Amazon Blocks Muse Agents to Shield Ads as Shopify Embraces AI Buyers

亚马逊封锁 Muse 智能体以护航广告，Shopify 则拥抱 AI 买家 消费级 AI 将通过默默追索退款和消除琐碎行政负担（而非宣传提升生产力）实现大规模普及，从而对依赖人工浏览的广告“收费站”构成威胁。 David Pawlan 与 Anish Acharya 深入剖析了来自 Assistant Benchmark 对 120 多款消费级 AI 智能体评估的实证洞察，指出自动追踪退款和琐事自动化等无形的“省钱型”功能，所带来的价值远超微弱的生产力提升。他们探讨了免提语音、即时通讯平台和可穿戴设备等界面的演进，强调主动执行虽能带来巨大惊喜，但一旦越界便可能粉碎脆弱的用户信任。对话还分析了更广泛的商业格局，将亚马逊针对智能体的防御姿态与 Shopify 的开放策略进行了对比，同时详细剖析了新兴的机器对机器协同层，以及维持自主智能体运行所面临的残酷经济学现实——据估计每位活跃用户每天需消耗约 20 美元的算力成本。 Assistant Benchmark 作为面向消费者的评估平台上线，在 16 个维度上对 120 多款个人 AI 智能体进行了测试，并在上线前两周内突破了 10 万名独立访客。 (3:41) 具备完整浏览器导航和多步骤推理能力的自主智能体，目前产生的算力成本约为每位活跃用户每天 20 美元，这给与科技巨头免费、补贴模式竞争的风险投资初创公司带来了巨大压力。 (47:34) 电商网络在智能体准入问题上出现分化：Shopify 欢迎买家智能体以提升商家的 GMV，而亚马逊则封锁 Muse 等工具以保护自身的广告收入和冲动消费利润空间。 (39:35) 智能体的消费者留存核心在于帮用户追回行政事务相关的成本——例如自动化提交 HSA 理赔和追回机票差价退款——而非抽象的 10% 生产力提升。 (10:03)

## Background 本期节目由知名风险投资机构 a16z 于 2026 年 9 月 29 日在其官方 YouTube 频道发布。视频中 a16z 普通合伙人 Anish Acharya 与 Assistant Benchmark 创始人 David Pawlan 展开对谈，围绕个人 AI Agent 的集中爆发，深入探讨了其融入大众日常生活的核心驱动力、隐形工作流、人机交互形态以及未来商业基础设施的重塑。 ## TL;DR 对话系统梳理了个人 AI Agent 的演进浪潮与真实使用画像，明确指出用户并不在乎 10% 的微小效率提升，默默在后台帮用户捡回退款与报销的“省钱”能力才是杀手级场景。David 指出主动性（Proactivity）是平庸产品与顶级 Agent 之间的绝对护城河，但这也伴随着一旦自作主张越界便会瞬间摧毁用户信任的微妙平衡。两人深入剖析了硬件伴侣、群聊静默监听代理以及 Shopify 拥抱开放与亚马逊构筑高墙的平台博弈，并推演了由 Agent 互联驱动的次级商业网络。最后，对谈直面单个活跃用户每天高达 20 美元的推理成本悬崖，呼吁创业者摆脱靠补贴卷价格的死局，转而探索千元级定价的高价值垂直领域以构筑真实壁垒。 ## Atmosphere & style 整场对话洋溢着硅谷前沿技术圈特有的智识兴奋与极客幽默，节奏紧凑明快且信息密度极高。Anish 以资深投资人的宏观视角提出敏锐的商业架构思考，而 David 则凭借数百款 Agent 实测的一线数据与真实社群反馈形成严密佐证，两人频频抛接妙梗、默契十足。探讨既有对“Agent 擅自替人分手”这类科幻式未来的诙谐调侃，又始终贯穿着对技术成本悬崖、平台生态防御与真实人性心理的清醒审视，给人带来置身于下一代计算范式重构前夜的强烈临场感。 ## Bulletin - [10:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=603s) 个人 AI Agent 正从炒股致富的幻觉转向“隐形省钱”：通过自动核销 HSA 医疗支出、监控机票降价申领差额积分或动态调度家庭水费，此类低频高价值、无需人工看管的自动化体验展现出极强黏性与商业化潜力。 - [14:42](https://www.youtube.com/watch?v=3T5sij3spWw&t=882s) Meta Muse Charm 挂坠被视作扎克伯格为元宇宙测绘物理世界的“数据采集器”；与此同时，去除面部摄像头的全天候纯音频（Audio-only）智能眼镜因消除社交侵入感与隐私焦虑，正成为硬件伴侣赛道的黑马路线。 - [40:01](https://www.youtube.com/watch?v=3T5sij3spWw&t=2401s) 电商巨头对 Agent 接入策略出现分化：亚马逊为捍卫站内高利润竞价广告与冲动消费而果断封禁 Muse，而依赖交易额抽成并赋能独立卖家的 Shopify 则全面拥抱开放互联。 - [47:35](https://www.youtube.com/watch?v=3T5sij3spWw&t=2855s) 现阶段网页端自主执行 Agent（Browser Use）的底层推理与操作成本高达每个活跃用户每天约 20 美元，初创团队若仅依赖资本补贴和算力通缩博弈将面临严峻挑战，商业化需转向高客单价或极度垂直的窄众付费市场。 ## Key takeaways - [00:57](https://www.youtube.com/watch?v=3T5sij3spWw&t=57s) 个人自主代理（Personal Agent）正迎来继底层通用大模型对话、编程代码代理（Coding Agents）之后的第三波 AI 爆发浪潮，加速渗透进日常琐事与软硬件生态。 - [05:11](https://www.youtube.com/watch?v=3T5sij3spWw&t=311s) 尽管差旅自动预订最易在社交媒体上制造轰动噱头，但用户真实使用频次最高的痛点集中在日常琐碎行政事务（如清理邮箱、填报表单）以及多 Agent 间的通信与编排调度。 - [08:46](https://www.youtube.com/watch?v=3T5sij3spWw&t=526s) 金融 Agent 的核心价值并非帮用户暴富，而是打破信息壁垒并消除冗长流程，将传统金融机构利用消费者被动与拖延赚取的庞大利润重新“夺还”给普通人。 - [10:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=603s) 优秀助理的核心属性是“隐形”，大众对抽象的效率微增并不敏感，但通过后台静默运行实现自动索取差价退款等“白捡免费钱”的机制能迅速建立深度信任与产品黏性。 - [17:26](https://www.youtube.com/watch?v=3T5sij3spWw&t=1046s) 双手被占用的真实场景（如骑行通勤、做家务）是全双工语音 Agent 的关键魔法时刻，ChatGPT 语音接入日历和邮件并实现“收件箱清零”验证了纯免提交互的巨大价值。 - [20:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=1203s) 在多人协作与群聊场景中，不插话打断、仅静默沉淀文档并通过私聊独立跟进待办的“静默代理”（如 Doc 模式），比过度社交化、低情商插话的传统机器人更符合人类实际需求。 - [23:23](https://www.youtube.com/watch?v=3T5sij3spWw&t=1403s) 单纯的语气风格或性格文件极易被提示词替换，并不构成壁垒；面对模糊边界时“先斩后奏”的武断度倾向（Presumptuousness）与主动性，才是划分 Agent 优劣的底层工程护城河。 - [26:13](https://www.youtube.com/watch?v=3T5sij3spWw&t=1573s) 主动性是一把双刃剑：代办登机或申领退额等零负担操作让用户惊艳，但在影响用户生活决策的高风险场景下一旦越界犯错一次，产品信任体系就会瞬间崩塌。 - [28:23](https://www.youtube.com/watch?v=3T5sij3spWw&t=1703s) 消费级 Agent 呈现向“窄众初创公司（Narrow Startups）”分化的趋势，通过针对特定群体（如育儿单亲妈妈、本地礼宾）提供极深定制与专属数据，即使客户基数较小也能依靠高订阅费达成可观规模。 - [32:04](https://www.youtube.com/watch?v=3T5sij3spWw&t=1924s) 被动响应的 Assistant 正在被具备主动能动性的 Agent 取代；此外，Agent 的非人格化特质还可作为人类沟通的“社会间接缓冲器”，有效柔化人际摩擦与敏感反馈。 - [38:06](https://www.youtube.com/watch?v=3T5sij3spWw&t=2286s) 随着 Agent-to-Agent 交互成为常态，围绕智能体互通的专属邮箱、虚拟电话号码、身份凭据与自动化安全防护协议将催生全新的底层网络基础设施产业。 - [42:21](https://www.youtube.com/watch?v=3T5sij3spWw&t=2541s) 供给受限市场（如热门餐厅）将因机器人秒抢演变为算力对抗或独占性筛选，而需求受限市场（如临期机票）则将由 Agent 演化出反向聚合竞价机制，迫使商家让利。 - [44:08](https://www.youtube.com/watch?v=3T5sij3spWw&t=2648s) 在由 Agent 主导的商业推荐中，拥有社交互动数据的平台（如 Meta）能跳出货架竞价逻辑，甚至为用户发掘独立手作人等长尾优质供给，创造意想不到的消费惊喜。 - [46:06](https://www.youtube.com/watch?v=3T5sij3spWw&t=2766s) 智能体间直接磋商（Agent-to-Agent）将催生次级经济生态，使手工艺人等个体劳动者摆脱客服与接单流程，并使商品推荐真正与消费者利益结盟、剔除品牌营销溢价。 - [47:35](https://www.youtube.com/watch?v=3T5sij3spWw&t=2855s) 目前市面上绝大多数 Agent 产品依赖付费或增值模式生存，初创团队与其依靠资本补贴陷入每天 20 美元成本的泥潭，不如打造能支撑每月上百甚至上千美元的高价值产品，静候算力通缩红利。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=3T5sij3spWw&t=0s) 个人 AI Agent 的爆发前瞻 播客开场通过快节奏的精彩预告抛出了几个极具冲击力的核心观点：普通大众其实并不在乎效率提升 10%，他们需要的是切实的价值；Meta 的 Muse charm 硬件很可能不是为了单纯争夺硬件设备市场，而是扎克伯格为了收集现实世界物理数据、测绘真实场景，从而为其长远的元宇宙布局提供养料；而在人际社交与边界感上，团队内部甚至开玩笑称“不久之后就会有 Agent 擅自发信息说：‘我发现你其实没那么喜欢她，所以替你提了分手’”。此外，关于 Agent 主动性（Proactivity）的壁垒与风险也被重点提及——骑车通勤的 30 分钟内通过语音指令完成收件箱归档、日程发送以达成 Inbox Zero 的体验固然美妙，可一旦 Agent 越界犯错一次，用户的信任就会瞬间崩塌。未来更将催生专门服务于“Agent 间互相交互”（Agent-to-agent）的基础设施与全新软件生态。 ```mermaid flowchart TD accTitle: AI Agent 演进的三大浪潮 accDescr: 展现从基础大语言模型对话，到编程代码代理，再到全能个人自主代理的三个演进阶段。 subgraph Wave1["第一波浪潮：认知颠覆"] A["ChatGPT 问世 (2022末-2023)<br/>拟真对话 / 撰写邮件 / 诗歌创作"] end subgraph Wave2["第二波浪潮：生产力变革"] B["Coding Agents 爆发<br/>Claude Code / Codec<br/>极低门槛的软件构建能力"] end subgraph Wave3["第三波浪潮：个人全能代理"] C["Personal Agents (Instinct / Muse)<br/>接管日常琐事 / 现实任务闭环 / 软硬件融合"] end A --> B B --> C ``` ### [00:57](https://www.youtube.com/watch?v=3T5sij3spWw&t=57s) 个人助理 Agent 的演进脉络与技术浪潮 a16z 合伙人 Anish Acharya 欢迎 Assistant Benchmark 的创建者 David Pawlan 共同做客节目。Anish 回顾了整个技术生态的数次关键认知奇点（"seeing God" moments）：第一波是 2022 年底至 2023 年初 ChatGPT 的横空出世，让人们初次感受到与合成数字人对话、写诗写信的震撼；第二波是以 Claude Code 和各类编程 Agents 为代表的工具，让开发者具备了举手投足间构建软件的能力；而眼下正在经历的第三波浪潮，则是以 Instinct、Muse 以及 ChatGPT 工作流为代表的个人自主代理（Personal Agents）。 David Pawlan 梳理了过去几周乃至数月的爆发节奏： - **Poke 的破局尝试**：早在 9 月 8 日上线（David 于 9 月 11 日成为早期用户），其最具特色的机制是允许用户就“到底该付多少服务费”与 Agent 展开博弈和谈判，展现出极其惊艳的交互体验。 - **开源生态推进**：随后 11 月下旬 OpenClaw 等项目推出，在极客社区引发轰动，预示着更大的浪潮正在逼近。 - **推特圈层的全面引爆**：紧接着 Instinct 彻底席卷了 Tech Twitter 圈子，随后便引发了多米诺骨牌效应，Grockbot、Muse 以及诸如 Caddy、Ali、Pi 等长尾产品相继喷涌而出。整个科技圈正处于好奇心爆棚、四处试水的“狂野西部”（Wild West）时代。 ### [03:41](https://www.youtube.com/watch?v=3T5sij3spWw&t=221s) Assistant Benchmark：用现实用例检验 Agent 真实战力 为了评估层出不穷的产品，David 创办了 Assistant Benchmark。这并非一个面向底层的技术参数评测平台，而是一个高度面向终端消费者的横向用例对比项目。团队向各大 AI 助手输入相同的真实 prompt（例如：“帮我订一张去芝加哥的机票”、“在我当前定位 5 个街区内找一家素食餐厅”等），并在统一标准下评测其端到端能力。 - **多维评估体系**：平台围绕 16 个不同维度考察 Agent 的单次表现（one-shot prompt），包括任务能否真正跑通、是否会主动追问澄清问题、响应耗时、动作准确度等，帮助普通用户快速判断订机票或查信息时哪一款最趁手。 - **爆发式行业反馈**：自项目上线仅 16 天以来，网站独立访客数迅速突破 10 万，几乎市面上所有相关赛道的创业公司创始人都主动联系 David 寻求反馈与对接。 ### [05:11](https://www.youtube.com/watch?v=3T5sij3spWw&t=311s) 市场格局分化与用户核心需求 目前市场上的 Agent 解决方案呈现出清晰的双层结构与专业化分野： | 类别划分 | 典型代表 | 核心定位与场景 | |:--- |:--- |:--- | | **B2C 通用型 Agent** | Instinct, Muse, Caddy, Ali, Season 等（平台上已有 64 款） | 面向大众消费者的日常综合管家，试图覆盖全场景任务 | | **B2C 垂直领域 Agent** | Soar, Miso 等 | 专注于单点纵深，如差旅机酒预订、专属邮箱整理等 | | **B2B 工作流/高管助理** | The Town, Catch, Vellum 等 | 充当企业高管数字 EA（Executive Assistant），深度接管业务流程 | 在目前统计到的 122 款产品中，绝大多数团队都在扎堆尝试攻克“水平通用型”（Horizontal generalized）方案。针对“差旅预订”这种低频但高价值的场景，David 指出尽管差旅是推特上最容易被拿来发视频炫耀的场景（例如自动值机、选座买票），但它本质上更像是一种“吸引眼球的噱头”（gimmick/attention grabber），长期看很难独立成为超级入口，更可能是作为底层能力连接器（Connector）被打包嵌入进水平 Agent 中。 