitely closer to crawling. And I think that's just like the stewardship responsibility we have sort of building a product that's being used by lots of people. Like, I don't think the long tail of customers are like ready to have AI running and just doing all the things like they probably they want to be in the draft. privacy, they're cautiously taking the first step. And I think the Google team and search is maybe the most quintessential example of this. I think they have a lot of responsibility to actually do that in a way that it brings people along and doesn't just change everything of how they interact with the internet and the way they associate with products and stuff like that. Yeah. Which products do you think are closest to the walk? That's a good question. I think Gemini app is definitely closest to walk. And so for Spark, I think having a 24-7 always-on agent, literally going and potentially doing a bunch of actions on your behalf is definitely one of the frontier use cases. And I think you'll see, I think like anti-gravity is another one where it's like you could have autonomous coding agents, rebuilding operating systems and doing billions of tokens and spending thousands of dollars on your behalf. And I think those are, again, like more... [10:00] like they're in GDM as well as another angle of this. So I think like GDM is taking like very much like a frontier look at this, where I think like the rest of Google's products, I think are like more incrementally getting there, which again, makes reasonable sense to me. Yeah. Do you think that Google ends up with one, two, three product surfaces for using AI or thousands? It's tough. I think a lot of this is actually baked in just like how humans consume products. And my sense is that there's something nice about having this compartmentalization and this specialization of products where it becomes, if you end up with a product that is doing everything for you, inherently there's more work involved in using that version of the product. I think I think would be like the default state. I think maybe somebody will spin together like the truly magic experience that doesn't make that true. But I think I think the long tail of folks end up having to spend more mental energy and more time to actually like get the general purpose product to do the thing that they actually want to do versus like something nice about I click my calendar app. It just shows me my calendar like I don't need to worry and deal with anything else. This is my hot take for why slide decks have existed for so long of just like you know the thing the piece of information you want it to be exactly in the same place And I think we as humans are just actually very used to that as opposed to the idea of a generative interface sounds so cool to me, but it's like, do our brains really, isn't it just more cognitive overhead for us? It definitely is in certain cases. And I think somebody needs to, again, there's a lot of incredibly smart people in the world. And so maybe somebody will find the experience that like makes it feel more natural. But to me right now, So I'm maybe not 10,000 is the extreme version. I'm guessing it looks more like more products going after sort of like different. And maybe the other answer is like, I don't know what it looks like for Google. For the ecosystem, it looks like a lot more products, I think. Like, and that's really exciting. I think like how Google will end up strategically deciding, like, do our customers want to deal with us having 10,000 products or would it be better to only have three? We'll come down to like a strategic decision for us. That totally makes sense. When I talk to companies in the enterprise, they say, you know, everyone's talking about agentic AI, but the only place they've seen agents really working is coding agents. Do you agree or disagree with that take? Yeah, I think it depends what your bar for working is, which I think is a lot of the nuance of this. I think if you truly trying to offload very complicated tasks for domains in which the models haven actually crossed the threshold of quality then like I think that definitely true Like it not going to solve the problem But this is something that I want. I wish we could like measure. A good example is like Open Router, for example, is like measuring, you know, the total token consumption that's happening. And so you can sort of like see these trends play out over time of like how much more intelligence is in the world, you know, now versus a year ago. in parallel the thing that i'm actually really interested to measure is like how long is the average like thing the average like agent run or the average task actually taking place and i don't think it's something that they publish but i feel like they probably have interesting data i'm sure there's others because because i i do think you're like seeing these like new model capability lands or new model drop and and it's like spiking up and and maybe the the curve is still like very low right now, but like you're seeing those like early signs of it spiking up or to like long running tasks and all the model labs are talking about like, we released this new model and it did, you know, three days of autonomous work or whatever it is. That's the extreme. But I think in practice, you're seeing that like trickling up like pretty, pretty quickly, which is really interesting. So even if the enterprises haven felt it outside of coding like they are going to like this year as sort of a bunch of those other use cases get much better as well From like a you know from the DeepMind perspective do you