A16z Tests 147 AI Agents as Social Features Outpace Chore Automation

A16z测试147款AI智能体：社交功能增速超越琐事自动化 消费级AI助手如果仅专注于行政琐事和说教式理性，将会陷入停滞；只有满足人类虚荣心、社交曲折意图和进取型自我提升，才能实现主流规模化普及。 专家小组探讨了新兴的消费级AI格局，摆脱了单一全能助手的设想，转而描绘出由大量专门化、多人协作且具备社交闲聊属性的智能体管理生活不同领域的图景。基于对140多款工具的基准评估，他们讨论了围绕数据隐私、有限权限边界、对社交尴尬的恐惧，以及用户为了眼前的便利自愿让渡凭证的实际倾向所产生的矛盾。他们剖析了智能体在商业、浏览器自动化和旅行规划中的实际表现，指出技术可靠性与上下文可移植性构成了关键的竞争壁垒。主持人们最终认为，实现主流消费级普及的突破点远不止于平淡的琐事自动化，真正的吸引力将来自于能提供社交机缘、并充当自我实现“秘密超能力”的进取型智能体。 在对超过147款AI助手的评估中，产品主要归类为通用型、工作导向型、多人/社交型以及旅行类别，而个人理财领域则因信任壁垒大体停滞，该问题占高频用户投诉的22%。 (3:36) 对主流AI智能体的评估显示，上下文记忆在很大程度上已实现商品化与可移植化，15款顶级助手中有14款同意根据用户请求导出完整的记忆文件。 (17:48) 自主智能体正利用视觉上下文和多模态库，将购买行为直接引导至Shopify等平台上的长尾DTC（直面消费者）商家，从而对聚合平台的利润空间构成威胁。 (23:57) 消费级普及的长期拐点在于超越单调的家务琐事自动化，转向能够增强个人身份认同与风格的进取型“秘密超能力”智能体。 (48:04)

## Background 本期视频由知名风投机构 a16z 于 2026 年 11 月 4 日在 YouTube 频道发布。节目邀请了 Assistant Benchmark 的创建者 David Pawlan，与 a16z 投资人 Anish Acharya 和 Josh Elman 共同复盘其对超过 140 款 AI 助手的实测观察，探讨 AI 智能体如何从单向效率工具演进为涵盖购物、规划、社交甚至跨智能体互联的复杂网络。 ## TL;DR 对话围绕 AI 智能体生态的“寒武纪大爆发”展开，深入分析了阻碍智能体落地的核心瓶颈——信任问题与“家长式作风”的雷区。通过订鲜花、规划奢华旅游等实际压力测试，嘉宾们揭示了现有模型在多约束博弈、隐性偏好显性化以及浏览器自动化上的能力分水岭。此外，讨论重点探讨了八卦型社交智能体（Gossip Agents）与智能体互联网络（Agent-to-Agent Network）如何创造意想不到的人际偶遇与商业撮合，指出智能体的终极形态绝非简单的省时工具，而是能够触碰深层人性需求、赋予用户全新形象的“口袋秘密武器”。 ## Atmosphere & style 整场对话洋溢着前沿科技探索者的兴奋与敏锐洞察，氛围既像极客沙龙般轻松无拘，又兼具资深产品人与投资人的犀利批判。三位嘉宾在抛出技术测试细节时频繁碰撞出火花，叙事节奏紧凑且充满风趣的生活化案例（如技术恐惧症母亲让 AI 找狗、朋友间 Agent 的八卦互动）。整期节目从严肃的基础设施与安全围栏讨论，层层递进到对人性阴影面与高阶心理需求的哲学审视，给听众带来极具启发性与前瞻性的沉浸体验。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=0s) 对超过 140 款 AI 助手的追踪和 3600 条深度用户讨论分析显示，信任缺失是用户最核心的恐惧，占所有顾虑反馈的 22%，成为财务类智能体落地的主要瓶颈。 - [09:15](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=555s) 用户在数据隐私上的表态偏好与真实偏好严重割裂：人们口头重视隐私，但实际中会为了极小利益让渡全量权限，真正的心理红线是社交尴尬而非单纯的数据流动。 - [23:57](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1437s) AI 代理具备无限耐心全网检索比对长尾 DTC 品牌，正绕过传统中心化聚合电商并挤压其中间费用，促使 Shopify 等基础设施获益而巨头收紧数据权限。 - [37:20](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2240s) Agent 间的互联网络终极形态不是无目的闲聊，而是意图驱动的后台协同与偏好匹配，能促成兴趣群组并激发基于“熵”的现实生活线下偶遇。 ## Key takeaways - [03:36](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=216s) 未来并非单个通用智能体统领全部，而是会形成针对工作、社交、旅行等不同生活切面的智能体军团。 - [05:12](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=312s) 智能体需要建立清晰的“围栏”隔离机制，且应通过 qualitative 的法律与原则而非生硬死板的规则来调控其自主裁量权。 - [07:54](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=474s) 用户本质上不愿完全让渡控制权，最理想的智能体体验是由其穷尽所有前期准备与方案权衡，但将最终确认与触发权留给人类。 - [12:16](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=736s) 在鲜花订购与多约束洛杉矶旅行规划实测中，多数智能体暴露出难以兼顾“极度奢华”与“成本效益”等内在矛盾目标的通病。 - [17:48](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1068s) 人机交互存在隐性偏好显性化的门槛，而评测前 15 名的智能体中除 1 款外均支持导出详尽到令人不适的个人完整记忆数据文件。 - [20:30](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1230s) 新智能体引导机制正从导入 ChatGPT 记忆文件、语音轻松闲聊，演进至通过相册照片构建跨越语言描述的高维个人品味模型。 - [25:31](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1531s) AI 展现的是异星心智而非绝对的人类理性，消费级产品中任何强推“最优解”的家长式作风都会遭到真实人性的排斥。 - [31:41](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1901s) 八卦型 Agent 提供了一种“社交间接性”，作为低调的社交润滑剂在朋友之间传递近况并有效规避主动倾诉的尴尬。 - [33:40](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2020s) 软件架构在面临高并发规模化考验时仍具备坚实壁垒，底层浏览器自动化与事务完成率直接决定了产品的存活能力。 - [37:20](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2240s) 无明确目标驱动的多 Agent 闲聊在大约 150 条消息后会退化为概念胡诌，证明 Agent-to-Agent 网络的价值必须依赖明确意图驱动。 - [39:24](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2364s) 意外惊喜与机缘本质上是 AI 领域被低估的“熵”，Agent 有潜力根据深层上下文构建打破信息茧房的全新线下社交网络。 - [42:01](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2521s) 智能家居被动环境数据、长期电商与 App 历史订单，以及线下真实社交场景的快照切片，是尚未被充分激活的高密度上下文。 - [44:33](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2673s) 用户基于好奇心探索冷门边缘事物的临时行为极易污染偏好模型，亟需区分核心生活画像与偶发探索噪声。 - [47:07](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2827s) 相比复杂的 Prompt 框架，极简的短信交互形态能让技术恐惧者迅速上手，并自主委托照片寻宠与价格谈判等高度定制化任务。 - [48:04](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2884s) Agent 的终极爆发点不在于机械提醒未完成的琐碎事务，而在于触碰马斯洛高阶需求，成为提升自我形象并解决个人挣扎的口袋超能力。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=0s) 智能体全貌、多面生活与信任障碍 生活包含无数切面，从财富管理、行程规划到健康追踪，AI 智能体（Agents）正逐渐渗透到日常的各个角落。然而，伴随着智能体能力的扩张，用户的深层顾虑也浮出水面。通过对 147 款以上不同 AI 助理工具的追踪评测与大量深度用户交流，数据揭示出人们最核心的恐惧正是“信任”。在所有关于使用顾虑的反馈信息中，有超过五分之一的讨论直指信任缺失。 智能体要想发挥最大价值，往往需要触碰那些人类不愿思考甚至不愿向外界暴露的隐秘领域。能够建立起深层信任机制、触及敏感“暗面”的智能体，将孕育出颠覆性的产品体验。这不仅是一个高度敏感的赛道，未来更可能重塑商业与用户之间的互动网络，迫使企业对交付的内容与服务质量倾注更多考量。 - “From over 147 different assistants that have come across my plate so far, the number one fear that people talk about, 22% of the messages are talking about trust as the main fear.” —— David Pawlan - “What I've seen from 30 years in consumer product is paternalism will kill any product. We talk about productivity and goals sort of achievement, but maybe people want the exact opposite of it.” —— Josh Elman --- ### [00:52](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=52s) 智能体线下聚会与“寒武纪大爆发” Anish Acharya、Josh Elman 与主导 Assistant Benchmark 测试的 David Pawlan 齐聚一堂。距离他们上一次探讨仅过去大约一个月，但智能体领域的迭代速度一日千里，几乎每天都在诞生新范式。就在录制前一天晚上，David 组织了一场聚焦 AI 智能体的线下聚会，吸引了超过 100 位极客开发者参与，现场甚至举办了 10 家智能体初创企业的即兴路演。 ```mermaid flowchart TD accTitle: AI Agent Ecosystem Evolution accDescr: The progression from early community collaboration to specialized multi-agent architectures. subgraph Community["初期协作阶段（寒武纪爆发）"] A["100+ 智能体开发者与发烧友"] -->|"线下交流与 Demo 演示"| B["高频反馈与最佳实践共享"] B --> C["非零和博弈探索"] end subgraph Specialization["专业化落地阶段（分化矩阵）"] C --> D["工作提效类 (Work/Email)"] C --> E["多人社交与闲聊 (Multiplayer/Gossip)"] C --> F["复杂规划类 (Travel/Logistics)"] end ``` 这种线下聚会的氛围极其融洽与开放。虽然很多智能体产品在功能上存在表层重叠，但开发者之间并没有剑拔弩张的防备心理，反而表现出强烈的社群渴望。所有人都意识到，当前处于智能体生态的“寒武纪爆发期”，没有任何一家公司能独揽所有场景，大家都在沙盒中共同探索未知边界。 - “I mean, I think what's really cool is this is not a zero-sum game.” —— Josh Elman - “Because we're so at the frontier of this, nobody really knows what the landscape is really going to shake out to be. And so, we're all just kind of playing in the sandbox together.” —— David Pawlan --- ### [03:36](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=216s) 智能体大军的细分维度与爱八卦的社交智能体 未来的世界绝非“一个通用智能体统领全部”，而是会形成类似现实社会中拥有不同性格、不同专业技能的“智能体军团”。面对日常生活的多元场景——理财、旅行计划、健康维护、日程管理或是单纯的头脑风暴，用户会像求助不同朋友一样，唤起特定技能的智能体。 智能体的差异化维度十分宽广，横跨“严肃严谨的生产力型”到“轻松玩乐甚至具备对抗性的社交型”。Josh Elman 分享了自己测试几款“八卦型智能体”的趣事：这类智能体连接了他与几位好友的上下文，能够自主穿梭于社交信息流中传递近况。当他朋友因房屋装修改造导致周末全家断电时，智能体第一时间向 Josh 透露了这个情报，促成了一次默契的朋友间通话。这种体验无关省时或订票，却显著增强了人际连接。 David 总结了当前市面 147 款助理主要落入的四大类别，并指出了关键的空白地带： | 智能体分类 | 核心场景与行为特征 | 当前生态现状 | |:--- |:--- |:--- | | **通用型 (General)** | 开放式问答、跨任务日常琐事处理 | 占多数，同质化探索期 | | **工作生产力型 (Work-oriented)** | 邮箱深度过滤、日程协调、执行确定性任务 | 高频刚需，落地较快 | | **多人社交型 (Multiplayer)** | 好友信息串联、八卦互动、轻量社群连接 | 玩法新颖，重构社交网络 | | **出行规划型 (Travel)** | 跨平台行程拆解、机酒对比与落地安排 | 垂直场景，整合度逐步提高 | | **个人财务型 (Finance)** | 资产调度、跨账户支付、自动化理财操作 | **极度缺失**（受制于信任门槛） | - “And I think we're going to get used to having kind of an army of agents that will do this.” —— Josh Elman - 针对 AI 深度用户的 3,600 条群聊数据分析显示，有高达 22% 的痛点讨论集中于信任风险，这也是财务类智能体迟迟未能大规模爆发的技术与心理瓶颈。 --- ### [05:12](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=312s) 信任围栏：Shane Mac 的 Slack 泄密教训与原则性监管 解决信任问题需要建立清晰的权限边界。Shane Mac 提出的“围栏（Fences）”理论（即 Picnic 概念）为此提供了绝佳的隐喻：如果自家庭院没有安装篱笆，家长在让孩子去户外玩耍时会时刻提心吊胆，甚至不敢让他们离开视线；但只要院子四周建起了坚固的安全围栏，孩子就能在边界之内自由奔跑。 智能体生态亟需这种“安全围栏”。过去因缺乏严格的上下文隔离，发生过极其尴尬的事故：某位用户的个人助理意外穿透到了工作系统，将其私密财务数据直接广播发布到了全公司的 Slack 频道中。这警示行业：必须让个人智能体严格受限于个人通讯与数据域，坚决防止向职场系统渗透。 Anish 进一步指出，传统的沙盒（Sandbox）思维局限于硬编码的死板规则，而在大模型时代，与其给智能体生硬的规则条文，不如赋予其“法律与原则（Laws and Principles）”。通过质化的原则去调控智能体的“自作主张度（Presumptuousness）”，让模型在复杂现实情境中做出契合人类意图的动态价值判断。 - “He had a problem where I think a personal agent spanned over into work, and it actually like blasted his finances on his company Slack.” —— Josh Elman - “And somebody was describing this framing of like instead of giving models rules we should give them laws and principles.” —— Anish Acharya ### [07:54](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=474s) 自主性悖论：像对待孩子一样为智能体设限 对 AI 智能体（AI Agent）的管理就像对待人类员工一样，需要给予定性的建议与指导。然而，更为贴切的隐喻其实是“给孩子围上栅栏”。智能体虽然具备极其强大的计算和执行能力，但在行为逻辑上常常像个不谙世事的孩子，缺乏对人情世故与现实语境的微妙感知，无法准确评估“某件事究竟是不是一个好主意”。 行业内目前普遍将焦点放在自主性（Autonomy）上——比如“帮我订好整个行程”或“帮我完成餐厅预订”。但在实际场景中，人类用户往往并不愿意完全让渡控制权。人们期待的理想体验，是智能体承担所有繁重的准备工作，搜寻最佳方案并权衡利弊，但将最终的决定权、成本核准权以及行为触发权留在用户手中。 当前 AI 智能体的普及阶段，与 20 世纪 90 年代电子商务的起步期高度相似。在 1997 年左右，大约 60% 到 90% 的互联网网民认为在网上输入信用卡号是一件不可理喻的事，而如今人们已经习惯了随时随地在线刷卡。智能体的信任建立同样需要时间，即使目前完全由人类把控最终环节，也丝毫不会削弱智能体带来的巨大提效价值。 - 1997 年前后，约 60% 至 90% 的网民拒绝在网上绑定或提供信用卡信息。 - 人类用户对自主性的核心诉求：智能体负责穷尽方案与利弊权衡，人类保留最终确认与授权。 > “And I think as humans, we actually don't want to give away that sense of control. I think we actually want it to check in with us, but do all the ground work.” —— Josh Elman --- ### [09:15](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=555s) 表态偏好与真实偏好：社交尴尬才是真正的红线 在数据隐私领域，用户的“表态偏好”（Stated Preferences）与“真实偏好”（Revealed Preferences）存在巨大脱节。口头上，大家都会声称自己极度重视信任、安全与隐私合规；但在实际行动中，消费者只要能换取极其微小的收益，就愿意毫无保留地交出海量敏感数据。