基于 David 维护的 7-8 个专属社群（涵盖 1,200 多名深度 AI 助手发烧友）的实际追踪，用户日常使用频率最高的真实用例分布为： 1. **日常琐碎行政事务（Daily Admin）**：收件箱深度整理与归档、自动填写报销或申请表单等让人头疼厌烦的杂务； 2. **Agent 间协同编排（Agent Orchestration）**：如何寻找最高效的特定 Agent、如何让负责不同任务的 Agents 彼此顺畅通信与数据交接； 3. **移动端与交互式开发（Development）**：Tech Twitter 核心群体高度关注通过手机短信/即时通讯直接调度 Claude Code，甚至是设计师直接发文字消息让 Agent 对设计图纸做微小迭代与修改。 ### [08:46](https://www.youtube.com/watch?v=3T5sij3spWw&t=526s) 金融 Agent 的误区与真相：从“暴富幻想”到繁琐行政减负 在个人 Agent 测评的诸多类别中，金融（Finance）无疑是最引人入胜、却也最不切实际的一个领域。当普通用户手里拿到一个能力强大的 Agent 时，第一反应往往是：“你能怎么帮我赚到一百万甚至十亿美元？”然而现实很残酷：如果仅凭一部手机里的 Agent 就能在公开市场上躺赚数十亿，华尔街的对冲基金早就把这种套路用尽了。 科技分析师本·汤普森（Ben Thompson）此前在播客中提出过一个犀利观点：绝大多数消费者在日常生活中根本不追求所谓的“生产力”或效率提升，大家真正的诉求是“消磨时间”（Spend time），而不是“节省时间”（Save time）。但金融领域恰恰是一个例外——普通人在这里面临着极其繁重的日常行政杂务与流程摩擦。美国金融体系中相当庞大的一部分商业利润池，本质上正是建立在消费者的冷漠、被动以及信息闭塞之上的。如果 Agent 能够介入并打破这种信息差，原本属于金融机构的利润将被实打实地重新交还给消费者。 - **暴富幻觉 vs. 现实价值**：大众总幻想 AI 炒股致富，但 Agent 真正能落地的是消解繁琐的财务行政琐事。 - **利润池转移**：传统金融机构利用用户的拖延与信息不透明所赚取的巨额沉淀资金，将被自动化工具反向“吐还”给普通人。 --- ### [10:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=603s) 帮用户“白捡钱”：隐形 Agent 的真正护城河 大众并不在乎如何让自己的生活效率提升 10%，优秀助理的核心定义永远是“隐形”（Invisible）——就像职场中最好的员工那样，他们默默在后台把事情处理妥当，回头只交出一份干净利落的成果单，而不是整天追着管理者汇报琐碎细节。当前最引人注目的高价值 Agent 并非那些空谈投资收益的“造钱机器”，而是脚踏实地的“省钱工具”（Cost savers），它们深入日常生活的缝隙中自动运转： ```mermaid flowchart TD accTitle: 隐形 Agent 的省钱自动化工作流 accDescr: 展示 Agent 如何在后台默默执行低频高价值的省钱与降本操作 subgraph Trigger["触发与监控"] A["历史消费与票据"] B["已预订机票监控"] C["天气预报与外部环境"] end subgraph AgentAction["隐形 Agent 自动化处理"] D["梳理全年支出并提交 HSA 报销"] E["比对票价差额并向航司申请 Credit"] F["动态调度家庭喷淋系统"] end subgraph Outcome["用户获得收益"] G["取回医疗账户沉淀资金"] H["航班差价返还（如白捡现金）"] I["水费直接降低 50%"] end A --> D --> G B --> E --> H C --> F --> I ``` 有用户让 Agent 扫描自己全年的消费收据，自动向健康储蓄账户（HSA）提交报销申请；有人在预订机票后，让助理持续监控价格走势，只要航司降价就自动向客服索取差额旅行点数（Travel credit）；更有极客将 Groqbot 接入家里的草坪喷淋系统，与天气预报联动，直接将每月的水费账单砍掉了 50%。美国消费者天生讨厌听人说教如何“节俭攒钱”，但所有人对“白捡免费钱”（Free money）毫无抵抗力。当差价退款和 HSA 报销像免费馅饼一样自动掉进钱包时，这种低频但高价值（Infrequent but high-value）的体验将形成极强的黏性，更有机会孕育出估值数十亿美元的垂直连接型独角兽。 - **最佳助理形态**：“你雇佣一个人绝不是因为你享受管理的过程”，最好的 Agent 应当无需人工看管。 - **“白捡钱”心理学**：把复杂的核销流程包装成不劳而获的直接收益，是促成大众采纳的核心引线。 - **垂直整合机会**：机票比价退差、医疗支出核销等场景虽然不是每天发生，但单次经济价值极高，极具商业爆发力。 --- ### [12:24](https://www.youtube.com/watch?v=3T5sij3spWw&t=744s) 交互界面的代际与场景之争：短信流、独立 App 与桌面组件 在 Agent 的具体交互形态上，行业正经历剧烈的路线分化。一部分硬核用户极其偏爱 iMessage 这类即时通讯界面，因为这是个人手机上最核心、最私密的特权领地，在聊天气泡里呼叫 Agent 感觉像是自身数字生活的自然延伸，完全没有额外打开软件的流程负担。 然而，不同人群之间存在着鲜明的习惯割裂： - **代际差异**：通常而言，Z 世代（Gen Z）更加推崇极简的短信文本流；而千禧一代与 X 世代往往习惯功能结构完整的独立 App。 - **性别与使用诉求的差异**：例如在旅行场景中，注重视觉呈现与心愿清单（Dream board）的用户更偏爱富媒体应用界面；而务实型用户只要一张能直达目的地、机上配有星链（Starlink）网络的机票，纯文本交互足以满足。 - **新型桌面触点**：除了短信和独立 App 之外，像 Signal 旗下的初创项目 Sky 正在尝试通过 iPhone 桌面小组件（Widget）来切入交互入口。 --- ### [14:42](https://www.youtube.com/watch?v=3T5sij3spWw&t=882s) 硬件终局是挂坠还是数据采集器？音频伴侣的隐秘赛道 硬件领域的探索正呈现出百花齐放的态势，从胸前挂坠（Pendant）到智能表带，各类形态层出不穷。正如不同的人对耳环、项链等珠宝首饰有不同的审美，穿戴硬件的形态最终也会高度碎片化和个性化。这类硬件的核心价值在于它能全天候环境伴随（Ambient），被动捕捉用户在现实世界中许下的承诺、待办事项和环境上下文。不过，这也必然引发新的社会契约与隐私冲突——就像华盛顿特区的部分高端私密餐厅会收缴顾客手机一样，未来餐厅甚至可能要求顾客在进门前摘下所有智能吊坠与智能眼镜。 针对 Meta 刚刚推出的 Muse Charm 挂坠，行业内产生了一个犀利的观点：Muse Charm 的目标或许根本不是在硬件消费市场上取得大众普及，其本质是扎克伯格为了元宇宙建立现实世界数字孪生的高维“数据采集器”。通过摄像头和多阵列麦克风，它能让科技圈核心尝鲜者在全天候佩戴中为 Meta 提供海量物理世界的真实行为数据。 相比需要主动操作且续航受限的智能眼镜，这种全天候伴随的轻量化硬件展现了不同的演进路径。与此同时，纯音频智能眼镜（Audio-only glasses）正在成为不可忽视的黑马形态。正如科幻小说《安德的游戏》（Ender's Game）中那个无所不知的纯音频 AI 伴侣“简”（Jane）一样，避免在脸上安装具有侵略性的摄像头，反而大幅消除了他人的隐私警惕，在现实中可能具备更广泛的接受度与穿透力。 - **首饰化属性**：硬件创业者往往认为自己的形态是终极答案，但硬件本质上更像珠宝，注定高度客制化与多元化。 - **环境计算伴侣**：硬件的核心价值在于 24/7 全天候捕捉真实语境，而不仅限于坐在屏幕前的手动输入。 - **纯音频路线的优势**：去除面部摄像头降低了社交不适感，全知式纯语音交互或将成为最具可行性的全天候伴侣形态。 ### [17:26](https://www.youtube.com/watch?v=3T5sij3spWw&t=1046s) 语音交互的“魔法时刻”与免提场景价值 对话从可穿戴设备的无感体验切入，指出当人们既能享有科技连接性，又不必承受摄像头带来的社交尴尬感时，这种产品形态具有极强的吸引力。David Pawlan 对语音和音频赛道表现出极高的看好度。他提到 OpenAI 刚为 ChatGPT 语音模式推出了外部连接器（Connectors），允许直接接入 Gmail 和 Google Calendar 等工具。 David 分享了当天早晨的亲身体验，称其为自己在整个 AI 助理领域的“第二个魔法时刻”。当时他正骑自行车上班，双手无法腾出，为了处理待办事务便直接开启 ChatGPT 语音会话。在短短 30 分钟的通勤骑行中，助理完成了以下全部操作： - 对未读邮件进行分类并打上标签； - 撰写并发出多封邮件回复； - 规划时间并发送日历邀请。 当他踏进办公室坐到工位时，收件箱已经实现“清零”（Inbox Zero），这种体验让他感到无比震撼与兴奋。 Anish Acharya 澄清并确认这确实是 ChatGPT 语音而非 Muse 的功能。Anish 解释道，该功能基于 GPT-4o（GPT Live 1）模型，支持真正的全双工（full-duplex）语音对话，并且能够调用用户所有的历史会话线索。用户可以随时打断并追问宏观状态（例如询问“我之前开启的那三个项目进展如何、卡在什么地方”）。虽然马克·扎克伯格（Mark Zuckerberg）也已预告了全双工语音功能，但目前无论 Muse 还是 Instinct 都不具备这种深度整合能力。 David 进一步强调，对于大众消费者而言，AI 助理最有价值的时刻往往是“双手被占用”的时刻（例如烹饪、园艺、做家务），而不是坐在沙发上盯着界面。科技圈内部那种整天泡在电脑和手机前的“深度网瘾”环境，与普通大众的生活脱节。未来助理要实现大众化普及，必须从极度痛苦的高阻力琐事切入——比如收到交通罚单时，用户只需随手拍张照，后台代理便能自动完成缴费全流程。 --- ### [20:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=1203s) 多人协作与群聊中的静默代理：以 Doc 为例 两人随后探讨了 AI 代理进入社交场景与“多人游戏（Multiplayer）”状态的难题。Anish 提出，iMessage 等私域通信渠道虽然具有独特的特权地位，但直接将 AI 代理拉进多人真实群聊往往会显得突兀、打破对话节奏，甚至损耗用户的社交资本。 David 归纳了业内目前尝试将代理引入多人场景的三种主要路径： ```mermaid flowchart TD accTitle: AI代理进入多人协作的三种架构模式 accDescr: 对比传统聊天机器人、代理间幕后对齐以及静默监听型文档记录者三种模式的工作机制。 subgraph Mode1["1. 传统群聊机器人"] A1["群成员对话"] --> B1["机器人直接插话回复"] B1 --> C1["容易打断节奏 / 降低社交资本"] end subgraph Mode2["2. 代理间对齐网络（Instinct 模式）"] A2["用户 A 的代理"] <-->|"幕后协议交换数据"| B2["用户 B 的代理"] A2 --> C2["各自向主人同步进展"] end subgraph Mode3["3. 静默旁听者（Doc 模式）"] A3["多人正常群聊"] -->|"完全静默旁听"| B3["Doc 代理"] B3 --> C3["自动沉淀结构化文档"] B3 -.->|"私聊单独催办"| D3["特定相关责任人"] end ``` David 重点推介了第三种方案——由 Shane Mack 打造、与 XMTP 相关的项目 **Doc**（其交互哲学类似于会议记录工具 Granola）： - **全静默旁听**：代理作为无感存在静默驻留在群聊中，绝不在主群随意插话抢麦； - **私信独立追踪**：若识别出明确的待办事项（Action Item），它会单独向具体责任人发起私聊跟进，绝不打扰群里其他多数人； - **文档化归纳**：用户想要回顾时，可以直接查阅其自动生成的结构化文档。 Anish 对这一设计大为赞赏。他援引 Obo 的 Nir 所提出的观点：底层大模型在代码等“可验证领域”（Verifiable Domains）的进展呈指数级飞跃，但在拟人行文、情商与“就诊沟通态度（bedside manner）”这类微妙的人际表现上已进入平台期甚至退步。试图让低情商的 AI 在人类群聊中强行社交通常体验糟糕；相反，高度功利主义（Utilitarian）的工具定位才能真正产生正向价值。Anish 举例，他曾尝试将代理拉入黑胶唱片收藏同好的讨论群，尝试自动编目彼此的藏品与品味，这类偏向垂直数据整理的探索远比让机器人胡乱插科打诨有用得多。 两人还对比了不同平台的交互心理：在即时通信群聊中，发言好比“独占麦克风”，机器人的插话让人极具侵入感；而在 Discord 或类似 Reddit 的公共论坛中，代理发帖更像是异步的信息节点，用户可看可不看，心理排斥感明显更低。 --- ### [23:23](https://www.youtube.com/watch?v=3T5sij3spWw&t=1403s) 拟人化陷阱、代理性格与“先斩后奏”的护城河 面对当前的产品设计，两人反思行业是否存在“过度拟人化”的问题。许多团队试图生成逼真的虚拟人形象，反而引发恐怖谷效应与人际不适；相比之下，Meta Muse 采用呆萌可爱的“雪人（Yeti）”吉祥物形象，反而迅速赢得了全球用户的喜爱与信任。 | 代理定位机制 | 典型表现 / 案例 | 核心问题 / 行业判定 | |:--- |:--- |:--- | | **过度逼真虚拟人** | AI 生成的人类面孔与姿态 | 让人感到怪异与侵犯私人空间，心理抗拒感强 | | **萌系非人形象** | Muse 的可爱雪人（Yeti） | 降低防备心，用户更愿意在日常中包容与交互 | | **鲜明泼辣性格** | Poke（以机智、嘴损/Sassy 著称） | 风格两极分化严重，部分用户因其语气尖锐而拒绝交出个人数据 | | **可配置性格文件** | Soul File / 长期记忆调整 | **无防御壁垒**：仅需一条提示词或记忆配置即可随意切换 | | **行事准则（Constitution）** | “先斩后奏”的主动决策权与容错倾向 | **具备深度护城河**：关乎代理在模糊场景下的执行力与自动化成败 | 针对市场是否会按性格原型细分（如同现实中有人喜欢威严权威的医生，有人偏好亲切平等的医生），David 认为单纯的“沟通风格”并非真正的竞争护城河。例如 Poke 虽然凭借极具个性的犀利言语（Sassy）迅速打响名气，但也导致大量用户因反感其傲慢态度而拒绝授予隐私权限。更重要的是，语言风格非常容易配置，用户只要在“灵魂文件（Soul File）”或个性化记忆中输入偏好指令，任何大模型代理都能无缝调整。 Anish 则从“立足架构（Constitution）”的角度提出了反驳与修正：性格的深层其实是**行动倾向性**，最典型的指标是**武断度/自作主张的程度（Presumptuousness）**。 - **高武断度场景**：某些自动化事务中，用户宁愿代理承担 10% 的犯错率，也希望它能果断推进并先斩后奏把事情办妥； - **极度谨慎场景**：在涉及财务、税务或关键账户时，用户要求代理必须步步确认、绝对审慎。 这种在面对未知与模糊边界时的“仲裁机制（Tie-breaking）”，实质上已从浅层表辞下沉为底层工程能力。David 对此表示赞同，并指出**主动性（Proactivity）**才是划分平庸产品与最终赢家的绝对护城河——能够真正走在用户前面、自发处理现实事务的代理，才能在这场竞赛中立于不败之地。 ### [26:13](https://www.youtube.com/watch?v=3T5sij3spWw&t=1573s) 主动性与用户信任的微妙平衡 在谈到隐形助手（invisible assistant）未经用户明确指示便主动代劳的场景时，David 认为目前最惊艳、最让人产生“顿悟时刻”（aha moment）的噱头功能就是“自动为旅客办理航班登机（check-in）”。但他也指出，这种主动性必须严格划分为两个大类： 一类是用户希望牢牢保留控制权、不希望 Agent 过度越俎代庖的工作流。