A16z Sees $20 Daily Compute Burn Squeezing Consumer AI Startups

A16z称每日20美元算力消耗正挤压消费级AI初创公司 面对大型科技公司的免费替代品，个人AI初创企业正面临不可持续的算力补贴危机，除非它们从广告变现模式转向能够支撑高溢价订阅定价的高效行政工作流。 David Pawlan与Anish Acharya基于Pawlan对120多个智能体产品展开的Assistant Bench评估，深入剖析了消费级AI助手的爆发式增长。他们指出，真正的用户需求并非预订旅行等花哨的偶发任务，而是集中在无感、后台化处理繁琐的日常生活行政事务以及减少财务阻力以节省开支。两位主持人探讨了产品界面在iMessage、独立App和环境音频穿戴设备之间的分化，同时审视了有益的主动性与失去用户信任之间脆弱的边界。展望未来，他们解读了智能体之间（agent-to-agent）协作的兴起，并评估了自主软件将如何重塑在线商务、绕过广告驱动的平台，以及对传统初创公司的单位经济模型提出挑战。 在2026年9月Poke、Instinct和Muse等产品集中发布后，消费级AI领域已迅速步入个人智能体时代；David Pawlan的Assistant Bench已对其中122款产品进行了追踪。 (3:20) 来自1200多名活跃社区成员的真实用户数据显示，清理电子邮箱收件箱、文书处理和表格填报等日常行政杂务是最普遍的用例，其热度超过了在网络上爆火的旅行预订。 (7:35) 目前支持复杂的消费级智能体工作流每位用户每天的算力成本估计达20美元，这导致了严重的资金消耗，迫使初创公司要么向高客单价定价方向实现差异化，要么寄希望于推理成本的迅速下降。 (45:00) 智能体商业正在不同平台间引发分化：Shopify接纳自主智能体以助力商家提升销量，而依赖广告收入的电商平台如亚马逊则选择封禁智能体，以保护广告展示次数和冲动消费。 (40:00)

Following conversational interfaces and developer coding tools, personal AI agents represent generative AI's third inflection point, moving from passive chat toward proactive task execution across consumer life. The Rise of Consumer Agents and Assistant Bench (4:46) Following conversational bots in late 2022 and coding agents, consumer personal agents expanded rapidly in September 2026 with releases like Poke, open-source OpenClaw tooling, Instinct, and Meta's Muse. To bring objective clarity to an influx of experimental products, David Pawlan developed Assistant Bench, an empirical consumer evaluation assessing agents across 16 operational dimensions such as task completion, clarification prompts, and latency. Pawlan tracks 122 agent products across two main categories: approximately 64 horizontal B2C agents (including Instinct, Muse, Caddy, Ollie, and Season) and specialized B2B workflow tools (like Catch and Vellum). Although episodic tasks like flight booking generate viral social media attention, Pawlan's data from group chats of over 1,200 users reveals that mundane daily administration—such as inbox triage and completing forms—is the most sustained use case, followed by multi-agent coordination. Invisible Economics and the Autonomous Trust Threshold (10:00) General partner Anish Acharya and Pawlan note that broad consumer interest is driven less by marginal productivity improvements than by background friction removal and recovering money lost to administrative inertia. Strong proactive use cases include auditing past receipts for Health Savings Account reimbursements, securing automated airline price-drop refunds, or optimizing home sprinkler schedules with weather data to cut water bills in half. Pawlan argues that proactivity separates market leaders from generic bots, categorizing actions into zero-friction background tasks and high-impact structural decisions that require explicit sign-off. Because autonomous overreach can cause immediate user churn, assistants must navigate a strict trust threshold: taking helpful initiative without crossing into presumptuous, unauthorized actions. Form Factors, Interfaces, and Multiplayer Coordination (12:35) Consumer access preferences currently split between native messaging like iMessage, standalone visual applications, and lock-screen widgets, with generational splits shaping interaction styles. While Pawlan views hardware like Meta's Muse charm as a passive sensor collection mechanism for physical-world mapping rather than a purely standalone hardware play, both speakers identify voice interfaces as a major breakthrough. In hands-free scenarios such as bike commutes, continuous voice sessions allow users to triage emails and organize calendars without screens. In shared social spaces, dropping autonomous bots directly into group chats often causes conversational friction because modern models still struggle with social nuance and conversational timing. Pawlan highlights alternative designs, including silent listener agents that track threads to log action items in external shared documents, and backchannel agent-to-agent networks that exchange information privately and deliver consolidated updates. Commercial Disintermediation and High Operating Burdens (40:00) The rise of agent-driven shopping exposes conflicting digital business models. Amazon has taken measures to block scraping agents like Muse to preserve human ad impressions and high-margin impulse purchases, whereas merchant-enablement platforms like Shopify welcome agentic integration to drive transaction volumes. Pawlan notes that booking services like Resy have similarly disabled bot-driven reservation scripts to prevent instant gridlock, foreshadowing a shift toward algorithmic guest screening or dynamic bidding. Simultaneously, agent startups face severe unit economic hurdles. Of the 122 startups tracked on Assistant Bench, 65 are paid and 30 use freemium tiers, with only 13 operating completely free. Acharya estimates ambitious browser-use workflows can cost roughly $20 per user daily—accumulating to potentially hundreds of millions of dollars annually—which makes subsidizing compute via venture capital unsustainable unless startups deliver high enough value to command premium pricing. While personal agents promise to absorb bureaucratic life overhead and unlock peer-to-peer commerce, startups face an unresolved challenge: sustaining steep browser-execution compute costs while navigating platform resistance from incumbent ad-driven gatekeepers.

a16z普通合伙人Anish Acharya与Assistant Bench创建者David Pawlan认为，生成式AI正从对话和代码工具快速演进为面向大众的个人自主智能体（Personal Agent）。该领域未来的竞争重心不仅在于基准模型性能，更在于主动化任务执行、新型交互形态以及智能体间协议对传统商业架构的重构。 从极客工具到日常琐事代理的真实需求 (1:54) 生成式AI在经历2022年底至2023年的对话浪潮（如ChatGPT）及随后的代码生成浪潮后，在2026年9月迎来以Poke、Instinct、Muse等为代表的消费级智能体爆发。针对市面上大量产品，Pawlan建立了包含16个现实测评维度的Assistant Bench，追踪了122款产品，旨在衡量其实际任务达成度、追问逻辑和响应延迟等真实体验，而非纯技术评测。在已追踪的122款产品中，约64款属于横向B2C综合助手，其余多为垂直或B2B流程助手。 尽管机票预订等旅游场景在社交平台上最具传播度，但Pawlan运营的1200多人社区数据显示，其实际使用量仅列第四。用户最强烈的真实痛点集中于繁重的日常行政杂务（如清理收件箱和填写表格），其次是多智能体调度与开发设计任务。Acharya指出，普通消费者对将效率提升10%并不敏感，但在能够省去繁琐手续或直接带来实质利益（如通过分析历史收据自动报销健康储蓄账户HSA，或在机票降价时自动索取差价退款）的场景中，智能体能展现出极高粘性。 交互界面的分化与隐形执行的信任边界 (12:35) 在交互入口上，市场目前分化为iMessage等即时通讯工具、独立应用程序、锁屏小部件以及硬件设备。Acharya与Pawlan认为，短消息占据着用户高频且私密的沟通界面，但代际偏好差异显著，视觉化规划或旅行常依赖独立App。对于Meta推出的Muse挂坠等硬件，Acharya对单一硬件形态持怀疑态度，认为设备形式取决于个人风格；Pawlan则推测其核心目的可能在于收集全天候的物理世界传感器数据。与此同时，语音交互在烹饪、通勤等双手受限的场景下展示出强大潜力，如Pawlan在30分钟骑行中通过ChatGPT语音指令配合Gmail和日历连接器实现了收件箱清零。 智能体的最大壁垒在于从“被动助理”向“主动代理”跨越时的尺度把控。Pawlan将主动行为划分为无需确认的零摩擦收益（如自动值机、追踪退款）与必须用户签字的高风险决策（如更换保险方案）。若在敏感领域越界或擅自决策，将瞬间破坏用户信任。此外，在多人聊天场景中直接接入智能体常会打破社交礼仪，因此通过静默监听、私下单聊派发任务或后台智能体互通（Agent-to-Agent），被认为是更适宜的协作架构。 商业解构：分发中介化与智能体原生生态 (40:00) Acharya将互联网划分为内容网、应用网和商业网，并指出商业网正面临智能体带来的去中介化冲击。当个人智能体绕过广告和搜索结果直接代客下单时，高度依赖消费者眼球与赞助广告的平台（如亚马逊）选择抵御或封锁自动化工具；相反，专注于赋能商户履约的平台（如Shopify）则积极拥抱智能体接入。对于餐厅预约（如Resy）等供给受限的市场，大量并发的自动化请求迫使平台从防范机器人攻击转向基于客户终身价值（LTV）的筛选或动态竞价机制。 展望未来，智能体间通信协议的建立将重塑服务流程，使直接面向智能体的独立标识、专属通讯渠道和安全权限机制成为刚需。在供需匹配上，摆脱广告导向的智能体能够根据用户的深度偏好直接对接长尾手艺人，促成长尾直连。然而，智能体初创企业的底层单位经济学依然严峻。Pawlan测试发现测评收录的122款产品中仅13款完全免费，35款设有直接付费墙。Acharya估算高水平运行此类代理每日每用户算力成本可能达20美元，靠风险投资补贴并非长久之计，初创企业亟需转向能支撑高客单价的高价值专业场景。 个人智能体正由概念验证走向真实工作流落地，但其规模化普及仍受制于高昂的运行推理成本、广告驱动型存量平台的防御壁垒，以及在极度追求主动性与避免用户信任崩塌之间的微妙平衡。

