Former Twitter CEO Launches Agentic Search Startup Parallel

Twitter 前 CEO 创立 Agent 搜索初创公司 Parallel Parag Agrawal 认为，除非开发者采用博弈论归因模型向内容创作者付费，否则 AI 网络 Agent 将摧毁由广告支撑的互联网模式。 Parag Agrawal 阐述了 Parallel Web Systems 的核心理念：AI Agent 代表了搜索基础设施的根本性转变，需要从底层重新设计，专门针对模型上下文而非人类点击进行优化。他解释了 Parallel 从深度搜索 Agent 产品演进至 200 毫秒 Turbo 索引的历程，以及其与 Google Cloud Platform 的接地检索（grounding）合作。此外，Agrawal 还探讨了随着 Agent 流量取代承载广告的人类页面浏览量，网络出版商所面临的经济危机，并提出了一种基于博弈论 Shapley 值的动态归因框架，以根据实际效用向内容创作者提供报酬。 Twitter 前 CEO Parag Agrawal 创立 Parallel Web Systems，指出人类点击数据对 AI 而言是“缺陷”（bug），且 AI Agent 的检索量将是人类的 1000 倍，亟需重构专属于 Agent 的搜索基础设施。 (0:00) Parallel 采取“先推搜索 Agent、后增量构建索引”的演进路径，最新 Turbo 产品将检索延迟压缩至 200 毫秒，相比传统搜索封装可降低一半以上的上下文 Token 消耗。 (7:52) Parallel 已直接集成至 Google Cloud Platform（GCP），作为与原生 Google Search 并列的企业级 AI 模型官方 Grounding 检索提供商。 (25:48) 针对 Agent 绕过广告导致互联网变现崩溃的困境，Parallel 引入博弈论的沙普利值（Shapley values）动态评估网页贡献度，预计在 12 至 24 个月内向内容出版商分配实际资金回报。 (40:48)

Parallel Web Systems founder Parag Agrawal argues that because AI agents will generate search traffic at a scale 1,000 times greater than humans, search architecture and internet monetization must be rebuilt from the ground up for machine consumption rather than human clicks. Engineering Search for Autonomous Machine Consumption (13:14) Traditional search engines rely on human clickstream data, autocomplete, and fast-loading summary pages designed to accommodate human impatience and imprecise keyword queries. Agrawal characterizes human click data as an architectural bug for agentic systems. Because AI agents formulate structured, long-context queries and process dense information without interface friction, they require direct access to authoritative, un-summarized sources rather than SEO-optimized web content. Incremental Indexing and the Latency Trade-Off (7:52) To bypass the prohibitive upfront capital costs of crawling hundreds of billions of web pages before launch, Parallel initially deployed deep research agents that traded immediate response times for crawl depth. Allowing queries to run for 1 to 10 minutes let the system crawl pages dynamically on demand, targeting complex, JavaScript-heavy pages that foundational model pre-training crawls typically ignore. The Multiplier Effect of Background Agents (27:39) Agrawal observes that while historical human usage averaged 20 to 30 queries daily, autonomous background agents execute hundreds or thousands of automated searches for tasks such as continuous portfolio risk audits or meeting preparation. Citing data from Cloudflare, he notes that AI page reads have already reached parity with human web traffic. Game-Theoretic Compensation for the Post-Ad Web (34:28) Because autonomous agents extract raw text directly into context windows without viewing advertisements or entering subscription funnels, the ad-supported business model of the open web faces structural collapse. Agrawal argues that existing bespoke, fixed-fee licensing agreements between AI labs and major publishers fail to scale as inference volumes multiply fiftyfold over two-year spans, leaving long-tail publishers uncompensated and incentivized to block scrapers. Agrawal envisions agents evolving from discrete tool users into continuous, event-driven networks that monitor web state changes in real time. However, this ecosystem remains contingent on deploying computationally feasible Shapley value models to fairly monetize web sources and prevent the open internet from locking behind anti-agent defenses.

前Twitter首席执行官Parag Agrawal创立的Parallel Web Systems提出，传统以人类点击和广告为核心的网络架构无法适配AI智能体时代。Agrawal认为智能体的网页访问与检索规模将达人类的1000倍，主张从底层重构低延迟、高信噪比的检索基础设施，并借助博弈论模型重塑网络内容的收益分配机制。 智能体驱动的千倍检索规模与信息范式转换 (0:00) 传统搜索引擎高度依赖人类点击流，但在面向AI智能体的检索中，人类点击数据被视为干扰而非有效特征。智能体与人类的搜索行为存在本质差异：人类习惯使用简短、模糊或包含拼写错误的关键词，容易受搜索引擎优化（SEO）和汇总浅层内容（Slop）影响；智能体则生成结构化且上下文丰富的长查询，需要系统直接从权威源头（如数十页的SEC申报文件）中提取高密度核心段落。根据Parallel的测试，其专用检索架构能将智能体的上下文token消耗降至传统搜索结果的一半以下，显著降低运行成本并提升准确率。 与此同时，自主运行的后台智能体正在推动检索调用量发生量级转变。以管理10,000笔小微企业贷款的信用风险监控或自动化日程准备为例，定期自动运行的智能体可将单个用户触发的检索调用从传统人类每天20至30次推高至成百上千次。Agrawal引用Cloudflare数据指出，AI在页面读取维度的流量已基本追平全人类的网页访问总量。 Parallel的多阶段架构与工程路径选择 (19:38) 为绕过构建全网索引的极高前期沉没成本，Parallel早期采取了“先做搜索智能体、后做搜索引擎”的策略，利用推理计算换取抓取深度。在深度调研场景下，用户对时延的容忍度由1秒放宽至1到10分钟，使系统能够在缺少完备万亿级索引的情况下即时动态抓取并积累实证评估数据。在完成基础数据与质量优化后，Parallel推出了响应时间仅200毫秒的Turbo产品，以适配高频实时的智能体操作循环。 在底层架构上，请求依次经过意图处理、多索引查询重写（包括大索引、实时索引、结构化索引及知识图谱）以及多阶段逐层重排，从数百亿URL筛选至数万个候选文档，最终向智能体输出约1000个高信号token。Agrawal强调，基础大模型厂商在预训练抓取时注重原始吞吐量，往往跳过渲染缓慢或重度依赖JavaScript的复杂网页；Parallel则专注于抓取这些被忽略的动态网页与实时知识，形成能力互补。目前，Parallel已作为官方检索与接地（Grounding）提供商接入Google Cloud Platform（GCP）。 传统广告变现崩溃与静态版权授权的局限 (34:28) 现代开放网络的商业闭环建立在人类注意力和广告点击的基础之上，高意向商业搜索产生的利润长期补贴了全网免费信息的生产与分发。然而，智能体直接抽取内容切片，既不浏览广告，也不进入订阅转化漏斗，导致内容发布商无法获得访问收益，进而促使网站广泛部署针对AI抓取工具的防御封锁。Agrawal警告，若无法确立新的经济回报模式，开放网络将面临大面积退入付费墙和全面封闭的风险。 与此同时，主流AI实验室与少数大型媒体之间签署的固定金额授权协议（涵盖训练数据、推理调用与免责保护）被证明不具备可持续性。在AI推理需求年增7倍、两年增长50倍的背景下，固定费用导致发布商的单位价值分配迅速贬值，且该模式完全无法惠及长尾创作者，脱离了具体任务中数据所创造的真实商业价值。 基于Shapley值的内容价值分配与网络演进 (40:48) 为量化单次智能体多重检索中各个网页的实际贡献，Parallel（原注册名Shapley Inc.）引入合作博弈论中的Shapley值概念：若剔除某个信息源导致答案质量下降，则该信息源的经济价值等价于开发者为弥补这一质量降级所需额外支付的算力或模型成本。鉴于全排列计算Shapley值的组合模拟算力成本过高，Parallel训练了专用评估模型进行快速估算，并结合任务商业价值实现差异化定价（例如金融分析场景的数据调取费用高于日常闲聊）。Agrawal估算，若将LLM推理支出的2%至10%分配给数据提供方，其规模将超越除封闭平台外的全网内容商业体量，并预测该机制将在12至24个月内为发布商带来切实回报。 面向未来，网络呈现“面向人类与面向机器”的双重发布形态。Agrawal将智能体演进分为三个阶段：第一阶段是单体智能体将搜索作为基础工具；第二阶段是多智能体协同运行复杂任务；第三阶段则是“事件驱动的推送网络”（Event-Driven Push Architecture），网络从被动查询转为常态化状态监听，一旦网页或外部现实世界数据发生变更，系统便会自动触发智能体进行决策与计算。 Agrawal所构想的新一代网络体系，建立在对传统爬虫架构重构与基于Shapley值精细化分成的双重支柱之上。然而，该模式的最终落地仍面临未经验证的挑战：其预测模型能否在全网范围内建立具有公信力的数据定价标准，AI应用层是否愿意为数据分成承担额外的推理成本，以及在开放网络全面筑起反抓取屏障之前，这一利益分配生态能否足够快地建立。

## Background 该内容由红杉资本（Sequoia Capital）于 2026 年 8 月 28 日在 YouTube 平台的《Training Data》播客栏目发布，由 Sonya Huang 与 Andrew Reed 共同主持。视频简介指出，前 Twitter CEO、现 Parallel Web Systems 创始人兼 CEO Parag Agrawal 正在押注一个颠覆传统搜索的前提：未来 AI Agent 查询网络的频率将比人类高出一千倍，而 Parallel 正通过将人类点击视作干扰、基于 Agent 反馈训练、开发 Turbo 产品实现 200 毫秒极速检索，并引入沙普利值（Shapley values）为内容所有者建立全新的利益分配机制。 ## TL;DR Parag Agrawal 深入剖析了创立 Parallel Web Systems 的底层技术思考与商业愿景，明确指出传统搜索引擎依赖人类点击和即时加载的逻辑根本不适配 AI Agent 的高频推理需求。Parallel 采取先做 Search Agent 再滚雪球式扩建索引的反直觉打法，将全网海量信息高效压缩为精准供给大模型上下文的 1000 个 Token，并通过新发布的 Turbo 架构将搜索延迟压降至 200 毫秒。面对广告模式随人类注意力转移而瓦解的困局，Parag 提出运用博弈论中的沙普利值量化网页贡献，预计在 12 至 24 个月内将推理算力预算回馈给内容创作者，推动网络从人类主导的“拉取式搜索”彻底进化为驱动多 Agent 自主运转的“推送式神经中枢”。 ## Atmosphere & style 对谈氛围兼具硬核技术深度与硅谷顶级创投沙龙的松弛感，主持人的敏锐追问与初次亮相的幽默互动，迅速打破了关于前巨头掌门人的严肃刻板印象。Parag 表现出极具张力的理性与笃定，在拆解 SEC 财报案例与自嘲早期公司命名尴尬时信手拈来，毫无浮夸的公关腔调。整场对话既有深入底层的分布式检索与博弈论数学推导，又充满对互联网商业根基重塑的宏大思考，给听众带来一种亲历下一代计算基础设施破土而出的智力激荡与前瞻兴奋感。 ## Bulletin - [00:57](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=57s) 前 Twitter CEO Parag Agrawal 创办 Parallel Web Systems，押注 AI Agent 未来的网络查询频率将比人类历史峰值高出 1000 倍以上，现有基于点击流的搜索与变现架构面临彻底重构。 - [11:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=663s) Parallel 视人类点击数据为优化 Agent 搜索的“Bug”，通过新发布的 Turbo 架构将全网深度检索延迟从 3 秒压缩至 200 毫秒，并将输出精准收敛至最具信噪比的 1000 个 Token。 - [25:48](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1548s) Parallel 宣布与 Google Cloud 达成战略合作，作为 GCP 企业级 Agent API 底层的搜索与 Grounding 引擎，与 Google 自有搜索并列供企业开发者选用。 - [45:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2746s) 针对广告模式在智能体时代失效的问题，Parallel 引入博弈论中的沙普利值（Shapley values）预估模型，计划在 12 至 24 个月内将知识工作推理预算的 2% 至 10% 分成给内容创作者。 ## Key takeaways - [00:57](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=57s) Agent 交互频率将比人类高出 1000 倍以上，要求搜索系统从服务低频懒惰的人类彻底转向承载高并发严谨 Agent 的新型基础设施。 - [03:25](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=205s) 网络搜索的核心本质是一个“十亿对十亿的匹配问题”，需在数百亿网页与跨时间维度的数百亿查询之间建立毫秒级撮合通道。 - [05:17](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=317s) 人类点击数据对 Agent 搜索而言是具有干扰性的噪音（Bug），LLM 强大的信息压缩与专家级评估能力使得脱离点击流构建全新反馈回路成为可能。 - [07:52](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=472s) Parallel 采取倒序战略，先做可容忍数分钟延迟的 Search Agent 并以推理算力换爬取时间，在为真实垂直业务替代外包人工的过程中滚雪球式扩建索引。 - [11:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=663s) Parallel 并非追求基础大模型研发的 Neo-lab，而是将自身定位为模型的“互补品与乘数”，专注于从全网万亿级网页中为模型提取最正确的 1000 个 Token。 - [13:14](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=794s) Agent 搜索通过专有召回与精简段落抽取，能使下游大模型消耗的 Token 数量减少一半以上，并在模型层、智能体层与搜索层之间实现算力的智能再分配。 - [22:22](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1342s) 大模型预训练爬虫追求原始 Token 吞吐而往往跳过加载慢的动态复杂网页，实时搜索索引必须扮演“完备主义者”，补足模型推理时对实时长尾网络的刚需。 - [27:27](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1647s) 企业后台自动化运行的后台 Agent（如每周对万家信贷客户进行全网风险扫描）正在释放数十万到百万倍的搜索查询倍增效应。 - [31:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1906s) 在质量、成本、延迟三维工程权衡中，Parallel 早期主动无视延迟死磕信息质量，待建立代差后再通过 Turbo 架构将检索耗时压缩至 200 毫秒。 - [34:21](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2061s) 基于人类眼球注意力的广告与差异化定价模式在 Agent 时代走向破裂，头部媒体固定总额买断协议也无法跟上 AI 推理调用的指数级膨胀。 - [40:48](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2448s) Parallel 基于沙普利值理论训练轻量估算代理模型，根据网页对最终推理任务的边际经济价值，在无需穷举排列组合的前提下实现公平价值归因。 - [45:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2746s) 知识工作场景下 2% 至 10% 的 LLM 推理预算若回流至开放网络，其资金池将超越传统开放网页广告总和，并预计在 12 至 24 个月内产生实质性美元分成。 - [47:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2866s) 互联网正在进入“双重发布”时代，高价值内容（如财报、API 文档）的第一受众正从人类读者转移为专为 AI Agent 优化的平行网络。 - [51:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=3063s) 网络范式正经历从主动“拉取”（Pull）向基于事件触发的“推送”（Push）演进，网络将从查询工具升级为驱动全球 Agent 实时运转的神经中枢。 ## Detailed walkthrough ### [00:57](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=57s) 从 Twitter CEO 到二次创业：为 AI Agent 重构网络与搜索 红杉资本（Sequoia Capital）的 Sonya Huang 与 Andrew Reed 共同主持了本期《Training Data》播客。节目开场，Sonya 隆重介绍了前 Twitter CEO Parag Agrawal。在将 Twitter 卖给埃隆·马斯克（Elon Musk）之后，Parag 重新踏上了创业旅程，创立了一家名为 Parallel Web Systems 的新公司，致力于为新兴的“Agentic Web”（智能体互联网）构建扩展性极强的 Agent 搜索基础设施。谈及创立 Parallel 的初衷，Parag 指出，以往人类依靠浏览器、搜索引擎以及点击和浏览来获取信息，而未来的 AI Agent 也必须以某种方式检索并消费网络内容。然而，Parallel 从第一天起就押注一个截然不同的前提：**未来 Agent 与网络交互并发出查询的频率，将是人类有史以来的 1000 倍以上**。面对如此庞大的调用规模，现有的搜索底层技术乃至维系其运转的商业模式都必须彻底重构。 当被问及从管理 Twitter 这样规模庞大的超级平台到二次创业有什么心得时，Parag 直言，创办 Parallel 迫使他主动“忘掉”（unlearn）许多过去在 Twitter 积累的管理经验。在 Twitter 担任高管期间，面对的是一个已经确立了产品市场契合度（Post-PMF）且体量惊人的成熟平台，每日有数以亿计的用户在上面平均停留超过 30 分钟，公司的迭代和运营高度依赖海量真实用户即时、密集的反馈闭环。相反，现在作为一家 Pre-PMF 初创企业的创始人，所有的假设都建立在“几年后互联网上将涌现出一种全新形态的用户（即 AI Agent）”。他们必须每天、每周不断学习与摸索，面向这个“尚未真正全面降临的客户群体”去提前构建下一代底层系统。 - **核心假设**：AI Agent 查询网络的频率将比人类历史峰值高出 1000 倍（1000x），彻底颠覆现有的检索负荷与交互逻辑。 - **创业心境转换**：从管理每日亿级日活、高粘性的 Post-PMF 巨头 Twitter，转向为“尚未登场”的 Agent 客户构建 Pre-PMF 基础技术设施。 --- ### [03:25](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=205s) 什么是网络搜索？“十亿对十亿”的匹配难题 在深入探讨人类与 Agent 搜索的本质差异前，Andrew Reed 幽默地插话提到，这是他第一次在《Training Data》播客上正式作为主持人亮相。他随即抛出了一个看似基础却切中要害的问题：硅谷的从业者虽然经常把网络爬取（Crawling）、建立索引（Indexing）和重排（Ranking）挂在嘴边，但对搜索本质的认知往往停留在“给出答案”这一表象上；网络搜索最底层的核心难题究竟是什么？ Parag 拆解道，搜索的基石诉求非常直观：当一个人或 Agent 想要寻找某个信息却不知道它藏在互联网的哪个具体角落时，搜索引擎需要在一个最便捷的位置精准呈现答案。为了在毫秒级内完成这项任务，搜索引擎绝不可能在收到查询请求后才现场打开网页。相反，它必须预先通过网络爬虫抓取全网几乎所有的 URL，解析并理解页面内容，并将这些数据高度结构化地存放在内存或超高速存储中——这便是建立“索引”。当一次查询输入系统时，系统会迅速理解意图，从数千亿个网页的浩瀚海洋中，经过多层检索（Retrieval）与精细重排，一路缩减至十几个、五个甚至唯一一个最关键的结果。Parag 将这一本质概括为一个**“十亿对十亿的匹配问题”（a billion to billion matching problem）**：在宏观尺度上，一边是网络上不断膨胀的数百亿、数千亿级网页，另一边是跨越漫长时间维度的数百亿级查询请求，系统必须在这两者之间建立极其高效精准的撮合桥梁。 ```mermaid flowchart LR accTitle: 传统网络搜索核心流程与漏斗 accDescr: 展现从全网数百亿网页预处理，到查询发生时的多阶段缩减与匹配过程。 subgraph Offline["离线预处理阶段 (Offline)"] A["全网爬取<br/>(Crawling Web)"] --> B["解析与组织<br/>(Pre-processing)"] B --> C[("超大规模索引<br/>(Index in Memory)")] end subgraph Online["在线查询与匹配 (Online Query)"] D["实时查询请求<br/>(Query Input)"] --> E["意图理解与多阶段检索<br/>(Retrieval)"] C -.