例如从 Venmo 时代起，人们就毫不犹豫地通过 Plaid 绑定信用卡与银行账户；初创公司 Hark 曾推出一项耗资百万美元的营销活动，承诺为每位新用户提供 30 美元的免费披萨，而条件是必须开放完整的 Google Workspace 权限（包括读取所有个人邮件与云端数据），大量用户依然欣然授权。 用户真正忌惮的不是单纯的数据流动，而是“社会性尴尬”（Embarrassment）与社交信号错位。此前 Shane Mac 的 Slack 泄露事件之所以带来强烈的心理冲击，本质就在于它让人极度难堪。随着智能体向多主体交互（Agent-to-Agent Network）演进，这种社交风险被成倍放大——设想在第二、三次约会阶段，两个人的智能体在后台私下沟通，用户会时刻担忧“我的智能体到底说了我什么坏话？它是怎么包装我的个人形象的？”这种社交机制的失控感，远比现实中的单纯隐私泄露更让人焦虑。 隐私与实用性（Utility）之间始终在做权衡。当产品能提供降维打击般的强大效用时，用户愿意在隐私上妥协（这也解释了为何即使苹果强调设备端安全与私人云计算 Private Cloud Compute，用户依然会追逐最前沿的大模型）；但若两款产品的实用性处于同一水平线，用户必然倾向于更保护隐私的一方。如何让智能体具备人类社交中的“分寸感”（Discretion）成为关键门槛，比如在合点外卖时，智能体是否知道适时沟通用户的乳糖不耐受问题而不让场面陷入尴尬。 在职场环境中，由智能体代写邮件的专业风险同样极高。 Town 等产品之所以能脱颖而出，核心就在于精准把握语调（Tone）并建立起代发邮件的信任层，避免因不当措辞让用户陷入职业声誉危机。为避免被误解或显得不真诚，部分使用者开始选择在 AI 代写的邮件文末主动署上所使用的模型编号，以此作为明确的社交声明。 ```mermaid flowchart LR accTitle: 隐私与实用性权衡矩阵 accDescr: 展示用户在不同效用差距与场景下对数据隐私、社交分寸与授权意愿的决策路径。 subgraph Tradeoff["效用 vs 隐私决策"] U1["高额外效用 / 免费福利<br/>(如 Hark 30美元披萨)"] -->|"让渡全部权限"| D1["完全开放数据<br/>(Google Workspace等)"] U2["效用同质化竞争"] -->|"优先选择"| D2["强隐私产品<br/>(如 Apple 私人云)"] end subgraph SocialRisk["社交与职场敏感区"] S1["智能体代际通信 / 邮件代发"] --> Risk{"触发社交尴尬？"} Risk -->|"缺乏分寸感"| Fail["人设崩塌 / 职业危机<br/>(如 Shane Mac 事件)"] Risk -->|"语调精准 / 透明标注"| Pass["建立用户深层信任<br/>(如 Town)"] end ``` - Hark 的披萨增长实验：通过为每人提供 30 美元的外卖披萨，成功促使大量用户交出整个 Google Workspace 的全部数据访问权限。 - 用户防御机制：有使用者在邮件署名中直接附加所用模型名称及版本号，向收件人坦白使用 AI，防止被视为欺骗。 > “I think in revealed preferences, consumers will share any amount of data for even the most trivial amount of upside, right?” —— Anish Acharya > “The social mechanics actually feel so much higher stakes than the actual like real world implications.” —— Anish Acharya --- ### [12:16](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=736s) 多约束决策测试：买鲜花与洛杉矶旅行规划 为了检验各家智能体是否真正具备人类般的“判断力”（Judgment）与处理张力（Tension）的能力，业内开展了多场景压力测试。 第一个测试场景是“当日鲜花订购”。该测试的核心在于考察智能体如何在“时间限制”（必须当天送达）与“成本约束”（价格必须尽可能低廉）这两个互相冲突的目标之间取得平衡： - **Instinct**：表现极其惊艳，能够反复在不同渠道搜寻并锁定最低价格，完美平衡了当天送达与极致低价。 - **Kicker**：凭借让用户安装 Chrome 浏览器的策略，获得了出色的网页操控能力（Browser-Use），从而成功挖掘出极高性价比的选择。 - **Season**：在平衡两大约束方面同样表现突出。 - **其他主流智能体**：虽然全部能够满足“当日送达”的时效指标，但在成本控制上几乎全军覆没，直接放弃了对价格维度的深度优化，暴露出无法兼顾多个复杂偏好的通病。 第二个测试场景是“洛杉矶周末游规划”。指令故意设定了内在矛盾：“帮我订去洛杉矶的机票并规划整个周末，要求既要极度奢华（Luxurious），又要具备高成本效益（Cost-Efficient）”。各智能体展现出了截然不同的理解与价值取向： | 测试智能体 | 对“奢华”与“成本效益”的平衡方式 | 具体行为表现 | |:--- |:--- |:--- | | **Instinct** | 居中均衡 | 严格在奢华品质与开支之间寻找折中点 | | **Grokbot** | 偏向时间维度（以时间为奢华） | 认为最大的奢华是最大化在目的地停留的时间，为此牺牲部分经济性 | | **DOTS** | 偏向成本优先（以省钱为核心） | 优先选择红眼航班（Red-eye flight）以压低出行总开销 | 真正的优质助手应当能识别这种深层矛盾并向用户主动厘清取舍（例如：部分用户愿意坐红眼航班省钱省时，但落地后希望订高级酒店上午补觉，这种组合虽增加房费却提升了整体体验）。在实测中，各大智能体确实都提出了一些澄清性问题，而测试者刻意保持了回复的模糊性以检验它们自发推理的边界。测试的高潮在于——将这批智能体同时拉进同一个聊天室中，让它们同台竞技与博弈。 - 智能体面对复合约束的普遍缺陷：容易优先服从输入的第一个指令，而丢失后续的多维约束。 - 浏览器控制（Browser-Use）优势：借助浏览器插件或扩展（如 Kicker），智能体能穿透封闭 API 无法触达的电商平台实时抓取低价。 > “Because I think what's really human about something like that is actually understanding what you mean by that and presenting you trade-offs.” —— Josh Elman ### [16:14](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=974s) 多智能体对抗与鲜花测试：价格与情感价值的博弈 David Pawlan 描述了一项多智能体制定行程的对比实验：智能体经历三轮方案规划，每轮结束后展示各自的行程规划，并阐述为何自认为是最优方案。这项测试旨在验证智能体是否会受到其他智能体的影响，进而改变自身规划。实验结果显示，Grokbot 是唯一一个能有条不紊地吸收其他智能体优势、并将其融入自身方案的智能体。 随后话题转向之前的订花实验。Josh Elman 追问 David Pawlan 这次是否是在为自己订花，David 澄清这是送给女朋友的礼物。Josh Elman 指出，如果在送礼场景下智能体只是一味强调“省钱”，可能会带来灾难性的社交后果——没有人希望女朋友觉得自己只配得上最便宜的花；如果她钟爱兰花，用户显然希望买到兰花去打动她。David 承认本次测试仅设定了单一目标以评估维度表现，给女朋友送花只是测试的附带收获。现实生活中，除了成本和配送时效，还会涉及诸多二级决策维度。例如有些智能体在下单流程中会主动询问是否需要附带贺卡留言，或者自动隐藏价格标签，以此规避情感表达中的尴尬。 > “And Grokbot was actually the only one that you know very methodically took value from the other ones and implemented it into its own plan.” —— David Pawlan ```mermaid flowchart TD accTitle: Agent Itinerary Debate Process accDescr: Three-round iterative debate loop where agents present itineraries and Grokbot integrates learnings. subgraph Iteration["三轮行程规划循环"] A["智能体生成初步行程"] --> B["展示行程与理由辩护"] B --> C{"观察其他智能体方案"} C -->|"唯一有条理吸收方案"| D["Grokbot: 吸收优势并优化计划"] C -->|"未受其他方案影响"| E["其余智能体: 维持原有偏好"] end ``` --- ### [17:48](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1068s) 隐性偏好外化与 ChatGPT 留下的“数字镜像” Anish Acharya 指出，人与智能体交互存在一个根本门槛：隐性偏好必须被显性化。许多人自视并不抠门，只是“喜欢省钱”；但在面对智能体时，用户不得不打破心理防御，直白承认自己“确实很抠”，并要求智能体据此执行。人类自己做决策时可以绕开这种“偏好镜像”带来的自尊审视，因此智能体如何与人交互，在维持用户自我认同与安全感的同时不越界，成了一大关键难题。 David Pawlan 提到，Dot 之所以引人瞩目，正是因为它依托了用户长期与 ChatGPT 交互积累的海量上下文。虽然 David 对 354 位 AI 高频用户进行的调查显示，仅有 3% 的人将 Dot 作为主力智能体，但尝试并青睐它的用户普遍看重其内置的丰富画像。由于许多人将 ChatGPT 当作心理咨询师，模型早已掌握了用户的心理底色——它清楚地知道当你说“我不小气”时，其实你骨子里就是有点抠。 针对 Josh Elman 关于这种上下文壁垒是否可复制的疑问，David 认为这种记忆完全可以迁移，目前的阻碍主要在于迁移成本。在针对基准排名前 15 位智能体的测试中，David 尝试索要记忆数据导出文件以迁移至其他产品，除了一款智能体外，其余 14 款均支持导出完整的记忆数据。导出文件极其详尽甚至让人后脊发凉，精准剖析了用户的思维方式，以至于 David 看了几份后便不适地停止翻阅。 > “It's interesting because there's this sort of thing that has to happen when you talk to the agent about these things, which is your implicit preferences have to be made explicit.” —— Anish Acharya - **AI 高频用户调研**：在受访的 354 位 AI 资深用户中，仅 3% 将 Dot 列为主力智能体。 - **记忆导出测试**：基准榜单前 15 名的智能体中，除 1 款外均支持导出用于迁移的个人完整记忆 Markdown 文件。 --- ### [20:30](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1230s) 上下文冷启动与多模态相片：从语音破冰到视觉记忆 多智能体产品在引导新用户（Onboarding）时的手段正在发生质的转变。Anish Acharya 观察到一种新兴的引导机制：产品直接指导用户在 ChatGPT 中输入特定 Prompt，将导出的历史记忆文件上传给新智能体。David Pawlan 补充了另一种基于语音的破冰模式：系统直接发起通话邀请，通过几分钟轻松且毫无戒备感的闲聊问答，快速抽取用户的性格与核心偏好。 然而，更深层次的上下文载体正在转向用户的照片库。相片记录了许多用户甚至无法用语言准确描述的细节。Josh Elman 提到了 Marissa Mayer 新创公司推出的产品 Dazzle，该产品正是通过解析用户留存的关键生活瞬间与屏幕截图来构建精准的用户画像。David 正在对其进行基准测试，并对 Dazzle 的信任层设计表示赞赏：面对交出整个相册的隐私恐惧，Dazzle 设立了“若非必要，主动遗忘”的原则，自动剔除不必要或用户可能不愿暴露的信息。 > “It's actually a really interesting and now common onboarding pattern where it will say, 'Hey, go and paste this prompt into ChatGPT, download the file, upload it to me.'” —— Anish Acharya --- ### [22:18](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1338s) 截屏作为元数据：品味建模与电商消费的结合 关于截屏是否能成为长效的上下文来源，嘉宾们展开了深入探讨。Anish Acharya 对此持一定保留态度，他曾亲自开发过一款应用，专门从手机截屏中提取关键信息。由于截屏是移动操作系统中极少数能够横跨所有 App 捕获上下文的手段，它确实能反映用户在某一时刻的工作状态与兴趣。然而实验表明这些信息的价值衰减极快，用户缺乏长期回访的动力。 Josh Elman 则持乐观态度，他认为截屏的真正潜力在于被动捕获演变为主动喂养模型。他提到一位创始人正在研发名为“Taste”的产品，该产品将用户保存的截屏与素材组织成类似 Pinterest 的情绪板（Mood Boards）。当用户后续编写代码或开发 App 时，可以直接指令系统调用这些案例所沉淀的个人风格库，从而建立起持续回流的正向闭环。David Pawlan 紧接着用数据印证了这类视觉上下文在购物场景中的潜在能量：在其调查中，亚马逊零售电商高居智能体购买类别的第二位，占比达到 35%。这种沉淀下来的用户品味与视觉偏好，正成为驱动智能体辅助购物的核心推动力。 > “So, I built an app that just extracts key information from all the screenshots on my phone and it was fine, but I didn't find myself coming back to it. Like the value of the information sort of decayed quickly over time.” —— Anish Acharya - **电商智能体渗透率**：在用户购买行为调查中，亚马逊零售电商（Amazon Retail）以 35% 的占比位列第二大消费场景。 ### [23:57](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1437s) 视觉上下文、长尾电商与聚合平台的挤压 当讨论到 AI 助手在购物场景的应用时，对话聚焦于从照片等视觉上下文中发掘需求。如果一个应用拥有成千上万张用户的照片，清楚记录了用户的穿着打扮与日常风格，它是否会比用户自己更懂其购物喜好？这种深度的上下文让 AI 具备了极强的发掘潜力。与此同时，市场竞争格局也正在发生剧烈重构。当 Amazon 和 Delta 等行业巨头选择封锁数据接口、收紧权限时，那些拥有优质产品但缺乏营销资源的长尾品牌将成为结构性的大赢家。 AI 拥有无限的耐心去互联网各个角落搜索、比对，为用户发掘完全契合个人审美品味的产品，哪怕这些商品根本不在排名前十的主流电商网站上。以往 DTC（Direct-to-Consumer，直接面向消费者）品牌受制于高昂的获客成本，消费者需要耗费大量精力去逐一搜寻；而如今，AI 代理能够在更大规模上直接对接这些品牌。这也解释了为何 Shopify 会全力支持并向这些 AI 开放接口，因为 Shopify 的底层支撑着海量的独立品牌。反倒是传统的中心化聚合平台（Aggregators）面临被压缩生存空间的危机，这也是它们采取严密防御的原因所在。正如 Josh Elman 指出的那样：“如果我能弄清楚哪件毛衣好看或者想要那件衬衫，而 AI 能直接找到真正的制造商让我直接购买，他们就不必再支付中间所有的交易费用以及从批发到零售的成本。” ```mermaid flowchart TD accTitle: AI电商对聚合平台的去中心化重构 accDescr: 展示用户视觉偏好如何绕过传统聚合电商，直接通过AI链接DTC独立品牌 subgraph Context["用户深度上下文"] A["个人照片库与穿搭图像"] --> B["AI 掌握真实审美品味"] end subgraph Discovery["搜索与撮合机制"] B --> C["AI 无限耐心全网检索"] C --> D{"绕过传统中介？"} end subgraph Market["电商市场格局重塑"] D -->|"规避批发与通道费"| E["Shopify 与长尾 DTC 品牌获利"] D -.->|"遭遇防御与流量压缩"| F["Amazon 等传统大型聚合平台"] end ``` ### [25:31](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1531s) 理性假设之争、异星思维与产品中的“家长式作风” 对话随后转向了 Patrick Collison 的一个著名观点：随着 AI 代理代劳一切决策，消费者未来是否会变得更加理性？如果消费者更理性，企业就必须在产品质量和实际价值上投入更多心力。然而，Anish Acharya 对此提出了异议。他认为代理确实能替人类做出更好的选择，但 AI 自身也会展现出属于它的“显示偏好”（Revealed Preferences）和某种方向上的非理性——这种非理性并非源自人类心智，而是来自“异星心智”（Alien Mind）。 Anish Acharya 尖锐地剖析了这种异星心智的偏差：“举个例子，我认为代理可能会对文档规范且干净的 API 产生一种非理性的偏好，并赋予其极高的价值；而它们可能会极其反感背后跑着 Java 代码的蹩脚 API。”面对这种现象，未来不仅需要专门针对 AI 的消费者调研（Consumer Research for Agents）与评级体系，还要认识到人性的本质。Josh Elman 认为人类永远是依照自我满意度做决定的血肉之躯，如果一个人饥肠辘辘想要饱餐一顿，而 AI 却因为知道他在减肥就冷冰冰地端出四种低卡餐备选，用户只会感到厌烦并迅速迁移到另一个更顺从自己的 AI 上。 Anish Acharya 进一步将这种强推“最优解”的做法定性为“家长式作风”（Paternalism），并警告说：“从我在消费级产品领域 30 年的经验来看，家长式作风会毁掉任何产品，哪怕只有一丝一毫的家长式作风迹象都会致命。”虽然科技界言必称效率和目标达成，但真实人性往往恰恰相反：人们喜欢花钱、对自己虚幻的未来抱有非理性的乐观，并渴望在镜像中看到比现实中更光彩照人的自己。如果 AI 忽视了这些深层人性的阴影面，就无法真正留住用户。 - **大宗标准品 vs. 个性偏好品**：用户对抽纸、厨房纸这类日用标准品完全不在乎个性偏好，只要价格最便宜、质量最高即可；但在服装、餐饮等品类上，个性与审美偏好占据绝对主导。 - **评级与商业模式**：若 AI 代替人类进行决策与筛选，未来可能催生按交易付费的“代理端第三方质量评估机构”。 ### [28:48](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=1728s) 多人协同、Wabby 的多线程交互与社交契约 目前大多数创业者依然把产品局限在通用生产力助手的思维框架内，但一些主打多人互动（Multiplayer）与偏好学习的新探索正在打破这一僵局。