Box CEO Sees Enterprise Apps Winning Trillions as Lab Subsidies End

Box首席执行官认为随着实验室补贴终结，企业级应用将斩获数万亿美元 Token补贴的瓦解与分散的企业权限体系，将使经济剩余从前沿模型实验室转向运行自主后台工作流的应用软件。 Aaron Levie认为，原始基础模型本身无法单独占领企业市场，并断言大部分价值将流向针对专有工作流、安全边界和搜索进行调优的应用层架构。他分析了为何基础模型的Token补贴在经济上不可持续，这使得模型不可知的路由机制和开放权重模型成为企业将成熟工作负载从昂贵的高级API中剥离的必备手段。通过将AI编程的普及与企业知识工作进行对比，Levie指出，尽管开发者受益于GitHub等统一环境，但更广泛的企业自动化却面临碎片化数据和严格的权限壁垒。因此，他预测在五年内，90%的企业Token消耗将由自主、异步的后台Agent驱动，而非面向人类的对话交互。 AI时代的职业发展轨迹在很大程度上由信息流速决定，能否紧密跟进实时的开发者与研究讨论，是拉开行业领袖与落后一年者差距的关键所在。 (0:00) 生成式AI在企业端最大的价值并不仅存在于原始基础模型中，更在于连接模型能力与特定企业工作流及权限系统的复杂应用层。 (1:54) 企业需要模型不可知的路由机制，因为基础模型提供商在帮客户优化成本与路由至自身的高级模型之间，存在天然的利益冲突。 (8:27) 前沿实验室人为补贴的API Token定价长期来看不可持续；开源模型引发的推理利润空间崩塌，将推动经济剩余向技术栈上层的应用提供商转移。 (9:58) 尽管AI代码生成因纯粹的功能实用性而受到欢迎，但AI生成的文本和演示文稿却面临质疑（被称为“工作废话”），因为书面成果往往作为评估作者批判性思维的社交衡量指标。 (18:42)

Box CEO Aaron Levie argues that enterprise AI value will accrue primarily to the applied application layer rather than foundation models, as translating raw machine intelligence into regulated workflows requires specialized agent harnesses, strict permissioning, and deep distribution. Value Shift to the Application Layer (1:54) Levie argues that dismissing generative AI applications as mere wrappers overlooks enterprise operational reality. Raw foundation models cannot independently navigate complex business processes, legacy IT architectures, or human approval gates. Furthermore, he asserts that commoditization pressures—driven by open-source weights from entities like Meta and hardware sales incentives from Nvidia—will erode pure inference margins over time, pushing the economic surplus upstack to application software. Drawing a historical parallel to the rise of public cloud infrastructure, Levie observes that while cloud providers created trillions in value, they enabled an equivalent market capitalization in application and data platforms like Snowflake and Databricks. He projects a similar trajectory for enterprise AI, arguing that independent software providers will capture trillions in value because frontier model labs cannot feasibly scale the massive consulting and implementation workforce needed to tailor models across diverse industries. Proprietary Harnesses and Dynamic Enterprise Permissions (24:08) Off-the-shelf frontier models lack the contextual awareness and access heuristics necessary for enterprise execution. Levie explains that baking organizational data directly into model weights is fundamentally unworkable for most corporate knowledge because employee access privileges and legal ethical walls change daily. Enterprise systems must instead rely on dynamic context retrieval paired with specialized agent harnesses. To solve this, Box developed an applied harness that operates directly on enterprise search, dynamic text chunking, and file permissions across millions of corporate documents. Levie notes this specialized harness outperforms generic model APIs on accuracy and latency. He also advocates for model-agnostic routing, pointing out that frontier labs face a conflict of interest in routing queries to their own expensive models, whereas enterprise applications can systematically benchmark competing engines to achieve exact thresholds—such as 98% extraction accuracy—at the lowest compute cost. Why Enterprise Knowledge Work Lags Behind Coding (48:03) Software engineering adopted generative AI rapidly because code is entirely expressed in text, aligns directly with foundation lab evaluation benchmarks, and is operated by technically proficient users capable of troubleshooting their own tooling. Centralized platforms like GitHub also provided a single, clean integration point for developer tools. In contrast, Levie emphasizes that broader corporate knowledge work diffuses slowly because no universal repository exists. Enterprise data remains scattered across fragmented, on-premises systems and complex permission hierarchies. Furthermore, while code is judged purely by its functional execution, human-written documents serve as proxies for interpersonal trust; AI-generated memos and presentations frequently trigger social resistance as 'work slop.' Finally, workflows like sales and operations are bound by external human rate limits, such as procurement schedules and budget approvals, preventing pure model speed from instantly translating into business velocity. Headless Systems and Asynchronous Background Agents (45:00) While conversational chat interfaces will remain permanent fixtures, Levie expects them to represent only a small fraction of enterprise token consumption. He forecasts that within five years, 90% of enterprise tokens will be consumed by autonomous, asynchronous background agents running continuous tasks—such as triaging security events, processing loan packages, and conducting document governance—delivering finished work to human queues for sign-off. To thrive in this environment, Levie contends that traditional systems of record must adopt a dual strategy. Platforms must become 'headless' by exposing their data via APIs and protocols like Anthropic's Model Context Protocol (MCP) to third-party agents, while simultaneously engineering highly specialized, native agents that outperform external generic tools on their home datasets. Levie concludes that because generative AI dramatically compresses raw software development cycles, long-term competitive moats will depend on enterprise trust, distribution channels, and compliance integration. However, full enterprise autonomy remains constrained by brittle browser automation and the painstaking operational friction of modernizing legacy corporate data.

Box 联合创始人兼首席执行官 Aaron Levie 认为，人工智能的核心经济价值正在从被逐步商品化的底层基础模型，转移至深耕企业工作流的应用层；由于企业内部存在复杂的权限、非结构化数据孤岛与组织惯性，AI 在企业中的扩散速度远比硅谷预期的漫长，但这也构成了数万亿美元规模的应用层软件机遇。 基础模型商品化与应用层的扩散经济 (1:54) Levie 指出，硅谷长期陷入“沉迷纯研究”（research-pilled）的盲区，误以为模型权重的原生智力可以凭空抹平企业的运营现实。然而，即使是超强模型也必须面对遗留 IT 系统集成、人工审批节点以及组织变革阻力。随着 Meta 开源模型权重、Nvidia 推动算力普及以及各自主权 AI 的竞争，底层模型推理成本将被压低至极薄的毛利率（甚至接近 10%），模型厂商单纯依靠高 API 利润补贴应用的商业模式难以在公开市场持续。因此，经济剩余必然向应用层迁移。 面对全球数以万计的银行、药企和律所，基础模型实验室不可能雇佣多达十万名系统集成人员去逐一定制业务流，这催生了庞大的“扩散经济”（diffusion economy）。Levie 进一步强调，模型厂商在销售自家算力时存在天然的利益冲突，无法充当客观的调度者；企业需要的是一个模型不可知（model-agnostic）的独立路由层，能够以最低成本将特定子任务精准分发给最优模型。 企业 Agent 护城河：权限体系与精度爬坡 (24:08) 在企业实际落地中，开箱即用的通用大模型通常只能在复杂金融借贷或法律文档分析中达到约 70% 的准确率，而自动化生产管线则要求 97% 至 98% 的确定性指标。Box 通过其内部前沿团队 Box Labs 构建专属的智能体框架（Agent Harness），利用并行检索、动态文本切片、即时嵌入向量与重新排序等专属机制，弥合通用模型与企业级可用性之间的断层。Levie 透露，在基准测试中，专有框架在准确率和延迟上显著优于直接调用第三方通用 API。 针对行业关于“知识应注入模型权重还是留存于检索上下文”的争论，Levie 强调了动态权限与合规隔离墙（ethical walls）的现实约束。在大型律所或金融机构中，员工权限每日变动，相邻工位处理相冲突业务的律师绝不可共享信息。因此，静态微调模型难以应对瞬息万变的权限控制，基于深度检索增强（RAG）与细粒度权限过滤的方案仍是企业生产环境的核心基础。在模型评测方面，Box 观察到大多数模型的文档处理能力与其代码能力强相关，但 Google 的 Gemini 在文档理解和工具调用上表现出高于纯代码榜单预期的实力。 记录系统无头化与异步后台智能体 (41:45) Levie 主张现有 SaaS 记录系统（Systems of Record）必须采取双重策略：一方面，打造深度优化、在自身数据架构上准确率显著超越通用方案的原生 Agent；另一方面，平台必须彻底走向“无头化”（Headless），通过模型上下文协议（MCP）等确定性接口全面向 Claude、ChatGPT 等外部客户端开放。开放非但不会蚕食业务，反而能激发更高频的数据调用与合规管控价值。 在人机交互层面，Levie 认为传统的对话框（Chat）仅是企业 AI 价值的一小部分，更广泛的生产力变革将来自异步运行的后台智能体（asynchronous background agents）。针对合同归档、服务器日志扫描、合规策略拦截和安全事件分流等高吞吐工作，用户需要的不是对话，而是结构化的工单队列与审批看板。Levie 预测，五年内 90% 的企业 Token 消耗将来自无需人类主动发起、自主在后台运行并仅在关键节点触发人工审核的业务流程。 知识工作与编程扩散速度的断层 (48:03) 软件工程工具在过去两年内实现爆炸式普及，而通用企业知识工作的渗透却相对滞后。Levie 分析，这是因为编程的核心产出高度表现为代码文本，开发者群体具备极高的自主排障能力，且全球绝大多数现代代码天然集中在 GitHub 中，形成了零摩擦的接入标准。相反，企业知识工作没有通用的“GitHub”，关键业务信息高度碎片化地分散在本地服务器、旧版文档系统和彼此孤立的软件中，面临复杂的跨部门权限索取门槛。 Levie 将知识工作描绘为一个自动化光谱：高度依赖屏幕文字处理的闭环工作（如软件开发、法律合同审查）能够迅速享受 AI 红利；而处在光谱另一端的开放式业务（例如企业销售与采购谈判）则受制于外部人类互动、客户预算审批以及会议排期等物理约束。无论内部生成技术多先进，这些职能的整体转化效率仍旧受到外部组织节奏的强力制约。 Aaron Levie 总结认为，虽然现代初创团队凭借 AI 获得了极高的初期研发杠杆，但软件构建门槛的骤降也导致同质化竞争空前激烈。传统企业 SaaS 虽因合规与安全要求而在开发速度上存在折价，却占据了关键的数据重力与渠道信任。AI 时代的企业竞争本质是一场漫长的“扩散战”，最终捕获万亿美元价值的赢家，属于那些具备深度垂直理解、耐心改造复杂企业流程并掌控核心分发渠道的应用层构建者。

## Background 本期内容由红杉资本（Sequoia Capital）于 2026 年 9 月 16 日发布在其官方 YouTube 频道上。视频记录了红杉与 Box 联合创始人兼首席执行官亚伦·莱维（Aaron Levie）的深度对谈，探讨一家拥有季度财报压力的成熟上市公司如何在 AI 时代完成自我重塑。页面简介指出，对话核心围绕应用层的独特商业价值展开，剖析为何基础模型的裸能力必须经由深度适配企业文件系统、搜索与权限的 Agent Harness 才能真正打通业务闭环，并对企业级技术扩散与长程 Agent 的未来给出了前瞻预测。 ## TL;DR 亚伦·莱维指出，硅谷普遍存在的“研究至上主义”低估了真实业务落地的难度，真正的万亿级商业价值沉淀在能够桥接基础模型与企业深水区工作流的中立应用层上。针对编程智能体光速渗透而普通知识工作滞后的现象，他深入剖析了两者的底层差异——代码具有高度功利性与闭环反馈，而常规业务流程既不存在统一的“GitHub”接口，又深陷老旧遗留系统与精细权限隔离的泥潭。同时，莱维论证了模型厂商的 Token 补贴不可持续且推理成本必将商品化，并大胆预测五年内 90% 的企业 Token 消耗都将来自后台自主运行的异步长程 Agent。整场讨论明确传达出一个核心判断：随着开发门槛被 AI 极大拉平，决定最终胜负的关键已不可逆地转向深入企业肌体的“扩散与渗透战”。 ## Atmosphere & style 整场对话语速极快、风趣幽默且信息密度极高，莱维展现出兼具“狂热技术极客”与“老道企业操盘手”的双重特质，频频抛出“根管治疗式访谈”、“让狐狸看守鸡舍”等生动比喻。主持人的敏锐追问与莱维坦率的自我解剖形成绝佳共振，既有每晚通宵刷推特、狂试新工具的兴奋与焦虑，又有对大企业合规治理与真实商业痛点的冷静敬畏。听众能强烈感受到一种历史交汇期的张力——老牌 SaaS 巨头既在为年轻初创团队惊人的研发杠杆感到震撼，又在万亿级企业应用渗透战前夕展现出极具底气与战备状态的自信。 ## Bulletin - [44:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2698s) Box 首席执行官 Aaron Levie 预测，未来五年内企业 90% 的 Token 消耗将来自后台异步长程任务，而非人类用户主动在对话框中触发。 - [09:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=598s) 模型厂商的低价 Token 补贴不可持续，随着开源模型与非经济理性参与者将推理毛利率压低，同等能力下的 Token 成本必然商品化，上万亿美元的核心经济价值将最终沉淀至应用层。 - [24:59](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1499s) Box 实测证明，基于自身企业级文件系统、权限隔离与专有搜索深度调优的 Agent Harness，在准确率与响应延迟上大幅超越直接调用 OpenAI 或 Claude 的原生 API。 - [52:19](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3139s) 知识工作领域缺乏像 GitHub 那样结构统一且授权简单的通用底座，企业数据高度受制于老旧系统与碎片化的细粒度动态权限，导致 AI 的企业渗透周期远长于编程场景。 ## Key takeaways - [01:54](https://www.youtube.com/watch?v=NE4CLThMPGU&t=114s) 应用层正在成为最关键的“新实验室”（Neolabs），填补基础模型与银行、药企或律所等真实业务工作流之间的深层工程鸿沟。 - [08:27](https://www.youtube.com/watch?v=NE4CLThMPGU&t=507s) 模型选型存在“让狐狸看守鸡舍”的利益冲突，企业需要完全中立的智能体路由层在保证准确率恒定的前提下跨模型进行成本最优化。 - [09:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=598s) 底层 Token 补贴属于阶段性现象，数百亿美元的资本开支压力迫使模型走向商业理性，而 Meta、中国厂商等多元力量将进一步把推理商品化。 - [16:43](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1003s) 企业级应用的最大全垒打场景在于将上百万份沉默的非结构化合同与研报转化为可结构化查询的资产，并由后台常驻的长程 Agent 接管端到端流程。 - [18:42](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1122s) 代码被视为纯粹的工程功利工具，社会接受度极高；而商业内容与汇报则是个人思考与信誉的代理变量，AI 生成痕迹极易诱发信任危机与“职场烂作”反弹。 - [37:28](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2248s) 在企业落地中，直接接入通用模型可能仅有 70% 准确率，而跨越到 97% 生产可用门槛的算法“爬山”正是应用层团队的立身之本。 - [33:23](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2003s) 律所等机构面临严格的“中国墙”与动态项目权限隔断，将内部私有数据直接微调进模型权重的方案极其棘手，动态上下文检索（RAG）仍是权限隔离的核心手段。 - [39:02](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2342s) 成熟 SaaS 厂商必须同时打造极度精调的第一方原生 Agent，并将自身记录系统通过 MCP 协议转变为无头架构，在外部模型高频调用中捕获新增量。 - [44:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2698s) 未来 AI 交互界面将从单次即时问答的对话框，全面转向以状态监控、任务排队和最终审查为主的异步后台批处理看板。 - [48:01](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2881s) 编程之所以实现闪电渗透，得益于纯文本可执行环境、密集代码预训练、极客自我排障能力以及实验室自身即用户的日复一日评测闭环。 - [50:18](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3018s) 知识工作的自动化程度呈现连续谱特征：代码与法律文书处于纯数字自包含的一端，而销售拓展等工种则受制于外部现实世界的沟通周期与预算决策。 - [01:01:51](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3711s) AI 让两人的初创团队具备过去 40 人的全栈研发能力，但时间窗口极度收窄，三天内即可涌现多家同类竞品，迫使竞争重心从“造出来”转向“企业渗透与分发”。 - [58:34](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3514s) 在内部 AI 推广中，监控 Token 消耗榜单的真正目的不是指标考核，而是甄别并提炼前沿员工的高杠杆实战工作流，并在数小时内全员推广扩散。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=NE4CLThMPGU&t=0s) 信息流与“接入网络”：在 AI 时代领先一年的关键 在对谈刚开始时，亚伦·莱维（Aaron Levie）半开玩笑地提到，他真想向全美所有大二、大三的在校大学生发送一条紧急避险警报，内容只有一句话：立刻去注册推特（Twitter / X），并死死关注那最核心的 20 个账号。在莱维看来，在当前的 AI 浪潮下，一个年轻人的职业生涯到底是领先时代整整一年，还是落后时代整整一年，完全取决于他的信息流（feed）质量。 莱维提到自己依然经常碰到一些二十出头的年轻人，对方会轻描淡写地说“平时偶尔会看一些文章”。这让莱维感到不可思议并吐槽道：“你怎么会用‘看一些文章’这种方式获取信息？这到底是什么意思？难道是碰巧有人给你发了封邮件附带链接，你才看一眼吗？”他认为在这个时代，保持敏锐的唯一方式就是彻底“接入网络”（wired in），亲自上手玩转、体验市面上涌现出的每一个新工具。 ### [00:50](https://www.youtube.com/watch?v=NE4CLThMPGU&t=50s) 开场寒暄：播客、根管治疗与企业家的痛苦 主持人正式引出嘉宾——Box 联合创始人兼首席执行官亚伦·莱维。Box 作为一家为现代企业提供内容管理与协同办公的 SaaS 巨头，在生成式 AI 的浪潮下焕发出了全新的生命力。主持人希望与莱维深入探讨 Box 的 AI 转型、对前沿技术趋势的宏观洞察，以及成熟创始人如何在 AI 时代完成自我重塑。 莱维幽默地回应称自己是该播客的铁杆粉丝，几乎一期不落。紧接着他话锋一转，打趣道：“我一直好奇，咱们这档播客什么时候也能聊聊‘根管治疗’（root canal）？”这暗指红杉合伙人道格·莱昂内（Doug Leone）此前在访谈中常用来比喻创业艰难的硬核老梗。莱维调侃称，大家完全可以搞一期类似吃辣翅访谈节目《Hot Ones》的企划：创始人躺在牙医诊疗椅上现场做根管治疗，忍着剧痛阐述自己的商业战略，而道格·莱昂内就在视频连线的另一头严肃审视。 ### [01:54](https://www.youtube.com/watch?v=NE4CLThMPGU&t=114s) 应用层公司是最新潮的“新实验室”（Neolabs）？ 话题切入核心技术与商业判断。主持人抛出一个极具争议的论点让莱维表态：“应用层公司正在成为当下最炙手可热的‘新实验室’（Neolabs），你同意还是反对？”莱维明确表示认同。他指出，如果两年前讨论这个话题大家可能不知所云，但现在伴随开源模型生态的爆发，原本被硅谷轻视甚至嘲笑的“LLM 包装套壳”（model wrapper）模式正在被彻底正名。 ```mermaid flowchart TD accTitle: 基础模型到企业工作流的鸿沟 accDescr: 展示基础模型与企业端落地之间的鸿沟及应用层软件所起到的桥梁作用 subgraph M["底层模型层 (Foundation Models)"] RawLLM["前沿大模型能力<br/>(GPT / Claude 等)"] end subgraph Bridge["应用层桥梁 (Application Layer / Neolabs)"] direction TB Conn["多源数据系统连接与权限管理"] HITL["人机协同与延迟处理 (Human-in-the-loop)"] Workflow["定制化业务逻辑与领域工程"] end subgraph Ent["企业实际业务 (Enterprise Workflow)"] Bank["银行合规审查"] Pharma["制药研发流程"] Legal["律所案件卷宗"] end RawLLM --> Bridge Bridge --> Ent ``` 当下市场实际上正围绕万亿美元级别的资本押注进行分歧博弈：一种观点认为，大模型与企业实际工作流之间的桥梁极窄，通用超级智能将直接通吃一切；另一种观点则认为，这座桥梁极其宽广且深不见底，需要深厚的组织内生能力去填补。莱维坚信后者。随着应用层公司不断跨越这条鸿沟，它们不仅能积累专属的业务数据和领域专业壁垒，最终还会水到渠成地拥有反哺自身专用模型的飞轮能力。谈及 Box 自身，莱维透露公司内部其实早已有类似“Box Labs”的机制，目前的重点是把 Agent Harness（智能体线束/控制框架）打磨到极致，使其在底层接入任何主流模型时都能表现优异，未来再针对具体客户和业务场景微调出专用模型。 主持人与莱维进一步归纳了这一趋势背后的两大驱动力： - **利益冲突防范（不让狐狸看守鸡舍）**：企业客户绝不希望由出售底层 Token 的模型厂商来裁定和计量何种业务该使用何种 Token。客户需要一个中立的模型路由层。 - **巨大的应用工程研究工作量**：把通用大模型对接到银行、律所或药企的深水区流程，需要做大量高度定制化的落地工程与应用研究。 ### [04:29](https://www.youtube.com/watch?v=NE4CLThMPGU&t=269s) 硅谷的“研究脑”与企业现实的鸿沟 莱维分析了当前硅谷普遍存在的认知偏差。硅谷有着极具吸引力但有时过于单一的“研究至上主义脑”（research-pilled），理所当然地认为前沿模型的原始通用智商是决定商业成败的唯一形态。然而一旦走进真实商业世界，就会发现现实远非单纯堆叠模型智力那么简单： - 企业流程必须对接纷繁复杂且未现代化的陈旧遗留系统（Legacy systems）。 - 业务流中存在大量必须由人类介入裁决（Human-in-the-loop）的节点。 - 业务流存在周期性延迟，智能体经常需要长期待机并管理状态。 - 涉及深刻的企业组织变革与流程重塑（Change management）。 这些脏活累活正是纯研究导向的通用模型实验室最不愿触碰的领地。莱维用云计算的历史演进做对比：十年前亚马逊 AWS、谷歌云 GCP 和微软 Azure 搭建了数万亿美元的基础设施层，但如果当时仅看云巨头自研的基础数据工具，没有任何人能预测到 Snowflake 或 Databricks 这样数十亿美元营收规模的应用级软件公司的诞生。大模型时代也将如此，底层模型价值连城，但把能力运用于高门槛行业的应用层软件同样将产生惊人的万亿级商业价值。目前模型实验室正面临激烈的战略两难：是冒险向上侵蚀生态、亲自去做具体软件，还是主动留出空间以换取开发者和企业客户的信任。 ### [06:56](https://www.youtube.com/watch?v=NE4CLThMPGU&t=416s) 历史重演：聊天机器人与 Agent 工作流的界限 面对这种应用层与模型层的博弈，风险投资机构正承受着前所未有的“存在主义焦虑”——红杉等一线机构每天都在纠结是该继续给顶级基础模型实验室砸十亿美元，还是重注垂直应用层。 莱维坦承自己由于掌舵 Box，在押注应用层上具有天然且不可分散的偏见，但他强调计算机软件 60 年的发展史从未偏离过这条轨道。当销售人员走进一家大型律所、跨国药企或投行时，对方要买的从来不是一个虚无飘渺的通用智力接口，而是一个深植于行业流程、能够精准访问其核心数据的业务软件。 莱维指出，这种本质差异最生动的例证就是“通用聊天机器人”（Chatbot）与“任务型智能体工作流”（Agentic Workflow）的界限： - 聊天机器人完全是横向、通用且平铺直叙的交互界面，什么都能聊两句。 - 真实企业的具体工作流却有着严苛的前置依赖、精细权限和任务闭环需求，这是单纯提供一个横向聊天窗口永远无法直接交付的价值。 ### [08:27](https://www.youtube.com/watch?v=NE4CLThMPGU&t=507s) 企业的扩散经济学与“让狐狸看守鸡舍”的中立性困境 在深入企业落地时，通用大模型往往无法直接介入具体的业务流程。AI 必须在正确的节点接入系统，或者调取聊天机器人原生无法触及的私有数据。