Fireside Chat Designing for the Agentic Era | Michael Martinez & Alex Cornell from Meta

炉边对谈：面向智能体时代的设计 | Meta 的 Michael Martinez 与 Alex Cornell Michael Martinez 与 Alex Cornell 回顾了自主个人智能体 Muse 的开发历程与外界反响。该智能体在专属云实例上持续运行，而非依赖同步的轮换对话模式。他们探讨了从确定性界面设计转向非确定性系统的范式转变，将现代设计师的角色比作大厨为用户备齐可自由搭配的灵活食材。Cornell 和 Martinez 强调，持续的后台主动性、快速迭代的原型制作，以及工程与设计之间传统角色的模糊化，正在重塑现代产品开发模式。最终，他们指出必须保持敏捷而非死守僵化的长期路线图，并得出结论：智能体工具正在拓展而非削弱人类的创造潜能。 Meta 的 Muse 打破了传统的轮换对话式聊天机器人范式，为每位用户分配一台专属且自主的个人云电脑，即便在应用关闭时也能在后台持续执行任务。 (0:04) Muse 从最初构想到正式推出历时约八个月，期间跨越了模型微调、UI 设计和分布式云基础设施等领域的紧密跨职能合作。 (4:14) 设计团队摒弃了长篇 Markdown 文章的形式，转而采用简洁的短信风格气泡和可定制的智能体个性，他们发现随着系统能力的提升，界面复杂度应当相应降低。 (5:19) 基础模型能力的快速变化使得死板的 18 个月产品路线图被淘汰，规划周期被压缩至大约六个月，且架构的敏捷性被置于优先地位。 (18:29)

In a fireside chat hosted by @Scale, Meta's Michael Martinez and Alex Cornell argue that agentic AI shifts software from static, turn-based utilities into proactive, open-ended companions, fundamentally restructuring product design and engineering workflows. From Turn-Based Chats to Dedicated Cloud Environments (2:57) Traditional generative AI chatbots operated on a synchronous, turn-by-turn basis where task execution ceased the moment an interaction ended or a browser tab closed. In contrast, Cornell explains that Muse assigns each user an individual virtual computer in the cloud, allowing the agent to persist and autonomously execute background tasks without continuous human prompting. Martinez notes that this architectural shift required Meta's infrastructure teams to develop new techniques for provisioning and scaling dedicated cloud environments for potentially billions of consumers, presenting operational hurdles far beyond the stateless inference endpoints used by earlier AI systems. Human Metaphors and Radical UI Simplification (5:19) Rather than adopting the prevailing late-2025 AI paradigm of dense, multi-page markdown reports loaded with toolbars and headers, the Muse team modeled the interface on standard text messaging. Cornell notes that as underlying system power expanded, the interface was intentionally simplified, using the guiding heuristic of how a user would communicate with a capable human assistant. Positioning and personifying the agent at the top of the interface—such as Martinez's shark avatar named 'M' or Cornell's agent named 'Jolly'—encouraged users to interact with the software as a collaborative partner. Cornell identified unexpected proactive notifications, such as Jolly unpromptedly analyzing past golf goals while the app was closed, as the key milestone that distinguished agentic interaction from reactive chatbot tools. Non-Deterministic Software and Additive User Workflows (12:34) Because Muse allows users to invent unanticipated pipelines—such as compiling and printing a daily physical family newspaper or chaining a concert tracker into automated ticket buying—product designers can no longer rely on rigid feature specifications. Cornell compares agentic product design to a chef preparing ingredients: designers supply the scaffolding, safe boundaries, and underlying modules, while end users determine the final assembly. Workflows in this paradigm become non-deterministic and inherently additive. Unlike conventional software releases that arrive at a fixed finish line, agent-built applications are continuously modified by users who incrementally introduce new telemetry, data inputs, and background automations over time. Blurring Craft Roles and Shortening Roadmaps (14:25) The rapid advancement of agentic tooling is eroding traditional organizational silos between engineering, product management, and design. Cornell notes that despite lacking a background in mobile engineering, he now spends most of his working hours writing code and maintaining an iOS sandbox prototype alongside the main application. Similarly, Martinez observes that technical leaders can now prototype directly rather than passively evaluating review documents. Consequently, long-range planning has fundamentally shifted. Martinez and Cornell agree that rigid 18-month roadmaps reviewed biannually are obsolete when underlying foundation model capabilities shift every few months; product planning is now effectively constrained to flexible horizons of roughly six months or fewer. While acknowledging the heightened tension of building without an established design playbook and the operational burden of scaling personal cloud infrastructure, Martinez and Cornell reject the premise that AI will automate away human developers, arguing instead that flexible human judgment remains essential for guiding non-deterministic systems.

Meta 的 Michael Martinez 与 Alex Cornell 阐述了个人代理产品 Muse 的研发实践，指出软件正从单轮被动对话演变为具备自主背景执行能力的代理形态。这种非确定性系统的兴起重构了界面设计、底层云架构与跨职能团队协作模式，要求团队以高敏捷度替代传统长期路线图。 超越对话轮次的主动式代理交互 (0:04) 传统生成式 AI 工具（如早期的 ChatGPT）受限于严格的回合制交互，用户在输入框键入提示词、等待响应并在会话关闭后终止联系。Alex Cornell 指出，Muse 的根本突破在于它是一个能够自主替用户办事的实体。通过为每个用户实例配置独立的虚拟计算机环境，系统可以在应用关闭或用户参与其他事务时持续执行后台任务，实现异步与长程生产力，彻底脱离了需要人工实时触发每一步的被动模式。 这种自主性催生出大量未预设的自发工作流。例如，有用户设置让 Muse 每日清晨自动抓取家庭日程、往年照片与新闻天气，自主排版并物理打印成一份定制的家庭晨报。Cornell 也分享了自己的实例“Jolly”会基于长期积累的喜好背景，在无提示的情况下主动推送针对高尔夫球技改进的练习建议。当代理具备长期上下文与主动推送机制时，用户的心智模型便从使用独立工具软件转向与随身教练或同伴协作。 极简人际隐喻与独立云环境架构 (5:19) 在交互设计上，团队拒绝了当时行业内充斥密集标题、分块排版与宽屏 Markdown 文档的静态报告样式，而是全面借鉴日常短信气泡模型。Cornell 认为，随着底层系统能力的增强，用户界面反而应当更加简化；复杂的真实人际交流绝不会以长篇大论的文档形式发生。团队在界面顶端设立具有个性化命名的代理身份（例如 Martinez 将其命名为性格干练的“M”），并以“高能力人类助理会如何应对”作为各项设计分歧的衡量基准，从而降低用户的上手认知门槛。 在底层架构层面，为了支撑长时运行与任务自主性，Muse 摒弃了传统的无状态共享推理接口，转向为每个用户分配专属的云端个人电脑。Martinez 承认，这一设计虽然保障了用户隐私和数据自主权，但也带来了极高的运营挑战，因为在云端为全球潜在数十亿用户分别部署和扩展虚拟计算机是一项艰巨的任务。这也迫使基础设施工程团队与核心产品设计人员必须展开无缝的跨领域协作，在模型调优、界面打磨和分布式资源调度上同步推进。 非确定性系统与模块化累加扩展 (10:39) 与传统软件明确的功能终点不同，代理型软件呈现出天然的“累加式”（additive）特征。Cornell 列举了演唱会追踪工作流的演变过程：用户最初仅建立喜爱艺术家的演出监测清单，随后顺畅扩展到本地场次提醒、自动票务购买以及与个人音乐库的深度联动；个人的高尔夫球或国际象棋分析系统亦能随时接入外部传感器日志。整个产品生命周期不再有固定的“完工”界限，用户可按需链式附加能力。 这种不可预测的扩展模式促使设计哲学发生根本转变。Martinez 对比了传统社交网络严格依赖数据定义静态功能集的研发范式，Cornell 则将新型设计角色比喻为“准备食材的厨师”：设计师不再是端出一盘摆盘固定的成品菜肴，而是搭建容器、底层架构与防护边界，提供丰富的原料和模块，让最终用户自主拼装定制方案。设计的核心焦点从而上移到更复杂的结构性问题，包括控制组件编排、约束条件设定以及保障自治组件间的可靠交互。 职能边界融合与缩短规划周期 新一代 AI 辅助工具极大地放宽了个人执行力，打破了设计、产品与工程之间原有的专业壁垒。Cornell 透露，尽管自己此前没有移动端开发背景，但如今大部分工作时间都在直接编写代码，并维护着一套配合 Muse 主应用的完整沙盒 iOS 原型。以往产品经理因工程带宽稀缺而不得不严格裁剪功能，如今团队成员则能绕过冗长的审批链路，直接以原型直观呈现构想，分工更多取决于个人在系统架构或后期增长上的偏好而非职位头衔。 这种生产速度也让传统的长周期研发规划失效。Martinez 与 Cornell 指出，在底层基础模型每隔数月便出现能力跃迁的背景下，以往制定 18 个月路线图、每半年校准一次的模式已经过时。除了保留少量初始未上线的高价值功能备忘外，团队将具体规划视界缩短至不超过 6 个月。研发策略的核心转向纯粹的架构敏捷性，即保持系统设计具有弹性，以便在新模型能力释出时能够立即无缝适配和调整方向。 Martinez 与 Cornell 将代理时代的应对准则归纳为果断（Boldness）、灵活（Flexibility）与敢于涉足未知风险（Taking Risk）。针对技术或将使软件创作者被淘汰的行业担忧，对谈强调人类的创意构想与判断力依旧是核心，自动化代理与高阶工具的普及正在前所未有地放大开发者与终端用户的创造力边界。