think Long Horizon agents is like a KPI that matters Is it the KPI that matters? It definitely matters. I think for DeepMind, like we're doing lots of things, which we can talk more about later. Like there's, you know, a huge portfolio of different bets that are taking place. Long-running agents obviously matters a lot. And I think also like specifically coding agents and that matters a lot. Like it clearly is an accelerant of like every other part of your business if you have a great coding model. And so making sure we have that, I think, is super top of mind. Got it. I'd love to shift gears a little bit and talk about coding. Yeah. OK, I'm going to ask a hard question. A lot of my developer friends were using Claude for a long time. Open AI saw that declared code red. Codex is now really good. I would say my friends are maybe split 50-50 now in using Claude and using Codex. I don't hear a ton of them using Gemini, which has always kind of puzzled me. What's going on with that? Yeah, it's a great question. I think there's one part of the story that I'll add, which makes it even more interesting, which is December, the narrative was that... [15:00] Google had won. And when we landed Gemini 3, I think it was like such a profound improvement from a model capability perspective. I think a lot of the narrative was like Google has taken a huge leap forward and made that happen. And I think what was interesting to see sort of as an ecosystem participant is like how not how quickly that narrative shifted, but just like the next wind of the narrative obviously was like all the agent encoding stuff that happened over the holidays and then into January and beyond. And that was not that long ago. And so it is a... I feel like we've been in warp speed ever since. Yeah, for sure. But it's a meta reminder of just how fast things can change. I think the observation is not unreasonable. I do think what's happening behind the scenes for us is trying to push the frontier as fast as possible on coding. And so I think anti-gravity actually like is an important part of that. I think one of the takeaways is that it's actually really hard to make a great coding model for this developer use case of like really long running SWE work if you don't actually have a product that does that. And so I think like Google realized that That why the sort of like windsurf deal happened It why those folks came over and then ultimately built anti and sort of we been using internally actually and soon our show this at IO just like the graph of growth of token consumption inside of Google So you sort of like you need that engine to spin and sort of the meta comment again is like the engine is spinning. It takes time in order to like actually make model progress. But I'm super confident. I think the folks, the group of folks who we have working on code is like, I describe it as like the Avengers of AI internally. And so it really is like some of the best people inside of Google trying to push the rock up the hill on this stuff and taking it super seriously and trying to push. I think 3Flash, notwithstanding some of the conversation about the price and stuff like that, is sort of a step towards actually starting to bring a lot of these capabilities and the fruits of that labor paying off. It's a Flash model that's better than any Pro model we've ever released from a coding standpoint. And the Pro models were really good before. So there's another thread of this also, which is like everyone forgets that there's like pre-training windows. And I wonder, somebody should track this online, which would be interesting to see. Meaning the big run what clusters have been available Exactly The big runs are like are an interesting thread of this And so it like it might look from an external perspective that like oh you you super behind in some way And like, actually, you miss all the context of like where the big runs are and where the large pre-training runs are. So I think that that also like obviously there's pre-training has historically been like a massive strength for DeepMind. Like we have some of the best people in the world. And so excited to see sort of the fruits of that labor and everything else that's happened. Like 3.5 flash was like all post-training gains, which is really cool. So a huge, huge testament to the team, the work that that team did to actually like make the level of gains and like surpass the previous pro model literally just with post-training, which is awesome. Hmm. How religious are you all about dogfooding internally? Like are, for example, are DeepMind folks still allowed to use other models or is it like you guys are using the Gemini harness now and we have to make this really, really good? Yeah, there's I mean, I think people it's so healthy to be using other models just because like it's sometimes hard to like actually grok what's happening in the ecosystem if you're not to like I use all the models. I use all the products. I think like, you know, folks across the rest of DeepMind are doing the same thing. You definitely have to use the Gemini models though It just like great from a feedback flywheel perspective And it part of how they get better is like DeepMind has and Google more broadly has like 100 plus incredible engineers who are using the models and giving feedback And like it should be a competitive advantage for Google because we have that scale of sort of engineering resources and like the depth of the talent and can run, you know, A-B tests and live experiments and all that stuff. So I think you have to use all the models. But I think for for the majority of folks, it's