这意味着必须有人深入企业内部进行配置，并将具体的垂直领域专业知识（Domain Expertise）注入模型，使其真正理解特定业务的运作逻辑。Aaron Levie 指出，除非人们假设几大前沿实验室会疯狂扩张到“毫不夸张地说多达 10 万名员工”的规模，否则这种技术扩散经济（Diffusion Economy）的体量将极为庞大。从 50 人的小团队到数十万人的跨国巨头，每一家企业都需要庞大的人力团队去协助推动组织变革管理（Change Management）。 随之而来的核心博弈在于模型选型的中立性。Aaron 提到了所谓的“让狐狸看守鸡舍”（Fox guarding the henhouse）困局——这也是中间应用层必须存在的最大单点理由。即使假定各大底层模型厂商毫无恶意、完全出于善意，商业常识依然决定了：当企业把任务委派给 Agentic（智能体）系统时，诉求是在“保证准确率恒定”的前提下实现成本最优化。而唯一能做到这一点的，只能是一家对底层 10 种不同模型毫无偏向、对调用哪家模型完全中立的公司。 ```mermaid flowchart TD accTitle: 模型中立的智能体路由架构 accDescr: 展示企业任务如何经由中立应用层评估，并根据成本与准确率动态分发给不同模型 subgraph Client["企业业务层"] Task["企业复杂任务 / 工作流"] end subgraph AgentLayer["中立智能体平台 (如 Box)"] Router["动态路由与成本优化引擎<br/>(保持准确率恒定)"] end subgraph FoundationModels["底层模型池"] M1["模型 A (高性能/高价)"] M2["模型 B (高性价比)"] M3["开源/专用模型 (低推理成本)"] end Task --> Router Router -->|动态选路| M1 Router -->|动态选路| M2 Router -->|动态选路| M3 ``` - **变革管理壁垒**：企业落地并非单纯接入 API，而是涉及流程重构、私有数据权限打通与专家经验微调，这催生了庞大的全职人力与配套生态。 - **利益冲突本质**：模型供应商天然倾向于消耗自身的高毛利 Token，而中立平台则致力于为客户跨模型实现成本与效果的最优解。 --- ### [09:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=598s) Token 补贴不可持续：非经济参与者与应用层的价值捕获 针对对谈者提出的“底层模型厂商可以通过价格补贴来压制中间层”的观点，Aaron 直言这种补贴不会长久。当这些前沿实验室走向公开市场上市时，它们同样必须屈从于资本市场的客观规律。在资本开支达到数百亿美元的体量下，巨额的训练成本和基础设施投资终究需要真金白银来偿还，“Token 补贴本质上只是一种阶段性现象”。 更深层的动态来自于底层推理的毛利平衡与生态中的“非经济理性参与者”： 如果模型厂商将 API 定价拉高以赚取高毛利来补贴自身产品，却把客户全数赶往第一方应用，API 收入就会随之萎缩，这一机制存在微妙的动态平衡。与此同时，Meta（通过开源生态）、SpaceX/xAI、中国厂商以及英伟达（Nvidia）等力量构成了完全不同的博弈维度。这四类群体并不依赖推理层维持超高毛利——他们甚至乐意将推理毛利率压低至 10%，仅仅为了跑满并消化庞大的算力集群。只要前沿机密没有形成绝对垄断，同等能力下的 Token 成本必然持续大幅下行，而释放出的绝大部分价值最终将沉淀到应用层。Aaron 强调，那种“一两家基础模型实验室通吃 95% 价值”的论调并不可靠；即便是对头部实验室自身而言，充分竞争也比因垄断而被国家层面收归国有（Nationalization）更加健康。 - **资本约束**：动辄数百亿美元的训练与算力支出，决定了模型厂商无法无限期进行无视利润率的低价补贴。 - **博弈异构方**：开源社区、芯片硬件巨头与特定地缘力量无须维系高推理毛利，其低毛利策略加速了单位 Token 成本的商品化。 - **价值流向**：基础智能成本的下降将直接拓宽上层软件的利润空间，推动价值向业务工作流层沉淀。 --- ### [13:00](https://www.youtube.com/watch?v=NE4CLThMPGU&t=780s) Box 的演进史：从云端非结构化数据到早期 AI 的探索与停滞 对话随后转向 Box 的发展轨迹。Box 最早的切入点极其直观：在云端安全地存储和共享数据。公司迅速向企业端转型（Aaron 在此特别向红杉资本的合伙人致意），捕捉到了历史性机遇——全球大型企业正从传统的本地部署系统（On-premises）全面迁移至云端，它们急需一个能安全存储、共享、协同和管理企业内部所有“非结构化数据”（Unstructured Data）的现代化平台，包括合同、财务报表、营销素材和研发文档。 事实上，Box 早在 2015 年前后便开始试水 AI 技术。在 2015 至 2018 年那场“以为爆发却最终冷却”的 AI 周期中，Box 曾利用早期的计算机视觉与 OCR 技术进行图像分类和文本提取。尽管这对企业资产打标大有裨益，但当时的技术路径存在致命缺陷：**成本高昂，且每一个微小的工作流都必须单独训练一个超专用模型（Hyper-trained model）**，通用性极差，Box 最终不得不将该方向暂时搁置。直到几年后 GPT 系列模型崭露头角，团队在内部黑客松中将其应用于笔记产品的智能输入预测（Type-ahead）以及安全层面的文本检测分类，才重新燃起火花；而真正引发团队认知剧震的，是 ChatGPT 的横空出世——它以一种交互式的对话形态，彻底打破了人们对机器处理复杂长文本能力的想象边界。 - **业务基石**：Box 承载了企业核心的非结构化数据资产，构成了日后 AI 理解企业深层业务的底座。 - **早期 AI 的局限**：2015 年代的深度学习方案受制于单一任务的训练成本与碎片化，无法形成平台化复用。 - **形态转变**：大语言模型（LLM）将孤立的模型统一为通用的交互界面，使连续处理高复杂度任务成为可能。 --- ### [15:08](https://www.youtube.com/watch?v=NE4CLThMPGU&t=908s) 全面转向 AI：坐拥数千亿非结构化文件的 Agent 落地 面对 ChatGPT 带来的产业拐点，Box 展开了一场极其果断的全面业务转型（Full Company Pivot）。管理层抽调核心精锐组建专门攻坚团队，推行日会机制追踪技术迭代，以严谨的标准逐步搭建起现代化的 Box AI 技术栈以及 Box Agent。 Box 拥有的核心壁垒在于其底层托管的“数千亿份企业文件”。这些文件蕴含着商业合同、临床试验研究、营销创意、贷款审批文件等企业最关键的命脉资产。长久以来的核心痛点是：**企业空有海量非结构化文档，却根本无法直观获知其内部沉淀的精确知识**，除非员工手动检索并逐页审读。通过部署针对这批数据的智能体架构，Box 正在重塑这一交互方式： 1. **精准即时问答**：将 Agent 派驻至海量文件堆中，针对具体业务细节进行跨文档提取与解答。 2. **知识结构化预处理**：自动解析文档并抽取关键元数据（Metadata），将杂乱的文本转化为结构化数据字段。 3. **工作流自动化流转**：借助 Agent 自主串联多步审核与审批业务逻辑。 在内容生成维度，Box 不仅在其协作产品中内置了支持无限量内容扩写的智能体，还与 OpenAI 及 Anthropic 展开了深度合作，推进复杂的企业级文档自动化排版与幻灯片（PowerPoint）生成，使数据不仅能被检索与提炼，更能直接反哺回工作成果的闭环创作之中。 ### [16:43](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1003s) 企业内容释放：从结构化提取到长时间运行 Agent Box 的核心口号是“企业生于内容，以 AI 释放潜能”（Your business lives in content, unleash it with AI）。谈及当下客户最为推崇的“全垒打”级应用场景（Hero Use Cases），亚伦·莱维（Aaron Levie）指出，对传统大型企业而言，最直接且屡试不爽的价值在于解决积压已久的非结构化文档处理难题。企业往往手握上百万份合同或研究报告，过去根本不可能雇佣足够的人力去逐页通读。除了少数像按揭贷款审批这样价值极高、不得不投入人工逐页核验的流程外，绝大多数企业数据都处于沉睡状态。如今，企业可以利用 AI 将上百万份文档中的核心结构化数据快速提取出来，存入数据库中进行灵活的查询、深度分析以及下游业务流自动化。 ```mermaid flowchart TD accTitle: 企业业务流程从传统人工向长时运行 Agent 演进 accDescr: 对比传统银行开户流程与基于长时运行 Agent 的新型自动化流程 subgraph Past ["传统企业流程"] A["非结构化文档<br/>（海量合同/研报/开户材料）"] --> B["人工逐份阅读与核查<br/>（成本过高、动辄耗时1至2周）"] B --> C["碎片化手动流转"] end subgraph Agentic ["长时运行 Agent 模式"] D["客户提交申请与材料"] --> E["Long-running Agent<br/>端到端自动化推进流程"] E --> F{"关键节点判断"} F -->|"异常或需复核"| G["人工快速介入校验"] F -->|"标准验证通过"| H["自动化完成入职/开户"] G --> H end ``` 放眼未来数年的发展，真正让亚伦·莱维感到兴奋的，是类似于“编程 Agent”逻辑在更广泛业务领域的映射——长时间运行的智能体（Long-running Agents）端到端接管复杂的业务流程。以银行的新客户开户入职（Onboarding）为例，理想状态下，Agent 会预先编排并自动执行几乎所有能自动化的步骤；只有在特定需要额外合规复核或身份验证的关键节点，系统才会跳出提示交由人工裁决。过去需要一到两周反复沟通扯皮的漫长流程，在 Agent 的驱动下可以在一个小时内迅速闭环。无论是加速客户接入、极速挖掘研发数据中的致命线索，还是实时响应安全事件，后台常驻且高度流程化的长效 Agent 正在成为下一阶段的核心形态。 - **核心价值点**：将数以百万计的沉默合同与非结构化研报，转化为可查询、可沉淀至数据库的结构化资产。 - **效率飞跃**：银行客户开户与准入流程从传统的 1 至 2 周拉扯，缩短至最快 1 小时内全自动与半自动混合完成。 --- ### [18:42](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1122s) 代码的实用主义 vs 内容的信任危机：“职场烂作”（Work Slop）之辩 针对 AI 在不同场景下的扩散鸿沟，对话转向了一个微妙而尖锐的话题：在编程领域，AI 工具（如各类代码生成助手）不仅被普遍接受，甚至被热烈拥抱；然而在 Box 所深耕的内容生产领域，AI 生成的内容却常常引发强烈的“过敏反应”，社交网络上关于 Pangram 等 AI 检测工具的声讨和对“职场烂作”（Work Slop）的嘲弄比比皆是，甚至有投资人抱怨如今收到的董事会汇报材料整篇充斥着浓重的 AI 腔调。 亚伦·莱维摘下 Box 创始人的官方身份，以一名企业经营者与内容消费者的视角解构了这种心理差异背后的深层原因： > “除了少数我们认识的顶尖工程师把写代码视为一门融合判断力与品味的崇高艺术，对全世界绝大多数人而言，代码纯粹是一种功利工具（Utility）——你只是想把某件事自动化，提供一个界面让别人按一下按钮跳到下一步。所以哪怕全套前端或后端都是 Agent 生成的，只要能用且没有漏洞，大家不仅完全接受，甚至拍手称快。” 但人类社会与大脑运转对“文字与演示文稿”的认知逻辑截然不同。在商业环境中，一份汇报材料不仅仅是信息的载体，更是评判一个人可信度、思考深度以及执行能力的“代理变量”（Proxy）。当受众一眼看出材料是 AI 批量炮制的“职场烂作”时，信任链条便瞬间崩塌：受众无法判断作者自己到底思考了多少，甚至陷入一种集体心理分裂——“我自己私下也用 AI 搞头脑风暴和辅助起草，可一旦别人端出一份 AI 味十足的东西给我，我心里立刻会打鼓：我还能信任你吗？” | 维度 | 代码生成（Coding Domain） | 商业内容与报告（Content Domain） | |:--- |:--- |:--- | | **社会属性定位** | 纯粹的工程功利工具（Utility） | 个人思考力、专业度与可信度的代理变量（Proxy） | | **受众接受阈值** | 极高：只要实现自动化、跑通逻辑即被拥抱 | 敏感易怒：出现 AI 生成痕迹容易被贴上“不负责任”的标签 | | **核心关注焦点** | 最终系统的功能可用性与安全性漏洞 | 作者本人究竟投入了多少独立思考与理解 | | **类似历史演进** | 类似标准库、编译器与现成组件的普及 | 类比财务模型中的宏与 Excel 公式，尚在建立共识的过渡期 | 这种认知撕裂在实际生活中引发了荒诞的体验。对话者提到传奇投资人斯坦·德鲁肯米勒（Stanley Druckenmiller）在《华尔街日报》发表的文章，即便在 AI 检测器中被打出 100% 的 AI 概率，读者依然因为对其本人的信赖而倍加赞赏。莱维坦言，如今在社交平台（X）上阅读长文已经背负了“双重计算开销”：既要耗费脑力理解文字本身的商业含义，还要在脑海中默默测算“这到底是他自己写的，还是把一段 Claude Prompt 原样粘贴出来的”，并据此艰难地校准对作者本人的评判。尽管有人将 AI 比作当年的电子计算器，但在商业内容制作上，这个类比正在失效——因为现代 AI 不是在做一道算术题，而是一口气接管了过去需要综合思考的十多项复合任务。 --- ### [24:08](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1448s) Harness 架构：企业文件系统与搜索认知的护城河 面对通用大模型厂商（Labs）能力的大幅提升，企业级应用层（Applied Layer）的核心壁垒到底在哪里？话题由此自然过渡到了 Box 自身构建的“Agent Harness”（智能体调用底座与环境支架）。 莱维解释道，通用的底层模型虽然强大，但在企业内部落地时，严重匮乏对特定企业环境的底层认知。Box 历经多年构建的企业级文件系统、复杂的权限层级体系（Permission Structures）以及专有企业搜索引擎，构成了极其深厚的领域知识。虽然 Box 非常乐意将自身对这套系统的理解与交互数据提供给各大前沿实验室去微调底层模型，以使外部智能体尽可能好用，但真正的核心在于“理解人类在企业内容环境中到底如何运作”——当用户在搜索结果的 10 个匹配文件中权衡时，他们内心依据怎样的隐性逻辑去挑选出真正相关的那份文件？这种在检索增强生成（RAG）、权限过滤和用户交互心智模型上的精细打磨，正是通用 API 无法直接取代的企业级护城河所在。 ### [24:59](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1499s) Box 的 Agent Harness 架构与检索精度优势 Box 团队基于自身对文件系统与文档启发式特征的深度理解，构建了一套专有的“智能体封装架构”（Agentic Harness）。这套 Harness 深度集成了 Box 的底层搜索体系和文件系统，内置了一整套工具集，能够在运行时按需提取文档纯文本、切分数据块（chunks）以及动态生成向量嵌入（embeddings）。以 Aaron Levie 自己的 Box 账户为例，经过 20 年的积累，其中存储了数以千万计的文件；如今借助 Box Agent，他可以对海量数据随意提问。系统会自动并发执行多重搜索、对结果重新排序（rerank）、闪电般提取出最核心的信息，并在需要时完整阅读全文。 这种针对自身工作流深度调优的专属 Harness，其实际表现大幅超越了直接把原始 API 交给外部大模型的做法。当 Box 将自己的 Harness 与“直接让 Claude 或 OpenAI 接入 Box 原始 API”进行对比测试时，发现无论在答案准确率还是响应延迟上，自研 Harness 都有着显著的优势。这是因为通用基础模型并不了解特定文件系统的组织规律与工作流上下文，而 Box 则清楚如何针对企业场景精细化调整每个处理环节。 ```mermaid flowchart TD accTitle: Box Agent Harness 检索与问答工作流架构 accDescr: 展示用户查询如何经过 Box Agent Harness 的多重搜索、动态嵌入重排与文档精读，最终输出高精度答案。 subgraph InputStage ["用户与数据层"] Q["用户复杂业务提问"] Repo[("Box 海量企业文件库<br/>数千万级历史文档")] end subgraph AgentHarness ["Box Agent Harness 调度核心"] Search["多路并发搜索与语义检索"] Process["动态文本切块与即时 Embeddings"] Rerank["领域启发式重排 Rerank"] DeepRead{"是否需要精读全文？"} FullDoc["全文深度分析与上下文整合"] end subgraph OutputStage ["结果交付"] Answer["高精度、低延迟回答与业务输出"] end Q --> Search Repo -.-> Search Search --> Process Process --> Rerank Rerank --> DeepRead DeepRead -- 是 --> FullDoc DeepRead -- 否 --> Answer FullDoc --> Answer ``` - **海量数据问答实测**：Aaron Levie 个人账户中沉淀了 20 年间积累的数千万份文件，Agent 可在几秒内调度多路搜索并完成综合分析。 - **自研 Harness vs. 原生 API**：直接向 Claude 或 OpenAI 开放底层 API 无法比拟专有 Harness，后者依靠自研的启发式文档定位与文件系统深度适配，在准确度和延迟上均实现质的飞跃。 --- ### [26:15](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1575s) 模型评估体系（Evals）与当前模型梯队格局 为了科学衡量底层大模型的真实能力，Box 建立了严密的评估体系，目前主要依赖两套核心评测集（Evals）： 1. **复杂工作评估集（Complex Work Eval）**：这是一套以文档为核心的高难度专业评测，覆盖生命科学、金融服务、公共部门以及高科技等垂直领域。测试设定典型业务场景（例如：“给出这 5 份专业文档和对应的问题集，模型如何解答？”），所有模型都会接入 Box Agent 进行跑分比对。 2. **留存实例评估集（Holdback Eval）**：基于 Box 内部员工在真实企业环境中使用数据的实际行为与交互记录，持续评估所有模型的演进。 Box 能够敏锐捕捉到模型在指标上哪怕“0.5 个百分点”的微小演进。在生产环境中，Box 会根据成本与准确率的综合平衡设定默认推荐模型，同时在“模型花园”（Model Garden）中把选择权完全下放给客户，允许企业自由挑选任意主流模型。在谈及各大模型厂商的角逐格局时，Aaron Levie 指出，模型在 Box 企业场景中的表现与它们的代码生成能力（Coding）高度正相关。唯一的例外是 Google Gemini——在某些特定知识工作和工具调用（Tool Use）用例中，Gemini 的实际表现甚至大幅超出了它纯写代码跑分所展现出的实力。当前最顶尖的头部梯队（包括最顶级的代码模型、Grok、Gemini 以及下一代前沿模型）整体处于极其胶着的并跑竞争状态。 - **微小增量监控**：测试流程对模型能力极其敏感，能够即时追踪到 0.5% 的性能涨跌，以此动态调整生产环境的最佳默认模型。 - **Gemini 的异军突起**：凭借出色的工具调用能力以及对通用知识工作场景的针对性优化，Gemini 在 Box 的部分业务测试中展现出了不成比例的超常水准。 - **代码能力作为风向标**：除少数例外，文档级企业复杂推理能力与模型的代码跑分（如各类前沿 Coding Eval）基本保持同步演进。 --- ### [28:35](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1715s) 客户使用习惯与开源权重模型的现实与未来 从客户实际使用行为来看，用户界面的交互搜索通常直接使用 Box 预设的默认模型，因为开箱即用且体验极佳。然而，消耗 Token 规模最大的场景并非个人端检索，而是企业后端的**工作流智能体（Workflow Agents）与批量数据抽取（Data Extraction）**。在这些场景中，企业会精打细算地跑评估，比如要求“在特定的成本预算下，数据抽取准确率必须达到 98%”。Box 甚至会派出全职工程师（FTE）深入客户的数据环境，横向对比测试五款不同模型，最终完全以实际跑分结果决定选型。 针对开源权重（Open-weight）模型的落地现状，Aaron Levie 给出了一个耐人寻味的评价：“普及度可能高于外界想象，但低于企业眼下的期望，更远远低于五年后的终极规模。”目前企业对开源模型的尝试，有超过 30% 是出于“想要尝鲜、求酷”的心理或是出于战略对冲考量——哪怕像 Gemini 或前沿闭源模型的降价轻量版本在性价比上完全可以满足需求，很多 Fortune 500 公司的 CIO 依然会执意表示“我们正在测试开源模型”。 目前开源模型在企业落地中仍存在不少现实短板：有时 Token 利用效率低下，甚至会出现推理到一半“莫名其妙突然开始吐中文”的诡异幻觉，这对于严谨的银行机构而言是难以接受的。然而长期来看，开源权重凭借纯粹的算力推理成本优势，必将逐步接管那些已经成熟、稳定的工作流。 ```mermaid flowchart LR accTitle: 闭源前沿模型与开源权重模型的双轨协同范式 accDescr: 展示复杂推理任务由闭源重型模型负责编排，而高吞吐长尾任务分配给开源轻量模型的协作架构。 subgraph Ingestion ["业务请求入口"] Req["复杂企业业务流"] end subgraph Orchestration ["智能编排层 (闭源前沿模型)"] HeavyModel["Claude / GPT 级重型模型<br/>负责意图拆解与高难决策"] end subgraph Execution ["规模化执行层 (开源权重模型)"] Worker1["开源模型 Worker A<br/>海量表单结构化抽取"] Worker2["开源模型 Worker B<br/>批量合同标准化校验"] Worker3["开源模型 Worker C<br/>海量日常长尾任务"] end Req --> HeavyModel HeavyModel -- 派发长尾低阶任务 --> Worker1 HeavyModel -- 派发长尾低阶任务 --> Worker2 HeavyModel -- 派发长尾低阶任务 --> Worker3 ``` - **开源与闭源的双重繁荣**：未来市场将呈现出一种双轨共存模式——企业用最顶级的前沿模型负责高难度的中枢编排（Orchestration），再将大量沉淀下来的长尾任务分发给廉价的开源模型；或者由轻量开源模型默认处理日常事务，仅在遇到攻坚难题时向顶级闭源模型求助。 - **Token 规模与预算分布的背离**：企业最终的 AI 预算支出可能在闭源与开源之间各占 50%，但开源模型承载的实际 Token 吞吐量将达到闭源模型的 10 倍以上。 --- ### [32:35](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1955s) 记忆机制、个性化与持续学习（Continual Learning）的前沿探讨 对话转向了 AI 的记忆（Memory）与个性化定制架构。当前主流的技术范式仍是基于检索增强生成（RAG）的上下文检索体系；即便开发者将 RAG 检索层做得极为复杂精致，本质上依然是在上下文中寻找信息，模型自身的权重参数并没有发生任何根本性的物理改变。 主持人提到，来自前沿实验室的研究人员越来越强调“持续学习”（Continual Learning）的理念，主张随着模型与用户交互的深入，模型自身的权重应当动态调整并沉淀记忆。播客节目此前曾邀请过致力于该方向的初创公司 Engram 创始人 Dan，探讨如何帮助企业客户直接将业务上下文“烘焙”（bake in）到模型权重当中。Aaron Levie 兴奋地表示自己也听了那一期播客，非常赞同其中的深度探讨，并透露自己正好马上就要与 Dan 进行一场深度沟通，共同探讨将知识直接内化于模型权重的技术走向。 ### [33:23](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2003s) 权重还是上下文：企业数据权限与模型的“中国墙”难题 在探讨将企业内部知识直接微调进模型权重（而非仅仅依赖检索增强生成 RAG）的思路时，Aaron Levie 表现出了浓厚的兴趣，但他立刻指出了学术研究理想与企业真实落地之间的巨大鸿沟。许多 AI 研究员在思考系统架构时，往往默认所有人拥有全量数据的访问权限——“我是个研究员，我能看到所有资料，如果有一个专门用我的世界训练出来的模型，那该多棒”。然而，一旦将这种架构带入现实企业，例如律师事务所，这种设想就会立刻撞墙。在律所内部，一位律师可能由于利益冲突限制（Ethical Wall / 中国墙），只能访问自己手头的五个项目，隔壁办公室的同事可能正代表直接竞争对手打官司，两堵墙之间决不能互通任何一份文档。如果每天员工的项目权限都在动态增加或撤销，针对单一员工或团队训练权重的方案在成本和权限控制上就会变得极其棘手。 针对“哪些信息进上下文（Context），哪些信息进模型权重（Weights）”的决策，Aaron 认为行业需要建立一张清晰的决策评估表（T-chart）。这取决于数据的更新迭代速率与将知识固化至权重所带来的准确率增益之间的权衡。Aaron 提到了 Andrej Karpathy 曾表达过的一个构想：如果挥动魔法棒，最理想的系统其实是剔除大模型权重中所有死记硬背的事实性记忆，仅保留纯粹的推理能力与思考风格（例如红杉资本做决策的底层哲学），再外挂一个强大的事实检索系统。然而在企业现实中，各家公司的核心竞争力往往不在于做事风格有何颠覆性的不同，而恰恰在于具体的业务专有知识（IP）——比如对某一特定法律案件的认知深度。 与此同时，这种权衡在不同行业差异巨大。例如像礼来制药（Eli Lilly）这样的医药巨头在进行药物研发时，内部并不存在严格的“政教分离”式权限隔离，他们希望尽可能把最广泛的科研数据向所有研发人员开放，这时采用针对专有领域深度定制权重的模型（如 Trajectory、Applied Compute 或 Prime Intellect 所探索的方向）就极具商业价值。 ```mermaid flowchart TD accTitle: 权重与上下文决策路径 accDescr: 展示企业在决定将知识注入模型权重还是放入上下文时的权衡机制 subgraph Decision["决策评估维度"] A["数据特征评估"] --> B{"数据更新频率与权限隔离"} B -->|"高频变动 / 严格权限隔断<br/>(如法律案件、金融风控)"| C["动态上下文检索 (RAG)"] B -->|"低频沉淀 / 全员共享专有知识<br/>(如制药研发、企业底层推理逻辑)"| D["模型权重训练 (Fine-Tuning/Pre-train)"] end subgraph Architecture["目标架构理想态 (Karpathy 构想)"] D -.-> E["纯粹推理核心与思考风格"] C -.-> F["外部动态事实检索库"] E & F --> G["高精度垂直领域产出"] end ``` - **权限复杂度**：企业级环境中最棘手的核心痛点在于动态权限与访问控制，员工每天的权限变更会直接打破静态微调模型的安全边界。 - **行业应用分化**：法律、金融等强监管、多防火墙领域严重依赖隔离的上下文检索；而生物医药研发等注重知识聚合的场景，更适合进行深度权重定制。 --- ### [37:28](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2248s) Box Labs 的研发重心：Agent Harness 与准确率“爬山” 谈及内部研发团队 Box Labs 时，Aaron Levie 幽默地调侃公司其实还没正式把这个名字大写，但它本质上就是 Box 的应用 AI（Applied AI）核心团队。该团队的工程力量主要聚焦在如何让 AI Agent 在面对特定业务场景时实现准确率的质变，特别是在 Agent Harness（外围框架与编排脚手架）层面进行算法级别的准确率“爬山”（Hill Climbing）。 Aaron 用银行业的真实业务举例：银行在审批贷款时需要处理动辄上百页的复杂贷款合同，如果直接调用市面上的开源或闭源现成基础模型，可能只能拿到 70% 的准确率，而要在商业流程中实现全自动化，系统必须达到 97% 以上的准确率。这从 70% 到 97% 的巨大跨度，决定了业务流程究竟是“无法使用”还是“完全替代人工”。Box Labs 的核心课题正是利用自动化研究（Auto-research）机制，根据客户的具体数据集构建问题拓扑图，持续测试、评估并迭代 Agent harness，直到将准确率提升到生产可用的临界点。 - **准确率鸿沟**：在企业场景中，70% 的开箱即用准确率与 97% 的企业级自动化要求之间存在巨大断层，这是应用层工程需要跨越的核心壁垒。 - **研发机制**：通过在 Agent 层面构建自动化评测与优化闭环，针对非结构化数据密集型任务（如百页合同审查）进行持续的准确率攻坚。 --- ### [39:02](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2342s) Agent 时代的记录系统：SaaS 巨头的“双重使命”与 Agentforce 现象 针对当前科技圈对“各大 SaaS 厂商都在兜售自己的专属 Agent，而客户只希望记录系统（System of Record）配合自己的第三方 Agent 工作”的质疑，Aaron 笑着抛出了“#Cloudforce”的调侃，并提到了 Salesforce 大举宣传的 Agentforce。他笑称刚看到这类营销的前三分钟觉得很滑稽，但看到股价反应后立刻觉得“这是天才之举，我们也要这么干”，尤其是当好莱坞影星马修·麦康纳（Matthew McConaughey）用其标志性的磁性嗓音念出广告词时，那简直成了向企业推销记录系统和 Agent 的“顿悟时刻（aha moment）”。 回到严肃的商业战略，Aaron 指出，对于拥有数据沉淀和业务工作流的传统 SaaS 平台而言，要想在 Agent 时代生存并获胜，必须同时做到两件事，任何只选其一的做法都注定失败： 1. **打造极致适配自身系统的第一方 Agent**：这个 Agent 必须是“针对评测极致优化（Eval maxed）”的，在操作自身平台时，必须比市面上通用的第三方开箱即用 Agent 高出 10 到 20 个百分点的表现。这并非靠给第三方设卡，而是源于自身对底层系统机制和上下文的极度精调。 2. **挖掘只有深耕领域的厂商才能洞察的隐性用例**：深耕垂直业务的 SaaS 厂商每年与成百上千家客户深度交流，能够发现通用模型开发者根本想不到的自动化痛点。 Aaron 分享了一个亲身经历的真实案例：他在一年中会与数百家客户沟通，其中至少有二十几次客户提出的场景让他拍案叫绝。例如，曾有客户提出希望在后台部署一个常驻 Agent，持续静默分析企业的所有文档，一旦发现某些文档符合合规治理策略，就自动判断是否需要将其打上法律冻结（Legal Hold）标签或移入归档库。这种极其贴合底层内容治理逻辑、在后台无声运转的 Agent，正是只有原生记录系统厂商才能创造的杀手级价值。 - **SaaS 厂商的胜负手**：必须兼顾“极致性能的第一方 Agent”与“深度理解业务的垂直原生场景”，单靠开放 API 或单靠封闭工具都无法维系竞争壁垒。 - **后台常驻 Agent 价值**：Agent 的形态正在从人机对话界面转向后台自主运行的工作流，例如自动执行合规审查与法律冻结判定。 ### [41:45](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2505s) 治理智能体与无头记录系统（Headless System of Record） 主持人与 Aaron Levie 聊到了企业内部对“治理”（governance）的态度。在很多人的刻板印象里，合规与治理往往显得死板，但对企业软件而言却蕴含着巨大的商业机遇。每家大企业都有首席合规官和治理负责人，过去他们只能自上而下进行宏观监督，根本不可能分身到每一处业务现场。