## Background 本期节目由 a16z 合伙人 Anish Acharya 与 Assistant Benchmark 创作者 David Pawlan 共同录制，于 2026 年 9 月 29 日通过 RSS 及播客平台发布。节目围绕个人 AI 智能体（Personal AI Agents）的爆发式增长展开，深入探讨了智能体如何融入日常生活、背后的交互形态与自主权边界，以及软件代人行动时对互联网和商业生态的深远影响。 ## TL;DR 对话从 David 追踪评测 122 款智能体的第一手观察切入，梳理了从生成式对话、编程助手到个人代理的三次浪潮，并指出智能体真正的杀手级场景在于自动报销、退差价等“隐形替用户省钱捡钱”的繁琐行政杂务。双方探讨了 iMessage、独立 App、纯音频眼镜与全双工语音（Full-Duplex）等不同交互载体，并剖析了智能体在主动行动（Proactivity）时不可逾越的信任红线。针对商业模式与底层网络演进，两人深入对比了 Shopify 与亚马逊对待 AI 代理的不同策略，展望了“代理对代理”（A2A）的全新原生网络，并指出初创公司应摆脱单纯的资本补贴，依靠垂直领域的专有品味与极致价值建立护城河。 ## Atmosphere & style 整场对话洋溢着硅谷前沿科技圈特有的兴奋感与极客探索欲，节奏紧凑且充满思想碰撞。Anish Acharya 展现出顶级风险投资人的宏观格局与商业敏锐度，擅长将零散的产品现象升华为行业假说；David Pawlan 则作为深度测试数十款前沿工具的硬核体验者，不断抛出接地气的评测数据与鲜活的个人生活案例。两人在默契的调侃与思想推演中，既勾勒了技术解放日常生活的浪漫近未来，又保持着对单用户高昂推理成本、平台防范封锁与信任危机等现实挑战的清醒审视。 ## Bulletin - `03:40` 个人 AI 智能体浪潮迎来爆发式增长，Assistant Benchmark 追踪的 122 款产品中，通用横向代理占比超过一半，共有 65 款包含付费门槛，而完全免费者仅有 13 款。 - `48:00` 高昂的算力与自动化成本构成严峻挑战，按目前顶级标准运行个人智能体，单用户每日运行成本高达约 20 美元，年化开销可达数亿美元，倒逼初创公司押注高客单价与极致市场契合度（PMF）。 - `40:00` 电商巨头在智能体入口的博弈呈现两极分化：Shopify 全面接入 Muse 等代理以推动交易去中心化，而亚马逊为保护高利润的展示广告与搜索流量则坚决封锁代理访问。 - `06:33` 相较于低频且仅具病毒传播效应的旅游场景，用户对智能体的核心高频刚需集中在日常行政杂务（如 HSA 自动报销、机票退差价）以及结合全双工语音的移动免手交互。 ## Key takeaways - `00:00` 个人智能体（Personal AI Agents）浪潮已继生成式对话大爆炸与编程智能体之后，成为 AI 演进的第三个关键转折点。 - `23:45` 智能体主动性（Proactivity）是核心护城河，但这是一条钢丝：无害的草拟邮件或退款可带来极大惊喜，但一旦越权擅作主张（如私自解约合同），就会彻底失去用户信任。 - `10:00` 普通消费者并不追求微小的效率提升，他们更看重后台完全“隐形”的行政代办，以及像漏报 HSA 报销、降价退差价这种能带来“捡到免费钱”心理刺激的财务场景。 - `18:00` 全双工（Full-Duplex）实时语音交互创造了免手操作的杀手级体验，在骑行、烹饪或园艺等无法触碰屏幕的碎片时间中能高效实现“收件箱清零”。 - `20:00` 将智能体直接拉入好友群聊往往会侵蚀社交资本；更优雅的落地方案是充当“静默监听者”，在后台整理纪要并通过私聊派发个人待办，避免公开刷屏。 - `30:00` “窄门初创”（Narrow Startups）假说证明了垂类代理的商业可行性：即使只服务极小众群体（如 3 岁以下幼儿的单身母亲），凭借独特品味与专有数据，收取每月 200-300 美元订阅费也能撑起规模化 ARR。 - `32:45` “助理”（Assistant）与“代理”（Agent）存在本质区别：助理仅按指令被动执行，而代理具备能动性，能在现实世界中主动促成好的改变，甚至能充当人际交流中的“社交间接缓冲层”。 - `35:20` AI 技术的终极目标不是削减成本，而是“可能性扩展”（Possibility Expansion），通过卸除人类在法律、文书及生活杂务上的官僚负担，释放认知带宽去探索新生活。 - `38:00` 全新的 Agent-to-Agent（A2A）机器网络正在成型，未来代理将拥有专属邮箱与电话号码，重构从服务预约到反向拍卖交易的底层商业链路，并催生代理原生网络安全新产业。 - `40:00` 基于广告变现的平台（如亚马逊）在代理时代遭遇商业模式冲击，因为代理代购消除了人类眼球注意力；而线下预订平台（如 Resy）则已开始封禁自动化浏览器代订工具以维护既有秩序。 - `45:00` 对话式代理有望重塑传统搜索与推荐引擎，摆脱过度营销与认知失调诱导，通过挖掘长尾独立手工制作者直接促成撮合，打开真正与消费者利益一致的互联网新入口。 - `48:00` 在巨头凭借雄厚资本推出免费代理的围剿下，创业团队不应寄希望于资本补贴，而应打造每月定价可达 1000 美元、满足深层渴望且具备极致 PMF 的高价值产品以求突围。 ## Detailed walkthrough ### `00:00` 个人智能体浪潮的爆发与“见证神迹”的时刻 对话在一段极具信息密度的开场集锦中拉开序幕：普通大众根本不在乎效率是否提升了 10%，更关键的变革在于技术如何以隐形的方式重构现实生活。嘉宾 David Pawlan 抛出了一个辛辣的观点：Meta 的智能眼镜与其说是为了赢得硬件战，不如说是为了在现实世界中采集海量数据，以此为扎克伯格构建未来元宇宙绘制真实物理世界的地图。团队内部甚至开玩笑说，按照现在的演进速度，要不了几天就会出现代理直接发信息给用户说：“我发现你没那么喜欢她，所以自作主张帮你提了分手。”两人强调，主动性（proactivity）蕴藏着极高的竞争壁垒，比如在骑车上班的 30 分钟里，戴着耳机与 ChatGPT Voice 随性交谈，代理就能自动将邮件归类、打上标签、发出日程邀请，让人刚到办公桌前就轻松实现“收件箱清零”（Inbox Zero）；但这同样是一条不可触碰的钢丝，“一旦你越过这条红线一次，你就会彻底失去用户的信任”。 回顾整个人工智能的发展弧线，a16z 的合伙人 Anish Acharya 指出，技术生态中的人们经历了数次“见证神迹”般的顿悟时刻。第一个关键拐点是 2022 年底至 2023 年初 ChatGPT 的横空出世，那是 AI 浪潮的“大爆炸”，人们第一次惊叹于一个能写邮件、作诗、且来回流畅对话的合成人；第二个拐点则是开发者社区中编程智能体（Coding Agents）的爆发，从 Claude Code 到 CodeX 彻底改变了软件开发逻辑；而当下随着 Instinct、Muse 以及 ChatGPT 针对日常任务处理的深化，行业正式迎来了个人智能体（Personal AI Agents）狂飙突进的第三次浪潮。 关于这一赛道中用户信任与边界的脆弱性，David Pawlan 直截了当地指出： > “But it will be a very fine line because if you cross that line once, you lose all trust with your user.” - **演进三阶段**：对话生成大爆炸（ChatGPT） $\rightarrow$ 开发者编程智能体（Claude Code / CodeX） $\rightarrow$ 消费级个人智能体（Instinct / Muse / ChatGPT）。 - **核心能力分水岭**：从被动等待用户输入 Prompt，转向后台主动感知（Proactivity）与自动化处理。 --- ### `03:40` Assistant Bench：为蛮荒西部建立消费级标尺 过去一个月内，个人智能体赛道呈现出指数级的井喷。David 回忆道，自己是较早入局的体验者之一，最初接触的是 9 月 8 日上线的消费级代理 Poke，他在发布后三天（9 月 11 日）便开始深度使用。当时 Poke 的核心噱头是让用户与代理讨价还价决定服务收费，体验极具冲击力。随后 OpenClaw 发布，再次引发关注；紧接着 Instinct 瞬间引爆了硅谷科技推特（X）圈，随之而来的便是一场多米诺骨牌式的爆发：Grokbot、Muse，以及 Caddy、Ollie、Pally 等一系列长尾玩家纷至沓来。然而，尽管产品层出不穷，但整个行业依然像没有规则的“蛮荒西部”（Wild, Wild West），谁也不知道终局在何方，大家都在摸着石头过河。 正是在这种信息极度不对称的背景下，David 创建了 Assistant Bench。这一评测并不是针对后台底层架构的技术参数测试，而是一个纯粹面向消费者的实用标尺——解决的是“如果我想订机票，到底该用哪一个？谁反应最快？谁能真正闭环解决问题而不是抛回一堆反问？”该基准从 16 个不同维度对智能体进行横向切片与打分。基准发布仅 16 天，访问量激增，几乎所有智能体初创公司的创始人都主动找上门来沟通，这充分反映出从业者在面对未知赛道时对标准化衡量体系的极度渴望。 对于当前行业普遍处于摸索阶段的现状，David Pawlan 生动地总结道： > “I think everyone's kind of in the same boat where nobody really knows. It's the wild, wild west. We're just playing around here.” ```mermaid flowchart TD accTitle: Assistant Bench Evaluation Workflow accDescr: A compact diagram showing how consumer intent is routed and evaluated across multiple dimensions. subgraph Input["用户真实任务"] T["复杂指令 (如：预订航班/清理收件箱)"] end subgraph Bench["Assistant Bench 评测体系 (16个维度)"] D1["一次性任务完成度 (One-shot)"] D2["多轮追问交互机制 (Follow-ups)"] D3["真实响应耗时 (Latency)"] end subgraph Output["横向评测结果"] R["综合表现排名与能力缺陷洞察"] end T --> D1 T --> D2 T --> D3 D1 --> R D2 --> R D3 --> R ``` --- ### `06:33` 市场格局分化与用户真实需求的四大阵营 目前 Assistant Bench 上已收录了多达 122 个独立智能体。David 将当前的版图划分为两大主营：一类是面向 C 端的通用型水平智能体（Horizontal General Agents），如 Instinct、Muse，以及长尾的 Caddy、Ollie、Season 等（仅通用大类就达 64 个）；另一类则是偏向 B2B 工作流与高管助理型的工具，如 Town、Catch、Vellum 等，其定位是充当数字化行政秘书。然而就竞争烈度而言，绝大部分团队仍扎堆在横向通用赛道，试图吞噬用户生活中的一切琐碎事务。 针对特定垂类（如旅游 Travel），Anish 提出疑问：旅行是一种低频但高价值的行为，垂直类旅行代理究竟会独立胜出，还是最终只会沦为通用横向智能体的一个外接插件（Connector）？David 认为，垂直领域或许能跑出一两家独立的赢家，但绝大多数都逃不过被通用平台整合为连接器的命运。从由 1,200 多名重度用户组成的讨论群组来看，旅行在讨论热度上仅排第四。虽然“帮我自动值机”在社交媒体上最容易制造病毒式传播，但由于普通人不会每天旅行，它本质上更像是一个抓人眼球的营销噱头（attention grab），而非粘性留存的核心抓手。 实际数据显示，真实用户群聊中探讨最为集中的前三大核心场景远比外界想象的更加务实： 1. **日常行政杂务（Daily Admin）**：收件箱整理、填表等真正令人心烦但无法回避的高频刚需。 2. **多智能体编排（Agent Orchestration）**：如何寻找最高效的智能体，以及如何让不同智能体彼此通信、协同作战。 3. **编程与设计迭代（Development & Design Focus）**：将开发能力迁移至手机端，例如通过给代理发送短信直接触发 Claude Code 跑代码，或者设计师直接发短信给代理对设计稿进行细微修改。 排名靠后的金融（Finance）则是最耐人寻味的场景。David 坦言，这几乎是最不切实际的方向，因为如果真有代理能帮人赚到十亿美元，对冲基金早就实现了，普通用户不可能单靠手机就暴富；但每当人们拿到一个强大的新工具，本能反应永远是：“你能帮我赚一百万吗？”针对这种矛盾，Anish 引用了 Ben Thompson 在 TBPN 播客中的核心论点，指出大部分普通消费者在生活中根本不追求效率与生产力，他们只想“消磨时间”（spend time）而不是“节省时间”（save time）；不过在金融与财务等涉及繁重行政负荷（administrative overhead）的领域，自动化依然存在巨大的刚需空间。 关于旅行类智能体在喧嚣背后的本质定位，David Pawlan 评价道： > “To your point, people aren't traveling every day. So it's more of a gimmick, right? It's an attention grab.” - **智能体总量分布**：全网追踪到 122 款不同智能体，其中 B2C 通用横向分类下占 64 款。 - **用户核心讨论榜单**： 1. 日常行政处理（Daily Admin） 2. 智能体相互编排与通信（Agent Orchestration） 3. 移动端编程与设计迭代（Dev & Design） 4. 旅游出行预订（Travel - 病毒传播性强但频次低） 5. 个人理财与套利（Finance - 幻想色彩最浓但实际可行性最低） ### `10:00` 隐形助手与“免费捡钱”：个人 Agent 的杀手级财务场景 谈到个人 Agent（智能体）的切入点，Anish Acharya 指出，现实中存在着大量由消费者的被动懈怠（apathy）或信息闭塞所堆砌起来的庞大利润池，如果 AI Agent 能够将这些利润重新返还到消费者手中，价值将极为可观。尽管大部分普通消费者上网是为了娱乐和消磨时间，但在美国消费者的日常生活中，沉重的行政杂务占据了大量精力，这正是 Agent 能够实现质的突破的领域。 David Pawlan 对此高度认同，并直言普通大众根本不在乎效率提升 10% 这种微小收益。人们雇佣助理的真正目的，是为了让对方变得“隐形”：不用主动去管理它，它自己在幕后搞定一切，并在事后汇报成果。David 观察到，目前最具前瞻性的场景大多集中在财务领域，但并非帮人花钱，而是帮人“省钱”。例如有人让 Agent 翻阅过去整整一年的收据并自动提交 HSA（健康储蓄账户）报销；还有朋友将 GrokBot 接入自家的草坪喷灌系统和天气系统，根据天气自动决定是否浇水，直接把水费砍掉了 50%。 Anish 敏锐地指出，对美国消费者来说，“省钱”往往不如“免费捡钱”有吸引力。Agent 带来的心理冲击正是让人感觉平白赚了钱：不管是补齐漏掉的 HSA 报销，还是自动监测机票降价并要求航空公司退还差价。这种低频但高价值的行为，正是垂直连接器（vertical connectors）能够切入并发展为数十亿美元业务的根据地。 > “The best employee I've ever had are the ones that are just doing their work, getting things done. They update you. Here's what I did. And you're like, wow, amazing. You don't hire someone because you enjoy managing them.” —— David Pawlan ```mermaid flowchart TD accTitle: Agent 从隐形运行到心理激励的价值闭环 accDescr: 展示 AI Agent 如何在后台接管繁琐行政任务，并将利益以“捡到免费钱”的形式返还用户。 subgraph Background["后台隐形自动化处理"] A["监测用户开销与日程"] --> B["扫描历史单据与机票价格"] B --> C["执行索赔/申请 HSA 报销/降价退差额"] end subgraph Impact["用户感知与心理反馈"] C --> D["水费减半 / 意外现金入账"] D --> E["心理感受：'白捡了一笔钱' (Free Money)"] E --> F["对 Agent 建立极高粘性与信任"] end ``` - **HSA 自动报销**：翻查过往一整年的消费收据，提取符合条件的健康储蓄账户项目并代为提交。 - **降价退差价**：自动化监控已购机票的价格波动，在降价时自动向航空公司申请退回差价。 - **智能节水喷灌**：将 GrokBot 与天气预报及喷水控制器联动，使家庭水费降低了 50%。 --- ### `13:17` 交互界面的博弈：iMessage、独立 App 与个性化偏好 关于用户究竟该通过什么界面与 Agent 互动，两人展开了探讨。David 提出，他和 Anish 倾向于生活在 iMessage 里，因为这种方式无需改变既有的日常工作流；但 Anish 的妻子和朋友则更偏爱独立的 App 界面。这种偏好差异很可能沿着代际和使用场景分化：Z 世代可能更习惯纯文本聊天，而千禧一代（Millennial）或 X 世代可能又有不同的诉求。 Anish 分享了家庭内部的鲜明对比：他的妻子喜欢做“愿景板”（dream board），聚焦于实现年度目标和长期人生规划，旅行安排也高度依赖视觉呈现，因此独立且富含视觉交互的界面对她必不可少；而 Anish 自己则完全是功能导向，只关心“飞机上有没有 Starlink 卫星网络”。Anish 认为，iMessage 对他而言是一个神圣且私密的位置，与 Agent 在短信里对话感觉更加亲密无间；但他也坦言，如果未来通讯录里每隔一条对话就是一个 Agent，这种特权地位可能会被打破。David 还提到，初创公司 Sky 正在尝试通过 iPhone 桌面小组件（Widget）来打造第三种交互入口。 > “I feel like the iMessage space is more precious. So it feels like a more personal interaction when I'm in that space.” —— Anish Acharya --- ### `15:00` 硬件形态的分化：Meta 的吊坠、实时数据收集与纯音频眼镜 硬件载体的讨论因 Mark Zuckerberg 及团队刚发布的 Muse 吊坠（Charm）而变得更为热烈。面对“某种硬件形态将成为终极载体”的论调，Anish 持怀疑态度。他认为硬件就像珠宝首饰，有人爱戴项链，有人爱戴耳环，最终形态将因人而异。不过，随身携带的无感硬件确实能够捕捉一整天的上下文、承诺与待办事项，尽管社会必须重新适应由此带来的隐私边界——比如首府华盛顿的部分餐厅已经开始没收手机，未来可能连吊坠也会一并收走。 David 提出了一个犀利的观点：Muse 吊坠的核心或许根本不是为了赢得硬件之争，而是 Meta 为了在现实物理世界中收集多模态数据的抓手。它配备了摄像头和多个麦克风，即便只被推特科技圈的小众核心用户日常佩戴，源源不断回传的高清现实世界数据也是极其惊人的。针对 Anish 关于“智能眼镜是否已经具备该功能”的疑问，David 指出眼镜不可能做到全天候常开，而吊坠则是 24 小时待命的环境伴侣。 随后，Anish 表达了对“纯音频眼镜”的高度看好，认为它是一匹黑马。这让人联想到科幻小说《安德的游戏》（Ender's Game）中的全知伴侣 Jane：既享受了随身 AI 的连接能力，又免去了在脸上架设摄像头可能给周围人带来的社交尴尬与不适感。 > “I have a hot take thesis that this Muse charm is actually less about trying to win the hardware game. And it's more about data collection in the real world for the Meta team in terms of how the real world actually functions...” —— David Pawlan --- ### `18:00` 骑车实现收件箱归零：语音 Full-Duplex 带来的“魔法时刻” 在所有交互方式中，David 对语音和音频表现出了极大的信心。他兴奋地分享了自己当天早上的亲身体验：ChatGPT Voice 新上线了连接 Gmail 和 Google Calendar 的功能，这构成了他在 AI 助手领域的第二个“魔法时刻”。 David 在早上骑自行车通勤的 30 分钟里，双手必须握把，无法触碰手机屏幕。他直接打开 ChatGPT Voice 开启对话会话。在骑行途中，Agent 顺畅地帮他分类了电子邮件、打上标签、起草并发送了回复邮件，甚至还发出了日历日程邀请。当他骑车抵达办公桌前时，原本堆积如山的邮箱竟然已经实现了“收件箱归零”（Inbox Zero），整个体验如同通电般令人震撼。 Anish 解释说，这套体验基于 GPT Live 1 模型（具备全双工 full-duplex 能力），能够打断、实时插话，并全面掌握用户的对话历史与上下文，因此用户随时能追问之前搁置的项目进度以及当前的阻碍。最后，两人总结道：对于普通大众而言，个人助手最不可或缺的时刻，恰恰是人们双手被占用的时候——比如在做园艺、烹饪或通勤时，用户根本无法操作屏幕。那些整天挂在电脑前、沉溺于网络世界的科技圈小众群体，其使用习惯与广大主流大众存在着根本性的错位。 > “Categorized my emails, submitted them to different labels, replied to different emails, sent out calendar invites, got to the desk, inbox zero. It was so electric.” —— David Pawlan > “...for the general consumer, an assistant is most helpful when you are most occupied. It's not like you're going to be sitting down on your couch and spending all of your time using your assistant to do stuff. It's most helpful when you're gardening or cooking or doing things where your hands are busy and you can't actually engage with an interface.” —— David Pawlan ### `20:00` 多人协作与群聊中的 AI 智能体探索 谈到 AI 智能体如何走向大众普及，David Pawlan 认为真正的切入点在于解决用户切身体会的极端痛点。例如吃到违章罚单这种人人都讨厌的繁琐杂事，用户只需随手拍一张违章停车罚单的照片，交给智能体代为缴纳即可。随着讨论转向社交场景和多人（multiplayer）互动，Anish Acharya 提出了一个尖锐的痛点：虽然 iMessage 为智能体提供了极佳的特权入口（例如 Instinct 的做法），但在实际体验中，把一个智能体直接拉进好友或家庭群聊，往往会让人感到被冒犯或入侵，甚至削弱了原本的社交资本（diminishes social capital）。 针对智能体在多人社交场景中的落地，David Pawlan 梳理了业界目前尝试的三种主要模式： ```mermaid flowchart TD accTitle: 三种多人群聊智能体交互模式 accDescr: 展示传统群聊模式、智能体后台通信网络模式以及静默监听者笔记模式的对比。 subgraph Mode1["模式一：传统群聊 Companion"] M1["智能体作为普通成员进群"] -->|"直接插话、公开回复"| U1["群内所有人类用户"] end subgraph Mode2["模式二：智能体后台网络 (Instinct)"] U2A["用户 A 的智能体"] <-->|"后台静默握手与协调"| U2B["用户 B 的智能体"] U2A -.->|"私下汇报结果"| UA["用户 A"] end subgraph Mode3["模式三：静默监听者 (Doc / XMTP)"] Doc["智能体静默旁听群聊"] -->|"实时记录并整理成文档"| DocLib["共享在线文档"] Doc -->|"涉及特定任务时私聊触发"| Indiv["独立私信相关责任人"] end ``` David 特别推崇 XMTP 生态中由 Doc（类似会议工具 Granola 的机制）所采用的第三种模式。在这种设计下，智能体进群后全程保持静默，只充当“隐形记录员”。它在后台记录对话要点，如果提炼出待办事项（action item），它绝不在大群里刷屏，而是私下单独给对应的人发消息确认；其他成员如果想看讨论脉络，直接查阅生成的文档即可。David 称赞这才是真正无感的“隐形智能体”（invisible agent），既干完了行政杂务，又完全不打扰群里的大多数人。 引用 Obo 创始人 Nir 的观察，Anish Acharya 指出当前基础模型在代码编写等具有可验证反馈（verifiable domains）的领域呈现出指数级增长，但在散文文学质量以及人情世故、社交分寸感（bedside manner）这类微妙维度上，进展却陷入停滞甚至有所下滑。正如 Anish 所断言：“我认为凭借当今的模型水平，再加上我们自身对这些社交动态的预期，智能体并没有太多途径能够真正促进或增强群体的社交连接。”Anish 分享了自己作为黑胶唱片收藏家的实验：他尝试把智能体引入同好收藏群，试图归档大家的藏品与审美偏好，虽然这一尝试尚未完全成熟，但这种偏重实用主义（utilitarian）的工具属性，远比一个低情商（low EQ）、到处乱插话的喋喋不休机器人要顺眼得多。 - **群聊话语权的心理机制**：Anish 提到本周听到的一种见解，即即时通信（Messaging）有着天然的排他性，“就像手里握着麦克风一样”，让机器人握着麦克风会让人极其尴尬不适；反而在类似 Discord、Reddit 或传统论坛这种发帖式的异步社区中，智能体以发帖者身份出现就自然得多。 - **人设形象的视觉鸿沟**：David 和 Anish 都认为当前产品有过度“拟人化”（humanize）的倾向。一些软件强行生成 AI 人类肖像反而引起用户的不适与排斥；相反，Muse 将智能体塑造成一只惹人喜爱的雪人（adorable Yeti），反而赢得了大众的喜爱。在用户的私密空间里，过于逼真的假人会让人感到冒犯，可爱的非人吉祥物反而更容易建立情感防线。 --- ### `23:45` 智能体性格、预判度（Presumptuousness）与主动性边界 关于智能体是否会像医生一样出现鲜明的原型（archetype）细分——比如有人喜欢严肃威严、说一不二的权威型医生，有人则偏好温暖友善、宛如平辈的朋友型医生——David 认为沟通风格确实因人而异。以 Poke 为例，它凭借鲜明且极具攻击性、爱挖苦（sassy）的性格一炮而红，但也有不少用户因为它的刻薄而拒绝将个人隐私托付给它。不过在 David 看来，性格风格本身并不是坚固的护城河，因为只要通过修改提示词、记忆模块或“灵魂文件”（soul file），就可以随时配置并切换语气。未来占统治地位的通用智能体必然高度可定制，用户可以自由调教。 Anish 则提出了更深层的一面：比表层语气更关键的是智能体的“宪章底层特质”（constitution）或决策倾向，其中最典型的指标就是**“自作主张度 / 预判度”（presumptuousness）**。在很多跑腿任务中，用户巴不得智能体极其专断，就算有 10% 的概率犯错，只要能省心办成事也无所谓；但在财务、健康等敏感领域，用户又要求它极度谨慎。 ```mermaid flowchart LR accTitle: 智能体主动行动的风险与授权决策边界 accDescr: 区分无需确认的纯收益行动与必须事前授权的高影响行动。 A["智能体发现优化机会"] --> B{"是否需要用户承受<br/>状态变更或直接后果？"} B -->|"否：纯收益、无破坏性"| C["高度主动执行 (Proactive)"] C --> C1["起草好邮件草稿"] C --> C2["自动申请航班退款/积分"] C --> C3["提前办理航班登机"] B -->|"是：涉及合同/金钱/人际"| D["谨慎征询确认 (Permission Required)"] D --> D1["发现车险过贵建议更换新保险"] D --> D2["涉及人际关系的重大沟通"] ``` David 指出，这种主动性（proactivity）才是真正拉开产品差距的决胜手：“我认为主动性拥有巨大的护城河优势；我认为这正是把最终赢家与平庸之辈区分开来的关键所在。”目前最惊艳的“顿悟时刻”（aha moment）莫过于智能体主动帮你办好值机手续。但 David 警告，主动性存在一条极其危险的红线： - **无须授权的安全场景**：例如默默草拟好邮件、帮用户自动争取到航司积分退款（flight credit）。这些举动即便做过了头，也没有产生不可逆的负担，纯粹在为用户省钱省时间。 - **必须确认的高风险场景**：例如智能体检测到现有车险太贵、想自作主张帮你换一家保险公司。这种涉及直接外部合同与个人义务的操作，一旦擅作主张且跨过信任红线，哪怕只错一次，就会瞬间输光用户的信任。 Anish 甚至调侃道，团队内部私下开玩笑说，按照这个演进速度，“再过几天我们就会看到智能体给主人发短信说：‘我注意到你其实没那么喜欢她，所以我自作主张帮你把她甩了’”。尽管如此，Anish 坚信魔力恰恰诞生于这种“擅作主张”之中；如果网上没人讨论智能体又越界干了什么惊天动地的事，反倒说明现在的开发者在主动性探索上还不够激进。 - **Bong Chang 登机幻觉事件**：网上曾热传一条推文，称 Instinct 在帮某用户值机时，幻觉编造了一个中间名“Bong Chang”，导致该用户直接无法登机。Anish 怀疑这可能只是推特上的搞笑段子（shitpost），但无论真假，这种极具戏剧性的错误不仅没有劝退核心用户，反而成了 X（Twitter）上引爆流量的绝佳谈资。 --- ### `28:10` 从 OpenClaw 到消费级普及的产品化蜕变 随着话题回到底层技术演进，Anish 指出当前智能体浪潮的核心脉络，在于过去“OpenClaw 时代”所积累的一整套底层逻辑和操作原语（primitives），如今终于被打包成了普通大众能够轻松消化和享受的成熟消费级产品。在此过程中，早期的先驱产品（如 Poke）展现出了非凡的巧思与灵气，可惜可能只是入局时间稍早了半个身位，否则在被 Cognition 收购前本有机会问鼎这一赛道的桂冠。 当 Anish 询问当下的消费级智能体究竟从那个充满极客极度折腾的 OpenClaw 时代汲取了哪些预示未来的关键遗产时，David Pawlan 给出了极其务实甚至略显清醒的评价：“老实说，在当下的阶段，我认为消费级智能体本质上就是 OpenClaw 的纯粹复制品，只不过进行了预先配置。”David 解释说，现在的个人助手在能力边界上其实并未超越当年的 OpenClaw 或 Hermes，最大的革命完全发生在工程化体验层面——用户不再需要自己去配置复杂的 API 密钥与运行环境，不必面对每隔七个小时就弹出一堆报错代码的崩溃局面，而是拥有了一个极其易用、优雅且即开即用的交互界面。当硬核工具的门槛被彻底抹平，真正的消费级普及才算拉开了序幕。 ### `30:00` 垂类初创公司与“Narrow Startups”投资假说 对话从个人 AI 的演进路径展开。David Pawlan 指出，整个领域正在不可阻挡地走向“主动性”（proactivity）与“超级专业化”（hyper specialization），这与 Anish Acharya 一直倡导的“窄门初创”（narrow startups）假说高度契合。David 举例设想了一种极端长尾的应用场景：团队不去做泛用型通用助手，而是将所有精力倾注在服务“家有 3 岁以下幼儿的单身母亲”这一个群体上；该助手通晓母亲日常运转的方方面面及儿童发育规律，能够精准预测需求，宛如拥有“读心术”一般。这种高度垂直的专家代理，未来甚至可以作为插件反向接入通用助手体系。 针对这一假说，Anish 深入阐释了其背后的商业逻辑与护城河构成。他认为软件开发范式已被彻底重塑，现在团队可以用极高野心为极小圈层打造无可替代的高价值产品。在消费者已普遍接受每月 200、250 甚至 300 美元订阅费的当下，企业仅靠数万名高黏性付费用户，就能撑起数千万乃至上亿美元年经常性收入（ARR）的健康业务，这在以往的技术周期中是前所未有的。 ```mermaid flowchart LR accTitle: 垂类代理壁垒与商业模型 accDescr: 展示垂直 AI 代理如何通过专有数据与品味建立壁垒，并以高客单价服务小众群体跑通单位经济。 subgraph Defense["核心壁垒 (Defensibility)"] A["独特品味与理念<br/>(Taste & School of Thought)"] B["专有场景数据<br/>(Proprietary Data)"] end subgraph Product["垂类 AI 代理"] C["高度专业化 Agent<br/>(例如：母婴/纽约本地向导)"] end subgraph Business["商业变现模式"] D["小规模高净值客群<br/>(数万名垂直用户)"] E["高客单价订阅<br/>($200 - $300 / 月)"] F["规模化 ARR 业务"] end A --> C B --> C C --> D D --> E E --> F ``` Anish 强调，垂直代理的核心比较优势源于“品味（taste）”与“专有知识（proprietary knowledge）”的结合： - **流派与主观品味**：就像孕妇挑选导乐（doula）会看重对方推崇的助产流派一样，未来的用户也会根据特定的哲学派系、行事风格或冷门专业经验来挑选个人代理。 - **专有场景认知**：类似于一位深谙纽约暗灯小馆、品味无可替代的顶级本地礼宾员（concierge），面对不同画像的人群，客观与主观层面的垂直细分都会蓬勃生长。 - 关于窄门商业的颠覆性，Anish Acharya 总结道：“The theory with narrow startups is that you can now build software that's incredibly ambitious and incredibly valuable for a very small number of people.”（关于窄门初创的理论在于，你现在可以为极少数人构建出极具野心且价值极其巨大的软件。） --- ### `32:45` Assistant 还是 Agent？界定主动权与社会角色 围绕科技圈与推特上持续争论的术语界限——究竟应该叫“个人助理”（Personal Assistant）还是“个人代理”（Personal Agent），两人展开了语义与本体论层面的辨析。Anish 明确表示，在他眼中“助理”属于“代理”的低阶形态：助理在定义上只能依照指令行事，本质是在辅助人类完成直接交代的事项；而代理则拥有内在的能动性（agency），能够主动促成现实世界中好事的发生。 这种演进轨迹直接呼应了企业级软件中 AI 的成长史：AI 最初入职时就像实习生，在证明能力后被逐步晋升去接管更高阶、更复杂的业务。同样地，在消费者的个人生活中，软件也在经历从被动助理到拥有更大自主裁量权的代理的跨越。David 顺势提出深层的社会学思考：称其为“Agent”是否赋予了恰到好处的拟人化？而“Assistant”是否带入了过多的人类从属属性，进而引发社会层面的心理不适？两人笑谈起地域认知差异：处于全球科技中心的旧金山身处最前沿的泡沫，纽约稍显滞后，全球其他地方则在随后跟进；但 Anish 纠正道，这不是泡沫，而是人类提前踏入了“近未来”。 更微妙的社会学效应在于代理可能充当人际交往中的“社交间接缓冲层”（social indirection）： - **化解正面社交尴尬**：代理作为非人类实体，可以作为一个没有情绪波动的缓冲区，代为转达群组中令人不适或尴尬的沟通需求。 - **非被动攻击的中立仲裁**：Anish 强调这并非倡导被动攻击型（passive-aggressive）的处事方式，而是提供了一个理性的社交出入口，帮助人们在充满张力的社交互动中降温。 - 关于助理与代理的本质划分，Anish Acharya 给出定论：“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.”（助理在定义上只能按你的吩咐做事，辅助你完成指定的工作；而代理拥有自主能动性，能够主动促成现实中好事的发生。） --- ### `35:20` 终极愿景：消解官僚重负与“可能性扩展” 放眼未来五年，AI 究竟如何让人类变得更加幸福？Anish 指出，技术的首要价值是卸下压在普通人身上的“日常官僚行政压力”（bureaucratic pressure）。他举出一个惊人的现实数据：全美每年有多达 150 万人因酒后驾驶（DUI）被捕，后续伴随着海量窒息的法律文书、行政审批与繁琐流程，而这当中绝大多数普通人根本负担不起专业律师团队。AI 的第一阶段，就是要让平民以极低门槛享受到以往富豪家族办公室（family office）或私人律师级别的统筹优化支持，把人们在财务、健康与日常行政中的精力损耗降到最低。 而在更深的精神维度，Anish 搬出了母亲的人生格言：“想要获得你想要的东西，最难的部分在于先弄清楚你究竟想要什么。”许多人终其一生都在探索自我认同、追求成为最真实的自己，个人代理与人类主观体验的深度交互，将成为驱动人类自我发展（self-development）的绝佳工具。David 对此深表赞同，他勾勒了一个极具画面感的未来场景：未来下一代看着如今街头行人的历史录像，一定会不可思议地质问“为什么你们整天低着头、伸长脖子死死盯着一块发光的玻璃屏幕？”——现实世界如此美妙，当无休止的“生活杂务”（life admin work）被彻底接管，父母才能真正陪伴孩子成长，二十多岁的年轻人才能全心投入深度的现实友谊中。 ```mermaid flowchart TD accTitle: AI 核心驱动从成本削减转向可能性扩展 accDescr: 展现 AI 如何接管行政琐事，将软件价值由纯粹的降本提效跃升至人类探索全新可能性的扩展。 subgraph AdminLayer["第一阶段：消除生活负荷"] A1["琐碎行政 (Life Admin)"] A2["法律与官僚审批 (如 DUI 等文书)"] A3["健康与财务日常优化"] A4["自动化执行与减负"] A1 & A2 & A3 --> A4 end subgraph ExpansionLayer["第二阶段：可能性扩展 (Possibility Expansion)"] B1["低门槛开发长尾创意项目"] B2["摆脱屏幕，回归真实人际联结"] B3["探索全新生活形态与自我认同"] end A4 -->|"释放人类时间与认知带宽"| ExpansionLayer ``` 谈及此处的底层逻辑，Anish 以编程代理（coding agents）为例打破了外界对“AI 抢夺人类岗位”的零和博弈偏见。他透露自己近期正在借助编码代理搭建一个可以真实走入逛店的“虚拟唱片店”（virtual record shop）Side Project。这类项目在以往根本不可能立项——他绝不可能在五年前自掏腰包雇佣专业程序员耗时一年去写代码，自己更不会耗费一年生命去手搓。编程代理所做的，根本不是在抢人类程序员的既有饭碗，而是在开拓无人涉足的新疆界。 - 关于技术的底层驱动力，Anish Acharya 总结道：“It's not about cost reduction. It's about possibility expansion.”（技术的真谛不在于削减成本，而在于扩展人类的可能性。） --- ### `38:20` Agent-Native 基础设施：机器对机器（A2A）的全新网络 话题最后切入当下正在浮现的近未来基建。结合 Moldbook 的启示以及 Instinct 正在探索的“网络互联（Instinct-to-Instinct Network）”，两人一致认为，纯粹服务于“代理对代理”（Agent-to-Agent，简称 A2A）交互的原生系统正在形成一个全新的计算范式。 David 激动地指出，这足以展开一场长达 10 小时的行业深剖： - **代理专属身份与通讯层**：未来每个代理都需要拥有专属的电子邮件地址（Agent Email）与电话号码（Agent Phone Number），以合规合法的实体身份穿梭于数字世界。 - **服务业链路的根本性颠覆**：商业交易链路正在从经典的“人对人（H2H）预订”，演变到前期的“代理帮人向人（A2H）预订”，并最终彻底洗牌为“代理直接向代理（A2A）预订”。服务业的撮合机制、定价逻辑与接口标准将被全盘重塑。 - **代理原生网络安全（Cybersecurity）**：历史正在按经典规律重演——互联网诞生之后孕育了现代网络安全产业；如今代理生态率先爆发，随之而来的必然是针对代理身份鉴权、越权控制与对抗欺诈的全新安全蓝海。 无论是重构服务业、重写商业交互逻辑，还是攻克全新的网络安全盲区，伴随这些极其垂直且未被开垦的基础设施痛点，新一代代理原生初创企业正在迎来爆发窗口。 ### `40:00` 代理商业的冲击：Shopify 的拥抱与亚马逊的封锁 Anish Acharya 提出了一个关于互联网构成的宏观视角，他认为网络内部实际上存在着三个不同的维度：“就像我内心住着两只狼一样，我认为网络是由三种形态构成的：商业网（commerce web）、应用网（application web）和内容网（content web）。”在 Anish 看来，内容网已经陷入困境多时，而自从编程智能体（coding agents）诞生后，应用网迎来了爆发式增长。至于商业网，其当下的博弈同样引人注目。最近最具代表性的事件莫过于 Shopify 选择全面拥抱 Muse 这一类 AI 代理，而亚马逊（Amazon）却坚决屏蔽了 Muse。这种应对策略的截然相反，揭示了随着新的流量入口出现，电商平台在利益重新分配中的阵营分化。 David Pawlan 指出，这两家巨头的决策差异根植于底层商业模式的根本冲突。Shopify 的目标是让商业民主化，赋能每个普通商家卖出更多商品，因此让代理进入平台完成购买是极具优势的；而亚马逊的核心利润引擎是庞大的广告业务。David 分析道：“整个商业的广告架构接下来将变得非常引人入胜，因为当你移除了人类的眼球，这就击溃了整个系统的根本目的。”如果任由 Muse 等代理绕过商品展示广告直接下单，亚马逊将失去最主要的利润来源。因此，在代理代为采购的新世界里，亚马逊势必需要建立一种全新的盈利范式。 ```mermaid flowchart LR accTitle: 对比 Shopify 与亚马逊应对 AI 代理的策略 accDescr: 图表展示了 Shopify 与亚马逊在商业模式及对待 AI 代理态度上的核心分歧。 subgraph Shopify["Shopify 模式"] S_Model["商业民主化 / 交易赋能"] --> S_Action["全面拥抱 AI 代理"] S_Action --> S_Outcome["降低交易门槛，提高销售总额"] end subgraph Amazon["亚马逊模式"] A_Model["高度依赖搜索与展示广告收入"] --> A_Action["全面封锁 AI 代理"] A_Action --> A_Outcome["保护人类眼球注意力与广告利润"] end ``` 这种平台防御正在向更多未被充分颠覆的领域蔓延，尤其是像餐厅预订这样的线下服务预订市场。David 提到，Resy 已经开始大规模封禁使用 Browser Use 自动化预订的用户账户，因为机器人代订打破了原本为人类设计的预订秩序。这引发了一个新的思考：当人人手握一个代理时，所有人都处于同一起跑线、在同一时刻蜂拥抢订，最终会变成代理之间的竞价大战，还是会让分配权力重新回到餐厅手中？ Anish 强调，在供给受限（如热门餐厅餐位）与需求受限（如临时飞往旧金山的旅客）的不同场景下，博弈机制截然不同。对于追求利润的餐厅而言，完美信息下的最优策略是筛选出客单价（AOV）和生命周期价值（LTV）最高的顾客，并维持近乎 100% 的上座率。这远比靠人类拼网速的 DDoS 式抢位要复杂得多。不仅如此，代理的普及可能催生反向拍卖机制：一群买家表达需求后，由供给端的商家主动竞价看谁能以最低价格提供最佳服务。谈及消费心理，Anish 认为亚马逊可能也是在担心失去“冲动消费”（impulse shopping），但代理本身由于追求前瞻性与惊喜感，并不完全理性，营销与冲动消费不会消失，而是会以全新的形式被重塑。 - Shopify 与亚马逊呈现鲜明对照：前者致力于交易赋能，后者严重依赖人类眼球带来的广告收益。 - Resy 开始封禁基于浏览器自动化（如 Browser Use）的订座行为，以维护其以人类交互为核心的业务秩序。 - 供给受限市场（如顶级餐厅预订）或将促使商家转向依据客户终身价值（LTV）与客单价（AOV）的数据驱动型筛选体系。 - 代理经济或催生“供给侧反向拍卖”：多个买家聚合需求，商户主动竞标以提供最优性价比体验。 ### `45:00` 推荐引擎重塑：从买白袜子到“代理对代理”的意外之喜 随着商业入口由搜索框转为对话式代理，传统的商品推荐引擎正面临全面重构。David 指出，大语言模型某种程度上“玩坏”了互联网的优质搜索体验。当用户仅仅对代理说一句“帮我买双白袜子”时，市面上存在着数万种不同的白袜子，代理根本无从得知用户的真实偏好。在这个维度上，像 Meta 的 Muse 或许具有先天优势，因为其坐拥用户的 Instagram 和 Facebook 数据，对真实品味了如指掌。 针对这一问题，Anish 构想了一种截然相反、充满意外惊喜（serendipity）的长尾电商形态。代理不再只是机械地抓取亚马逊销量第一的标品，而是可能挖掘出内布拉斯加州某位老奶奶在 Etsy 上手工编织的高品质白袜子。Anish 兴奋地展望，随着技术向更具未来感的发展阶段迈进，这种交互将彻底打通“代理对代理”（Agent-to-Agent）的真实世界协作网络。例如，买家的代理直接去给那位编织毛袜的老奶奶的代理发短信，询问对方是否愿意接单制造。David 笑着打趣道：“我现在脑子里就在想象，一个内布拉斯加州的老奶奶收到一条短信说：‘嘿，旧金山有个家伙想让你给他织双白袜子，你愿意干吗？’” Anish 强调，这种基于真实产品质量而非营销操纵的机制，让激励体系重新回归正轨。代理凭借自身的评估能力去筛选好产品，即使用户坚持选择 Nike 等大品牌，也是基于明晰的个人偏好，而非被认知失调式的洗脑营销所诱导。对此，David 总结道：“它打开了一扇通往真正与消费者利益一致的互联网大门，而这恰恰是我们已经缺失了相当长一段时间的东西。” ### `48:00` 智能体的经济账：免费巨头围堵下的创业公司生存之道 谈及 AI 代理的商业化现状，David 分享了他在助理基准测试（Assistant Benchmark）中梳理出的第一手调研数据。在他追踪的 122 个代理项目中，他亲自深度测试了 26 个。在整个样本池中，商业模式的分布呈现出向收费端倾斜的格局：在这 122 个代理中，多达 65 个是付费的（其中 35 个必须全额付费使用，30 个采用达到用量上限后收费的免费增值 Freemium 模式），而真正完全免费的仅有 13 个。面对 Muse 和 Instinct 这类背靠巨头且完全免费的头部玩家，以及传闻中 OpenAI 即将推出的重量级产品，David 发出疑问：初创企业在面对巨头免费赠送的挤压下，究竟该如何突围？ ```mermaid flowchart TD accTitle: 122 个 AI 代理商业模式分布结构 accDescr: 展示 David Pawlan 调研的 122 个 AI 代理项目中，全额付费、免费增值模式以及完全免费工具的具体比例分布。 Total["122 个 AI 代理项目调研"] --> PaidAll["完全付费模式\n35 款"] Total --> Freemium["免费增值 Freemium\n30 款"] Total --> Free["完全免费模式\n13 款"] Total --> Others["其他混合/未明确模式\n44 款"] PaidAll -.-> SummaryPaid["付费类工具共计 65 款"] Freemium -.-> SummaryPaid ``` Anish 算了一笔残酷的成本账，揭示了初创公司的艰难处境：“根据我们的估算，以一种极具野心的方式运行这项业务，每位用户每天的成本大约高达 20 美元，这对一家初创公司而言每年可能意味着数亿美元的支出。”他认为，创业者目前主要寄希望于两个积极趋势：一是底层浏览器自动化技术（Browser Use）的成本通缩，Token 价格和推理成本在以指数级迅速暴跌；二是产品本身的价值壁垒。 然而，Anish 明确反对将单纯的“客户补贴”作为核心价值主张。他更推崇能提供极致价值的产品：“我希望能看到那种每月收费高达 1000 美元，但消费者依然极度兴奋并心甘情愿买单的代理。”这类代理服务的往往不是基础的刚性需求，而是人类更深层的强烈渴望；唯有这种产品，才能在享受算力成本剧烈通缩红利的同时，真正凭借极高水准的“极致市场契合度”（extreme market fit）建立起不可替代的壁垒。 - 在被统计的 122 个 AI 代理产品中，共有 65 个产品包含付费门槛（35 个全付费，30 个 Freemium），完全免费的产品仅有 13 个。 - 按照顶级标准运作个人代理，目前单用户每日的推理与运行成本约达 20 美元，年化开销可达数亿美元规模。 - Anish 警示将资本补贴当成主要卖点的危险性，看好每月可定价高达 1000 美元、能满足高级欲望并具备极致产品市场匹配（PMF）的代理方案。 ### `50:00` 尾声：每周巨变的代理赛道与免责声明 在对谈的尾声，Anish 对 David Pawlan 及其发起的 Assistant Benchmark 项目表达了感谢，并感慨个人 AI 代理领域日新月异的发展节奏：“我觉得我们甚至可以每周都聊一次，老兄，因为每天都有全新的事情发生。”两位主持人呼吁观众关注他们在 X（原 Twitter）上的账号，随时跟进前沿动态，共同在社区中观察这一波浪潮的演进。 播客最后附带了标准的法律声明：本节目内容仅供参考与信息交流，不构成任何法律、商业、税务或投资建议；节目讨论内容不得用于评估任何投资或证券，亦不针对 a16z 旗下任何基金的现有或潜在投资者。同时，a16z 及其关联方可能在节目中提及的某些公司中持有投资头寸，相关投资组合的完整披露链接可在 a16z 官方网站的合规披露页面查询。 ## Notable quotes > `35:20` It's not about cost reduction. It's about possibility expansion. — Anish Acharya > `40:00` 整个商业的广告架构接下来将变得非常引人入胜，因为当你移除了人类的眼球，这就击溃了整个系统的根本目的。 — David Pawlan > `32:45` 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. — Anish Acharya > `30:00` The theory with narrow startups is that you can now build software that's incredibly ambitious and incredibly valuable for a very small number of people. — Anish Acharya > `48:00` 我希望能看到那种每月收费高达 1000 美元，但消费者依然极度兴奋并心甘情愿买单的代理。 — Anish Acharya > `10:00` The best employee I've ever had are the ones that are just doing their work, getting things done. They update you. Here's what I did. And you're like, wow, amazing. You don't hire someone because you enjoy managing them. — David Pawlan ## Mentioned Anish Acharya, David Pawlan, Assistant Benchmark, ChatGPT, ChatGPT Voice, OpenAI, Claude Code, CodeX, Instinct, Muse, Poke, OpenClaw, GrokBot, Caddy, Ollie, Pally, Season, Town, Catch, Vellum, Sky, Mark Zuckerberg, Meta, Hermes, Granola, Doc, XMTP, Obo, Nir, Cognition, Shopify, Amazon, Resy, Browser Use, Etsy, Ben Thompson, TBPN, Ender's Game ## 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. This episode was recorded on September 24, 2026. [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/db7a374f-e7e8-4561-875a-fde28e631933/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=db7a374f-e7e8-4561-875a-fde28e631933&feed=JGE3yC0V)

[35:20] Anish Acharya: It's not about cost reduction. It's about possibility expansion. [40:00] David Pawlan: 整个商业的广告架构接下来将变得非常引人入胜，因为当你移除了人类的眼球，这就击溃了整个系统的根本目的。 [32:45] 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. [30:00] Anish Acharya: The theory with narrow startups is that you can now build software that's incredibly ambitious and incredibly valuable for a very small number of people. [48:00] Anish Acharya: 我希望能看到那种每月收费高达 1000 美元，但消费者依然极度兴奋并心甘情愿买单的代理。 [10:00] David Pawlan: The best employee I've ever had are the ones that are just doing their work, getting things done. They update you. Here's what I did. And you're like, wow, amazing. You don't hire someone because you enjoy managing them.

[40:00] David Pawlan: The whole ad structure of commerce is going to be really fascinating to watch because when you remove human eyeballs, it defeats the entire purpose of the system. [25:00] David Pawlan: I think there is massive defensibility around proactivity. I think that's actually going to be something that separates the winner from the meat of the pack. [40:00] Anish Acharya: Amazon probably is rightly taking the bet that they don't want to be disintermediated and that they'll lose not just advertising revenue, but impulse shopping. [30:00] Anish Acharya: The theory with narrow startups is that you can now build software that's incredibly ambitious and incredibly valuable for a very small number of people. [32:32] 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. [45:00] Anish Acharya: I'd love to see the agent that costs $1,000 a month that the consumer is so excited about that they're willing to pay.

