Exa Bets GPU Cluster on Search Built Exclusively for AI Agents

Exa押注GPU集群打造AI Agent专用搜索 传统广告驱动的搜索引擎无法满足AI系统需求，随着AI搜索量预计将激增至人类使用量的1000倍，Exa正着手构建专用查询基础设施。 Will Bryk阐述了为何传统广告驱动型搜索引擎无法满足现代AI的需求，指出AI需要的是海量、高密度且无偏见的上下文，而非蓝色链接和广告。随着自动化AI搜索量预计将远超人类搜索量，Exa专门针对AI Agent构建了网络级专用搜索引擎，索引了数百亿份权威文档。Bryk解析了AI原生搜索的五大核心支柱——包括低延迟、Token密度优化和外部数据市场集成，同时解释了动态网络检索为何是突破LLM参数化记忆限制的必备手段。最后，他将完美的信息检索定位为应对即将到来的社会变革与技术转型的关键支柱。 Exa 成立于 2021 年，致力于为 AI 构建完美的信息引擎，目前已服务超 5,000 家企业与 40 万名开发者；传统搜索引擎受广告驱动，无法满足 AI 对客观中立上下文的需求。 (0:04) AI 搜索调用量正在呈指数级增长，预计将在 2026 年底前超越人类搜索总量并在数年内达其 1,000 倍，促使搜索商业模式转向以数据准确度为导向的 API 按量计费机制。 (3:06) 前沿大模型参数容量（约 10TB）无法存储数以亿计 TB 的全球动态网页数据，且 AI 智能体需要即时处理复杂查询并吞吐海量原始数据，使动态检索成为绝对刚需。 (4:23) AI 原生搜索必须满足五大核心支柱：权威无幻觉、复杂查询解析、低于 200 毫秒的低延迟、Token 传输高密度优化，以及通过 ExaConnect 接入 SimilarWeb 和 Crunchbase 等第三方专有数据源。 (6:22) Exa 的 GPU 集群规模相较 2021 年扩容约 100 倍以训练专有嵌入模型，当前已对全网扫描并建立了覆盖 800 亿份高质量文档的检索索引。 (11:44)

Will Bryk, CEO of Exa, argues that artificial intelligence systems require an entirely new search architecture because legacy ad-supported search engines are structurally misaligned with the operational demands of autonomous agents. Because large language models face strict parametric memory limits, ongoing AI development depends on dynamic, high-density retrieval systems optimized for accuracy rather than human attention. The Conflict Between Ad Monetization and Information Utility (1:28) Bryk contends that legacy search platforms have failed to fulfill their original missions of organizing global knowledge because their business models rely on advertising. By monetizing user attention, page views, and personal data, legacy search engines are commercially incentivized to optimize for engagement and clicks rather than delivering objective, complete answers. This dynamic has populated the web with search engine optimization (SEO) inflation, spam, and fragmented information. In contrast, Bryk argues that modern search engines must operate as utilities. An enterprise search API charging per query aligns a provider's revenue directly with the accuracy and utility of the data returned to the client, removing the structural pressure to manipulate results for ad impressions. The Transition From Human Queries to Agent Retrieval (3:06) According to Bryk, human search activity currently generates approximately 15 billion queries per day, but automated searches generated by AI agents are expanding exponentially. He forecasts that automated AI search queries will surpass total human search volume by the end of 2026 and eventually reach 1,000 times human volume within a few years. This shift fundamentally changes technical requirements. While human searchers enter short keyword phrases and manually browse blue links, AI agents function as programmatic consumers. Autonomous systems generate intricate multi-sentence queries, ingest thousands of pages simultaneously, and demand dense, structured data payloads that can be parsed instantly. Parametric Memory Limits and Information Theory (4:23) From an information theory perspective, Bryk maintains that frontier foundation models cannot internalize all external knowledge. The public web contains trillions of documents and billions of terabytes of evolving data, whereas next-generation model weights (such as anticipated frontier architectures) cap out around 10 terabytes of parametric memory. Static weights cannot permanently store global human knowledge, making external retrieval indispensable. Because external data access dictates system output quality, Bryk asserts that pairing smaller or open-source models with high-grade retrieval pipelines can match or exceed the performance of larger, costlier models running on inferior search tools. Architectural Pillars for AI-Native Search Engines (6:22) Bryk identifies five core technical characteristics necessary for an AI-native retrieval engine. First, engines must verify sources and filter out SEO spam to eliminate enterprise hallucinations. Second, systems must execute multi-variable logical queries (such as cross-referencing executive roles, hiring status, and funding rounds) and deliver structured results in half a second. Third, latency must scale down to sub-200 milliseconds for interactive voice agents, while retaining batch depth for asynchronous research. Fourth, engines must optimize token economics: delivering roughly 100 dense tokens per result instead of thousands of raw page tokens dramatically cuts downstream model inference costs. Finally, platforms require integrated access to external, non-indexed databases (via integrations such as ExaConnect with SimilarWeb or Crunchbase) to supply private market and traffic data alongside web results. To support this retrieval architecture, Bryk reports that Exa expanded its GPU infrastructure approximately 100 times between 2021 and 2026, curating an index of 80 billion high-authority documents from trillions of scanned URLs. Bryk envisions this infrastructure providing reliable context across critical societal events, including the 2028 U.S. presidential election, though the long-term feasibility of the ecosystem depends on whether retrieval infrastructure can economically absorb automated query volumes projected to reach thousands of times current human scale.