like Gemini as the daily driver, which is great. Do you believe in this narrative around like a like a soft takeoff of like once you have a good enough agentic coding model, then it accelerates the pace of research progress and like it's a self-reinforcing cycle? It seems obvious that that's true, but maybe I'm I'm too I've drinking too much Kool-Aid that that's that's the case. Are you seeing the signs of it yet? Yeah, I mean, you definitely see some signs of this. I think the signs that are still early is doing this from a model perspective. And I think part of the context of that is the resource allocation for some of these larger training runs is just significant. And so you definitely still have a human in the driver's seat of making those decisions because you're not going to accidentally take 10,000 TPUs to go kick off some job that actually doesn't make that much sense. [20:00] A product perspective, you for sure see it. Like, I think we're seeing this on our team. Like, we've built mobile apps using anti-gravity and, like, we'll launch them to the world, like, faster than I think any team at Google has ever built a mobile app. Josh's team did this with the Gemini macOS app and sort of, like, end-to-end delivered an app sort of faster than any team had ever delivered a Mac app at Google. And it's because of identity coding. And so it's great from a product perspective. I think you've said in the past that if you could have a system that could build anything with code. Humans can't compete on the same level and that's narrow super intelligence. Do you think we've reached that point? It is interesting. I think this narrow super intelligence example is interesting to see. Obviously, it kind of feels that way for coding right now where coding is just so good that it does kind of feel like narrow super intelligence. I don't know. It depends how you actually end up the details of quantifying this. But I think the The important thing is like, to your point earlier, it works incredibly well for code. And so it would be great if it did a bunch of other things, but it's actually just like so impactful that it can be great at code And so I spent a lot of time just like letting that fact sort of just like wash over me because I think it like obviously building AGI is super important and very interesting but like building AGI, if it sort of like takes away from the story of like the current present capability of the technology, I think is actually like kind of a bad sort of like trade-off. And so I'm trying to like always hold these two things in my head equal at the same time, which is we need to build general purpose technology, but obviously it's so impactful to have this thing. And it feels like it hasn't taken away sort of, it's been one of the best positive outcomes is that I feel like it hasn't taken away from like human developers. It really does feel like an accelerant of what human development, like I, as a human developer feel like I have more agency in the world. I feel like I can tackle, this is my personal experience. I feel like I can tackle more ambitious problems. I feel like I used to kick around ideas and they were like slightly out of reach. And I would just be like, ah, wouldn't it be nice? And now I have the opposite problem, which is I'm kicking around an idea and I'm like, I could probably make this even more ambitious. And sort of it does it adds a different layer of sort of responsibility or like some different layer of burden actually because I like oh I can just like do the the sort of mvp of this like i actually need to like go 10 steps further because the technology enables me and like resetting my my level of ambition i think is something that i i've also spent a bunch of time thinking about but i think that will happen in other these like vertical super intelligence domains um which will be interesting and it feels like we're going to get a bunch of those before we've like solved, like it's almost like jagged, like jagged super intelligence, I think is what we'll end up with. What verticals do you think we'll get super intelligence at next? That's a great question. I do spend a lot of my time, too much time probably thinking about coding these days. So I'll think for a second of like the other, the other domains. I think part of this is like things that have like better verifiability, obviously are like the ones where you'll see the gains happen more quickly. So like things with like math and finance, actually, like science could be a really interesting one, like it would be fascinating to see, like some of these domains where there's some level of verifiability, like actually like really start to take off, which would be cool. And I also think like an important thing in this like broader narrative about just like what what impact AI is having on the world like you almost like want that to be the case in the sequencing of things that work you want a lot of these really really good impactful positive things for the world to happen as early on as humanly possible so that folks understand what the potential positive impact of the technology is. So I think science could be a really interesting one. Obviously, there's all this stuff happening right now with math proofs and stuff like that, which I'm not a mathematician, so it's somewhat over my head. but um i saw a great tweet the other day uh why did air those have so many problems exactly that's a that's a good one i like that that is that's a good like t-shirt um so funny okay i but speaking of twitter i went through your twitter before this so i'm gonna read back another tweet