## Background 该视频由 YouTube 频道 @Scale 于 2026 年 10 月 30 日发布。视频内容为 Scale 举办的一场炉边对谈，主讲嘉宾为来自 Meta 的 Alex Cornell，由 Michael Martinez 担任主持，深入探讨智能体时代的软件设计。页面未提供具体的视频简介说明。 ## TL;DR 对谈围绕 Meta 个人智能体产品 Muse 的研发故事展开，探讨了从传统单轮对话工具迈向拥有独立云端计算环境的主动型 Agent 的范式转变。Alex Cornell 阐述了系统能力越强、界面越应极简的设计哲学，并强调设计师的角色已从确定性的组件排布转向为非确定性系统搭建容器与脚手架。同时，对话揭示了 AI 工具如何消融设计师与工程师的职责边界、将传统 18 个月的路线图压缩为极致敏捷的原型迭代，最终总结了打造此类产品所需的胆识、弹性与冒险精神。 ## Atmosphere & style 整场对谈洋溢着技术突破带来的极高兴奋感与隐隐的紧绷张力，两位从业者仿佛是在未知荒原拓荒后进行推心置腹的经验复盘。对话氛围真诚且充满共鸣，从日常生活的奇妙用例自然延伸到深层的系统架构哲学，既有管理层亲自写代码验证原型的敏锐，也有面对行业“换挡加速”时的敬畏。听众能清晰感受到一种技术乐观主义的情感底色，见证开发者在被大模型重塑工作流后迸发出的职业新生与创造力。 ## Bulletin - [02:24](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=144s) Meta 团队在 8 个月内极速打造出新一代个人智能体 Muse，其核心技术分水岭在于为每位用户在云端分配独立运行的计算机实例，支持后台静默执行与全天候主动服务。 - [00:04](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=4s) Muse 彻底颠覆了单轮对话框范式，用户自发组合出“全自动家庭日报”、动态音乐巡演追踪等未被列入官方路线图的常态化用例。 - [04:14](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=254s) 系统设计呈现反直觉趋势：随着底层 Agent 能力增强，界面反而大幅简化为即时通讯气泡与常驻人格化实体，从“确定性像素摆放”转向“非确定性架构与食材准备”。 - [14:25](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=865s) AI 原型工具打破了设计、工程与产品经理的传统职责壁垒，使产研路线图从传统的 18 个月长周期收窄为聚焦近期的极致敏捷迭代。 ## Key takeaways - [02:24](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=144s) Muse 的产品本质是拥有独立计算机的个人智能体（Personal Agent），能够跳脱出被动单轮问答，在后台持续自主运转与推进任务。 - [00:04](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=4s) 强大的 Agent 催生了自下而上的用户创新，例如用户自主组合出包含日程后勤、家庭老照片与天气的“家庭日报”，绕过了传统产品立项流程。 - [04:14](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=254s) 面对全新技术类别，设计原则是“越强大的系统界面越应极简”，团队摒弃冗长排版，采用拟人化短消息气泡与可命名的实体形象降低认知门槛。 - [08:12](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=492s) 为全球数十亿用户在云端实时调配与维护隔离的独立计算实例，对底层模型推理、虚拟化算力调度与基础设施提出了史无前例的工程挑战。 - [10:17](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=617s) Agent 时代的用例具有“永不完结的累加特性”，用户可以像搭积木一样无限接入新数据源与外部 API，功能边界伴随需求持续自然演进。 - [13:15](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=795s) 在非确定性（Non-deterministic）系统下，设计师的角色从规定确定性像素与路径，转变为定义容器层级、控制灵活度边界与提供“食材”的架构师。 - [14:25](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=865s) AI 工具赋能全员独立端到端制作原型，打破了研发角色壁垒，但同时也让底层架构与系统协调性面临更深层次的挑战。 - [18:29](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1109s) 软件研发节奏进入了前所未有的加速档位，面对高速迭代的模型能力，传统的 18 个月固定路线图失效，团队必须保持高度敏捷（nimbleness）。 - [21:35](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1295s) Agent 带来代际颠覆的核心转折点是“主动性（proactivity）”，即便应用未被打开，系统也会在后台结合上下文独立思考并主动推送可执行建议。 - [24:15](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1455s) 在 Agent 时代成功探索新范式依赖于三大特质：敢为人先（Bold）、保持弹性（Flexible）与勇于承担风险（Risk-taking）。 - [24:15](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1455s) AI 的普及并没有削弱人类的价值，反而彻底释放了开发团队与终端用户的个体创造力，人的参与依然起决定性作用。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=4s) 突破性产品 Muse 的登场与意料之外的家庭报纸 在这场对谈接近尾声的环节，主持人 Michael Martinez 表达了他对本次对话的期待。作为 AI 行业大热产品 Muse 的核心参与者，来自 Meta 的 Alex Cornell 与 Michael 一同深入探讨了这款产品的诞生故事与本质特征。Michael 坦言自己从产品内部试用（dogfooding）阶段一直用到公开发布，日常生活中早已离不开它；在他眼中，Muse 就像是一个无所不知的朋友，宛如经典益智问答节目《谁想成为百万富翁》（Who Wants to Be a Millionaire）中那个随时可以场外连线的智慧锦囊。它无处不在、全天候在线，主动替用户化解生活中的各种琐事，甚至他的妻子现在买任何东西之前，都非要先听听 Muse 的意见不可。 除了这些日常陪伴，最令 Michael 震惊的是社区用户发掘出的各种神级用例。其中一个让他印象极为深刻的案例是：有一位用户让自己的 Muse 每天清晨为全家自动生成并排版一份专属的“家庭日报”。这份报纸详尽列出了当天的行程后勤、孩子们要去参加哪些体育项目，附上了去年今日的老照片回顾，甚至还汇编了早间头条新闻以及出门穿衣参考的天气预报；每天清晨全家人醒来时，这份报纸已经由家里的打印机静静打印完毕。Michael 特别指出，这个令人惊艳的功能从未列入团队的产品路线图（roadmap），也根本没有经过任何用户的需求提报，“它就是这么自然而然地被用户搭建出来了”。这也顺势引出了当天对谈的核心命题：对于台下坐满的工程和产品人员而言，这究竟意味着传统软件开发模式的终结，还是全新时代的序幕？ - Muse 的日常渗透感极强，甚至演变为用户家庭消费决策中必不可少的前置环节。 - 用户自建的“家庭日报”包含四大核心模块：孩子当日运动与日程后勤、家庭历史照片回顾、精选头条新闻、出门天气预报。 - 这一用例完全跳过了传统的产品立项、排期与功能开发流程，完全由用户通过智能体自主组合实现。 --- ### [02:24](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=144s) 重新定义智能体：告别单轮对话，走向拥有独立计算机的 Agent 为了给现场可能尚未使用过 Muse 或未密切追踪行业动态的观众解惑，Alex 简洁明了地给出了 Muse 的官方产品定义：“Muse 是一个能力极为强大的个人智能体（personal agent）。”Alex 进一步阐释，用户之所以能清晰感知到它与普通工具的区别，关键在于它能够真正替用户办实事，并且能够在后台持续运转——而支撑这一切的底层根基，在于 Muse 拥有独属于它自己的一台计算机。这种系统不仅具备前所未有的执行力，更打破了以往被动等待输入的僵局，能够主动（proactively）推进任务。 在对比早期的 AI 应用（例如普及度极高的 ChatGPT 等对话产品）时，Alex 一针见血地指出了两者的代际差异。在过去的一整套范式中，用户面对的本质上只是一个输入框，系统以严格的回合制（turn-by-turn）运作：用户提一个问题，模型给出一个回答，一旦对话会话（session）关闭，整个交互便彻底终结。然而，以 Muse 为代表的新一代智能体完全颠覆了这种工具属性；无论是晨间定制简报、充当私人高尔夫教练，还是自动生成文件，即使用户正坐在台下开会，后方的 Muse 依然在后台独立执行任务。谈及交互形态的演进，Alex 深刻地总结道：“即使与它交互的方式依然是通过自然语言，但当你真正与一个实体对话，而不是仅仅面对一个工具时，那种交流是截然不同的。” ```mermaid flowchart TD accTitle: 对话型工具与主动型智能体的交互范式对比 accDescr: 对比传统回合制AI工具与拥有专属计算环境的主动型个人智能体的工作流 subgraph Traditional["前一代 AI 应用：单轮工具"] A["用户在文本框输入问题"] --> B["模型逐字单轮推理"] B --> C["输出静态文本结果"] C --> D["会话结束 / 状态重置"] end subgraph Agentic["新一代个人智能体：Muse"] E["用户自然语言意图 / 上下文"] --> F["个人智能体拥有独立计算机"] F --> G["后台静默执行与主动工作"] G --> H["跨任务协同：日程/教练/文件生成"] H -.->|"主动提醒与长期记忆"| F end ``` --- ### [04:14](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=254s) 八个月极速研发：越强大的系统，界面反而越应极简 谈及产品的研发周期，Alex 透露团队从年初才正式着手推进，从最初立项到推向大众大约只花了八个月左右的时间。作为一个设计师，Alex 坦言这段经历充满了未知的挑战：在以往的产品设计中，设计师总能找到成熟的界面模式（design patterns）或行业先例作为参照物；但面对 Muse 时，他时刻感觉自己是在一片完全未开垦的处女地中拓荒。团队在过程中曾多次以为产品已经就绪可以发布，但随即又发现总有更深层的边界需要继续推进打磨。 在处理全新技术类别的用户体验时，Alex 并没有堆砌复杂前卫的控件，而是反其道而行之，大胆借用了人类最熟悉的人际交往隐喻作为框架。面对棘手的设计抉择时，团队最常用的思维基准就是：“如果屏幕另一端坐着一个能力极强的真人助理，你会怎么做？”基于这一原则，Muse 摒弃了当时行业普遍流行的长篇排版——到去年年底时，大量 AI 产品趋向于输出类似完整 Markdown 文章的长屏排版，充斥着大标题、插图、多段落的长篇大论，而这在人与人之间真实的即时沟通中是绝不可能出现的。Alex 提出了一个极具穿透力的设计洞见：“我从一开始就觉得，随着系统能力的不断增强，界面反而可以大幅简化。”正如面对世界上最聪明的人时你只需通过短消息就能办成所有事一样，Muse 采用了极其清爽的短信式气泡，并在屏幕顶部直观展示一个可以被赋予名字和自定义外观的实体（entity）。这种“实体而非工具”的视觉与交互心理暗示，正是 Muse 迅速打破认知壁垒、赢得广泛受众的关键所在。 - 研发节奏：从年初从零起步，历时约 8 个月完成探索并上线。 - 界面反思：摒弃全宽（full-blade）、多级标题排版的多页文章式 AI 回答，重塑为轻盈自然的即时通讯质感。 - 实体感打造：通过在屏幕顶部常驻可命名、可装扮的形象标识，唤醒用户潜意识中的人际交往模式。 --- ### [08:12](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=492s) 云端独立计算机与规模化基础设施的竞速 听到 Alex 谈及界面的实体感，Michael 深有感触地分享了自己的个性化设置：他的 Muse 名叫“M”，头像是一只鲨鱼，性格比自己本体更加直接、干脆利落，就像是自己的数字化“另我”（alter ego）。正是因为 M 始终端坐在界面顶部，每一次打开应用都让人感觉是在与一位真实的伙伴聊天，而不是在唤醒后台冷冰冰的脚本代码。 然而，这种流畅无缝的拟人化体验背后，承载着底层系统架构的重大颠覆。Michael 指出，Muse 最核心的技术分水岭在于“每位用户都拥有属于自己的独立计算机”。在当今公众对数据隐私和数据掌控权极其敏感的背景下，为用户分配隔离的计算环境是迈向保护个人隐私的巨大飞跃，但这也意味着难以想象的工程复杂度。要在云端为全球数以亿计甚至未来数十亿的用户实时调配、运行并维护独立的计算机实例，是一项史无前例的技术壮举。为此，Meta 的底层基础设施团队必须与上层产品设计团队保持并驾齐驱的极速狂飙，在算力调度、虚拟化与网络架构领域同步攻克大量前所未见的工程技术难关。 ### [09:31](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=571s) 超大规模扩展与未开垦的设计边界 Michael Martinez 谈到，尽管团队在快速进行区域和国家级的产品灰度扩展（rollouts），但外界并没有出现关于系统性能崩盘的负面报道，整体舆论焦点主要集中在用户发掘出的各种惊艳用例上。Alex Cornell 对此表示赞同，并指出这不仅仅是设计维度的突破，而是整个研发链路都在向前推移。要承载如此庞大的智能体（Agentic）工作流，团队必须在模型侧专门优化智能体任务的推理能力，在底层基础设施侧打造高并发与稳健的承载底座，整个项目呈现出一种全栈协同探索的状态。 谈及跨学科推动前沿的紧迫感与兴奋感，Alex Cornell 直言：“I think in all categories we were sort of needing to push the various boundaries. you know everything from the design but also the model itself being able to like be optimized for agentic tasks and then the infrastructure as well.” 面对这一完全没有参考范式的未知领域，整个跨职能团队共同承受挑战，也因此体验到极大的研发热情。 ### [10:17](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=617s) 动态进化用例：从个人教练到永不完结的音乐追踪系统 在谈到用户自发探索出的具体场景时，Alex Cornell 分享了自己如何将 Muse 当作私人专属教练：他将自己的高尔夫数据和国际象棋棋局持续同步给 Muse，系统会在每天清晨发送结构化的复盘分析与改进建议。此外，还有用户构建了一套极其灵活的个人演唱会追踪流程：Muse 不仅全面汇总该用户喜爱的所有活跃艺人名单，还会自动抓取这些艺人即将在用户周边地区举办的巡演，跟踪门票持有状况，形成一个动态的线下演出看板。 这类用例最令 Alex 震撼的是其“永不完结”的累加特性（additive nature）。