如果现在能让一个 AI 智能体像合规官一样“坐在员工身边”，在员工即将做出违背公司治理政策的行为前实时提醒：“嘿，你这么做会违反治理规定”，就能在源头掐断风险，免去员工费心去猜测和理解繁琐合规条款的负担。这类直接内置于产品之中的原生智能体，能比外部工具更早、更准地发现风险，甚至完成此前在平台之外根本无法实现的操作。 与此相对，记录系统（System of Record）本身必须走向“无头化”（headless）。企业应用必须将自己的 API 开放给 Claude、ChatGPT 等各类前沿模型平台，无论是提供确定性 API 让外部智能体通过 MCP（Model Context Protocol，模型上下文协议）调用，还是直接将无头智能体暴露在主流生态中，这都是不可逆的趋势。有些人曾经对这种开放持悲观态度，担心自身价值被侵蚀，但 Aaron 认为这种担忧严重低估了新用例带来的纯增量空白市场（white space opportunities）。以 Salesforce 为例，Aaron 直言自己现在的 Salesforce 使用量比以往高出了整整一个数量级，因为他每天都在通过 MCP 从 Claude 或 ChatGPT 里频繁查询 CRM 系统的数据。 ```mermaid flowchart LR accTitle: 无头记录系统与外部智能体协同架构 accDescr: 外部通用模型通过 MCP 协议调用底层无头系统的确定性 API，并由内置治理智能体进行实时合规拦截。 subgraph External["前端模型与交互层"] A["通用智能体<br/>Claude / ChatGPT"] end subgraph CorePlatform["记录系统 (System of Record)"] B["MCP 协议接口<br/>(确定性 API)"] C["内置治理智能体<br/>(实时合规巡检)"] D["企业核心数据源<br/>(CRM / 知识库 / 文档)"] end A -->|"发起查询与操作"| B B --> C C -->|"放行合规请求"| D C -.->|"拦截违规操作"| A ``` 针对主持人调侃“记录系统是否会蜕变成收费站模式（toll booth business）”，Aaron 笑称自己并不喜欢“收费站”这个说法，因为“谁在收费站都没遇到过愉快的体验”，但背后的商业逻辑确实需要向基于用量（volume-oriented）的模式演变。记录系统的核心价值在于深度组织工作流、管理与保护数据并设定安全边界，只要真正为客户解决问题，商业化自然水到渠成。Aaron 甚至开玩笑说，他曾跟 LinkedIn 的产品经理讲过，如果 LinkedIn 能支持通过 MCP 自由调用、让他随时洞察某位 CIO 的最新动向以便跟进，“我愿意多掏十倍的钱，求你直接收下我的钱吧”。 - **实时合规干预**：传统合规只能事后抽检，而嵌入工作流的治理智能体可在违规发生前精准预警，无需员工死记硬背合规细则。 - **MCP 带来的使用量暴增**：无头化集成非但没有分流流量，反而让核心记录系统被外部前沿模型高频调用，用量激增一个数量级。 - **高商业价值数据资产**：企业不仅愿意为数据治理与工作流托底买单，甚至愿为更灵活的 API 数据接入支付数倍溢价。 --- ### [44:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2698s) 超越通用聊天框：异步后台智能体与应用层未来 当被问及通用的聊天对话框（chat box）是否会是未来 AI 应用的主导 UI 时，Aaron 给出了非常清晰的判断：通用的问答式聊天界面必然会长久存在，就像“每款产品都有一个搜索框”一样自然，横向通用产品和纵向垂直产品都会标配这种交互，用于处理单次的、即时的提问或按需生成。然而，企业运作的本质并不是即时的一问一答，而是大量在幕后持续运转的复杂流程与工作流——有的靠机器自动化运行，有的依赖人工处理，还有大量本该有人做却因成本太高而从未被执行的任务。 这就引出了完全不同的产品形态。未来的企业核心不是让员工坐在屏幕前反复与 Chatbot 对话，而是让智能体在后台全天候、异步地工作：读完每一份合同、排查每一条系统日志、分类处理每一个安全突发事件。用户最终面对的不是空白输入框，而是整合后的看板（dashboard）、工作流看板、任务排队列表（queue）与待办清单；聊天界面最多只是在需要快速追溯上下文时充当辅助抓手。如果横向通用产品试图把所有垂直领域的这些工作流、看板和按钮全都生搬硬套进一个庞大的界面里，整个软件会变得异常臃肿复杂，彻底丧失原有的轻巧与优雅。 因此，垂直应用层公司迎来了巨大的发展空间。垂直厂商深谙特定行业的业务流程，能够精准设计符合该场景的操作按钮、标签页与专业术语。无论是在法律领域已经崭露头角的 Harvey 和 Leya、在安全运维领域涌现的专用工具，还是像 Cognition 和 Factory 这类专注于长时间运行的编程智能体，都在验证这一逻辑。Aaron 由此做出了一项大胆的预测：“五年之内，企业中 90% 的 Token 消耗，都将来自人类用户从未主动手动触发的工作任务。”员工只需在早晨打开电脑，审查后台智能体已经默默处理完的一项项任务结果。 - **UI 范式转移**：从“人问机器答”的同步交互，转向以任务队列、状态看板为主的异步后台批处理交付。 - **垂直软件的不可替代性**：通用大模型界面无法承载垂直业务的沉重工作流，行业专属的标签、状态机与按钮必须由深耕领域的应用层打造。 - **五年内 90% 消耗无人触发**：绝大部分企业 Token 消耗将由自动化巡检、合同分析、日志排查等后台长程任务驱动，人类转变为最终审批者。 --- ### [48:01](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2881s) 为什么代码生成能光速渗透，而其余知识工作却步履蹒跚？ 谈及 AI 的扩散速度（diffusion），主持人提出：编程智能体在极短时间内便呈现爆发式渗透，仿佛“2026 年 1 月 1 日一到，人类历史上最快的一波技术经济渗透就此发生”，但 AI 在其他日常工作领域的渗透却明显缓慢滞后。Aaron 深入拆解了编程场景与普通知识工作之间截然不同的底层逻辑。编程之所以成为 AI 扩散的绝对先锋，核心在于代码的“效用”（utility）几乎 100% 能够被生成的文本量与逻辑质量直接度量。尽管在写代码前凝聚了无数会议、架构设计和业务知识，但最终交付的价值载体就是那份能跑通的程序文本。一个顶尖程序员如果能不吃不喝全天候编码，其价值产出与写出好代码的行数与时间高度正相关。 围绕这一点，Aaron 总结了编程赛道独占的五到六重得天独厚的优势： 1. **纯文本与可执行性**：代码的本质就是文本，其正确性有编译器和测试套件提供确定性的即时反馈。 2. **训练数据高度饱和**：前沿模型在大规模开源代码和严谨逻辑上经过了极其密集的超量训练。 3. **顶会与实验室的核心基准**：各大 AI 实验室将代码能力视作最核心的竞技指标，争相在各大公开 Benchmark 上刷新成绩。 4. **实验室员工自身就是用户**：AI 实验室的研究员和工程师每天都在用模型辅助开发，形成了日复一日闭环的自我评测（self-evals）。 5. **极其强悍的技术受众（极客自愈力）**：程序员面对 Agent 运行报错或 MCP 服务器返回“连接无效（connection invalid）”时，绝不会停下来打电话给 IT 支持求助，而是心领神会地意识到“噢，刚才那个端口没开，我去改一下就行”，他们具备当场排查解决管线故障的硬核能力。 6. **行业薪资极高、ROI 极其显著**：软件开发属于高薪技术工种，哪怕 AI 只能带来 10% 到 20% 的效率提升，其换算出的财务价值和成本节约也是惊人的，更不用说带来数倍生产力跃升时的夸张回报率。 相比之下，其他非技术工种既没有标准化的文本产出标准，也没有能快速自我排障的终端用户，导致 AI 魔法向更广泛知识工作领域的扩散必然经历更长期的适配与磨合。 ### [50:18](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3018s) 知识工作的自动化连续谱：从编程到销售 Aaron Levie 深入剖析了为什么 AI 在不同工作场景下的渗透速度截然不同。如果根据自动化属性画一张直方图，以“与软件编程的相似度”作为横坐标展开各种知识工作，法律领域与编程颇为相近：律师整天坐在电脑前审阅法律文件、撰写合同、处理海量信息，文本高度集中，价值创造直接发生在人机交互界面上，因此 AI 在法律行业的应用迅速爆发。 然而，沿着连续谱越往后走，工作模式就与编程相去甚远。以销售代表（Sales Rep）为例，销售的核心价值在于说服外部客户采购软件、技术甚至卡特彼勒（Caterpillar）重型卡车，其生产力完全受制于外部变量：客户何时回复、客户愿意今天见还是下周二见、客户今年是否有采购预算等。即便为销售配备了全世界最顶尖的自动化系统，他们依然受制于人际交互与客户决策的节奏。这就勾勒出了知识工作的完整光谱：一端是整天坐在电脑前敲击代码、极易被自动化的工程师；另一端则是高度受制于外部现实世界的销售人员。要把智能真正注入实体商业工作流中，就必须尊重各行各业特有的业务形态，做好落地变革管理（Change Management），并将数据整理并置入能够与 AI 顺畅协同的环境。 ```mermaid flowchart LR accTitle: 知识工作自动化潜力连续谱 accDescr: 展示从完全受限于机器文本的编程工作到受限于外部现实世界的销售工作的光谱 subgraph 纯数字与文本密集型 A["软件工程 / 编程<br/>(环境高度标准化，自包含)"] B["法律审查 / 文书处理<br/>(处理海量文本，人机闭环)"] end subgraph 现实与人际强依赖型 C["研发 / 跨职能运营<br/>(多系统、跨部门协同)"] D["销售与外部拓展<br/>(受制于客户预算、回复周期)"] end A --> B --> C --> D ``` --- ### [52:19](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3139s) 知识工作没有“GitHub”：企业级权限与遗留系统鸿沟 在落地部署环节，软件工程与常规企业业务之间存在巨大断层。过去开发者推出一款编程 Agent 时，甚至不需要用户经历复杂的注册流程，只需一句“把你的 GitHub 仓库授权给我们（Give us your GitHub）”即可一键接入。然而在广大知识工作领域，根本不存在通用的“GitHub”。当主持人调侃“只要交出 Box 就行（Give us your Box）”时，Aaron 幽默而清醒地指出，Box 当前的年化营收跑道（ARR）虽然达到了 13 亿美元，但这依然意味着市场上绝大多数企业并未采用 Box。 海量企业数据仍沉睡在本地机房系统（On-premises）、老旧的文件共享服务器（File shares）以及根本无法与现代 Agent 通信的遗留基础设施中。此外，两者的访问控制权限模型（Access Controls）存在天壤之别： - 工程师在 GitHub 中往往直接拥有整个仓库甚至相关服务绝大部分代码的读写权限。 - 普通企业员工每天的工作日常却是：“嘿，Sally，你能帮我把那个共享文件夹的权限开一下吗？这个项目我还没被加进去。” - 如何让 AI Agent 在遵循这种动态、细粒度且碎片化的企业权限体系下准确获取所需数据，是企业落地的巨大阻碍。 硅谷必须认清现实：AI 在传统产业的扩散（Diffusion）绝不会一蹴而就，其耗时将远超科技圈内部的乐观预期。但这恰恰孕育了“应用层（Applied Layer）”的万亿美元机遇。那些拥有耐心、深耕垂直行业专有知识（Domain Expertise），并且愿意亲自走访客户、实打实做地推落地的团队，将把技术真正交付给律师、销售、生命科学研究员或客服主管，捕获最为丰厚的核心价值。 --- ### [54:32](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3272s) Aaron Levie 的信息获取之道：全天候浸泡于 Twitter 当主持人问及 Aaron 为何几乎出现在硅谷所有热门 AI 初创公司的股东名册（Cap table）上、如何始终立于 AI 浪潮中心时，Aaron 解释这源于双重因素：首先，Box 专注非结构化数据长达 20 年，早就在等待 Agent 时代的到来，产品形态天然契合这一波技术变革，甚至在八到十年前他们就曾尝试推进类似对话；其次，他本人对 AI 技术有着纯粹的极客热爱。 被问及日常在哪里学习最前沿的 AI 知识时，Aaron 坦言除了高质量的播客（如主持人与 Dwarkesh Patel 的播客）之外，95% 的信息其实都来自 Twitter（X）。他每晚睡前都有一个固定习惯，就像网络上的悲伤表情包（Sad meme）一样，整夜不停地在信息流中疯狂滑动屏幕，多方印证比对行业动态。在他看来，Twitter 就是全球 AI 的中心广场： - 建议所有大二、大三的在校大学生立即注册 Twitter 并关注最核心的 20 个 AI 账号。 - 信息流的质量决定了你的行业认知究竟是领先别人一年，还是落后别人一年。 - 听到很多 20 岁的年轻人还在依靠“偶尔看到几篇推送文章”来了解行业动态时，Aaron 感到非常不可思议：“难道获取信息全靠别人随缘给你发邮件吗？” 针对主持人极力推荐的个人助手产品 Instinct，Aaron 坦白承认由于自己待测试的个人助手积压了三款之多，还没来得及使用邀请码体验，但他对这类产品充满期待。不过他也指出，现阶段基于浏览器操作（Browser-use）的 Agent 仍面临底层限制，整个互联网或许需要为各种产品开发专门的命令行接口（CLI），底层基础设施仍有大量硬仗要打。日常工作中，Aaron 表现得像一个典型的“AI 重度依赖者（AI-pilled）”，每天会向 4 到 5 个不同的 AI 系统抛出 20 到 30 个问题，用于行业调研、人才物色、竞品追踪以及市场拓展研究。 ### [58:34](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3514s) Box 的 AI 内部实践：非结构化数据底座与“Token 消耗榜” 谈到企业内部如何践行“AI 优先”（AI-first），Aaron Levie 指出 Box 拥有一个得天独厚的优势：公司自创立以来，几乎所有业务问题涉及的非结构化数据——无论是会议记录、项目计划、产品路线图、演示文稿，还是财务规划文档——百分之百都沉淀在 Box 内部。“我们从不让员工使用其他杂乱的工具，因此我们坐拥一套为了这种场景而天然调优的数据架构。”加上 Salesforce 等核心业务系统，Box 保持着极佳的数据治理水平，使得在顶层运行 AI Agent 变得格外顺畅。为此，Box 的 CIO 对 AI 极为上瘾（"AI-pilled"），公司不仅成立了 AI 卓越中心（Center of Excellence），还专门招聘了全职 AI 工程师（FTEs）来梳理高杠杆的工作流。 当被问及内部是否有员工 AI 使用排行的“排行榜”时，Aaron 笑称他们并不推崇盲目的“刷 Token”（token maxing）。虽然后台确实有一份按 Token 消耗量排序的员工名单，但其目的是为了业务审查与经验复盘。他坦言，排名前三的高频用户里，往往有一个人浪费了一半的 Token，而另外两个人则摸索出了革命性的工作法。两周前就发生过一次典型案例：Aaron 发现某位员工的用法极其惊艳，当即要求该团队立刻召集所有人观摩学习。短短 6 个小时后，全组人就聚齐在会议室里，那位名叫 Mick 的员工直接现场做了一场完整的 AI 工作流演示，这种即时的自发培训正是 Box 试图在全公司普及的运作模式。 - **非结构化数据归集**：Box 将 100% 的内部文档、计划和财务资料统一保存在自身平台，为构建 Agent 提供了现成的精准语料与权限边界。 - **Token 榜单的真实意图**：排查 Token 消耗并非为了考核指标，而是甄别哪些人在“低效浪费”，哪些人沉淀出了值得全公司推广的最佳实践。 - **内部 AI 推广机制**：通过 CIO 牵头的卓越中心与一线员工的高效实操演示（如员工 Mick 的 6 小时内快速落地分享），加速技能扩散。 --- ### [01:00:27](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3627s) 开发范式的快速迭代：从 Cursor 到 Slack 协同代理 在衡量产研提速时，Aaron 强调他根本不在乎代码行数增加了多少，唯一的核心指标是“是否为客户交付了更多真正需要的功能”。在 Box 的部分技术栈中，交付给客户的实际产品功能已经实现了两到三倍的增长。然而，每当他与 Anthropic 的朋友交流时，对方展示的前沿水准依然让他感叹“我们还有很长的路要走”。开发工具的“最佳版本”（meta）几乎每隔几个月就被颠覆：两年前人们还在 IDE 里使用简单的自动补全插件；紧接着公司全面推行 Cursor；而转眼之间，工程师们甚至已经不再频繁打开本地编辑器，而是直接在 Slack 频道里通过 @ 机器人代理来协作完成代码编写。 对于企业内部的“Slack 协同 Agent”，Aaron 表示 Box 已经在尝试多种形态，例如通过 Claude 的标签与代理在团队对话流中直接调用。尽管目前团队在 Slack 上的激进程度或许还没达到 Marc Benioff（Salesforce CEO）所期许的极客标准，也没完全达到 Anthropic 或 OpenAI 内部的高度，但整个产研团队正在全速朝这个方向倾斜，重新定义人机协作的组织架构与数据交互范式。 - **业务交付衡量准则**：拒绝“代码行数虚荣指标”，只聚焦于面向客户的实际功能交付速率（已在部分业务线提升 2-3 倍）。 - **工具链迭代演进**：传统 IDE 补全插件 $\rightarrow$ 引入并普及 Cursor $\rightarrow$ 转向 Slack 内以 @ 提及（@-mentioning）为主的自主 Agent 编码流。 - **工作流组织重塑**：逐步验证“Slack 协同工作代理”与标签交互机制，让开发与业务流程直接在聊天信息流中闭环。 --- ### [01:01:51](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3711s) 现代创始人的“甜蜜与残酷”：两人的团队与五倍的对手 面对当下正在被 AI 重塑的商业世界，主持人和 Aaron 都感叹未来的商业史必定会像重写《孙子兵法》一样记录这个时代。谈及当下年轻创始人与 Box 当年创业时的异同，Aaron 坦言自己的心态是“既嫉妒又庆幸”。他最近观摩了一个仅有两人的初创团队的产品演示，对方做出的系统功能极其完整、每个页面标签背后都有扎实可用的业务逻辑支撑，这在五年前或者 Box 刚起步时，至少需要 40 人的研发团队才能做出来。这种从第一天起就以 AI 为核心设计原则（AI-first）所带来的组织杠杆，令成熟企业非常羡慕。作为一家必须对海量客户数据安全和合规性负责的上市企业，Box 绝不可能像初创团队那样激进地取消代码审查（Code Review），这种坚守合规与安全的代价，客观上会让大企业承担一定的效率折损。 但硬币的另一面是极其残酷的竞争环境。Aaron 庆幸当年 Box 起步时拥有长达数年的平静期，团队可以心无旁骛地打磨产品与体验；而现在的创业者每提出一个绝妙的想法，几乎在三天之内就会冒出五个同样被 Sequoia 或 Benchmark 注资的竞品，整个赛道瞬间沦为白刃战。 ```mermaid flowchart TD accTitle: AI 时代初创企业与成熟企业的优势与制约对比 accDescr: 对比新兴 AI 团队极高的研发杠杆与激烈的外部竞争，以及成熟企业在合规与分发上的护城河。 subgraph EarlyStage["现代初创团队 (AI 原生)"] A1["极高研发杠杆<br/>(2人可做40人项目)"] A2["无历史包袱<br/>(原生 AI-First 架构)"] A3["残酷竞争红海<br/>(数天内涌现多家同行)"] end subgraph Incumbent["成熟企业 (如 Box)"] B1["合规与安全枷锁<br/>(必须保留代码审查/合规流)"] B2["数据与信任基础<br/>(承载核心非结构化数据)"] B3["客户分发壁垒<br/>(直达政企大型组织渠道)"] end A1 -->|研发降本| Comp["市场竞争白热化"] A3 --> Comp B2 --> Diff["企业级渗透优势"] B3 --> Diff ``` - **初创公司生产力跃迁**：极小编制团队凭借 Agent 能够实现相当于过去 40 人的全栈产品开发体量。 - **成熟企业不可妥协的底线**：受制于企业客户的数据安全、审计与隐私合规要求，不能盲目跟风取消代码审查机制。 - **时间窗口的急剧压缩**：过去企业有数年时间验证 PMF，如今优质赛道在几天内就会被顶级风投资本迅速塞满竞争者。 --- ### [01:03:59](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3839s) 终局判断：应用层的万亿争夺与“企业渗透”定胜负 随着构建软件的技术门槛与研发周期被 AI 大幅压扁，竞争的终局正不可逆转地从“谁能造出来”转向“谁能将它送达客户手中”。Aaron 明确表示，业内有远见的创业者（如 Scott 或 Matan 等创始人）都已经达成了高度共识：当前的核心命题是一场彻底的“企业渗透战”（enterprise diffusion play）。 在企业软件的应用层，目前存在着“真正意义上数万亿美元的商业价值等待瓜分”。单纯沉迷于技术模型或功能实现，而忽视企业级分发、合规接入与工作流融合的团队，终将输掉整场战争。未来唯一的赢家，必然是那些深刻理解如何构建复合团队，并具备将 AI 技术稳稳嵌入大型企业复杂肌体能力的公司。 - **核心重心位移**：软件生产速度的通胀导致技术实现的稀缺性降低，触达客户的“分发通道”与“部署能力”成为决定性护城河。 - **万亿应用层空间**：AI 落地创造的最大经济价值不在基础模型层，而在面向垂直业务场景、连接系统与工作流的应用层。 - **企业渗透战役的铁律**：决定最终胜负的不是谁的 API 接口更潮，而是谁能解决大型机构的合规、权限、业务习惯与组织协同，将能力深植于企业日常运转之中。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=NE4CLThMPGU&t=0s) 你怎么会用‘看一些文章’这种方式获取信息？这到底是什么意思？难道是碰巧有人给你发了封邮件附带链接，你才看一眼吗？ — Aaron Levie > [16:43](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1003s) 企业生于内容，以 AI 释放潜能 — Aaron Levie > [18:42](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1122s) 除了少数我们认识的顶尖工程师把写代码视为一门融合判断力与品味的崇高艺术，对全世界绝大多数人而言，代码纯粹是一种功利工具（Utility）——你只是想把某件事自动化，提供一个界面让别人按一下按钮跳到下一步。所以哪怕全套前端或后端都是 Agent 生成的，只要能用且没有漏洞，大家不仅完全接受，甚至拍手称快。 — Aaron Levie > [28:35](https://www.youtube.com/watch?v=NE4CLThMPGU&t=1715s) 普及度可能高于外界想象，但低于企业眼下的期望，更远远低于五年后的终极规模。 — Aaron Levie > [44:58](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2698s) 五年之内，企业中 90% 的 Token 消耗，都将来自人类用户从未主动手动触发的工作任务。 — Aaron Levie > [52:19](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3139s) 把你的 GitHub 仓库授权给我们（Give us your GitHub） — Aaron Levie > [41:45](https://www.youtube.com/watch?v=NE4CLThMPGU&t=2505s) 我愿意多掏十倍的钱，求你直接收下我的钱吧 — Aaron Levie > [58:34](https://www.youtube.com/watch?v=NE4CLThMPGU&t=3514s) 我们从不让员工使用其他杂乱的工具，因此我们坐拥一套为了这种场景而天然调优的数据架构。 — Aaron Levie ## Mentioned Aaron Levie, Doug Leone, Box, Sequoia Capital, Twitter / X, Claude, ChatGPT, OpenAI, Anthropic, Meta, SpaceX, xAI, Nvidia, Snowflake, Databricks, AWS, Google Cloud Platform (GCP), Microsoft Azure, Pangram, Stanley Druckenmiller, The Wall Street Journal, Google Gemini, Grok, Dan, Engram, Andrej Karpathy, Eli Lilly, Trajectory, Applied Compute, Prime Intellect, Salesforce, Agentforce, Matthew McConaughey, MCP (Model Context Protocol), LinkedIn, Harvey, Leya, Cognition, Factory, Caterpillar, GitHub, Dwarkesh Patel, Instinct, Mick, Cursor, Slack, Marc Benioff, Benchmark ## Video description Starting a company is hard. Reinventing your company for AI as a public company with quarterly earnings results is even harder. Aaron Levie has pulled off the transition with Box and offers hard-won advice for founders. The cofounder and CEO of Box argues the value isn't only in the model; it's in the bridge from a model's raw capability to the actual workflow inside a bank, a law firm, or a pharma company. That's the case for the application layer, and Box is building it: an agent harness tuned so tightly to its own file system, permissions, and search that it beats handing the raw API to Claude or ChatGPT on both accuracy and latency. Aaron explains why token subsidies from the labs can't last, why you want a model-agnostic company routing your tokens rather than the one selling them, and why coding diffused fast while the rest of knowledge work won't. (There's no "give us your GitHub" for a sales rep.) His prediction: within five years, 90% of enterprise tokens go to work no human user ever initiated. [Source](https://www.youtube.com/watch?v=NE4CLThMPGU)

[0:00] Aaron Levie: 你怎么会用‘看一些文章’这种方式获取信息？这到底是什么意思？难道是碰巧有人给你发了封邮件附带链接，你才看一眼吗？ [16:43] Aaron Levie: 企业生于内容，以 AI 释放潜能 [18:42] Aaron Levie: 除了少数我们认识的顶尖工程师把写代码视为一门融合判断力与品味的崇高艺术，对全世界绝大多数人而言，代码纯粹是一种功利工具（Utility）——你只是想把某件事自动化，提供一个界面让别人按一下按钮跳到下一步。所以哪怕全套前端或后端都是 Agent 生成的，只要能用且没有漏洞，大家不仅完全接受，甚至拍手称快。 [28:35] Aaron Levie: 普及度可能高于外界想象，但低于企业眼下的期望，更远远低于五年后的终极规模。 [44:58] Aaron Levie: 五年之内，企业中 90% 的 Token 消耗，都将来自人类用户从未主动手动触发的工作任务。 [52:19] Aaron Levie: 把你的 GitHub 仓库授权给我们（Give us your GitHub） [41:45] Aaron Levie: 我愿意多掏十倍的钱，求你直接收下我的钱吧 [58:34] Aaron Levie: 我们从不让员工使用其他杂乱的工具，因此我们坐拥一套为了这种场景而天然调优的数据架构。

[0:00] Aaron Levie: you're either like a year ahead or a year behind simply based on your feed. [1:54] Application companies are the hottest Neolabs. Agree or disagree. [8:27] Aaron Levie: honestly like I'm not exaggerating, like a 100,000 employees. [9:58] Aaron Levie: at some point you'll just be nationalized if you're the only thing that exists as intelligence. [18:42] Aaron Levie: When you see work slop, you're like, I'm losing my ability to know for a fact how much of the thought process was them versus how much was the AI. [24:59] Aaron Levie: meaningfully better results on accuracy and latency [29:25] Aaron Levie: randomly just like speak Chinese mid-chain, [41:45] Aaron Levie: nah, that’s going to break your governance policy [54:42] Aaron Levie: Probably unfortunately for brain cells, it's probably 95% Twitter. I have a routine where like at the end of each night I just go through the feed... I look like some sad meme probably, of just scrolling and scrolling and scrolling and attempting to triangulate all the information. [56:56] Aaron Levie: The entire internet needs just like a CLI for their product [58:34] Aaron Levie: Probably our top, you know, in the top three all AI users at Box, like probably one of them is wasting, you know, half the tokens and then two of them are like, 'Oh whatever they're doing, we need to go like do like an internal training session for everybody else.' [60:32] Aaron Levie: You're walking around, you're like, 'You're not only working from Slack, just at-mentioning bots doing your code? Like, what are you doing?' It's like we're constantly just changing what are the workflow paradigms on this. Aaron Levie: Our customers can't possibly entrust us with their data security and compliance if we don't take that seriously. So we're always going to have a little bit of a discount on the productivity because of where we are in the stack, what we do as a business.