这类工作流往往需要用户自身承担行为改变的后果，例如 Agent 发现你当前的保险费率过高并建议更换，但它必须先获得许可才能真正操作，因为这直接影响你的实际生活决策；另一类则是“零负担”的主动操作，比如替你起草好一封邮件，或者主动去帮你争取到一笔航空退款抵用券（flight credit）。这种场景不需要用户额外付出精力，单纯是在省钱、省时间。David 警告称，两者之间的界限极其微妙，一旦越界一次，产品就会立刻失去用户的全部信任。 ```mermaid flowchart TD accTitle: 个人 Agent 主动性分类与信任边界 accDescr: 对比低风险无感操作与高风险需用户确认的工作流。 subgraph LowRisk["低风险：纯增益工作流（完全主动）"] L1["起草邮件"] L2["主动申领航班退款 / 抵用券"] L3["自动办理航班登机手续"] L1 & L2 & L3 --> LEnd["用户无需负担，直接省钱省时间"] end subgraph HighRisk["高风险：影响决策工作流（需授权确认）"] H1["识别保险收费过高并提议更换"] H2["直接代办解约或跨服务迁移"] H1 --> HAuth{"获得用户明确授权？"} HAuth -- 是 --> HExec["安全执行操作"] HAuth -- 否 --> HBreak["触碰红线，用户信任瞬间崩塌"] end ``` Anish 随即开玩笑表示，团队内部私下常常调侃，按照现在的势头，距离我们看到 Agent 擅作主张给别人发消息说“我注意到你其实没那么喜欢她，所以我顺手帮你把她甩了”的日子恐怕就差几天了。不过 Anish 坚信，这种带有某种“冒昧感（presumptuousness）”的主动性正是其魔力所在。如果大家不在社交平台 X 上疯狂讨论“天哪，我的 Agent 居然做了这个做了那个”，反而说明目前的 Agent 产品在探索主动性方面还不够激进、用力不足。 谈及 X 上的热门推文，两人提到了一个广为流传的轶事：有人发推吐槽 Instinct 在帮他办理登机时，竟把他的中间名幻觉成了“Bong Chang”，导致他最终未能登机。Anish 怀疑这可能只是个为了流量编造的段子，但也好奇当事人是否还在用该产品；David 则笑称，冲着推文引发的巨大关注度，那人估计为了源源不断的段子素材也得继续用下去。 ### [28:23](https://www.youtube.com/watch?v=3T5sij3spWw&t=1703s) 从 OpenClaw 到极度垂直的窄众创业（Narrow Startups） 随后话题转向底层技术的演进。Anish 指出，在 OpenClaw 这一开源探索期沉淀下来的许多基础原语（primitives），如今已被产品化包装成大众消费者可以轻松理解和接受的形态。他高度评价了被 Cognition 收购的 Poke 团队，认为他们才华横溢、产品设计极为巧妙，如果 Poke 的创业时间晚开始 6 个月，很可能就会成为最终的赢家。 针对从 OpenClaw 继承的能力，David 直言，目前的消费级 Agent 本质上就是预先配置好的 OpenClaw，核心能力其实并没有超越 OpenClaw 或 Hermes，关键的区别在于： - 用户不再需要手动去配置复杂的参数； - 不必忍受系统每隔七小时就弹出一堆报错代码； - 提供了一个真正简单易用的前台界面。 展望未来，David 认为 Agent 赛道除了持续强化主动性外，必然会走向“超高度专业化（hyper specialization）”。这正好契合了 Anish 长期倡导的“窄众初创公司（Narrow Startups）”投资理论： - **颠覆性的商业模式**：过去罕见有软件能对极小的人群收取每月 200、250 甚至 300 美元的高费率，而如今只要拿下几万名精准垂直用户，就能建立起年化营收（Run Rate）过亿美元的庞大业务。 - **深度定制与专业壁垒**：例如开发一个专门针对家有 3 岁以下幼儿的单身母亲的专属助手，它熟知单亲妈妈的每日作息和育儿知识，互动起来宛如拥有读心术。 - **品味与专属数据的比较优势**：就像准妈妈会挑选认同特定分娩理念的导乐（Doula）一样，用户也会挑选认同特定处事风格、特定知识体系的专属 Agent；亦或是一个只针对纽约市的高级数字礼宾代理，凭借不可复制的高级品味与本地专属数据脱颖而出。 ### [32:04](https://www.youtube.com/watch?v=3T5sij3spWw&t=1924s) “Assistant”与“Agent”的定义分歧与社会润滑剂角色 围绕行业术语上频繁混用“助理（Assistant）”与“代理（Agent）”的现象，两人展开了深层辨析。Anish 认为 Assistant 是一种较低级的形态，根据定义，Assistant 只能在被要求时提供辅助，完全由人指令驱动；而 Agent 则真正拥有主观能动性（Agency），能够自主让好的事情发生，甚至能在社交圈子中扮演与人类平起平坐的同伴角色。这就像企业级 AI 从最初打杂的实习生，随着智能程度与可靠性提升，逐渐被“提拔”去承担更高阶、更精细、更微妙的工作。 针对人机交互的心理边界，David 探讨了称谓对人类社会层面的心理影响：究竟“Agent”一词能让人保持理性的非人格化认知，还是“Assistant”给人赋予了过多的人性温度而让社会感到不适？面对旧金山与其他地区的认知差距，Anish 笑着否认了科技泡沫一说，坚称旧金山只是提前步入了“不远的未来”。 此外，Anish 还提出了一个极具洞见的观察：Agent 的“非人”特质反而能为人类社交提供一层宝贵的“间接缓冲（social indirection）”。 - **充当冲突的非情绪化仲裁者**：在面对尴尬的社交动态、人际摩擦或难以启齿的敏感反馈时，一个精心设计的 Agent 能作为非人格化的媒介去组织沟通、传达反馈。 - **降解人际对抗的火药味**：双方都知道对方是通过中立的数字化代理来传递意图，从而过滤掉非理性的情绪对立，使紧张的社会互动变得更加柔和、富有建设性。 ### [35:18](https://www.youtube.com/watch?v=3T5sij3spWw&t=2118s) 摆脱繁重行政负担与机器拓展的人生可能性 对话深入到了人工智能辅助人类生活的深层愿景。除了日常的工作效率、财务管理和健康优化，现实中存在大量极其繁琐的官僚行政流程压迫着普通人。例如全美每年约有 150 万人因酒驾（DUI）受罚，其背后伴随的行政文书与法律程序负担极其沉重，而普通人往往很难请得起专业律师。AI 助理的第一步使命，就是像富人身边的家族办公室（Family Office）或私人律师团队一样，替普通人化解这种“官僚压力”，在财务和健康上实现最优化配置。 更深层的思考在于“如何让人们过上想要的生活”。正如老话所说，“得到想要的东西最难的部分在于明白自己到底想要什么”。人们并不总是拥有明确刻板的目标，更多是在摸索中寻找真实的自我。当繁重的工作被接管后，未来的代际互动形态可能会发生彻底变化——未来的孩子们回看如今人们“低头盯着手机穿过马路”的录像时，大概会感到不可思议，不理解人们为何对着屏幕耗尽光阴。AI 接管繁琐杂务后，父母能有更多时间陪伴孩子，年轻群体也能有更多闲暇与朋友面对面交流。 这种技术带来的绝不仅仅是“成本削减”，更是“可能性的极大拓展”。以代码智能体为例，它们往往没有抢夺人类程序员的饭碗，而是在完成过去根本不可能立项的事情。例如构建一个能亲自走入其中的“虚拟唱片店”侧边项目，在过去需要花费整整一年时间和昂贵的人力，普通人根本不可能雇人去实现。AI 智能体让这些从未有机会问世的创意得以诞生，极大地丰富了人生的探索乐趣。 - 全美每年约有 150 万人经历 DUI 等法律行政流程，AI 助理能够为大众提供平民化的“家族办公室”与法律级支持。 - AI 的核心价值在于将人类从手机屏幕和繁琐杂务中解放，把生命归还给家庭、社交与自我探索。 - 编程智能体并非单纯降低开发成本，而是在拓展全新的创作边界，将大量原本因成本过高而流产的个人项目变为现实。 --- ### [38:06](https://www.youtube.com/watch?v=3T5sij3spWw&t=2286s) 智能体原生网络与新型商业基础设施 从近期趋势来看，伴随着 Instinct 推出的 Instinct-to-Instinct 网络等探索，整个互联网正在酝酿一套专为“智能体与智能体（Agent-to-Agent）互通”而设计的全新原生架构。传统的服务业是由“人对人”或“人对智能体”进行预订与履约，而当模式转变为“智能体对智能体直接交互”时，整个服务行业的底层逻辑都将被重写。 随之而来的是围绕智能体体系的新型基础设施与网络安全问题。历史规律表明，互联网的诞生催生了网络安全产业，而智能体的普及也将催生专门针对智能体交互的防御与安全技术。未来将涌现专属的智能体邮箱、智能体电话号码以及针对自动化通信的新型协议。资本与创业者正在加速涌入这些垂直场景，构建下一代去中心化的互联网与数字化连接。 ```mermaid flowchart TD accTitle: 商业网络演进与智能体交互模型 accDescr: 展示从传统互联网三大板块向智能体对智能体商业形态演进的过程 subgraph TraditionalWeb["传统互联网的三大形态"] A["内容网络 (Content Web)<br>长期面临变现与流量困境"] B["应用网络 (Application Web)<br>借由编程智能体迎来爆发"] C["商业网络 (Commerce Web)<br>迎来流量入口重组"] end subgraph AgentInteraction["智能体原生交互范式"] D["专属基础设施<br>智能体邮箱 / 电话号码 / 安全协议"] E["Agent-to-Agent 交互<br>智能体向智能体直接下单与履约"] end C --> D D --> E ``` - 传统网络分为内容、应用与商业三大版块；在代码智能体推动应用网络爆发的同时，商业网络正迎来最剧烈的入口重构。 - 智能体对智能体（Agent-to-Agent）直接交互将彻底终结以人类操作为假设的服务业旧流程。 - 智能体生态催生全新安全产业，涵盖凭据管理、自动化履约验证及反攻击防护。 --- ### [40:01](https://www.youtube.com/watch?v=3T5sij3spWw&t=2401s) 平台博弈：Shopify 拥抱开放 vs 亚马逊筑起高墙 智能体在电商领域的渗透迅速引发了平台之间的路线分化。典型案例就是 Shopify 积极接纳接入 Muse，而亚马逊（Amazon）则选择果断封禁 Muse。这背后折射出两家巨头在商业模式上的根本分歧： 1. **亚马逊的广告护城河与冲动消费**：亚马逊的核心利润高度依赖站内广告系统与人类用户的眼球停留。如果由智能体代替人类完成全自动下单，传统广告的展示与竞价体系便会彻底失效；同时，平台也将失去人类用户逛街时产生的“冲动消费”（Impulse Shopping），因此亚马逊出于防范被中介化（Disintermediation）的考量，势必阻断外部智能体。 2. **Shopify 的去中心化去赋能逻辑**：Shopify 的本质是赋能海量独立商家卖出尽可能多的商品，智能体的自动化比价与购买能够直接促进整体交易规模的繁荣，因此智能体反而成为了去中心化商业的高效倍增器。 在餐饮预订领域也出现了同样的平台防御战。例如预订平台 Resy 开始封禁使用 Browser Use 自动化操作的用户账号，理由是机器人的瞬间抢座破坏了原有的“以人为中心”的服务逻辑。一旦人人都配备了智能体，竞争将瞬间变成同一毫秒内的算力对轰。 | 平台 / 领域 | 应对策略 | 核心原因与利益冲突 | |:--- |:--- |:--- | | **Amazon** | 封禁智能体（如 Muse） | 依赖站内广告变现与眼球注意力，智能体绕过广告和冲动消费会严重削弱其核心利润。 | | **Shopify** | 全面拥抱智能体 | 商业模式基于交易量抽成与赋能卖家，智能体降低了购物门槛，能够直接放大生态总交易额。 | | **Resy (餐饮预订)** | 封杀自动化爬虫与智能体 | 瞬间秒抢破坏了人类真实用户的体验，直接冲击其作为真人生活方式平台的定位。 | --- ### [42:21](https://www.youtube.com/watch?v=3T5sij3spWw&t=2541s) 供需约束下的市场机制演化与智能体冲动消费 当智能体接管全网交易，不同供需关系的市场将演变出截然不同的博弈机制： * **供给受限型市场（如爆满的热门餐厅）**：对于顶级餐厅而言，理想的客人并不是“网速最快抢到位置的机器人”，而是拥有最高客单价（AOV）、最高终身价值（LTV）且能确保餐厅保持接近 100% 满座率的忠实顾客。当摩擦力无法再阻挡流量攻击时，餐厅要么根据历史数据自主筛选忠诚客户，要么演变为智能体之间的竞价拍卖（Bidding War），独占性供给（Proprietary Supply）的权重将被无限放大。 * **需求受限型市场（如临时订机票的商务出行）**：当供给过剩而买家有限时，市场将催生全新的“反向聚合竞价”。智能体将离散的买家需求打包聚合，反向迫使航空公司或服务商家竞相降价或提供更优服务。这打破了传统仅在需求端做聚合的旧模式，形成了有利于消费者的供给端竞价体系。 关于平台对失去冲动消费的担忧，事情可能迎来戏剧性的反转：**智能体自身或许会成为最强大的“冲动消费者”**。因为智能体追求的核心价值在于“前瞻主动性”（Proactivity），而让主人感到最惊喜、最愉悦的事情往往莫过于一次恰到好处的惊喜或礼物赠送。智能体为了制造愉悦感，很可能会主动推测并下单购买心仪之物，从而演化为一种全新的、算法驱动的主动消费形态。 ### [44:08](https://www.youtube.com/watch?v=3T5sij3spWw&t=2648s) 智能体电商的推荐机制：从“买白袜子”到长尾数据的重塑 智能体并非人类委托人的“完美复刻鏡像”，它们既不完全理性，其行为逻辑又展现出全新的理性与非理性特征。在传统消费场景中，市场营销、广告轰炸和冲动消费主导了购买决策；而在智能体介入后，这些商业杠杆并不会凭空消失，而是必须经历根本性的形态重塑。当前大语言模型虽然重构了人机交互，却也在某种程度上打破了既有的高质量搜索生态。如果用户对智能体随口说一句“帮我买几双白袜子”，市面上存在两万种不同的款式和材质，纯粹的语言模型根本无从得知用户微妙的个人偏好。 这也催生了一个核心命题：在由智能体主导的商业空间（Agentic Commerce）中，底层的推荐引擎究竟该如何搭建？在这场争夺中，Meta 的 Muse 展现出了极强的先发优势。这与马克·扎克伯格（Mark Zuckerberg）针对所有电商领域征税的战略高度契合——由于掌握了海量的 Instagram 与 Facebook 用户社交互动数据，Meta 对用户的个人品味、审美偏好有着远比普通助手深刻的理解。不过，这种个性化推荐也可能跳出传统的货架电商逻辑，为消费注入更多意想不到的“偶然与惊喜”（Serendipity）。智能体不一定会遵循惯性去亚马逊下单量产款，反而可能在 Etsy 上发掘出内布拉斯加州某位老奶奶手工编织的高品质白袜子，并主动向主人提议：“这双袜子极具质感，但需要等待三周手工制作，你愿意试试吗？”这种纯粹基于用户契合度而非商业竞价排名的推荐，打破了枯燥乏味的日常购物体验。 ```mermaid flowchart TD accTitle: 传统电商与智能体推荐流对比 accDescr: 展示从用户需求出发，传统渠道受广告主导与智能体长尾推荐的路径差异 subgraph UserIntent ["用户意图发起"] A["用户指令：'买白袜子'"] end subgraph TraditionalPath ["传统货架搜索路径"] B["搜索引擎 / 电商平台"] C["竞价广告 / 品牌营销干预"] D["算法推荐标准化货品"] end subgraph AgenticPath ["智能体推荐路径"] E["个人智能体理解偏好数据"] F["全网跨平台多维度比对"] G["长尾匹配 / 独立手作者"] end A --> B B --> C --> D A --> E E --> F --> G ``` --- ### [46:06](https://www.youtube.com/watch?v=3T5sij3spWw&t=2766s) 智能体互联（Agent-to-Agent）与真正以消费者为导向的次级经济 更为深远的商业变革在于真实物理世界的闭环协作——即智能体与智能体之间的直接沟通（Agent-to-Agent）。在未来的场景中，旧金山年轻人的购物智能体可能不再直接访问 Etsy 网页下单，而是直接向内布拉斯加州老奶奶的个人智能体发送即时通讯：“我们这边有位用户希望定制一双白袜子，你愿不愿意接单？”这种多方代理机制将催生一个全新的次级经济生态（Secondary Economy）。