传统软件的功能边界在上线时即被固化，而在 Muse 中，任何小工具都可以像积木一样无限延伸。比如音乐追踪器后续可以继续接入自动购票 API 或关联流媒体曲库；高尔夫分析助手除了处理下场正赛数据外，随时可以再接入练习场的击球雷达数据。Alex Cornell 总结道：“And then now it's like you're always additive in some way, which is I think really interesting.” 用户再也不用等待官方发布下一个更新版本，功能会伴随实际需求自然演变。 ```mermaid flowchart LR accTitle: Muse 动态用例演进架构 accDescr: 展示用户如何从基础数据输入开始，经过核心分析，再不断叠加外部能力与新数据源的无限延伸机制。 subgraph Input["数据接入"] A["高尔夫球局数据 / 棋谱"] B["关注艺人列表"] end subgraph Core["Muse Agent 核心处理"] C["早晨专属复盘报告"] D["本地巡演与门票追踪"] end subgraph Extension["无限叠加拓展"] E["练习场挥杆传感器"] F["一键购票与流媒体曲库"] end A --> C B --> D C -.->|"自由扩展"| E D -.->|"自由扩展"| F ``` ### [13:15](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=795s) 非确定性系统下的设计范式：从“摆放组件”到“准备食材” 面对传统基于指标分析和固定生命周期的产品开发模式，Michael Martinez 提问：这种“永不结束、持续生长”的机制，是否会限制开发团队在规划下一代功能版本时的想象力？Alex Cornell 解释说，当前软件设计中最剧烈的心智转变在于应对系统的**非确定性（Non-deterministic nature）**。过去，界面设计师清楚地知道每一个组件和方框会固定渲染在屏幕的何处；而现在，设计师的角色更像是为用户备好丰富食材的厨师，由用户根据自身口味去烹调最终的盛宴。 在这种形态下，设计师的核心任务变成了搭建稳健的脚手架、定义容器层级、控制灵活度边界。Alex Cornell 指出，这种架构让所谓“个人超级智能（Personal Superintelligence）”真正照进了现实：“I think quite a lot about like how the whole app is designed, where the containers are, what goes where, and where there's flexibility. But I think having quite a lot of personalization and malleability within that is part of what makes this like really delivering on like we hear personal super intelligence all the time.” - 传统开发：确定性像素排布、固定交互路径、封闭功能模块。 - 智能体开发：关注上游架构、控制权与边界定义、极高可塑性（malleability）。 ### [14:25](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=865s) 角色界限消融与上游系统架构的挑战 Alex Cornell 坦言自己的日常工作形态相比一年前已经发生了翻天覆地的重塑。以往不会写代码的他，现在每天大部分时间都在编写代码，甚至维护着一个全功能的 iOS 原型沙盒，与 Muse 应用并行验证各种想法。AI 工具极大地放大了每一个个体的技术执行力，从而使设计师与工程师之间的传统壁垒逐渐消失。 Michael Martinez 补充分析道，以往的组织架构中，设计师专注于体验旅程，工程团队死磕高可用与吞吐量，PM 则苦于在有限的工程带宽中做艰难的取舍（bet on a few things）。而现在的协作模式彻底松绑：早期原型验证可以由任何人独立完成，后续再由专项专家跟进打磨。Alex 强调，虽然每个人现在都能借助 AI 自行搭建专属工具（比如无需官方设计团队介入就能做出演出追踪器），但这也导致底层设计问题变得更深、更靠上游，整体系统控制力与协调性的要求成倍增加。 当被问及这段经历是否“毁掉（ruined）”了他参与传统软件研发的兴致时，Alex Cornell 断然否认，并直言自己没有受到任何消极影响，反而感到前所未有的精力充沛与兴奋，准备将这套系统性思考和敏捷原型能力融入到后续更加广阔的产品探索中去。 ### [18:29](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1109s) 软件开发进入“全新档位”与路线图规划周期的缩短 Alex Cornell 回顾了自己的职业生涯，感慨自己从事软件开发已经很长时间。他原本以为随着年龄增长和从业时间变长，人会自然而然地感到疲惫，最终产生一种“什么都见过了”的心态。但在打造该 Agent 产品的过程中，他多次与身边的同事对视并感叹彼此创造出的东西太不可思议。这种体验是他过去未曾有过的，甚至在此前都不知道软件开发还能达到这样的境界——将如此令人兴奋、新颖有趣的想法真正赋予生命。Alex 提到，他直到现在才意识到软件开发还存在这样的“加速档位”（gear），这一发现极其鼓舞人心，让他感受到了整个职业生涯中前所未有的专注度与兴趣。 Michael Martinez 随后表达了强烈的共鸣，称这种状态带有一种“神经紧绷的兴奋感”（nervous energy）。作为支持多个工程团队的管理者，Michael 以往大部分时间都是坐在会议室里审阅评审文档、做决策，很少有机会亲自上手；而现在借助强大的 AI 工具，他能够随时亲自动手快速制作原型，直接向团队展示自己的设想。但与此同时，这种兴奋感也伴随着不安，因为行业变化太快，很难“看清下一个转角”。过去行业通常会制定 18 个月的路线图，并每 6 个月复盘校准一次；而如今，随着底层模型能力不断涌现、应用场景更加多变，规划的时间跨度已被大幅压缩。面对 Michael 提出的“未来走向如何”的疑问，Alex 表示团队在短期内有着清晰的规划，把一些首发时未上线但确定有价值的功能列入了近期路线图；但在 6 个月以上的远期规划上，没有人能完全看清，唯一的应对策略是保持极致的敏捷性（nimbleness），随时根据底层模型的能力变化与产品诉求调整姿态。 Alex Cornell 在阐述这一心态转变时说道：“I think I feel more engaged and interested than I have in in my whole career probably you know and I think that that is speaks to maybe just like the capabilities of these systems but also where we are as an industry and I think I'm really look forward to you know seeing what's next for.” - 传统产研规划周期通常长达 18 个月并每半年对齐一次，而当前大模型驱动的敏捷开发使这一时间跨度大大收窄。 - 面对模型能力的高速迭代，产品团队必须具备快速重塑与变通的姿态（nimbleness）。 ```mermaid flowchart TD accTitle: 传统研发与 Agent 时代研发模式对比 accDescr: 对比传统模式长周期、重评审与 Agent 时代短周期、强敏捷原型的流程差异 subgraph Past["传统软件研发模式"] A["18 个月长期战略路线图"] --> B["每 6 个月校准与评审"] B --> C["管理层审阅评审文档，脱离实际动手"] end subgraph Present["Agent 时代研发模式"] D["短期核心功能路线图"] --> E["高频跟踪底层大模型迭代"] E --> F["管理层与个人直接原型验证"] F --> G["保持敏捷姿态（Nimbleness）灵活转向"] end ``` --- ### [21:35](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1295s) 原型突破时刻：从被动工具到主动式 Agent（以 Jolly 为例） Michael Martinez 询问 Alex，在项目早期的哪些具体瞬间让他笃定这项工作将演变为某种划时代的大事件，以及如何识别这类带来重大突破的机会。Alex 坦言，人很容易对新技术习以为常，但如果回看今年年初，当时的交互体验发生了根本性的颠覆。其中最核心的转折点在于“主动性”（proactivity）——让用户真切感受到即便自己没有在使用应用，Agent 也一直在后台思考该如何为用户提供帮助。 Alex 分享了一个具体的切身体验：他在测试中为自己的 Agent 取名为 Jolly。当时 Jolly 在没有被即时唤醒的情况下，自主基于他们早先的闲聊内容去检索了资料，随后主动推送了一条消息。尽管 Alex 自嘲虽然喜欢高尔夫却根本没时间去打，但当屏幕上跳出通知写着“嘿，我刚才一直在思考你的高尔夫目标……”并附带具体的行动建议时，那种震撼感是前所未有的。随着使用时间的增长，上下文更加丰富，这种主动思考与建议变得越来越精准实用。Michael 赞同道，这彻底打破了传统“点开应用”的孤立工具模式，Agent 变得像是一位随时在身旁陪伴、辅导和鼓励你的真实朋友。 Alex Cornell 回忆那个震撼的瞬间时说道：“And I think the first time that I got a message that was from my muse about that where it had gone and thought about it and then come back to me with something actionable, that's like a wild feeling the first time, you know, and then it's only gotten more contextrich now.” - 主动性（proactivity）是 Agent 与传统被动式对话界面的核心分水岭。 - 上下文越丰富，后台自主思考生成的行动建议对用户而言就越有价值。 - 案例中名为 Jolly 的 Agent 能够追踪用户的高尔夫目标并在后台独立推理后推送通知。 --- ### [24:15](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=1455s) 总结回顾：胆识、弹性、冒险与人的不可替代性 在对话的尾声，Michael Martinez 对整场对谈的核心心智模型进行了系统提炼，总结出推动此类前沿创新不可或缺的三项特质： 1. **敢为人先（Being bold）**：当下拥有前所未有强大的 AI 工具，正是大胆设想和突破常规的时刻。 2. **保持弹性（Being flexible）**：路线图、业务策略乃至团队中每个人的角色分工都必须具备高度弹性，根据项目需要灵活调用工具和转变职责。 3. **承担风险（Taking risk）**：进入一个全新的市场、探索大众用户尚未完全熟悉的新型交互范式，必须有直面未知的勇气。 Alex 对这一总结表示完全赞同，并进一步指出，开发产品与身处当前行业的心态是相通的：虽然基础架构和结构化设计必不可少，但唯有保持极致的弹性，才能适应这场“全方位的档位切换”（full gear shift）。今年是整个行业与技术全面升档的关键年份，不仅产品能力的边界被重构，从业人员自身潜能的释放也进入了新维度。Michael 最后特别强调，尽管外界常有声音认为在 AI 时代人类将变得无关紧要，但这次对谈充分证明了“在背后创造的人”（people working on it）依然起着决定性作用；AI 不仅重塑了软件开发，更彻底释放了开发人员和终端用户的创造力。 Michael Martinez 针对人才价值坚定地指出：“Yeah, I completely agree and I really like that you said people working on it because I think you know there there's a lot of other perspectives that suggest that people aren't as important but I think the conversation here proves that the the people stay still play a very important role in in in where things go both in terms of the development of product but then also as a user you're just un your creativity is unleashed and so I'm really excited about that and I really appreciate the time today it's like really fun product and it's also really fun to talk about.” - 打造前沿 Agent 产品的三大成功法则：Bold（大胆创新）、Flexible（角色与战略保持弹性）、Risk-taking（勇于承担新市场风险）。 - 无论是在产品形态层面还是人才本身的工作方式上，当前技术革命的核心价值都在于释放人类的创造力，而非淘汰人的角色。 ## Notable quotes > [04:14](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=254s) 随着系统能力的不断增强，界面反而可以大幅简化。 — Alex Cornell > [02:24](https://www.youtube.com/watch?v=Fn6F-A3wIZc&t=144s) 即使与它交互的方式依然是通过自然语言，但当你真正与一个实体对话，而不是仅仅面对一个工具时，那种交流是截然不同的。 — Alex Cornell ## Mentioned Michael Martinez, Alex Cornell, Meta, Muse, ChatGPT, Scale, Jolly, iOS, Who Wants to Be a Millionaire [Source](https://www.youtube.com/watch?v=Fn6F-A3wIZc)

[4:14] Alex Cornell: 随着系统能力的不断增强，界面反而可以大幅简化。 [2:24] Alex Cornell: 即使与它交互的方式依然是通过自然语言，但当你真正与一个实体对话，而不是仅仅面对一个工具时，那种交流是截然不同的。

[5:19] Alex Cornell: I think the thing that I felt towards the the beginning was that like as the power of the system was increasing, the interface could simplify.