[0:00] I want to send an emergency alert to [0:01] like everybody who's like a sophomore or [0:03] junior in college and just be like [0:05] follow these 20 accounts on Twitter and [0:08] and also join Twitter because like this [0:10] will just help your career. You're [0:11] either like a year ahead or a year [0:12] behind simply based on your feed and my [0:14] feed is like so so wired in. I'll still [0:17] talk to to 20-year-olds that are like [0:19] yeah like you know I I see some articles [0:22] and I'm like what do you mean how do you [0:23] see articles? Like I don't even know [0:25] what that means. Like do you just like [0:27] get lucky that somebody emailed you an [0:28] article? You just have to be wired in. I [0:30] I enjoy it. It's a lot of fun. I play [0:32] with everything. [0:36] [music] [0:48] [music] [0:50] Today I'm excited to welcome Aaron Levy, [0:52] founder and CEO of Box. Box is building [0:54] a collaboration platform and content [0:56] management platform for the enterprise [0:58] that has now taken on really new life [1:00] with AI. And I'm excited to chat with [1:02] you today about Box and AI, about your [1:05] general thoughts on AI because you are [1:06] such a thought leader and and about how [1:08] founders can reinvent themselves for [1:09] this AI wave. So, thanks for joining us. [1:11] >> Thanks for having me. I'm a big fan of [1:12] the the podcast. Um I religiously uh [1:16] watch every episode and [laughter] um [1:17] and so great work. I am wondering [1:19] though, I know this is a different [1:20] podcast, but when do you talk about root [1:21] canals in this? Oh my gosh. [1:23] >> Is that Have we call in Doug Leone? Have [1:26] we figured out how to weave that in yet [1:27] or [laughter] or not? I mean, I'm happy [1:29] to talk about different kind of pain and [1:30] suffering that uh that I've been [1:32] through. [1:32] >> We should just do this pod with a root [1:33] with a live root canal. Let's do it [1:34] [laughter] [1:35] >> and see how it goes. Actually, that [1:36] would it be like hot ones, but you get a [1:38] root canal and you have to talk about [1:40] your strategy while you're in a dentist [1:41] chair [laughter] [1:43] and Doug is just on the other end of it. [1:45] >> Amazing. Amazing. Doug is so pleased [1:47] with himself, by the way, right now. Um, [1:49] okay. Um, [1:50] >> has he seen every tweet? [1:51] >> Oh, yeah. [1:51] >> Okay. Oh. Oh, yeah. He's he's very [1:53] pleased with himself. [1:54] >> Good. [1:54] >> Uh, okay. Let's start with I'm curious [1:57] your take on this. Application companies [1:58] are the hottest Neolabs. Agree or [2:01] disagree. [2:02] >> Um, you know, two years ago, I think it [2:04] would have not made that much sense as [2:05] like what does that mean? Um, but but [2:07] very clearly I I think this is what [2:09] what's playing out in the market and um [2:12] it's all working out mostly because of [2:14] open source. But uh what's what's pretty [2:16] amazing is right now you know the the [2:18] whole concept of being a a sort of LLM [2:21] rapper or model rapper is actually [2:22] working out because uh what what what I [2:26] think people underappreciated was that [2:28] in the real world in the enterprise uh [2:30] what you need is some bridge from the [2:33] models capability to the actual workflow [2:35] that the enterprise has and that bridge [2:38] you basically like probably a trillion [2:40] dollars has been bet on on basically one [2:44] of of two outcomes. Either that bridge [2:46] is is very kind of limited or that [2:48] bridge is actually very vast and and [2:50] needs to be able to to you know take on [2:53] you know a lot of depth within [2:54] organizations. Um and that bet basically [2:56] looks like are you only long you know [2:59] kind of the the model itself and super [3:00] intelligence or are you long this sort [3:02] of application tier or you know maybe [3:04] previously would have been just pure [3:06] neolab um but uh but I think it's very [3:08] clearly playing out that that actually [3:10] there's a lot of gap between [snorts] [3:11] the model and the workflow and as you [3:13] bridge that gap over time you get to a [3:16] point where you realize oh I should also [3:18] do the model and and then you have [3:20] enough data you have enough sort of [3:21] domain expertise where that becomes its [3:23] own you know sort of flywheel So I'm [3:25] very bullish on this um on this idea and [3:28] what's cool is it's opening up like [3:29] multiple layers of of you know sort of [3:32] opportunity for startups because you [3:33] could either be the actual applied [3:35] company itself i.e. the Neolab or you [3:37] could be the infrastructure provider to [3:39] the Neol and you have like multiple [3:41] layers of of you know going and [3:42] attacking that space but uh I think a [3:44] huge update for for the market and [3:46] everybody's kind of view of it. [3:47] >> Box Labs let's go. [3:49] >> We we already have it. It's a little bit [3:50] it's a little bit secret but uh we're we [3:52] we we pay very close attention right [3:54] now. the main focus is let's make the [3:56] agent really really good on any uh model [3:59] but over time obviously you would peel [4:01] off certain use cases um either on a per [4:03] customer basis or or kind of across the [4:05] the whole data set [4:06] >> yeah I mean it seems like there's two [4:07] forces that are happening one is people [4:10] don't want the fox to be guarding the [4:11] hen house they don't want the seller of [4:13] the token to be the one that is also [4:14] metering and gating you know what is the [4:16] best token for each use case yeah [4:18] >> uh and then the second is you know [4:19] there's actually a lot of work to do on [4:21] that bridge to cross and there's real [4:22] research involved in it And um it's [4:25] pretty bespoke to the exact uh the exact [4:27] workflow and the exact end customer that [4:28] you have. [4:29] >> Yeah. I think I think what what I tend [4:31] to see happen in the valley is and and [4:33] for very good reason and it's actually [4:35] why these companies have been so [4:36] successful is like everybody is so kind [4:37] of research pilled uh which is again [4:40] totally awesome big fan. It's it's led [4:42] leads to all these breakthroughs. Um but [4:44] the the sort of sense that okay the [4:46] model and the intelligence in the model [4:48] is kind of the only form factor that [4:50] matters. And then you go to the real [4:52] world and you you sort of see how [4:54] intelligence actually gets rolled out in [4:56] people's workflows and the model could [4:58] be the most intelligent you know super [5:00] intelligence in the world but that [5:02] workflow still requires you to connect [5:03] up to other data systems still requires [5:06] these moments where there's a human in [5:07] the loop interaction. There's delays in [5:09] the workflow and so the the sort of [5:11] agent has to sit idle. Um there's change [5:13] management of the actual business [5:15] process. There's legacy systems that [5:16] haven't been modernized. So you kind of [5:18] go through these five or 10 things that [5:20] are way sort of, you know, much more [5:23] operational, much more blocking and [5:24] tackling than just the pure super [5:26] intelligence of the model. And the last [5:28] thing that I think a classic sort of [5:29] research organization wants to go do is [5:31] go attack every single one of those [5:33] things. And this is sort of not a [5:35] there's like this is not a sort of a [5:37] one-off in history. Like we've always [5:38] had this relationship between [5:40] infrastructure and application. like you [5:42] know obviously AWS or GCP or Azure have [5:45] created you know trillions of dollars in [5:46] value of of kind of market cap of [5:48] infrastructure but guess what there's [5:50] also trillions of dollars of value in [5:51] software that only exist because of that [5:54] infrastructure and if you were to go [5:55] back [5:56] >> you know 10 years ago and you were to [5:57] look at what was happening in the data [5:58] space and you looked at what what you [6:00] know GCP early kind of versions of GCP [6:02] was building or or AWS is building I [6:05] guarantee you would not have predicted [6:06] Snowflake or data bricks existing you [6:08] would have been like the infrastructure [6:09] just already does that why would you pay [6:10] another $10 billion of revenue to all [6:13] these other products that are just [6:14] making it so you can work with your [6:16] data. The same thing is going to be true [6:17] for intelligence which is the models [6:19] will be insanely valuable but the [6:20] application of bringing those models [6:22] into real workflows in banking and life [6:24] sciences and healthcare and government [6:26] that's just going to be a lot of [6:27] software. Now the sort of challenge for [6:30] the next let's say two to five years is [6:32] how much do the model providers need to [6:34] move up that stack and also try and you [6:36] know go and compete at that layer or do [6:38] they either leave open intentionally or [6:40] accidentally that entire space for the [6:42] application sort of ecosystem and [6:44] actually to some extent this is like a a [6:46] big question strategically for them [6:47] because on one hand you want to be [6:49] closer to the customer on the other hand [6:51] you also want to be able to have an [6:52] ecosystem so people trust you so so [6:54] that's going to be like a really [6:55] interesting tension over the coming [6:56] years. [6:56] >> How do you think that'll play out? I [6:57] feel like that's the question we spend [6:58] every day wrestling with. [7:00] >> Yes. Uh you can Yeah. You know, I don't [7:02] want to speak for for Sequoia, but you [7:04] can sense some of the existential dread [7:06] of a VC right now is like, should I just [7:08] put another billion dollars into [7:09] Enthropic or, you know, should I attempt [7:12] to sort of see what what at the applied [7:14] layer is going to is going to play out. [7:15] So, nobody's, you know, that envious of [7:18] uh of your guys position having to to [7:20] figure that out. But but obviously, it's [7:22] even harder for the entrepreneur. Um uh [7:24] but you know I I am you know I'm pretty [7:27] long obviously the application layer [7:28] like I'm also equally very biased like [7:31] I'm I'm I have a very concentrated bet u [7:33] with very limited diversification on uh [7:36] on it working out that you still want to [7:37] buy technology that that sort of [7:39] understands your workflow and can get to [7:41] the the core enterprise data. But I [7:44] don't know that that there's I I just [7:46] don't see a different event happening [7:48] than the than all of history. I mean, [7:50] you know, we only have like 50 or 60 [7:52] years of of of computer software [7:53] history. Um, but basically, when you go [7:56] to that law firm and you go to that, you [7:58] know, pharma company and you go to that [8:00] bank, they need something that bridges [8:02] the core technology to their workflow in [8:04] their business process. And AI has not [8:06] meaningfully sort of changed the need or [8:08] the shape of what that looks like. And [8:10] the and the best sort of manifestation [8:12] of this is um sort of this chatbot [8:15] versus kind of agentic workflow, you [8:17] know, demarcation. So the chatbot can be [8:20] totally universal and can be totally [8:21] horizontal, but you know, all of a [8:24] sudden you kind of look at that and you [8:25] say, "Well, my workflow kind of needs [8:27] something to kind of ping me at the [8:28] right time in the process or it needs [8:30] access to a certain kind of data that [8:32] that just, you know, the the chatbot [8:34] can't natively get access to. So [8:36] somebody has to go into that [8:37] organization like get it set up and [8:38] somebody has to go and provide domain [8:40] expertise to this model so it really [8:42] understands our particular business [8:44] process." Um, and and so then unless you [8:46] really just underwrite, you know, the [8:48] big labs at at honestly like I'm not [8:50] exaggerating, like a 100,000 employees, [8:53] if you don't underwrite that, then then [8:55] the diffusion kind of economy is going [8:57] to be massive because every single one [8:59] of those companies, whether that's a 50 [9:01] person firm or certainly a multiund,000 [9:04] person firm, is going to need an army of [9:05] people to go in and help them with that [9:07] transformation, that change management. [9:09] And so this is, you know, like whether [9:10] it's the FTE phenomenon or just again [9:12] understanding that domain expertise that [9:14] that's going to be a very big deal. And [9:15] then you you kind of alluded to to this [9:18] point um uh Fox kind of guarding the hen [9:20] house and um uh and that that might [9:23] actually be singularly the biggest sort [9:25] of reason this has to happen which is [9:27] even even under all sort of like [9:29] complete benevolence and like nobody's [9:31] actually doing something um uh in any [9:34] kind of like you know sneaky way it just [9:36] stands to reason that if I'm going to [9:37] sort of give a task to an to an a an [9:39] agentic system I just want that task to [9:41] be cost optimized like with accuracy as [9:43] as kind of holding constant [9:45] And so who can do that? It's it's the [9:47] company that doesn't care among you know [9:49] 10 different models which model is is [9:51] performing that task. So like [9:52] definitionally you would want that to be [9:54] the the the the company