at you the good thing about twitter is there's a public record of all your predictions i need to turn on that auto deleting feature or whatever it is um last october you tweeted everyone is going to be able to vibe code video games by the end of 2025. Yeah. Did that end up being true? It feels close. And I think there's, I mean, obviously not AAA games, like you're not building the next Call of Duty or GTA yet. But I think it feels closer than it's ever been. And I think actually a lot of the interesting bit about video games is you actually need to end up building a lot of this [25:00] other stuff like models and we were talking off camera before this like 3js is a great example of this like 3js makes a lot of things possible that weren't before but there's still all these like rough edges that like just a coding agent doesn't solve and so you need like you know sprite generation and like the models aren't very good at doing that natively and so you need like some orchestration layer and tooling in order to make that happen there's a bunch of other things like that that like are core to like the gaming video game experience um that need to have a high degree of reliability that I think it feels like it's within reach, but actually like requires a lot of like product scaffolding work in order to create experiences that are like reusable and replayable and sort of like have the level of depth and requires a little bit of taste in there. Do you see people making a lot of video games inside AI Studio and the other developer surfaces that you have? Yeah. And so this was actually based on like us looking at the early data. And there was something like in AI Studio at the time, it was like 20% of all apps that folks were making were actually games. Like people are trying to build games. A lot of it. Is that the most popular category? It's not the most popular category anymore. Um, just cause I think like the, the ecosystem has shifted and like the user base has shifted, but it was a lot of a lot of games Um which is the most popular category I think it was like it like 20 percent like finance related stuff 20 percent Wow People like counting their money that much People like I think it it something around crypto actually I think is what people are doing a lot of stuff with with finance. A lot of like personal productivity things and a lot of media stuff, actually, because obviously the Google suite of media stuff is amazing. Yeah, has done a great job. But I also think GDM has sort of like a, obviously Demis cares a ton about games and sort of like started his career and doing AI stuff because of games. And so I think we'll have some interesting swings at this. And our team actually in Kaggle, which is sort of a bunch of the AI benchmarking stuff we do in GDM, sort of works with GDM to build this game arena, which is sort of our way of sort of like testing progress towards AGI, like using games as a proxy, which again is like very deeply rooted in GDM's history. So how close do you think we are to, you know, rando off the street with a good idea, can vibe code a really fun playable game? I want to say this year, I actually, I think it's, I think the model capability makes it possible. I think this is where like, I've gotten excited on the product side. And, you know, we were, again, we were also talking off camera about sort of like the startups in this ecosystem because it feels like it possible It doesn feel like there a gap in model quality It feels like there a gap in like you someone who knows what it takes to build a great game actually like putting the scaffolding together in the right way to make that possible. I think there are folks who are doing this right now. And so some of it is like a discoverability and awareness thing that like people just don't even know that they can do that. And some of it is just like, maybe certain categories of model capabilities are just like slightly off and we're like, you know, weeks or months away from like that chasm being crossed and then it just like working for most people. And so this is a good segue into, I want to ask you about world models next, but do you think vibe coded video games is more likely going to be, you know, game engine plus coding agents based, or do you think it's more likely to be world model based? Yeah, I think the, what will end up happening is the definition of world models will blur which we should talk about with Omni. And it will still, I think the coding agent will look like some sort of world model type system. But you actually do need to make world models useful for like real things you need like scaffolding And so I think there again there actually a bunch of interesting startups like doing work like figuring out what is the scaffolding for world models so that you can take them from these like very open inherent design of world models, very open-ended spaces and like do it in a tangible way so that it's like grounded in a use case that like you could use in a reoccurring way. That could be somebody maybe will figure out the scaffolding for world models to make games possible. but like the inherent nature of world models right now, I think make it so that it's like actually not well suited for like games in the current form. But the progress has been crazy. So who knows, maybe in like two years, the versions will be able to, but at least in the short term, it's like coding agent plus some sort of game engine, I think is like where you'll see way more alpha from a game's perspective. That makes sense. Okay. So you said the definitions of world models are blurry. Can we unpack that? Yeah. I