[0:00] The general population does not care about being 10% more efficient. I have a hot take thesis that this muse charm is actually less about trying to win the hardware game, and it's more about data collection in the real world to fuel Zuckerberg's future metaverse of mapping out the actual world. We were joking internally, like we're days away from an agent messaging someone saying, I noticed you weren't that into her, so I went ahead and broke up with her. There is massive defensibility around proactivity. start talking to ChatGBT Voice, and throughout my 30-minute bike ride to work, categorized my emails, submitted them to different labels, sent out calendar invites, got to the desk in box zero, but it will be a very fine line because If you cross that line once, you lose all trust with your user. It does feel like there's going to be infrastructure and products that exist only for agent-to-agent interactions that doesn't exist today. What do you think? It is inevitable that we're going to see personal AI agents have gone from experiment to one of the fastest moving areas in consumer AI. But what would it actually take for one to become part of everyday life? In this episode, A16Z's Anishacharya sits down with assistant benchmark creator David Pollitt, who has been testing dozens of personal agents and financial admin. They discuss why the best agent might be the one you barely notice, proactively handling the tedious parts of life. without needing to be managed. They also get into how much autonomy will actually give these systems, whether we'll interact with them through text, voice, apps, or wearables, and where specialized agents could still beat the big general purpose players. And looking further ahead, They explore what happens when agents start transacting with other agents, reshaping everything from commerce and restaurant reservations to how we spend our time online. part of the ecosystem has sort of seen God. And ChatGPT was a big one, I think, for a lot of folks in November 22 and a bit of 23. It was like, what is this thing that can write emails and poems and start to have a conversation back and forth and feel like you're interacting with the synthetic person? That was, of course, the beginning, the big bang. The second moment really was coding agents and everything that developers have been obsessed with around sort of quad code and codecs and recently with Instinct, Muse, and ChatGPT work, we've had this extraordinary enthusiasm around personal agents and what they can do for us as consumers. So we're here to talk about all things personal agents. We're going to go as deep as we can. We're going to assume that everybody has seen many of the same things that we have seen, and we're going to make this a first in a series of conversations. Welcome, David. Thank you. Thanks for having me. Excited to be here. Let's like zoom out and maybe talk about the arc of personal agents over the last few months, the last few weeks. What have you been seeing? What's your sort of view on things? Yeah, so this space has exploded over the past four weeks. We can roll it back. I was an early adopter of Poke. So I want to say Poke, which was kind of one of the first consumer agents to really hit the market. That launched, I believe it was September 8th. And then I started using it September 11th, three days later. And I remember it so vividly because it was just an unbelievable experience. Their whole gimmick was this negotiation with the agent on how much you're open claw gets released and that kind of takes the actual world by storm of like there is something coming. And then a few other have popped up here and there, but most recently, we saw Instinct really blow up the tech Twitter bubble. Just instantly. absolutely wrote a wave, went nuts. And then it was just dominoes. You have Grokbot, you have Muse, you have a bunch of long tail ones like Caddy or Ollie or Pally. And It's really been amazing to see this whole tech bubble is just exploding with interest and an intellectual curiosity for what can these agents do for us on a day-to-day basis. I think everyone's kind of in the same boat where... Nobody really knows. It's the wild, wild west. We're just playing around here. I totally agree. And we should take a moment to talk about Assistant Bench, which is this new, really important and discussed benchmark that you're the author of. Maybe tee up Assistant Bench a little bit and I'll ask you a few questions. one-shot problem. [5:00] Does it actually work? Does it ask follow-up questions? How does it perform across 16 different dimensions? and we compare those also. We use the word benchmark. Think of it more as like a consumer facing benchmark, not necessarily technical behind the scenes. But if I want to book a flight, which one should I actually use? Who is performing the best? How fast are they actually replying? Got into this whole comparison because I was just doing market research myself, trying to figure out how are these all comparing against one another? What are the gaps? And published a bit of that. And it was almost two weeks ago. It's 16 days ago today that we and visitors have had outreach from basically every single founder. And I think it really just goes to show like, Everyone's curious. People are looking for answers to where this industry is moving. And so talk us through, David. That's incredibly exciting. It's cool, actually. It's funny how benchmarks have become such a staple now of the conversation. So there's so many launches, it's hard to try everything, which is also an exciting signal. Maybe talk us through everybody who's not instinct and muse. We're definitely going to deep dive on those later. But who else do you think is doing interesting work? And what are the kind of directions of specialization that you're seeing? Totally. So we've got a bunch of long tail solutions out there. I think I would categorize it into two groups. You have the B2C consumer facing generalized agents. That's like the Instinct or Muse. Some of the longer tail ones, you have Caddy, you have Ollie, you have Season. There's, I believe, 64 right now, just in the general category on the site. And then we can break that down even further into a little more specialized within that. So you can have travel The other side of the spectrum is going to be more of like your B2B workflow type assistance. And those are anyone from the town to catch to vellum. And they're all essentially doing the same thing, right? They're your assistant that's actually... like more of like an executive assistant, not the consumer facing approach to it. And yeah, That's kind of like the general landscape at the moment. So far, to my knowledge, there are 122 different ones across all the categories. Incredible. And what about areas of concentration? Where is the most sort of interesting competitive focus? So everyone right now, by and large, is focusing on the general. Horizontal. Yep. Yep. Just trying to do everything. And when you think about something like travel, just to pick on that, that's an infrequent high value behavior. Do you think that that category ends up being organized more as a connector into the horizontal agent? Or do you think they have the right to win as a general horizontal agent as well? I think you're going to maybe see potentially one winner as an independent agent. But by and large, I really think it is going to be more of a connector. with just individuals obsessed with assistants talking. So there's now over 1,200 people across all the group chats just chatting together. over and over and over. Travel is the fourth most talked about use case, which is really interesting because when you look at like virality, it tends to be the number one that people are talking about. If it checked into my flight for me, I was able to book my flight. But then you actually look at the day-to-day conversations of what people are really using it for and most curious about. To your point, people aren't traveling every day. So it's more of a gimmick, right? It's an attention grab. What are use cases one to three? So number one is more just daily admin type stuff, whether it's cleaning out your inbox, whether it's filing forms for you, the things that are really annoying. And I think that gets to the broader thesis of where I think the winner of this space will go that we can touch on. The second is going to be more agent orchestration, which is really interesting. I think these group chats are a little biased as the third one is development focus. But people are talking a ton about agent orchestration and how do you find the most efficient agent or let your agents talk to each other. Development agents, how does this now translate to this coding world on your phone, whether it be using cloud code from a text component or even take it one step further of, okay, if I'm a designer, can I now make edits to my design files by just like texting my agent of like little small iteration changes? Yeah. And then the last category being finance. I think that's the most fun one to do. look into because it is, in my opinion, the least practical. There's so many decisions. And if you could have an agent make a billion dollars for you, all the hedge funds would be doing it. It's not something that everyone's going to succeed at with their own phone. But it is the most interesting that you give someone, an agent, a powerful tool right in their hands. And the first thing they think is, okay, how are you going to make me a million dollars? Yeah. It's really interesting because one of the critiques Ben Thompson was on TBPN a few days ago talking about how most consumers aren't looking for productivity in their lives. They're looking to spend time, not save time, which I agree with. Yeah. But it does feel like finance is an area in which there is a ton of administrative overhead for the average consumer. And if you look at the number of financial markets or financial. [10:00] sort of profit pools that are defined by consumer apathy or being uninformed. That's a lot of profit dollars that could be delivered back to the consumer, right? So I do think that the, you know, while I generally agree with Ben that most consumers are looking for entertainment and to spend time, I think there are sort of defining aspects of the U.S. consumer's life that are, you know, heavily administrative that these agents can really be a breakthrough for. Totally. And I think that that kind of touches on the general thesis to where I think the winners will sit. And that is that, like to Ben's point, I agree. I think the general population does not care about being 10% more efficient. Yeah. You hire an assistant to allow it to be invisible. The best employee I've ever had are the ones that are just doing their work, getting things done. They update you. Here's what I did. And you're like, wow, amazing. You don't hire someone because you enjoy managing them. Yeah. And, and, The best, most proactive use cases I'm seeing, I think, do fall in the finance category. And I think we're going to see an explosion in a world, though, where it's not cost makers, it's cost savers. And so it's going to be workflows where, like some cool use cases I saw, people are using it to go through their past year of receipts and file HSA reimbursements. credit. And I have a friend, this is my favorite use case I think I've seen yet. He hooked up his GrokBot to his sprinkler system at his home and connected it to the weather system. And so now his sprinklers are dependent on the weather and it's dropped his water bill by 50%. And so I think these invisible type things, agents that are going to be doing these cost-saving projects. type workflows fall within that finance category, but they're really going to lean into these cost-saving activities that create a world that is a lot more consumer-aligned. Yeah, it'll be interesting too. I think the American consumer, nobody wants to hear about saving money. People want to hear about spending money and free money. And I actually think the hook that's going to get a lot of people here is it's going to feel like free money. Going back through and submitting HSA requests. sort of refund requests is going to be free money. You know, getting all of your airline tickets where the price is dropped, you know, to refund you the delta is going to be free money. So I think that's a hugely powerful hook. It's also really, really interesting because this is a sort of infrequent but high value behavior. And that's where my head goes on both travel and finance. Like maybe these things will be really powerful, you know, billion dollar businesses that are still vertical connectors rather than general horizontal agents. Totally. It's going to be a very interesting world in terms of the application phase, which I think then... Also, something that we talked about last time we went for the walk was, okay, I'm here's what it can do, but now how do you actually interact with that agent? And it was something that we chatted on. I'm curious your thoughts. We both were in the world of... iMessage. It's where we live. It's easier for us. You mentioned your wife and her friends prefer a separate application. And so where do you see this world moving? Do you believe that multiple surfaces can exist as a winner or there's going to be one dominant lane? Yeah, it's a great question, man. I mean, I think this is why there's so much consumer preference that's going to matter here. Because I've heard the argument that for folks that are Gen Z, they're more interested in just texting. And for folks that are millennial or Gen X, fragment by generation. Part of it may actually fragment by gender as well to some extent. I mean, it's a generalization, but the types of things I'm doing versus the types of things my wife is doing, you know, she's big on building a dream board and trying to achieve all of her one-year goals and her lifelong goals. So she really prefers a visual interface. You know, she does a lot of sort of, travel is very visual driven for her, whereas for me, it's much more functional. Like, hey, Starlink on my plane, which works better for me. I feel like the iMessage sort of The iMessage space is more precious. So it feels like a more personal interaction when I'm in that space. And now, of course, that changes if every second thread is with an agent. But for now, I feel like it kind of lives in a privileged place on my phone where, you know, the app on the app grid just doesn't. What do you think? Yep. I completely agree. I think the... The personalization that comes with iMessage, just because I already live there, it feels like it's something that's already connected to me. It doesn't feel like an additive workflow. And I really have only, though, seen, by and large, two services appear. It's either the iMessage or a separate application. I've seen a third signal startup, actually, Sky. They have like an iPhone widget that they're trying to introduce. [15:00] There's a different surface which is Pretty interesting. But then even yesterday, right, we saw Zuckerberg and the Muse team dropped the Muse charm. Yeah, that's now an entirely different use case for for how we interact with the agent. What do you think about hardware? I think it's fascinating. I mean, a smart person said this, which is that every person who's working on a piece of hardware thinks that their hardware is the kind of be all end all. Like the pendant is the ultimate thing. You know, the watch strap is the ultimate thing. So I don't think that that's right. I think, again, it's going to be very personal preference driven, you know, just as jewelry is like some people like to wear necklaces and some people like to wear earrings. And I think the same way this kind of form factor is going to manifest is going to be very customer specific. And having something that is sort of with you and ambient and can capture all the kind of contacts and requests and, you know, promises and commitments and everything that happens through the day. I mean, I think we're going to have to sort of figure it out together in terms of privacy expectations and society. You know, you go to restaurants in Washington and they take away your phone. Maybe they're going to be taking away your phone and your pendant. Right. But I think that there's actually something there with hardware. And I just like that we're being ambitious in that direction. from working on hardware. Now we've got a lot of interesting new consumer efforts. Right. I think it's going to be super interesting. I have a hot take thesis that this Muse charm is actually less about trying to win the hardware game. And it's more about data collection in the real world for the meta team in terms of how the real world actually functions, given it has a camera and multiple microphones. And it's primarily, I don't really see it being a mass