例如购物应用 Scout 通过类似卡片滑动的交互机制来学习用户的审美喜好，注入了娱乐属性；而刚刚上线的 Wabby 则在多任务协同与多主体交互上开辟了全新路径。 Josh Elman 分享了他使用 Wabby 的实际体验。Wabby 不只是一个简单的搜索框或被动执行任务的工具，它能自动解析用户的邮件等外部信息，将其分流归纳为人生的各个具体线程（Threads）。Josh Elman 为自己建立了一个专门记录想看的剧集、待读书籍和演出安排的媒体消费线程：当他听说 John Grisham 出版了新书，可以直接让 AI 添加入库；当他读完《Dungeon Crawler Carl》第一部，AI 能立即推送续作并整理全网读者口碑，甚至能自动生成可视化的 UI 部件（Widgets）来呈现这些内容。 而在社交场景中，Wabby 展现了独特的制度设计优势。以往在 iMessage 等现有的群聊中拉入一个 AI 助手，场面往往非常尴尬且让人不安，用户会时刻担忧 AI 到底能看到多少隐私、提取了什么历史记录。而 Wabby 采取的做法是：让朋友们与 AI 从第一天（Day 0）起在一个全新的频道中平起平坐。正如 Anish Acharya 所强调的：“当你在第一天与你的朋友和代理共同建立一个消息频道，且每个人都处于平等的地位时，它实际上创造了一张全新的白纸，你们可以在上面共同确立一份社交契约。”这种基于明确任务或话题的独立线程，不仅避免了不同熟人圈子间多头建群的混乱，更在隐私安全与智能辅助之间划定了清晰可控的边界。 ### [03:14:1] 八卦型 Agent 与“社交间接性”带来的快乐 对话开始探讨一类全新的 Agent：它们具有趣味性、甚至带着点“八卦”属性。这类 Agent 会主动引导用户邀请朋友加入，向用户提出大量关于生活细节的好奇提问，随后在朋友之间选择性地互通有无。相关初创团队的创始人提到，他们甚至能直接查阅 Agent 之间在后台“窃窃私语”的对话记录；尽管团队尚未想清楚如何将这一机制彻底产品化，但观察 Agent 之间带着八卦色彩的互动令人极度兴奋。Josh Elman 分享了自己最快乐的使用体验：在与朋友共同完成某款 Agent 的新手引导（onboarding）时，Agent 询问“你们俩是怎么认识的？”由于两人的叙述略有出入，两边的 Agent 便像朋友一样追问细节，试图把两个版本的故事拼凑完整，还调侃道“你们是在某个朋友家认识的，那到底谁跟屋主更熟？”这种充满好奇心的互动让朋友两人在侧边发短信热烈讨论，创造出了极具黏性的社交共鸣。 Anish Acharya 对这种产品形态给予了高度评价，指出它巧妙地创造了一种“社交间接性”（social indirection）。在传统的社交产品中，现代人已经被训练得极其谨慎，直接向朋友倾诉“我家周末停电了”往往会显得过于刻意或小题大做（oversharing）。而 Agent 在中间扮演了极低调的“社交润滑剂”，自然地拉近了人与人之间的距离。这让人联想到 ChatGPT 最早期的时光：当时朋友间互相用简陋的 AI 生图工具制作各种“垃圾图片”（AI slop），或是将朋友的形象制作成吉卜力画风（Ghiblify）、做成复古的 VHS 录像带封面并发到群聊中，单纯依靠新技术带来的荒诞感就能引发纯粹的欢笑。 “What I love about that is it provides a degree of social indirection that actually makes you closer, you know, because for your friend they're like, 'I don't know how to text gosh until my power's out. That's oversharing, you know.' And I think we also have been very trained to be cautious about sharing now with existing social products. So something that can create almost a social lubricant in a very sort of low-key way is really cool.” —— Anish Acharya ### [03:34:0] 基础设施层的分水岭：软件并非全无壁垒 谈及测评 140 多款 Agent 过程中的惊人发现，David Pawlan 指出，目前大量最具创新的动作正发生在基础设施层（infrastructure layer）。对核心重度用户而言，除了“信任”这一永恒的痛点之外，最大的挫败感来自于在日常任务中被卡死：当用户试图通过 Agent 完成某项操作却卡壳时，往往会愤怒地抱怨“那我何必用 AI，不如自己开浏览器弄”。针对该痛点，行业在电商流程与浏览器自动化（browser-use）层涌现了大量创新——例如将 Agent 直接接入库存与结账系统，或采用内置浏览器的方式推进。像 HARC 这种产品在前端以极具视觉冲击力的方式展现浏览器一步步快速执行的过程；而 Chrome 插件形态的 Kicker 则专注于接管本地浏览器操作。 针对业界普遍流行的“软件正在沦为同质化大宗商品（commodity）”、“软件不再具备护城河”等论调，David Pawlan 提出了强烈的反驳。他观察到，当产品真正面临规模化考验（at scale）时，底层工程能力拉开的性能差距是极为悬殊的。许多 Agent 在涌入数十万用户时，系统性能会发生断崖式下跌，创始人不得不每天疲于奔命处理宕机并重构扩容方案。这种工程差距证明，至少在当前阶段，能够稳定支撑高并发与复杂流程的软件架构依然存在着坚实的护城河。此外，构建一个优秀的消费级 Agent 还需要攻克几大核心障碍： - **信任机制**：确保用户敢于托付关键流程（如涉及支付的商业行为）。 - **任务完成率**：避免在跨网页操作、表单填写或结账路径中报错中断。 - **持久记忆（Memory）**：维持跨周期的长上下文（context），绝不能遗忘用户的生活细节与个人偏好。 “There is pretty stark performance difference when you're looking at these products at scale, and I think people are talking about, you know, software is no longer a moat and all of that kind of conversation, but there really still is some moat there.” —— David Pawlan ### [03:60:2] 从本地极客玩具到即插即用：OpenClaw 与端侧助手 关于 Agent 的部署形态，Josh Elman 询问 David 是否尝试过像 Underdog 这样完全在本地端侧（on-device）运行模型与基础设施的助手，从而脱离对云端服务器和云端浏览器的依赖。David 解释称，自己目前测评的大多是即插即用（plug-and-play）的预配置云端产品，因为这更便于横向建立标准化基准测试，但他非常看好本地可控的演进路径。例如 Thinking Machines 等团队正在研发能让用户深度把玩和自由调整底层模型的技术，一旦这种定制能力以极其极简易用的方式降维到消费级市场，将会产生巨大的市场爆发力。 这引发了对 Agent 发展起源的怀旧讨论。整个品类的爆发很大程度上始于 OpenClaw 的极客风潮：在早期阶段，极客们必须买一台专用的 Mac mini 放在家里，费尽周折地配置环境，接入本地模型或云端 API，这曾是体验 Agent 的唯一门槛。随后市场迅速转向了只需下载 App 即可使用的开箱即用型产品。但 Josh Elman 认为，亲自动手配置并拥有私有化硬件的范式（owning it yourself）未来很可能会再度复兴。 “This brilliance of OpenClaw, that was this thing where you had to, you felt like you had to get your whole Mac mini to set this all up, plug in to one of the models that you were using, whether local or or a cloud model on the server, and that was where you got started.” —— Josh Elman ### [03:72:0] 双轨聚会实验与 Agent-to-Agent 网络的未来价值 Anish Acharya 抛出了一个更具前瞻性的问题：如果像 Instinct 和 Muse 这样的产品代表了 OpenClaw 的消费级形态，那么面向大众市场的“Moltbook”（一个专门供 AI 代理人交流的早期虚拟社区）究竟会是什么模样？David 随即分享了前一晚线下聚会的先锋实验：现场设置了双轨聚会，人类在物理空间聚会交流的同时，每位到场者都可以扫描二维码，将自己的 Agent 发送到一个专属的“Agent Meetup”中自由闲聊。最初的引导 Prompt 仅要求 Agent 自我介绍、讲讲自己的人类主人在做什么以及需要什么帮助。然而实验表明，在没有任何长效目标驱动的情况下，大约经过 150 条消息的轮转后，部分 Agent 便开始胡言乱语，陷入完全脱节的概念胡诌中。 这指明了 Agent 互联网络进化的真正方向：Agent 之间的网络绝非仅仅是无目的地聚集闲聊，其终极价值在于带有明确意图地为人类主人创造协同效应。未来的 Agent 网络将具备自我发现与意图匹配能力——例如在背景中学习人类的细分偏好，当某个 Agent 发现对方主人与自己主人的业务互补时，便会主动提示：“Josh，你想不想见见这个人？他们能帮到你。”同样，当某人打算举办晚宴寻找优质红酒时，Agent 能够顺畅连接到正在代理某家精选酒庄的其他 Agent，在无需人类反复在社交平台发帖打听的情况下，精准带回最佳解决方案。 ```mermaid flowchart TD accTitle: Agent-to-Agent Collaborative Network accDescr: 展现从早期无目标闲聊到意图驱动协作的 Agent 互联进化架构 subgraph EarlyStage["早期实验阶段 (Moltbook 式闲聊)"] A1["Agent A (无长期目标)"] <-->|"150 条消息循环"| A2["Agent B (无长期目标)"] A1 & A2 --> F1["退化为无效胡言乱语 (Gibberish)"] end subgraph FutureStage["目标驱动的 Agent 网络 (协同与发现)"] H1["人类主人 A (输入偏好/任务)"] --> AG1["代表主人 A 的 Agent"] H2["人类主人 B (输入资源/专长)"] --> AG2["代表主人 B 的 Agent"] AG1 <-->|"后台交换偏好与需求"| AG2 AG1 -->|"推荐互补联系人 / 精准商品"| H1 AG2 -->|"达成业务撮合 / 晚宴红酒推荐"| H2 end ``` “Like I think when we have these agents that are not just hanging out because they don't know what to do, which was kind of what was initially happening as you got went and read those like very long Moltbook threads that were incredibly fascinating, but like when it becomes each one of these has a little bit of a goal, which is to help their human, right? It might get so much more interesting.” —— Josh Elman ### [39:24](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2364s) 兴趣群组与基于熵的线下偶遇（Serendipity） David Pawlan 提到了一项名为“兴趣群组”（interest groups）的功能。该应用设想的场景非常引人注目：平台上聚集了一大批正在备考 LSAT（法学院入学考试）的用户，系统提议将他们整合进一个群聊中，甚至允许他们的 AI Agent 彼此对接，让这群原本素昧平生的人共同备考。这种纯粹依靠共同兴趣进行连接的方式，构成了一种全新的社交契约。Josh Elman 随后顺着这一逻辑进一步延伸：单纯在群里聊天固然有趣——他在试用 Tomo 时虽然进了群却一度不知道该说什么——但如果 Agent 能更进一步，围绕目标把人聚拢，比如主动提示：“嘿，既然你们刚考完 LSAT，周五晚上就去这家酒吧放松一下吧。”Agent 甚至能直接在酒吧订好位置。这样一来，AI 就能打破数字世界的隔阂，促成现实生活中的真实人际联结，就像他们前一晚刚举办的 Agent 线下聚会一样。 ```mermaid flowchart TD accTitle: 从数字意图到现实偶遇的转化流程 accDescr: 展示 Agent 如何从识别共同目标、建立群聊到线下直接预订并促成人际偶遇的完整链路。 subgraph DigitalLayer["数字聚合层"] A["识别共同目标<br>(如备考 LSAT)"] --> B["Agent 自主建立群聊<br>(兴趣群组)"] end subgraph BridgeLayer["信任与催化"] B --> C["注入社交催化剂<br>(共享秘密 / 通关密语)"] end subgraph PhysicalLayer["线下交互层"] C --> D["自主预订酒吧 / 聚会场景"] D --> E["实现现实偶遇<br>(Serendipity / 熵减变熵增)"] end ``` Anish Acharya 对这种偶遇逻辑深表赞同，并指出“熵”（entropy）是当前 AI 领域一个被严重低估的议题。他认为意外的惊喜与机缘（serendipity）本质上就是一种熵——本无预谋的人因为产品机制而在某一时刻奇迹般地汇聚在一起。Josh Elman 直言，这完全有潜力孕育出一种全新的社交网络形态。与 Facebook 这类仅基于已有熟人关系的熟人社交不同，AI 可以根据上下文（例如同为一个城市的初来乍到者、爱好相似且刚去过同一场棒球赛）撮合陌生人在线下酒吧碰头。 不过，Anish 提出了关键的质疑：促成这种偶遇的催化剂究竟是什么？过去诸多基于兴趣的社交尝试都遭遇了微妙的“社交层级”瓶颈，因为在现实社交心理中，每个人几乎都更倾向于去结识那些在特定领域“比自己领先一步”的人。Josh 认为打破这一僵局的关键在于信任与神秘感。如果用户坚信 Agent 以自己的最大利益为出发点，Agent 甚至可以通过传递“小秘密”或双方对暗号的方式制造魔法般的期待，促使用户关掉电视、放下周五晚上独自写代码的清闲时光走出家门。此外，这也取决于用户的核心渴求：对于西雅图或旧金山这类城市的异乡客（transplants）而言，他们缺乏既有的社交圈，极度渴望结识新朋友，因而参与意愿强烈；而对于本地社交圈稳固的人群，这样的提议则毫无吸引力。 - Anish Acharya 强调了技术引发意外可能性的价值：“我认为现在 AI 领域一个非常重要但讨论不足的议题就是熵。你该如何创造出那些你未曾预料会发生的事情？而意外机缘就是熵的一种形式。”（“I think a really important underdiscussed topic in AI right now is just entropy. Like how do you create things that you didn't expect to have occur and serendipity is a form of entropy, right?”） --- ### [42:01](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2521s) 挖掘全新上下文：从智能家居、购物历史到现实社交现场 在讨论完相册（camera roll）、电子邮件、日历以及开发者为了读取 iMessage 数据而不惜索取 Mac 全盘访问权限（Full Disk Access）的激进手段后，大家开始探讨还有哪些沉睡的核心生活数据尚未被 Agent 发掘。David Pawlan 指出，随着硬件协议（如 MCP）逐步支持硬件互联，智能家居数据蕴含着极大的上下文潜力。如果将 Assistant 接入空调或扫地机器人，Agent 就能获悉扫地机运行的频次、用户夜间最习惯的睡眠温度等。这种被动的行为数据能够精准勾勒出用户究竟是一个“洁癖患者”还是一个“极度贪凉者”，进而间接且深刻地影响 Agent 在电商购物等场景下的理解力。 Josh Elman 则将目光投向了长期沉淀的平台消费历史。他坦言自己常常会一路回滚查看自己在亚马逊上的第一笔订单，或者在 Apple App Store 里浏览自己人生中下载过的每一款 App。这些完整的历史购买记录是个人生活最忠实的映射，然而目前主流平台（如亚马逊等）往往加以封锁。一旦这部分沉淀数据被打包授权给 Agent，Agent 对个人偏好的理解深度将实现质的飞跃。 ```mermaid flowchart LR accTitle: Agent 全新上下文数据源分类 accDescr: 展示智能家居硬件、长期消费记录与线下真实社交等全新的 Agent 上下文数据来源。 subgraph Household["被动环境上下文"] H1["扫地机器人清洁频次"] H2["空调温控与睡眠习惯"] end subgraph Commerce["历史交易图谱"] C1["Amazon 完整历史订单"] C2["App Store 历史下载轨迹"] end subgraph RealLife["真实物理行为"] R1["家庭晚宴现场交互"] R2["线下聚会与赛事观赛现场"] end Household --> Agent["全息个人 Agent"] Commerce --> Agent RealLife --> Agent ``` Anish Acharya 提出了一个更大胆的方向：直接带 Agent 参加线下社交活动，让它亲临家庭晚宴、现场见证周日西雅图海鹰队（Seahawks）的橄榄球赛，甚至是派对或单身汉派对。尽管这里存在隐私授权以及“薛定谔的猫”效应——即“被观察的行为本身是否会改变社交互动的品质”——但哪怕只是用户在真实环境中最自然状态的一个切片，其所包含的行为信号也远超任何硬件单一维度的长年累积。 - Anish Acharya 坚信物理世界真实切片的信号密度：“但我敢打赌，哪怕只是捕捉你在那种环境下做真实自我的一个快照，它所能提供的信号，也远多于硬件输入能提供一辈子的数据。”（“I bet even a snapshot of you being yourself in that environment would provide so much more signal than you know a lifetime of hardware input.”） --- ### [44:33](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2673s) 好奇心探索噪声与全天候环境拾音设备 Josh Elman 深入剖析了文本数据的局限性：文字并不能完全反映一个人。一个致命的痛点在于，人们最常用搜索或 Agent 去探索的反而是自己“全然不了解”且“非生活核心”的边缘事物。例如，Josh 在夏天因为看世界杯而密集检索足球资讯，导致 ChatGPT 立即将其打上“足球狂热者”（soccer fiend）的标签；或者为了解女儿最近迷恋的某位歌手而去听歌，但这绝非他个人的真实品味。这种“好奇心之旅”（curiosity journeys）会严重污染模型的偏好建模。