mean, I think like Omni is an example of this. We launched this at I.O. You can take in any input, create any output. And I think Demis framed it to the world, rightfully so, as a world model because of just the level of understanding that it has of the world. I think that technically looks different than, and I'm not an architecture expert on the way that we've done world models before, but it is different from an architectural standpoint. [30:00] standpoint than what's happened in the past um which i think is positive because it's getting closer to like some of the ways in which it might actually be more scalable um and historically like it's been like super not scalable it's like very very expensive to run yeah traditional like online world models like genie being like yeah okay so if you think of traditional world models as being like an action conditions video model almost then like right now what we're when we when we say world model, what we actually mean is a model that has some understanding of the world as opposed to being strictly technically a action condition video model. Yeah. And so the interesting thing, though, is like it has understanding of the world, but then it also has that like really great. And that's where like the line is blurry to me, where it's like it can do a lot of those same use case. It's not real time right now, but like it can do a lot of those same use cases that you would describe or like visually could create with that same exact world model, which I think is what's most interesting to me. So I do feel like this like world model video model thing is gonna change and play out in a different way than was obvious before And how does it work under the hood Like whatever you able to share like is it Gemini plus video models Is it something different entirely It is a single model which I think is the important part This was actually part of the original desire was you were training eight different models to do all of those things historically It like you have a text model with the baseline Gemini model You have audio You have music models with Lyria You have Nano Banana You have VO video models We have a whole suite of audio models And like, it would be great for us, our customers, if you just had a single model to do all those things. So it is like a new setup that sort of makes that possible. It's not like routing to a bunch of different models, which like we you could have imagined we could have done something like that actually before and done like a Gemini Omni model. But this is like a true Omni model. and it's starting with like the the use case that works the best right now which is the why it's the one that's available is this like video editing capability the technically it's like functional with the other things it's just like the quality isn't isn't like perfect and it's not state-of-the-art so we haven rolled that out yet it also just like the first crank of the model turn on Omni it the Omni flash model the first iteration and so we have like much much more capable powerful versions which will be exciting to see So we could edit this set So it looks like we should Yes Yeah I want this Again we were talking off camera, like we should do that for the intro because I think it just like makes all this stuff more capable. And I've seen these examples of like such subtle nuance that make me appreciate that it's like the world understanding playing out. I was giving a talk and was on stage with with my friend Tulsi who leads the model team who I don't know if you've ever had on before but she's amazing I love Tulsi and in I mentioned to someone in the crowd to like edit the video and they literally like took the picture edited with Omni in real time and this like dog came on the stage and like the other in the edited version the other guests sort of like look down and see the dog they like chuckle a little bit this is while I'm like opining about whatever AI nonsense. Yeah, that it was not my jokes. They laugh at the dog coming up, it jumps onto my lap. I sort of like acknowledge the dog. I keep talking. I'm like petting it or whatever. And just like, there's like so much subtle subtlety in getting that right. And the model crushed it And it just it very interesting and like still trying to like absorb and digest like what that means for you know the way we make content and all these other things that so interesting yeah i i the biggest bull on generative media and what it means and i mean one of the things we've thought about for our podcast is the visuals matter as much as the content for sure um that's how you catch people's attention in the first place right and and so okay i'm excited to i'm excited to play with Omni. I'm excited too. And I think the, and I think you probably feel this way as somebody who makes content, but I've historically like been like very for myself personally, like I don't use AI to make any content that I produce. Like it's all my words. It's always my voice. It's always my image and picture showing up. Like I just, I feel like there's just like so much alpha and authenticity. And so like, I would much rather it be me than some AI version of me. What I like so much about Omni is that it's like not changing me. It is like changing a bunch of these other bits, which are not me. Like I didn't choose any of the like set around us or the coffee table. It's like so our words can stay the same and like you can change these bits that are like not personal and do something more interesting with them, which I think is really, really cool and feels. It feels like the version of what I want sort of