[0:04] We're toward the end of the day here. [0:05] It's been quite a day. Uh, a lot of [0:07] great conversation and talk, practical [0:09] uses of AI. I'm I'm excited for our [0:11] discussion though uh because you, Alex, [0:14] have been part of uh breakout success uh [0:17] in our industry uh with the Muse [0:20] product. And so I'm excited to kind of [0:22] go deep into um you know what Muse is, [0:26] but also like how it came to be. Uh from [0:28] my perspective, uh I've been using Muse [0:32] for a while now, both started with dog [0:33] fooding and then of course with the the [0:35] public release and you know it just [0:37] really feels like a friend that knows it [0:40] all, right? Uh if you think back to like [0:43] who wants to be a millionaire? It'd be [0:44] the friend that I phone. Uh because it's [0:47] it's always there. It's like always on. [0:49] It's proactively solving all sorts of [0:51] different things uh in my life, which is [0:53] great. Um, you know, my wife doesn't let [0:56] me buy anything now without consulting [0:58] Muse first. Uh, so I think that that's a [1:01] big big congratulations to you. Uh, but [1:04] the other really interesting thing about [1:06] Muse is all of the use cases that people [1:09] are bringing up. Uh, I think one of the [1:11] more interesting ones that I heard about [1:14] uh, was someone has actually had their [1:17] Muse generate a daily newspaper for [1:19] their family. It has the logistics, what [1:21] what sports the kids need to be at, some [1:24] pictures, uh, you know, looking back [1:27] from last year. Uh, it has kind of some [1:30] top news headlines that are going on, [1:31] the weather to help get the kids out the [1:33] door. It's on the printer every morning [1:36] when they wake up, right? And I think [1:38] what's really interesting about this [1:41] particular use case is it didn't hit [1:43] your road map. It didn't require uh [1:47] someone to submit a feature request. it [1:49] it's just built. And so I think that [1:51] that's really the driver of our [1:54] conversation today, which is, you know, [1:56] how do we see, you know, for a room full [1:59] of product folks on the engineering and [2:01] and and other sides of the business, [2:03] like is this really, you know, kind of [2:05] the the end of traditional uh [2:08] development or is it just the beginning? [2:10] Uh but maybe for those in the room that [2:12] might not have used Muse or or or maybe [2:15] aren't participating in in uh kind of [2:17] all of the headlines that are out there [2:18] right now, could you tell us what Muse [2:20] is? Because I think my my explanation [2:22] might not be the product definition. [2:24] >> Sure. Yeah. No, absolutely. Um well, I [2:27] think Simply Muse is a very capable [2:30] personal agent and uh I think the way I [2:33] usually explain that is it can do [2:35] actually do things for you. It can work [2:36] in the background and it does all that [2:38] because it has its own computer. So in [2:40] that sense it's quite unique. Um it can [2:43] also work proactively which again is a [2:45] pretty new thing for systems like this. [2:47] Um I guess now I can maybe also add that [2:50] it will be the person that you call if [2:51] you're ever on who wants to be a [2:52] millionaire. Um but very capable [2:54] personal agents probably the best way to [2:56] understand it. [2:57] >> Okay. Yeah. I mean it really is a [2:59] different category though, right? I I [3:01] think you know chat GPT you know became [3:04] that kind of default and that's really [3:05] the chat interface and that the agent is [3:08] kind of this new category that we're [3:09] we're moving into. Um but what's what [3:12] also is interesting like were there [3:15] other things before it that were similar [3:16] or is this just something that we we [3:19] came up with? [3:20] >> Yeah, I think um it's you know you [3:22] mentioned the sort of the previous era [3:23] of of AI AI apps. I think the way that [3:26] they typically worked up to really this [3:29] year was basically you have a text box [3:31] and in a very turnbyturn based way you [3:33] ask a question and you get an answer and [3:36] when you're done with your that session [3:38] it's kind of over that's it. Um when I [3:41] think about really what makes agents [3:42] like Muse unique is that they really can [3:44] do things for you. You know and many [3:47] many things and like you said it can be [3:49] anything from a newsletter that you get [3:51] every morning. It can be a golf coach. [3:53] It can make files for you. lots of [3:55] different things working in the [3:56] background when you're like right now, [3:58] you know, like I have mine doing things [3:59] even while I'm sitting here. And that's [4:01] not the way that it used to work. Even [4:03] though I think like the way that you [4:05] interface it through natural language is [4:07] still the case, it's a very different [4:09] exchange when you're actually talking to [4:12] something as opposed to just a tool. [4:14] >> Interesting. I mean a product like Muse [4:17] you say really just since the beginning [4:19] of the year is is that really the time [4:21] horizon is like you know idea to [4:24] production is within this year. [4:26] >> Yeah. Uh it was a very quick process uh [4:29] I guess relatively when you think about [4:31] it. Um we started working towards the [4:33] beginning of the year and I think at the [4:35] very beginning it was uh it was [4:38] definitely very clear that it was [4:40] different and that it was a distinct [4:42] model and it was really exciting. Um, [4:45] but it was kind of uncharted territory, [4:47] I think, which as a designer was [4:49] somewhat um unusual for me. Like usually [4:52] there's some patterns or something that [4:54] you can kind of look to or reference [4:55] when you're working on something. And [4:57] with this, I constantly felt like we [4:59] were sort of exploring a new category [5:01] completely. Um, which is really [5:04] fulfilling and interesting to work on as [5:05] a designer for sure. Um, but yeah, it [5:08] was really quick. I think we thought at [5:10] various times that we were ready to go, [5:12] but there was always something more to [5:13] keep pushing on. And so, you know, I [5:15] think it took the better part of eight [5:17] months or so to get it out there. [5:19] >> That's amazing. It's it's amazing from [5:22] many different angles, like one new new [5:24] product category that you're working [5:26] towards. So, there isn't really a [5:28] reference to copy. And so, I I I would [5:30] like to tease apart maybe how you [5:32] approach design because the app overall [5:34] is is very approachable. I mean when you [5:36] bring in a a new category [5:40] there's uh training or and learning that [5:42] you have to you know educate the users [5:44] on but in this case it kind of just like [5:46] blew