-> E E --> F["高精重排<br/>(Ranking)"] F --> G["最终精准呈现<br/>(Top 1 ~ 10 结果)"] end ``` --- ### [05:17](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=317s) 为什么敢挑战巨头做新索引：“人类点击数据其实是一个 Bug” Andrew 紧接着提出质疑：抓取全网内容并构建完整索引需要极其庞大且昂贵的底层基础设施，长期以来业界普遍认为只有 Google 和微软 Bing 这种万亿市值的超级巨头才拥有足够的算力和资本去运行真正 Web 规模的爬虫与索引。作为一家刚刚起步的年轻公司，凭什么有底气向这块巨头盘踞的领域发起冲击？ Parag 解释说，如果站在三年前来看，后来者无法与巨头竞争的根本原因，表面上是基础设施成本，更深层的壁垒其实是**高质量的反馈回路**——即判断“搜索结果 A 是否比搜索结果 B 更好”的评价体系。巨头垄断了海量真实人类用户的点击流数据（Click Data）以及昂贵的人工标注系统（Human Ratings）。然而，大语言模型的突破彻底打破了这一游戏规则：现在的先进模型不仅具备极其卓越的“信息压缩”（Compressing Information）能力，更能以极低成本生成堪比领域专家的评估与评分数据。 更具颠覆性的是 Parallel 对点击数据的底层认知。Parag 直言：**“在 Parallel 看来，人类点击数据其实是一个 Bug（Human click data is a bug）。一个借助搜索来完成复杂工作的 Agent，应当依赖的是 Agent 的反馈，而不是人类的反馈。”** 传统搜索引擎根据人类容易被标题党吸引、注意力短暂且充满偏见的点击行为进行优化，而这对执行严谨逻辑推理的 Agent 而言完全是干扰噪音。当把 Agent 本身视为独立的技术实体和全新的“目标客户”时，构建全网索引的问题在算法和逻辑上都变得更具可行性（Tractable）。此外，Parallel 找到了分步推进的商业化路径，打破了“必须先砸下数十亿美元建好全网索引才能服务第一个客户”的资本魔咒，而是随着为客户解决的问题不断增多，逐步滚雪球式地扩建更大、更复杂的动态索引。 --- ### [07:52](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=472s) 倒序打法：先做 Search Agent，再建搜索引擎与增量索引 面对 Sonya 关于“客户通常在第一天就要求全网覆盖，如何做到渐进式扩建”的追问，Parag 揭示了一个违背直觉的技术打法：**Parallel 并没有在第一天发布搜索引擎，而是先发布了一款“搜索智能体”（Search Agent）产品**。 在传统的即时搜索中，索引本质上是一种用存储换延迟的“延迟优化机制”（Latency Optimization）——为了在 200 毫秒内出结果，必须把一切都预先备在内存里。但在需要深度研究（Deep Research）的复杂工作流中，用户对等待时间的容忍度大幅提升，往往可以耐心等待一分钟乃至十分钟。如果用户愿意等待，且系统拥有一张足够精准的先验网络拓扑图来指导“优先爬取哪些高价值页面”，系统完全可以在收到查询请求**之后**，再去实时爬取网络。Parag 形象地指出：在项目早期，要在第零天直接挑战 Google 这种即时搜索引擎几乎是不可能的；但如果去竞争“雇佣外包人类坐在搜索引擎前花几十分钟搜集整理数据”的场景，Agent 的优势就显露无遗。通过用推理期计算（Inference-time Compute）换取爬虫索引的时间窗口，Parallel 成功避开了前期的天价算力墙。 在早期落地中，大量企业客户利用 Parallel 的 Search Agent 替换繁重的人工网络调研： - **保险与理赔工作流**：在核保（Underwriting）和复杂的理赔调查中，自动化穿透网络抓取涉案主体与资产的历史证据。 - **销售线索数据富化**：自动检索目标潜在客户的公司架构、动态与背景数据（Sales Data Enrichment）。 - **金融建模数据预处理**：过去金融分析师为了构建量化模型，通常需要将数据搜集任务夜间外包给海外的人工团队，第二天早上再把整理好的数据喂入模型；如今这些流程直接被 Parallel 的搜索智能体所取代。 通过在真实垂直场景中替代外包人工，Parallel 得以积累来自 Agent 的真实经验评测（Empirical Evals），从而明确摸清“专为 Agent 设计的搜索到底该长什么样”，而非停留在学术式的理论评估上。 --- ### [10:17](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=617s) 撕掉 Neo-lab 标签：Parallel 的终极交付物绝不是模型 针对 Parallel 的定位，Andrew 提出了一个观察：在业内开始流行“Neo-lab”（新型 AI 实验室）这一概念之前，Parallel 就已经聚集了大量顶尖的 AI 研究人员，同时也打造了极硬核的网络系统（Web Systems）。外界在初看 Parallel 时，往往很难一眼分清它究竟是一家纯粹的底层网络系统公司，还是一家前沿 AI 研究实验室；而随着交流深入，其技术架构呈现出极强的研究驱动属性。 面对“Parallel 是否算得上一个 Neo-lab”的提问，Parag 明确给出了否定的回答。他坚定地表示：**“我们不把自己归类为 Neo-lab。”** 尽管团队中拥有极高密度的人工智能科学家，且深度利用了大模型压缩与表征技术，但 Parag 认为，Neo-lab 的核心产出和商业终局通常是一个大基础模型本身。而对于 Parallel 而言，模型只是达成更优搜索与数据索引的强力工具，公司的真正交付物是面向 Agent 生态的高性能搜索与网络基础设施，绝非单纯对外提供一个通用的参数模型。 ### [11:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=663s) 作为模型的乘数而非 Neo-lab：从万亿网页中精准提取 1000 个 Token Parag Agrawal 明确划分了 Parallel 与当前一众前沿 AI 实验室（Neo-labs）的本质区别。在他看来，模型实验室的核心产出是基础大模型，而 Parallel 的定位是“模型的互补品与乘数”（complement and multiplier）。每当业内有人发布更强大的基座模型时，竞争对手面临的是更高难以逾越的高山；但对 Parallel 而言，外部模型的每一次升级，都自动解锁了三四个全新的落地场景，让其搜索能力的价值成倍放大。 尽管 Parallel 内部也进行大量前沿研发，但他们绝不涉足万亿参数大模型的预训练，其真正的技术攻坚点在于“模型极度压缩”与快速排序。Parag 将其定义为“千亿对千亿的匹配难题”：当一个 Agent 抛来请求时，本质上是在对搜索引擎说——“请从全网万亿级网页中，给我挑出最正确的 1000 个 Token”。就在节目录制的前一周，系统在读取时还需要耗费大约 3 秒钟的重度算力；而团队刚刚发布的 Turbo 产品，直接将这一过程骤降至 200 毫秒（200 ms）。这种极致的延迟压缩，依赖于跨内存层级的精细化数据组织，以及在极度受限的延迟与算力预算下，对预训练模型进行架构改造与精准调度。 - **模型乘数效应**：不争夺模型皇冠，而是成为赋予 Agent 超能力的底层加速器。 - **数据规模与目标**：从包含数万亿网页的庞大空间中，精准锁定最高信噪比的 1000 个 Token。 - **延迟跃升**：通过 Turbo 架构，将 Agent 搜索的端到端耗时从 3 秒一举压缩至 200 毫秒。 --- ### [13:14](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=794s) Agent 搜索对决传统 Google：Token 减半与算力智能再分配 针对主持人提问目前以 OpenAI Deep Research 为代表的深度研究智能体往往只是用模型进行推理并循环调用 Google，Parag 深入剖析了为什么直接套用传统搜索引擎是远远不够的。在深度推理循环中，如果直接调用 Google 或其代理，Agent 确实能获取一些线索，但如果改用专为 Agent 设计的 Parallel 搜索，Agent 消耗的 Token 数量普遍可以减少一半以上，整体准确度更高，端到端执行速度也更快。 节省一半的 Token 对 Agent 架构意味着质的飞跃。在长上下文和显存容量受限的现实约束下，更紧凑的信息输入意味着相同的上下文窗口可以承载更复杂的推理链条，或者以十分之一的成本跑完同样的工作流。在 Parag 眼中，面对 AI 对相关信息永无止境的吞噬欲望，整个系统的底层逻辑归根结底是“质量、成本与延迟”的三方博弈。Parallel 的核心赌注正是在模型层、Agent 层与底层的搜索层之间实现“算力的智能再分配”，将纯净信号剥离出来直接喂给大模型。 --- ### [15:12](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=912s) 交互接口的范式转移：懒惰的人类与精确的 Agent 面向 Agent 与面向人类用户的搜索，其交互界面正经历一场底层范式的转移。人类在向 Google 提问时习惯采用短促的关键词，主持人调侃人类大脑或许只能在短期记忆中暂存几个词，Parag 则幽默地反驳道：“不，我们的大脑能容纳很多词，我们只是纯粹太懒了。”人类日常交谈能长篇大论，可一旦把手放在搜索框上，就只想敲两个错漏百出、带着拼写错误的残缺词汇，甚至指望下拉菜单的自动补全来省下打第三个词的力气。 相比之下，AI Agent 完全没有人类的这种劣根性。Agent 发送给搜索引擎的查询没有错别字，语义定义严谨详尽，长度往往更长，且能明确交代上下文背景。这意味着搜索引擎不再需要费尽心机去“盲猜”用户到底想要什么，原本在传统搜索中用于消除人类歧义的复杂算法可以彻底让位，去解决完全不同维度的高阶检索难题。 --- ### [16:21](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=981s) 解构“SEO 垃圾信息”：从 SEC 财报 PDF 到 Agent 的上下文直达 针对主持人提到 Google 搜索结果中充斥着各类前 AI 时代的“SEO 垃圾站（pre-AI slop）”与联盟营销页面，Parag 展开了一段极富洞察力的辩护与反思。他举了一个生动的经典案例：如果用户想查询某家上市公司的最新核心营收数据，最权威的数字无疑静静躺在公司提交给 SEC（美国证券交易委员会）的一份 300 页官方财报 PDF 的第 73 页。 ```mermaid flowchart TD accTitle: 人类搜索与 Agent 搜索路径对比 accDescr: 对比传统人类搜索需要通过 SEO 汇总页获取数据，而 Agent 搜索能直接穿透至权威信源并提取摘要。 subgraph HumanFlow["传统人类检索路径"] A["查询上市公司营收"] --> B["Google 搜索结果"] B --> C["SEO 二手摘要网页<br/>（1秒加载/首屏提炼/99.99%准确）"] C --> D["人类因懒惰放弃翻阅<br/>300页 SEC 财报 PDF"] end subgraph AgentFlow["Parallel 智能体检索路径"] E["Agent 发送精准 Query"] --> F["Parallel 架构多级召回"] F --> G["定位 SEC 官方权威 PDF<br/>第73页核心财务段落"] G --> H["精准提取 1000 Token<br/>直接置入 Agent 上下文窗口"] end ``` 如果 Google 严格恪守“权威至上”，把这份 PDF 排在第一位，普通用户点开后需要忍受长达 3.5 秒的漫长加载，接着还要像在终端敲 grep 指令一样在数十页密密麻麻的条款中翻找，用户立刻就会感到挫败。正因如此，互联网上才诞生了一整套 SEO 衍生生态——他们把核心数据做成首屏展示、1 秒即可加载完毕的摘要网页。这些页面虽然充斥着广告，但切实解决了懒惰人类的核心痛点。然而 Agent 既不需要点击浏览，也不会因翻看 PDF 而烦躁。Parallel 能够直接穿透到全网最权威的源头，切出最核心的原文段落，直接送入 Agent 的上下文窗口，彻底粉碎了因迎合人类弱点而不得不做出的妥协。 - **权威与体验的割裂**：权威源（如 SEC 300 页财报 PDF）加载慢、信息埋藏深；人类因懒惰拒绝手动翻阅。 - **SEO 网页的真实价值**：在信息呈现上将 20 个关键数据置于首屏（above the fold），以牺牲部分纯净度为代价换取即时性。 - **Agent 的降维打击**：Parallel 消除了对中间二手摘要站的依赖，直接将原始权威出处的精确切片送达 Agent 上下文。 --- ### [19:38](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1178s) 揭秘 Parallel 架构内部：从百亿文档到 1000 Token 的多级过滤流水线 当开发者向 Parallel 的 Agent 搜索 API 发送一条请求时，背后是一套严密的现代信息检索与模型调度网络。第一阶段，系统调用专门的模型解析并丰富该 Query，判断其意图，并将其重写分发至底层多个异构子索引中——这包括全网主索引、实时新鲜索引（Fresh index）以及专门的结构化知识图谱（Knowledge Graph）。 ```mermaid flowchart LR accTitle: Parallel Agent 搜索处理全流程 accDescr: 从开发者请求进入，经过意图分析、子索引检索、粗排重排，最终输出1000个高价值 Token。 subgraph Ingestion["阶段一：分析与重写"] Q["开发者 Query"] --> Enrich["Query 解析与意图增强"] Enrich --> Rewrite["生成多路专用 Query"] end subgraph IndexRouting["阶段二：异构子索引并行召回"] Rewrite --> IdxMain["全网主索引<br/>(Big Index)"] Rewrite --> IdxFresh["实时索引<br/>(Fresh Index)"] Rewrite --> IdxGraph["结构化知识图谱<br/>(Knowledge Graph)"] end subgraph RankingPipeline["阶段三：多级漏斗重排"] IdxMain & IdxFresh & IdxGraph --> Ret["检索层粗筛<br/>(数百亿级 -> 数万篇)"] Ret --> Rank1["初级轻量重排"] Rank1 --> Rank2["深层大模型段落重排<br/>(提取核心 Excerpts)"] end Rank2 --> Out["最终输出：1000 Token<br/>最高信噪比直接回传 Agent"] ``` 在多路召回之后，流水线进入极为密集的层级筛选机制。系统必须在极短时间内将数百亿乃至上千亿的候选网页与文档，迅速削减至数万篇，再进一步切碎成具体的段落切片（Excerpts）。在漏斗的每一级，系统部署了体积由小到大、架构各异的模型提取特征并做重排。不同版本的 Parallel API（例如 Turbo 与标准版）本质上是在这一流水线的不同阶段调配不同的算力配比，以完美命中客户特定的成本与延迟指标。当主持人追问其衡量质量的“北极星指标”时，Parag 坦言，百亿对百亿的匹配是一个“永恒的问题（forever problem）”，真正的关键不在于是否存在一个绝对完美的打分，而在于算力边际投入的临界点——何时再追加算力优化已经不再划算。 ### [22:22](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1342s) 预训练爬虫 vs 实时搜索索引：为什么大模型公司不会直接通吃 随着模型智力每隔几个月就变得更强、更便宜，整个技术栈的核心正演化为三件事的结合：成本低廉的顶尖大模型、高质量的外部信息，以及在具体需求与海量内外部数据之间做精准匹配。投资人经常提出一个针锋相对的质疑：抓取网页与构建索引是大模型预训练的基础设施，大模型厂商为了预训练本就投入了数十亿美元做网页抓取，这天然构成了他们的“事实来源”（Source of Truth），他们难道不会顺理成章地通吃实时搜索？ 对此，Parag Agrawal 明确表示，从一线实践的经验数据来看，预训练爬虫与实时网页索引完全是两码事。预训练追求的是原始 Token 规模（比如数万亿 Token），极度注重单位算力的吞吐效率。面对那些加载缓慢、充斥着动态 JavaScript 的现代网页，模型公司的爬虫根本没有耐心等待，因为“为了抓取这些页面所消耗的算力与产出的 Token 数量相比，效率差了一到两个数量级”。相反，Parallel 所做的是大模型的互补品（Complement）：Parallel 必须是一个追求极致覆盖的“完备主义者”（completionist），专门去抓取大模型预训练时由于性价比过低而放弃的长尾动态网页。 大模型在处理绝大多数实际业务时，不可能单靠静态权重活着。Parag 认为，在企业购买大模型用于实际工作的场景中，“十之八九的用例都需要赋予模型访问实时网络的能力”，这使得针对 Agent 优化的搜索基础设施成为所有 LLM 推理的刚需邻近层。面对这个庞大的专有需求，模型厂商未来必然面临“自建还是外采”（Build vs. Buy）的抉择。 ```mermaid flowchart TD accTitle: 预训练抓取与 Agent 搜索抓取的模式对比 accDescr: 对比大模型预训练爬虫与 Parallel 完备主义爬虫在目标、策略与执行上的核心差异。 subgraph PreTrain["大模型预训练爬虫"] P1["目标：追求万亿级 Token 吞吐"] --> P2["遇到复杂慢速 JS 网页"] P2 --> P3["放弃抓取（算力产出比差 1-2 个数量级）"] end subgraph AgentSearch["Parallel 实时 Agent 索引"] A1["目标：实时性与全网完备覆盖"] --> A2["遇到动态渲染与复杂交互"] A2 --> A3["耐心完整渲染并建立索引（补足模型盲区）"] end ``` --- ### [25:48](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1548s) 重磅合作：成为 Google Cloud 企业的搜索与 Grounding 引擎 虽然 Parag 对与其他前沿模型实验室的未公开谈判保持克制，但他现场公布了一项重磅合作：Parallel 已正式与 Google Cloud（GCP）达成合作，成为其企业级 Agent API 的底层搜索与知识对齐（Grounding）服务商。这意味着当企业客户在 GCP 上使用 Gemini 或其他开源模型构建智能体或对话应用并需要连接实时网络时，系统内置的可选引擎只有两个：Google 自身的搜索，或者 Parallel 搜索。 这项集成并非简单的 API 转发，而是两家技术团队在底层系统、模型训练、产品及商业化维度全面深度打磨的结果，确保 Gemini 搭配 Parallel 能够输出极具确定性的检索与推理结果。面对主持人的惊叹——在搜索领域的老牌霸主 Google 面前拿下这一阵地，称得上是不可思议的突破——Parag 幽默地回应道：“Google 可是最原始的‘新潮实验室’（Google's the original Neolab）。”这场合作印证了即便强如超大规模云厂商，在针对 Agent 的专用检索架构上也正在向第三方专业生态敞开大门。 - **GCP 落地场景**：涵盖 Gemini 模型的 Grounding、企业级 Agent 开发以及各类生成式 AI 推理任务。 - **深度工程优化**：Parallel 团队与 Google 研发团队深度协同，专门针对模型读取网页结构与内容召回进行了通道级优化。 --- ### [27:27](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1647s) 从 10 倍到百万倍：Agent 带来的搜索查询倍增效应 当业界谈论 Agent 将带来“千倍于人类的搜索量”时，这并非一句空洞的口号，而是正在真实发生的工作流演变。Parag 将这种搜索需求的激增拆解为不同层级的放大乘数（Multipliers）： 最基础的层级是交互式对话 Agent。用户以往在 Google 上输入一个关键词只触发一次检索，但在使用 ChatGPT 等产品的深度研究或推理功能时，模型为了回答一个稍微复杂的 Prompt，哪怕在几秒钟内完成响应，后台也会自动并发执行 5 到 20 次检索；如果开启深度调研模式，查询次数会瞬间飙升到数百甚至上千次。仅仅是将人类意图翻译为模型动作，搜索量就已经发生了一个数量级的放大。 真正产生质变的是自主运行的后台企业级 Agent（Background Agents）。Parag 举了一个极具代表性的金融风控案例： > 一家机构为 10,000 家小微企业提供信贷支持。过去，他们每月依赖人工流程去网上翻查企业动态、评估违约风险；而引入开发者编写的自动化 Agent 后，这套基于全网信息抓取的风控流程直接由月度巡检变成了每周自动跑批。为了生成全景仪表盘和异常预警清单，Agent 每周在全网发起的搜索量高达数十万乃至数百万次。 在这个场景下，一个程序员写下的几行调度代码，便直接在底层激发出上百万倍的搜索乘数效应。 --- ### [30:18](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1818s) 会议筹备 Agent 与算力价值的天花板 谈及个人日常工作流，主持人调侃 Sequoia 内部也有专门的投前调研 Agent（甚至开玩笑提到他们优秀的年轻投资人 James Flynn 就是他们的“真人 Agent”），而 Parag 则分享了他在 Notion 内部署的定制智能体。他只写了一次 Prompt，配置了一个能够同时打通 Parallel 内部企业知识库以及 Parallel 外部网页搜索 API 的会议筹备 Agent。如今，他日程表上的每一场会议召开之前，Agent 都会自动在后台发起数十到上百次全网检索，抓取参会人员背景、最新动态与业务交集，输出结构化的前置备忘录。 这类后台 Agent 不断延伸、无休无止地为人类工作，其调用的天花板究竟在哪里？Parag 认为，唯一的硬边界是“产出价值与算力消耗的经济账”（Value vs. Spend）。目前许多人甚至没有完全理性地去精打细算成本，因为单次调用的绝对单价极低，“虽然我现在花在上面的钱可能超过了理性限度，但总额并不大，所以我暂时根本不在乎”。未来伴随企业对 ROI 认知的成熟，后台 Agent 会理性扩展到所有“边际产出大于边际算力成本”的领域，而每一次扩展都伴随着底层搜索 API 的海量并发消耗。 --- ### [31:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1906s) 质量、成本与延迟：Turbo 的研发路径与 Agent 搜索的未来 在公司成立的前两年，Parallel 确立了技术攻坚的三维坐标体系：**质量（Quality）、成本（Cost）、延迟（Latency）**。在资源有限的创业初期，团队做出了一个看似反常的战术决策：**完全无视延迟，集中全部精力死磕质量与成本**。Parag 解释说，模型蒸馏、系统工程剪枝和吞吐量压榨是“已知的工程技术”（known art），而如何在无界网络中获取最高质量、对 Agent 最有效的信息，则是充满未知探索的研究难题。 一旦他们在各种定价区间内确立了质量与成本的绝对代差，团队立刻将重心转向延迟优化，并推出了名为 **Turbo** 的产品。这项技术将原本需要数秒的 Agent 级深度搜索压低至前所未有的极速响应，成为市场上速度最快、质量最高的 Agent 专用网络检索产品。 在被问及“目前互联网上的 Agent 查询总量是否已经超越人类查询”时，Parag 保持了客观与清醒： - **大众市场仍处于极早期**：硅谷容易陷入信息茧房，走出这个圈子，普通大众甚至连当下最新的 AI 工具名词都没听过，因此总体查询量目前仍由人类占大头。 - **重度用户的千倍跃迁**：对于 Parag 这样的从业者，三年前每天大约只做 20 到 30 次 Google 搜索；而今天加上其名下所有自动化 Agent 和公司后台进程，搜索频率已轻松达到当年的 100 到 1,000 倍以上。 - **流量构成的临界点**：根据 Cloudflare 等网络基础设施厂商的公开监测，目前 AI 爬虫与机器人的全网页面读取量（Page reads）在体量上已经与人类浏览流量分庭抗礼，这预示着底层网络数据消费结构的根本性拐点正在迅速逼近。 ### [34:21](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2061s) 广告模式的破裂：当智能体取代人类眼球 随着 AI 智能体（Agent）正在接管互联网访问，原先建立在“人类注意力稀缺性”与“人类眼球点击”之上的互联网经济模型正面临全面崩塌。Parag Agrawal 坦言，这正是促使他创立 Parallel Web Systems 的核心动因之一。虽然大众普遍反感广告，但曾在 Twitter 负责广告系统的 Parag 强调，从底层逻辑上看，广告模式其实非常值得尊重——正是因为广告的存在，Google 搜索、Twitter 以及全网海量的优质内容和技术工具才能向全人类免费开放。 