like GenMedia to be, which is like not a bunch of [35:00] AI avatars. No Fruit Island videos? Exactly, truly. It really is. It's the original content. It's the person. It's like the personhood is there. It's just different and amplified. Super interesting. Okay, I'm excited to play with it. Yeah, we should send some prompts right after this and try some things. I don't mind the fruit videos though. I'm happy for a world of both. On the coding side, you launched the ability in AI Studio for people to vibe code Android apps. Yeah, yeah. I'd love to hear how that's going so far and where you want to take that. Yeah, it's super exciting. I think one of the strategic things for AI Studio, and actually this is based on a lot of the feedback from the ecosystem and actually from developers, from others, it's like so many Google products. There's so many different ways in which you touch Google through all these different journeys of building a startup or bringing an idea to your life. And so we have this like first class principle of like, how do we bring things into AI studio that make it so that you are exposed to other parts of the Google ecosystem without having to like go through nine different UIs across Google. And so Androids are like a great example not only of that but also of enabling people who wouldn have otherwise built an Android app And so I literally built my first Android app in AI studio Very cool to see It what is it Yeah, I just did like a plan, not a crypto app, just a plant one. I was planting trees in my backyard. Yeah. And so it was just like playing around with a gardening app as I as I was kicking the tires. I haven't had my like breakthrough idea yet of what I want for a mobile app, but I'm going I'm going to come up with something and see, go compete on the app store. Have you seen anything vibe coded, like really fly in the app store yet? That's a good question. It'd actually be interesting to like see some analysis. I don't know. I'm sure it's like accelerating a lot of things on the app store, but I don't know how much, like, I don't know anyone like personally who's, who's done that. It is interesting. And I was going to make the observation too, that I think the last time I checked the numbers we were viewing it this morning, it was like 350,000 Android apps built in AI studio since last week, which is crazy. And like, excitingly, it's like 350,000 apps that like probably no one was going to build before. A lot of these are personal too. And so this is where I think this like maybe GenUI is like farther out there. But I think like the idea of you building software to solve your personal problem is like very real right now And like people are doing that It like one of the most common use cases of a lot of these products And being able to like unlock a bunch of the native capabilities of the phone I think is also really interesting because you just have so much context that like in different places. So I'm getting very excited about sort of that opportunity. And Android feels like it's becoming the platform for builders. Does it matter that something is an app versus just like the web is so powerful now? Yeah, it's also very interesting to see that play out. Web is definitely powerful. There are certain things that the operating systems have that you just can't unlock, like lots of native richness that actually make experiences feel so much richer. I think about this for text messaging, actually, that the text messaging experience in all the main operating systems feel way richer to me than any AI chat app that I've ever used. Like if I could just talk to AI and whatever texting app I use, like I would be way happier than having to go to some other app. Because I think we're also just like conditioned on like the operating systems. So yeah, makes sense. Okay. I want to ask about the model eats the harness or the model eats the scaffolding. What are your thoughts? Yeah, I think it's true. And I think part of this is like what we have historically thought of as the model is not the model anymore Like when I think like two years ago when LLMs were popular it was like the model was like actually just a set of weights It was a set of weights and it was like really like how can you like as simple as possible send tokens in and get tokens out. And I think we've just like progressively step by step by step. We still call it the model. We still call it, you know, Gemini 3.5. if you still call it GPT, whatever, and Claude, whatever. But like, it's actually not just the weights anymore. It's like an entire expanding sprawling system that's built around the weights that sort of like enable a lot of these like next generation experiences from agentic tool calling to tool, you know, like all these hosted tools, search, code execution, et cetera. You know, the models are now being spun up in containers and sort of have an agent harness and all that stuff. So the scaffolding is oftentimes a couple of steps ahead of what is baked directly into the model. And then what ends up happening is the model eats that scaffolding and it becomes part of the native model system. And there's still value in having the external scaffolding in certain cases. And search maybe is an example of this. There's lots of folks who use different search providers and there's different use cases that you want. [40:00] The model can natively use search, but you also want something else. Code execution, another example of that. But it does feel like maybe the agent harness is the quintessential example of this right now, where everyone's like, we got to go build a harness. And