Exa 联合创始人兼首席执行官 Will Bryk 指出，前沿人工智能的发展受制于模型参数存储与传统广告驱动型搜索的缺陷，AI 智能体需要一种全新的、以准确度和数据密度为导向的检索基础设施。 传统搜索模式与现代信息需求的脱节 (1:28) Bryk 认为，主流媒体、社交网络和传统搜索引擎最初都不是为了提供完备、客观的知识而设计的，大部分公众在面对国际争端或移民政策等重大问题时往往只能获得不完整的事实。这一缺陷源于传统引擎的商业模式：例如 Google 虽然早年提出了整理全球信息的愿景，但最终转向了基于广告曝光和用户停留时长的变现逻辑。这种机制促使平台优化用户点击率而非信息客观度，导致搜索不再作为中立的公用基础设施运作，也加剧了虚假信息泛滥的风险。 智能体驱动的查询规模与商业模式重构 (3:06) 人类目前的日均搜索量约为 150 亿次，但自 2022 年以来由 AI 智能体发起的自动化查询正呈指数级增长。Bryk 预测，到 2026 年底 AI 的搜索量将超越人类总和，并在几年内达到人类查询规模的 1000 倍。这种需求侧的剧变重构了搜索的技术与经济模式：由于 Exa 采用按 API 查询次数计费的商业模式，其商业利益与为系统提供精准上下文深度绑定，而非依赖向人类展示广告。只要模型获得更可靠的高质量数据，搜索服务商的收入就能持续增长，从而消除了传统广告模式带来的利益冲突。 信息论视角的模型瓶颈与检索架构差异 (4:23) 从信息理论来看，前沿大语言模型受物理限制无法脱离外部检索系统。公开网络涵盖数万亿份文档以及数以十亿计的太字节动态数据，而像 GPT-5 这样下一代顶级模型的参数记忆上限预计仅在 10 太字节左右，无法完全容纳人类全部知识体系。与此同时，AI 智能体与人类的检索行为截然不同：人类通常输入简短的模糊关键词并人工筛选几个链接，而智能体则是信息吞吐极高的数据处理机器，能够自动构建复杂结构化语句并同时解析数千份文档。测试表明，通过结合低成本的开源模型与高质量检索系统，智能体的实际表现可以媲美搭配劣质搜索工具的昂贵前沿模型。 原生面向 AI 搜索的五大技术支柱 (6:22) 为了向自动化工作流提供可靠支持，AI 原生搜索引擎必须具备五项核心能力。第一是来源可信与权威，通过过滤垃圾内容、SEO 操纵和虚假信息，杜绝模型幻觉。第二是复杂查询理解，能够在半秒内处理多重逻辑过滤并返回符合数据库规范的结构化结果。第三是低延迟与高度定制，即语音交互等实时场景需低于 200 毫秒的响应，而非实时批量分析则优先追求深度。第四是经济效率，通过将单条结果从原始网页的上千个词元压缩至约 100 个高密度词元，大幅降低下游推理费用。第五是外延丰富度，通过 ExaConnect 市场整合类似 SimilarWeb 和 Crunchbase 等付费或私有数据源，形成统一的数据交付网络。 基础设施扩张与社会过渡期定位 (11:44) 构建网络级搜索引擎的技术门槛极高，Bryk 强调全球拥有独立搜索工程能力的主体甚至少于国家级航天计划。自 2021 年成立以来，Exa 将最初半数的种子资金用于采购 GPU 集群，到 2026 年其计算集群规模已扩大约 100 倍，用来训练专为理解全网设计的嵌入模型。目前 Exa 从扫描的数万亿网址中筛选并索引了 800 亿份高质量文档，服务超过 5000 家企业客户与 40 万名开发者。Bryk 规划在 2028 年美国大选前夕为数十亿用户提供透明、客观的信息基础设施。面对不可逆转的 AI 技术浪潮，他认为除了算法专家之外，法律等非技术专业人士在维护社会制度稳定中同样关键，而 Exa 的核心使命在于为这一过渡期构建近乎完美的信息底座。 Bryk 认为，前沿模型参数规模的物理局限与智能体查询量的指数级爆发，使检索成为决定人工智能实际应用效能的核心瓶颈。然而，将庞大的动态公网以及私有商业数据高密度压缩为权威、低幻觉的信息输入，仍高度依赖长期的底层基础设施投入与持续优化的模型架构。

## Background 该视频由频道 Ai4 于 2026 年 8 月 16 日发布在 YouTube 平台。视频简介指出，当前企业知识工作的核心瓶颈在于上下文数据供给，Exa 联合创始人兼 CEO Will Bryk 将在演讲中探讨为何必须从底层为 AI 重构搜索基础设施，以支撑未来规模达到人类 1,000 倍的智能体搜索需求。 ## TL;DR Will Bryk 深刻阐述了 Exa 打造“面向 AI 的完美信息基础设施”的战略蓝图。他指出传统搜索引擎受困于广告变现模式与面向人类的点击交互，且根据信息论原理，大模型有限的静态权重永远无法包容动态互联网，必须依赖专业搜索引擎实时供给精准上下文。针对 AI 智能体的大规模、结构化数据吞吐需求，Exa 确立了可信、智能、低延迟、高性价比及打通私有数据的 ExaConnect 五大核心支柱。依托 800 亿级核心文档索引底座与庞大算力集群，Exa 致力于消除信息失真，不仅赋能全球 Agent 落地，更为解决人类重大社会协同难题提供客观信息基石。 ## Atmosphere & style 整场演讲激情澎湃、叙事宏大且兼具深厚的技术理想主义情怀。Will Bryk 开篇以旧金山反乌托邦废墟图幽默切入，随后运用硬核的信息论逻辑与商业利益分析层层推演传统搜索的困境，兼顾硬派技术架构与敏锐的商业洞察；结尾处他分享了多年前与身为律师的母亲之间的温情对话，将底层搜索技术的演进升华至各行各业共同维系社会秩序、平稳度过 AI 历史大转型的哲学高度，极具思想穿透力与感染力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=4s) Exa 致力于构建 AI 时代面向智能体的原生信息基础设施，目前团队超 150 人，已赋能超过 5,000 家企业（包括 Cursor、Cognition 等）与 40 万名开发者。 - [00:04](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=4s) 当前大模型在基础学科与数学推理上呈指数级突破，而人类依赖的主流媒体与传统搜索等信息获取工具却停滞不前，二者脱节加剧了全球信息失真风险。 - [01:05](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=65s) 传统搜索引擎依赖广告变现商业模式，其本质是将用户的注意力出售给广告主，这种利益机制天然与提供“客观、纯粹的完美信息”背道而驰。 - [03:06](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=186s) 搜索主体正经历历史性逆转：AI 智能体发起的搜索量预计在 2026 年底超越人类每日 150 亿次的总量，并在未来数年内达到人类搜索规模的 1,000 倍以上。 - [03:06](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=186s) 从信息论物理极限来看，全网数十亿 TB 的动态数据无法全部压缩写入仅约 10 TB 的模型权重中，外挂搜索引擎是 LLM 认知现实世界的终身刚需。 - [05:02](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=302s) 人类搜索偏向少关键词与点击链接，而 AI 则需要瞬时吞吐数百篇纯净数据；当前限制智能体落地能力的瓶颈已从“模型智商”转移为“上下文数据质量”。 - [06:22](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=382s) Exa 提炼出面向 AI 的完美信息五大支柱：信源权威可信、智能理解长复合指令、超低延迟与定制化、高性价比 Token 抽取，以及公私数据全面连通。 - [06:22](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=382s) Exa 专注于数据深度清洗，可将抓取网页的 1,000 个 Token 噪音压缩为 100 个高浓度的“黄金 Token”，从源头大幅降低下游 LLM 的推理成本。 - [06:22](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=382s) 通过推出 ExaConnect 数据网络，智能体可直接打通 SimilarWeb、Crunchbase 等非公网私有数据库，实现一站式多源商业情报聚合与收益分成。 - [11:44](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=704s) Exa 自研大规模向量表征与检索模型，GPU 算力集群较初期扩容上百倍，当前活跃索引库已覆盖 800 亿份核心权威文档，并持续向数千亿级扩展。 - [11:44](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=704s) 流行病应对、气候治理及政治互信本质上都是信息与社会协同问题，Exa 设定了在 2028 年美国大选前让数十亿人通过 AI 获取近乎完美信息的长远愿景。 - [14:38](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=878s) AI 引发的社会大转型不可逆转，各行各业维系社会秩序与前沿研究同等重要，Exa 将专注做好数据检索基础设施以支撑这场时代变革。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=4s) 完美信息的愿景与落后的信息工具危机 Will Bryk 以极其高昂的情绪开场，向全场观众宣布他将分享一个关于“AI 时代完美信息（Perfect Information）”的构想。他幽默地打预防针说，自己接下来可能会发表许多“听起来很疯狂的观点”，但他真正的目标是让现场所有人都能像他一样对完美信息的未来感到兴奋。回顾 Exa 的发展历程，公司始于 2021 年，最初的愿景就是构建面向未来的完美信息基础设施。如今 Exa 已经拥有超过 150 名员工，服务于超过 5,000 家企业客户以及 400,000 名开发者，客户群体不仅涵盖 Coda、Cursor、Cognition、Neo Labs 等顶尖 AI 原生团队，还包括众多对冲基金、私募股权机构等金融巨头。在经历了数年的积累后，Exa 终于获得了充足的资金储备，得以放手去实现 2021 年成立之初设想的所有宏大蓝图。 面对“为什么当下必须重构搜索引擎”的质疑，Will Bryk 在大屏幕上展示了一张 2035 年旧金山沦为废墟的科幻反乌托邦假想图，并开玩笑说：“虽然旧金山现在算不上最完美的城市，但它还不至于变成大屏幕上这副废墟模样——不过，如果我们无法解决日益严重的信息失真与错误信息问题，这迟早会成为现实。”其背后的核心矛盾在于：AI 技术的演进速度与人类信息获取工具的进化水平之间出现了灾难性的脱节。一方面，AI 正在以不可思议的速度爆发，大模型已经在前沿数学领域实现重大突破，并即将席卷物理学等所有基础学科，甚至引发行业大佬签署公开信试图踩刹车；另一方面，人类获取信息的工具却依然停留在 20 年前的水平——高度受控的主流媒体、充斥算法噪音的社交平台，以及早已僵化的传统搜索引擎。在当今世界，无论是面对 AI 技术的最新进展，还是国际战争与国内移民政策等复杂政治议题，几乎没有人能够真正掌握完整、准确的信息，全人类几乎都在依托残缺不全的信息碎片来理解世界。 - **Exa 规模与生态**：团队规模突破 150 人，赋能 5,000+ 家企业与 400,000+ 名开发者，覆盖 Cursor、Cognition、Coda 等头部 AI 编程与 Agent 产品。 - **技术与认知断层**：AI 在数学与基础科学的突破呈现指数级加速，而主流媒体、社交网络和传统搜索引擎等信息工具却长期停滞不前。 --- ### [01:05](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=65s) 传统搜索引擎的困境：广告商业模式与“完美信息”的背离 为了彻底扭转这一局面，Exa 树立了明确的企业使命：“组织全球所有信息，并按人们和企业所需的形式精准交付”。Will Bryk 坦言，这句使命听起来非常耳熟，因为它几乎完全复刻了 Google 当年立下的宏愿——“组织全球信息，使人人皆可访问并从中受益”。Google 当年确实取得了无与伦比的成就，但最终却未能真正达成“完美信息”的终极目标。导致这一偏差的根本原因在于，Google 在探索商业化的过程中发现了一个过于赚钱的模式：网络广告。 ```mermaid flowchart TD accTitle: Google Ad Business Model vs Exa Search Engine Model subgraph GoogleModel["传统搜索模式 (面向广告变现)"] A["免费搜索服务"] --> B["聚集海量人类眼球"] B --> C["优化点击率与广告位 (SEO/竞价)"] C --> D["将用户注意力作为产品出售给广告主"] end subgraph ExaModel["AI 原生搜索模式 (面向真实数据)"] E["按 API 调用/查询付费"] --> F["追求极致准确与高权威度数据"] F --> G["直接向 AI 智能体输出结构化上下文"] end ``` 一旦搜索引擎的盈利核心转向广告，优化点击率与广告收益的商业逻辑就与“向用户提供纯粹、完美的信息”发生了不可调和的利益冲突。Will Bryk 一针见血地指出：“Google 卖的从来不是搜索技术，他们卖的其实是用户自己。如果你没有为一款产品付费，那么你自己就是被出售的产品。”这种以眼球经济为核心的旧商业范式，在当前向 AI 时代剧烈转型的历史节点上，暴露出了巨大的局限性与副作用。 - **根本利益错位**：广告模式驱动的搜索引擎天然倾向于迎合 SEO、延长用户停留时间并展示广告，而非提供直截了当的客观真相。 - **“你就是产品”**：传统搜索将人类流量打包变现，使得搜索结果充满商业推广、信息茧房与冗余内容。 --- ### [03:06](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=186s) AI 搜索爆发：信息论视角下大模型对外部搜索的必然依赖 随着 AI 技术的到来，搜索领域迎来了全新的商业机遇。在过去 30 年中，互联网搜索几乎 100% 由人类发起，目前全球人类每天的搜索量约为 150 亿次。然而自 2022 年起，AI 智能体发起的网络搜索开始迅猛抬头。作为 2023 年首个推出面向 AI 的专用搜索 API 的团队，Exa 见证了这一指数级暴涨。Will Bryk 预测，到 2026 年底，来自 AI 系统的搜索量将彻底超越全人类的搜索总量；在未来几年内，AI 的搜索量将达到人类搜索规模的 1,000 倍以上。在商业模式上，Exa 直接向 AI 系统提供信息并按查询请求计费，因此客户每天都在推动 Exa 提升信息准确度：“如果你想预测一家公司的最终走向，直接看它的商业激励机制就可以了。” ```mermaid flowchart LR accTitle: Information Theory Weight Capacity Gap subgraph WebData["动态全网数据 (Trillions of Docs)"] W1["海量网页与非公开数据库"] W2["数十亿 TB (不断实时更新)"] end subgraph ModelCap["前沿模型权重 (e.g. GPT-5)"] M1["模型参数上限 ~10 TB"] M2["静态固化、缺乏实时上下文"] end WebData -- "无法全部压缩写入" --> ModelCap WebData -- "通过 Exa 智能搜索实时提取" --> Context["精准 Agent 上下文窗口"] ``` 从信息论（Information Theory）的基本原理来看，大语言模型（LLM）永远无法摆脱对专业搜索引擎的依赖。人类互联网包含数万亿份文档、数十亿 TB 的庞大动态数据，而即便是像 GPT-5 这样极其庞大的下一代前沿大模型，其参数权重容量也仅在 10 TB 左右。在物理极限与数学逻辑上，静态的模型权重根本不可能容纳整个世界的信息；加之现实世界的数据每分每秒都在更新变动，模型必须实时调用外部搜索才能获取最新的动态上下文。 - **搜索主体大反转**：人类每日搜索量约 150 亿次，预计 2026 年底 AI 搜索量将超越人类，并在数年内达到人类搜索量的 1,000 倍。 - **物理与容量局限**：互联网数十亿 TB 的动态数据无法塞入仅约 10 TB 的模型权重中，外挂搜索是 LLM 认知现实世界的终身刚需。 --- ### [05:02](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=302s) 智能体与人类搜索的本质差异：信息瓶颈与架构重构 人类与 AI 在搜索行为上存在着天壤之别。传统搜索引擎是为“偷懒的人类”量身定制的：人类习惯输入两三个简单的关键词，浏览几条带有摘要的网页链接，然后挑一两篇点进去阅读。然而，AI 智能体是“吞噬信息的巨兽”：它们能够瞬间生成高度复杂的结构化检索指令，一键抓取并即时解析成百上千篇原始网页与深层数据。如果继续沿用为人类点击设计的旧搜索架构来服务 AI，无异于缘木求鱼。 ```mermaid flowchart LR accTitle: Human Search vs AI Agent Search Paradigm subgraph Human["人类搜索模式"] H1["输入 2-3 个简单关键词"] --> H2["传统搜索引擎"] H2 --> H3["展示 10 条网页链接与广告"] H3 --> H4["人工挑选并点击 1-2 个网页"] end subgraph Agent["AI 智能体搜索模式"] A1["生成长句、多约束复合 Query"] --> A2["Exa 原生信息引擎"] A2 --> A3["直接返回高价值 Token / 数据库级数据"] A3 --> A4["Agent 瞬间消化数百条纯净知识"] end ``` 在开发 Agent 产品的过程中，决定智能体能力上限的往往不是模型本身的智商，而是其能调用的信息质量。如今即使是低成本的开源模型或轻量级前沿模型，在处理企业级知识工作（如 GTM 市场情报分析、自动化财务尽调等）时也已经具备了足够的推理智商，真正的系统瓶颈在于外部信息输入的广度与精度。将高性价比的模型与 Exa 这样的高质量信息引擎相结合，所达成的任务效果完全可以媲美甚至超越“笨重的大模型 + 劣质搜索引擎”的传统组合；而若将顶级前沿模型与 Exa 结合，则能释放出终极的智能体表现。 - **交互范式转变**：人类搜索看重网页链接与界面交互，AI 搜索则要求纯粹的数据吞吐与结构化直出。 - **破除信息瓶颈**：当前大模型在多数垂直领域不再缺乏智商，决定 Agent 落地胜负的关键已全面转移至上下文数据的精准供给。 --- ### [06:22](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=382s) 赋能 AI 的五大信息支柱：可信、智能、低延迟、高性价比与全面性 为了适应 AI 时代的基础设施需求，Exa 将“面向 AI 的完美信息”提炼为五大核心维度： ```mermaid flowchart TD accTitle: Five Core Pillars of Information for AI Root["面向 AI 的完美信息架构"] Root --> P1["1. 可信度 (Trusted)<br/>权威来源验证，根除幻觉与噪音"] Root --> P2["2. 智能化 (Intelligent)<br/>理解复合多句指令，实现秒级深度调研"] Root --> P3["3. 低延迟与定制化 (Low Latency)<br/>毫秒级实时响应与深度批处理灵活切换"] Root --> P4["4. 高性价比 (Economical)<br/>精炼关键 Token，大幅缩减 LLM 推理开销"] Root --> P5["5. 全面性与 ExaConnect (Comprehensive)<br/>公网数据与私有垂直数据库一体化打通"] ``` 1. **可信度（Trusted）**：当企业依赖 AI 进行数百万美元的投资或关键战略决策时，输出的每一个 Token 都必须建立在经过权威验证的真实信源之上。Exa 优化模型算法以过滤 SEO 垃圾站和错误信息，直接锚定最权威的信源。 2. **智能化（Intelligent）**：AI 提出的检索需求往往包含复杂的长句和多层过滤条件（例如：“帮我寻找设有首席信息安全官（CISO）且当前正在招聘安全岗位的公司，要求位于美国东海岸且近期完成过新一轮融资”）。Exa 不返回广告或网页链接，而是直接输出如同数据库查询般的结构化结果，实现“半秒钟完成一整年研究调研”的极致体验。 3. **低延迟与定制化（Low Latency & Customizable）**：针对语音交互等实时 Agent，Exa 打造了低于 200 毫秒的超快检索能力；与此同时，面对不需要实时性却追求极致深度的批量处理任务，Exa 也支持高度定制的离线高质量扫描。 4. **高性价比（Economical）**：为了解决企业在 AI Token 上的高昂开销，Exa 在数据清洗上下足了功夫。以往抓取网页通常需要塞入上千个包含布局噪音的 Token，而 Exa 能够直接抽取出最核心、最具价值的 100 个“黄金 Token（Juiciest Tokens）”，从源头大幅降低下游大模型的推理费用。 5. **全面性与 ExaConnect（Comprehensive）**：智能体需要打破公网与私有数据的边界。Exa 推出了 **ExaConnect** 数据交易网络，使 SimilarWeb（网站访问量）、Crunchbase（融资数据）等非公网私有数据源能够直接对接给 AI Agent，数据提供商也能以此从庞大的 Agent 查询流量中持续获得丰厚的分成收益。 - **复合查询示例**：通过集成 ExaConnect，智能体可以一条指令搞定“调研 AI 基础设施公司、拉取 SimilarWeb 流量、调取 Crunchbase 融资数据并汇总最新产品发布动态”。 - **Token 降本增效**：从单页抓取 1,000 Token 压缩至提取 100 个高浓度信息 Token，显著压缩 LLM 上下文负载。 --- ### [11:44](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=704s) 从 GPU 集群到 800 亿索引：重构搜索底座与社会协同愿景 打造一个具备上述全部特性的新型搜索引擎，工程难度极其苛刻。Will Bryk 回顾了团队早期的硬核创业史：2021 年 Exa 刚拿到数百万美元融资时，他们顶着巨大风险把一半以上的资金直接砸去采购 GPU 集群。“虽然我不建议其他初创团队盲目效仿，但这确实让我们走通了道路。”到了 2026 年，Exa 运行的 GPU 集群算力规模已经扩张了上百倍，专门用于训练顶尖的检索 Embedding 模型来理解和表征整个互联网。 ```mermaid flowchart LR accTitle: Exa Search Engine Scale and Indexing Pipeline A["扫描数万亿 (Trillions) 级互联网 URL"] --> B["大规模智能过滤与去噪清洗"] B --> C["沉淀 800 亿 (80B) 级核心权威文档索引"] C --> D["持续向数百亿 (Hundreds of Billions) 级扩展"] D --> E["统一信息底座赋能万千行业 Agent"] ``` “世界上拥有独立航天工程的国家数量，比真正自研底层搜索引擎的团队还要多。”