广告之所以是一套极其高效的商业化机制，关键在于它具备无与伦比的“差异化定价”（Differential Pricing）能力。在 Google 搜索中，绝大多数普通查询其实是在赔钱的，但少部分极具商业意向的高价值查询却能赚回暴利，从而支撑起极高的利润率。主持人调侃道，热衷网购的用户相当于在自掏腰包补贴全网获取免费信息的人。然而，这一逻辑成立的前提是“人类有限的注意力能最终转化为商业转化”。一旦访问网络的不再是人类，而是代替人类执行任务的智能体，整个经济循环便彻底脱节。 ```mermaid flowchart TD accTitle: 传统广告模式与智能体网络经济模式断裂对比 accDescr: 对比人类注意力驱动的广告/订阅模式与无名智能体访问导致商业闭环断裂的过程 subgraph Traditional ["传统互联网经济模型（以人为核心）"] H["人类用户（有限注意力）"] --> Search["搜索引擎 / 社交平台"] Search --> Ads["高精准广告与差异化定价"] Search --> Sub["内容网站转化为订阅（如 1000 人转化 20 人）"] Ads & Sub --> Revenue["支撑免费内容与高利润商业体"] end subgraph AgentBreakdown ["智能体驱动的网络（商业闭环断裂）"] AH["人类委托代理"] --> Agent["AI 智能体并发检索"] Agent -.->|"访问并抓取内容"| Web["内容发布商 / 高价值网页"] Web --> Block["无法衡量转化 / 无名智能体被封禁"] Block --> Crisis["内容方拒绝智能体访问，旧变现体系瓦解"] end ``` - **注意力经济瓦解**：内容方过去面对人类访客，有着成熟的转化模型（例如吸引 1000 个真实人类，能转化出 20 个月度付费订阅者）。 - **智能体的“身份盲区”**：智能体访问网站时，发布商既无法追踪统计，也不知道这些“无名智能体”是否会带来订阅转化，甚至感觉自身内容只是在被单向度地“白嫖”。 - **相互博弈的困境**：内容创作者陷入两难——他们一边花钱做 SEO 吸引人类，一边又设立防火墙切断智能体的爬取。尽管智能体最终代表的是人类意图，但在缺少商业闭环的情况下，内容方根本无法变现这类流量。 --- ### [37:20](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2240s) 头部买断协议的死局与新型激励对齐机制 在 Parallel 出现之前，高价值内容版权方如果想通过 AI 获得收益，唯一的路径就是作为极少数的“头部（Head）媒体”，与 OpenAI 等头部基础模型实验室签署固定金额（Fixed-fee）的授权协议。Parag 回忆起自己当年在 Twitter 主导向 OpenAI 出售数据交易的经历，并指出这种买断合同主要捆绑了三项诉求：一是用于模型预训练的数据授权；二是为过去未经授权的抓取提供法律责任豁免（Liability Protection）；三是推理阶段（Inference Time）的实时数据检索。 但 Parag 直言，这种“固定总价”的合作模式不仅绝大多数长尾内容方根本无法参与，即便对头部媒体而言也是完全不可持续的死胡同。随着 AI 推理调用量呈指数级爆发（比如今年增长 7 倍，明年再增长 7 倍，实现 50 倍增长），没有任何一家媒体在签署为期两年的固定合同后，会相信自己在续约时能拿到同比例增长 50 倍的分成；相反，内容方在整个价值链中的份额势必被急剧稀释。真正合理的解决方案必须具备以下关键特质： - **双重维度的差异化定价**：不仅要针对极其优质的内容支付高昂溢价，更要根据智能体调用该内容时所执行任务的高价值程度（Value of Work）进行浮动定价。 - **可规模化运转**：摆脱依靠法务与商务团队针对单个机构进行低效、定制化商业谈判的旧路径，转为全自动化的大规模分配协议。 - **彻底的激励对齐（Incentive Alignment）**：让内容所有者不仅有动力向模型开放内容，还能从 AI 推理生态的指数级增长中持续获得合理分成。 --- ### [40:48](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2448s) 沙普利值归因：如何量化每一篇网页的贡献 在一次复杂的智能体工作流中，即使一个极其简单的用户提示词，底层也可能触发十次以上的网页检索，并最终浓缩成短短一段文字总结。面对这种错综复杂的协作，AI 公司究竟该如何精确计算每个来源的贡献并合理分账？Parag 提出了一套基于博弈论的理论模拟实验：假设从检索数据池中撤掉某个特定网页，观察智能体的输出质量下降了多少；如果要将质量拉回到原有水平，系统可能需要调用更大尺寸的模型或消耗更多的推理算力（例如需要额外花费 1 美分的算力才能弥补这一质量损失）。那么从最终用户价值的角度看，该网页来源的实际经济价值就大约等同于这 1 美分。 这种思想在博弈论中的严格数学形式即为**沙普利值（Shapley Values）**。沙普利值专门解决“多人合作创造出大于各部分简单相加的总体价值（Pie）时，如何公平分配收益以使所有人自愿保持合作”的难题。机器学习领域常用它来做特征重要性评估（Feature Importances），但直接计算完整的沙普利值在现实中计算成本极高，因为必须穷举并模拟所有参与者子集排列组合的平行世界。 | 对比维度 | 纯数学理论沙普利值 | 传统广告竞价（第二价格拍卖） | Parallel 估算模型解决方案 | |:--- |:--- |:--- |:--- | | **核心机制** | 穷举所有子集排列组合，严格计算边际贡献 | 依赖博弈论设计真实报价机制，现代演化为基于 ROI 的自动出价（Auto-bid） | 采集海量模拟评估数据，训练专用估算模型替代穷举 | | **计算成本** | 极高（算力开销远超支付给内容方的金额） | 极低（毫秒级竞价算法） | 极低（模型即时推理预估） | | **应用痛点** | 算力成本倒挂（为了计算 1 美元分成可能要花数美元算力） | 无法衡量非眼球经济与复杂的长链路生成内容 | 依赖高质量的 Eval（评测系统）与模拟数据质量 | | **最终目标** | 理论上绝对公平的博弈均衡 | 广告主利益最大化 | 激励对齐，最大化内容方与 AI 智能体的全网协作 | Parag 指出，如果为了确定给某位内容创作者支付 1 美元，系统却不得不消耗数美元的算力去跑全量沙普利值模拟，这在商业上显然是荒谬的。因此，Parallel 的解法是利用强大的评测体系（Evals）针对“撤除特定 URL、特定域名或组合”等场景运行离线模拟，收集高质量数据，进而训练出能精准预测沙普利值的高效代理模型。如同早年广告系统从机制设计（第二价格拍卖）逐步走向完全基于 ROI 反馈回路的“自动出价（Auto-bid）”一样，Parallel 致力于通过扎实的博弈论底层框架实现激励对齐，让内容创作者与 AI 使用方在透明对称的体系中实现全网良性协作。 ### [45:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2746s) 基于沙普利值的内容激励与宏观商业闭环 在谈及让内容创作者和发布商接受沙普利值（Shapley values）分配机制时，Parag Agrawal 坦言，坐在内容所有者面前去解释背后的沙普利数学原理“确实需要花点时间”，但只要把核心特征讲透，对方就会豁然开朗：创作者直接参与价值分配，提供越独特、越具差异化的数据，赚得就越多。同样是一次阅读，如果是一位手握高薪的银行家在处理高价值业务时被 Agent 读取了你的数据，相比于 Parag 退休的父亲随手翻阅，前者的阅读会带来高得多的经济回报。 从宏观经济模型的角度来看，这套逻辑完全能够自洽。未来人类在知识工作场景下的 LLM 推理（inference）上将投入庞大的资金，若将其中的 2% 到 10% 重新分配给开放网络上的数据提供方，其规模将远超如今除 Facebook、LinkedIn 等围墙花园（walled gardens）之外的所有传统网页数据商业模式。随着网络上的 Agent 数量以每年一个数量级的速度爆发式增长，Parag 计算认为，只需 12 到 24 个月，这套沙普利分配数学就能转化为实打实的真金白银，惠及极广泛的 Web 内容创作者。 面对 Stack Overflow 流量断崖式下跌、传统“人类互联网”因激励机制失效而瓦解的现状，主持人 Sonya Huang 感慨，看到 Parallel 致力于重构让人们持续创作输出的激励机制，非常振奋人心。Parag 则明确回应：“这就是我们创办这家公司的初衷。” - 核心收益原则：数据的独特性与使用场景的商业价值直接决定分成，高价值商业工作中的 Agent 调用将支付更高溢价。 - 宏观测算：知识工作 LLM 推理预算的 2%–10% 流向开放 Web 数据，其资金池将超越当前开放互联网广告模式的总和。 - 落地时间表：预计在 12 到 24 个月内，沙普利值分成模型将开始为广泛的网络内容拥有者提供可观的真实美元收益。 --- ### [47:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=2866s) 从 Shapley Inc. 到 Parallel：专为 AI 演进的双轨互联网 当被问及公司为何命名为“Parallel”时，Parag 透露了一个有趣的内幕：公司最早注册的名字其实是“Shapley Inc.”。当时他全神贯注于解决价值归因的核心问题，虽然明知“Shapley”作为 B2B 产品名称体验极差，但因一时找不到更好的词便先以此注册，顺手拿下了 Shapley.ai，而 Shapley.com 实际上是一个停泊的色情域名。为了避免尴尬，在随后长达 6 到 8 个月里，他对外绝口不提 Shapley 这个名字，参加各类活动的胸卡上也只写着“NCO”（New Company）或“隐形期初创公司”（Stealth Company）。 最终选定“Parallel”，一方面契合了他们底层系统高并发、并行处理的工程特质；更重要的原因在于团队构想的全新范式——一个专为 AI 打造的“平行网络”（Parallel Web）。过去人们发布网页只考虑人类读者；而现在创作者开始处于“双重发布”（dual publishing）的现实中：既要考虑人类眼中的页面长相，又要确保 Agent 能够高效读取与解析。 主持人 Andrew Reed 补充道，许多高价值领域早已完成向“Agent 优先”的倾斜，例如上市公司财报电话会（earnings call transcripts）：通过 AI Agent 获取和消化财报的人数，早已远超真正听音频直播或逐字阅读原始速记稿的人。如果今天由他担任上市公司 CEO，他会全力确保发言能被 Agent 精准转录和理解。Parag 赞同并指出，Parallel 自身在发布 API 文档和 SDK 时也是如此测试的——阅读文档的主要是客户用于构建 AI 系统的 Agent，而非手忙脚乱在页面上搜寻的人类开发者，因此 Agent 才是文档的第一受众。 ```mermaid flowchart TD accTitle: 双重发布模式结构图 accDescr: 展示内容创作者同时面向人类 Web 和 Agent 平行 Web 发布内容的双轨流程。 subgraph Origin["内容创作源头"] A["高价值内容发布<br/>(API 文档 / 财报 / 专业知识)"] end subgraph DualWeb["平行 Web 架构"] B["人类网络 (Human Web)<br/>视觉排版 / 交互动效 / 广告呈现"] C["平行 Agent 网络 (Parallel Web)<br/>结构化数据 / 低延迟检索 / 高度语义可读"] end subgraph Consumers["消费端"] D["人类读者<br/>(低频浏览与阅读)"] E["自主 Agent / 子 Agent<br/>(高频调取 / 自动化分析 / 执行工作流)"] end A --> B A --> C B --> D C --> E ``` --- ### [51:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=3063s) Agent 演进的三重境界与从“拉取”到“推送”的未来 Web 在畅想 Parallel 的终极愿景时，Parag 将 Web 这一轮基于 Agent 的演进划分为三个阶段： 第一阶段（当前主流阶段）：开发者构建相对简单的单体 Agent，将网络单纯视作传统的外部搜索工具（Web Search Tool）。这是因为当前的大模型主要基于人类既有工具的习惯进行微调，直接套用人类搜索模式来完成任务。 第二阶段（前沿探索阶段）：一部分先锋客户开始构建复杂的“多 Agent 系统”（Multi-Agent Systems）。这类架构中包含了分工明确的子 Agent（sub-agents），它们彼此唤醒、归档信息或并行探索。从当下代码编写 Harness 里的子 Agent 协作，到能够进行长时间、重算力投入的“AI 科学家”（AI Scientist）系统，都体现出这一特征。 第三阶段（未来愿景：Web 从 Pull 转向 Push）：目前无论前两个阶段如何演变，本质都是 Agent 遇到需求时主动向网络“拉取”（Pull）信息（“立刻帮我找这个”）。但在未来两三年内，Web 的运行逻辑将被彻底反转为“推送”（Push）机制——用户向 Parallel 声明：“如果网络上发生了某种特定事件，请立刻呼叫我，以便我的 Agent 或人类去跟进处理。” Parag 常对团队强调一个基本逻辑：在算力充沛的前提下，任何今天能做且有价值的事，Agent 会立刻执行完毕，绝不会推迟到明天。因此，未来的 Agent 每天要执行的新任务，必然是被外界的变化所触发——可能是卫星图像的更新、客户发表的新评论、另一个 Agent 跑完的计算结果，或是全网信息流的最新异动。当网络成为驱动全球 Agent 自动化运转的实时感知源时，Parallel 将代表所有客户持续把算力分配在全网动态的监控与理解上。“一旦符合条件的事件发生，立即通知我并唤醒 Agent”，这将把 Web 从即时查询工具，升维为驱动全自主智能社会的神经中枢。 ## Notable quotes > [00:57](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=57s) 未来 Agent 与网络交互并发出查询的频率，将是人类有史以来的 1000 倍以上。 — Parag Agrawal > [03:25](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=205s) 在宏观尺度上，一边是网络上不断膨胀的数百亿、数千亿级网页，另一边是跨越漫长时间维度的数百亿级查询请求，系统必须在这两者之间建立极其高效精准的撮合桥梁。 — Parag Agrawal > [05:17](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=317s) 在 Parallel 看来，人类点击数据其实是一个 Bug。一个借助搜索来完成复杂工作的 Agent，应当依赖的是 Agent 的反馈，而不是人类的反馈。 — Parag Agrawal > [10:17](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=617s) 我们不把自己归类为 Neo-lab。 — Parag Agrawal > [11:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=663s) 请从全网万亿级网页中，给我挑出最正确的 1000 个 Token。 — Parag Agrawal > [15:12](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=912s) 不，我们的大脑能容纳很多词，我们只是纯粹太懒了。 — Parag Agrawal > [22:22](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1342s) 十之八九的用例都需要赋予模型访问实时网络的能力。 — Parag Agrawal > [25:48](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1548s) Google 可是最原始的‘新潮实验室’（Google's the original Neolab）。 — Parag Agrawal > [30:18](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1818s) 虽然我现在花在上面的钱可能超过了理性限度，但总额并不大，所以我暂时根本不在乎。 — Parag Agrawal > [31:46](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=1906s) 完全无视延迟，集中全部精力死磕质量与成本。 — Parag Agrawal > [51:03](https://www.youtube.com/watch?v=fUcnE6pjq5w&t=3063s) 如果网络上发生了某种特定事件，请立刻呼叫我，以便我的 Agent 或人类去跟进处理。 — Parag Agrawal ## Mentioned Parag Agrawal, Sonya Huang, Andrew Reed, James Flynn, Elon Musk, Parallel Web Systems, Parallel, Twitter, Google, Google Cloud, Microsoft Bing, Sequoia Capital, OpenAI, ChatGPT, Gemini, Turbo, Deep Research, SEC, Cloudflare, Stack Overflow, Facebook, LinkedIn, Notion, Shapley Inc. ## Video description Parag Agrawal is making a bet that goes against two decades of web search: agents will query the web a thousand times more than humans ever have, and the infrastructure built around human clicks is wrong for them. The former Twitter CEO, now founder and CEO of Parallel Web Systems, explains why Parallel treats human click data as a bug and trains on agent feedback instead. He unpacks the counterintuitive choice to ship a search agent before a search engine, building an index incrementally, and how the new Turbo product cut agentic search to 200 milliseconds. But the problem Parag keeps returning to is economic: the ad-supported internet collapses when agents show up instead of people. His fix draws on Shapley values to pay content owners for the value their pages provide agents, with real dollars reaching publishers, he predicts, within 12 to 24 months. Hosted by Sonya Huang and Andrew Reed, Sequoia Capital 00:00 Introduction 03:25 What Is Web Search 05:17 Why Start a New Index 07:52 Search Agents First 10:17 Not a Neolab 13:14 Agents vs Google Search 19:38 Inside the Search Stack 28:59 Search Multipliers With Agents 30:21 Meeting Prep Agent Workflows 31:46 Quality Cost Latency And Turbo 32:42 Are Agents Overtaking Humans 34:28 Ads Model Meets Agent Web 37:20 New Incentives For Content 40:48 Shapley Values Attribution 47:46 Parallel Web And Future Vision [Source](https://www.youtube.com/watch?v=fUcnE6pjq5w)

[0:57] Parag Agrawal: 未来 Agent 与网络交互并发出查询的频率，将是人类有史以来的 1000 倍以上。 [3:25] Parag Agrawal: 在宏观尺度上，一边是网络上不断膨胀的数百亿、数千亿级网页，另一边是跨越漫长时间维度的数百亿级查询请求，系统必须在这两者之间建立极其高效精准的撮合桥梁。 [5:17] Parag Agrawal: 在 Parallel 看来，人类点击数据其实是一个 Bug。一个借助搜索来完成复杂工作的 Agent，应当依赖的是 Agent 的反馈，而不是人类的反馈。 [10:17] Parag Agrawal: 我们不把自己归类为 Neo-lab。 [11:03] Parag Agrawal: 请从全网万亿级网页中，给我挑出最正确的 1000 个 Token。 [15:12] Parag Agrawal: 不，我们的大脑能容纳很多词，我们只是纯粹太懒了。 [22:22] Parag Agrawal: 十之八九的用例都需要赋予模型访问实时网络的能力。 [25:48] Parag Agrawal: Google 可是最原始的‘新潮实验室’（Google's the original Neolab）。 [30:18] Parag Agrawal: 虽然我现在花在上面的钱可能超过了理性限度，但总额并不大，所以我暂时根本不在乎。 [31:46] Parag Agrawal: 完全无视延迟，集中全部精力死磕质量与成本。 [51:03] Parag Agrawal: 如果网络上发生了某种特定事件，请立刻呼叫我，以便我的 Agent 或人类去跟进处理。

[0:00] Parag Agrawal: human click data is a bug, not a feature, when building search for AI agents. [0:00] Parag Agrawal: AI agents will perform web searches and navigate pages at a scale 1,000 times greater than humans ever have. [3:25] Parag Agrawal: billion-to-billion matching problem [7:52] Parag Agrawal: Search Agents First, Search Engines Second [10:17] Parag Agrawal: 3 seconds down to just 200 milliseconds. [13:14] Parag Agrawal: under half the context tokens [13:14] Parag Agrawal: Quality, Cost, and Latency [13:14] Parag Agrawal: lazy searchers [23:34] Parag Agrawal: completing component [40:48] Parag Agrawal: If a banker in an expensive job reads your data versus my retired dad reads your data, the banker ends up paying more for that read because it's a part of high-value work. [47:46] Parag Agrawal: Shapley Inc. [47:46] Parag Agrawal: parallel web designed explicitly for AI agents alongside the human web