that does not [9:56] have a sort of a preference. [9:58] >> And then the and then the force you have [9:59] going against that is that the model [10:00] companies can subsidize their their [10:02] models or offer them at different [10:03] different rates. [10:04] >> I don't know how long that lasts though [10:06] because when when these companies become [10:07] public, I think they will be held to [10:10] basically the same laws of of capitalism [10:12] that everybody else is. So, so the [10:14] subsidization is working very well up to [10:16] a certain threshold of spend and we [10:19] might have we may have exceeded that [10:20] spend when you're at like tens of [10:21] billions of dollars of of kind of [10:23] capital and in fact if anything it might [10:25] even be worse because uh eventually you [10:27] have to go in and sort of pay for your [10:29] training runs as well. So like they they [10:32] like the subsidization of tokens is a I [10:34] I think it just has to be a temporary [10:36] phenomenon. [10:37] >> You mean from a gross margin perspective [10:39] or from like an antitrust perspective? [10:41] >> Entirely gross margin. Like if I [10:43] >> but they have such high gross margins on [10:44] inference right now. [10:45] >> But then what are they subsidizing the [10:47] like then then it's actually then [10:49] they're then they're charging actually [10:50] like you know decent rates. [10:52] >> Yeah. But they can afford to price the [10:53] API higher than their own first party [10:55] products. Right. [10:56] >> So that part to totally fair. You know [10:58] there's an interesting kind of you know [10:59] dimension which is well like if API is [11:02] the high margin thing that is sort of [11:04] paying for the subization but you've [11:05] moved all your customers over to the [11:07] applied product and there's no API [11:08] revenue. So like there is like an [11:09] equilibrium you have to strike with [11:11] this. [11:11] >> Totally. Um and um and so then all the [11:13] while if you have some sort of either [11:16] non-economic actors or people just with [11:19] a totally different game theory in this [11:20] you know Meta being one maybe SpaceX [11:22] being one China certainly being a giant [11:24] one even Nvidia being one like that [11:27] changes the calculus as well which is [11:29] those four kind of cohorts don't [11:32] necessarily need to make money on on [11:34] inference in the in in sort of at least [11:36] at the same margin structure at like [11:38] what anthropic or openi needs. So they [11:40] would might be fine to bring down [11:42] inference to 10% margin because they [11:44] just want to basically you know pay for [11:46] massive compute clusters and then as [11:48] long as that happens and as long as like [11:50] there's not insane proprietary sort of [11:52] you know you know sort of closely held [11:54] secrets then no matter what you're going [11:56] to have kind of cost per token go down [11:58] on a on a like forlike basis all of [12:00] which means more value acrru to the [12:02] application layer which is I think like [12:04] a very long-winded way of saying I think [12:06] there's just value in kind of everybody [12:07] in the stack. I just don't know that [12:09] that the I I the only thing I probably [12:11] wouldn't bet on is just okay, one or two [12:14] labs get 95% of the value creation. I [12:17] think there's just going to be a much [12:18] more dynamic environment. And honestly, [12:20] if I were one of the, you know, two or [12:21] three biggest labs, I think I'd prefer [12:23] this outcome too because back to your [12:25] antitrust point, like at some point [12:26] you'll just be nationalized if you're [12:28] the only thing that that is is sort of [12:29] exists as intelligence. So, you kind of [12:32] want a little bit of healthy competition [12:33] in this ecosystem anyway. [12:34] >> Yep. Totally. Um let's let's transition [12:37] talking about bucks. I'm going to come [12:38] back to talking about tokens and Jeban's [12:39] paradox and China and all these all this [12:41] stuff, but let's let's talk about Box [12:42] for a second. [12:42] >> I love that topic, too. [12:43] >> Um, give folks I mean I'm guessing a lot [12:46] of people that listen to this podcast [12:47] use Box, but give give people like a [12:49] brief kind of explanation of the history [12:51] of Box and how how you're reinventing [12:53] yourself with AI. [12:54] >> We started the company as a way to be [12:55] able to kind of securely store and share [12:57] data in the cloud. Um and um uh and it [13:02] was a very simple idea, but but we we we [13:04] kind of just kind of cracked a a nut or [13:07] struck a nerve. Um and um that you know [13:10] for for Doug out there uh we uh we were [13:13] able to sort of scale up quickly. We [13:15] pivoted rapidly in the enterprise and [13:17] the idea was enterprises would be moving [13:20] from on- premises systems to the cloud [13:22] and they would need a a better way to be [13:24] able to sort of store, share, [13:25] collaborate, manage their all of this [13:27] unstructured data, their their corporate [13:28] documents, their financial documents, [13:29] their marketing assets, their research [13:31] materials in the cloud securely. So that [13:33] was the the kind of company and we we [13:35] had been sort of flirting with AI kind [13:38] of products and and sort of experiences [13:40] really since like 2015. um if you [13:43] remember like the first kind of rise of [13:45] the you know maybe the the at least in [13:47] modern times the the AI winter that [13:49] happened in like the 2015 to 2018 period [13:51] which is like we think it's going to [13:52] happen now and then it didn't um and uh [13:56] uh that was a period where we were like [13:57] okay these very you know sort of early [14:00] AI models were showing us signs that [14:02] okay if you if you you know looked at an [14:03] image and you could classify the image [14:05] well that's pretty useful if you're in [14:06] an enterprise because now maybe you take [14:08] all of your image data and sort of label [14:10] it or maybe you'd OCR something and you' [14:12] be able to kind of pull out, you know, [14:13] kind of, you know, sort of the the the [14:15] text in there. That's enormously [14:17] helpful. The problem was is insanely [14:18] expensive and you had to have a model [14:20] for every single use case required, you [14:23] know, kind of a hyper trained model for [14:24] each workflow that you wanted to do. So, [14:26] we kind of shelved it. Um, you know, few [14:29] years later, uh, you know, started [14:31] paying attention to the GPTs. We had [14:32] some hackathons where people were like, [14:33] "Oh, we could do like, you know, type [14:35] ahead in one of our kind of, you know, [14:36] note-taking products." And, and that was [14:38] that was sort of early kind of versions. [14:40] We did some some early work in in sort [14:42] of text detection and classification [14:45] which helped with security use cases. [14:47] Then obviously catchb moment sort of [14:49] hits and um and that was the the big [14:51] sort of you know head exploding moment. [14:53] If for no other reason why you know then [14:55] then they kind of figured out a form [14:56] factor that opened up everybody's mind [14:58] to oh these could be these interactive [15:00] systems that you just like ask a [15:02] question get an answer back of of sort [15:05] of you know kind of increasing [15:06] complexity and length. Yep. [15:08] >> So we looked at that very quickly. We [15:09] jumped all in. We sort of did the whole [15:11] company pivot like like everything was [15:14] was kind of exactly like academically [15:16] what you should do. We had a team carved [15:17] out. We we put you know the best people [15:19] on the team. We we you know met every [15:21] day looked at the updates and then [15:23] slowly but surely sort of built out what [15:25] what today is is kind of our you know [15:27] our AI stack and then basically the box [15:29] agent and um and for us you can imagine [15:31] the use case is is you know very [15:33] straightforward. We sit on hundreds of [15:36] billions of files. Every single one of [15:38] those files contains critical [15:39] information for an enterprise. That [15:42] could be their contracts, their research [15:43] files, their marketing assets, their [15:45] loan documents, like all of this [15:47] critical information. The problem is [15:49] they rarely know what's actually inside [15:50] of it. So they they they unless you [15:52] literally look at the document and and [15:54] kind of search and find it, you just [15:55] don't know what's inside of it. So now [15:57] agents can go and basically be farmed [16:00] out to go and answer questions about [16:02] that data. they can pre-process it and [16:04] extract metadata from those documents [16:06] and turn it into structured data. You [16:07] can use agents to automate sort of steps [16:10] and workflows. So, we've built a [16:12] platform that basically lets you deploy [16:14] agents against all of that unstructured [16:15] data. And that's been the the kind of [16:17] core focus. [16:17] >> Are you using agents to create new [16:20] content? [16:20] >> We are. Um there's a couple modalities [16:22] where that shows up. One is we have [16:24] again an online sort of collaborative [16:25] product that that an agent can just like [16:27] generate any amount of content in it. [16:28] And then um and then we've done most of [16:30] the I think probably more exciting work [16:32] with with OpenAI and uh Enthropic on [16:36] just how do you do like advanced [16:37] document creation, PowerPoint creation. [16:39] We've we've decided that their tech is [16:41] is you know at this point can always be [16:43] frontier. So so we have an agent that [16:45] goes and interacts with those systems uh [16:47] to produce you know a high quality [16:49] PowerPoint etc. [16:50] >> Awesome. Your tagline your business [16:52] lives in content unleash it with AI. So [16:55] what are the hero home run use cases for [16:56] how people are unleashing it today? [16:58] Yeah. [16:58] >> And then if you had to fast forward a [17:00] few years, what do you think people will [17:02] be doing with AI in your products in a [17:04] few years? [17:04] >> Yeah. So the the probably the easiest [17:06] hero for again more of a traditional [17:08] enterprise to think about is is just as [17:10] simple as you have a million contracts. [17:12] Why don't you find out what's inside [17:14] them? Or you have a million research [17:15] documents, be able to go and pull out [17:17] all of the critical structured data, put [17:19] that into a database, and then be able [17:21] to query, analyze, automate workflows [17:23] around that. So that that's kind of the [17:25] thing that that just knocks it out of [17:26] the park every single time because it's [17:27] been a long-standing problem that people [17:30] have never been able to go and and and [17:32] sort of apply human, you know, sort of [17:34] labor to because it's just too expensive [17:35] to read every every contract, every [17:37] research document. You know, maybe you [17:39] could do it if you had like a loan [17:40] document process, but most other data [17:42] just never gets read at that at that [17:44] scale. And then I think the the stuff [17:46] that that we're probably you know uh as [17:49] much if not more excited by is is really [17:52] the equivalent of what we see with let's [17:53] say coding agents or other other other [17:55] complex agents which is you have [17:56] longunning agents that are just [17:58] executing your entire kind of workflow [18:00] or process. And um and this would be in [18:02] the form of you go to a bank and you're [18:04] onboarding at a bank and they've [18:06] basically like automated every step that [18:08] is possible to automate and then sort of [18:10] jumps out to a person in the in the [18:12] steps in the process for extra review or [18:14] extra verification. But now instead of [18:16] that sort of one or two week back and [18:18] forth, it just happens in like an hour. [18:20] Like that's the dream state of most of [18:22] these enterprise workflows is what if we [18:24] could onboard a client faster? What if [18:26] we could discover kind of critical data [18:28] inside of our research much more [18:29] quickly? What if we can alert to a [18:30] security event much more much more [18:32] quickly? So to do that, you need these [18:34] sort of background agents or workflows [18:36] that are that are sort of [18:37] pre-established uh for those processes. [18:39] >> Totally. It's the year of the long [18:41] running a agent. [18:42] >> It is. It is. [18:42] >> Yeah. Yeah. Um I'm curious like you made [18:45] the analogy to cloud code. It seems to [18:46] me that in the coding domain, uh using [18:49] AI is like not only accepted, it's [18:51] embraced. [18:51] >> Yes. [18:52] >> Uh in the content domain, which is I [18:54] think a where a lot of the content in in [18:56] box sits. Yep. um using AI to produce [18:59] content at least. It's just like there's [19:01] this almost this allergic reaction to it [19:02] like all the panggram stuff on Twitter. [19:04] There's the um you know it's like this [19:07] concept of work slop. I'm curious what [19:09] you think about work slop and like will [19:11] this still be a thing in a few years? [19:13] >> I I'm going to sort of separate the box [19:14] corporate hat and just now kind of maybe [19:17] riff as a as a as a as a consumer of of [19:21] >> gosh I wish there was a better term but [19:22] work slop as you know inside of an [19:24] enterprise context. Um, [19:26] >> I get board decks that are entirely [19:27] written by AI these days and it kills [19:29] me. [19:29] >> So here here's the difference I think uh [19:31] on the acceptability. [19:33] Um, so there's probably like more um [19:37] symbolism to this actually topic than [19:40] than than just like the this the slop [19:43] element, but like actually like [19:44] diffusion of AI in general sort of [19:45] almost ties to this code like other than [19:49] you know the the top engineers that we [19:51] hang out with that like are like they [19:53] have you know