Will Bryk 感慨道。目前 Exa 的实时索引库已经收录了超过 800 亿份核心文档，并且正在向数千亿规模全速扩容。后台引擎持续扫描数万亿条 URL，严密过滤掉低质内容，仅保留最权威的知识。 在演讲尾声，Will Bryk 强调 Exa 的长远目标不仅是服务企业，更是为整个人类社会提供解决重大问题的“信息基础设施”。他向团队定下一个里程碑：在 2028 年美国总统大选之前，让数十亿人能够通过 AI 获得近乎完美、客观的信息，从而做出更加明智的决策。无论是应对全球流行病蔓延、重建公众对领导者的信任，还是应对气候变化保护地球，本质上全都是“信息与人类协同问题”。AI 的计算能力必然会变得无比强大，但全人类能否在未来利用完美信息实现良好的社会协同，将决定文明的走向。 - **底层算力与模型**：从 2021 年豪赌 GPU 集群到如今算力规模扩大 100 倍，专门自研超大规模文本表征与向量检索模型。 - **海量索引吞吐**：日常扫描全网数万亿 URL，当前活跃索引达 800 亿份文档，并全速向数千亿量级迈进。 - **社会愿景**：立足 2028 年大选与 2035 年长远未来，致力于通过消灭错误信息来化解人类社会的协作难题。 ### [14:38](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=878s) 每个人在 AI 历史大转型中的角色与 Exa 的使命 在演讲行将结束之际，Will Bryk 分享了一段发生在约十年前的个人趣事。当时他刚刚对人工智能——或者说当时大家更常称呼的“神经网络”（neural networks）——产生极其浓厚的兴趣与狂热。他回忆起自己当年激动地给母亲打电话时的情景，他在电话里对母亲说：“妈妈，这场 AI 大变革太不可思议了，我们每个人都必须尽自己的一份力量，帮助引导整个人类社会度过这场转型。”然而，他的母亲面对这番宏大叙事却给出了极其务实且略带困惑的反应：“威廉（William），我是一名律师。你到底觉得我能做什么呢？” 当时被问懵了的 Will Bryk 一时不知道如何作答，只能先挂掉了电话。但他反复琢磨着母亲的反问，不久后又重新把电话打了回去，对母亲说道：“妈，你是一名律师，你在确保整个社会不至于陷入崩溃。如果我们失去了所有的律师，那后果绝对比失去所有 AI 研究人员要严重得多。”母亲对这个回答显然非常满意。Will Bryk 借此故事强调，这一幕背后的深刻道理在于：当下正在发生的 AI 变革规模极其浩大，人类正处于历史长河中最关键的几年，这场大转型无论如何都会发生，而每个人都有责任维持社会的良好运转，确保转型能够尽可能平稳顺利地推进。 针对 Exa 自身的定位，Will Bryk 表示团队将全力以赴做好自己的那份职责——那就是打造极致的基础设施，努力让供给给智能体的信息检索变得尽可能完美。但他同时向台下的所有与会者呼吁，在场的每个人都在各自的公司中打拼、推进着各自的项目，所有人都在这场大变局中扮演着至关重要的角色，希望大家在大会闭幕之际都能深刻意识到自身工作的非凡意义。 - **母子对话的启示**：社会正常运转的基石（如法律制度）与前沿技术同样重要；AI 转型不仅依赖 AI 专家的突破，更依赖于各行各业守住社会运作的基本秩序。 - **历史关口的必然性**：AI 引发的社会大转型具有不可逆的必然性，当前正处于人类历史上最为关键的节点，平稳度过转型的关键在于全社会的协同参与。 - **Exa 的核心责任**：在整个 AI 时代的技术版图与生态中，Exa 专注的任务是攻克搜索与数据检索环节，为智能系统提供最精准、最完美的信息输入。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=4s) 虽然旧金山现在算不上最完美的城市，但它还不至于变成大屏幕上这副废墟模样——不过，如果我们无法解决日益严重的信息失真与错误信息问题，这迟早会成为现实。 — Will Bryk > [01:05](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=65s) 组织全球所有信息，并按人们和企业所需的形式精准交付 — Will Bryk > [01:05](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=65s) Google 卖的从来不是搜索技术，他们卖的其实是用户自己。如果你没有为一款产品付费，那么你自己就是被出售的产品。 — Will Bryk > [03:06](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=186s) 如果你想预测一家公司的最终走向，直接看它的商业激励机制就可以了。 — Will Bryk > [11:44](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=704s) 世界上拥有独立航天工程的国家数量，比真正自研底层搜索引擎的团队还要多。 — Will Bryk > [14:38](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=878s) 妈妈，这场 AI 大变革太不可思议了，我们每个人都必须尽自己的一份力量，帮助引导整个人类社会度过这场转型。 — Will Bryk > [14:38](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=878s) 威廉，我是一名律师。你到底觉得我能做什么呢？ — Will Bryk 的母亲 > [14:38](https://www.youtube.com/watch?v=NIMi4t3N5gg&t=878s) 妈，你是一名律师，你在确保整个社会不至于陷入崩溃。如果我们失去了所有的律师，那后果绝对比失去所有 AI 研究人员要严重得多。 — Will Bryk ## Mentioned Will Bryk, Exa, Google, Coda, Cursor, Cognition, Neo Labs, Kimi K3, GPT-5, ExaConnect, SimilarWeb, Crunchbase ## Video description Will Bryk - CEO & Co-founder, Exa Why? Because even opensource models, like Kimi K3, are now intelligent enough for most knowledge work in the enterprise. The differentiating factor between companies and products is the data these intelligences have in their context. And search is meant to filter all the world's data into exactly the context that's needed. The problem is that traditional search engines were not built for AI, they were built for humans. Humans search for links and click on a couple of them. In contrast, agents want to search for raw data and can consume thousands of results instantly. And then there's the scale that's required -- the number of searches from AI systems will be 1000x more than humans in a few years. The world needs search rebuilt from the ground up for AI. This talk covers how we've already started to do that, how thousands of companies are using this infra, and how perfect search is ultimately what every enterprise needs. [Source](https://www.youtube.com/watch?v=NIMi4t3N5gg)

[0:04] Will Bryk: 虽然旧金山现在算不上最完美的城市，但它还不至于变成大屏幕上这副废墟模样——不过，如果我们无法解决日益严重的信息失真与错误信息问题，这迟早会成为现实。 [1:05] Will Bryk: 组织全球所有信息，并按人们和企业所需的形式精准交付 [1:05] Will Bryk: Google 卖的从来不是搜索技术，他们卖的其实是用户自己。如果你没有为一款产品付费，那么你自己就是被出售的产品。 [3:06] Will Bryk: 如果你想预测一家公司的最终走向，直接看它的商业激励机制就可以了。 [11:44] Will Bryk: 世界上拥有独立航天工程的国家数量，比真正自研底层搜索引擎的团队还要多。 [14:38] Will Bryk: 妈妈，这场 AI 大变革太不可思议了，我们每个人都必须尽自己的一份力量，帮助引导整个人类社会度过这场转型。 [14:38] Will Bryk 的母亲: 威廉，我是一名律师。你到底觉得我能做什么呢？ [14:38] Will Bryk: 妈，你是一名律师，你在确保整个社会不至于陷入崩溃。如果我们失去了所有的律师，那后果绝对比失去所有 AI 研究人员要严重得多。

[11:44] Will Bryk: there are more space programs in the world than there are true search engines [14:38] Will Bryk: Mom, you're a lawyer. You ensure that society doesn't collapse. If we lost all the lawyers, it would be way worse than if we lost all the AI researchers.