[0:00] Our view at parallel is that human click [0:03] data is a bug and agent [0:08] doing work with search should rely on [0:10] agent feedback not human feedback. We [0:13] believe that these models are really [0:16] good at compressing information and we [0:18] can benefit from a lot of the research [0:20] that have gone into building models and [0:22] apply it to search indexing and ranking. [0:25] And so you can now make many many [0:28] arguments and that's the arguments we [0:30] made back then that actually now it's [0:32] way more tractable as a problem because [0:35] of the existence of agents not just as [0:37] in technology but as a distinct customer [0:47] [music] [0:55] >> [music] [0:57] >> Parog, thank you so much for joining us [0:59] today. Uh we're delighted to have you on [1:01] the show. For those who don't know, [1:02] Parog of uh Twitter CEO fame was was CEO [1:06] of Twitter um before selling it to Elon. [1:09] Uh and is now back on the founder arc. [1:12] Um you founded a company called Parallel [1:14] Web Systems, which is scaling up Agentic [1:17] Search uh for the Agentic Web. Um, we're [1:19] very excited to have you here today to [1:21] talk about the future of search and the [1:22] future of the internet. So, let's get [1:24] started. What's Parallel? [1:25] >> Thank you, Sonia, for having me. Thanks, [1:27] Andrew, for joining us. I'm really [1:29] excited about this conversation. At [1:31] Parallel, we're building uh a bunch of [1:34] technology in order to allow agents to [1:38] search and use the web. So just like [1:40] humans forever have figured out how to [1:42] use browsers and search engines and [1:44] clicking and browsing around the web to [1:46] access information, uh agents need to do [1:49] the same things. We started parallel [1:51] with the bet that agents would do it a [1:52] thousandx more than humans ever have. [1:54] And as a result, we need to reinvent the [1:58] technology that can power search for [2:01] agents, but also the business models [2:04] that go alongside it. And that's what [2:06] we're doing at Parallel. [2:08] Okay, I want to go back into what's what [2:10] makes human and agent search so [2:12] different. But before we get there, [2:14] you've told us that, you know, you're [2:16] you're unlearning a lot of the lessons [2:17] that you learned from running Twitter as [2:19] you build build parallel. Why is that? [2:21] Listen, when I was at Twitter in [2:24] leadership roles, Twitter was a post [2:28] product market fit extraordinarily [2:30] scaled business where your feedback [2:34] loops were from the hundreds of millions [2:37] of customers using the product for 30 [2:40] plus minutes every day. Right? In that [2:44] world, you operate differently than a [2:48] pre-product market fit company based on [2:51] the premise that in a few years a new [2:56] customer is going to show up on the [2:58] internet and let's go build technology [3:00] for the not yet here customer that we [3:04] are all learning every day and every [3:06] week. [3:07] >> Fantastic. Okay, let's talk about let's [3:09] talk about parallel. Um, [3:11] >> I have a question actually. Can I can I [3:12] jump in? No, no, you cannot, Andrew. [3:14] >> I'm Andrew. I didn't get an [3:15] introduction, but I'm also happy to be [3:17] an inaugural guest on the Training Data [3:19] podcast. [3:19] >> Andrew is making his podcast debut on [3:21] training data. We are we are honored to [3:23] have you finally here. [3:24] >> Thank you. Um I feel like all of us in [3:27] Silicon Valley have a very surface level [3:30] idea of search like what is search? Like [3:33] the job to be done is, you know, getting [3:35] the answer makes sense. But we all know [3:36] there's crawling and there's an index [3:38] and there's ranking. But maybe let's [3:40] take a big step back and explain what is [3:42] the problem of web search be it for a [3:45] human or an agent and then we can dive [3:47] into the differences. [3:48] >> Um the problem of web search and we all [3:52] know it and experience it. When we want [3:54] to find something and we do not know [3:55] where it is on the web, we go to a [3:58] search engine like Google and the search [4:00] engine then hopefully surfaces the [4:02] answer to us in the most convenient of [4:04] locations. That's the base problem. Now, [4:07] in order to do this, what is the search [4:09] engine doing? The search engine is going [4:11] and crawling the web, which is finding [4:15] every URL that's out there, trying to [4:18] read it, trying to organize that [4:20] information in what might be called an [4:22] index. So that by having all of this [4:26] stuff in memory in one location, you [4:30] don't have to once the query comes in, [4:31] you don't have to spend time loading up [4:33] pages because you already have them. [4:35] you've already done pre-processing to [4:37] organize them and then when the query [4:38] does come in, you quickly understand the [4:40] query. You find the most relevant [4:42] results and then there are many many [4:44] because you're essentially taking [4:45] hundreds of billions of pages and [4:47] narrowing it down to what five or 10 or [4:50] ideally one in terms of what someone is [4:54] looking for right now. So you go through [4:55] these multiple stages of retrieval and [4:58] ranking in order to surface the most [5:01] relevant result. So that's the broad [5:04] problem. One way to think about it is [5:06] it's a billion to billion matching [5:09] problem, right? So you have hundreds of [5:10] billions of pages and hundreds of [5:12] billions of queries over time and you [5:15] need to figure out how to match make [5:16] across these two. [5:17] >> That sounds like an enormously expensive [5:20] infrastructure challenge and I think for [5:22] the longest time basically only Google [5:25] and Bing had you know done full full [5:29] webcale crawls and indices. Why did you [5:31] think about you know you as a young [5:33] company could go off and tackle that [5:34] problem? It seems like a problem of the [5:36] giants. [5:37] >> So it is expensive in the long term. So [5:40] what's actually interesting is that when [5:42] I was starting the company 3 years ago, [5:45] you could imagine a world where one some [5:50] of the reasons it was difficult for [5:51] others to compete which was not having [5:54] access to great feedback in terms of is [5:57] this a better search result than this? [6:00] To collect that feedback at scale, there [6:02] was a problem around human ratings. [6:05] There's a problem around click data. You [6:07] need to access those. Now it is of [6:09] course expensive to crawl the entire web [6:12] and index it. Uh but as you think about [6:15] what agents and the large models that we [6:18] now have access to have enabled is the [6:22] ratings data can be now created by [6:25] experts way more cheaply. Our view at [6:28] parallel is that human click data is a [6:32] bug. An agent [6:35] doing work with search should rely on [6:38] agent feedback, not human feedback. We [6:41] believe that these models are really [6:43] good at [6:45] compressing information and we can [6:47] benefit from a lot of the research that [6:49] have gone into building models and apply [6:52] it to search indexing and ranking. And [6:54] so you can now make many many arguments [6:58] and that's the arguments we made back [6:59] then that actually now it's way more [7:02] tractable as a problem because of the [7:05] existence of agents not just as in [7:07] technology but as a distinct customer [7:11] and then we figured out a way to go [7:13] about building this business which did [7:16] not require us to go spend all of the [7:19] money on infrastructure upfront before [7:22] we could service a customer. Right? So [7:24] if you can paint a path of incrementally [7:26] being able to build a increasingly [7:30] larger and more sophisticated index over [7:32] time as you solve problems for more and [7:36] more customers that's the insight that [7:39] actually convinced me that this was a [7:41] viable problem for us to take on. [7:44] >> And how do you make that happen? Because [7:45] I imagine these this is one of those [7:47] things where customers want full [7:49] coverage day one. So how do you go about [7:51] making that happen? When when we first [7:53] launched the product, right, we launched [7:55] an we did not launch a search product [7:57] first. We launched a search agent [8:00] product first. Our search agent could go [8:05] essentially crawl the web after a query [8:07] arrived. So if you're doing deep [8:08] research, you have patience to the [8:11] extent of a minute. And we've had [8:15] products which sometimes take do 10 [8:16] minutes of research. That's a lot of [8:18] time to be able to crawl a lot of pages [8:20] if only you have enough of a map to know [8:23] what to prioritize crawling, right? So [8:24] you can make up for shortcomings like [8:26] index is often times you can think of it [8:28] as a latency optimization. So if you [8:30] give up on that dimension, if you're [8:32] competing with humans, that's why our [8:34] search agents were competing with the [8:36] alternative being outsourcing to humans [8:40] to curate amazing data. Right? So we [8:43] said it seems like humans sitting on [8:46] search engines are far way easier to [8:49] compete with than a search engine on day [8:53] zero. So by building a product that was [8:56] a search agent to do real work on top of [8:58] web data. We were able to incrementally [9:02] go build our index. [9:03] >> H were there some examples of what [9:05] people built with your search agents? In [9:07] the very early days uh there were search [9:10] agents being built with us for doing [9:14] some kinds of insurance underwriting [9:16] workflows and claims processing [9:18] workflows. People in uh sales were doing [9:22] all kinds of sales data enrichment. [9:25] people in finance uh who would [9:28] previously and historically go [9:32] to build a model, collect data by [9:35] sending it overnight to a set of humans [9:38] who would curate data that would feed [9:40] into a modeling exercise they would do [9:43] would start using us to do that instead. [9:45] And so we were going initially after [9:50] replacing where there was outsourced [9:53] human work on top of web data in order [9:57] essentially to collect evals to run [10:00] agents [10:02] to figure out what search for agents [10:05] should look like in the first place with [10:08] empirical use cases instead of [10:10] theoretical evals. I see you're trading [10:12] off the crawl for inference time compute [10:15] effectively. [10:16] >> Yeah, you guys were started before the [10:19] term like Neolab came into ex existence [10:22] and you have you know web systems [10:24] product but you also have like a bunch [10:26] of AI people from the outside before we [10:29] met for the first time like it wasn't [10:30] obvious to me how much of a web systems [10:32] company or an AI company that you've [10:35] built. the more we spent together, more [10:37] time we spent together, the more clear [10:38] it's like by some definitions this is a [10:40] neolab. Do you want to talk about like [10:42] where the research side of what you're [10:44] doing comes in, not just the like [10:46] infrastructure and system side of what [10:48] you're doing? [10:49] >> So we don't categorize ourselves as a [10:52] new lab. [10:53] >> Well, of course you would not. [10:54] >> We [10:56] no because I