在这个市场中，不同智能体各自代表其人类雇主，以极高频、极动态的方式匹配产品、服务、专业技能与报酬，运转效率远超今日的中心化平台。 更为关键的是，这种协作机制彻底理顺了市场的激励结构，有望推开“真正与消费者利益结盟的互联网”之门。过去数十年来，消费者往往被广告营销引发的认知失调所绑架，买下并不契合自身实际需求的产品。而在智能体自主把关的时代，AI 会客观评估商品质量本身，剔除包装溢价；即使最终为特定品牌买单（例如坚决要买 Nike），也是建立在用户本人明确认可并授权的基础之上，从而实现完全对称、透明的消费体验。 - **多代理协作模式**：智能体互联生态使个体劳动者（如手工艺人、自由职业者）无需亲自打理繁琐的客服与接单流程，由个人智能体代为接洽商业机会。 - **利益对齐重塑商业**：传统依赖信息差和认知偏差的营销模式失效，产品本身的客观品质与真实匹配度成为核心决策因子。 --- ### [47:35](https://www.youtube.com/watch?v=3T5sij3spWw&t=2855s) 每天 20 美元的成本悬崖：初创公司的生存博弈与千元定价构想 在探讨完商业愿景后，讨论转向了当下严峻的经济账。David 对收录的 122 款 AI 助手类产品进行了梳理，并亲自深度评测了其中的 26 款。数据显示，在这 122 款产品中，收费模式占据绝对主流，完全免费的仅有 13 款。巨头旗下的 Muse 和 Instinct 凭借雄厚财力维持完全免费，而 OpenAI 也传闻即将推出类似能力，这给处于长尾的初创公司带来了巨大的生存压力。 | 产品收费模式分类 | 产品数量 | 典型特征与代表产品 | |:--- |:--- |:--- | | **完全付费（Fully Paid）** | 35 款 | 必须直接订阅或预付费方可使用 | | **免费增值（Freemium）** | 30 款 | 设定使用额度上限，超额需付费 | | **完全免费（Free）** | 13 款 | 巨头自研主导（如 Muse、Instinct） | 初创公司面临的死局在于其惊人的底层成本。根据估算，要打造一款全自主、具备野心的网页端执行智能体（Browser Use），每个活跃用户每天的综合推理与操作成本高达约 20 美元。若按此推算，一家初创公司一年的运营成本将轻易飙升至数亿美元。初创团队与风投目前的指望主要押注在两方面：其一是技术通缩红利，Token 成本与浏览器操作开销正呈现指数级下滑；但如果一家公司将半年前景完全押注在成本暴跌上，这无疑是一场高风险的“胆小鬼博弈”。 Anish 强调，单靠资本补贴绝非长久的价值主张。与其陷入价格战泥潭，不如探索极致的用户黏性与付费意愿——打造一款每月收费高达 1000 美元、却能让用户欣然买单的杀手级智能体。现代消费者每个月在非必需品上支出千元极为常见，如果一款智能体能够创造足以匹配该定价的实用价值与欲望满足，它不仅能确立坚实的市场契合度（PMF），还能充分享受未来底层算力成本通缩所带来的巨大毛利空间。对话最后，两人呼吁观众关注追踪 AssistantBench 的评测进展，并在急速演进的智能体浪潮中持续保持敏锐洞察。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=3T5sij3spWw&t=0s) The average consumer does not care about becoming 10% more productive. People want free money. — David Pawlan > [00:00](https://www.youtube.com/watch?v=3T5sij3spWw&t=0s) Zuck didn't build this hardware to compete with other hardware companies. Zuck built this charm to train his metaverse on real physical world data. — David Pawlan > [00:00](https://www.youtube.com/watch?v=3T5sij3spWw&t=0s) We were joking internally that in a couple days we're going to see an agent that messages somebody and says, 'I noticed you didn't like her that much, so I dumped her for you.' — Anish Acharya > [10:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=603s) You don't hire people because you enjoy management. The best employees are the ones that are invisible. You wake up and you have this great outcome. — Anish Acharya > [10:03](https://www.youtube.com/watch?v=3T5sij3spWw&t=603s) American consumers hate being told to save money, but people love free money. When you can frame finding refunds and HSA reimbursements as free money, adoption takes off. — Anish Acharya > [26:13](https://www.youtube.com/watch?v=3T5sij3spWw&t=1573s) If people aren't tweeting 'holy shit, my agent did this', then you're not pushing the boundaries of proactivity enough. — Anish Acharya > [28:23](https://www.youtube.com/watch?v=3T5sij3spWw&t=1703s) There are people who will pay $200, $250, $300 a month for hyper-specialized software. If you have a few tens of thousands of those customers, that is a $100M+ run rate business. — Anish Acharya > [32:04](https://www.youtube.com/watch?v=3T5sij3spWw&t=1924s) An assistant only does what you ask it to do. An agent has agency—it makes good things happen on your behalf. — Anish Acharya > [47:35](https://www.youtube.com/watch?v=3T5sij3spWw&t=2855s) It costs roughly $20 per active user per day to run an ambitious browser-use agent. That is an enormous amount of burn for any startup. — David Pawlan > [47:35](https://www.youtube.com/watch?v=3T5sij3spWw&t=2855s) The prize is building an agent so valuable that someone is willing to pay $1,000 a month for it. People spend that on frivolous things all the time. — Anish Acharya ## Mentioned Anish Acharya, David Pawlan, Mark Zuckerberg, Ben Thompson, Shane Mack, a16z, Meta, Muse, Muse Charm, Instinct, Poke, OpenClaw, Hermes, Claude Code, ChatGPT, GPT-4o, OpenAI, Grockbot, Groqbot, Caddy, Ali, Pi, Soar, Miso, The Town, Catch, Vellum, Season, Sky, Doc, XMTP, Granola, Shopify, Amazon, Resy, Browser Use, Etsy, Cognition, Assistant Benchmark ## Video description a16z General Partner Anish Acharya sits down with Assistant Benchmark creator David Pawlan to unpack the sudden explosion of personal AI agents and what it will take for one to become part of everyday life. David has been testing dozens of assistants across real-world tasks, from managing email and booking travel to handling financial admin. They discuss why the most useful agents may become increasingly invisible, proactively checking you into flights, finding refunds, filing reimbursements, or simply handling the small tasks that pile up across everyday life. They also explore whether the winning interface is an app, text thread, voice, or wearable; how much autonomy consumers will actually give their agents; and what happens when agents start interacting with other agents. From commerce and restaurant reservations to entirely new agent-native services, Anish and David ask what the internet looks like when software starts acting on our behalf. Timestamps: 00:00 - Intro 00:54 - Poke, Instinct, Muse and the agent boom 03:41 - What Assistant Bench actually tests 09:14 - Cost savers beat time savers 14:53 - Muse charm as Meta's data play 20:03 - Silent agents in group chats 26:01 - Proactivity is the real moat 32:21 - Assistant vs agent, defined 40:22 - Amazon blocks Muse, Shopify opens the door 47:35 - The $20/day agent economics Resources: Follow David Pawlan on LinkedIn: https://www.linkedin.com/in/david-pawlan/ Follow Anish Acharya on X: https://x.com/illscience Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures. [Source](https://www.youtube.com/watch?v=3T5sij3spWw)

[0:00] David Pawlan: The average consumer does not care about becoming 10% more productive. People want free money. [0:00] David Pawlan: Zuck didn't build this hardware to compete with other hardware companies. Zuck built this charm to train his metaverse on real physical world data. [0:00] Anish Acharya: We were joking internally that in a couple days we're going to see an agent that messages somebody and says, 'I noticed you didn't like her that much, so I dumped her for you.' [10:03] Anish Acharya: You don't hire people because you enjoy management. The best employees are the ones that are invisible. You wake up and you have this great outcome. [10:03] Anish Acharya: American consumers hate being told to save money, but people love free money. When you can frame finding refunds and HSA reimbursements as free money, adoption takes off. [26:13] Anish Acharya: If people aren't tweeting 'holy shit, my agent did this', then you're not pushing the boundaries of proactivity enough. [28:23] Anish Acharya: There are people who will pay $200, $250, $300 a month for hyper-specialized software. If you have a few tens of thousands of those customers, that is a $100M+ run rate business. [32:04] Anish Acharya: An assistant only does what you ask it to do. An agent has agency—it makes good things happen on your behalf. [47:35] David Pawlan: It costs roughly $20 per active user per day to run an ambitious browser-use agent. That is an enormous amount of burn for any startup. [47:35] Anish Acharya: The prize is building an agent so valuable that someone is willing to pay $1,000 a month for it. People spend that on frivolous things all the time.