up very quickly and so how did you [5:49] approach the design for a completely new [5:51] category of technology? Yeah, you know, [5:53] I think like the um the category itself [5:56] was new and the experience definitely [5:58] was new, but I think we were able to [6:00] really lean on familiar metaphors both [6:04] in the framing but also in the design [6:06] language itself. And by that I mean like [6:08] we were treating it really like you were [6:10] talking to a person that kept being like [6:12] a very familiar um framing that we could [6:16] lean back on. Whenever there was a [6:17] question like how do we handle this? [6:18] It's like, well, how would you do this [6:19] if it was just you on the other end, [6:21] like a very capable assistant? Um, and [6:24] you see that throughout the design. You [6:26] know, I think something simple, just [6:28] like the way the messages are rendered, [6:30] they look like a text message app. You [6:32] know, they're not like this kind of uh [6:34] markdown widescreen full-blade kind of [6:37] article style because I think towards uh [6:39] certainly towards the end of last year, [6:41] we were starting to see a coalescence [6:43] around this very article-like format for [6:45] all answers. you know, you'd get a big [6:47] header, text, pictures, text. It'd be [6:50] multi-pages long, you know, and if you [6:52] think about when you're talking to [6:53] somebody, that's never how you exchange [6:56] information, no matter how complex the [6:59] exchange. And I think the thing that I [7:01] felt towards the the beginning was that [7:03] like as the power of the system was [7:05] increasing, the interface could [7:07] simplify. And you know, in the same way [7:09] that if you had like the smartest person [7:11] in the world on the other end of your [7:12] exchange, you could handle just about [7:14] everything over chat. Um, and so that [7:17] was a really comforting and familiar [7:19] thing to continuously look to just [7:21] whenever we sort of had a question. And [7:23] I think like I said, you see that with [7:25] the way the messages are rendered, but [7:26] also something very simple like the [7:28] entity itself being at the top of the [7:30] screen, something that you name and that [7:32] in some cases you can customize to look [7:35] however you want. That kind of [7:37] familiarity, I think, is something that, [7:39] you know, consciously or not, people [7:40] will feel very familiar with the way [7:42] that you would then interface with [7:44] something like that. And I see that [7:45] whenever I see people using it, the [7:47] types of things that they say, the way [7:49] they respond, you can tell they're [7:51] treating it a bit more like an entity [7:53] than a tool. Um, so I think even though [7:56] it was very much a new design paradigm [8:00] for us and a new kind of era, I think [8:03] that taking from familiar patterns that [8:06] we have in other areas was really [8:08] effective. And I think part of maybe why [8:10] it's had some universal appeal. [8:12] >> I I didn't spend much time thinking [8:14] about that, but I So my muse is M. He's [8:17] a shark. Uh, he's kind of like [8:19] >> an uh alter ego of myself. is a little [8:22] bit more direct, maybe a little bit more [8:24] punchy. Um, but you're right, like it's [8:26] sitting right at the top of a makes it [8:28] feel like it's a chat message and and so [8:30] it does like embody, you know, the [8:32] conversation a lot better where it is a [8:34] friend uh and and not just, you know, [8:36] this this agent working in the [8:38] background. What what's interesting [8:41] about this product is not only was it [8:43] new category and took some different [8:44] approaches and it really is a new [8:46] experience that you know behaviorally uh [8:49] you know it kind of changes how people [8:50] engage because an agent can work when [8:53] you're not prompting it or when you're [8:55] not there but then technically you [8:57] mentioned at the beginning you have your [8:59] own computer and that I think is also [9:01] very different right in this current um [9:04] you know the world that we're in there's [9:05] a lot of questions about privacy and [9:07] control of your data Uh and I think that [9:09] that is like a giant leap forward for [9:11] the company. Um also like a very complex [9:14] one because scaling a computer for [9:16] everyone in the cloud for billions [9:18] potentially billions of users around the [9:20] world would would be uh uh quite a feat. [9:22] Uh and so it is impressive that our [9:24] infrastructure teams also had to kind of [9:26] race along with the product development [9:28] to come up with a lot of new techniques [9:31] uh to hypers scale this product and have [9:33] been doing quite well. I know a lot of [9:35] regional rollouts, regions being, you [9:37] know, country rollouts still continuing [9:39] on. Um, but no headlines about the the [9:42] performance really. It's it's really [9:44] focused on on all of the the great use [9:46] cases that are people are finding. [9:48] >> Um, yeah. [9:49] >> Yeah. No, I mean I think that that [9:51] reminds me that, you know, we were [9:53] talking about the design sort of being [9:54] an unprecedented area, but I think in [9:56] all categories we were sort of needing [9:58] to push the various boundaries. you know [10:00] everything from the design but also the [10:02] model itself being able to like be [10:04] optimized for agentic tasks and then the [10:06] infrastructure as well. So it really was [10:08] I think a team effort in that sense [10:09] across the board is because it was kind [10:11] of uncharted territory for everybody um [10:14] but also really exciting as a result. [10:17] >> Yeah. Interesting. I I I think another [10:19] thing I I would be curious in we we [10:21] hinted at a couple of different use [10:23] cases. I I heard the one example I gave [10:25] at the beginning. You know there's the [10:27] shopping use case which seems to be like [10:28] out there a lot. you kind of hinted at [10:30] golf a little bit, but are there other [10:32] use cases that were kind of unforeseen [10:34] or or predicted that that you've been [10:36] impressed with or the the broader team [10:38] has been talking about? [10:39] >> Yeah, I mean, I think that's probably [10:41] one of the most fun parts about working [10:43] on a product like this is that you're [10:45] always kind of surprised how people use [10:47] it and what they get value out of. And I [10:49] think, yeah, I I use it a lot for [10:51] hobbies, certainly for golf and for [10:52] chess. I kind of like having both of [10:54] those things, treating my Muse almost [10:56] like a trainer because it can keep an [10:58] eye on whether it's my chess games or my [11:00] golf