think I don't think our [10:58] output is a model. Like I somehow maybe [11:01] my definition is broken. My definition [11:03] of a new lab is an output is a model. I [11:06] think our output is a complement to a [11:08] model. What we build is something that [11:10] multiplies on top of a model in order to [11:15] give either you can call it the model [11:17] gets better or the agent built with the [11:19] model gets better and has superpowers. [11:21] Right? So we've always want to be in a [11:23] place where whenever someone ships a [11:25] better model, a Neol for example, if [11:27] somebody else ships a better model, they [11:29] have a higher hill to climb. for us [11:32] somebody ships a better model they have [11:35] now unlocked four more use cases where [11:37] we can be valuable. Now whether we need [11:42] to do work that might be framed as [11:46] research [11:47] that remains. So in that sense we do [11:50] have to do research but we're also not [11:53] uh pre-training extraordinary large [11:54] models. In fact, our job is to [11:57] compress things down into tiny ranking [12:00] models, right? Like I think if you go [12:01] back to my framing around like this 100 [12:03] billion to 100 billion matching problem, [12:06] every query is essentially give me a [12:09] thousand tokens from [12:13] a trillion web pages on the web and make [12:15] sure they're the right thousand tokens. [12:17] Like that's the query that we are [12:19] getting effectively from to our search [12:21] engine. And so now what you have to do [12:22] and until last week we would give [12:25] ourselves 3 seconds to throw a bunch of [12:26] compute at read time to do it. This week [12:30] last week we shipped a product which now [12:31] does it in 200 millconds. Right? So you [12:33] have now that much time to go figure out [12:36] how to pick those right. So you will do [12:39] we're trying to organize information in [12:42] memory across the memory hierarchy in a [12:45] way that we can access it fast. we're [12:47] trying to figure out how to train what [12:49] model uh to select the BF thousand [12:53] tokens. And so there's a large amount of [12:56] research that goes into figuring out how [12:57] to take a pre-trained model of some [12:59] kind, adapt its architecture for this [13:02] new problem and allocate effectively [13:06] compute in a way that produces the best [13:10] output with a limited compute/ latency [13:14] budget. So a lot of these use cases seem [13:16] like the deep research kind of shaped [13:18] use case. And when I think deep [13:19] research, at least in the earlier [13:20] incarnations, it was effectively like an [13:23] agentic loop with the model reasoning [13:25] and then basically just calling a search [13:27] engine calling calling Google or some [13:29] proxy to Google. Yeah. [13:31] >> Right. Why is that insufficient in your [13:33] eyes versus what I'm hearing from you is [13:35] agentic search is kind of a a net new [13:38] capability. So [13:41] calling [13:42] Google for every query in a deep [13:45] research thing, it gets you [13:47] >> somewhere, right? [13:49] >> Okay. [13:49] >> If you use parallel search, [13:53] >> you will for the most part use under [13:57] half the tokens in your agent, it will [14:00] become more accurate and to be faster [14:03] end to end. [14:05] And so you can think of it as but every [14:08] time if you use only half the tokens if [14:10] your model is context limited or memory [14:14] limited you can now do [14:17] more problems you can do the same [14:20] problems cheaper or faster. So [14:22] everything to me when you have a [14:26] infinite appetite for information and [14:29] relevant information for all kinds of [14:31] work it's at its core an optimization [14:34] problem around quality cost and latency. [14:38] Every model advancement is about how do [14:40] you squeeze out more intelligence and [14:43] then how do you detail it down into keep [14:46] most of it at a tenth of the cost. [14:48] Right? And that applies to search as [14:52] well. And every time you can produce all [14:55] of the signal with less noise in your [14:57] search results to give to a model, [15:00] you now let give the model the ability [15:02] to do more. And so a lot of the bet here [15:05] is intelligent compute allocation across [15:09] the model layer and the agent layer down [15:11] to the search layer. The other [15:12] interesting thing to observe is the [15:15] interface actually changes uh when you [15:18] think about serving agents versus [15:21] humans. Humans rely on keyword search. [15:24] With agents, we've had to innovate quite [15:26] a lot on the what does an agent tell our [15:30] search engine. The more we know, the [15:32] better we can do. [15:33] >> Interesting. So like humans can only [15:35] hold a few words in in memory basically [15:37] versus an agent. [15:38] >> No, I think we can. We're just lazy. [15:40] >> Yeah. [15:41] >> Right. like we we can we can have a [15:44] conversation. We can hold a lot of [15:45] words, but when you start typing, we we [15:48] want to tell Google like two incomplete [15:50] words with a typo in there and hope for [15:52] the best or rely on some form of a drop [15:56] down autocomplete thing to avoid typing [16:00] those three words. [16:01] >> Yeah. [16:01] >> In the first place. So, we're [16:03] fundamentally lazy. Turns out like [16:05] models less. So fewer typos, better [16:09] specified queries, perhaps longer [16:11] queries, less for the search engine to [16:14] guess what the agent might want. Uh so [16:18] you get to solve a different class of [16:19] problems. One other thing about humans [16:21] is I feel like we know how to decipher [16:26] the, you know, pre-ai slop that [16:29] populates many of the top ranking Google [16:32] results, you know, for best product for [16:35] X, Y, and Z. and you get all the [16:36] affiliate advertising things. Um, and [16:39] yet obviously they rank highly. [16:41] >> I'm very susceptible to it. [16:42] >> Are you? Yeah. I'm like I feel like I [16:44] have this incredible I get to see coming [16:46] a mile away. How do agents deal with [16:48] things like that, right? Where these [16:49] are, you know, well traffic pages. They [16:51] have seemingly good answers, but you [16:53] just know they're not real. [16:56] >> Perhaps with parallel, agents have to [16:57] deal a little bit less with that. [16:59] >> Let me tell you why those pages exist in [17:03] the first place. Right. Let's work [17:05] through a simple example. [17:07] If you ask for a public company's most [17:11] late most recent financials like just [17:14] the headline revenue number, we can sit [17:16] here and know that there exists an [17:17] authoritative filing with the SEC which [17:20] has that number perhaps on page 73 of a [17:24] PDF. [17:26] But that is the authoritative number. [17:28] Right now imagine Google decided like I [17:31] care about authority and whenever you [17:34] ask that query that's the first result [17:36] you see right and then you click that [17:39] this PDF takes what like 3 and 1/2 [17:42] seconds to load instead of the second [17:44] that the another page does you're [17:46] already frustrated and then you see a [17:49] earnings page [17:52] >> that's the alternative but here you have [17:54] to now grab your way to page 77 to find [17:56] your answer. [17:58] >> We're lazy. We're not going to do the [18:00] work. Right? So now there exists an [18:03] entire class of content on the web which [18:06] is like okay this information is needed [18:10] by a lot of people. It is worth putting [18:14] it on a page that loads fast where this [18:16] information is above the fold. All of [18:19] you can go there. It's 99.99% [18:23] right. So you're not like so skeptical [18:25] when you go there that this is 100% [18:28] wrong. And it's added real value in the [18:31] process because it identified out of the [18:34] 300page [18:36] earnings report 20 bits of information [18:39] that should be above the fold. [18:41] >> Yeah. [18:42] >> Right. And so you can call it slop, pre- [18:45] AI human slop, or you can call it [18:49] catering to a lazy human and being [18:51] successful at SEO, right? The good news [18:54] is with agents, we're not making the [18:56] agent click around and fumble around and [18:59] crap a PDF, right? [snorts] [19:00] We're taking a excerpt from the most [19:04] authoritative place on the web and [19:06] trying to bring it to the agent's [19:08] context window. And so we aren't forced [19:12] into this weird trade-off. And this [19:16] trade-off existed with humans in the [19:17] first place because like with with [19:18] standard browsers and protocols and [19:20] everything, we just didn't figure out [19:22] how to have us point exactly [19:26] consistently across everywhere on the [19:28] web to the exact right highlighted tiny [19:32] paragraph. Right. With agents, we get to [19:35] bring that to the agents context and [19:37] then let it figure out what's next. Can [19:38] you walk us through what actually [19:40] happens when when your developer sends a [19:42] query off to your agentic search API? [19:45] >> It's in some sense pretty standard. So [19:48] we run some models to figure out what [19:50] this query is and enrich it to figure [19:51] out how it will flow into the rest of [19:53] our system. We have a bunch of indexes [19:56] which organize different subsets of the [19:58] in of the web in different ways. And so [20:02] the first layer will essentially craft [20:04] queries for each of these different [20:06] systems. Each one of these systems [20:09] perhaps is our sort of big index. One of [20:11] these is perhaps our fresh index. One of [20:13] these is perhaps like you can some [20:15] people will describe it like a knowledge [20:17] graph. Some people will describe it like [20:18] a structured index. Like there are a [20:22] bunch of these, right? So you're now [20:25] deciding which ones this query needs to [20:26] go to. You're figuring out what is the [20:29] query rewrite for each of these. Then [20:31] each of these has like a big retrieval [20:33] layer and then a ranking layer and then [20:36] more ranking layers. So, you're going to [20:38] try to boil down tens, hundreds of [20:40] billions of URLs or documents [20:44] uh down into [20:47] thousands, tens of thousands down into [20:50] specific excerpts and paragraphs in [20:51] those tens of thousands with more and [20:54] more bigger models running at each stage [20:58] with different architectures [21:00] pulling more features to ultimately get [21:03] down to here are the thousand tokens I [21:05] want to bring back to this AI. I which [21:08] has the highest signal, [21:09] >> right? And now if you look at our [21:11] various versions of our search API, they [21:14] just throw for different latency and [21:17] cost constraints, different amounts of [21:20] compute [21:22] at various points in this journey to hit [21:25] those limits. [21:27] So in the abstract, it's [21:29] simple, right? What's interesting is [21:33] what each of these models that I [21:34] described, how you curate and collect [21:36] the training data for them, right? How [21:40] you build those models and optimize [21:43] them. The index itself, how you use the [21:46] the memory hierarchy to store it to be [21:49] able to hit certain [21:52] cost, quality, latency thresholds. [21:54] >> What's the north star from a quality [21:56] perspective? like you know in Google [21:57] there's that you know did you get the [21:59] result you wanted in the first three [22:00] answers or something like what is the [22:02] equivalent north star for you [22:03] >> so I think of I don't know if I'm right [22:06] on this but like my take is that a [22:09] billion to billion matching problem is a [22:10] forever problem and so the real question [22:14] is at what point incremental [22:16] optimization isn't worth the squeeze [22:20] right so I don't think there is a thing [22:22] as like okay we're done on improving [22:26] this [22:27] The question is at some point it's going [22:28] to get harder and harder to improve this [22:30] thing and it just won't be worth it, [22:35] right? But I'm hoping that we don't get [22:39] there actually because like if you think [22:41] of what we're doing with AI, we will [22:43] have more intelligence that gets cheaper [22:47] every few months [22:49] and as a result of it, it'll come down [22:52] to [22:53] having great models which are cheap, [22:57] having great information, [23:00] matchmaking across a need, and all of [23:02] the information available to you, [23:03] whether it's your own or on the [23:06] and doing something unique and [23:08] differentiated with it to produce more [23:10] knowledge, right? And anytime you can do [23:13] something 20% better than somebody else, [23:16] that might give you an edge. So why [23:18] wouldn't you, [23:20] right? So in the super AGI build [23:24] worldview, it feels like if you can push [23:28] on quality across web search and in the [23:31] model layer, why wouldn't you? I've [23:34] heard a point of view that this is so [23:36] fundamental to the model companies that [23:38] they're just going to own it and and in [23:40] part because as they're collecting data [23:43] for pre-training that's a very very [23:44] expensive infrastructure exercise that [23:47] kind of is your source of truth for the [23:48] for the crawl. Um what do you think of [23:51] that? [23:52] >> I don't see empiric data on the ground [23:54] to support that view to build a fresh [23:57] web index. [23:59] I don't think that crawl is particularly [24:02] useful [24:04] and let me frame why. So if you think of [24:07] what we are building, we're building a [24:09] complement to models. We like to crawl [24:12] things that [24:15] people don't like to crawl for [24:17] pre-training because like if the model [24:19] already was trained on it, like it's not [24:22] useful. Model companies for the training [24:26] aren't patient enough to go in a [24:29] completionist way try to wait for really [24:32] slow random JavaScript to load because [24:36] the number of tokens you get per amount [24:38] of compute you throw at it is like one [24:41] order of magnitude so two order of [24:43] magnitude too bad and so like is it [24:46] worth the extra effort to get these [24:48] tokens for us it is because we're [24:50] completionist right for them it's [24:53] I'll take x trillion tokens. So that's [24:55] one. Now I do think this is a core part [25:00] of every agent, right? I my worldview is [25:03] that if you're buying LLMs [25:07] for doing work [25:09] for [25:11] nine out of 10 use cases, you will want [25:15] them to have access to the web and [25:18] create search infra optimizing for [25:21] agents. So it is a real adjacency [25:25] for all kinds of LLM inference and that [25:30] supports your view that model companies [25:34] could should have the best-in-class [25:36] infra for it. So now they can build it [25:40] or they can buy it and that's the [25:42] conversation and the we will see who [25:45] builds and who buys and who partners and [25:47] how things evolve. Are you partnering [25:49] with any of the model companies that you [25:51] can share? [25:52] >> We want to we will I can't share [25:55] anything on that. We did announce and I [25:58] don't know how your definitions of model [26:00] companies uh we announced today actually [26:03] that we are working with uh Google cloud [26:08] to be a search and grounding provider [26:12] for their enterprise agent APIs. So if [26:15] you think of grounding Gemini models or [26:18] other models available on GCP when you [26:20] build agents on GCP or chat apps on GCP [26:23] or do any other inference with LLMs on [26:25] GCP uh when you attach web search to it [26:30] uh your options [26:32] are Google search or parallel search and [26:36] parallel search is product integrated [26:39] the integration is optimized we've spent [26:42] time with technical teams and training [26:44] teams and product teams and commercial [26:46] teams to make sure that when people use [26:50] Gemini models with parallel like they [26:53] get exceptional and great results. So [26:55] yes, there are these partnerships now [27:00] emerging. Uh I bet that there'll be [27:04] several of these. They will all look [27:06] somewhat unique. [27:08] >> That's a big deal from the search king. [27:09] Congratulations. [27:10] >> Google's the original Neolab. I I don't [27:13] know how to frame Google as a model lab [27:15] versus a hyperscaler and I don't know [27:18] what precise lessons to learn from this [27:20] one whether it applies to other labs or [27:23] to hyperscalers [27:27] and so we will see the way that I use [27:31] voice agents and I need to get [27:33] reservation for dinner at night what [27:34] restaurant should I go to and then that [27:36] request gets fulfilled by the agent. [27:39] When we talk to a lot of the parallel [27:41] customers, there's this, you know, [27:44] background agent, whether the monitor [27:45] product, these agents that sort of are [27:47] always watching the world or watching [27:49] the web and when something happens, then [27:50] they go off and take actions and do [27:52] something with it. Um, it might be worth [27:55] because if we think about what does a [27:56] thousandx more mean, there's like the [27:58] depth of research and then there's just [27:59] like the the you know, what is actually [28:01] initiating the tasks? Is it a human [28:03] initiating the search or is it agent [28:05] itself? You want to talk about that [28:06] dynamic a little bit? [28:07] >> Yeah. So there's a bunch of dimensions [28:10] here. Let's go back to for a moment on [28:12] search agents, right? If you run a [28:15] typical search agent and even without [28:17] doing deep research, it'll it'll do [28:20] somewhere between 5 to 20 searches even [28:23] if it answers within a few seconds [28:25] because why not, right? So already if [28:29] you transition from using a [28:33] chat GPT medium or and all the [28:38] categories but like somewhere not on the [28:40] high tier like instant low medium every [28:43] time you write a prompt to it in chat [28:45] GBT it will do five to 10 searches as [28:48] you dial it up it'll do hundreds and [28:51] thousands of searches. So one [28:53] interesting thing to observe is like a [28:55] human action to a multiplier on number [28:58] of searches that happened right so just [29:00] by [29:02] using a AI app you're kind of [29:06] multiplying your way to perhaps one [29:09] order of magnitude more searches now a [29:13] lot of our initial takes on the product [29:16] and the market were to go after bigger [29:20] multipliers than even that right so we [29:23] were much more interested when you said [29:25] I have a [29:27] portfolio of [29:30] 10,000 small businesses where I have [29:32] given out credit to for all of them [29:36] every month I have this human process [29:39] that runs [29:40] to feel out how my risk is going up or [29:43] down now can we and it relies on a bunch [29:46] of web data we're trying to use agents [29:48] for doing this so here a developer is [29:50] effectively [29:52] The multiplier there is hundreds of [29:55] thousands or a million in terms of the [29:57] number of web searches that happen [29:59] because a human goes and programs that [30:00] instead of now doing this process every [30:02] month, we can do it every week, right? [30:05] So it's doing a lot of searches when [30:08] this agent runs every week to create a [30:11] dashboard on the portfolio and a [30:12] collection of action items that somebody [30:14] needs to look at. Right? Um, now you can [30:18] go to another example which is even more [30:20] interesting. I don't know. Do you use [30:22] something to do uh meeting prep [30:26] documents for you all? That's an agent. [30:28] >> There's a Sequoia one. There's a Sequoia [30:30] agent. [30:30] >> Yeah. Yeah. I was I was going to say [30:32] James Flynn, but [laughter] [30:35] he's one of our great young guys. So I [30:37] use um notions uh agent and you can [30:42] build custom agents which look into all [30:44] the internal data that [30:47] you we have at parallel plus all of the [30:50] web data using parallels APIs to create [30:54] meeting prep docs. [snorts] One time I [30:56] went and created one prompt to build [30:58] this custom agent. Now it does tens and [31:02] hundreds of web searches for every [31:03] meeting I have. Every time I build a new [31:06] agent for a new use case, that keeps [31:09] multiplying. So I think the path to [31:12] these background agents doing more and [31:14] more and more work all the time for us. [31:18] It's only going to be bounded by value [31:21] versus spend. [31:22] >> You know, like I don't think it's [31:23] rational right now. Like I'm probably [31:25] spending more on it than I should be, [31:28] but it's not too much. So I don't care [31:30] yet. So there'll be some rationalization [31:32] in all of these agents. Um but I think [31:35] we will deploy background agents to the [31:38] extent that there is incremental value [31:40] in doing that compute and the same thing [31:43] applies to searches. All of these [31:44] background agents would do a bunch of [31:45] searches. [31:46] >> Mhm. And so our first set of products [31:48] were really obsessively focused on these [31:51] in part because we were building growing [31:53] the index and we decided that our [31:55] company was based on three dimensions [31:58] quality cost latency and for the first [32:02] couple of years we said let's focus [32:05] let's ignore latency and let's just nail [32:08] the other two because optimizing systems [32:11] distilling to smaller models is much [32:15] more known art than unknown research. [32:20] >> Right? So once we achieved [32:23] the best quality search and search agent [32:27] products at every price point, we [32:30] started working on latency and that's [32:32] what we shipped with a product we call [32:34] Turbo. Now it is the fastest, highest [32:38] quality agentic web search on the market [32:41] by a lot. Do you think there are more [32:44] agentic queries than human queries on [32:46] the web now? [32:47] >> I don't think yet. I don't think we [32:49] >> just given some of the multipliers you [32:50] mentioned and