[0:04] I'm going to tell you about perfect  information for an AI future [0:07] and I'm going to say a lot of crazy things  here so that energy will help. My goal [0:10] is to get you all as excited about  perfect information as I am. Okay. [0:16] Well, just a quick note on Exa. Uh we started  in 2021 to build perfect information uh and a [0:22] lot has changed. You know, now we're over 150 uh  people. We serve over 5,000 companies and 400,000 [0:27] developers. Uh you know, companies like, you know,  Coda, companies like Cursor, Cognition, Neo Labs, [0:32] a bunch of financial institutions like hedge  funds, private equity, and everyone in between. Uh [0:36] and now finally we have enough funding uh to build  all the crazy things that we wanted to do in 2021. [0:43] Okay. Well, you might be wondering why do  we even need a new kind of search engine? [0:47] Well, it's because if I think if we don't build  one, we get a future that looks like this. [0:53] So, this is San Francisco in 2035. Now, I know  San Francisco's not the best city, but it doesn't [0:59] look like this yet. Um but I think it will if  we don't uh solve this misinformation problem. [1:05] And the reason is because as you all know from  this conference, AI is progressing extremely [1:10] quickly and your information tools are not  progressing at the same rate. So, on the AI side, [1:15] you know, as of last week, AIs are making major  discoveries in mathematics and that's going to [1:18] continue and accelerate. You know, uh companies  are now trying to sign letters to slow this [1:23] stuff down. We're going to get this sci-fi future  first in math, then physics, and everything else. [1:28] At the same time though, our information  tools are stuck like they were 20 years ago. [1:31] Right? Like where do we get our information? From  the mainstream news and that has pretty, you know, [1:36] bad properties. Social media, that's got even  worse properties. And then traditional search [1:39] engines. Like none of these tools were built to  provide perfect information to us. And I would ask [1:44] you, do you feel like you have perfect knowledge  of even just like what's going on in AI right now? [1:49] If you're like me, you don't. Um do you have  perfect knowledge about political issues like [1:54] what's going on in some war, uh across the  world or even immigration at home. Again, [1:59] we all don't. We all pretty much walk around with  incomplete understanding of nearly everything [2:04] because our information tools were never  optimized uh to to give us perfect information. [2:10] So, the way to solve this and what X's mission  is is to organize all the world's information [2:14] and provide it back to people and companies  however they need it. It might sound familiar um [2:20] uh because you know, there's another company that  had this mission to organize world's information, [2:23] make it universally accessible and useful. You  might know what company I'm thinking about. [2:27] But they never really achieved their mission.  So, uh you know, this is Google by the way. Uh [2:33] and like the Google, you know, had big  aspirations to organize world's information [2:36] and they did a fantastic job, but they never  really got there because at some point along [2:41] the way they discovered a business model  that was so good that ad like ads uh and [2:46] you know, optimizing for ads is not aligned with  optimizing for perfect information uh for users. [2:52] And the reason is because Google was never really  selling search. Ultimately, they were selling you. [2:57] Uh you know, if you're not paying for a product,  you are the product. Um and and that has real big [3:01] big implications now that we're transitioning  into this crazy AI world. Luckily, there is a [3:06] new kid on the block uh that enables a new type of  business model for search. I'm talking about AI. [3:13] So, this is a graph of uh the number of searches  per day from humans and AIs over the past 30 [3:18] years. Obviously, in 1998 uh there was no AI  searching it was all humans. And even now there [3:23] we they you know, humans search on average 15  billion times a day. Uh but since 2022, AIs have [3:29] started to search the web. And it's growing really  fast. By actually by end of 2026, uh we expect the [3:35] number of searches from AI systems to exceed the  number of searches from humans. I know we've been [3:39] on the front lines like we were the first search  API for AI in 2023. So, we've seen this growth [3:44] and it's going to continue like crazy. Uh you  know, from our projections, we expect the number [3:48] of searches from AIs to be a thousand times the  number of searches from humans in a few years. [3:53] Uh this is real. This is crazy. By the way, like  this is not uh to scale. It was hard to make this [3:57] graph. You know, the blue would actually be like a  you know, at the top of like a four-story building [4:01] or something. Um and and why does this matter  with regard to perfect information? Well, because, [4:05] you know, our business model is we serve AIs uh  information. We get paid per query. So, you know, [4:11] like you know, we wake up in the morning and all  these customers are like, "Why can't your search [4:14] do this? Why can't it do that?" They want perfect  information. And so, we are incentivized to build [4:18] it. And if you want to know where a company's  going to go, just look at the incentives. [4:23] Okay, you might be wondering, why do  AIs even need search in the first place? [4:28] Um well, there's a simple information theory  argument, which is that they can't fit [4:32] the whole world's information into their  weights. So, here's the web. It's around Oh, [4:36] there's something behind me. Here's the web.  