data that's coming into it and [11:03] analyze it and send me some nice and [11:05] instructive feedback in the mornings. I [11:07] like I like it for that. I think um I [11:10] heard the other day somebody told me [11:11] about this use case I hadn't thought of [11:13] but now really want to try which is [11:15] basically tracking all of the music that [11:17] they have and all the artists that are [11:18] active today and then they use that to [11:22] it searches for anytime that they're [11:23] going to come and do a show nearby and [11:25] then they can keep track of whether or [11:27] not they have tickets and they can use [11:28] that to basically manage their entire uh [11:31] concert tracking system which is which [11:33] is great and it's like something I [11:34] hadn't thought of but you know I would [11:36] definitely do and I think the coolest [11:38] part is anytime somebody is telling me [11:40] about a use case or something that [11:42] they've had Muse build for them or for [11:44] me, there's always this sort of like and [11:46] then I'm going to do this or and then [11:48] I'm going to add this next thing because [11:50] anything that you make with Muse, it's [11:52] never really done. You can always add [11:53] one more thing. So like for the music [11:55] one, the way they had built it was it [11:56] was keeping track of all the shows for [11:58] the artists that they like, but it [11:59] wasn't hooked into like buying the [12:01] tickets. It wasn't hooked into like [12:03] their music catalog, but you could go do [12:05] that if you want to. You know, my golf [12:07] one currently will take my my golf [12:09] rounds that I play and send me an [12:11] analysis the next day, but I could also [12:13] hook it into like my actual practice [12:15] sessions. You know, when I go practice, [12:16] I get a lot of data, too. I could do [12:18] that. And it's like there's always one [12:19] more thing that you can add on. And I [12:22] think that's the coolest part because it [12:23] usually it would be like you're kind of [12:25] finished. It's like this is the this is [12:27] the thing. This is what it does. And [12:28] then now it's like you're always [12:30] additive in some way, which is I think [12:32] really interesting. It's it's like [12:34] really hard to process that because I I [12:36] think now back to you know I spend a lot [12:38] of time working with uh FB and and IG [12:41] product teams and uh you know there [12:44] there's a a process that we follow right [12:46] we analyze data use cases and and that [12:49] helps figure out like the product [12:50] development life cycle but what you're [12:52] talking about is like a product that [12:54] never ends uh right features that just [12:57] keep building or or evolving. And so [13:00] what's it like in the in the room of the [13:03] Muse team when you think about like what [13:05] the next release is? Um because I think [13:08] to some degree that could constrain, you [13:10] know, the creativity that that users are [13:12] bringing to the table. How do you how do [13:14] you think about that? [13:15] >> Yeah, I think um I think probably the [13:18] thing that has taken the most getting [13:19] used to in this new phase of design and [13:22] development is the non-deterministic [13:24] nature of the system. You know, so it [13:26] used to be that when you would design [13:28] something, you knew for sure exactly [13:30] what box was going to go where and it [13:32] would sort of always be there. You know, [13:33] that was the nature of software [13:35] development. And I think now I feel a [13:37] little bit more sort of like a uh a chef [13:41] preparing lots of ingredients and other [13:43] people can make whatever meal they want [13:44] to make, you know, and where it used to [13:46] be that like everything was exactly just [13:48] so. And I think the the fun thing about [13:50] that is there are times where we need to [13:54] spend a lot of focus around like the [13:56] structure and the architecture of the [13:58] system. So I think quite a lot about [14:00] like how the whole app is designed, [14:02] where the containers are, what goes [14:04] where, and where there's flexibility. [14:07] But I think having quite a lot of [14:09] personalization and malleability within [14:11] that is part of what makes this like [14:13] really delivering on like we hear [14:15] personal super intelligence all the [14:16] time. And I think like the way that this [14:19] unfolds really is able to deliver on [14:21] that in a way that I think never before [14:23] was really possible. [14:25] >> Yeah. What so do you think it's the end [14:28] of that traditional development cycle or [14:31] and and what you're describing here is [14:33] is what we need to all start moving [14:35] toward or [14:36] >> I do think I mean it's it's certainly [14:38] the case that it's a new phase [14:40] definitely and I feel that every day [14:42] like when I think about what I used to [14:44] spend my time doing as a designer [14:47] what I was doing a year ago is very [14:49] different than what I'm doing now you [14:50] know and that I think like that will [14:52] probably continue to be the case but [14:54] what I'm finding and I I hear this a lot [14:56] with other colleagues as well is that [14:58] like more and more there's more for each [15:01] of us to do within our area just because [15:03] now we can do so many more things. You [15:05] know, I didn't used to be able to code [15:07] for example. I spend most of my day [15:08] coding now, you know, and I have like a [15:10] a prototype that sits alongside the Muse [15:12] app that I can maintain and it's like a [15:14] full sandbox iOS prototype. I never knew [15:18] how to do iOS development before and now [15:20] I can. And so I think that like in my [15:22] experience so far, it's just been sort [15:24] of more of an explosion of what we're [15:26] each capable of and what we can do. I do [15:29] think though on software development and [15:31] design, it will probably become a bit [15:34] more of that scaffolding and [15:35] architecture construction that you focus [15:37] more on like I was talking about just [15:39] because when you can really build as a [15:42] user a use case for yourself, you don't, [15:46] you know, you don't need me to build the [15:47] concert tracking software for you. you [15:49] can do that and make it very specific to [15:51] your own use case and I think that's [15:52] actually really good but the controls [15:56] and like how it all works cohesively all [15:58] of that becomes that much more important [16:00] and I think actually is in some ways [16:03] that much harder just because of how far [16:06] much farther upstream it is. So I have [16:08] found that the problems have gotten