deep taste in the code [19:55] like like you know and and the judgment [19:56] is incredible and and like it is them as [19:58] much an art as is a science. So take [20:01] that group aside. For most of the world, [20:03] code is a utility. Yeah. It is just [20:05] trying to accomplish something. You [20:06] you're just trying to automate [20:07] something. You're just trying to put a a [20:10] sort of interface up there that somebody [20:12] presses a button and moves to the next [20:13] step. So for most of the world, the [20:16] value creation of code has been to [20:19] automate things and to be able to have [20:21] it as a utility. So, so at the end of [20:24] the day, uh, like like we're and we'll [20:27] probably still use the term slop for a [20:29] while because because there's taste in [20:31] kind of front-end design and there's [20:33] taste in in sort of systems and you [20:35] don't want to have vulnerabilities in [20:36] your code. So, that's going to exist for [20:38] for a while. But, at the end of the day, [20:39] if you can tell an agent like, please go [20:41] and generate my entire backend system or [20:44] my front-end system, like it's just it's [20:46] not only acceptable, it's it's [20:47] preferable because it's just like that [20:48] was the thing that was blocking us from [20:50] moving forward. So, we need to go do [20:51] that. [20:52] At least the way society functions and [20:54] the way the world works and our brains [20:56] work at the moment. Maybe this changes. [20:58] You know, when you when you get a [20:59] presentation from somebody, there's [21:01] still this association which is like I'm [21:03] trying to decide if I can trust that [21:05] person to go and execute on that thing [21:07] or deliver that result or understand [21:09] that topic. And so if when you see work [21:11] slop you're like I my I'm losing my [21:14] ability to to sort of know for a fact [21:16] that like like how much of the of the [21:19] thought process was them versus how much [21:21] was the AI? How much should I even care [21:23] about that because I myself am doing the [21:25] same thing. So like we have this weird [21:27] like it's a it's this very weird sort of [21:29] like collective issue that we have which [21:31] is like which is like I'm doing work [21:33] slop for some of my you know brainstorms [21:35] and decisions but when I get it from [21:36] somebody else I'm like hm should I trust [21:38] you? uh and uh and I don't know I you [21:41] know I mean it just might be a thing [21:43] that as a society we have to kind of [21:45] keep cranking through over the next kind [21:47] of three to five years and and end up at [21:49] the other at the other side like like [21:51] you know I hate to use like these like [21:52] totally busted analogies but you know [21:54] obviously you don't care when you see [21:56] somebody's financial model you're like [21:58] yeah that was generated clearly by like [22:00] a macro uh or or you know that was like [22:02] not you did not personally go and [22:04] compute all of that but but you're [22:06] showing it to me and we're talking about [22:07] it So, why can't the same exist for a [22:10] strategy deck or or whatnot? But I think [22:12] right now we're going through this [22:13] evolution of like what is the person's [22:16] role? What is the content a proxy for? [22:18] Is it supposed to be a proxy for how [22:20] much that person knows? Is it a proxy [22:22] for what we think that they can go and [22:24] execute on? I think I think we're just [22:26] in this very messy period where we have [22:27] to kind of figure that out. [22:28] >> Totally. Um, did he read the Stan Ducken [22:31] Miller Wall Street Journal? I I read the [22:33] uh I read the discussion about it. [22:35] >> Yeah. The [laughter] the uh the reaction [22:37] to it, but I actually I I didn't I [22:39] didn't read it. But was it like very [22:40] sloppy? [22:40] >> I didn't think it was love. I loved it. [22:42] And so to me, it was just a nice counter [22:44] example of I have this like allergic [22:46] reaction. [22:46] >> How many how many it's not exits wise [22:48] were in there? [laughter] [22:49] >> I don't think there were but it [22:50] [clears throat] but it does show up as [22:52] 100% AI in in Pangram. [22:53] >> Okay. And dashes [22:55] >> there. I think they're okay. You can't [22:57] do that. Yeah. Yeah. But it's like it it [22:59] was a nice counter example to me because [23:01] normally I read something that's clearly [23:03] written by AI and I just have this [23:04] allergic reaction whereas with the Stan [23:06] piece I didn't. Yes. And I'm not sure [23:07] how much of that was just you know it's [23:08] Stan therefore I trust in Stan versus [23:10] >> Yeah. No but it's it is psychologically [23:12] kind of like weird because I I'll read [23:14] these like you know um X articles and [23:17] I'm like now like doing 2x the amount of [23:19] work to read these. I'm reading it one [23:21] for the substance and I'm also reading [23:22] it two for the calculation of like did [23:25] the person write it or am I just [23:26] literally reading like a claude prompt [23:28] and then like I'm like I'm literally [23:30] like my mental processing is now like [23:32] should I now does that upweight or low [23:34] or like lower my my sort of judgment of [23:36] the person or the post and I think we're [23:38] yeah we're in for some weird times [23:39] because of this. I would hate to be a [23:41] college professor. I would just I would [23:42] totally quit because you're just like [23:44] you're like I don't I don't know anymore [23:45] what you did. I I like what what's [23:47] >> it does seem like the calculator is the [23:49] closest analogy though. [23:50] >> Yeah, it except it's just like that was [23:52] like more like finite in terms and you [23:55] still had to piece together so many more [23:56] things [23:57] >> like we like the and some of these [23:59] analogies are breaking down of like it's [24:01] just a task because it's like well at [24:03] some point like this thing is doing like [24:04] at least 10 tasks at once. Yeah. [24:06] >> Um but uh but yeah, [24:08] >> let's talk about harnesses. How does the [24:10] uh how does the box [24:11] >> Great transition to harnesses. [24:12] [laughter] Okay, [24:13] >> speaking of calculators, let's talk [24:15] about harnesses. Yeah, [24:16] >> this is back to the kind of Neolab kind [24:18] of applied applied layer. Uh there's [24:22] there's a bunch of things that that we [24:24] know about our file system, permission [24:27] structures, our search our our search [24:29] engine that that you know certainly and [24:32] by all means we actually would love the [24:33] the we'd love all the labs to train on [24:36] our understanding of this because it [24:37] would only make external agents you know [24:38] as good as possible. Yeah. like like we [24:40] we we always talk to labs like hey we'll [24:42] give you as much data as you want about [24:44] about kind of how the system works but [24:46] but let's just say like that aside we we [24:48] have a lot of depth of understanding of [24:50] what do people do in box how do they [24:51] search box how do they decide when they [24:54] look through 10 files which is the one [24:55] to go pick how do what what is the in [24:57] their internal kind of calculus or [24:59] heristic on on sort of figuring out the [25:01] most relevant document to look at. So we [25:03] know all of that and we basically you [25:06] know have have built an agentic harness [25:08] that that you know attempts to kind of [25:10] understand that set of domain [25:12] understanding about our system. Uh it it [25:14] obviously has access to our search [25:16] system our file system. It has a a bunch [25:19] of kind of mechanisms for just pulling [25:21] out just the text of a document just [25:23] pulling out chunks from the document [25:25] doing embeddings on the document on a [25:27] fly. So there's a set of kind of tools [25:28] that it can use. Um, and effectively [25:31] it's a harness for asking questions of a [25:33] large data set. So in my box account I [25:35] have I don't even know the latest number [25:37] but on the order of tens of millions of [25:39] files. Um, just because it's like it's [25:40] like every everything that has ever kind [25:43] of accumulated over over 20 years. But I [25:45] can now ask any question of of all that [25:47] data set using the box agent and it it [25:50] goes around it does it does a multiple [25:51] searches in one. It reranks it. It then [25:54] very quickly sort of pulls out the most [25:56] relevant information. And then it'll in [25:57] some cases read the full document, you [25:59] know, does all the steps [26:00] >> and and then we compare that against [26:01] like, well, what if we just gave, you [26:03] know, claude our API or gave OpenAI our [26:05] API and we see like meaningfully better [26:07] results on accuracy and latency because [26:09] again, we kind of know exactly how to [26:11] how to tune it for our workflows. Uh, so [26:13] that that's effectively the the harness [26:14] that we built out. [26:15] >> And then what EBLs matter the most to [26:17] you? [26:17] >> Um, so I have a couple I have a couple [26:19] like, you know, just funny personal ones [26:21] that I just keep track of of like my own [26:22] use cases. But we do we have I don't [26:24] know hundreds of different tests that we [26:26] do on every single model. Uh we actually [26:28] have two evals at the moment. One is uh [26:30] we put out a thing called the complex [26:31] work eval which is a which is a set of [26:34] domain specific work in life sciences [26:37] financial services public sector uh tech [26:40] etc. And it's it's kind of exactly what [26:41] you'd think of as a as a document [26:43] ccentric eval. So given these five [26:46] documents and this set of problems like [26:47] what would your answers be? And we test [26:49] every single model against those with [26:50] our agent. And then we have a hold back [26:52] eval which is um actually the first one [26:54] is hold back also but but the second one [26:56] is just like then our box instance and [26:59] how box employees use use their data and [27:01] then we eval every model again on that. [27:03] Um so we're able to kind of roughly keep [27:05] track of of of all the incremental [27:07] progress like we we see when things move [27:09] by half a point uh in terms of model [27:11] improvement. Um and then we roll out [27:13] sort of default models um based on [27:16] different kind of cost and accur [27:18] accuracy thresholds. And then we let [27:19] customers also choose any model they [27:21] want from effectively our model garden. [27:23] >> What's your current view of the race and [27:25] like where all the horses are in terms [27:26] of model performance on your use case? [27:28] >> They more or less closely correlate code [27:31] with one exception which is actually in [27:33] some of our use cases Gemini uh is [27:36] disproportionately better than than what [27:38] you would see from coding. and uh and it [27:41] might be, you know, sort of just better [27:43] tool use. Um, you know, given given kind [27:45] of, you know, the the the Gemini [27:47] ecosystem and what they need to build [27:48] for um it it it solves, you know, kind [27:51] of a a strong set of sort of general [27:53] knowledge work use cases as well. Um, [27:55] but I think mostly correlating to to [27:57] kind of code. So, Fable 5.1 was clearly [28:00] kind of, you know, state-of-the-art um [28:02] and the best model that that we've seen. [28:04] There's obviously rumors about other [28:06] models and so we'll see how the kind of [28:08] race y you know kind of you know [28:09] continues on this on this front but but [28:11] basically by and large like when you [28:13] look at GDP vow um Merkore has their [28:16] Apex uh eval uh these things will all [28:21] generally follow the the coding uh [28:23] models and and so I think it's we're [28:25] just neck and neck on like gro muse the [28:28] fable class and uh you know GP56 [28:32] whatever they're building next like [28:33] these are just It's a total race right [28:35] now. [28:35] >> And do your customers typically express [28:37] a preference on which model they want to [28:38] use or do they just use your default? [28:41] >> So they they you know kind of by volume [28:43] they use our default because it's just [28:45] easy and and it works extremely well and [28:47] it's tuned for uh there's a a few ways [28:49] our our kind of agent manifest. So like [28:51] the way that you'd most commonly [28:53] experience it as an end user is you [28:55] would just be searching and and kind of [28:57] you know asking questions of your data. [28:59] the but by volume the the volume of [29:01] tokens tends to go through more of our [29:03] workflow agents or data extraction. [29:05] That's where actually you you have [29:06] customers actually doing eval basically [29:08] saying okay I want to really make sure [29:11] that at this cost profile I can get 98% [29:13] accuracy on data extraction and that's a [29:15] place where like we'll have an FTE that [29:16] goes in and helps you you know helps [29:18] understand your data environment tests [29:20] against five different models and then [29:22] you're just basically at the mercy of of [29:24] the eval. [29:25] >> Yeah. What are you seeing in terms of [29:26] the adoption of openw weight models in [29:27] your customer base? [29:29] >> Um, so probably higher than people [29:32] think, [29:33] lower than what enterprises actually [29:36] want and much much much much much lower [29:39] than what it'll be in five years. So, so [29:41] like some mix of that would be the kind [29:43] of message like [29:44] >> and is primarily cost that's driving [29:46] that decision? I I you know I have to [29:48] probably attribute 30 plus percent to [29:50] just kind of the the sexiness of of of [29:53] like [29:54] >> I want to try GLM. [29:55] >> Yeah. Yeah. I think there I think [29:56] there's that like I've heard [29:57] >> I've heard you know CIOS of Fortune 500 [29:59] companies say we're we're playing with [30:00] open source here and I look at that and [30:02] be like well I know