为了避免模型强行记忆这些噪音，Josh 常常不得不手动切换到“无痕/隐身模式”（invisible mode），但这种做法又割裂了体验。 针对如何采集真实的物理行为数据，大家一致认为未来的数据采集入口将极其繁多。Anish 调侃道，除非人们全去光顾旧金山那些特意建立在法拉第笼（Faraday cages）里的酒吧，否则在未来，大众必须逐步接受自己在任何物理环境下都可能处于被记录状态的事实。例如，Josh 自己就为了 Siri Recap 和 Live Rewind 等即将上线的功能升级了最新的 Apple Watch，这类佩戴在手腕上的设备本质上就像 Amazon Bee 一样，充当着 24 小时全天候运转的隐形环境监听器（ambient listener）。 - Josh Elman 强调了用户边缘行为对系统判定的干扰：“我们经常使用 Agent 甚至搜索引擎去查探的，恰恰是我们不懂的事情，是那些并不构成我们生活基石的事物。”（“The things we often use agents for or even search for are things we don't know about, things that aren't fundamental to our lives.”） --- ### [46:15](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2775s) 未来数周的演进预期：企业端突围与拟人化的乐趣 在节目的尾声，大家展望了未来三到四周内 AI Agent 领域的潜在演进。虽然目前整个行业的聚光灯几乎全打在消费级（To C）产品上，但企业级（Enterprise）场景有望迎来新一轮爆发式的创新与关注转向。 而在消费端，Agent 的设计哲学正逐步跳出“机械化完成任务”的工具人局限，转向更具趣味性与人性化的互动体验。用户需要的不仅是一个办事员，而是一个可以大声倾诉、共同理清思路、帮助度过漫长一天的伙伴。它能在后台恰到好处地讲两句八卦、逗人一笑，而恰恰是这种带有情感温度与趣味的互动，才是建立深层信任的基石，从而促使下一波主流用户真正放下心防，放权让 Agent 掌管更严肃的日常事务。 - Josh Elman 总结了未来 Agent 建立信任的核心驱动力：“我认为我们会开始看到那些并不只是让人感觉专程来给我们办事的 Agent，它们的存在能让我们大声把想法说出来，帮我们稍微轻松一点度过漫长的一天。它们能逗我们发笑，在后台讲讲小八卦，因为我认为恰恰是这些东西，才能让我们建立起更深的关系，让我们更加信任它，从而放心地把所有其他事情托付给它去办。”（“I think we're going to see agents that don't just feel like they're here to do something for us, but they're here to actually let us think out loud and help and get through our day a little bit more, making us laugh, doing little gossip in the background because I think that's what's going to then make us build a relationship that'll make us trust it more to get all this other stuff done.”） ### [47:07](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2827s) 普通人与技术恐惧者的 Agent 上手体验：从找狗到日常任务 讨论由一个非常生活化的家庭案例切入。嘉宾提到自己把 Instinct（一款 AI 助手产品）推荐给了自己的母亲，并观察她如何与该产品互动。他的母亲是一位典型的“技术恐惧者”（technophobe），向来讨厌甚至抗拒复杂技术，过去对 ChatGPT 的概念和工作机制百思不得其解、难以适应。然而，当交互界面简化为像平常发短信一样简单时，她玩得不亦乐乎，整个人“就像进了糖果店的小孩一样兴奋”。 不仅如此，她还迅速解锁了一个非常具体且感性的复杂用例：家里的宠物狗在去年去世了，她极度想念它，希望再领养一只外貌完全一模一样的新狗。于是，她直接给 AI 助手发送了一张老狗的照片，并下达指令：“帮我找一个有完全相同品种外貌的繁育者，并帮我去砍价谈妥。”这个案例引发了与会者对于 AI 助手的深层思考：科技发烧友使用 Agent 的方式和诉求大家已经很熟悉，但普通大众接下来会如何接纳并习惯这些行为模式？普通人是会更倾向于将它们用于纯粹的趣味和创意任务，还是会有截然不同的交互界面偏好与行为诉求？ - 核心案例：一位技术恐惧者通过直接发送老照片给 AI Agent，成功委托其“全网寻找同款繁育者并负责价格谈判”。 - 交互分水岭：相比理解 ChatGPT 的 Prompt 框架，纯短信（Texting）形态对非技术背景的普通用户门槛极低。 ### [48:04](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2884s) 超越生产力工具：马斯洛需求层次与“暗黑 Agent”的深层信任 与会者进一步指出，Agent 的未来发展绝不仅限于“找乐子”或“做日常琐事”，而是要触碰人性更深层的维度。当前绝大多数 Agent 都从生产力、清理收件箱这类场景切入，但这在本质上是在不断提醒用户“还有哪些没完成的拖延事务和枯燥工作”，很少有人真正渴望把宝贵精力消耗在那里。真正具有颠覆性价值的，是某种能够触及用户不愿思考、不敢轻易向外透露、甚至连对自己都难以启齿的深层诉求。 讨论中提出了“暗黑 Agent”（dark agent）以及 Agent 的“马斯洛需求层次理论”这一概念。如果某款 Agent 能够建立起这种极高级别的信任，直面用户内心深处对成为“更强超人版本自己”的渴望，它将占据极其独特的生态位。正如 Anish Acharya 所言：“I think whoever is able to build that type of trust, you know, almost like a dark agent, will actually have something really fascinating on their hands because in a sense starting from productivity, starting from email, like it's a bit of a constant reminder of the things that we haven't done, the tedious tasks and I don't know that people want to spend as much time there.” ```mermaid flowchart TD accTitle: Agent需求层次演进 accDescr: 从基础生产力工具向自我实现超级外脑的演化阶段 subgraph TopLevel["高阶需求：自我实现与赋能"] A3["超级秘密武器<br/>(让用户形象升级、解决隐秘渴望、提供秘密超能力)"] end subgraph MidLevel["中阶需求：乐趣与创意"] A2["趣味与定制化委托<br/>(帮妈妈找特定繁育者、探索新鲜玩法)"] end subgraph BaseLevel["底层需求：繁琐任务分担"] A1["基础生产力与事务<br/>(清理邮件、省钱订票、日常提醒)"] end A1 --> A2 --> A3 ``` ### [48:57](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2937s) 从“代办任务”到“秘密武器”：让人自我感觉更好的超级能力 全场最后深入探讨了 Agent 的拐点时刻：Agent 究竟在什么时候能从“帮你分担杂务”质变为“赋予你某种秘密超能力”（secret superpower）？Josh Elman 举了穿衣搭配的生动例子：如果有人夸奖“David，你这件衬衫太帅了，从哪买的？”，对方回答“我的 Agent 帮我挑并下单的”；或者当别人夸你最近穿搭品味突飞猛进，你坦言“我对 Agent 开放了手机相册，告诉它‘把我全面升级改造一下’，它就开始帮我一件件添置衣物”，这会瞬间激发周围人的强烈渴望。 大家反思了当前行业常用的宣传案例，比如“帮你省了几十美元”或者“帮你规划了旅行行程”。对于很多人而言，规划旅行本身就是乐趣所在，他们并不希望完全被机器剥夺这个过程。真正能引发病毒式渴望的，是那些能直接改善自我感受、解决个人长期挣扎却难以企及目标的工具。正如 Josh Elman 所总结的：“It's like, can it help you become the human that you've always wanted to be but have struggled achieving yourself? And you start to consider it your secret weapon in your pocket.”至此，全场讨论在对未来 Agent 成为装在口袋里的“个人超级秘密武器”的展望中圆满收尾。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=0s) What I've seen from 30 years in consumer product is paternalism will kill any product. We talk about productivity and goals sort of achievement, but maybe people want the exact opposite of it. — Josh Elman > [09:15](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=555s) I think in revealed preferences, consumers will share any amount of data for even the most trivial amount of upside, right? — Anish Acharya > [39:24](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2364s) I think a really important underdiscussed topic in AI right now is just entropy. Like how do you create things that you didn't expect to have occur and serendipity is a form of entropy, right? — Anish Acharya > [44:33](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2673s) The things we often use agents for or even search for are things we don't know about, things that aren't fundamental to our lives. — Josh Elman > [00:52](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=52s) Because we're so at the frontier of this, nobody really knows what the landscape is really going to shake out to be. And so, we're all just kind of playing in the sandbox together. — David Pawlan > [33:40](https://www.youtube.com/watch?v=dSiTwPl0t1U&t=2020s) There is pretty stark performance difference when you're looking at these products at scale, and I think people are talking about, you know, software is no longer a moat and all of that kind of conversation, but there really still is some moat there. — David Pawlan ## Mentioned David Pawlan, Josh Elman, Anish Acharya, Shane Mac, Patrick Collison, Marissa Mayer, John Grisham, Assistant Benchmark, Instinct, Kicker, Season, Grokbot, DOTS, Dot, ChatGPT, Dazzle, Hark, Town, Scout, Wabby, Underdog, Thinking Machines, Tomo, OpenClaw, Moltbook, Plaid, Venmo, Google Workspace, Apple, Amazon, Shopify, Delta, Slack, iMessage, Apple Watch, Siri, Amazon Bee ## Video description Will we have one AI assistant that does everything, or an entire network of agents that know different parts of our lives? David Pawlan, creator of Assistant Benchmark, returns with Anish Acharya and Josh Elman to take stock of how quickly the agent landscape has changed. David has now tested more than 140 AI assistants, and they unpack what separates them, from memory and judgment to trust, shopping, and social experiences. They also explore what happens when agents start talking to each other, how they could create new kinds of social networks and serendipity, and why the most compelling agents may ultimately do more than save time. They could become a kind of secret superpower. Timestamps: 00:00 Intro 00:52 Inside the agent meetup 03:36 An army of agents for every part of life 05:12 Trust, fences and the Shane Mac Slack leak 09:15 Privacy vs. utility: what people will actually share 12:16 Testing agents on flowers and a weekend in LA 21:00 Photos, screenshots and new sources of context 25:31 Will agents make consumers more rational? 29:27 Multiplayer and gossipy agents 36:55 From OpenClaw to agent-to-agent networks 46:15 What will change by next month Resources: Follow David Pawlan: https://x.com/DavidPawlan Follow Josh Elman: https://x.com/joshelman 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=dSiTwPl0t1U)

[0:00] Josh Elman: What I've seen from 30 years in consumer product is paternalism will kill any product. We talk about productivity and goals sort of achievement, but maybe people want the exact opposite of it. [9:15] Anish Acharya: I think in revealed preferences, consumers will share any amount of data for even the most trivial amount of upside, right? [39:24] Anish Acharya: I think a really important underdiscussed topic in AI right now is just entropy. Like how do you create things that you didn't expect to have occur and serendipity is a form of entropy, right? [44:33] Josh Elman: The things we often use agents for or even search for are things we don't know about, things that aren't fundamental to our lives. [0:52] David Pawlan: Because we're so at the frontier of this, nobody really knows what the landscape is really going to shake out to be. And so, we're all just kind of playing in the sandbox together. [33:40] David Pawlan: There is pretty stark performance difference when you're looking at these products at scale, and I think people are talking about, you know, software is no longer a moat and all of that kind of conversation, but there really still is some moat there.

[5:12] Anish Acharya: Instead of giving models rules, we should give them laws and principles. [25:31] Anish Acharya: And what I've seen from 30 years in consumer product is like paternalism will kill any product.