more [16:10] complex and more actually as as we've [16:13] gone which as a designer I think is uh [16:17] yeah maybe just very interesting and [16:18] definitely fulfilling too. Yeah, that [16:21] the roles kind of blending into each [16:23] other is a really interesting, you know, [16:25] development that we have here because [16:27] everybody has their core strengths. You [16:28] know, you you have a long career of a [16:30] lot of different product experiences and [16:32] how users interact and engage and [16:34] navigate. On the engineering side, you [16:36] know, it's around scaling and [16:38] reliability and the and the quality and [16:40] performance elements. And then in the [16:42] the PM side, it's like really where are [16:44] we going? Because we have to choose, you [16:46] know, we have to choose a few things to [16:48] bet on uh because we only have so much [16:50] of the other uh you know functions time [16:53] to deliver. But but here we're just kind [16:55] of unconstrained in where we go because [16:57] everyone can kind of fill in in [16:59] different areas whether it be early [17:01] stages of prototyping and then have an [17:03] expert come in and kind of modify uh or [17:06] maybe eventually the agents kind of [17:08] running along and and then we're being [17:09] selective of of where we go. [17:12] >> Um so it is really a an interesting time [17:15] uh that we're going through and and and [17:17] what we're able to to do. Yeah, I find [17:20] that you know people have natural uh [17:24] expertise in various areas like for me [17:27] my I think best place in a process is [17:31] generally towards the beginning that's [17:32] the system thinking that I tend to enjoy [17:34] and engage with the best and I think [17:36] that like even as the roles kind of [17:38] become one people's focus and the things [17:41] that they enjoy doing are sort of what [17:43] then delineate what people work on you [17:45] know and so I have always found myself [17:47] towards the beginning of a project [17:49] working on the architecture and you know [17:51] then there are other people who really [17:52] like feature development growth that [17:54] kind of thing and whether you're a [17:56] designer or a PM or an engineer matters [17:58] less now I think [18:00] >> yeah uh has your experience on muse kind [18:03] of ruined you to work on any other [18:04] product I mean the this the the quick [18:06] scale I think is is is what's really [18:09] exciting about the Muse product but then [18:10] also everything we've talked about here [18:12] so is there are you just ruined or is [18:14] there one one exciting learning now that [18:16] you want to go take on to the next [18:18] adventure venture uh or next product [18:20] area you take on. [18:21] >> Yeah. No, I don't think certainly not [18:22] ruined. I feel [18:23] >> okay. Good. Think good. [18:25] >> I think like I think if anything like [18:27] more energized than than I have been, [18:29] which is surprising because I have been [18:30] working in software for a really long [18:33] time. And I think like maybe naturally [18:35] thought that as I got older and I worked [18:37] longer, I would get more tired and [18:39] eventually sort of feel like, you know, [18:41] I've seen it all. And I think like there [18:43] were so many times in this project that [18:46] I can remember back to where I would [18:48] look at somebody else on the project and [18:50] say like this is a crazy thing that we [18:52] have just concocted or built, you know, [18:55] in a way that like I've never really had [18:58] that experience before. And I didn't [19:00] really even know was possible in in [19:02] software development, you know, like [19:03] really like bringing to life something [19:05] that you found so exciting and novel and [19:08] interesting. And I think like I didn't [19:10] really realize that that gear was [19:13] possible and I think have found that the [19:16] fact that it's there just to be really [19:18] uh I really motivating you know and so I [19:21] think I feel more engaged and interested [19:24] than I have in in my whole career [19:26] probably you know and I think that that [19:28] is uh speaks to maybe just like the [19:31] capabilities of these systems but also [19:33] where we are as an industry and I think [19:35] I'm really look forward to you know [19:37] seeing what's next for [19:38] Yeah, I I also think it's it's almost [19:40] like a nervous energy like I'm I'm more [19:42] excited than I've been in a very long [19:44] time because I I feel kind of [19:46] unconstrained, right? Before it's, you [19:48] know, for me, I I support many different [19:50] engineering teams and so you sit through [19:52] many review documents kind of making one [19:54] one decision or another, but not really [19:56] hands-on anymore. Whereas now, I have [19:59] the ability to be very hands-on and I [20:01] can I can almost prototype things myself [20:02] and be like, "Hey, this is what I'm [20:04] thinking, right?" Uh and so it does feel [20:06] like a very exciting time. Nervous [20:08] because it's much harder to see around [20:10] the corner. So if you think about, you [20:12] know, the road maps of of the past, you [20:15] know, it'd be very common for us to have [20:16] a 18-month road map and then we true it [20:18] up every 6 months just based on kind of [20:20] how things are going. It seems like that [20:23] time horizon has shrunk uh for how far [20:25] you can go because the tools are just [20:27] emerging and the the use cases are much [20:29] more funible. And so how how do you [20:32] think this all ends up? What does the [20:34] future look like here? [20:35] >> Well, I think you know fortunately we we [20:39] had a lot of things that we knew we [20:41] wanted to do that we have serving as a [20:43] roadmap that we're excited about that [20:45] are areas that I think were concretely [20:47] valuable and we knew we should have but [20:50] didn't weren't part of the initial [20:51] release. So I do think that like the [20:52] near term we are excited about what [20:54] we're working on and know what's coming. [20:57] But I think like one thing that we've [20:58] learned as you've noted is that this [21:00] industry moves so fast everything's [21:01] always changing that like I don't think [21:04] you know concretely about six months in [21:07] advance just because I know that the [21:09] landscape will change and capabilities [21:10] will change as well. I think the one [21:12] thing that we can that we always do is [21:14] just like really plan for that kind of [21:16] like nimleness you know being able to [21:18] really adjust quickly depending on what [21:20] comes and depending what the models [21:22] provide and what we want it to do. And [21:25] so I think