It's like uh you know, right now you see 10,000 [4:40] documents on a screen. The web is trillions  of documents, billions of terabytes of data. [4:46] But, you know, our biggest models, our  biggest language models are like GPT-5 or [4:49] like around 10 terabytes. So, you physically  cannot fit the world's information into these [4:54] uh models. And also, the world's information is  constantly updating. It's constantly changing. [4:58] And so, LLMs will always have to rely  on a search engine. It's good for us. [5:02] Okay, and what's interesting is these AIs don't  search uh in the same way that humans do. So, this [5:08] look might look familiar. This is what you  look like when you're searching the web. [5:10] Uh you know, you're a lazy human who types in a  few keywords. And you know, traditional search [5:15] engines were optimized for this creature. Uh you  get a few links. You click on the links. That's [5:19] what search was. But, AIs are not like humans.  AIs are these like information-guzzling machines [5:24] that could you know, instantly output complex  queries and get a list of a thousand uh results [5:28] and instantly uh consume all that information. It  would be crazy if the same search that was optimal [5:33] for humans was optimal for AIs. It's not. You have  to build a whole different type of search engine. [5:39] One thing that's cool and probably relevant to  all of you, uh you you know, a lot of you are [5:42] building agents. Well, your the performance of  your agents is not just about the intelligence of [5:46] the model. And I think people are coming around to  this. I I've been tweeting about this for years, [5:49] but I think people are starting to realize that  you know, the agents are already smart enough for [5:52] a lot of tasks. Whether you're doing go-to-market  intelligence or or or you want a financial agent, [5:57] they're smart enough. And actually, what  they're bottlenecked by is the information [6:00] they have access to. And so, what we've been  noticing is that a lot of uh companies are [6:03] using these cheap open-source models and combining  it or or cheap frontier models and combining it [6:08] uh with a super high-quality uh information  engine like Exa, and you get the same quality [6:12] as if you were using a bigger model and a worse  search engine. Of course, you could now combine [6:16] the best uh information engine with the  frontier models to get the best performance. [6:22] Okay, so I want to just walk through, you  know, what is super high-quality information [6:26] um for a for this AI world. What does it even  mean? And you know, there's a lot uh uh of things [6:30] that are changing in search, but we tried to boil  it down into the five uh most important aspects [6:35] of information for this new AI world that enable  us to get both perfect information for society, [6:39] but also, you know, make everyone's businesses  do really well. And they're perfectly aligned, [6:43] so that's great. Uh okay, so first is trusted.  Making information trusted. So, by the way, this [6:48] little robot thing is your AI, and whatever the  the environment or the thing it's holding is uh [6:53] the information engine. Uh so, yeah, information  needs to be trusted, right? Um when you're talking [6:58] to your AI system and, you know, maybe you're  making a multi-million-dollar investment based [7:02] on what the AI outputs, you want to make sure  uh that every like token that's coming out of [7:07] that thing is correct. And so, we all want and we  should have uh every token is is based in some, [7:13] you know, uh source that is verified. Um and  and you could do that with a search engine, [7:17] but it's not just about having a source. You  need to make sure that's the right source, [7:21] right? There's lots of misinformation out there.  Um there's lots of non-authoritative sources. So, [7:25] you know, at Exa, we've been trying to optimize  our search engine uh not for human clicks, [7:28] but for, you know, the most authoritative  sources, the least about of misinformation. [7:33] It's a very hard problem, but you  could get pretty close uh to perfect. [7:37] Okay, another big quality of uh uh of the best  information for AIs is to make them make the [7:43] information intelligent. And what that means is  these AIs are going to be making very complex [7:47] queries. If you actually look at the queries  that are coming into Exa, they look really weird. [7:51] They're like multiple sentences, they have all  these filters because you have a very intelligent [7:55] uh creature, the LLM, that's trying to make  really complex queries. And so we've designed [8:00] Exa with that in mind. Uh this is not the typical  query you would make to Google, but this is the [8:03] kind of query that AIs like to make. Like find  me companies with a chief information security [8:07] officer that's hiring for security roles. And they  just want exactly the information they asked for. [8:12] They don't want SEO, they don't want ads,  they just want a database basically response [8:16] of exactly what