given the background [32:51] agents it seems. [32:53] >> Yeah. I I think I don't have to remind [32:55] you but we are early like we are very [32:58] early in agent adoption like you go step [33:00] outside of our bubble people haven't [33:02] heard the word fable like so we are very [33:07] very early. I think there are now people [33:10] like me who are probably operating at [33:13] the [33:15] thousandx. [33:17] I don't know. What do you think? How [33:19] many Google searches a day did you do 3 [33:22] years ago before charg [33:24] 30? [33:24] >> Yeah, I would have guessed something [33:25] like 20 30. I think today if you just [33:29] look across all of my agents like I I [33:33] bet they're doing thousandx more than [33:36] that. [33:38] maybe 100 to a thousand. [33:42] Um, if you count some of the things that [33:44] happen at my company, which isn't like [33:47] assigned to a human. [33:48] >> Yeah, [33:50] >> it might easily be more than a [33:51] thousandx. But I think we are the [33:53] outliers rather than the norm. [33:57] >> So I think we're very very early on this [34:00] journey. I I do think I recently saw I [34:03] think it was Cloudflare that said that [34:05] in their monitoring of web traffic [34:08] >> the AI traffic is about the same as [34:10] human traffic [34:12] >> in terms of [34:14] >> page reads which is slightly different [34:16] from searches because it includes [34:19] perhaps all crawlers that are out there [34:21] and a bunch of other stuff but I think [34:24] it's it's going to happen. So maybe this [34:28] is a good segue to talk about a topic [34:30] that I know you are passionate about [34:32] which is the economics of the internet [34:36] as we know them. Some of the kind of [34:37] fundamental assumptions there you know [34:40] human eyeballs scarcity of attention [34:42] seem to be falling right in front of us [34:44] right now. Are the economics of the [34:46] internet broken now and what's going to [34:47] happen? [34:48] >> Yeah. Yeah. No, I think this was perhaps [34:50] part of the thing that totally convinced [34:53] me that this needed to be done by me and [34:55] build parallel. Uh I know most people [34:58] hate ads. I used to do ads and build [35:01] systems for ads. [35:02] >> I love ads. [35:03] >> Oh wow. Huh. [35:04] >> I love shopping. [35:05] >> Okay. Yeah. So you get good ads. I don't [35:08] know if I love ads, but I intellectually [35:12] love ads because ads is the reason that [35:15] so much amazing content and technology [35:20] is available for free to all of us, [35:23] right? Like Google search wouldn't be [35:24] free without ads. Twitter wouldn't be [35:26] free without ads. And these are truly [35:29] useful pieces of technology. A lot of [35:32] content on the out on the web out there [35:34] that we can access for free wouldn't be [35:37] free if not for ads. Um, and so ads is a [35:42] very efficient monetization scheme. It's [35:44] a very efficient monetization scheme [35:47] because it is exceptional at [35:49] differential pricing, right? So most [35:53] queries Google loses money on. Some of [35:56] them make it up and it's an [35:57] extraordinary business with [35:58] extraordinary margins. Same with [36:00] Twitter. like most users. [36:03] >> Exactly. [36:04] >> You're welcome. I'm subsidizing the free [36:06] the free information you guys are [36:07] getting. [36:07] >> You are subsidizing [laughter] all of [36:09] us. Thank you. Um but I think ads is [36:12] extraordinarily efficient at [36:14] differential pricing and monetization on [36:16] the web which is why it has been a [36:18] dominant business model. Now the core [36:21] assumptions as you noted going into it [36:23] around [36:24] >> limited human attention to translate uh [36:28] into outcomes. if humans don't show up [36:31] and their agents show up on the web like [36:33] what does this mean? How does the [36:37] business work? Um, and yeah, I think [36:40] this was the if we don't figure out a [36:43] new business model, we're seeing it [36:45] already, right? Like people are going to [36:46] say, okay, I don't want my content to be [36:49] accessed by an agent because I have a [36:51] business model for humans. So I actually [36:53] want to go [36:55] pay someone to SEO optimize myself so [36:58] more humans show up but then if their [37:00] agent shows up I'm going to cut it off [37:02] right which seems confusing and [37:05] disconcorted like okay ultimately this [37:08] agent is acting on behalf of a human but [37:11] we haven't found business model [37:14] alignment [37:15] >> in a way [37:17] >> you can't monetize that visit [37:18] >> you can't monetize that visit [37:20] >> so let's say you are in the business of [37:22] subscriptions, right? So, you get a [37:24] bunch of you get a thousand humans and [37:25] you convert 20 of them into monthly [37:28] subscriptions, right? You don't know how [37:30] that loop works yet with agents. You [37:35] can't distinguish. You don't have [37:37] statistics. You don't know if these will [37:38] lead to subscriptions or they'll just [37:39] keep stealing your content as nameless [37:42] agents. And so there are real challenges [37:44] around the old business models breaking [37:47] and us not figuring out real scalable [37:49] new business models. Right? So if you [37:51] own highquality content and you [37:53] currently put it on the web, [37:56] your one available business model [37:59] pre-parallel was to be in the head and [38:03] be able to transact with the lab on some [38:06] fixed fee contract. That's literally [38:09] which includes some amount of training [38:11] and liability and then [38:14] inference time access right that was [38:16] your one option that option is not [38:18] available to [38:20] most content on the web it's a very head [38:24] phenomena and even for the head it is a [38:29] broken [38:30] business model because when AI's or [38:35] inference grows let's say 7x this year [38:38] and another another 7x the next year. On [38:42] this 50x, their deal size is not growing [38:45] 50x. Like none of them after signing a [38:47] 2-year deal believes that their share [38:50] isn't going to decline materially [38:54] at renewal. And so these are like fixed [38:57] price constructs in the world of AI [38:59] inference. And so which doesn't drive [39:02] for sustainability for for all of these [39:04] businesses. Now our solution is trying [39:09] to learn all the lessons from my work at [39:12] ads on having transacted like I sold [39:15] Twitter data to open AI having [39:18] transacted on that side uh to figure out [39:21] what actually might work and be [39:23] incentive aligned. So what might work? [39:26] >> Efficient differential pricing. [39:29] Paying for differentially a lot for [39:32] extraordinarily high value content. [39:34] Paying a lot for [39:37] high value work accessing the same [39:39] content. So it's differential on both [39:41] dimensions. Quality and value of work [39:43] being done with it. Uh and a way of [39:48] doing this scalably and not just with [39:50] bespoke deals. So those are the [39:53] properties needed for any reasonable [39:56] solution and the biggest property of it [39:59] all is incentive alignment right like at [40:01] what point do [40:04] people want to collaborate into this [40:06] enterprise? [40:07] >> Why do the model companies need to pay [40:09] anything at all? [40:10] >> Um you should ask them. Um my [40:13] understanding [40:15] is one you want [40:19] fresh data during inference time to be [40:21] able to display it in products like [40:23] chargot. Two you want training data and [40:28] three you want some liability protection [40:30] for training that you already did. [40:34] And so the payments are some combination [40:40] of these three things. And I can't be [40:44] sure of like how they value each of [40:47] these three. [40:48] >> You said, you know, even a very simple [40:50] query will go do 10 searches. And so how [40:53] do you do attribution between all the [40:54] different sources that you know boil [40:55] down to one paragraph response? [40:58] >> Yeah, at parallel we like to build [40:59] models. [41:01] >> I heard. [laughter] [41:02] So uh no I think the the it's a it's a [41:05] it goes back to my point around [41:07] incentive alignment. So what are you [41:08] let's before we try to build a model [41:10] let's try to figure out like if you were [41:11] going to try to do this intellectually [41:13] theoretically how do we go do it. So you [41:16] have to ask the question of okay how [41:18] much incremental value did somebody's [41:20] content add right so you can you can run [41:25] all of these simulation exercises you [41:27] take one piece of content out of the [41:31] corpus and then you say let's run the [41:32] agent let's see if the quality of our [41:34] outputs declined how much to claw back [41:38] that quality perhaps if I threw a little [41:40] bit more compute or a better model in [41:42] some way could I claw back that quality [41:46] Oh, I could. It cost me a scent. [41:50] Huh? I could get the quality I lost by [41:54] not having this source. My alternative [41:57] was to throw more compute, [41:59] a scent worth of compute to get that [42:02] same quality. It might be a different [42:04] result, but as far as the end [42:06] [clears throat] customer value is [42:07] concerned, it's the same quality. And so [42:09] you're like, okay, this source is worth [42:11] close to a scent. Feels like, right, [42:13] that's intuitive. Now the formalization [42:16] of this kind of an intuition [42:18] is the the the core framework we use [42:21] it's called shappley values. Uh [42:23] >> yes let's go. What's a sharply value? [42:26] >> It's a [42:26] >> music's my ears. [42:27] >> It's a game theoretic. You're you're a [42:29] game theorist. [42:30] >> A bunch of game theory. Yeah. [42:31] >> Yes. [laughter] Shley values is this [42:33] very theoretical mathematical concept. [42:36] It is used in sharap values in feature [42:38] importances for those who are ML folks [42:41] here. Let's say the three of us [42:43] collaborate on something and the whole [42:46] is bigger than the sum of parts uh in [42:49] that moment. The theoretical question is [42:51] okay how do I divide up this sort of [42:54] bigger pie that we created by [42:56] collaboration. So all three of us have [42:58] incentive to collaborate. Right? And [43:01] Sharappley values is a mathematical way [43:03] of effectively answering this question. [43:06] Right? Now that sounds amazing, right? [43:08] Like you could if you're co-founders, [43:10] you could figure out how to divide [43:11] equity. [43:13] Unfortunately, it's not that useful. [43:15] It's not that useful because in order to [43:17] compute Sharley value, you need to [43:19] simulate all words where [43:22] some subset of the two of us [43:24] collaborated but the third one doesn't. [43:26] and play out those realities to then [43:28] impute back to today in terms of how we [43:31] should divide the pie which in practice [43:33] you can't do most places right in ML [43:37] models when you do feature importances [43:38] you kind of can you can hold a feature [43:40] back run your model and see how well it [43:42] did in web search we can run [43:47] simulations of if we did not have access [43:49] to this URL or this domain or this [43:51] collection of them how would the agent [43:54] perform we can if you're if you're good [43:55] at eval else. If you're good at [43:56] assessing quality, you can build that [43:59] data by running various scenarios, [44:02] collect a bunch of this data and then [44:04] you can train models to [44:07] >> your favorite thing. [44:08] >> Favorite thing. So the challenge with [44:10] Sharappley values is computing Sharapply [44:12] values in our context [44:14] is way more expensive than the amount of [44:17] dollars [44:19] we spend on the agent. Forget the amount [44:20] of dollars we want to pay a publisher. [44:22] Right? So to compute a content owner to [44:24] compute that a content owner gets a [44:26] dollar. If I decide to do the full [44:28] sharply value computation that might [44:31] take several dollars, right? So it [44:33] doesn't make any sense except you can [44:37] really well estimate them if you build [44:39] the right kind of data and models around [44:41] it. But we have confidence that our [44:44] estimations are good and that this is [44:49] sound theoretically that if there was [44:52] perfect information symmetry, people [44:55] would want to collaborate. So the same [44:57] way in the old ads days, people did [45:00] second price auctions and believed for [45:02] better or worse that people would reveal [45:05] their true bids in in an ad auction and [45:09] end up paying less than that. [45:11] uh once feedback loops get established [45:14] in a market like today everything is [45:15] like an auto bit right like most people [45:17] are measuring ROI when they're doing [45:19] advertising and running on auto bit [45:21] instead of making a bits uh I think a [45:24] solid foundation based on incentive [45:28] alignment that chaplat drives will [45:31] ultimately maximize participation [45:34] of content owners uh into this as well [45:39] as content seek speakers via AI in an [45:42] optimized system and we have some [45:46] positive evidence to support it. Like [45:48] we've had some interesting partnerships [45:49] that we've been able to do and announce [45:52] uh and yeah like imagine sitting with [45:56] content owners and explaining Sharley [45:58] math takes a moment. Uh but ultimately [46:01] once once you pull out the properties [46:04] right that you participate in the value [46:07] if you have unique differentiated data [46:10] you get paid more. If [46:13] a banker in an expensive job reads your [46:17] data versus my retired dad reads your [46:19] data, the banker ends up paying more for [46:23] that read, right? because it's a part of [46:25] high value work and [46:28] the the macromath also seems to work [46:31] like if you're going to spend a lot of [46:33] money on inference on LLMs for knowledge [46:36] work [46:37] if we allocate it like what 2 to 10% of [46:41] it to data on the web that's way bigger [46:47] than all web data business models today [46:54] outside of like walled gardens like [46:57] Facebook and LinkedIn like so the [47:00] macromath supports it. It is a scalable [47:03] approach and as [47:06] agents on the web grow an order of [47:09] magnitude year on year. It's like by my [47:12] calculations like we're 12 to 24 months [47:16] from this math being able to give [47:21] meaningful dollars for a very wide range [47:24] of content owners on the web. [47:27] >> That's very exciting especially at a [47:29] time when you know you see the traffic [47:32] data the stack overflow plummeting. You [47:35] see a lot of the human internet as we [47:36] know it going away because of [47:38] incentives. It's very exciting to see [47:41] how you're thinking about incentive [47:42] alignment for people to keep publishing. [47:44] >> Yeah, it's it's why we started the [47:46] company. [47:46] >> Why is a company called Parallel? [47:49] [gasps] [47:50] >> The company's original name was Sharley [47:54] Inc. [47:55] >> Really? [47:56] >> It was when [47:57] >> Did you know this? [47:58] >> I knew that. Yeah. [47:59] >> When I incorporated, [48:01] I'm telling you, I was obsessing about [48:04] everything to do with the problem space. [48:06] So while knowing that the first set of [48:08] also one sharing is a terrible name for [48:12] a B2B product, it was not going to be [48:15] the long-term name. So we we in but I [48:18] incorporated as Shappley Inc. for lack [48:20] of a better word and Shappley.ai [48:22] happened to be [48:24] available. Shappley.com is a parked pawn [48:29] domain. So [48:32] it was not going to be a name. I went [48:34] around not talking about my company as [48:36] Shappley and my badges at all events [48:39] used to call it NCO or stealth company. [48:44] So and it took us almost six 8 months to [48:48] figure out what the real name of the [48:49] company will be. We ended up at parallel [48:52] in part because it like one at the time [48:55] we're doing a lot more in parallel and [48:57] two we started visualizing this sort of [49:00] a parallel web built for AIS and how its [49:04] properties are different and this [49:07] metaphor that when you publish you're [49:09] now thinking of as we all are now of two [49:13] audiences okay I'm going to create a [49:14] page [49:16] I know humans will read it what should [49:19] it look like to them And then how should [49:21] I make sure that agents can read it too? [49:25] So it feels like you're dual publishing [49:27] to two audiences. And so we had this [49:29] like this parallel web for agents will [49:31] emerge. And so that's why we started [49:34] liking parallel. [49:35] >> Love it. [49:36] >> I actually was thinking about this [49:37] recently with the um you know things [49:39] like earning transcripts, right? Like I [49:41] feel very confident that more people are [49:44] consuming earnings transcripts through [49:47] agents than are actually certainly [49:49] listening to the audio and honestly [49:51] probably been reading the transcript [49:52] itself. So there are these like very [49:54] high value portions of the web that have [49:56] flipped already to being agent first and [49:59] obviously the communication has not yet [50:01] flipped. Um, but I think if I were a [50:03] public company CEO today and I was doing [50:06] an earnings report, I would make very [50:09] clear that the thing that I'm saying [50:11] will be transcribed and interpreted [50:13] correctly by the agents, not just by the [50:15] people listening. Theoretically, [50:17] >> yeah, same for us when we're publishing [50:20] docs for our APIs. Like, [50:22] >> yeah, [50:23] >> it's like our customers are building AI [50:26] solutions. [50:28] They are using AI to do it. It's their [50:31] agents reading our docs [50:35] and code in our FDKs. It's not humans [50:39] fumbling around docs pages for the most [50:41] part. So in fact for us the primary [50:43] audience is a agent and that's how we [50:46] test our docs. [50:47] >> A parallel web for agents. It's very [50:49] cool. [50:50] >> Um maybe close us out. tell us just give [50:52] us a snapshot of where where parallel is [50:54] today and if everything that you hope to [50:58] build comes true what does parallel look [51:01] like what does the world look like [51:02] what's your role in it [51:03] >> I think of three levels in this journey [51:07] level one is people are building simple [51:09] agents that use [51:12] the web more as like a tool like a web [51:15] search tool because it's familiar [51:17] because the first set of agents we built [51:19] to your point was just like model and [51:20] give it the same tools Because these [51:22] models have been trained to use all the [51:24] tools human have been humans have been [51:26] used to and let them do work. [51:29] By and large if you think of like most [51:31] work it's there today of or inference [51:33] it's there today. We have now started [51:36] seeing some subset of customers who are [51:38] in this world where they're building [51:41] more sophisticated [51:44] multi- aent systems which use sub aents [51:47] and have agents wake each other up or [51:50] archive crate uh in interesting ways [51:53] right people are seeing that as sub [51:54] aents most familiar one is sub aents [51:56] with encoding agent harnesses but we see [51:59] a lot more of that like if you're [52:00] building an AI scientist uh some of [52:03] those systems are very interesting, very [52:06] sophisticated, very longunning [52:09] and like just throw large amounts of [52:12] compute and data at a really hard [52:14] problem. Uh and I think the third layer [52:18] for the web specifically is the web goes [52:22] from pull to push. So today by and large [52:25] across the first two modalities I [52:27] described either an agent is calling a [52:30] tool or a sub agent but it is telling it [52:33] a request saying go find this for me [52:37] right now and I think where we will end [52:40] up in a couple of years is a variety of [52:44] use cases will be [52:47] the web parallel [52:49] call me if this happens so that my agent [52:55] can do some work or a human can do some [52:57] work. So I use this line with my team [53:00] all the time which is like imagine like [53:03] agents are everywhere and they can do a [53:04] lot of things right if there is [53:07] something you can do today and it's [53:09] worth doing [53:11] and we still have GPUs available [53:15] we'll just go do it right like we won't [53:17] like we won't be like oh let's just do [53:18] this tomorrow for the sake of it if it [53:21] can be done today do it so what will we [53:23] do tomorrow we will do tomorrow work in [53:27] response [53:28] to something that circles through either [53:30] another agent's work or something [53:33] changes in the world as visible in [53:36] satellite imagery or some customer [53:40] commentary that happens. Some agent [53:41] finishing some compute, a human having a [53:44] new insight to trigger work. But there [53:45] are going to be a few feeds like this [53:48] which will drive new agentic work [53:52] tomorrow. And one of those feeds is [53:55] going to be everything that changed on [53:57] the web, which is really exciting for me [54:00] because then you're framing not a point [54:04] in time need [54:06] a long-term [54:08] here is what is actionable for me. If [54:11] something like this happens as evidenced [54:13] by all of the information on the web [54:16] that is changing all the time, call me, [54:19] right? And then I'll run my agent on it. [54:21] And so we get to then allocate compute [54:25] onto the entire web all the time on [54:29] behalf of all the customers and that's [54:31] really exciting. [54:32] >> That's awesome. Really ambitious vision. [54:36] You're clearly extremely passionate [54:37] about about this and exciting to see you [54:39] building. Thank you so much for joining [54:41] us today, Prague. And thank you for [54:42] joining us, Andrew. [54:43] >> You're very welcome, Sonia. Thanks. [54:45] >> Thank you, Andrew, for your debut. [54:49] >> [music] [55:06] [music] [55:12] [music]