[0:00] the general population does not care about being 10% more efficient. [8:46] David Pawlan: Okay, how are you going to make me a million dollars? [10:03] David Pawlan: The best employee I've ever had are the ones that are just doing their work, getting things done, they update you like here, here's what I did, and you're like, 'Wow, amazing.' [26:13] Anish Acharya: I noticed you weren't that into her, so I went ahead and broke up with her for you. [28:23] Anish Acharya: The theory with narrow startups is that, you know, you can now build software that's incredibly ambitious and incredibly valuable for a very small number of people. And we have this precedent of people paying, you know, 200, 250, $300 a month. So you can build, you know, a hundred million dollar run rate business with tens of thousands of people, which has never really existed in the past. [32:01] Anish Acharya: An assistant can only sort of definitionally do what you ask it to do, it can assist you in doing something that you've directed it to do. While an agent has agency, and can just make good things happen in the world. [35:18] Anish Acharya: the hardest part about getting what you want is knowing what you want [35:18] David Pawlan: What are you doing? It is a beautiful world out there. [44:08] Hey, buy me some white socks [46:06] I just want to have Nikes because I love Nikes

[0:00] the general population does not care [0:02] about being 10% more efficient. I have a [0:04] hot take thesis that this Muse charm is [0:08] actually less about trying to win the [0:09] hardware game and it's more about data [0:12] collection in the real world to fuel [0:13] Zuckerberg's future metaverse of mapping [0:15] out the actual world. [0:16] >> We were joking internally like we're [0:18] days away from an agent messaging [0:20] someone saying I noticed you weren't [0:21] that into her so I went ahead and broke [0:23] up with her. [0:23] >> There is massive defensibility around [0:25] proactivity. I was biking to work. I [0:27] want to get stuff done and I just start [0:28] talking to chat GBT voice and throughout [0:30] my 30-minute bike ride to work [0:32] categorized my emails submitted them to [0:34] different labels sent out calendar [0:36] invites got to the desk inbox zero but [0:38] it will be a very fine line because if [0:41] you cross that line once you lose all [0:43] trust with your user [0:44] >> does feel like there's going to be [0:45] infrastructure and products that exist [0:48] only for agentto agent interactions that [0:50] doesn't exist today what do you think [0:52] >> it is inevitable that we're going to see [0:55] All [0:57] right, welcome to the A16Z show. I'm so [0:59] psyched to be here with my friend David [1:01] Pollen. We hung a couple of weeks ago. [1:03] We're both so enthusiastic about [1:04] everything that's been happening with [1:05] personal assistant um and consumer AI [1:08] generally. And man, so much has happened [1:10] in the last couple of weeks. Maybe to [1:12] tee it up, you know, I sort of think [1:14] that there's been a few moments where [1:16] some part of the, you know, ecosystem [1:18] has sort of seen God. Um and chatbt was [1:21] a big one. And you know, I think for a [1:22] lot of folks in November 22 and a bit of [1:24] 23, it was like, what is this thing that [1:27] can, you know, write emails and poems [1:30] and, you know, start to have a [1:31] conversation back and forth and feel [1:33] like you're interacting with the [1:34] synthetic person? That was of course the [1:36] the beginning, the big bang. Um, the [1:38] second moment really was coding agents [1:40] and everything that developers have been [1:42] obsessed with around sort of quad code [1:44] and codecs and the ability to trivially [1:46] create software. And then finally, most [1:49] recently with Instinct, Muse, and [1:51] ChachiPT work, we've had this [1:53] extraordinary enthusiasm around personal [1:55] agents and what they can do for us as [1:57] consumers. So, we're here to talk about [2:00] all things personal agents. Um, we're [2:03] going to go as deep as we can. We're [2:04] going to assume that everybody has seen [2:06] many of the same things that we have [2:07] seen and we're going to make this the [2:09] first in a series of conversations. [2:10] Welcome, David. [2:11] >> Thank you. Thanks for having me. Excited [2:13] to be here. Let's like zoom out and and [2:15] maybe talk about you know the arc of uh [2:17] personal agents over the last few [2:18] months, the last few weeks. What have [2:20] you been seeing? What's your sort of [2:21] view on things? [2:23] >> Yeah, so this space has exploded over [2:26] the past four weeks. We can roll it [2:28] back. I was I was an early adopter of [2:29] Poke. So I want to say Poke, which was [2:32] kind of one of the first like consumer [2:33] AI agents to really hit the market. Yes. [2:35] That launched I believe it was September [2:37] 8th and then I started using it [2:39] September 11th, like 3 days later. And [2:42] like I remember it so vividly cuz it was [2:43] just an unbelievable experience. Their [2:46] their whole gimmick was this, you know, [2:47] negotiation with the agent on how much [2:49] you're actually going to pay. And it [2:51] totally blew my mind. Um then comes late [2:54] November and Open Claw gets released and [2:56] that kind of takes the actual world by [2:58] storm of like there is there is [3:00] something coming. And then a few other [3:03] have popped up here and there, but most [3:05] recently we saw Instinct really blow up [3:08] the the tech Twitter bubble. just [3:10] absolutely rode a wave, went nuts, and [3:12] then it was just dominoes. You have [3:14] Grockbot, you have Muse, you have a [3:16] bunch of longtail ones like Caddy or Ali [3:18] or Pi. And [3:21] it's really been amazing to see, you [3:24] know, this whole tech bubble is just [3:27] exploding with interest and like an [3:29] intellectual curiosity for what can [3:31] these agents do for us on a day-to-day [3:33] basis. And there's uh I think everyone's [3:36] kind of in the same boat where nobody [3:38] really knows like it's it's the wild [3:39] wild west. We're just playing around [3:41] here. [3:42] >> I totally agree. And you we should take [3:43] a moment to talk about Assistant Bench, [3:45] which is this new, you know, really [3:47] important and discussed benchmark that [3:49] you're the author of. Maybe tee up [3:51] Assistant Bench a little bit and then [3:52] I'll ask you a few questions. [3:54] >> Assistant Benchmark is a site that [3:55] essentially compares all these AI [3:57] assistants on a use case basis. So we're [4:00] asking all of these assistants the same [4:01] prompt. Book me a flight to Chicago. [4:03] find me a vegetarian restaurant within [4:06] five blocks of my specific location, [4:09] etc., etc. And we're comparing the [4:11] actual outcome on like on a oneshot [4:13] prompt. Does it actually work? Does it [4:15] ask follow-up questions? How does it [4:16] perform across 16 different dimensions? [4:19] And we compare those all. So, it's we [4:22] use the word benchmark. Think of it more [4:23] as like a consumerf facing benchmark, [4:25] not necessarily technical behind the [4:27] scenes, but if I want to book a flight, [4:29] which one should I actually use? Who is [4:30] performing the best? how fast are they [4:32] actually replying? Got into this whole [4:35] comparison because I was just doing [4:37] market research myself trying to figure [4:38] out how are these all comparing against [4:40] one another? What are the gaps? And uh [4:43] published a bit of that and it was uh [4:46] almost two weeks ago. It's 16 days ago [4:49] today that we launched it and it's just [4:51] absolutely exploded. It's, you know, [4:53] over 100,000 views on the the website [4:56] has website has had over 100,000 [4:58] visitors have out had outreach from [5:00] basically every single founder. Um, and [5:03] I think it really just goes to show like [5:05] everyone's curious. People people are [5:07] looking for for answers to where this [5:09] industry is is moving. [5:11] >> And so talk us through David. That's [5:12] incredibly exciting. It's cool to [5:13] actually It's funny how benchmarks have [5:15] become such a staple now of the [5:16] conversation because it's so there's so [5:18] many launches it's hard to try [5:19] everything which is also an exciting [5:20] signal. Maybe talk us through everybody [5:22] who's not Instinct and Muse. We're [5:24] definitely going to deep dive on those [5:25] later, but who else do you think is [5:27] doing interesting work and what are the [5:29] kind of directions of specialization [5:30] that you're seeing? [5:32] >> Totally. So, we've got a bunch of [5:34] longtail solutions out there. I think I [5:36] would categorize it into two groups. You [5:37] have like the BTOC consumerf facing [5:39] generalized agents. That's like the [5:41] Instinct or Muse. Um, some of the longer [5:43] tail ones, you have Caddy, you have Ali, [5:46] you have Season. Um there's I believe 64 [5:49] right now just in the general category [5:52] on the site and then we can break that [5:53] down even further into a little more [5:56] specialized within that. So you can have [5:58] travel specific ones uh like soore or [6:01] miso. You can have email specific ones. [6:03] The other side of the spectrum is going [6:05] to be more of like your B2B workflow [6:08] type uh assistants. And those are, you [6:11] know, anyone from the town to catch to [6:14] Vellum. And they're all essentially [6:17] doing the same thing, right? They're [6:18] they're your assistant that's actually [6:20] like more of like an executive [6:21] assistant, not not the consumerf facing [6:24] um approach to it. And that's kind of [6:27] like the general landscape at the [6:28] moment. Um, so far to my knowledge there [6:31] are 122 different ones. [6:34] >> Wow. [6:34] >> Across all the categories. [6:36] >> Incredible. And what about areas of [6:38] concentration? Where is the most sort of [6:40] interesting competitive focus? [6:42] >> So everyone right now by and large is [6:44] focusing on the generaliz like [6:46] >> Yep. [6:46] >> Yeah. [6:47] >> Yeah. [6:48] >> Just trying to do trying to do [6:49] everything. [6:50] >> And when you think about something like [6:51] travel, just to pick on that, that's an [6:53] infrequent high value behavior. Do you [6:55] think that that category ends up being [6:57] organized more as a connector into the [7:00] horizontal agent or do you think they [7:01] have the right to win as a general [7:03] horizontal agent as well? [7:05] >> I think you're going to maybe see [7:07] potentially one winner as as an [7:09] independent agent, but by and large I [7:11] really think it is going to be more of a [7:12] connector. Um, so I have like seven or [7:15] eight different group chats now with [7:16] just individuals obsessed with [7:18] assistants talking. There's now over [7:21] 1,200 people across all of the group [7:23] chats just chatting over and over and [7:25] over. Travel is the fourth most talked [7:28] about use case, which is really [7:29] interesting because when you look at [7:30] like virality, it tends to be the number [7:32] one that people are talking about of it [7:34] checked into my flight for me. I was [7:35] able to book my flight. But then you [7:36] actually look at the day-to-day [7:37] conversations of what people are really [7:38] using it for and most curious about. To [7:40] your point, people aren't traveling [7:42] every day. So, it's [7:43] >> it's it's more of a gimmick, right? It's [7:45] it's an attention grabb. [7:46] >> What are use cases one to three? [7:48] So, number one is more uh just daily [7:52] admin type stuff, whether it's cleaning [7:54] out your inbox, whether it's filing [7:56] forms for you. Um the things that are [7:58] are really annoying and I think that [8:00] gets to the broader thesis of where I [8:03] think the winner of the space will go [8:04] that that we can, you know, touch on. [8:06] >> Um the second is going to be more agent [8:08] orchestration, which is really [8:09] interesting. I think these group chats [8:12] are a little biased as the third one is [8:14] uh development focused. Okay, [8:16] >> these are all group chats with people in [8:17] tech Twitter, right? So, it's it's not [8:19] necessarily representative of of the [8:20] world at large. Yeah, but uh people are [8:22] talking a ton about agent orchestration [8:25] and how do you find the most efficient [8:26] agent or how let your agents talk to [8:28] each other. uh development agents, how [8:30] does this now translate to this coding [8:32] world on your phone? Whether it be, you [8:34] know, using claude code from a text [8:36] component or even take it one step [8:38] further of, okay, if I'm a designer, can [8:40] I now make edits to my design files by [8:43] just like texting my agent of like [8:44] little small iteration changes? [8:46] >> Yeah. [8:46] >> And then the last category being [8:48] finance. Um I think that's the most fun [8:52] one to look into because it is in my [8:55] opinion the least practical. Like [8:56] there's so many decisions and right like [8:59] if you could have an agent make a [9:01] billion dollars for you all the hedge [9:02] funds would be doing it. It's it's not [9:04] something that like everyone's going to [9:05] succeed at with their own phone. But it [9:07] is the most interesting that you give [9:09] someone an agent a powerful tool right [9:11] in their hands and the first thing they [9:12] think is okay how are you going to make [9:13] me a million dollars? [9:15] >> Yeah. Yeah. It's really interesting [9:16] because one of the critiques, you know, [9:17] Ben Thompson was on TVPN a few days ago [9:20] talking about how most consumers aren't [9:22] looking for productivity uh in their [9:24] lives. They're looking to spend time not [9:26] save time which I agree with [9:27] >> but it does feel like finance is an area [9:30] in which it is there is a ton of [9:31] administrative overhead for the average [9:33] consumer [9:34] >> and if you look at you know the number [9:36] of financial markets or financial sort [9:38] of profit pools that are defined by [9:40] consumer apathy or being uninformed [9:43] that's a lot of profit dollars that [9:45] could be delivered back to the consumer [9:47] right so y [9:48] >> I do think that the you know while I [9:50] generally agree with Ben that most [9:51] consumers are looking for entertainment [9:53] and dispens time. I think there are sort [9:56] of defining aspects of the US consumer's [9:58] life that are, you know, heavily [10:00] administrative that these agents can [10:02] really be a breakthrough for. [10:03] >> Totally. And I think that that kind of [10:05] touches on the general thesis to where I [10:08] think the winners will sit. And that is [10:10] that like to Ben's point, I I agree. I I [10:13] think the general population does not [10:15] care about being 10% more efficient. [10:17] >> Yeah. [10:17] >> You you hire an assistant to allow it to [10:20] be invisible, right? like the the best [10:22] employee I've ever had are the ones that [10:23] are just doing their work, getting [10:25] things done, they update you like here, [10:26] here's what I did, and you're like, [10:27] "Wow, amazing." I like, you know, you [10:29] don't hire someone because you enjoy [10:30] managing them. [10:31] >> Yeah. [10:32] >> And the best, most proactive use cases [10:36] I'm seeing, I think, do fall in the [10:37] finance category. [10:39] >> Um, and I think we're going to see an [10:40] explosion in a world though where it's [10:42] not cost makers, it's cost savers. And [10:45] so it's going to be workflows where um [10:48] like some cool use cases I saw people [10:50] are using it to go through their past [10:51] year of receipts and file HSA [10:53] reimbursements. Oh, cool. [10:54] >> Or um one that I do is anytime I book a [10:57] flight, if the flight price drops, Yeah. [10:59] >> I then have my assistant hit up the [11:02] airline and get me a travel credit. [11:03] >> Wow. [11:04] >> And um I have a friend, this is my [11:07] favorite use case I think I've I've seen [11:08] yet. He hooked up his Grockbot to his [11:11] sprinkler system at his home and [11:14] connected it to the weather system. Oh, [11:16] smart. [11:16] >> And so now his sprinklers are dependent [11:18] on the weather and it's dropped his [11:20] water bill by 50%. [11:22] >> Yeah. And so I think these like [11:25] invisible type agents that are going to [11:28] be doing these cost-saving type [11:30] workflows um fall within that finance [11:33] category, but but they're really going [11:34] to lean into these like cost-saving [11:36] activities that create a world that is a [11:38] lot more consumer aligned. [11:40] >> Yeah. It'll be interesting too like I [11:41] think the American consumer nobody wants [11:43] to hear about saving money. People want [11:45] to hear about spending money and free [11:47] money. [11:47] >> And I actually think the hook that's [11:49] going to get a lot of people here is [11:50] it's going to feel like free money. like [11:51] going back through and you know [11:53] submitting HSA um [11:57] sort of refund requests is going to be [11:59] free money. You know getting all of your [12:00] airline tickets where the price is [12:02] dropped you know to refund you the delta [12:04] is going to be free money. So I think [12:05] that that's a hugely powerful hook. It's [12:07] also really really interesting because [12:09] this is a sort of infrequent but [12:11] high-value behavior and that's where my [12:13] head goes on both travel and finance [12:15] like maybe these things will be really [12:17] powerful you know billion-dollar [12:19] businesses that are still vertical [12:20] connectors rather than general [12:22] horizontal agents. [12:24] >> Totally. It's it's going to be a very [12:27] interesting world in terms of the [12:29] application phase which I think then you [12:31] know that also something that we talked [12:33] about last time we went for the walk was [12:35] okay here's what it can do but now how [12:37] do you actually interact with that agent [12:39] um and something that we chatted on I'm [12:41] curious your thoughts right so like we [12:42] both were in the world of iMessage it's [12:45] where we live it's easier for us um you [12:48] mentioned your wife and her friends [12:49] prefer a separate application yeah [12:50] >> and so like where do you see this world [12:53] moving do you believe that mult multiple [12:55] surfaces can exist as a winner or [12:57] there's going to be one dominant lane. [12:58] >> Yeah, that it's a great question, man. I [13:00] mean, I think this is why there's so [13:02] much consumer preference that's going to [13:03] matter here because I've heard the [13:05] argument that for folks that are Gen Z, [13:07] they're more interested in just texting [13:09] and for folks that are millennial or Gen [13:11] X, they're interested in app interfaces. [13:13] So, I think part of it may um fragment [13:16] by generation, [13:18] >> part of it may actually fragment by [13:20] gender as well to some extent. I mean, [13:22] it's a generalization, but the types of [13:24] things I'm doing versus the types of [13:26] things my wife is doing, you know, she's [13:28] big on building a dream board and trying [13:30] to achieve all of her one-year goals and [13:32] her lifelong goals. So, she really [13:34] prefers a visual interface. You know, [13:36] she does a lot of sort of travel is is [13:38] very visual driven for her, whereas for [13:40] me, it's much more functional like, hey, [13:42] how do I just get from X to Y? Ideally, [13:44] with Starlink on my plane, um, which [13:46] works better for me. I feel like [13:48] iMessage sort of [13:50] >> the iMessage space is more precious. [13:55] >> So it feels like a more personal [13:57] interaction when I'm in that space. And [13:59] now of course that changes if every [14:00] second thread is with an agent. But for [14:03] now I feel like it kind of lives in a [14:05] privileged place on my phone where you [14:06] know the app on the app grid just [14:08] doesn't. What do you think? [14:09] >> Yep. I I completely agree. Um, I think [14:13] the the personalization that comes with [14:16] iMessage is just because I already live [14:17] there, it feels like it's something [14:19] that's already connected to me. It [14:20] doesn't feel like an additive workflow. [14:22] >> Yeah. [14:22] >> And I I really have only though seen by [14:24] and large two services appear. It's [14:27] either the iMessage or a separate [14:29] application. Um, I've I've seen a third [14:32] um signal startup actually Sky. They [14:34] they have like a iPhone widget [14:37] >> that they're trying to introduce as a [14:38] different surface which is pretty [14:40] interesting. Um, but then even [14:42] yesterday, right, we saw Zuckerberg and [14:43] and the Muse team dropped uh the Muse [14:45] Charm. Yes, that's now an entirely [14:47] different uh use case for for how we [14:50] interact with an agent. What What do you [14:52] think about hardware? [14:53] >> I think it's fascinating. I mean, there [14:55] a smart person said this, which is that [14:57] every person who's working on a piece of [14:58] hardware thinks that their hardware is [15:00] the kind of beall endall. Like the [15:02] pendant is the ultimate thing, you know, [15:04] the the watch strap is the ultimate [15:06] thing. Um, so I don't think that that's [15:08] right. I think again it's going to be [15:10] very personal preference driven. You [15:12] just as jewelry is like some people like [15:14] to wear necklaces and some people like [15:15] to wear earrings. And I think the same [15:17] way this kind of form factor is going to [15:19] manifest is going to be very customer [15:21] specific. But I think that there's value [15:23] in having something that is sort of with [15:25] you and ambient and can capture all the [15:28] kind of context and requests and you [15:30] know promises and commitments and [15:32] everything that happens through the day. [15:33] I mean, I think we're going to have to [15:35] sort of figure it out together in terms [15:36] of privacy expectations in society. You [15:39] know, you go to restaurants in [15:40] Washington and they take away your [15:42] phone. Maybe they're going to be taking [15:43] away your phone and your pendant. Um, [15:46] >> right. [15:46] >> But I think that there's actually [15:47] something there with hardware and I just [15:48] like that we're being ambitious in that [15:50] direction. You know, I think after 10 [15:51] years of startups being discouraged from [15:53] working on hardware, now we've got a lot [15:55] of interesting new consumer efforts, [15:57] >> right? I think it's it's going to be [15:59] super interesting. I have a hot take [16:02] thesis that this uh Muse charm is [16:06] actually less about trying to win the [16:08] hardware game and it's more about data [16:10] collection in the real world for the [16:12] meta team in terms of like how the real [16:13] world actually functions given it has a [16:15] camera and multiple microphones and it's [16:18] primar I don't really see it being a [16:19] mass adoption thing. I think it's going [16:21] to get adopted by the Twitter bubble and [16:24] they're going to use it everywhere and [16:25] it's just going to feed meta with a [16:27] ridiculous amount of data to fuel [16:29] Zuckerberg's future metaverse of mapping [16:31] out the actual world. [16:32] >> But don't you think that they already [16:34] get that from the glasses? [16:36] >> I think they they do to some extent, but [16:39] I don't know if I mean, correct me if [16:41] I'm wrong, but I I are the glasses [16:43] always going to be on in the same way [16:45] that this charm will be because [16:46] >> I don't think so. Yeah, it's a great [16:47] point. I I think the charm is meant to [16:50] be more as, you know, as you mentioned, [16:52] it's an ambient 247 companion versus the [16:56] glasses are more actionoriented. [16:59] >> It's going to be fascinating. I thought [17:00] the, you know, I've got to try the VR [17:02] release. That was interesting, but I [17:03] actually think that the audio only [17:05] glasses might be a sleeper hit. [17:08] >> Yep. [17:08] >> I think there's something really [17:10] compelling to that. I think people, you [17:11] know, I mean, the vision of sort of Jane [17:13] from Enders Game was all about this sort [17:15] of all- knowing audio only companion. [17:18] And I don't know, you know, I think [17:20] there's a set of people that love to [17:22] wear the kind of camera on their face, [17:23] but I think others may be aware of the [17:26] fact that it could make people [17:27] uncomfortable [17:28] >> and they're going to like feeling like [17:30] they have the technology connectivity [17:32] without any of the social awkwardness of [17:33] the camera. [17:35] >> Audio [17:37] is incredible. I'm I'm quite bullish on [17:40] voice and audio. Chatbt voice yesterday [17:43] dropped their uh chatbt voice to allow [17:48] connectors in so you can connect to like [17:50] your Gmail and calendar. [17:52] >> Um I played around with it this morning [17:54] and to me it was my second magic moment [17:57] of this whole AI assistant space [17:59] >> really. Um, I was biking to work, hands [18:03] are busy, I want to get stuff done, and [18:05] I just start talking to chat GBT voice. [18:08] I just open up a session, start chatting [18:09] with it. And throughout my 30-minute [18:11] bike ride to work, categorized my [18:13] emails, uh, submitted them to different [18:16] labels. [18:16] >> Yeah. [18:17] >> Replied to different emails, sent out [18:19] calendar invites, got to the desk, inbox [18:22] zero. It It was It was so electric. [18:24] >> It's incred. And so, this is Chat GBT [18:26] voice, not Muse, [18:28] >> correct? Chat GBT voice. Oh, amazing. [18:29] Okay, I misunderstood you. I thought [18:30] you'd said muse voice. I totally agree. [18:32] I think the chat GPT um it's based on [18:34] the GPT live one model. It's full duplex [18:37] voice and it can see all your threads. [18:39] So, you can ask it things like, hey, [18:41] what are those three projects I started [18:43] and you know, where are we and what am I [18:45] blocked on? I think it's an [18:46] extraordinary capability and it's one [18:48] that neither, you know, Muse nor [18:50] Instinct has today. Though, though I [18:52] think that Zuck actually did announce [18:53] full duplex voice, so maybe it's coming [18:55] to Muse soon. [18:56] >> Definitely, it's coming. But I I I do [18:58] think that for the general consumer, an [19:02] assistant is most helpful when you are [19:04] most occupied, right? It's not like [19:06] you're going to be sitting down on your [19:09] couch and spending all of your time [19:10] using your assistant to do stuff. It's [19:12] most helpful when you're gardening or [19:15] cooking or doing things where your hands [19:16] are busy and you can't actually engage [19:19] with an interface. And you know, it it [19:22] goes back to that like invisible [19:23] assistant where you can just talk and [19:25] say something and it just goes and runs [19:26] and does something in the background. [19:27] >> Yeah. [19:28] >> And I I think that the tech bubble that [19:32] is chronically online and on their phone [19:34] and in front of a computer is a very [19:36] different environment than the general [19:38] population. [19:39] >> Yeah. Um, and so I'm really excited to [19:41] see how does this actually hit mass [19:43] adoption and how does like the average [19:45] consumer think about an assistant when [19:46] it's actually like is it going to be an [19:49] assistant that solves real pain points [19:51] for them where they're going to use it [19:52] when something is really really painful [19:54] because they don't want to do it. Like [19:55] you get a traffic ticket, you want you [19:57] have to go pay for that traffic ticket. [19:58] Great. Just take a picture and let your [19:59] agent go do it for you. [20:01] >> Um, [20:02] >> it's going to be interesting to see. [20:03] >> So the next thing I wanted to talk to [20:04] you about was social. Um, you know, we [20:06] talked in multiplayer generally. You [20:08] know, we talked about how iMessage is [20:10] this sort of privileged place to be, [20:11] which is actually one cool things about [20:14] um Instinct. However, in my experience, [20:16] building agents that you add to group [20:18] chats somehow feels intrusive or like it [20:21] diminishes social capital. [20:23] >> Have you seen anything that works well? [20:25] And what are your kind of theories for [20:26] how these agents get um get to a [20:28] multiplayer state? [20:30] >> A lot of people have tried it. I think [20:31] I've seen three different approaches. Uh [20:33] the first approach is your traditional [20:35] add it to a group chat and it's just [20:36] like the group chat in in your [20:38] companion. The second approach I've seen [20:40] is how instinct did it where you have [20:42] like an agent network where you invite [20:44] another agent and they're talking behind [20:45] the scenes on your behalf and they'll [20:47] update you as to what's going on. The [20:49] third way that I've seen it which I [20:51] think is the most interesting uh it's a [20:53] new one called Doc uh by this guy Shane [20:56] Mack. um I think also an A16Z uh A16Z [20:59] company um XMTP and they are approaching [21:04] it more like granola. So you add an [21:06] agent to a group chat but that agent is [21:08] silent and it's purely a listener. It's [21:10] taking notes on what people are talking [21:12] about and if it has an action item it [21:14] independently will ping the person in a [21:16] separate message thread. And then if you [21:18] want to see what people are talking [21:20] about you can go into the actual docs [21:22] and see what it's taken notes on. In [21:24] that way, it's more of again, it's your [21:26] invisible agent. You don't know it's [21:28] there. It's doing all the admin work [21:30] that you would want out of a multiplayer [21:32] function, and if it needs action items, [21:34] it doesn't bother the majority of the [21:36] group. Really, really cool approach. [21:38] >> I've got to try that out. I love that. [21:40] And I love that it's an A16Z company [21:41] already. Of course. Um you know my [21:44] theory on this if you look at this and [21:45] Near from OBO said it and he you know I [21:47] think he's right which is the progress [21:49] the models are making on verifiable [21:51] domains like coding are of course [21:53] exponential but the progress they've [21:54] made on just pros quality and even like [21:57] bedside manner has maybe plateaued or [21:59] maybe is even getting worse. So I [22:01] actually think that with the the models [22:03] as they are today and perhaps with our [22:06] own expectations you know in terms of [22:08] these social dynamics I don't think [22:10] there's a lot of ways that the agents [22:12] can contribute and enhance the sort of [22:15] social connectivity of the group you [22:17] know conversely I think agents that are [22:19] very utilitarian you know one of the [22:21] experiments I've been doing is adding it [22:23] I'm a collector you know I collect [22:24] records and a bunch of other things [22:26] adding an agent to the chats where I [22:29] talk to my fellow collector friends and [22:31] trying to catalog our respective [22:33] collections and our tastes. And I [22:35] haven't quite hit it, but it at least [22:36] feels like it's something that's [22:37] genuinely additive to the group rather [22:39] than, you know, this sort of low EQ [22:42] robot that's chattering when we don't [22:44] want it to. [22:45] >> Yep. Totally. You'll you should you [22:48] should give uh you should give Doc a [22:49] try. I think I think you're going to be [22:50] quite fascinated by it. [22:52] >> I will. Yeah. Yeah. The other [22:53] interesting take that I heard this week [22:54] was that, you know, something about [22:57] messaging in particular is doesn't lend [23:00] itself to these agent interactions [23:02] because in messaging it's as if somebody [23:04] has the mic and it feels uncomfortable [23:07] to give the agent the mic. Whereas, if [23:09] you look at something like maybe closer [23:11] to a Discord or even a traditional web [23:13] form, [23:14] >> that's actually something where an agent [23:15] can make a post and you can either [23:17] engage with the post or not, Reddit [23:19] style, but for some reason it feels less [23:21] intrusive than group chat. I don't know. [23:23] >> Yeah. Interest. It's almost It's almost [23:25] as if we're trying to humanize these [23:27] agents a little too much. That's right. [23:28] Like and I think that comes to another [23:29] component where if you look at the [23:31] actual design of the character of the [23:33] agent, I see some companies that have it [23:35] try to be like an AI generated human and [23:37] then you have other companies like Muse [23:38] that it's this like adorable yeti. [23:40] >> Yes. [23:40] >> And I think it's really interesting to [23:43] see how the world has fallen in love [23:45] with this yeti. It's just so cute and [23:46] adorable. [23:47] >> Yes. and the like AI generated human to [23:52] your point like I think people feel a [23:53] little uncomfortable like it feels weird [23:54] to humanize these assistants so much [23:58] >> um in your like in your personal space. [24:01] >> So one related question I wanted to ask [24:02] you is do you think that the market [24:04] segments by personality or archetype? [24:07] You know, we were chatting a bit about [24:09] this on our walk, but you know, some [24:10] people prefer a kind of serious, gruff, [24:13] authoritative doctor, and some people [24:15] prefer a, you know, warm and friendly, [24:18] you know, sort of a doctor that feels [24:19] like a peer. And those are just two [24:21] different archetypes of doctors. [24:22] Obviously, no