that uh having that posture [21:27] has benefited us really well so far and [21:30] will continue to. Um and I think that's [21:33] probably the way that we'll approach it. [21:35] >> Okay. Um maybe a fun question for you as [21:39] we we kind of wrap this this [21:40] conversation up because I I would love [21:42] to just sit here and talk about all the [21:43] use cases that you've seen because I [21:45] think that's what's what's really [21:46] interesting right now. But one of the [21:48] the more fun questions that I have uh is [21:51] what were the experiences that you kind [21:53] of note at the early side of this [21:54] project that you knew that this was [21:56] going to lead to something big? Are [21:57] there ways that you can identify either [22:00] new problems that you go to take on or [22:02] or uh a different approach to how you [22:04] work so that you can kind of identify uh [22:08] things to to to grab on to to to have uh [22:12] that that will lead to you being part of [22:13] something very big and successful. [22:15] >> Yeah. Yeah. Yeah, I mean there were [22:17] there were a few times numerous times I [22:19] think through the whole process where it [22:21] felt like this could be big and could be [22:24] really like a a big hit. And I think the [22:26] first time and it's hard to remember [22:28] back like it we take it for granted [22:31] immediately you know like we're used to [22:32] it already but like in you know towards [22:35] the beginning of the year this was like [22:37] exceedingly different in the way that [22:39] you would interact. it felt so [22:40] different. And I think one of the main [22:42] ways was proactivity and the way that [22:46] what it feels like to have an agent [22:48] that's basically thinking about how to [22:50] help you even when you're not using it, [22:52] you know, and the first time that you [22:54] get a message that is the result of that [22:56] process where it's thought about [22:58] something that you were talking about [23:00] before, done a little bit of research or [23:02] something, usually in my case it is [23:03] golf, although I always say like [23:05] ironically like I have no time to play. [23:06] So it's usually just say the next time [23:08] you get to play like you should try [23:09] this. But I think the first time that I [23:11] got a message that was from my muse [23:14] about that where it had gone and thought [23:16] about it and then come back to me with [23:17] something actionable, that's like a wild [23:20] feeling the first time, you know, and [23:21] then it's only gotten more contextrich [23:24] now. Like as I've used it more and it [23:25] knows much more about me, those moments [23:27] are now just that much more useful [23:28] because it knows more about me. But I [23:31] think I I can still remember the first [23:32] time that I ever got one of those and [23:34] thinking that, okay, this this is very [23:36] different when you see a notification on [23:38] your screen from Jolly and it said it [23:42] says like, hey, I've been thinking about [23:43] your golf goal and it goes on. It's like [23:45] that's a that's a really uh [23:47] fundamentally different feeling than [23:49] what had come before. [23:50] >> Yeah. I mean, it's when the product is [23:52] just kind of so integrated. It's not [23:54] just an app that pops up. It's it it is [23:57] like that friend that's kind of there or [23:59] coaching you or being supportive. [24:01] >> Uh and and I think that that is what's [24:03] really special about this this phase of [24:05] of uh of AI and and also this product. [24:09] So so congratulations on that like very [24:11] well done to you and the team. Uh it's [24:13] very impressive. Um maybe just to [24:15] summarize what I took away from the [24:16] conversation. uh because you've clearly [24:20] operated differently, you've thought [24:22] differently. And so it it seems really [24:24] like being bold right now uh now is the [24:28] moment to do that because we just we [24:29] have we have a lot of tools and and [24:32] power with us uh with with the the AI [24:34] tools that are there. Um another thing [24:37] that really stands out is the the being [24:39] flexible piece, right? The dynamics of [24:41] road maps and planning and and strategy [24:43] and even role. So being flexible in in [24:47] what you bring to the table and and what [24:48] tools you use in in different projects. [24:51] Um and then I I think the the last thing [24:54] that really stands out to me is taking [24:56] risk because again you entered a new [24:59] market and and and and a new way to [25:01] engage with AI which is still relatively [25:04] new to the you know general users that [25:06] are out there. Um but when you bring [25:08] that recipe together of being bold and [25:10] flexible and taking risk really great [25:12] things can happen. Mhm. [25:14] >> Um do you have any other way that that [25:16] something you would add or or something [25:17] you would summarize differently? [25:19] >> That's a great way great way to sum it [25:20] up. I think that you know in some ways [25:22] the the approach both to the product and [25:25] also your role kind of in the industry [25:27] today is similar. you know, like you [25:29] said, where there's areas where there is [25:32] a fair amount of structure that's [25:33] useful, but then being much more elastic [25:36] and flexible both in the product and [25:37] your understanding of what you work on, [25:39] I think is the right approach now for [25:41] what really feels like to me like a full [25:44] gear shift, you know, like if there's [25:45] any feeling I have from this year is [25:47] like this is the year where things [25:49] changed and we went into the next year, [25:52] you know, and I both on the what the [25:53] products are capable of, but also what [25:55] we're capable of uh as people working on [25:57] them, you So I think uh it's really [26:00] exciting and I couldn't be happier to be [26:02] in the middle of it. Yeah, I completely [26:04] agree and I really like that you said [26:06] people working on it because I think you [26:08] know there there's a lot of other [26:10] perspectives that suggest that people [26:11] aren't as important but I think the [26:13] conversation here proves that the the [26:14] people stay still play a very important [26:16] role in in in where things go both in [26:20] terms of the development of product but [26:21] then also as a user you're just un your [26:24] creativity is unleashed and so I'm [26:26] really excited about that and I really [26:28] appreciate the time today it's like [26:30] really fun product and it's also really [26:32] fun to talk about. So, thank you very [26:33] much. I appreciate it. [26:34] >> Thank you for having me.