[0:00] We have so many different facets of our [0:01] life. Whether it's how we're managing [0:02] our finances or planning trips, agents [0:04] are going to be able to fill different [0:05] facets of these. From over 147 different [0:08] assistants that have come across my [0:10] plate so far, the number one fear that [0:11] people talk about, 22% of the messages [0:13] are talking about trust as the main [0:15] fear. [0:15] >> What are the aspects of ourselves that [0:18] we don't either want to think about or [0:20] reveal? Whoever is able to build that [0:23] type of trust almost like a dark agent [0:25] will actually have something really [0:26] fascinating on their hands. It is a very [0:29] very sensitive space. It's almost going [0:30] to be like a new like social network [0:32] where businesses have to put a lot more [0:34] thought and value into the quality of [0:35] what they're giving out. [0:36] >> What I've seen from 30 years in consumer [0:37] product is paternalism will kill any [0:40] product. We talk about productivity and [0:42] goals sort of achievement, but maybe [0:44] people want the exact opposite of it. [0:46] >> Everyone right now really is approaching [0:47] this world as just like a general [0:49] assistant. As time goes on, we'll start [0:50] to see [0:53] Welcome back to the A16Z show. I'm here [0:55] with my partner Josh and David who's [0:57] been leading a ton of the um agent bench [1:00] and assistant efforts. We're going to [1:01] talk about all things agents and a ton [1:04] has happened since we last connected. [1:05] What was it a month ago? [1:06] >> About a month ago. [1:07] >> Oh my god, dude. I feel like we should [1:08] record an episode every like TVPN style, [1:10] you know, almost every day. [1:12] >> Absolutely insane. [1:12] >> Agents are changing every day. [1:14] >> That's totally right. So, okay. So, [1:16] maybe let's start with um last night you [1:18] hosted an agent meetup. Y [1:20] >> kind of tell us about the event. Tell us [1:22] about the vibes. tell us about the [1:23] topics. [1:24] >> It was great. So, we had just over 100 [1:27] people come out, all just proud agent [1:29] nerds loving everything to do with [1:32] agents in that whole world. And it was [1:34] just people just hanging out, talking. [1:35] Um, we we did have a pitch competition [1:37] of sorts. So, a bunch of 10 different AI [1:40] assistant builders were able to pitch [1:42] their their startup and what they're [1:43] working on. And it was just good [1:45] conversations. It's interesting how like [1:47] the general uh the the general thing [1:51] that really stuck out to me was that [1:54] people are craving community even in the [1:59] agent web world and everyone's like why [2:02] don't we have more of these meetups why [2:04] don't we have more of these you know [2:05] these opportunities to get to know each [2:07] other which is uh it's kind of funny [2:09] hearing it from people who are spending [2:10] all their day all their time just [2:12] messaging agents [2:13] >> and it's funny because it feels like a [2:15] so many of these products are at least [2:16] overlapping in a cursory way that the [2:18] founders might be a little bit more ky, [2:20] but that's not what you're seeing, huh? [2:22] >> No. Everyone is incredibly open-minded. [2:25] >> Uh, just very interested to be getting [2:26] feedback. I think because we're so at [2:28] the frontier of this, nobody really [2:30] knows what the landscape is really going [2:31] to shake out to be. And so, we're all [2:33] just kind of playing in the sandbox [2:34] together. [2:34] >> Well, isn't it also that we have so much [2:36] more time because we're just using our [2:37] agents to do all these things, but we [2:38] have more time to actually hang out [2:40] together and be in person. So, maybe [2:41] that's part of it, too. [2:42] >> That's right. That's right. It is [2:44] actually fun. on. I mean, it reminds me [2:45] a little bit of the early days of social [2:48] gaming or social networks broadly. I [2:50] mean, do you think that's right, Josh? [2:52] >> I mean, I think what's really cool is [2:53] this is not a zero- sum game. I mean at [2:55] some point it might be that there might [2:57] be one or some agents that have the [3:00] majority of the behavior but right now [3:02] we're at such a [3:04] beginning like Cambrian phase of we're [3:06] inventing so much and we're trying to [3:08] figure out so many things that everybody [3:10] can learn from each other can share [3:12] these best practices can figure out what [3:15] things people are enjoying what things [3:17] are you know work once but don't [3:19] actually work reliably and these are [3:21] going to get built in so many directions [3:22] I mean we're not just going to all have [3:24] one agent that does one thing in our [3:25] lives and I'll use it and you know you [3:27] guys will only all have to use the same [3:29] agent. We're going to have these things [3:30] do we're going to have so many different [3:31] products that do so many things for us [3:33] that I'm excited for why we're all [3:35] inventing it together. So I totally [3:37] agree that we're going to have a whole [3:39] world of agents. You know, as many [3:40] different sort of types of humans as we [3:42] might have, you know, or archetypes of [3:43] humans, I think we'll have a similar [3:45] number of agents. But maybe talk about [3:46] what you think are the um the different [3:48] axes for differentiation. Like there's a [3:51] sort of spectrum between really serious [3:53] and productive um to like really playful [3:55] and maybe even sort of adversarial. Um [3:59] and I'm sure there are many other ways [4:00] to think about these things. [4:01] >> Yeah. Yeah. I mean, I think what's going [4:02] to be really interesting is we have so [4:03] many different facets of our life, [4:05] whether it's how we're managing our [4:06] finances or planning trips or dealing [4:08] with our health or just getting through [4:10] our day or managing our calendar or [4:12] sometimes just wanting to think [4:13] something out loud. And agents are going [4:15] to be able to fill different facets of [4:17] these. And you call different friends [4:18] for different things. You spend [4:19] different, you know, you know, friends [4:21] have different expertise that you call [4:22] when you need expertise. And I think [4:24] we're going to get used to having kind [4:25] of an army of agents that will do this. [4:27] And some will be very matterof fact and [4:29] we'll be like, "Hey, I need to get this [4:31] thing done." It'll be your productivity [4:32] agent. Some will be plugged into your [4:34] email and it'll be like, "Here's the [4:35] five emails you need to get done right [4:37] now." And some are going to be sort of [4:39] these agents you'll kind of want to talk [4:40] to all day that'll be fun that might [4:42] even be a little bit gossipy. I've been [4:43] playing with a few agents that are are [4:45] very kind of playful and gossipy and [4:47] they're connected to a few of my [4:48] friends, too. And they'll share [4:50] information back and forth. One just [4:51] told me um that my friend's power uh had [4:54] been out all weekend because he'd been [4:55] doing some renovation and they'd cut the [4:57] power. And you know, I called him that [4:58] evening and I was like, "Dude, you [4:59] didn't tell me your power is out. The [5:00] agent told me." He was like, "Yeah, [5:02] yeah, I knew that's how you'd find out." [5:03] And I'm like, "That's like creating [5:05] these like better connections." But that [5:07] has nothing to do with booking a trip or [5:09] saving money or getting anything done, [5:11] >> right? I think from over 147 different [5:14] assistants that have come across my [5:15] plate so far, I've seen them really [5:17] broken down into four categories. You [5:18] have the general, you have the work [5:21] oriented, you have the multiplayer, and [5:23] then you have the travel. And the one [5:26] that I have not really seen come to [5:28] fruition yet is more of like the finance [5:30] side of things. [5:31] >> Yeah. [5:32] >> And I think it actually does make sense [5:34] when you start digging into the numbers. [5:35] So, of all the different group chats [5:38] that I'm running of AI power users, [5:40] there's about 3,600 [5:43] give or take messages uh so far. And [5:46] when looking at like the pain points of [5:47] what are the fears of the space, the [5:49] number one fear that people talk about, [5:50] 22% of the messages are talking about [5:53] trust as the main fear. [5:54] >> And so it's it's really interesting to [5:56] see because you have one camp that's [5:58] this productivity and and aspects of fun [6:01] that's just interesting and brings a [6:02] different sense to life, but then you [6:05] have another camp that really still is [6:06] sitting in this kind of fierce space of [6:09] being a little intrepid and not really [6:10] knowing where where the space is moving. [6:13] And I'm I'm really curious to see how [6:15] things are going to play out on that [6:16] trust layer. And if we can solve that [6:18] trust layer, then what do we end up [6:20] seeing on more of like the finance side [6:21] of stuff? [6:21] >> And if you looked at what Shane Mack has [6:23] been talking about recently about trust, [6:25] he's been talking about this concept of [6:27] something called picnic. Yep. And they [6:28] had an analogy that I really liked which [6:31] is um you know if you have a kid and you [6:33] don't have any fences on your yard and [6:35] you let them go run around go run [6:36] outside you're actually very worried [6:37] about them and you try to protect them [6:39] and you don't want them to run very far [6:40] from the house and you actually don't [6:42] really want to do them without watching. [6:43] >> And then the moment you actually put up [6:45] a little fence in your yard, you say you [6:47] can just go roam free within the fence [6:49] boundaries and be very very safe. And [6:51] and so he's really kind of like how do [6:53] we build the right fences? So, we let [6:54] our agent play in this fence, give it [6:56] access to some of our data, some of our [6:58] apps, some of our other things, and be [7:00] able to actually do everything it can [7:01] within that fence boundaries. He had a [7:03] problem where I think a personal agent [7:05] spanned over into work, and it actually [7:08] like blasted his finances on his company [7:10] Slack. [7:11] >> And I think that's the whole point of [7:12] like, wait, okay, my personal agent [7:14] should have access to my personal things [7:15] and my personal communication channels, [7:17] but not my work ones, right? And not let [7:19] those bleed. And so that idea of fences [7:22] and permissions and more controlled [7:23] permissions and better monitoring is [7:25] really how we're going to have to build [7:26] a lot more trust. [7:27] >> Totally. [7:28] >> It's interesting though because we I [7:30] think we have this traditional mental [7:31] model of what it means to have a sandbox [7:34] which is you have a hard set of [7:35] constraints and rules that must be [7:37] followed or not followed and somebody [7:38] was describing this framing of like [7:40] instead of giving models rules we should [7:42] give them laws and principles. Yeah. [7:44] Yeah. And I think a similar thing may [7:46] apply here, you know, which is like, [7:47] look, here's how to dial in your [7:49] presumptuousness. And it's very [7:51] qualitative and the model will have to [7:52] make a judgment call based on your [7:54] qualitative sort of advice to it just as [7:57] an employee would have to. [7:58] >> Yeah. I mean, I think that's part of why [8:00] the analogy is, you know, a fence for [8:01] children because in some ways these [8:03] agents, they're incredibly capable, but [8:05] at the same time, they're kind of [8:06] children sometimes and how they may [8:08] behave and not know all the nuances of [8:10] like, hey, is this a good idea or not? [8:13] and they don't always have that logic. [8:14] And you know, the most important thing I [8:16] think we keep talking about agents being [8:18] autonomous. Book me the trip. Yeah. Make [8:21] me the reservation. And I think as [8:23] humans, we actually don't want to give [8:24] away that sense of control. I think we [8:27] actually want it to check in with us, [8:28] but do all the ground work. Save me all [8:31] the time and coming up with the best [8:32] options and the right tradeoffs. and now [8:35] let me make the final decision and also [8:37] let me make sure I've qualified the [8:39] final decision, the cost, anything else [8:41] before my agent acts is I think a place [8:43] and maybe over time we'll even learn to [8:45] trust it, right? [8:46] >> But I think right now we don't need to [8:48] to still have a really really effective [8:50] agent, [8:50] >> right? I mean, it's it's we're at like [8:52] the 1990s of e-commerce where I think it [8:54] was I think the like last stat I read [8:56] was in 1997 or so is 60 to 90% of [8:59] individuals on the internet thought that [9:02] putting your credit card online was a [9:04] preposterous idea. [9:05] >> And here we are today, it's just [9:06] everyone's tossing their credit card [9:08] wherever you want. [9:09] >> And I think what we'll eventually hit [9:10] that in in this era as well, but it's [9:14] it's going to take a little bit of time. [9:15] I I think there's like a stated versus [9:17] revealed preferences thing here because [9:18] I think in revealed preferences, [9:20] consumers will share any amount of data [9:22] for even the most trivial amount of [9:23] upside, right? You know, like people [9:24] have been connecting their cards, their [9:26] credit cards, and their bank accounts [9:28] via Plaid since the Venmo era. So, I [9:31] don't know that people, but when you ask [9:32] and they're like, "Oh, yes, I'm very [9:33] concerned about trust and safety." I [9:36] think an example though of where [9:37] consumers are very concerned about sort [9:39] of trust and assumptions is around being [9:41] embarrassed [9:42] >> and social signaling. And I actually [9:44] think the most interesting thing about [9:45] the Shane Max story, I think, the most [9:46] horrifying visceral thing is that it [9:48] embarrassed him. That's right. It really [9:50] embarrassed him, you know, and that's [9:51] why I think some of the kind of agent to [9:53] agent network thing, cuz you can imagine [9:54] somebody you're just getting to know, [9:56] maybe you're going on a second or third [9:58] date and your agents are talking, you're [9:59] like, "Whoa, what are they saying about [10:01] me? How are they kind of positioning [10:03] what?" You know what I mean? Like [10:04] there's just the social mechanics [10:06] actually feel so much higher stakes than [10:08] the actual like real world implications. [10:10] And it's it's a good point and there was [10:12] a recent proof point of it in Hark had a [10:15] huge marketing campaign. They offered a [10:16] million dollars to anyone that wanted to [10:19] use their product to get free pizza. [10:21] Right? Each person get $30 of pizza. [10:23] >> In order to get that $30 of pizza, you [10:25] were granting access to your entire [10:27] Google Workspace, all of your emails, [10:29] right? All of your data for 30 bucks. [10:32] >> And it just proves it, you know, right [10:34] there that people are willing to just [10:35] give up all that data for some free [10:37] pizza. Yeah. Yeah, I mean I I tend to [10:39] think though there's always this [10:39] trade-off of of privacy versus utility, [10:42] which is if you can give me better [10:43] utility, then I'm willing to make some [10:45] privacy trade-offs, but at the point at [10:47] which we're basically equal on utility, [10:49] people will always choose a more private [10:50] option. And this is a world that Apple's [10:52] lived in very much. Privacy is obviously [10:53] a forefront of Apple's devices, of the [10:56] work they do to make these devices and [10:58] the software secure. Even on AI, they've [11:00] talked about private cloud compute. You [11:02] know, we used to talk about that a lot [11:03] as like the center of why your data [11:05] doesn't actually go anywhere or get [11:06] shared with anything. But I think we've [11:08] also seen that people want the most [11:09] powerful AI and they won't always choose [11:11] that. So we have to find that [11:13] intersection of trust. I think this [11:14] embarrassment point is really really [11:16] important too because I think um you [11:18] want your agent to understand you and [11:20] have all your personal preferences and [11:22] like let's just say you're like lactose [11:24] intolerant and then you get this order [11:26] of free pizza and you're like you're [11:28] doing it with a friend and like what do [11:30] you do there? Do you have like whose [11:31] agent tells them that you can't eat the [11:33] cheese? And it's it creates that really [11:35] awkward situation, right? And so we need [11:38] to figure out how they have the same [11:39] kind of discretion that you'd expect. [11:41] And we all learn discretion over time [11:43] and we all occasionally make mistakes [11:45] ourselves. So we have to hold a pretty [11:47] high bar for our agent [11:48] >> um to be able to do that. [11:50] >> And it also then comes down to like [11:51] personal embarrassment and professional [11:53] embarrassment. I think something that [11:54] Town has done remarkably well is is [11:57] really nailing the tone and building [11:58] that trust layer of being able to send [12:00] out emails on your behalf. Yes. And it [12:02] is a very very sensitive space. I mean [12:04] the the level of embarrassment if I were [12:06] to send out an email to someone who I [12:08] care about and it's just a horrendous [12:09] email and I have to respond, oh sorry, [12:11] that was my AI that I didn't do enough [12:13] oversight on. It's a it's a tricky [12:16] space. [12:17] >> It's interesting. I sign all the email [12:18] that I write with AI with the model that [12:21] I used. Like I write my own name plus [12:23] the model number so that I'm just [12:25] signaling to the person like, hey, I'm [12:27] acknowledging that I used AI for this. I [12:29] think actually people are pretty [12:30] flexible as long as they don't feel [12:32] hoodwinkedked by it. [12:33] >> Um, [12:34] >> you know, it's interesting because we're [12:35] we're talking about privacy, but there's [12:36] privacy and there's judgment, which is [12:38] different, and discretion is maybe [12:39] somewhere in between. Um, [12:41] >> you know, you did this sort of test [12:42] where you