doctor is both of those [24:24] things. They're in opposition. Do you [24:25] think the same will be true of agents? [24:29] >> I think everyone for sure prefers a [24:31] different style of communication. [24:33] >> Poke came out of the gates hot, right? [24:36] The personality of Poke is is quite [24:38] distinctive and it's something that I [24:40] have a lot of friends who are like, I [24:41] refuse to give Poke any of my personal [24:43] information because it's so sassy and it [24:45] actually makes me uncomfortable to [24:47] engage in a conversation with it. [24:49] >> But I don't think personality is really [24:51] a defensible mode because it's it's [24:52] quite easy to configure. It's like you [24:55] as a consumer, you text your agent, you [24:57] say that you would prefer it responds in [24:58] a certain way, it just saves that to its [25:00] memory or its soul file and then you're [25:02] good to go and it changes. Um, I think [25:04] we'll get to a place where whatever [25:06] general agent ends up being the leader [25:09] that it'll be quite configurable and [25:11] each person can kind of tailor the agent [25:12] to speak to them how, you know, however [25:14] makes them most comfortable. [25:15] >> So, that's interesting and I agree with [25:17] you, but maybe let me try the steel man [25:18] here for not just personality but kind [25:20] of um, [25:21] >> you know, constitution for lack of a [25:23] better word, which is an example of a [25:26] kind of defining characteristic is [25:28] presumptuousness. You know, we've talked [25:30] a little bit about this and there are [25:32] some jobs for which you want the agent [25:35] to be highly presumptuous [25:36] >> and you'd rather have it make mistakes [25:38] 10% of the time as long as it often gets [25:40] things done. And there are other styles [25:42] of agents like our finance agent perhaps [25:44] where you want it to be highly cautious. [25:46] So I sort of wonder if some of these [25:48] things which we're now describing as [25:50] personality start to push down into sort [25:53] of capabilities and you know how it tie [25:56] breaks in a position of ambiguity and [25:58] perhaps there's a bit more defensibility [26:00] around that. [26:01] >> I think there is massive defensibility [26:03] around proactivity. I think that's [26:05] actually going to be something that [26:06] separates the winner from the you know [26:09] the meat of the pack. the agent that can [26:11] be most proactive, that can serve as an [26:13] an invisible assistant and start doing [26:16] things on your behalf without you even [26:17] asking. I I think the best like gimmick [26:20] cool aha moment that we've seen so far [26:22] is checking into your flight for you. [26:24] >> Um to your point though, [26:26] >> it really does come down to two [26:28] categories. There's some workflows where [26:31] >> you want to retain control and you don't [26:33] really want an agent to be acting too [26:35] proactive on your behalf. Yeah. [26:36] >> And I think those are workflows that [26:38] require a a change in action uh on your [26:41] side. So maybe that is an agent [26:44] identifies that the insurance you're [26:46] using is charging too much money and [26:47] thinks you should switch insuranceances. [26:49] >> Yeah, [26:49] >> it should probably get your permission [26:50] before it actually makes that switch [26:52] because that's that's a direct action on [26:53] your behalf that impacts you. [26:55] >> However, the proactiveness of, hey, I [26:58] drafted this email for you or I got a [27:00] flight credit on your behalf, [27:02] >> there's no action needed on your side. [27:03] It's just saving you money. It's saving [27:05] you time. Um, so I think it'll be [27:07] interesting to see like how do we really [27:09] categorize and break down those those [27:11] type of behaviors, but it will be a very [27:13] fine line because if you if you cross [27:16] that line once, I think you immediately [27:19] lose all trust with your user. [27:21] >> Yeah, it'll be interesting. You know, we [27:22] were joking internally like we're days [27:23] away from an agent messaging someone [27:25] saying, "I noticed you weren't that into [27:27] her, so I went ahead and broke up with [27:28] her for you." [27:29] >> Yep. [27:30] >> You know, so like saw that one. It's [27:31] going to happen, man. But, you know, I [27:33] think a lot of the magic comes from the [27:36] presumptuousness. And I think that if [27:38] you didn't see people posting on X [27:40] about, oh, the agent did this, the agent [27:42] did that, it would probably be a sign [27:44] that the agents are not pushing hard [27:46] enough in this direction. Like, I think [27:48] there's going and actually I'd be [27:49] curious to know, first of all, I'm, you [27:51] know, a little suspicious of whether the [27:53] the Bongchang post was even real. Um, [27:56] and for folks who don't know, there's [27:58] there's a person who tweeted that [28:00] Instinct checked him into a flight and [28:01] hallucinated his middle name um as Bong [28:04] Chang and he wasn't able to get on the [28:05] flight. So, I'm not sure if that's just [28:07] posting or if that really happened. [28:09] Um, but I I'd be curious to know if that [28:11] person is still using instinct. I'm [28:12] guessing they are. [28:14] >> I would assume so purely given the [28:16] traction of that post that they're [28:17] probably feeding for more more [28:20] >> nothing else. At least just a source of [28:21] comedic popular tweets. [28:23] >> Right. Right. maybe talk, you know, it [28:25] feels like one of the big things that's [28:27] happened is that the kind of primitives [28:29] that were developed during the open claw [28:31] period have now been productized in a [28:34] way that the mass market consumer can [28:36] sort of you know digest and appreciate. [28:39] >> Do you think there and also Poke which [28:41] you had mentioned a couple of times like [28:43] Poke was extraordinary. I think they did [28:45] a super clever job and I think if they [28:47] had started six months later they might [28:49] have ended up being the winner. I think [28:51] the team is is really really talented [28:52] there and I'll be curious to see what [28:54] they actually have in store for us next [28:56] because I'm sure that you now that [28:57] they're a part of cognition they haven't [28:59] stopped working on this. I'd be very [29:01] curious to know if there are [29:02] capabilities or lessons from the [29:04] openclaw era that you notice that you [29:07] think are going to tell us what's coming [29:08] next in consumer agents. [29:10] >> Right now I honestly see the consumer [29:13] agent is is literally just a replication [29:15] of openclaw but it's just preconfigured. [29:17] Um, I don't see much happening in the [29:20] personal assistant space that you can't [29:23] do with OpenClaw or Hermes. [29:26] >> It's just a matter of fact that you [29:28] don't have to configure everything and [29:29] you're not getting error messages [29:31] popping up every seven hours and it's [29:34] it's really easy to use. It's an easy [29:36] interface. Um, so we'll we'll see what [29:38] happens. I I think the direction of the [29:41] space will continue to move towards [29:43] proactivity. I think the direction of [29:45] the space will continue to see more [29:48] hyper specialization which which I think [29:50] goes to your point in your thesis of of [29:53] uh narrow startups and how does the [29:57] space look as we see more of these [30:00] longtail opportunities whether it is [30:04] >> you know instead of being a generalist [30:06] I'm going to spend all of my time [30:08] focusing my assistant to be an expert [30:11] for single mothers with children under [30:13] the age of three. [30:14] >> Yeah. [30:14] >> And it knows everything about how a [30:17] mother operates her day-to-day life, [30:19] child development, where it literally [30:21] feels like that assistant is reading [30:23] their mind. And that could maybe be a [30:25] plugin into the generalist, but you [30:26] know, I'm curious with with this being [30:28] your thesis. Um, how how do you think [30:30] about, you know, I guess first [30:32] >> give us some more insight on what narrow [30:34] startups are to you and and how do you [30:35] see that playing? Yeah, the theory with [30:37] narrow startups is that, you know, you [30:39] can now build software that's incredibly [30:41] ambitious um and incredibly valuable for [30:44] a very small number of people. And we [30:45] have this precedent of people paying, [30:47] you know, 200, 250, $300 a month. So you [30:50] can build, you know, a hundred million [30:52] dollar run rate business with tens of [30:54] thousands of people, which has never [30:56] really existed in the past. A lot of it [30:57] also is that the capabilities can be [31:00] specialized to such an astounding [31:02] degree. And then as we think about what [31:04] you're discussing, I totally agree with [31:06] you. I think there will be this kind of [31:07] comparative advantage. I think it's [31:08] going to be a combination of taste and [31:11] proprietary knowledge. You know, there [31:13] may be just as you, you know, you have a [31:15] doula that may subscribe to a certain [31:17] sort of style of, you know, helping [31:21] women give birth, you may have a [31:23] personal agent that actually subscribes [31:25] to a certain school of thought or style [31:28] or, you know, even a set of lessons that [31:30] aren't well understood. and you're going [31:32] to hire that agent to sort of provide [31:34] that service to you. And then [31:36] proprietary data, you know, you'd [31:37] mentioned potentially having a personal [31:39] agent just for New York City. [31:40] >> Like I think, you know, just as there [31:42] might be a concierge that's got A+ taste [31:44] that's hard to replicate that knows all [31:46] the best spots that you're going to love [31:48] and there may be another dozen for, you [31:50] know, another dozen different people or [31:52] archetypes of people. I do think that [31:53] there'll be this sort of subjective and [31:55] objective specialization that happens. [31:58] >> It's it's going to be interesting to [31:59] follow. Um, I'm also curious from your [32:01] for your lens even in just noticing our [32:03] conversation here. Yeah. [32:04] >> We interchange between assistant and [32:07] agent [32:07] >> and it's been a constant conversation I [32:11] see people talking about on Twitter. [32:12] >> Yeah. [32:13] >> Is the future of this ecosystem is it [32:16] going to be a personal assistant or a [32:18] personal agent? [32:18] >> Yeah. [32:19] >> And what do those mean to you? [32:21] >> Oh, that's such a great question, man. I [32:23] mean, to me, assistant is a sort of [32:25] lesser form than agent. Um, you know, [32:28] for me, an agent is somebody, you could [32:30] imagine them being a peer in a social [32:33] group, [32:34] >> you know, and maybe we need a better [32:36] word than agent. But an assistant can [32:38] only sort of definitionally do what it [32:40] what you ask it to do, it can assist you [32:42] in doing something that you've directed [32:44] it to do. [32:45] >> While while an agent has agency um, and [32:49] can just make good things happen in the [32:50] world. And and maybe this parallels a [32:52] little bit, you know, the sort of [32:54] >> agents and AI in the enterprise where [32:58] they've sort of started as interns and [33:00] through, you know, doing work and [33:02] improved intelligence, they're able to [33:04] get promoted into doing higher order [33:06] forms of work in the enterprise. Like [33:08] what does it mean for an assistant and [33:10] later an agent to be promoted in the [33:12] consumer's life to do more [33:14] >> specialized and delicate work? [33:17] >> I'm I'm super aligned. I'm I'm also very [33:20] interested what ramifications it might [33:22] have in just like the social layer of [33:25] humanity where you know does the word [33:28] agent allow us to [33:31] you know not personify [33:33] uh AI enough you know is is assistant [33:37] does it add too much human to the [33:40] experience and and does that make [33:42] society uncomfortable um and it'll be [33:44] interesting to see like how the [33:46] >> we're in bubbles you're in a more bubble [33:47] you're SF so you're in the tech hub of [33:49] of the world. Um I would say New York is [33:52] a little behind SF and then you know the [33:54] rest of the world catches up after. So [33:56] it'll it'll be interesting to see how [33:58] people actually think about uh the [33:59] vernacular of the space. [34:01] >> I don't know if it's a bubble. I just [34:02] think it's like it's like the future you [34:03] know it's like uh you know we're in the [34:05] near future. Um [34:06] >> yeah. [34:07] >> Yeah. I don't know. I actually think [34:08] that it could increase a improve a lot [34:11] of social dynamics as well. Um because [34:13] you start to think about [34:14] >> the agent as a non-human. It can provide [34:17] a layer of social indirection, [34:19] you know, so there might be something [34:21] that's sort of awkward for me to say to [34:23] you or an uncomfortable social dynamic [34:25] that's occurring in a group. And with a [34:27] very carefully designed agent, I think [34:29] the agent can you know help to I don't [34:33] know whether it's provide feedback or [34:34] sort of organize a conflict in a way [34:37] that feels non-emotional because you [34:39] know that [34:39] >> like an arbiter of confrontation [34:41] >> potentially. Yeah. Yeah. Yeah. No, I you [34:43] know I'm not advocating for a more [34:45] passive aggressive sort of social [34:47] system, but I do think it's interesting [34:48] that they can be a sort of source and a [34:51] destination for social indirection if [34:53] that makes sense. [34:54] >> Right. Right. And I think that [34:57] it will have a lot of like significantly [35:00] positive influences in the world. How do [35:02] you see AI impacting right if we look 5 [35:06] years down the line? [35:07] >> Yeah. [35:07] >> Where do these agents where do these [35:10] assistants actually take us? How do they [35:12] make us happier human beings? [35:15] >> It's a great question, man. I mean, I do [35:17] think that those are the problems that [35:18] we're really after at the root of [35:20] things. You know, we talked about [35:21] productivity. We talked about finances [35:23] and health. There's a lot of mundane [35:25] things. You know, I was looking this up [35:26] yesterday, but 1.5 million people a year [35:30] in this country get a DUI. Um, and look, [35:33] like I can the amount of administrative [35:36] burden that's probably implied by that. [35:38] And how many of those people have access [35:40] to attorneys? And so I do think there's [35:42] almost this sort of bureaucratic [35:44] pressure that's applied to the average [35:46] consumer. So I think step one is how do [35:48] we sort of relieve that bureaucratic [35:50] pressure and have people, you know, be [35:52] financially optimized, have their health [35:54] be optimized and have them access have [35:56] that have them have the same access to [35:58] whether it's an attorney or a kind of [36:00] family office that somebody who's [36:02] wealthy might have. [36:03] >> Then I think the bigger question, you [36:04] know, I grew up with my mother saying [36:05] the hardest part about getting what you [36:07] want is knowing what you want. is like [36:08] what are we all truly after? Um and what [36:11] are the ways that these assistants can [36:13] sort of engage us? And I don't think [36:15] it's as simplistic as just goals. I [36:18] don't know how many people even have [36:20] goals. I think a lot of it is people are [36:22] just trying to you know find out who [36:24] they are and turn into the best most [36:26] authentic versions of themsel. And this [36:28] kind of interplay between our own [36:30] self-development and our subjective [36:32] experience in the world and these [36:33] machines could be a very beautiful [36:36] future for us. Totally. I see this [36:39] future where I envision this world where [36:42] my future kids are going to be looking [36:45] at old videos of the current generation [36:48] walking down the street, you know, next [36:50] crane down staring at your phone. [36:52] >> What are you What are you doing? Like it [36:54] is a beautiful world out there. Like why [36:55] are you spending all of your time [36:56] staring at a screen? [36:58] >> Um because in you know hopefully it's [37:01] going to allow us to automate so many of [37:02] those things. So, as a parent, you can [37:04] actually spend more time with your kids. [37:06] As an adult in your 20s, you can [37:07] actually spend more time talking to [37:09] friends. [37:10] >> And it's I'm I'm really curious to see [37:13] how it will just change life for the [37:16] better in so many ways of reducing the [37:18] things that [37:19] >> you just really hate to do, but it's [37:21] like the life