the harness is where the alpha is. And I think that perhaps won't be true, at least in the way that we think of the harness today in 12 months. I think the models will have sort of just digested a bunch of that. It'll be upstreamed into the model. and the alpha will be somewhere else now. It won't be in sort of trying to spin your own harness because the model just like does it natively. But I thought that the part of the reason why people are building their own harnesses is because if you use a harness from any given model provider, you're locked in, right? So a lot of the application companies want flexibility, which is why they're building their own harnesses. Yeah, and I think that's part of the scaffolding story is like that starts out perhaps true, but then as the model capability improves, it becomes less true over time, actually. I think the model, that like you you don't have a generalized model if it can't use another harness and so it is and is important and i i mentioned this uh in another conversation with someone a few weeks ago but we need something like harness bench which is like actually measuring like how good are all these different models at adapting to all the different harnesses i feel like that seems like a reasonable thing we should measure as an ecosystem And I be curious to see what models are actually best But I think over time you expect they be able to use every harness unless you're completely out of distribution, which in that case, you're still going to be completely out of distribution, even if you're using your own harness. So not sure it matters much. Fair enough. What about the application layer? How do you think about where independent companies can have a hope of surviving when the model eats the harness and eats the stuff around it? Yeah. It's an interesting story that both of these things feel true. Both on one hand, everywhere I look, I'm like, there's never been more opportunity to go and build something. At the same time, obviously, the models are doing more than they've ever done before. I think there's like, you know, there's that thread of capability overhang, which I think there's a huge amount of alpha and there's the thread of the model companies are like going after these like very general problems. And there's just like so much value in these like verticalized domains. If you have expertise in that domain, you sort of like know the customers, you know, the ecosystem, like it's just you can really like run laps around even the best model labs because like focus is the like super power of startups Like if you can focus you can do anything And if you look at all of the companies that are big or doing lots of stuff, like there's just not a lot of focus. And for some, for some reasons, like rightfully so, because maybe I'm, I'm overly justifying, you know, Google strategy, but like, we just have a lot of products. We have a lot of users. We have a lot of different things going on. And so like, we actually can't focus in one domain. We have an obligation to do a bunch of things as a big company. I think that's not true for startups. And so I think like 24 months ago, we were all asking ourselves like, oh, wow, it seems like it seems like the opportunity space is shifting. And maybe it's possible one of the outcomes is there's less opportunity for startups in the future. That feels like so far in a way, not what has ended up playing out, which is really positive. If anything, it feels like there's just even more opportunity than there was. Like now coding has helped you like close the gap on like larger companies that have like established code bases and all this other stuff because you can just like run way faster and write software quicker um the agentic like primitive is like a new category that you can sort of build products around that like actually in a lot of cases to the conversation about like the risks involved with building like there risk involved and so like what your like the risk appetite of different companies is different And so if you willing to take more risk in some domains like you can win a user cohort who like interested in also taking risk There so much opportunity Awesome. I'd love to talk about Google DeepMind's culture. And I'm curious, what does it feel like to be inside GDM right now? You know, we had Demis at AI Sends. He was so inspiring. I've heard Sergei's back. You guys have Noam Shazir back. Walk me through what it's like to be at GDM right now. It's incredible. I do try to take it all in because it's like a moment. I try to reflect as much as possible in the chaos of all the things that are happening, just because there's so much cool stuff going on. GDM's culture is interesting and maybe three observations. One, back to this thread of focus, we're doing a lot of things. And so I think you see sort of, I think about this a lot, from a portfolio perspective, I think we have one of the strongest portfolios, which is really exciting. But you do see these moments where another lab or another company, whatever it is, will pull ahead in a certain area where we underinvested, just like hadn't been focused enough in that domain. And it's cool to see like the the way we go about trying [45:00] like close that gap. I very much, I very much appreciate it. I think I've, I've watched the Demis thinking game documentary a few times. And like, you see sort of like a lot of like details of that, like original culture and just like the way that strikes work and all this stuff, which is actually really similar today is like, you just get a bunch of smart people together and like go solve the