adoption thing. I think it's going to get adopted by the Twitter bubble and they're going to use it everywhere. And it's just going to feed meta with a ridiculous the actual world. But don't you think that they already get that from the glasses? I think they do to some extent, but... I don't know if, I mean, Correct me if I'm wrong, but are the glasses always going to be on in the same way that this charm will be? I don't think so. Yeah, it's a great point. I think the charm is meant to be more as, you know, as you mentioned, it's an ambient 24-7 companion versus the glasses are more action oriented. It's going to be fascinating. I thought the, you know, I've got to try the VR release. That was interesting. But I actually think that the audio only glasses might be a sleeper hit. Yep. I think there's something really compelling to that. I think people, you know, I mean, the vision of sort of Jane from Ender's Game was all about this sort of all knowing audio only companion. And I don't know, you know, I think there's a set of people that love to wear the kind of camera on their face, but others may be aware of the fact that it could make people uncomfortable and they're like feeling like they have the technology connectivity without any of the social awkwardness of the camera. Audio is... is incredible. I'm quite bullish on voice and audio. ChatGPT voice yesterday dropped their... uh, chat GPT voice to allow connectors in so you can connect to your Gmail and calendar. I played around with this morning and To me, it was my second magic moment of this whole AI assistant space. Really? I was biking to work. Hands are busy. I want to get stuff done. And I just start talking to ChatGBT Voice. I just open up a session, start chatting with it. And throughout my 30-minute bike ride to work, categorized my emails, submitted them to different labels, replied to different emails, sent out calendar invites, got to the desk, inbox zero. It was so electric. So this is ChatGPT voice, not Muse. Correct, ChatGPT voice. Oh, amazing. Okay, I misunderstood you. I thought you'd said Muse voice. I totally agree. I think the ChatGPT, it's based on the GPT Live 1 model. It's full duplex voice, and it can see all your threads. So you can ask it things like, hey, what are those three projects I started? And where are we? And what am I blocked on? I think it's an extraordinary capability, and it's one that neither Muse nor Instinct has today. Though I think that Zuck actually did announce full duplex voice, so maybe it's coming to Muse soon. Definitely is coming. But I do think that... for the general consumer. an assistant is most helpful when... you are most occupied. It's not like you're going to be sitting down on your couch and spending all of your time using your assistant to do stuff. it's most helpful when you're gardening or cooking or doing things where your hands are busy and you can't actually engage with an interface. And, you know, it goes back to that like invisible assistant where you can just talk and say something and it just goes and runs and does something in the background. Yeah. And I think that the, tech bubble that is chronically online and on their phone and in front of a computer is a very different environment than the general market. [20:00] population. And so I'm really excited to see how does this actually hit mass adoption and how does the average consumer think about an assistant when it's actually like, is it going to be an assistant that solves real pain points for them where they're going to use it when something is really, really painful because they don't want to do it? You get a traffic ticket, you have to go pay for that traffic ticket. Great. Just take a picture and let your agent go do it for you. Yeah. It's going to be interesting to see. So the next thing I wanted to talk to you about was social. We talked in multiplayer generally. We talked about how iMessage is this sort of privileged place to be, which is actually one of the cool things about Instinct. However, in my experience, building agents that you add to group chats somehow feels intrusive or like it diminishes social capital. Have you seen anything that works well? And what are your kind of theories for how these agents get to a multiplayer state? A lot of people have tried it. I think I've seen three different approaches. The first approach is your traditional added to a group chat and it's just like the group chat and your companion. The second approach I've seen is how instinct did it, where you have like an agent network where you invite another agent and they're talking behind the scenes on your behalf and they'll update you as to what's going on. XMTP, and they are approaching it more like granola. So you add an agent to a group chat, but that agent is silent and it's purely a listener. It's taking notes on what people are talking about. And if it has an action item, it independently will ping the person in a separate message thread. And then if you want to see what people are talking about, you can go into the actual docs and see what it's taken notes on. And that way, it's more of, again, it's your invisible agent. You don't know it's there. It's doing all the admin work that you would want out of a multiplayer function. And if it needs action items, it doesn't bother the majority of the group. Really, really cool approach. I've got to try that out. I love that. I love that it's an A16Z company already, of course. My theory on this, if you look at this, and Nir from Obo said it, and I think he's right, which is the progress the models are making on verifiable domains like coding are, of course, exponential. I'm going to say made on just prose quality. And even like Bedside Manor has maybe plateaued or maybe is even getting worse. So I actually think that with the models as they are today, and perhaps with our own expectations, you know, in terms of these social dynamics, I don't think there's a lot of ways that the agents can contribute and enhance the sort of social connectivity of the group. You know, conversely, I think agents that are very utilitarian, you know, one of the experiments I've been doing is adding it. I'm a collector. records and a bunch of other things, adding an agent to the chats where I talk to my fellow collector friends and trying to catalog our respective collections and our tastes. And I haven't quite hit it, but it at least feels like it's something that's genuinely additive to the group rather than, you know, this sort of low EQ robot that's chattering when we don't want it to. Yep, totally. You'll... You should give Doc a try. I think you're going to be quite fascinated by it. I will, yeah. The other interesting take that I heard this week was that something about messaging in particular doesn't lend itself to these agent interactions because in messaging, it's as if somebody has the mic and it feels uncomfortable to give the agent the mic. Whereas if you look at something like maybe closer to a Discord or even a traditional web forum, that's actually something where an agent can make a post with the post or not, Reddit style. But for some reason, it feels less intrusive than group chat. I don't know. Yeah, interesting. It's almost as if we're trying to humanize these agents a little too much. That's right. And I think that comes to another component where if you look at the actual design of the character of the agent, I see some companies that have it try to be like an AI-generated human. Yeah. And then you have other companies like Muse that it's this adorable Yeti. Yeah. And I think it's really interesting to see how the world has fallen in love with this Yeti. It's just so cute and adorable. Yes. And the... like AI generated human, To your point, I think people feel a little uncomfortable. It feels weird to humanize these assistants so much in your personal space. So one related question I wanted to ask you is, do you think that the market segments by personality or archetype, you know, we were chatting a bit about this in our walk, but some people prefer a kind of serious, gruff, authoritative doctor, and some people prefer a warm and friendly, you know, sort of a doctor that feels like a peer. And those are just two different archetypes of doctors. Obviously, no doctor is both of those things. They're in opposition. Do you think the same will be true of agents? I think everyone for sure prefers a different style of communication. Poke. came out of the gates hot, right? The personality of Polk is, is, [25:00] quite distinctive. And it's something that I have a lot of friends who are like, I refuse to give poke any of my personal information because it's so sassy. And it actually makes me uncomfortable to engage in a conversation with it. But I don't think personality is really a defensible moat because it's quite easy to configure. It's like you as a consumer, you text your agent, you say that you would prefer it responds in a certain way. It just saves that to its memory or its sole file. And then you're good to go and it changes. I think we'll get to a place where whatever general agent ends up being the leader, that it'll be quite configurable. And each person can kind of tailor the agent to speak to them however makes them most comfortable. So that's interesting. not just personality, but kind of constitution, for lack of a better word, which is... An example of a kind of defining characteristic is presumptuousness. You know, we've talked a little bit about this. And there are some jobs for which you want the agent to be highly presumptuous. And you'd rather have it make mistakes 10% of the time as long as it often gets things done. And there are other styles of agents like our finance agent, perhaps, where you want it to be highly cautious. So I sort of wonder if some of these things, which we're now describing as personality, It tie breaks in a position of ambiguity. And perhaps there's a bit more defensibility around that. I think there is massive defensibility around proactivity. I think that's actually going to be something that separates the winner from the meat of the pack. The agent that can be most proactive, that can serve as an invisible assistant and start doing things on your behalf without you even asking. I think the best... gimmick, cool, aha moment that we've seen so far is checking into your flight for you. To your point, though, it really does come down to two categories. There's some workflows where... you want to retain control and you don't really want an agent to be acting too proactive on your behalf. And I think those are workflows that require a change in action on your side. So maybe that is... an agent identifies that the insurance you're using is charging too much money and thinks you should switch insurances. It should probably get your permission before it actually makes that switch because that's a direct action on your behalf that impacts you. However, the proactiveness of, hey, I drafted this email for you or I got a flight credit on your behalf, there's no action needed on your side. It's just saving you money. It's saving you time. So I think it'll be interesting to see how do we really categorize and break down those type of behaviors. But it will be a very fine line because if you cross that line once, I think you immediately lose all trust with your user. Yeah, it'll be interesting. We were joking internally, like we're days away from an agent messaging someone saying, I noticed you weren't that into her, so I went ahead and broke up with her for you. Yep. Saw that one. It's going to happen, man. But I think a lot of the magic comes from the presumptuousness. And I think that if you didn't see people posting on X about, oh, the agent did this, the agent did that, it would probably be a sign that the agents are not pushing hard enough in this direction. all, I'm a little suspicious of whether the Bong Chang post was even real. And for folks who don't know, there's a person who tweeted that Instinct checked him into a flight and hallucinated his middle name as Bong Chang and he wasn't able to get on the flight. So I'm not sure if that's just shit posting or if that really happened. But I'd be curious to know if that person is still using Instinct. I'm guessing they are. I would assume so, purely given the traction of that post. They're probably feeding for more content there. If nothing else, at least just a source of comedic, popular tweets. Right. Right. Maybe talk, you know, it feels like one of the big things that's happened is that the kind of primitives that were developed during the open claw period have now been productized in a way that the mass market consumer can sort of, you know, digest and appreciate. Do you think there are, and also POKE, which you had mentioned a couple of times, like POKE was extraordinary. I think they did a super clever job. And I think if they had started six months later, they might have ended up being the winner. I'll be curious to see what they actually have in store for us next, because I'm sure that now that they're a part of Cognition, they haven't stopped working on this. I'd be very curious to know if there are capabilities or lessons from the open claw era that you noticed that you think are going to tell us what's coming next in consumer agents. Right now, I honestly see the... consumer agent is literally just a replication of OpenClaw, but it's just pre-configured. I don't see much happening in the personal assistant space that you can't do with OpenClaw or Hermes. It's just a matter of fact that you don't have to configure everything and you're not getting error messages popping up every seven hours. And it's really easy to use. It's an easy interface. So we'll [30:00] We'll see what happens. I think the direction of the space will continue to move towards... proactivity. I think the direction of the space will continue to see more hyper specialization, which I think goes to your point in your thesis of narrow startups. And I think, How does the space look as we see more of these long tail... opportunities, whether it is, you know, instead of being a generalist, I'm going to spend all of my time focusing my assistant to be an expert for single mothers with children under the age of three. And it knows everything about how a mother operates her day-to-day life, child development, where it literally feels like that assistant is reading their mind. And that can maybe be a plug-in into the generalist. But I'm curious with this being your thesis, how do you think about, I guess first, give us some more insight on what narrow startups are to you and how do you see that Yeah, the theory with narrow startups is that 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 $200, $250, $300 a month. So you can build a $100 billion run rate business with tens of thousands of people, which has never really existed in the past. A lot of it also is that the capabilities can be specialized to such an astounding degree. there will be this kind of comparative advantage. I think it's going to be a combination of taste and proprietary knowledge. You know, there may be just as you, you know, you have a doula that may subscribe to a certain sort of style of, you know, helping women give birth, you may have a personal agent that actually subscribes to a certain school of thought or style or, you know, even a set of lessons that aren't well understood. And you're going to hire that agent to sort of provide that service to you. And then proprietary data, you know, you had mentioned potentially having a personal agent just for New York City. I think, you know, just as there might be a concierge that's got A plus taste that's hard to replicate, that knows all the best spots that you're going to love. And there may be another for another dozen different people or archetypes of people, I do think that there'll be this sort of subjective and objective specialization that happens. It's going to be interesting to follow. I'm also curious from your lens, even just noticing our conversation here. We interchange between assistant and agent. And it's been a constant conversation I see people talking about on Twitter. Is the future of this ecosystem, is it going to be a personal assistant or a personal agent? Yeah. And what do those mean to you? That's such a great question, man. I mean, to me, assistant is a sort of lesser form than agent. You know, for me, an agent is somebody you could imagine them being a peer in a social group. You know, and maybe we need a better word than agent, but 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. And maybe this parallels a little bit, you know, the sort of agents and AI in the enterprise where they've sort of started as interns. get promoted into doing higher order forms of work in the enterprise? What does it mean for an assistant and later an agent to be promoted