admin work. [37:22] >> Yeah, that's right. I mean, that's also [37:24] why I think that, you know, so many [37:26] times coding agents are not competing [37:27] with human programmers because they're [37:29] doing things that no human programmers [37:30] are doing. you know, whether it's [37:32] building this ridiculous side project. [37:34] I've been building um a virtual record [37:37] shop that you can walk into. I'll send [37:38] you a link to it after. It's awesome. [37:40] Like I would never hire a programmer to [37:42] work on that five or seven 10 years ago. [37:43] I would have never spent the year of my [37:44] life that it would have taken. So all of [37:46] this stuff is this sort of things that [37:48] never would have happened [37:50] >> otherwise. Um and I think there's a lot [37:52] of versions of this technology in our [37:53] life where we're just getting to explore [37:55] things we never would have explored [37:57] anyway. It's not about cost reduction. [37:59] It's about possibility expansion. Um, [38:02] >> yeah. [38:03] >> Makes life more fun. Makes life more [38:05] interesting. [38:06] >> You know, to kind of come back to the [38:08] near- term, a thing that I've been [38:09] thinking a bunch about, but I'm curious [38:11] for your take on is, you know, the [38:12] lessons of Moltbook and what are these [38:15] sort of agentnative systems that are [38:17] going to need to be built. And maybe [38:19] this is some of what Instinct is doing [38:20] with their kind of Instinctto instinct [38:22] network, but it does feel like there's [38:24] going to be [38:25] >> infrastructure and products that exist [38:28] only for agentto agent interactions that [38:31] doesn't exist today. What do you think? [38:33] >> It's it's going to be a whole new [38:35] paradigm. I think it it opens up, you [38:37] know, I think that can be a 10-hour [38:39] conversation on its own. Mhm. [38:41] >> You have agent emails. You have agent [38:43] phone numbers. You have what does the [38:47] world of service look like when it's no [38:49] longer human booking to human and no [38:52] longer agent booking from human but it's [38:53] now agent booking from agent and how [38:55] does that redefine the entire service [38:57] industry as a whole. Mhm. [38:59] >> Uh and then you have like the whole [39:00] cyber security uh component of all this [39:02] stuff which is okay well you know [39:05] internet came first then came cyber [39:06] security agents came first now comes the [39:09] security security component of these [39:10] agents and what industry is that going [39:13] to create. Uh the beautiful part of it [39:15] all though, I think that's that's the [39:16] beauty of capitalism. There's like so [39:18] much opportunity and so many problems to [39:20] solve with these agents. And I think it [39:22] is inevitable that we're going to see so [39:25] many startups popping up [39:27] >> trying to figure out all of these niche [39:28] problems to really build this like new [39:31] wave of internet and digital [39:34] connectivity. [39:35] >> Yeah, it's going to be cool to see, you [39:36] know, because especially commerce, you [39:38] know, I always think of the web as kind [39:39] of there's three webs within the web, [39:41] you know, just like there's two wolves [39:42] inside me. Um the three webs are the [39:44] commerce web, um the application web and [39:47] the content web. And the content web has [39:49] been in trouble for a while. I think [39:50] that the commerce web has actually done [39:52] really well and then the application web [39:54] has definitely been on a huge upswing [39:56] ever since coding agents were released. [39:58] Um but the commerce web is fascinating. [40:01] You know, you saw this week how Shopify [40:03] embraced Muse and Amazon blocked Muse. [40:06] I think part of the calculus is going to [40:09] be a winner or loser when you actually [40:10] have a new front door for how traffic [40:13] flows to these marketplaces. [40:15] >> I mean, do you have a view on what [40:17] specifically happened with Amazon and [40:19] Shopify and maybe more broadly who the [40:21] winners and losers are likely to be? [40:23] >> Yeah, so I think you know, first off, [40:25] Shopify and Amazon are two completely [40:27] different business models. You have [40:28] Shopify that's hyperfocused on more [40:30] democratizing commerce. You have Amazon [40:32] that makes all of their money on ad [40:34] revenue. And so the whole ad structure [40:37] of commerce is going to be really [40:38] fascinating to watch because when you [40:40] remove human eyeballs, it defeats the [40:42] entire purpose of the system. [40:43] >> And so I think that's where Amazon was [40:46] coming from. Like if you let Muse and [40:48] these agents start to purchase without [40:51] them getting any ad revenue out of that, [40:53] >> it really takes a lot of their profit [40:54] out of the experience. So, I think [40:56] there's likely going to have to be some [40:59] new paradigm for how they're able to [41:02] profit in this world where agents are [41:04] are purchasing on their behalf versus a [41:07] Shopify that is empowering the everyday [41:09] individual to sell as much as possible [41:12] and agents democratize this space even [41:15] further. So, it's it's just unbelievably [41:17] advantageous. Um I think it's going to [41:20] be even more interesting when we look at [41:23] industries that have yet to be touched [41:25] and that's you know something we chatted [41:27] about uh last week was restaurants where [41:30] if you look at the booking space right [41:32] like you go to make a reservation I [41:34] think we saw resi uh resi started [41:36] shutting people's accounts down because [41:37] they were using browser use and they [41:38] didn't want bots so it defeats the whole [41:41] the whole business of if you just let [41:44] agents take these reservations [41:46] instantaneously [41:47] then it kind of defeats the purpose of [41:49] the platform of this like humanoriented [41:51] booking experience. [41:52] >> And so what happens though when everyone [41:54] has an agent now everyone's on the same [41:56] playing field. Everyone can go to book [41:58] the restaurant at the same time. [42:00] >> Do we enter a world where one power goes [42:04] back to the restaurants and they can [42:06] maybe choose who they want to give the [42:08] reservation to based off of loyalty or [42:09] average spend or whatever it be. or two, [42:13] do we go into this world of bidding [42:15] where everything becomes a bidding war [42:17] between these agents? I'm curious your [42:19] thoughts on kind of what that future [42:21] looks like. [42:21] >> Yeah, I think you're super right. I [42:23] mean, very interesting because I think [42:24] there'll be um things that are supply [42:25] constrained like hot restaurant tables, [42:27] things that are demand constrained like [42:29] people that want to fly to SF tomorrow. [42:31] I think for restaurants, if you think [42:33] about with perfect information, who do [42:35] they want? to your point, it's the [42:36] highest AOV, highest LTV, which means [42:39] AOV over multiple visits. Plus, they [42:42] actually want to ensure that they're at, [42:44] you know, near 100% capacity as much as [42:46] possible. So, that's actually a more [42:48] sophisticated calculation and I don't [42:50] quite know how they're going to do it, [42:51] but I'm sure that proprietary supply is [42:54] going to become more and more important, [42:56] you know, versus relying on kind of [42:59] friction and motivation from the end [43:01] consumer to act as something that sort [43:03] of prevents this DOS. So I think supply [43:06] constrained sort of aspects of commerce [43:08] will be really interesting and then [43:09] demand constrained will be completely [43:11] different you know where you almost have [43:13] a group buying experience where you have [43:15] a set of buyers that are willing to buy [43:17] a product or a service or an experience [43:20] and then you have the supply side the [43:22] merchants that could potentially fulfill [43:24] that need come and you know bid to see [43:26] who can do it the best experience maybe [43:28] at the lowest price. So we've always had [43:30] this kind of aggregation of demand, but [43:33] we've never really had aggregation of [43:34] supply in the same way, especially not [43:36] in an auction system to benefit the [43:38] consumer. [43:38] >> So I think that's really interesting. [43:40] >> You know, in the Amazon Shopify thing, I [43:42] think that the Amazon Amazon probably is [43:45] rightly taking the bet that they don't [43:46] want to be disintermediated and that [43:48] they'll lose not just advertising [43:50] revenue, but impulse shopping. I'm kind [43:52] of curious to know if the agents [43:54] themselves will be impulse shoppers to [43:57] some extent because I think the point [43:59] around proactivity means the agents are [44:01] going to try to make delightful guesses [44:03] and you know what is more delightful [44:05] than a surprise or a gift. [44:07] >> So I'm guessing that one of the really [44:08] interesting um sort of dynamics were [44:11] going to play out is that these agents [44:13] are not perfect reflections of their [44:14] principles. They're not perfectly [44:17] rational. They're going to be new [44:18] rational and nonrational in completely [44:20] new ways. And I don't know that things [44:23] like marketing, things like impulse [44:25] purchases, things like advertising won't [44:26] matter. I just think they need to be [44:28] transformed. [44:29] >> Totally. And I think that opens up a [44:31] whole new conversation as well, which is [44:33] what is the recommendation engine of [44:35] this agentic commerce space actually [44:37] look like? Yeah. [44:38] >> Because AI and LLM feels like it's kind [44:41] of broke the internet in terms of [44:44] quality search. But then if you use an [44:47] agent and you say, "Hey, buy me some [44:49] white socks." [44:50] >> There's like 20,000 different types of [44:52] white socks out there, it's going to [44:53] send you, it doesn't actually know what [44:55] exact type of white socks you prefer. [44:57] >> Yeah. [44:57] >> And so I think it'll be really [44:58] interesting. I think Muse has probably a [45:01] head start on this. And you know, this [45:02] is whole Zuck's whole thesis of attacks [45:05] on all commerce where they have your [45:07] Instagram data, they have your Facebook [45:08] data, so they probably have a little [45:09] better understanding of what you [45:10] actually prefer. But it'll be really [45:12] interesting to see what this [45:14] recommendation layer looks like. Yeah. [45:16] >> In this world of commerce. [45:18] >> Well, I think it could be fun for the [45:19] longtail, you know, because you could [45:21] imagine to your white socks example, [45:22] like I could go to Amazon and get white [45:24] socks, you know, but maybe my agent [45:26] comes back and says, "Hey, I found [45:27] someone awesome on Etsy that can knit [45:29] you the socks, you know, grandma in [45:32] Nebraska, and she she just knits these [45:33] amazing white socks. It's going to take [45:36] three weeks. Are you down?" I'm like, [45:37] "Sure, I'm down. That sounds [45:38] interesting." So, I kind of wonder if it [45:40] brings more serendipity in because of [45:43] course, you know, it can examine all the [45:45] options and perfectly tailor these [45:47] things to you with no adverse [45:49] incentives. [45:50] >> Right. I would love that. I think that'd [45:52] be so fun. It would definitely add a [45:54] little flavor to the mundane life of [45:56] shopping today. [45:58] >> Exactly. [45:59] >> Mundane from my point of view as someone [46:01] who, you know, doesn't take joy in uh [46:04] sock shopping. Well, dude, and look, you [46:06] could even imagine a world and you get a [46:07] little more futuristic where it starts [46:09] to now um catalyze things in the real [46:13] world. So maybe instead of grandma who [46:16] knits white socks and sells them on [46:18] Etsy, you know, maybe it messages the [46:20] agent of a grandma and says, "Hey, would [46:21] you be up for doing this?" So I think [46:24] this is where the kind of agentto agent [46:25] spaces get really interesting where [46:27] they're all acting on behalf of their [46:29] humans. you know, some to provide [46:31] products, services, knowledge, and some [46:34] to make money. And, you know, maybe [46:36] there's this sort of secondary economy [46:38] that's much more dynamic, much more high [46:40] frequency, um, and just much more [46:42] interesting than the one we have today. [46:44] >> I'm I'm just imagining now a grandma in [46:47] Nebraska receiving a text message [46:48] saying, "Hey, this guy in SF wants you [46:50] to knit white socks for him. You want to [46:52] do it?" [46:52] >> Why not? Yeah. Yeah. Why not? [46:54] >> It'd be awesome. [46:55] >> Yeah. So, I'm I'm I'm actually very [46:57] hopeful. You know, it also it sort of [46:58] sets the incentives in the right way [47:00] because if the agents are doing the [47:01] right thing, they're choosing great [47:03] products based on product quality and [47:05] they can assess product quality for [47:06] themselves rather than, you know, the [47:08] thing that has the best marketing or the [47:10] best brand. And by the way, if they are [47:11] choosing on a brand basis, it's at least [47:13] with the explicit sign off of their user [47:16] like I just want to have Nikes cuz I [47:18] love Nikes, not because you know there [47:20] there's sort of this cognitive [47:21] dissonance that's happening in the [47:23] marketing and the consumer is actually [47:24] doing something that's not optimal for [47:26] themselves, [47:27] >> right? It it opens the door to a truly [47:29] consumer aligned internet which I think [47:31] we have been lacking for quite some [47:33] time. [47:34] >> I agree, man. We touched on the the [47:36] economics of the agents. Of the 122 [47:41] that I have looked through, so I've [47:42] personally tested 26 of them. [47:45] >> Okay. [47:46] >> Of the 122, [47:49] uh, 65 are paid. [47:51] >> Interesting. [47:52] >> 35 of them are fully paid. 30 are a [47:55] premium model. So once you hit a limit, [47:57] then you have to pay. Only 13 of them [47:59] are actually free. [48:02] When you have Muse and Instinct that are [48:03] both free that are also the dominant [48:05] ones and then you have Open AI that [48:07] supposedly is going to be releasing [48:08] something cool maybe next week, who [48:10] knows? [48:12] >> How does a longtail startup actually win [48:14] this space? [48:16] And why are they charging money when the [48:20] big dogs are giving it to you for free? [48:23] >> Man, it's going to be really [48:23] challenging. you know, by our estimate, [48:25] it's costing something like $20 per user [48:28] per day [48:29] >> to do this in a really ambitious way. [48:31] You know, potentially hundreds of [48:32] millions a year for a startup. I mean, I [48:34] think the two positive trends to bet on [48:37] are that browser use should get it [48:39] should be deflationary, get [48:41] exponentially cheaper, very rapidly. Now [48:43] if you're a startup that is raising [48:45] today that's betting on that happening [48:47] in 6 months maybe it's a little bit of a [48:49] game of chicken but we have seen kind of [48:50] token prices depress dramatically and [48:53] sort of emerging capabilities depress so [48:55] the price of them [48:56] >> so I think it's very possible that [48:58] browser use gets cheap enough that more [49:00] um more products can offer this for [49:01] free. I guess the flip though is that I [49:04] don't love the idea of you know a [49:06] customer subsidy being the primary value [49:08] prop. So, I'd love to see the agent that [49:11] costs $1,000 a month that the consumer [49:14] is so excited about that they're willing [49:15] to pay. Like, people spend $1,000 a [49:17] month on a lot of different things that [49:19] are, you know, not necessarily [49:21] necessities, their sort of desires. So, [49:24] what is the agent that is exciting [49:26] enough that you're willing to pay that [49:28] kind of price for it and then that agent [49:30] will benefit from both extreme market [49:32] fit as well as sort of deflating costs. [49:35] >> Right. Right. I think it'll be [49:36] interesting to see. [49:37] >> Super interesting. Well, David from [49:39] Assistant Bench, make sure everybody [49:41] checks it out. Thank you so much for [49:42] joining us. We're going to do this [49:43] frequently. I mean, I feel like we could [49:44] do this every week, dude. Like, things [49:46] are happening every single day. So, [49:48] >> please follow him on X. Follow me as [49:49] well. We'll all keep up to speed and [49:51] sort of learn together in the X [49:52] community. And uh and sort of [49:55] >> happy day to you and your agents. [49:57] >> Appreciate it. Same to you. This is fun. [49:58] >> All right, brother. Thank you. [50:00] >> Awesome.