you asked for. And often they make  more complex queries than this. They might say, [8:20] "Find me companies with these properties, but uh  you know, that are based on the East Coast or that [8:24] recently raised a funding round." And you want a  search engine that can handle all these things. [8:28] And so we we we now have a search engine that can  do that. Our goal is I I like to say that a year [8:32] of research in half a second. That's what a search  engine an intelligent search engine should feel [8:35] like. Which is really cool. Imagine like your a  you or your AI is typing something and it goes off [8:40] and within half a second it's almost like it did a  year of research. That's what it should feel like. [8:44] Okay. And related to that is making information  low latency. Our AIs are moving extremely quickly. [8:50] You know, whether you know, let's say you're  building a voice AI, you're talking to your AI, [8:53] it might need a search in the background  multiple times and then respond to you. Well, [8:57] you know, it can't just be half a second like the  Google was made to be half a second cuz that's [9:01] what humans like. But these AI systems, they want  faster than Google. And this is what it feels like [9:05] to have, you know, sub 200 ms search that  we've built. It's like It's like, you know, [9:10] you you you can't really This is obviously not  built for a human, but this is what AIs like. [9:15] Well, certain AIs like that. Actually, you  know, some use cases they require in in in AI [9:20] agents that people build require the fastest  possible search. But other AI agents, they [9:24] don't care about latency at all, right? If you  know, now we have customers who want to run like [9:28] crazy batch jobs over the world information  and they don't care at all about the latency. [9:32] They just want the highest quality. And so the  right information engine for this AI world is [9:36] actually like very customizable. It could be super  fast or it could be super slow and high quality. [9:42] Another component of information for AIs is to  make [clears throat] them economical. We're all [9:47] uh dealing with the problems of like our  organization spending way too much on AI. [9:50] Uh well, search can actually help with that. Um  and we put a lot of effort into making the highest [9:57] uh at every quality level, the the cheapest  search possible. So, you save the most money [10:01] when your agent is using the search. And beyond  that, we actually have put in a lot of effort into [10:06] returning the least amount of tokens to your  agent, like the highest quality tokens, so that [10:10] your total LLM costs are a lot lower, right? So,  in the past, like years ago, we would return a [10:14] lot more tokens, maybe 1,000 tokens per page.  Now, we can return like 100 tokens that are the [10:19] the juiciest tokens, the ones that contain all the  value, and that could save you a lot of uh money. [10:24] Lastly, a a perfect information engine uh  needs to be comprehensive. These AI agents, [10:28] they don't care if something information's on the  public web or not on the public web. They just [10:31] want all the information out there. They have the  patience to analyze as much data as you give them [10:36] uh unlike humans do. Uh and so, we put a lot of  thought into how do you enable uh these agents to [10:41] have access to all the world's information?  And so, something we recently released, [10:45] that we call it ExaConnect, is this idea  where, you know, the companies that have data [10:50] uh can now provide that data to agents who want  that data. So, basically, there's a marketplace, [10:54] and we're trying to enable a marketplace where  the the the providers of that data can actually [10:57] make money from the agents. And you know, as the  amount of searches from agents goes exponential, [11:02] this could actually be a humongous amount  of revenue uh for all these data providers [11:06] in a way that ads never actually achieved. Uh  so, that's pretty cool. So, we basically like [11:10] we're starting to gather now like the public  web, but also way more data beyond the web. [11:15] And that allows you to do queries like this.  Like research AI infra companies, get me the [11:19] monthly website visitors with SimilarWeb, which is  a non-public data source, fundraising total from [11:24] Crunchbase, uh not another private data source,  and the latest product releases. This is a complex [11:28] uh query that involves public and private data,  and this is what it would look like with Exa. [11:35] Right. You get a database response of of all  these companies and all that information, [11:39] right? That's That's quite valuable.  It's exactly what the agent wants. [11:44] Cool. Okay, so those are the properties uh of  information of perfect information for agents and [11:49] for the world. Um and I'll just quickly tell you  about like the journey to get there. So, this is [11:54] in 2021 when we started Exa, we actually we raised  a couple of million dollars and immediately spent [11:58] half of it on a GPU cluster, which I don't  recommend for any startups out there. Um [12:02] but it ended up working out pretty well and um and  now in 2026, you know, we have a a much bigger GPU [12:07] cluster which enables us which we recently got.  