in the consumer's life to do more specialized and delicate work? I'm super aligned. I'm also very interested what ramifications it might have in just like the social layer of humanity where – you know, does the word agent allow us to... personify AI enough is assistant... Does it add too much human to the experience? And does that make society uncomfortable? And it'll be interesting to see how the... We're in bubbles. You're in a more bubble. You're in SF. So you're in the tech hub of the world. I would say New York is a little behind SF. And then the rest of the world catches up after. So it'll be interesting to see how people actually think about the vernacular of the space. I don't know if it's a bubble. I just think it's like the future. It's like we're in the near future. Yeah. Yeah, I don't know. I actually think that it could increase and improve a lot of social dynamics as well. Because you start to think about the agent as a non-human, it can provide a layer of social indirection. So there might be something that's sort of awkward for me to say to you or an uncomfortable social dynamic that's occurring in a group. No. [35:00] unemotional? Because you know that it's not... It's like an arbiter of confrontation. Potentially, yeah. Yeah. Now, I'm not advocating for a more passive-aggressive sort of social system, but I do think it's interesting that they can be a sort of source and a destination for social indirection, if that makes sense. Right. Right. And I think that it will have a lot of significantly positive influences in the world. How do you see... AI impacting, right? If we look five years down the line, where do these... agents, where do these assistants actually take us? How do they make us happier human beings? It's a great question, man. I mean, I do think that those are the problems that we're really after at the root of things. You know, we talked about productivity. We talked about finances and health. There's a lot of mundane things. You know, I was looking this up yesterday, but 1.5 million people a year in this country get a DUI. And look, like I can, the amount of administrative burden that's probably implied by that and how many of those people have access to attorneys. And so do you think there's almost this sort of, that bureaucratic pressure that's applied to the average consumer. So I think step one is how do we sort of relieve that bureaucratic pressure and have people, you know, be financially optimized, have their health be optimized and have them have the same access to whether it's an attorney or a kind of family office that somebody who's wealthy might have. Then I think the bigger question, you know, I grew up with my mother saying the hardest part about getting what you want is knowing what you want. It's like, what are we all truly after? And what are the ways that these can sort of engage us. And I don't think it's as simplistic as just goals. I don't know how many people even have goals. I think a lot of it is people are just trying to find out who they are and turn into the best, most authentic versions of themselves. And this kind of interplay between our own self-development and our subjective experience in the world and these machines could be a very beautiful future for us. Totally. I see this future where I envision this world where my future kids are going to be looking at old videos of the current generation walking down the street, you know, next crane down, staring at your phone. I mean, what are you doing? Like, it is a beautiful world out there. Like, why are you spending all of your time staring at a screen? Yeah, I love that. Because hopefully it's going to allow us to automate so many of those things. So as a parent, you can actually spend more time with your kids. As an adult in your 20s, you can actually spend more time talking to friends. And I'm really curious to see how it will just change life for the better in so many ways of reducing the things that you just really hate to do. But it's like the life admin work. Yeah, that's right. And that's also why I think that so many times coding agents are not computing with human programmers because they're doing things that no human programmers are doing. Whether it's building this ridiculous side project, I've been building a virtual record shop that you can walk into. I'll send you a link to it after. It's awesome. I would never hire a programmer to work on that five or seven, ten years ago. I would have never spent the year of my life that it would have taken. this technology in our life where we're just getting to explore things we never would have explored anyway. It's not about cost reduction. It's about possibility expansion. makes life more fun makes life more interesting You know, to kind of come back to the near term, a thing that I've been thinking a bunch about, but I'm curious for your take on is, you know, the lessons of Moldbook. And what are these sort of agent native systems that are going to need to be built? And maybe this is some of what Instinct is doing with their kind of instinct to instinct network, but it does feel like there's going to be infrastructure and products that exist only for agent to agent interactions, right? Totally. That doesn't exist today. What do you think? It's going to be a whole new paradigm. I think it opens up, and I think that can be a 10-hour conversation on its own. You have agent emails, you have agent phone numbers, you have... What does the world of service look like when it's no longer... human booking to human and no longer agent booking from human, but it's now agent booking from agent. And how does that redefine the entire service industry as a whole? And then you have the whole cybersecurity component of all this stuff, which is, okay, well, internet came first, then came cybersecurity, agents came first, now comes the security component of these agents. And what industry is that going to create? The beautiful part of it, although I think that's problems to solve with these agents. And I think it is inevitable that we're going to see so many startups popping up trying to figure out all of these niche problems to really build this new wave of internet and digital connectivity. Yeah, it's going to be cool to see, especially commerce. [40:00] I always think of the web as kind of, there's three webs within the web, just like there's two wolves inside me. The three webs are the commerce web, the application web, and the content web. And the content web has been in trouble for a while. I think that the commerce web has actually done really well. And then the application web has definitely been on a huge upswing ever since coding agents were released. But the commerce web is fascinating. You saw this week how Shopify embraced Muse and Amazon blocked Muse. is who's going to be a winner or a loser when you actually have a new front door for how traffic flows to these marketplaces. I mean, do you have a view on what specifically happened with Amazon and Shopify and maybe more broadly who the winners and losers are likely to be? Yeah, so I think... First off, Shopify and Amazon are two completely different business models. You have Shopify that's hyper-focused on more democratizing commerce. You have Amazon that makes all of their money on ad revenue. And so... The whole ad structure of commerce is going to be really fascinating to watch because when you remove human eyeballs, it defeats the entire purpose of the system. And so I think that's where... Amazon was coming from. If you let Muse and these agents start to purchase without them getting any ad revenue out of that, it really takes a lot of the profit out of the experience. I think there's likely going to have to be some... new paradigm for how they're able to profit in this world where agents are purchasing on their behalf versus a Shopify that is empowering the everyday individual to sell as much as possible and agents democratize this space even further. So it's just unbelievably advantageous. I think it's going to be even more interesting when we look at industries that have yet to be where if you look at the booking space, right, like you go to make a reservation, I think we saw Resi started shutting people's accounts down because they were using browser use and they didn't want bots. It defeats the whole business of... If you just let agents take these reservations instantaneously, then it kind of defeats the purpose of the platform of this human-oriented booking experience. And so what happens though, when everyone has an agent, now everyone's on the same playing field, everyone can go to book the restaurant at the same time. Do we enter a world where one, power goes back to the restaurants and they can maybe choose who they want to give the reservation to based off of loyalty or average Do we go into this world of bidding, where everything becomes a bidding war between these agents? Curious your thoughts on what that future looks like. Yeah, I think you're super right. I mean, very interesting, because I think there'll be things that are supply constrained, like hot restaurant tables, things that are demand constrained, like people that want to fly to SF tomorrow. I think for restaurants, if you think about, with perfect information, who do they want? To your point, it's the highest AOV, highest LTV, which means AOV over multiple visits. We're at near 100% capacity as much as possible. So that's actually a more sophisticated calculation, and I don't quite know how they're going to do it. But I'm sure that proprietary supply is going to become more and more important, you know, versus relying on kind of friction and motivation from the end consumer to act as something that sort of prevents this DDoS. So I think supply-constrained sort of aspects of commerce will be really interesting. a group buying experience where you have a set of buyers that are willing to buy a product or a service or an experience. And then you have the supply side, the merchants that could potentially fulfill that need, come and bid to see who can do it, the best experience, maybe at the lowest price. So we've always had this kind of aggregation of demand, but we've never really had aggregation of supply in the same way, especially not in an auction system to benefit the consumer. So I think that's really interesting. You know, in the Amazon Shopify thing, I think that the Amazon probably is rightly taking the bet that they don't want to be disintermediated and that they'll lose not just advertising revenue, but impulse shopping. I'm kind of curious to know if the agents themselves will be impulse shoppers to some extent, because I think the point around proactivity means the agents are going to try to make delightful guesses. And, you know, what is more delightful than a surprise or a gift? perfect reflections of their principles. They're not perfectly rational. They're going to be non-rational in completely new ways. And I don't know that things like marketing, things like impulse purchases, things like advertising won't matter. I just think they need to be transformed. Totally. And I think that opens up a whole new conversation as well, which is what is the recommendation engine of this agentic commerce space actually look like? [45:00] because... AI and LLMs feels like it's kind of broke the internet in terms of quality search. But then if you use an agent and you say, hey, Buy me some white socks. There's like 20,000 different types of white socks out there. It's going to send you, it doesn't actually know what exact type of white socks you prefer. And so I think it'll be really interesting. I think Muse has... probably a head start on this. And this is Zuck's whole thesis of attacks on all commerce, where they have your Instagram data, they have your Facebook data. So they probably have a little better understanding of what you actually prefer. But it'll be really interesting to see what this recommendation layer looks like in this world of commerce. Well, I think it could be fun for the long tail, because you can imagine to your white socks example, I could go to Amazon and get white socks. But maybe my agent comes back and says, hey, I found someone awesome on Etsy that can know, grandma Nebraska. And she just knits these amazing white socks. It's going to take three weeks. Are you down? I'm like, sure, I'm down. That sounds interesting. So I kind of wonder if it brings more serendipity in because, of course, you know, it can examine all the options and perfectly tailor these things to you with no adverse incentives. Right. I would love that. I think that'd be so fun. It would definitely add a little flavor to the mundane life of shopping today. Exactly. Mundane from my point of view, as someone who doesn't take joy in sock shopping. Well, dude, and look, you can even imagine a world to get a little more futuristic where it starts to now... catalyze things in the real world. So maybe instead of grandma who knits white socks and sells them on Etsy, you know, maybe it messages the agent of a grandma and says, hey, would you be up for doing this? So I think this is where the kind of agent to agent spaces get really interesting, where they're all acting on behalf of their humans, you know, some to provide products, services, knowledge, and some to make money. And, you know, maybe there's this sort of secondary and just much more interesting than the one we have today. Right. I'm just imagining now a grandma in Nebraska receiving a text message saying, hey, this guy in SF wants you to knit white socks for him. Why not? You want to do it? Why not? Yeah. Yeah. Why not? That'd be awesome. Yeah. So I'm actually very hopeful. You know, it also, it sort of sets the incentives in the right way because if the agents are doing the right thing, they're choosing great products based on product quality and they can assess product quality for themselves rather than, you know, the thing that has the best on a brand basis, it's at least with the explicit sign-off of their user. Like, I just want to have Nikes because I love Nikes, not because, you know, there's sort of this cognitive dissonance that's happening in the marketing and the consumer is actually doing something that's not optimal for themselves. Right. It opens the door to a truly consumer-aligned internet, which I think we have been lacking for Quite some time. I agree, man. We touched on the economics of the agents. Of the 122... that I have looked through. So I've personally tested 26 of them. Okay. Of the 122, 65 are paid. Interesting. 35 of them are fully paid. 30 are a freemium model. So once you hit a limit, then you have to pay. Only 13 of them are actually free. Thank you. When you have Muse and Instinct that are both free, that are also the dominant ones, and then you have OpenAI that supposedly is going to be releasing something cool maybe next week, who knows? How does a long tail startup actually win this space? And why are they charging money? When... the big dogs are giving it to you for free. Man, it's going to be really challenging. By our estimate, it's costing something like $20 per user per day to do this in a really ambitious way, potentially hundreds of millions a year for a startup. I think the two positive trends to bet on are that browser use should be deflationary, get exponentially cheaper very rapidly. Now, if you're a startup that is raising today, that's betting on that happening in six months, maybe it's a little bit of a game of chicken, seen token prices depress dramatically and emerging capabilities depress, so the price of them. So I think it's very possible that browser use gets cheap enough that more products can offer this for free. I guess the flip, though, is that I don't love the idea of a customer subsidy being the primary value prop. So I'd love to see the agent that costs $1,000 a month that the consumer is so excited about that they're willing to pay. not necessarily necessities, they're sort of desires. So what is the agent that is exciting enough that you're willing to pay that kind of price for it and then that agent will benefit from both extreme market fit as well as sort of deflating costs? Right, right. I think it'll be interesting to see. Super interesting. [50:00] Well, David from Assistant Bench, make sure everybody checks it out. Thank you so much for joining us. We're going to do this frequently. I mean, I feel like we could do this every week, dude. Like things are happening every single day. So please follow him on X. Follow me as well. We'll all keep up to speed and sort of learn together in the X community and sort of happy day to you and your agents. Appreciate it. Same to you. This was fun. All right, brother. Thank you. Awesome. you As a reminder, 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. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures. Thank you. Bye.