for a fact like [30:04] >> like Gemini or Muse would have been just [30:06] fine at that particular cost profile [30:07] that you're trying to do or probably [30:08] even like you know you know 56 Luna or [30:11] Terror or whatever whichever you know [30:13] one had the had the crazy discounting [30:14] they just did like it probably would [30:16] have been totally fine but you want to [30:17] be able to be like okay I'm a little [30:18] hedged like it's cool like like we're at [30:21] that phase still um over time I think it [30:24] stands to reason that that you'll see [30:26] meaningful different costs because [30:27] you'll be able to peel off workloads [30:29] that that just only make sense at at [30:32] sort of you know grinding down to the [30:34] cost of inference in which case open [30:36] open weights will have the sort of [30:37] economic advantage. Right now there's [30:39] you know this challenge of like [30:40] sometimes it's more token inefficient [30:42] you know sometimes like randomly like [30:44] I've heard stories like randomly it'll [30:46] just like speak Chinese like like mid [30:48] midchain so you're like okay well [30:50] that'll be weird for a bank. Um, so [30:52] funny. So like we need to like probably [30:53] work on some of those things. But [30:55] longterm I think it has to be, you know, [30:56] the case that that you're going to [30:57] you're going to peel off those [30:58] workloads. You know, one one of the more [31:00] interesting posts I think on this that I [31:01] totally subscribe to is Jesse at at [31:03] Decagon. Do you probably read that post [31:05] of like of like this paradox of like [31:07] we're going to have like you're going to [31:08] see clothes just, you know, continue, [31:11] you know, go exponential. But what's [31:12] going to happen is each use case that [31:14] kind of matures you can peel off to open [31:16] source. And once you have kind of [31:18] stability in that use case, it starts to [31:20] make sense to veer it toward an open [31:22] weights model, assuming one of two [31:24] things is true. One, that that it's [31:26] actually literally cheaper, or two, [31:27] having some post- training gets you x x% [31:30] more performance. And and so I think you [31:33] will just be in a reality where we will [31:36] we will and this is going to be very [31:37] confusing probably for like the press [31:38] more than people in the valley because [31:40] you'll be like, wait a second, like the [31:42] revenue of entropic, opening eye, etc. [31:44] are like off the charts but somehow open [31:47] weights is like also growing [31:48] exponentially and you're like how is [31:49] this like how is open routing so fast? [31:51] >> Yeah. And it's like the pie is growing [31:52] so fast [31:53] >> but what's happening is is actually [31:54] there's there's an interesting duality. [31:56] It's it's it's not even just like like [31:58] rising tide lifts all boats. It's like [31:59] no no like we either use Fable or or 56 [32:03] for orchestration and then we farm out [32:06] all these longtail tasks to a cheaper [32:08] model. Um or the opposite is true. um [32:11] like you use you have some kind of [32:13] orchestration agent that like by default [32:15] does the cheaper stuff but occasionally [32:18] sort of sees some something that is just [32:20] way too hard and then it pops it out to [32:22] to one of these heavier models. And so [32:24] you might have blended 50% spend on each [32:27] but 10 times the amount of tokens, you [32:29] know, on the open weights model. And so [32:31] like everybody's kind of winning, but [32:32] there's an there's an interplay between [32:34] why they're winning. I'm curious about [32:35] how you think about memory and uh [32:37] customization or personalization and [32:39] where that's going to go because it [32:40] seems like today the dominant [32:42] architecture is kind of like rag based [32:44] system still like you get fancy on the [32:46] rag but it's still context lookup where [32:48] the weights themselves aren't [32:50] fundamentally changing. Yeah, [32:51] >> it seems like I mean if I listen to my [32:52] friends at the labs like continual [32:54] learning this idea of like the models [32:56] weights should adapt as it gets to know [32:57] you. [32:58] >> Yeah. [32:58] >> Um we had engram on the podcast. I don't [33:00] know if you if you know Dan like [33:02] >> I I just got introduced to him. I listen [33:04] to the podcast. Soing I would have loved [33:06] to been the [laughter] fourth person in [33:07] the room or [33:08] >> Amazing. Yeah, they they I I think [33:09] they're working with customers that to [33:11] to help basically bake in some of the [33:14] context into the weights themselves. [33:16] What direction do you think this going [33:17] to go? [33:18] >> You know, you're catching me at a time [33:19] right before I'm actually doing my call [33:20] with Dan. So So I I wish I I could have [33:23] talked to him first and then I'll have [33:24] like a way more eloquent answer. Um I'm [33:27] like extremely fascinated by by the [33:29] approach. like I I have no reason for [33:32] not wanting to it to work and and exist. [33:35] We live in a world at Box where we see [33:37] these sort of the the the high degree of [33:40] complexity on permissions and access [33:42] controls and data that that tends to be [33:45] sort of the rub on a lot of these types [33:47] of of approaches. And I'm I'm going to [33:49] put engram aside because like I'm sure [33:51] they've already thought this through. So [33:52] I'm I'm going to talk more generic phil [33:54] philosophically. I think sometimes you [33:56] will talk to a researcher that you know [34:00] sort of imagines the world working the [34:01] way they work which is like I'm a [34:03] researcher I have access to everything [34:04] and so if I had a model that was trained [34:06] just on my world this would be amazing [34:08] and then you're like let me introduce [34:09] you to a lawyer and the lawyer like has [34:13] this tiny little you know access point [34:15] of just like the five projects they're [34:17] working on because somebody right you [34:20] know one door over is working on the [34:21] competitive project to to a another [34:25] company in the space that they can't [34:27] have any sort of overlap with with what [34:29] they see or what they know and there [34:30] can't be a single document that passes [34:32] between those two walls and they have to [34:34] be these kind of hard barriers. So, you [34:37] know, so like sure, like you could still [34:39] you could still now train a model just [34:41] for that one user, but like what happens [34:42] if every single day they get added or [34:44] removed from something that that adds [34:46] important context to to sort of what [34:48] they need to understand. And and and you [34:50] know, again, like I think there's going [34:51] to be probably breakthroughs in [34:52] continual learning that that sort of all [34:53] resolve this. But like this is why [34:55] previously it was just like, you know, [34:57] there was no way you could pull this off [34:58] 5 years ago because it would be insanely [35:00] expensive, impossible to kind of wrap [35:01] your head around around how those access [35:03] controls are supposed to work. But, you [35:05] know, obviously with like as the cost [35:07] curve goes down, as open weights, you [35:09] know, get, you know, cheaper, smaller, [35:10] faster, better, um, I think this becomes [35:12] super interesting. One thing on on the [35:14] podcast that I found very fascinating [35:15] and I just need like a T-chart honestly [35:17] is just like, you know, like what is the [35:20] decision point of what goes in context [35:22] and what goes in the weights? Like you [35:24] you have to be a little bit thoughtful [35:25] about like where is the the massive [35:28] performance gain that you get by by [35:30] baking in the weights. And there's [35:32] probably some like incredible like like [35:34] um calculation of like like when the [35:36] rate of change of the data is not you [35:38] know so far but the upside of the of the [35:41] weights you know you know dramatically [35:43] change the accuracy of the model like [35:46] you know you'd have to kind of land on [35:47] some sort of you know rubric like that. [35:49] I mean, if you could wave a magic wand, [35:50] it almost seems like I think Karpathi [35:52] has said this in some prior interviews, [35:53] like if you could almost remove all the [35:56] memorized information from the models [35:58] and just have have it encapsulate the [36:00] specific [36:01] >> reasoning capabilities like the ethos of [36:04] how we do things for example at Sequoa [36:06] and then you have all the you know [36:07] actual content and a lookup system that [36:10] almost feels like the if you could wave [36:11] a magic wand that's what the system [36:12] would look like [36:13] >> and so so that one's super interesting [36:14] and the I I think the the the question [36:16] will be like how much are enterp [36:18] enterprise is different at that level [36:20] versus it's actually their their literal [36:22] IP that is what makes them different [36:24] like how many different types of styles [36:27] of execution are there in the world [36:28] versus no it's just like the depth of [36:31] knowledge about that particular legal [36:32] case and and how do I apply it to this [36:34] other project I'm working on that's [36:36] where so much of the value sits so so [36:39] but again if you can just like wait till [36:40] my Zoom call with Dan and then I'll [36:42] really know the answer but like I'm a [36:44] fan because no matter what there's going [36:46] to be like like I've jumped right into [36:48] like the individual, you know, but like [36:50] no matter what, like at a firm level, [36:51] there's probably ways to take this [36:53] approach. Like I'm a big sort of fan of [36:54] what Trajectory or Applied Computer [36:56] doing or Prime Intellect because because [36:57] I think there's like there's there's no [36:59] question that if you're Eli Liy, you [37:02] want a model for how you do drug [37:04] discovery and that probably does need to [37:06] go like farther or deeper or more sort [37:08] of specific than what you're getting at [37:10] off the shelf. And there's not a lot of [37:11] sort of, you know, kind of church and [37:13] state problems for for drug discovery, [37:15] you know, workflows. they they probably [37:17] want as much of that information [37:18] available to as many people as possible. [37:19] So I think it's going to be like domain [37:20] specific and you know you're just going [37:22] to have different outcomes based on [37:23] which vertical or or you know type of [37:25] use case and where the firewalls need to [37:27] be in that process. [37:28] >> Makes sense. Okay. So you hinted at the [37:30] beginning that there's a Box Labs. What [37:32] what type of work is Box Labs doing? [37:34] >> Uh so it's the equivalent of Box Labs [37:36] like I don't know if we've used capital [37:37] L yet. Um but uh but basically you know [37:39] it's our it's our applied sort of AI [37:41] team. And [37:43] >> what research areas are most interesting [37:44] to your your team right now? [37:45] >> Yeah. So, so the the the biggest areas [37:48] that uh that the most sort of research [37:50] kind of or you know of the continuum of [37:52] engineers there's you know some cluster [37:54] that is sort of more on the research [37:55] bent and uh of that cluster the things [37:58] that we spend time on still again is at [38:00] the kind of applied layer but but it's a [38:02] lot around how do you how do you take [38:04] agents and make them you know another 10 [38:06] points of accuracy improvement given x [38:09] problem. Um so what what is the you know [38:12] how do you build a map of the of the [38:14] problem set you know with a given set of [38:16] data to sort of best execute on that [38:18] task. So we spend a lot of time on on [38:20] that style of work. We have a team for [38:22] instance working on how do you do [38:24] effectively at the a agent level at the [38:26] harness level you know some form of kind [38:28] of auto research on on sort of hill [38:31] climbing on accuracy of answering [38:33] questions or set sets of problems on on [38:35] given a kind of a set of client data. So [38:38] if you're a bank, you have a bunch of [38:40] loan documents coming in and these are [38:41] like 100page documents like whether [38:44] you're getting 70% accuracy with an [38:46] off-the-shelf model or like 97% is like [38:49] basically obviously a world of [38:51] difference in can you actually go and [38:53] automate that process. So somehow you [38:55] have to hill climb from the base model [38:57] to the 97% and there's a lot of work [38:59] going into the system to to basically [39:01] pull that off. Maybe zooming out, what [39:03] do you think of as, and this can be a [39:05] box question or a non-box specific [39:07] question, the role of systems of record [39:10] in a world with agents? And I'm sure you [39:12] saw some of the Twitter discourse on [39:13] like, you know, every software company [39:14] is trying to sell me their own agent [39:15] right now. I don't want another agent [39:17] from them. I want I want their system [39:18] record to work well with my agent. And [39:20] so, like, how do you think about that? [39:21] >> Hashtagcloud force. Um, [laughter] so, [39:23] uh, [39:24] >> catchy name, by the way, very catchy. [39:25] >> I I mean, literally, it's like one of [39:27] those things where where like the first [39:29] three minutes you're like, man, that [39:30] seems funny. And then four minutes [39:32] later, especially when you see the [39:33] stock, you're like, "Ah, brilliant [39:34] move." Like like this is great. We're [39:37] doing this. Um, and then when you I [39:39] think somebody somebody actually uh said [39:41] this the best. They were like when they [39:42] heard Matthew McConna say it out loud, [39:44] that was like really the that sealed the [39:46] deal. That [39:46] >> was the aha moment. [39:47] >> That was the aha moment. I was like, [39:48] man, he can sell software. Like it's [39:50] actually incredible. Like his voice