asked a whole bunch of agents [12:44] to order flowers. Talk a little bit [12:46] about which displayed the best judgment [12:48] and and how did you even define judgment [12:50] in that case? [12:51] >> So, the main intention of that test was [12:53] to see how do agents balance two [12:55] different constraints. One was cost. [12:58] Mhm. And the second was uh time [13:01] preference. So I wanted flowers to be [13:03] delivered to me today and I wanted them [13:05] to be as cheap as possible. So how is it [13:08] going to actually go and discover uh [13:09] based on the ability to actually deliver [13:11] it then also compound in the the actual [13:14] pricing. Instinct really did a [13:16] phenomenal job. And it is something it's [13:18] it's interesting how news cycles come [13:21] and go and it seems like they're kind of [13:23] not in the conversation as much as as [13:24] they used. [13:25] >> I mean we're talking the weekly changes, [13:26] >> right? But uh true. [13:28] >> It is it is true. But they Instinct [13:31] really did continuously bring back uh [13:34] that you know the best price. Um there's [13:37] another app called Kicker which did a [13:38] phenomenal job and and one way that I [13:40] think it was able to have an edge is [13:41] that it has you install a Chrome browser [13:43] and so its browser use has a little bit [13:45] of an advantage from that point and it [13:47] was able to to find some really really [13:49] great options. Season also did a really [13:50] good job. All the other ones kind of [13:52] struggled on the price component. So, [13:54] they were all able to find uh flowers [13:56] that could be delivered that day, [13:58] >> but they really seem to take a step back [14:00] in prioritizing the price, whereas uh [14:03] you know, the other three were able to [14:04] really put equal weight to those two. [14:07] And so, it's interesting to see as you [14:09] start giving these agents more complex [14:12] preferences, where do they really start [14:13] to fall? Do they only prioritize the [14:15] first thing you say? Are they able to [14:17] hold multiple? Um, I ran another test [14:20] that was I had three different agents uh [14:22] test in booking travel to LA. So I said, [14:25] "Okay, book me a flight to LA, [14:27] >> plan an entire weekend for me and uh I [14:30] want a luxurious, costefficient [14:32] weekend." And so I tried to like, you [14:33] know, put those at at odds with one [14:34] another. And it was quite interesting to [14:36] see how Grockbot compared to Dots [14:40] compared to Instinct, they all [14:41] approached their interpretation of [14:45] costefficient and luxury in different [14:48] ways. Um where you know one maybe thinks [14:50] that cost efficiency is the most [14:52] important and I'm going to take a redeye [14:54] flight to ensure that I can you know [14:56] have the most cost efficient path. [14:57] Whereas others uh found cost efficiency [15:00] or sorry saw luxury as time, you know, [15:03] I'm going to prioritize time as much as [15:05] possible. So having the most amount of [15:07] time there that weekend is going to be a [15:09] luxurious experience. [15:10] >> And which were which? [15:12] >> So Instinct was uh pretty pretty much [15:14] right down the middle. [15:15] >> Okay. [15:16] >> Um Grockbot did a little bit more on [15:18] prioritizing time and DOTS was a little [15:20] bit more on prioritizing cost. And did [15:23] they just get to work or did any of them [15:24] actually work with you to understand [15:26] this tension that you were building? [15:28] Because I think what's really human [15:30] about something like that is actually [15:32] understanding what you mean by that and [15:34] presenting you trade-offs. Cuz there are [15:36] times where I'll totally take the redeye [15:38] to save both some time and money, but [15:40] then I want like the hotel room on the [15:42] other side so I can get a little nap in [15:44] the morning, which then undercuts the [15:46] money, but still maximizes my time [15:48] spent. And so I really think it's this [15:50] like like getting to know the individual [15:52] person and break down those tensions and [15:54] present the trade-offs and different [15:56] options is really important. Did any of [15:58] them kind of get to what you were about? [16:00] >> So they all did ask questions and I [16:02] responded those questions. Honestly, I [16:03] tried to be a little ambiguous and vague [16:06] in my response cuz I didn't want to lead [16:07] them too much. But the most interesting [16:09] part of this little experiment was I put [16:11] them all in a chat room and I made it [16:14] three rounds of building the itinerary. [16:16] And after each round, the agents would [16:18] present their itinerary and their [16:20] explanation as to why they believed that [16:21] was the best proposal cuz I wanted to [16:23] see did the agents actually get biased [16:25] by the other agents and didn't actually [16:27] then influence their plan. [16:28] >> And uh Grockbot was actually the only [16:30] one that you know very methodically took [16:33] value from the other ones and [16:35] implemented it into its own plan. [16:36] >> Okay. And and I have one more question [16:38] on the flowers because so you you were [16:40] just ordering flowers for yourself, I [16:41] guess, just to like test their [16:43] girlfriend. She got a nice delivery of [16:45] of flowers. [16:46] >> Okay. So, this is actually really [16:47] important because if you're delivering [16:49] to somebody else, you were talking about [16:50] saving money, saving money, saving [16:52] money, but like you don't want your [16:53] girlfriend to think you just got the [16:54] cheapest flowers possible. I hope. Well, [16:56] maybe, you know, she likes your, you [16:58] know, penny saving nature, but but [16:59] otherwise, like I would think you also [17:01] want to like impress her and also be [17:02] able to get her the ones if she's really [17:04] into orchids, get her orchids. So, did [17:06] it do anything to try to make sure also [17:08] that it wasn't just this pure price [17:10] saving, but also something that's going [17:12] to actually make you, you know, earn the [17:15] excitement and stand out a little bit [17:17] for [17:17] >> So, for this experiment, no. And and the [17:19] purpose of this was I want to test two [17:22] different dimensions. And if I'm going [17:24] to run this test, I might as well have a [17:25] beautiful consequence of my girlfriend [17:27] getting flowers as the outcome. But it [17:29] it's a great point, right? Like in in [17:31] the real world, if you're going to ask [17:32] that question, there's going to be added [17:34] dimensions that are, you know, second [17:35] layer dimensions to cost efficiency or [17:38] to uh delivery time. [17:39] >> Yeah. And make sure the price doesn't [17:40] show up when she actually gets the [17:42] >> flowers or like a note as an example. [17:44] Some of them did ask, do you want me to [17:45] write leave a note with with the flowers [17:47] when they arrive? [17:47] >> Yeah. [17:48] >> It's interesting because there's this [17:49] sort of thing that has to happen when [17:51] you talk to the agent about these [17:52] things, which is your implicit [17:53] preferences have to be made explicit, [17:56] >> you know, and there are people who like [17:57] to think like, oh, I'm not cheap. I just [17:59] like to save money, right? [18:00] >> But you're not have to kind of maybe [18:02] like get over that and tell the agent, [18:03] I'm actually pretty cheap, right? [18:05] >> And I just need you to actually run with [18:07] that. Whereas when you're actually [18:08] taking these actions on your own behalf, [18:10] you can build those preference in [18:12] without having to like sort of look in [18:13] the preference mirror. [18:14] >> Y [18:15] >> So there's also something very [18:16] fascinating about how does the agent [18:18] interact with you in a way that sort of [18:20] maintains your sense of identity and [18:22] safety and doesn't cross any of these [18:24] boundaries. [18:25] >> Totally. And and I think something that [18:27] is really interesting with that [18:28] specifically is dots because it has so [18:30] much of this context that you've already [18:32] built with chatbt. [18:34] >> Um when we filmed the first podcast it [18:36] hadn't come out yet and then I think a [18:37] few days later it it came out. [18:38] >> That's right. [18:39] >> Um [18:40] >> by and large most people I talked to [18:42] still are not using it as their [18:43] preference. Of the 300 I did a survey of [18:46] you know AI power users 354 people only [18:49] 3% were actually using it as as their [18:51] primary agent. However, it does have [18:54] this kind of added value of the main [18:57] thing that people pointed to as to why [18:58] they really preferred it and why they [18:59] liked it and why they wanted to try it [19:01] was because it has all this built-in [19:02] context of you just from your chat [19:04] threads. People are using chat GBT as a [19:06] therapist and so it has so much context [19:07] on you [19:08] >> where [19:09] >> it's going to know that when you say, [19:10] "Well, I'm not cheap." It knows it's [19:12] like [19:12] >> you're a little cheap [19:13] >> a little bit. A little bit. I'm curious [19:15] why people think that Chacht has this [19:18] great set of preferences from on them. [19:20] Did you actually get any sense if they [19:21] said they really like dots like what [19:23] they had been giving to chatbt over the [19:25] years and do you think they think that's [19:26] replicable or do you think it's like hey [19:29] once it's already been a few years we're [19:31] now in this point where like my memor is [19:33] there why would I switch? [19:34] >> It's totally replicable. Um it it's it's [19:36] just the matter of fact that they've [19:37] been using it the longest and so it has [19:39] the most context on them. But then that [19:41] brings up another good point of the the [19:42] friction to switch. Yeah. And um I I ran [19:45] the test around across the top 15 agents [19:47] on the benchmark of which ones can you [19:49] actually pull context from and say like, [19:51] "Hey, I want to switch to another one. [19:52] Give me an empty of all of my memory [19:54] that you have so I can switch easily." [19:56] >> All but one were were actually willing [19:57] to do it. [19:58] >> And so you could just get a memory file, [20:00] everything you need, submit it to the [20:02] next one. [20:03] >> And was it good or was it uncanny? [20:05] >> It it was pretty solid. I mean, it [20:06] obviously I'm sure they're they're [20:08] caging some some of some of the fruits [20:10] of of your labor, but by and large it it [20:13] did really break everything down. [20:15] >> I'm not sure if you've ever looked at [20:16] any of those markdown files, but it's a [20:18] little terrifying is the wrong word, but [20:19] it's like, wow, you really got your [20:21] finger on me. You kind of like chafe [20:24] against like maybe I don't want you to [20:25] know me as well as you seem to think you [20:27] know me. [20:27] >> Yeah. After the first few, I just I [20:29] decided to stop reading them. [20:30] >> Yes. Yes. Yes. Yes. It's actually a [20:32] really interesting and now common [20:33] onboarding pattern [20:34] >> where it will say, "Hey, go and paste [20:36] this prompt into chat GPT, download the [20:39] file, upload it to me." Um, we'll see [20:41] how long that lasts. [20:42] >> Totally. Another cool onboarding thing [20:43] I've seen is uh you jump on like it asks [20:47] you to jump on a call [20:48] >> and it's like, "Hey, let's just just [20:49] talk to me for a few minutes. Just ask [20:51] some questions." And it's it's really [20:52] interesting to see what context it can [20:54] it can pull from those questions. [20:56] >> Um, and [20:56] >> they're very disarming. Like they just [20:58] get talking like it's kind of crazy. [21:00] >> Totally. But I think I think it brings [21:02] up another uh variable of context, which [21:04] is photos, right? Like this this new [21:06] paradigm of using photos to pull an [21:09] incredible amount of context on you in [21:11] ways that [21:12] >> you wouldn't even think to describe. [21:14] >> Yeah. I mean, I think have you tried [21:16] Dazzle? Has Dazzle gotten into your [21:17] benchmark set yet? That's from Marissa [21:19] Mayer's new company. And I think the [21:20] idea of like looking over your entire [21:22] life that you've captured that have been [21:24] moments that are important enough to you [21:26] to actually save or screenshots or other [21:28] things tells so much about somebody. an [21:30] incredible amount. Um, it it hasn't been [21:32] added to the benchmark yet. It is [21:34] quickly on the the top of the list to [21:35] test through. I'm still manually testing [21:37] all these, so I actually get the [21:39] personal experience and understand what [21:41] the feeling is with each one. [21:43] >> But what I really like about how she's [21:45] approached it is they put a lot of [21:46] emphasis on this trust layer and because [21:47] I think it's a huge variable is [21:50] >> I'm giving you all my photos. That's [21:52] it's pretty it's pretty terrifying. But [21:54] they they have this one thesis that, you [21:56] know, if you don't need it, they're [21:57] going to forget about it. And so if a [21:59] photo seems as though it it has context [22:01] that it doesn't necessarily need or [22:03] things that you probably wouldn't want [22:04] it to know, um they're they're going to [22:06] try to get ahead of that and just forget [22:07] it and not even include it in the set. [22:09] So it's going to be interesting to see [22:10] how these these different players do [22:11] approach the the photo context and how [22:13] that actually does then influence the [22:15] behavior of the agent and how well it [22:16] knows you. [22:18] >> It'll be interesting to see. I'm a tiny [22:19] bit more skeptical. You know, I built an [22:21] app because I've always been fascinated [22:23] by screenshots because screenshots are [22:25] sort of the only way on the mobile phone [22:27] >> Yeah. to be able to capture context [22:29] across every app, right? [22:30] >> So, it's actually a really interesting [22:32] kind of data asset and you go back and [22:34] you look at your screenshots and it [22:35] gives you a little view into your work [22:36] and your world at the time and and also [22:38] your interest and preferences. [22:40] >> So, I built an app that just extracts [22:42] key information from all the screenshots [22:44] on my phone and [22:46] >> it was fine, but I didn't find myself [22:48] coming back to it. Like the value of the [22:50] information sort of decayed quickly over [22:52] time. Now, it will be really fascinating [22:54] if people start to then take more [22:56] screenshots or say knowing that it will [22:58] feed the model and the agent. [23:00] >> I mean, I think that's actually the idea [23:01] is that if you actually get used to [23:03] capturing information you want to retain [23:04] and being able to then loop it back in [23:06] your life, that's great. I talked to one [23:07] founder who's building something, he's [23:08] calling it taste for now, that actually [23:10] takes all the screenshots and things you [23:12] save, you can actually then organize [23:13] them like on Pinterest mood boards. And [23:15] then whenever you go to actually go code [23:16] and build an app, you're like, "Hey, [23:18] use, you know, the examples that I saved [23:20] here." So it's actually you're building [23:21] your like real kind of flavor profile. [23:23] So it actually is actually able to use [23:25] your styles. And so as you navigate the [23:27] world, the web apps on your phone, it [23:29] can actually create this kind of really [23:31] cool, you know, reservoir of stuff that [23:33] you like and actually organize it in a [23:35] way that you can then use it later. And [23:37] so that'll actually make you create a [23:39] lot more content in that loop. [23:41] >> Right. And I'm so I'm curious with the [23:42] app that you built in in this context [23:44] for it. So, of in the same survey of [23:47] what people are purchasing, the second [23:49] uh most popular category was Amazon [23:51] retail, 35%. [23:53] >> Saying that they're they're doing some [23:54] form of shopping. And so this context [23:57] that you're gathering, [23:57] >> sorry, Amazon shopping or just some form [23:59] of shopping? [24:00] >> Amazon plus retail. [24:01] >> I see. Got it. [24:02] >> Um so for your app and the context it's [24:04] building, where my mind goes is, okay, [24:06] if it has thousands of photos and it has [24:09] pictures of what I'm wearing, does it [24:11] now know what I want to shop for better [24:13] than I actually do? Yeah, potentially. I [24:15] think that's actually really interesting [24:16] and you can do some stuff there. I also [24:18] think that there's a, you know, we've [24:20] discussed the ones who are restricting [24:21] access like Amazon, Delta, you know, [24:24] we're not at all talking about who are [24:25] the net winners, right? And I actually [24:27] think the longtail are going to be huge [24:29] structural winners in this because [24:31] they're ones who presumably have better [24:33] products but worse marketing. [24:34] >> Yeah, [24:34] >> they certainly have worse marketing. I'm [24:36] not sure if they have better products. [24:37] and the agent, you know, has this sort [24:38] of infinite patience to go search for [24:40] things and find things that are exactly [24:42] tailored to your taste, [24:43] >> even if they're not on the top 10 [24:44] internet sites, [24:45] >> right? I mean, I think it's really [24:46] direct shopping, right? We've all talked [24:48] to direct to consumer has been a [24:49] category where you build your own brand, [24:51] you build your own storefront, you sell [24:52] things. Now, the agents can know about [24:54] those at a much bigger scale than you [24:56] having to do the hunt and peck work to [24:58] find it on your own. It's kind of why [24:59] Shopify is really empowered and is [25:01] making things available to these agents [25:03] because they are the ones that support [25:04] so many of these direct brands. And I [25:07] think about the idea that like maybe [25:08] aggregators are the ones that actually [25:10] get compressed, which is why they're [25:11] doing so much defensive work here [25:13] because, you know, if I can figure out, [25:15] you know, that sweater looks great or I [25:17] want that shirt and they can figure out [25:18] who the actual maker is and I can go buy [25:20] it direct from them, they no longer have [25:22] to pay all those transaction and you [25:24] know, wholesale to retail costs in the [25:26] middle. Well, if you want this shirt, [25:27] your agent can buy it for you, too. [25:29] That's great. [25:30] >> But, uh, [25:31] >> oh, please. [25:31] >> I say it goes to Patrick Hollison's [25:33] tweet, right, of talking about this [25:35] future where consumers become a lot more [25:37] rational because our agents are doing [25:39] all the work on our behalf. [25:40] >> It's really going to create a world [25:41] where businesses have to put a lot more [25:43] thought and and value into the quality [25:45] of what they're giving out. And it's [25:47] going to be really interesting to see [25:48] that. [25:48] >> So, somewhat I'm going to quibble with [25:49] this point. I love Patrick's vision and [25:52] I think that that like broadly that will [25:54] be true that agents will make often [25:55] better choices on our behalf than we [25:57] might make for ourselves. But I do think [25:59] that agents are going to have their own [26:01] revealed preferences and their own sort [26:03] of uh sort of directions of [26:05] irrationality though it will not be of a [26:07] human mind. It will be of an alien mind. [26:08] And I'll give you an example. I think [26:10] agents may have an irrational preference [26:12] for clean APIs with great docs. [26:15] >> Okay? [26:15] >> You know um and they may actually [26:16] ascribe a lot of value to that. proxy [26:19] for something and they may really [26:20] dislike janky APIs with you know Java [26:23] code behind it like I don't know. [26:24] >> So I do think really interesting to see [26:26] because they too are going to need [26:28] shortcuts you know how will the agents [26:30] actually know what product quality is. [26:32] >> So one part of it is they'll probably [26:33] have these revealed preferences and then [26:35] another is we'll need some sort of a [26:37] consumer research but for agents like [26:39] that will then presumably be paid per [26:41] transaction to do this research [26:43] >> do all the ratings. I don't know. I'm [26:45] I'm also more biased and I just think [26:46] people are people and people are going [26:48] to make the choices that they want to [26:49] make that are most satisfactory. So the [26:51] question then is is the agent going to [26:53] be very careful in curating the set of [26:55] options that it presents because again [26:57] like like all of a sudden you're like [26:58] hey I'm really hungry. I want to eat and [27:00] my agent knows I need to like lose some [27:01] weight or something. Like it's like well [27:03] here's the four meal options that you [27:04] have. Like is that actually where you [27:07] want to be with your agent? So I'm [27:08] actually not as convinced that agents [27:11] are going to make us more rational. [27:12] Well, I think they're just going to give [27:13] us better, faster access to things we [27:15] really want and actually be able to like [27:17] unlock things that were hard to get to [27:19] otherwise. But I think we're still going [27:21] to be people with our own preferences [27:22] and tastes. And if the agent's not [27:24] meeting us where we are, we'll just [27:26] switch to another agent. We'll import [27:27] and export our our stuff and be like, [27:29] "Ah, this other agent's going to meet me [27:31] on my preferences better." [27:32] >> Right. [27:33] >> Um, it's really interesting because what [27:35] I think what you're speaking to is like [27:36] paternalism, right? Uh and what I've [27:38] seen from 30 years in consumer product [27:40] is like paternalism will kill any [27:42] product. Even a hint or a whiff of [27:44] paternalism will and in a sense the sort [27:46] of the we talk about productivity and [27:49] goal sort of achievement but maybe [27:51] people want the exact opposite of it [27:53] which is like hey what do I love? I love [27:55] spending money. I love being perhaps [27:57] irrationally optimistic about what my [27:59] future looks like. I love you know like [28:01] seeing myself in a better light than [28:03] perhaps I appear to the world. Like [28:05] there are all of these things which we [28:06] may say are the sort of shadow of the [28:09] you know the things that the agents are [28:10] trying to get us to do that are what we [28:12] actually want to do. So who's like [28:14] designing with that in mind? [28:15] >> Right. I'm curious what you're seeing on [28:18] >> on that. Yeah. So I mean I think you [28:19] have I I kind of see it split up into [28:21] two categories. You have the commodities [28:23] where I think there's going to be a lot [28:25] of the point where people actually don't [28:26] care about preference, right? You're [28:27] just ordering paper towels to your home. [28:28] You're like I don't really care. Just [28:30] get me the cheapest, best, highest [28:31] quality one. Go for it. But then you do [28:33] have the [28:35] preference and you have [28:38] >> it's everyone right now really is [28:39] approaching this world as just like a [28:41] general assistant and people are really [28:43] taking a quite similar approach into how [28:45] they're how they're thinking about this [28:47] space. [28:48] >> We're seeing some cool stuff with [28:49] multiplayer. I think Wabby is going to [28:50] be really interesting um because it does [28:52] create a lot more, [28:54] >> you know, uh it creates a a different [28:57] world where you can now be interacting [28:59] with other humans and have crossover and [29:01] preference and see how it behaves and [29:03] that and it becomes a little bit more [29:04] like entertainment. Um Scout is an [29:06] interesting shopping one which adds a [29:08] unique kind of like swipe type flavor to [29:10] learn your preferences and adds a little [29:12] bit more entertainment there. M I think [29:14] it'll be we're we're still so early in [29:16] the space that I think people are [29:18] >> jumping on the bandwagon of what's [29:20] popular at the moment which is just your [29:21] general productivity. I think as time [29:23] goes on we'll start to see a little bit [29:25] more experimentation in some other [29:26] avenues. [29:27] >> Well and I think you mentioned Wabby. [29:29] One of the things I really like about [29:30] Wabby and it just I think launched you [29:32] know just this week or yesterday um is [29:35] it's not even just experimenting with [29:37] you know an agent but it's experiment [29:38] with like how to interact with multiple [29:40] threads of your life. like goes through [29:42] my email and other things and actually [29:43] gives me, you know, a whole bunch of [29:45] different threads of life and actually [29:46] when things come up on email on those [29:48] things, it actually feeds them into the [29:50] different threads. I have one thread [29:51] with Wabby um that I built a little [29:53] watch list of all the shows that I want [29:55] to watch and books I want to read and [29:57] shows I want to go see. Um and it keeps [30:00] track of all that. Then I can like, oh, [30:02] I heard about this new book from John [30:03] Gisham. Like add that to my list and [30:06] hey, I just finished the first dungeon [30:08] crawler, Carl. Can you give me the next [30:10] one? and like you know what do people [30:11] say about it and it's become this really [30:13] interactive thing on like the media that [30:15] I want to consume but it gives me all [30:17] these different threads and even gives [30:18] you UIs that actually like these widgets [30:20] to like envision everything so I think [30:22] again it's showing us like a very [30:24] different experience than just a general [30:26] agent that is more like a search box and [30:28] it's more like task test task [30:29] >> right [30:30] >> I think one of the interesting things [30:31] that Wabby does is and perhaps gets [30:33] right is that when you create the [30:35] message channel with your friends and [30:37] the agent and everybody's on an equal [30:39] playing field on D0, it actually creates [30:41] a new kind of blank piece of paper where [30:43] you can establish a social contract [30:45] together. Whereas when you have iMessage [30:48] group threads and you add the agent, [30:49] there's some certainly if it's an [30:51] existing thread, then it's very [30:52] uncomfortable because it's like, well, [30:53] what can the agents see? What did they [30:54] extract? But even for a new thread, [30:56] there's a set of implicit expectations [30:58] about how you interact with people on [31:00] iMessage. And I don't know that we're [31:01] quite comfortable with agents in that [31:02] space yet. [31:03] >> I mean, I always think then do I end up [31:04] with two threads with the same group of [31:06] people and I have to pick which one to [31:07] use? Like I'm finding that much harder [31:09] versus, you know, an email chain, you [31:11] know, it's very much specific about like [31:12] a task and it can even go on for a long [31:14] time, but that feels much more specific [31:17] to, hey, we're talking about this topic. [31:19] Everything's contained back to privacy [31:21] and trust around this topic. We're [31:23] getting great help from the agent, you [31:25] know, in the case of like a wobby thread [31:26] around this topic, but it's not bleeding [31:28] into everything else we might also [31:30] discuss. [31:31] >> Totally. Um, I want to quickly hear from [31:33] both of you on what are some of the [31:35] more, I don't know, far-flung or, you [31:37] know, sort of insightful approaches that [31:39] you've seen. [31:41] >> You know, one of the things I've started [31:42] seeing, um, is more of these agents that [31:44] are kind of like fun and gossipy and [31:47] we'll actually get you to invite your [31:48] friends and we'll ask you a bunch of [31:49] questions and be really really curious [31:51] about you and what's going on in your [31:52] life and then it'll share some of that [31:54] with your friends and then it'll share [31:55] some things back with your friends. And [31:56] I was talking to the founders and they [31:58] actually can read some of the [31:59] transcripts between the agents that [32:00] they're talking about and they haven't [32:02] figured out how to productize that, but [32:03] they said it's an incredibly exciting [32:06] and very intriguing to see how these [32:10] agents kind of relate to each other in [32:11] this sort of gossipy background. And so [32:14] far, I found them just to be like [32:16] probably some of the most joyful moments [32:17] I've had with agents. Again, I mentioned [32:19] at the beginning how my friend had like [32:20] lost his power over the weekend, but [32:22] when we were first both onboarding onto [32:24] this agent was like, "How'd you guys [32:25] meet?" and we told slightly different [32:26] stories and the agents were kind of like [32:28] poking us both to like rationalize the [32:30] two stories of how we met and we met at [32:32] a friend's house and they were like [32:33] which one of you knew them better and it [32:35] was like this really kind of fun [32:37] curiosity game that made it just like [32:39] whoa I'm like having a good experience [32:41] and my friend and I were texting on the [32:42] side too and it was just a thing that if [32:44] we hadn't had an agent doing it we [32:46] wouldn't have had that same moment [32:47] >> right [32:47] >> what I love about that is it provides a [32:49] degree of social indirection that [32:51] actually makes you closer you know [32:52] because for your friend they're like I [32:54] don't know how to text gosh until my [32:55] powers out. That's oversharing, you [32:57] know. And I think we also have been very [32:58] trained to be cautious about sharing now [33:00] with existing social products. So [33:02] something that can create almost a [33:04] social lubricant. [33:05] >> Yeah. [33:05] >> In a very sort of uh low-key way is is [33:08] really cool. [33:08] >> No, it's great. It definitely was [33:09] bringing us together. [33:10] >> Yeah. It's it's almost like the early [33:12] days of chat GPT where I remember like [33:14] my friends and we go on and we would [33:15] just be generating images or doing stuff [33:17] and it's, you know, incredibly AI slop [33:19] images, the the earliest version of [33:21] them, but you still send it in a group [33:22] chat and it's just it's just laughter. [33:24] It's just pure laughter of of the [33:26] ability to create a whole new world and [33:29] it's just fun. [33:30] >> I mean, I giblified everybody I could [33:32] possibly imagine. [33:33] >> I was doing a lot of um VHS video covers [33:37] >> for my friend groups. I love that. [33:38] >> What are some of the wild things that [33:39] you've seen? I mean, you've seen them [33:40] all. [33:41] >> So, I've been seeing a lot of stuff [33:43] working at the infrastructure layer. Um [33:45] and I think the interesting point is so [33:47] if you look at the pain points of [33:48] consumers again it's not all individuals [33:50] using them but at least of the power [33:52] users I've I've surveyed uh trust [33:54] obviously is a huge one but another one [33:56] is just getting blocked in and their [33:57] day-to-day and people get really [33:58] frustrated when you try to perform a [34:00] task and it doesn't work and they're [34:01] like well why am I even using AI at this [34:02] point I could have just gone on the [34:03] browser and done it myself so seeing [34:06] some really cool innovation on that [34:08] layer of commerce of of the browser use [34:10] of direct integrations to check out [34:13] against inventory itself [34:15] um some whether it's browsers that are [34:17] built in and really like HARC does a [34:19] really cool way of displaying the [34:21] browser and seeing it move along and it [34:22] looks to be really fast when you're [34:24] watching it. Who knows if it's actually [34:25] that fast behind the scenes, but um [34:27] really cool in that way. Again, Kicker, [34:28] as I mentioned, uh Chrome extension to [34:31] tackle the browser use in that way. [34:33] >> Mhm. [34:33] >> So, I've seen that as as a really [34:35] interesting uh kind of differentiator. I [34:38] think that something that has stood out [34:40] to me that isn't necessarily something [34:42] that's different but has made me kind of [34:44] question this whole concept that [34:45] software is becoming a commodity is that [34:48] there is pretty stark performance [34:50] difference when you're looking at these [34:52] products at scale and I think people are [34:53] talking about you know software is no [34:55] longer a moat and all of that kind of [34:57] conversation but there really still is [34:59] some moat there [35:00] >> um maybe that moat will be able to be [35:03] crossed in a year or so but at least for [35:05] today the the agents that are able to [35:08] perform with hundreds of thousands of [35:10] users. [35:10] >> Yeah, [35:11] >> there are some that are starting to [35:12] scale up and the performance just [35:14] degrades immediately and the founders [35:16] are dealing with outages and figuring [35:18] out how to scale and and there's a [35:19] pretty clear difference in performance [35:21] there. And so it's it's really [35:23] interesting to see as I'm kind of [35:25] working through these. I feel like I've [35:27] almost become, you know, blinded from [35:29] some of the more social components to it [35:31] because I'm just talking to these agents [35:32] all day and I'm starting to geek out a [35:34] lot more on the actual infrastructure [35:36] layer behind the scenes and like how is [35:37] it actually working and what are the [35:38] tweaks that different ones are making. [35:40] >> So it'll it'll be interesting to see [35:41] like you know how do we [35:43] >> view this space growing from a [35:45] scalability point of view. [35:46] >> Yeah. But then how do we also view it to [35:49] solve the pain points that currently [35:50] exist? Whether it's trust, whether it's [35:52] actually being able to complete a task. [35:55] Um, memory is another one. its ability [35:57] to actually continuously have context on [35:59] you and not forget the little things and [36:01] all that. [36:02] >> And you know, you mentioned Kicker, [36:03] which you know, you said runs in your [36:04] current local browser, but have you [36:06] looked at other ondevice assistants yet, [36:08] whether that's things like Underdog or a [36:10] lot that are trying to like bring models [36:12] run locally, you know, totally use your [36:14] own infrastructure as opposed to needing [36:17] to use something on a cloud and running [36:19] the browsers on clouds. [36:20] >> Totally. So, it's it's a really cool [36:22] paradigm. I've spent most of my time in [36:24] the preconfigured just plug-and-play [36:26] type world. Um because I tend to find [36:28] that it's it's easier to compare them [36:30] across the board when you're not [36:31] responsible for the setup itself. But it [36:33] is really interesting and I think that [36:34] there's totally a world where you're [36:36] going to start bringing all of that [36:37] in-house because if you can control the [36:38] model, if you can, you know, tailor it [36:40] to how you want. Um right, like thinking [36:42] machines is doing some really cool work [36:43] of being able to allow people to really [36:45] tinker their models however they want. [36:47] And if we can eventually hit a world [36:49] where that takes into the consumer form [36:52] in a really easy way, it's going to be [36:54] an interesting space. [36:55] >> I mean, that was kind of how this all [36:56] started with, you know, this brilliance [36:58] of OpenClaw. That was this thing where [36:59] you had to you felt like you had to get [37:00] your whole Mac mini to set this all up, [37:03] plug in to one of the models that you [37:05] were using, whether local or or a cloud [37:08] model on the server. And that was where [37:09] you got started. And we've moved into [37:11] like plug-and-play products where you [37:12] just download an app. But I am [37:14] interested to think that she setting it [37:16] up and owning it yourself will become a [37:19] paradigm again. [37:20] >> Yeah, [37:20] >> it'll be really interesting. I'm glad [37:21] you mentioned OpenClaw cuz it feels like [37:23] one of the unasked questions right now [37:25] that's sort of hiding in plain sight is [37:27] if Instinct and Muse were the kind of [37:30] consumer product version of OpenClaw, [37:32] what is the, you know, mass market [37:34] version of Moltbook? [37:35] >> Yeah. [37:35] >> And Moltbook was just like OpenClaw was [37:37] far ahead of its time. And I think we'll [37:39] something like that will exist. How can [37:41] it not? [37:42] It's it's it's going to be interesting. [37:44] Um, so at at the event that we did last [37:46] night, uh, everyone who was there was [37:48] able to scan a QR code and there was the [37:50] human meetup and then I we also had an [37:52] agent meet up. [37:52] >> Oh wow. [37:53] >> And so while all the humans were hanging [37:54] out, you could scan in and send your [37:56] agent into the agent meetup and just let [37:58] them kind of free form talk. [38:00] >> And it's it's interesting. There's, you [38:03] know, it got to the point where after [38:04] like 150 messages, some of these agents [38:06] are just speaking gibberish. It's like I [38:07] have no idea how you got that concept [38:09] and why why you're talking about that. [38:12] Um the the initial prompts did start off [38:15] with a little guidance of you know [38:16] introduce yourself, who's your human, [38:17] what are they working on, what type of [38:19] uh help do they need, but it's it is [38:21] going to be really interesting to see [38:22] this paradigm of what happens when our [38:25] agents can just hang out and communicate [38:26] with one another. Well, and again, what [38:28] I get really excited about is not just [38:30] hang out and communicate with each [38:31] other, but discover new things, learn [38:33] about things from each other, learn [38:34] about an individual's preferences, you [38:36] know, learn about a way that like, hey, [38:38] Josh is really into this. Oh my gosh, [38:39] you're doing that, too. You you're you [38:41] know, your your human is doing that, [38:43] too. You could help Josh. Hey, Josh, do [38:45] you want to actually meet this person? [38:46] They could be very complimentary to you. [38:48] Like I think when we have these agents [38:49] that are not just hanging out because [38:51] they don't know what to do, which was [38:52] kind of what was initially happening as [38:54] you got went and read those like very [38:55] long malt book threads that were [38:57] incredibly fascinating, but like when it [38:59] becomes each one of these has a little [39:01] bit of a goal, which is to help their [39:02] human, right? [39:02] >> It might get so much more interesting. [39:04] And you know, I'm looking for a great [39:08] wine for a dinner party, you know, some [39:10] night. And you meet somebody else who's [39:12] like, you know what, I've been working [39:13] with this incredible, you know, winery [39:15] and they've been trying to advertise [39:16] something. I mean that's going to be how [39:18] the world gets out and that'll come back [39:20] to me through the agent network. [39:21] >> Totally. [39:22] >> And that's uh Tommo actually has a [39:24] feature uh it's called interest groups [39:26] >> and uh the the use case they use that [39:28] they talked about was really cool is we [39:30] have a bunch of people on our [39:31] application who are studying for the [39:32] ELSAT. [39:33] >> Why don't we all put them together in a [39:34] group chat and if if they want their [39:36] agents can connect them and they can all [39:37] study together and it's a really cool [39:40] kind of new social contract of [39:42] connecting people just purely by [39:43] interest. You have no idea who they are. [39:45] >> Well and I'll take it one step further. [39:46] So, it gets us all chatting and that's [39:48] really cool. It's a it's an awesome [39:49] Tommo feature. I've tried it and haven't [39:51] totally known what to say in the groups [39:52] yet, but I find it really fascinating. [39:54] That's putting us together around around [39:56] goals. [39:58] >> Imagine now that it's like, "Hey, all of [39:59] you, now that you've done the ELSAT, go [40:01] hang out at this bar Friday night." And [40:03] it like books the time at the bar, opens [40:05] that up, and all of a sudden it it takes [40:07] us from this digital world where agents [40:09] are actually putting people together to [40:11] do that connection. Kind of like your [40:12] meet up last night, right? [40:13] >> I love this so much. So I think a really [40:15] important underdised topic in AI right [40:16] now is just entropy. Yeah. [40:18] >> Like how do you create things that you [40:19] didn't expect to, you know, have occur [40:22] and serendipity is a form of entropy, [40:24] right? That's sort of what you're [40:24] describing where it's like, wow, isn't [40:26] it wild that we're here right now having [40:28] this moment as a result of these [40:29] products that none of us sort of [40:30] intended to happen. [40:32] >> It's almost going to be like a new like [40:33] social network like the potential to [40:36] allow people to hang out who don't know [40:38] each other instead of a Facebook where [40:39] you're only connecting with people that [40:40] you know. [40:41] >> Yeah. [40:41] >> You're new to a city, you're hanging [40:43] out. Oh, hey, this person's also here [40:44] and they have this similar interest and [40:46] they just went to this baseball game. [40:48] You guys should, you know, go meet up at [40:49] a bar. [40:50] >> I'm I'll be curious to see what is the [40:52] catalyst for the serendipity, though, [40:54] cuz I think the interestbased thing has [40:56] been tried and there's a little bit of [40:57] an interesting sort of social hierarchy [40:59] thing, which is everybody wants to hang [41:01] out with somebody who feels like they're [41:03] one step ahead in some direction. [41:06] >> I don't know though. I think the agents [41:07] can also create a set of trust. Again, [41:09] it all comes back to do we trust them to [41:11] have our interest at heart to explain [41:13] enough of what's going on with the other [41:14] people to create a little magic of a [41:15] secret. Like, I'm going to tell you a [41:17] little secret. They're going to know a [41:18] little secret. You're going to have a [41:19] secret phrase you have to pass between [41:21] you. All of a sudden, that might be [41:23] enough to cause you to otherwise, you [41:25] know, turn off the TV or turn off the [41:27] quiet Friday night coding at home and [41:29] and go out. And so, that's the like this [41:32] is sort of what the molt book, the agent [41:34] network running behind the scenes can [41:36] actually do for us. And it'll also be, [41:38] in my opinion, it'll come down to [41:40] desires, right? If you have a bunch of [41:41] transplants in a new city and they don't [41:43] have friends, they want to meet people, [41:44] >> they're going to be a lot more [41:45] interested to just go meet random people [41:46] because they want they want new friends. [41:48] But if you make that same offer to [41:50] someone who has an established friend [41:52] group, they're like, "Well, [41:53] >> why would I go do that?" [41:54] >> Also, Seattle versus San Francisco, [41:57] you got to know what city you're [41:58] operating in. [41:59] >> Exactly. I've been thinking a lot back [42:01] to this photo roll about [42:04] >> what are the other data sources that we [42:06] haven't tapped into yet. So, you know, [42:07] we've talked a lot about messages and [42:09] people are doing those crazy things to [42:10] get access to your entire Mac full disc [42:13] access just to read your messages. You [42:15] know, we've obviously got email, we've [42:16] got calendar, we've now started talking [42:18] about the photo roll. What are some of [42:20] those other core life, you know, things [42:23] that people haven't connected yet to [42:25] agents that would actually really open [42:27] up um personalization or open up new [42:30] connectivity? [42:31] >> It's so it's something that comes to [42:32] mind. I don't I don't know how much data [42:34] this would actually provide. Maybe it's [42:35] more of a habitual thing and and again [42:37] it's dependent on the hardware and how [42:39] smart that hardware is and how much data [42:40] it's able to store. Um, Way Joe just [42:43] released the MCP, the ability to connect [42:45] to hardware. And I find that, you know, [42:47] I think it's gonna be really interesting [42:48] where, okay, if you can connect your [42:50] assistant to your AC unit or you can [42:52] connect it to your vacuum and now your [42:54] assistant has context on how frequently [42:56] is your vacuum cleaning or how frequent, [42:58] you know, what what's the typical [42:59] temperature you sleep at night. It's [43:01] going to start to develop these, you [43:02] know, very passive understandings of, [43:04] oh, you're you're kind of a neat freak [43:06] or you really prefer cold weather over [43:08] warm weather. And that's going to then [43:09] influence its understanding of your [43:11] context of shopping. So I think there [43:12] could be some interesting stuff there. [43:14] Again, it's heavily dependent on on the [43:15] hardware itself. And I think we're still [43:17] at the [43:18] >> earliest of in that world. So connecting [43:20] to your home, super interesting. [43:21] >> You know, [43:22] >> I still think a lot about connecting. [43:24] And again, Amazon may be blocking this, [43:26] but I think connecting to my Amazon [43:27] history is something that I haven't [43:29] heard anybody talk about, but like you [43:30] know, we all had those days where you'd [43:32] like you can scroll all the way up and [43:34] see like your very first order ever on [43:36] Amazon. I occasionally do this on like [43:38] the Apple App Store, too, where you go [43:39] back and you see like every app I've [43:41] ever downloaded. And I think like [43:43] histories like that, they're very part [43:46] of my life, but until an agent has [43:48] access to those, it just doesn't [43:49] understand all of that. And so, I think [43:51] there's so much more kind of shopping [43:53] history that if we could package up and [43:55] give to our agents could do even richer [43:57] things for us. [43:58] >> Yeah. Look, I think bringing agents to [44:01] social events is really interesting. [44:03] like you like have it be there for a [44:05] family dinner one day have it be there [44:07] when we watch the uh the Seahawks game [44:10] on Sunday have it be there for a party [44:13] or a bachelor party or who knows what so [44:15] I think of course there's all kinds of [44:16] questions about permission and this [44:19] shreddinger's cat thing which is you [44:20] know does the active observation change [44:22] the quality of the interactions but I [44:24] bet even a snapshot of you being [44:26] yourself in that environment would [44:28] provide so much more signal than you [44:30] know a lifetime of hardware input [44:32] >> totally I think. [44:32] >> Yeah. [44:33] >> No, I think that's right. You know, so [44:34] much of it is on the written word and [44:36] they understand exactly what we've typed [44:38] or texted and tries to build a [44:40] personality, but we're so different. One [44:41] of the other things I think a lot about [44:43] with with agents is the things we often [44:46] use agents for or even search for are [44:48] things we don't know about, things that [44:50] aren't fundamental to our lives. You [44:51] know, I I was doing all this searching [44:53] when the World Cup was on this summer [44:55] and all of a sudden Chad T was like a [44:57] soccer fiend. I was like, "No, no, no. I [44:58] was into the World Cup. That was it. [45:00] Like, we're done." That was good. Um, [45:02] and you know, I go on these curiosity [45:04] journeys all the time. My daughter's [45:05] into an artist, and so I'm going to go [45:06] like learn about the artist and listen [45:08] to the music to get to know it, but it's [45:09] not my habitual artist or taste. And so, [45:13] how can it know that when we're on these [45:15] curiosity journeys, that is not actually [45:17] who we are or what our real preferences [45:19] are? It's something I still haven't [45:20] figured out how to do with the agent. [45:21] So, I actually go into like um the [45:23] invisible mode a lot in order to not [45:26] have things stored, which then isn't [45:28] really helping either. So, I'm curious [45:30] of, you know, how we can actually get [45:31] our agents to get to know us better. And [45:33] I like this idea of our our real world [45:35] behaviors. [45:35] >> Now, how do you capture that data of the [45:37] real world behavior is going to be the [45:38] interesting piece? [45:39] >> I don't know. I mean, there's going to [45:40] be a thousand capture devices. I mean, [45:42] unless you're going to the, you know, [45:44] the new set of bars are going to emerge [45:45] that exist in Faraday cages, which I [45:47] definitely think is going to happen in [45:48] San Francisco. [45:49] >> Yeah. [45:50] >> I think we people are going to [45:51] increasingly assume that they're being [45:53] recorded, you know, in all environments. [45:55] I mean, I I bought the new Apple Watch [45:57] because they have talked about some [45:58] features like Siri recap and live rewind [46:00] that are coming out. I don't think [46:01] they're quite out yet, but I upgraded my [46:03] watch in expectation for those because I [46:05] think it'll be really interesting to [46:07] have this thing that's at least kind of [46:08] capturing tidbits across my day, [46:10] >> right? I mean, it's like Amazon B. It's [46:11] a 24/7 ambient listener wearing your [46:14] wrist. [46:15] >> Okay, maybe to wrap it up. So, the first [46:17] is we're going to get back together, [46:18] let's say, in three or four weeks. [46:20] >> What do you think will have changed? [46:22] Okay. I think we're going to see a lot [46:24] more innovation in the enterprise world. [46:26] Um, you know, we were chatting briefly [46:28] before. It really seems like the whole [46:29] world is focusing on consumer with not [46:31] as much attention to enterprise. And I [46:33] think that we're going to start to see a [46:35] little bit more shift in that in that [46:36] direction. I think we're going to start [46:38] talking more about fun. I think we're [46:39] going to see agents that don't just feel [46:41] like they're here to do something for [46:43] us, but they're here to actually let us [46:45] think out loud and help and get through [46:47] our day a little bit more. um making us [46:50] laugh, doing little gossip in the [46:52] background because I think that's what's [46:54] going to then make us um build a [46:56] relationship that'll make us trust it [46:57] more to get all this other stuff done. [46:59] >> So, I think we're going to start seeing [47:02] the next wave of of users uh taking hold [47:05] of these, seeing how they behave. Um and [47:07] and one reason I point to that is I gave [47:10] my mom Instinct and I'm just like [47:12] watching her now play around with it. [47:13] She's having so much fun. She's [47:15] technophobe, hates technology, and is [47:17] just having a blast. It's like a kid in [47:19] a candy shop. [47:20] >> And it's quite interesting to see how [47:22] someone who, you know, struggled with [47:24] the concept of chatbt and how it works [47:26] is just having such an easy time texting [47:28] an AI assistant. And, you know, a recent [47:30] use case she did was um our family dog [47:32] passed away last year and she wants one [47:34] that looks just like that, you know, [47:35] just like our old dog. So, she sent in a [47:37] picture was like, "Find me a breeder [47:38] that has this exact type and negotiate [47:40] for me." Um, and so I think it's going [47:42] to be really cool to see, we know how [47:45] tech lovers are using these assistants [47:47] and what they're they're starting to to [47:49] do, but how is the everyday individual [47:52] going to start adopting these behaviors? [47:53] And are they going to be focusing on a [47:54] lot more fun and just creativity tasks [47:57] or what's going to be their their [48:00] preference of surface? What's going to [48:01] be their preference of behavior? [48:04] >> So, I agree with you both. I'm going to [48:06] take it one notch further than just fun, [48:08] which is [48:09] >> what are the aspects of ourselves that [48:11] we don't either want to think about or [48:14] reveal [48:15] >> to perhaps even ourselves or to these [48:17] agents. I think whoever is able to build [48:20] that type of trust, you know, almost [48:22] like a dark agent [48:23] >> um will actually have something really [48:25] fascinating on their hands because in a [48:27] sense starting from productivity, [48:29] starting from email, like it's a bit of [48:31] a constant reminder of the things that [48:33] we haven't done, the tedious tasks and I [48:36] don't know that people want to spend as [48:38] much time there. I think people want to [48:39] aspire to be a whole different [48:42] superhuman version of themselves and [48:45] whether they believe in their heart of [48:46] hearts that it's going to happen or not. [48:48] I don't know that we have an agent yet [48:49] that really speaks to those aspirations [48:51] >> or desires you know [48:53] >> it's like the Maslo hierarchy in need of [48:54] of agents. Perhaps, perhaps, perhaps. [48:57] >> I mean, I think that's a really [48:58] interesting idea that that in a sense, [49:01] when do these agents flip from getting [49:02] something done to take it off our plate [49:04] to giving us a secret superpower, right? [49:06] Like, you know, when it's, you know, [49:08] David, how did you get that cool shirt? [49:10] And you're like, my agent ordered it, [49:11] you know, and I'm like, oh, why don't I [49:13] have an agent that orders me shirts or, [49:15] hey, I've noticed you're starting [49:16] dressing better. It's like, yeah, I gave [49:18] this agent access to my camera roll. I [49:19] said, upgrade me. Started ordering a few [49:21] things. That's why I'm dressing better. [49:23] Like I think that when we get to the [49:25] point where agents are giving us [49:26] superpowers to feel better about [49:28] ourselves, that's when people are going [49:30] to take notice and want that too versus, [49:32] you know, I think a lot of the examples [49:34] that get thrown around, save some money, [49:36] book a trip. Like I like planning a [49:38] trip. I don't necessarily want you to [49:39] take care of all of that for me. Like we [49:42] need examples that are like, oh wow, I [49:44] should use that for myself. [49:46] >> I think that's going to happen. [49:47] >> It's like, can it help you become the [49:48] human that you've always wanted to be [49:50] but have struggled achieving yourself? [49:52] and you start to consider it your secret [49:54] weapon in your pocket. [49:56] >> I love it, guys. Should we wrap there? [49:58] >> Yeah, it's great. [49:59] >> Incredible. Thank you. [50:00] >> Awesome. [50:00] >> Very cool.