problem. And I love that. And it's like very cool to be a part of. Another one is this, I think you see the culture permeate from like who the leaders are. And as I maybe this isn't like a perfect characterization of the ecosystem, but like Demis is a Nobel Prize scientist and like the sort of OG of a lot of this stuff. And sort of you feel that in the deep mind culture. I think like Sam is like the, you know, maybe one of the world's best businessmen ever. And like you sort of see that in the open AI culture and the way that they go about the world. I don't have a strong sense of who Dario is, but I think Anthropic is a very interesting place. And you sort of, at least as an external observer, he seems like an interesting guy and so somewhat esoteric And so it seems like they sort of like that in the DNA and the culture of the company You know the other labs are interesting But I like this like very scientific approach to the world and the way that like Demis looks at like the reason he's doing this and the reason they started this mission was like literally to like solve disease and all these things. And it's like so easy to get. And again, I'm always trying to pull myself out of the moment, but like it's so easy to get lost in this like competitive race of who's pushing a number higher on sweet bench or whatever it is it's very easy to lose sight lose sight of like the reason we're doing that is so that like we're can solve problems that humans actually have and there's a uh my my favorite quote from all of silicon valley is something like you know we can't let other people make the world a better place more than we can which is like what this moment feels like and i just the gavin belson quote and i think about that all the time and it's like we're all fighting over who can make the world better more than the other person which just like when you frame it like that seems really goofy to me um and so it's very much not zero sum um and i think that's like a a way of looking at the world i think the last thing about deep minds culture is like we're very we're sort of the engine room of google uh which i think is like literally the twitter bio now of the the deep mind twitter account which I love You man the DeepMind Twitter account I don I don want any responsibility manning other people accounts online too much responsibility to do that But it does feel like that, too. So it's like on one hand, you have sort of like the deep rooted lab culture. On the other hand, you have sort of like all of these partners across the Google ecosystem that we're collaborating with. everybody from Android that we talked about earlier to Google Cloud to, you know, Gmail, to Workspace, et cetera, et cetera. And so it's an interesting blend of like, I think there's lots of research work happening, but like there's tons of applied work that's happening to like, actually like work with some of the, like the forefront customers, like deploying Gemini to billion user products is a problem that like only two companies in the world have. And we have 13 of those products. And like we, the, you know, Google goes through this all the time now. And it's such an interesting place to like see that happen and see the innovation that takes place in order to make that actually possible. And I feel like it's you can only do that inside of inside of Google, which is really cool. Beautifully said. Did they did it give them a lot of heartburn when you joined and were tweeting a lot? That's a good question. I get sign off from I very one of the the silver linings to my Google experience has been just like how great that group of like folks across marketing comms are to work with And I think like, you know, their job is protect Google, make sure we tell the right story and make sure a bunch of bad things don't happen. And so I have a ton of appreciation and partnership with them. But it's been an incredible experience to like be able to go try to tell the story that resonates with developers in a way that feels authentic and not have a huge amount of, you know, I don't have to get my tweets approved all the time and all this stuff. Like it's a very, very positive culture. And I think hopefully I am always trying to walk the line of not not burning the trust and goodwill that I've accumulated with those folks. But it's been super positive because ultimately, I think it's like it's really hard for Google to tell this like authentic story. So there's like it's a big company. There's a lot of people. There's a lot of opinions. And so you take the like magic of Google and you water it down through like a lot of people and a lot of process. and you actually, you miss the beautiful story, which is like Google's doing the most interesting technology in the world and like helping our users with some of the hardest problems in the world. And it feels, it's a privilege to like get to help tell that story. So it's a lot of fun. [50:00] I love what you're doing. I love what Josh is doing. You guys have put a really sincere human touch on, as you put it, the most important problem of our time. Thank you. Wonderful, Logan. Thank you so much for joining me today. This is a very far-ranging conversation, everything from agents and coding to world models and harnesses and GDM culture and lots of nuggets here. Thank you for joining me today. This is a ton of fun. Thank you for having me. And I'm excited to see what the folks cook up of where we've been sitting this whole time, maybe in front of us. Maybe there'll be a dog. A dog. You can make my dog dreams come true. I love it. Awesome. Thanks, Logan. Of course. Thank you.