Uh you know, it's like I don't know, maybe a 100x [12:12] bigger uh that allows us to do all the crazy uh  training uh that we want to do. Cuz to build a [12:16] search engine that has all these properties, you  need to train uh embedding models like extremely [12:21] high intelligence models that then process the  whole web. I mean, we could go into it at some [12:25] point later. But okay. Uh and yeah, we had to  build a search engine from scratch. So, it takes [12:29] years to build a search engine. This is You can't  actually read this, but the point of this slide [12:33] is to show you that it's really complex to build  a search engine and it takes years of research [12:36] and development. I I like to say that there's  like there's more uh space programs in the world [12:40] than there are true search engines. Uh it's it's  really a really hard piece of software to build. [12:45] And it takes massive scale engineering. Uh so,  we we recently announced the size of our index. [12:49] It's, you know, we serve 80 billion  documents right now. We are rapidly [12:52] expanding that to hundreds of billions,  uh which is very fun for the engineers [12:55] at Exa. And we actually scan trillions of URLs  and filter to the highest quality information. [12:59] Remember, we're only trying to give uh  these AIs the most authoritative sources. [13:05] And yeah, it's been really fun over the past 5  years to work with all sorts of cool companies [13:08] and build the search engine with them.  There's a lot of learnings that go into what [13:12] uh you know, what is perfect for AI agents. It's  very weird and and and counterintuitive. Uh it's [13:17] been really cool to work with all these companies.  Of course, if any of you are building agents, [13:20] you can email me at will@exa.ai and and and we'll  talk. Hopefully, I get a thousand emails after [13:24] this. Um and yeah, I think what's also interesting  is like you could have one information layer that [13:28] serves all of the different types of agents. And  we're in the very early days, right? Like we, [13:33] you know, we built a good search engine that's  very valuable for a lot of companies, but it's not [13:37] perfect search yet. And so there's a lot more to  go. Um, and you know, it's a pretty big deal. Uh, [13:43] I I tell I tell the company uh, that like one one  clear goal is like there's the 2028 presidential [13:49] election is coming up. We should enable billions  of people to have near perfect information [13:54] by that date. Cuz then, you know, they can  make informed decisions and and and we'll [13:57] make better decisions as a society and all these  things. And and and information really does help, [14:01] you know, solve big problems, whether it's  like uh, you know, coordinating pandemics [14:04] or trusting in leaders or protecting the  planet. These are information problems. [14:08] They like the AI is going to get really really  good. What what's what's not necessary or not uh, [14:12] inevitable is how do we coordinate as a species?  And so information is very critical to that. [14:17] And so, I think the world could look like  this. Uh, this is San Francisco in 2035. [14:22] Uh, a world filled with perfect information, not  misinformation. And I truly think like it it will [14:26] help us get to this world. Um, people will make  informed decisions as individuals, as a society, [14:32] and we'll navigate this crazy AI transition  really well. I'll just end with uh, a story. Uh, [14:38] you know, I got super excited about AI or what  it was called neural networks at the time, [14:41] like 10 years ago. And I remember calling my  mom on the phone and saying, "Mom, like the this [14:47] AI transition is crazy and you know, we all have  to do our part to help navigate society through [14:51] it." And my mom was like, "William, I'm a  lawyer. Uh, what do you expect me to do?" [14:56] And I I I didn't know the answer and so at some  point I hung up the phone. But I thought about it [15:01] and I called her back and I was like, "Mom,  you're a lawyer. You ensure that society doesn't [15:05] collapse. If we lost all the lawyers, uh, it  would be way worse than if we lost all the [15:09] AI researchers." And she liked that answer. And  and and the lesson here is that like, you know, [15:13] this is a massive thing that's happening. This  is the most critical years in human history [15:18] and we all have a role to play um, Um, ensure  that society is functioning well uh, so that this [15:23] transition that is going to happen no matter what  goes as as well as possible. You know, we're going [15:26] to do our part at XA trying to make information  as perfect as possible. But, you know, you're [15:31] all working at companies, you're all working at  your projects and and and I just hope, you know, [15:35] you see how important it is, um, you know, as  we end this conference. Okay. Thank you all.