is [39:52] so good for selling [laughter] systems [39:54] of record and agents. Um if you're in [39:57] like like our sort of you know [40:01] contemporary group uh of like you built [40:05] a SAS platform and you have some set of [40:07] data and workflow that that you know [40:10] your customers kind of operate in [40:11] there's there's effectively two things [40:12] you just have to do and I I think [40:14] anybody attempting to do one over the [40:16] other is just going to lose. Um you have [40:19] to build an agent that is insanely great [40:21] at your product. like that agent has to [40:23] be you have to provably be like 10 or 20 [40:26] points better than an off-the-shelf [40:28] agent at at using your system. Not [40:30] because it's hobbled the other side. [40:31] It's just like you are so Eval maxed and [40:33] you're like so tuned to your particular [40:35] workflow that that you can improve your [40:38] system. You have to have that and you [40:41] probably because you understand your [40:42] domain unless you're like totally asleep [40:43] at the wheel. you probably have use [40:45] cases that no one has has sort of [40:47] thought to build products around because [40:48] you you talk to customers every day and [40:50] you see what they run into and you're [40:52] like, "Oh, we could just like have our [40:53] agent go do that for you." Like I've had [40:56] at least a dozen I mean, so I probably [40:58] talked to a couple hundred customers a [40:59] year in in variety capacities. I've had [41:02] at least two dozen times where the [41:04] customer has a use case that is like a [41:06] breakthrough moment for me of like, [41:08] that would be actually totally [41:10] insane. like what's example? [41:12] >> Um, unfortunately, since you put me on [41:14] the spot, I don't know if my example [41:15] will pay off the the level of excitement [41:18] I just had because it'll be [41:20] >> put you on the spot. [41:20] >> No, no. I It's just like I think like [41:22] the thing I was thinking of is just like [41:24] I think it's going to be a wamp wamp for [41:25] the podcast. But um uh there was there [41:28] was basically uh a customer had this [41:30] idea of they wanted an agent in the [41:32] background trying to sort of figure out [41:34] when documents sort of met their [41:36] governance policies and like does [41:38] something need to go into some kind of [41:39] archive or something into some kind of [41:41] legal hold or whatnot. Oh, cool. [41:43] >> And and [laughter] see exactly see [41:45] that's exactly the voice that was [41:47] exactly the voice I was worried about. [41:48] Yes. Yeah. No, you couldn't even you [41:50] couldn't even pull it off. Okay. So, so [41:52] but in our world, this is awesome [41:54] because you're like, "Oh, [41:55] >> yeah." [41:55] >> Like, no, because think about it. Every [41:58] company has a head of governance. [41:59] >> I meant Cole sincerely. [42:00] >> Okay. I know. I believe you. Uh I mean, [42:02] listen, you do enterprise, so like I [42:04] think that it was at least half serious. [42:05] Um imagine you're an enterprise. You [42:07] have a head of compliance and a head of [42:08] governance. Okay? They can only be like [42:12] overseeing the whole sort of enterprise. [42:15] They've never been able to be everywhere [42:17] at once. Now imagine if they could sit [42:19] next to the employee and and basically [42:22] be able to be like, "Oh, you're about to [42:24] go do something that breaks our [42:25] governance policy." So So like the idea [42:27] was like, "Oh, what if there was just an [42:29] ongoing agent that just like [42:31] automatically was just like, nah, that's [42:32] going to break your governance policy." [42:34] Instead of the user having to like try [42:35] and predict or understand this stuff. So [42:36] anyway, those are the kind of things [42:38] where if you have an agent within your [42:40] product, you're going to be able to [42:41] identify sort of sooner and better than [42:43] than the rest of the market and or just [42:45] like do things that maybe would be [42:46] impossible to do off platform. On the [42:48] other hand, it's just like obviously you [42:51] have to go headless. You literally have [42:52] to make sure that your APIs are exposed [42:54] to cloud and chatbt and and you know all [42:57] the different platforms. And you have to [42:58] make sure that you have a either a [43:00] direct way into deterministic APIs so [43:03] that those agents can use your your APIs [43:05] to make calls via MCP or whatever or at [43:07] least make your agent be headless and be [43:09] exposed in those systems. And then the [43:11] only reason maybe this is like a like a [43:13] even remotely a hard debate is you have [43:15] to just make sure as a as a system of [43:17] record that you can find a way where [43:19] commercially it sort of makes sense on [43:21] the other side and and is sort of [43:23] valuable and interesting. And the reason [43:25] why I think a lot of people got that [43:26] wrong that that weren't in these [43:28] companies was just underestimating the [43:31] amount of of sort of new use cases that [43:33] are just total upside. They're like [43:34] complete whites space opportunities for [43:36] these systems of record. So like in the [43:38] Salesforce example, I use Salesforce [43:40] more today, probably by an order of [43:41] magnitude than I ever have because I MCP [43:44] into it via Claude or Chatbt and so and [43:47] so I just am always asking questions [43:49] about the data inside of our CRM system. [43:51] >> And do you think that means the systems [43:52] of record become more toll booth [43:54] businesses then to make sure that [43:55] they're capturing the opportunity? [43:56] >> I don't love that term. Um because like [43:59] no one's had a good experience at a toll [44:00] booth. [44:00] >> I love toll booth. [44:01] >> Yeah, exactly. You love it more than [44:02] governance agents. Um [laughter] so uh [44:04] so so I would say that that because they [44:07] they sort of have a have a depth of [44:10] purpose of organizing the workflow, [44:12] managing the data, securing the data, [44:14] providing guard rails, uh then it's [44:16] really just yeah, you need to you have [44:18] to have some kind of volumeoriented [44:19] business model on that other side. And [44:22] uh and I just think there's like if [44:23] you're solving real problems for [44:24] customers, it'll just like make money. [44:26] I've like like this is so cheesy but [44:28] like I've I've told like LinkedIn [44:31] product managers I'd probably pay 10x [44:33] more for LinkedIn if I just could MCP [44:35] into it. If I just had a way of just [44:36] like always understanding like like okay [44:39] this CIO is doing this thing and I need [44:41] to reach out or whatever I'll I'll like [44:43] take my money. Yeah. [44:44] >> So so there's these systems actually [44:45] have a tremendous amount of value based [44:48] on the data that they that they have and [44:50] customers will absolutely sort of find [44:52] some way to reward you for that value [44:54] creation if you're doing a good job. [44:55] >> Super interesting. Um maybe related [44:58] let's talk about product UI. [44:59] >> Yep. [45:00] >> Big you know generic chatbot chat box [45:03] agent. Um is that going to be the [45:06] dominant UI for how people are using AI [45:09] in the future especially when it comes [45:10] to the application layer? This is where [45:12] why I think the applied layer, you know, [45:14] has so much run u room to run is because [45:18] probably the the sort of, you know, [45:21] universal chat system that you ask a [45:23] question to, you get an answer back or [45:24] it does sort of some work in the [45:25] background. I think it's going to be [45:26] obviously like that's going to be a [45:28] mainstay that that that UI will always [45:30] exist. It'll be incredibly powerful. The [45:32] horizontal products will have it. The [45:34] vertical products will have it. [45:35] Everybody will have it. It's just like [45:37] >> your product has a search box. [45:38] Obviously, it does. Yeah. So, so that [45:40] that's always going to be here for these [45:42] sort of one-off asks of an agent or go [45:44] find this thing or answer this question [45:46] or produce something for me on demand. [45:48] But most of the enterprise uh is sort of [45:50] made up of these processes and workflows [45:53] that are kind of just happening behind [45:54] the scenes. Sometimes they're happening [45:55] with computers and computers are running [45:57] these things or sometimes they're [45:58] happening with other people that are [46:00] doing these things or sometimes they [46:01] should be happening with people but you [46:02] could never afford to have them happen [46:04] with people so they just didn't happen. [46:06] And so that's that sort of is a slightly [46:08] different kind of, you know, metaphor [46:11] than than a chatbot where you ask a [46:13] question and it comes back w with an [46:15] answer. That's like, okay, I I kind of [46:16] want agents in the background to do [46:19] things for me, read every contract, look [46:21] at every log, triage every security [46:23] incident in incident, and then instead [46:25] of me chatting, maybe I'll chat as a as [46:28] a means of doing kind of a catch-up. I [46:30] want a dashboard. I want a workflow. I [46:32] want a queue. I want a task list. So [46:35] then the the the the challenge becomes [46:38] well does the horizontal product kind of [46:41] take on every one of those those [46:43] components and and manifest every one of [46:46] those experiences in one in which case I [46:48] think you'll start to be like man that [46:49] thing is like really like a that's [46:51] pretty heavy and like and then we'll [46:52] start to like be like oh this is no [46:54] longer this simple easy delightful thing [46:56] anymore. So then the vertical players [46:58] actually like like they actually sort of [47:00] understand the process and can manifest [47:02] all the right buttons and tabs and the [47:04] names of the things for that particular [47:06] workflow. So I think it puts I think as [47:08] you have agents that are doing more work [47:10] in the background doing more you know [47:12] async work that is just like I farmed [47:14] out a bunch of agents to review things [47:16] as they happen or whatnot that leans [47:18] more toward the applied companies that [47:21] can understand those workflows that can [47:22] understand those processes. I think [47:24] you're going to have these in every [47:25] field. you're going to certainly we we [47:27] already know that you know how they're [47:28] going to look in legal with Harvey Lor [47:30] etc. um we are seeing them start to [47:32] emerge in areas like security. We've [47:35] seen them start to emerge in the [47:36] longunning kind of coding agents with [47:38] cognition and factory. Um so I think [47:40] that that will be you know one of the [47:43] bigger kind of applied AI sort of use [47:45] cases and ultimately like in five years [47:47] from now I would bet like 90% of all [47:50] tokens in the enterprise are things that [47:52] a user never kicked off and they just [47:54] see a result. They just they see a task [47:57] show up and they have to go review it [47:58] and it's just like it's just happening. [48:01] >> Yeah. Yeah. Makes sense. Okay. Let's [48:03] talk about AI diffusion. Um coding [48:05] agents it was like you know boom January [48:07] 1st 2026 happened and like the fastest [48:10] diffusion of anything into the economy [48:11] we've ever seen has happened. [48:13] >> Um the diffusion of you know the rest of [48:15] the AI magic into the rest of our jobs [48:17] seems like it's been a lot slower. [48:19] >> Yeah. [48:20] >> What are your thoughts on that? And [48:22] where are the areas where you think [48:23] we're going to see faster diffusion and [48:24] how is that going to happen? [48:26] >> Yeah, so you always have to kind of [48:27] compare and contrast coding versus [48:29] everything else to really understand the [48:31] the dissimilarity. Um so in coding [48:34] essenti and this is back to the kind of [48:36] utility point on on you know slop like [48:39] the utility of code is almost 100% [48:42] represented by the amount of text that [48:45] you can generate like like all obviously [48:48] an insane amount of value went into the [48:49] text but but like and like knowledge and [48:52] expertise and meetings and everything [48:53] but like ultimately like the text is the [48:56] thing that produces the the program that [48:58] is actually the thing that you're trying [49:00] to do. You know, if you could have the [49:02] world's greatest programmer like and [49:05] never had to sleep, never had to eat, [49:06] they they could intuit it what to build [49:09] and they could just sit on a computer [49:10] all day long, your value creation would [49:12] be 100% correlated with how many hours [49:14] they could sit at that computer and type [49:16] code. Like lines of code, ideally, [49:19] ideally good code is the most thing that [49:21] will be correlated to whether you [49:22] produced software that that people [49:24] wanted. So it's all text. The models are [49:26] hyperrained on these. Everybody in AI [49:29] labs treat coding as as a competitive [49:32] benchmark to constantly try and exceed. [49:35] They they get to do their own evals on [49:37] it every single day because they are the [49:38] ones coding the the models themselves. [49:41] >> And it's the most technical audience of [49:43] all time where when they deploy an [49:45] agentic system and they run into a [49:48] either a bug or a problem or like some [49:49] MCP server comes back with like [49:51] connection invalid, they fix it. [49:52] >> They know how to triage the problem. [49:53] They don't call it. They just like, "Oh [49:55] yeah, no, I didn't open up that port. [49:57] Sorry. I'll go fix it. So, that's like [49:59] five things. Oh, and maybe like the six [50:02] like it's just like a very very [50:03] highpaying vertical that that like it's [50:06] just like automatically valuable if you [50:07] could get 10% or 20% productivity gain, [50:10] let alone 5x productivity gain. So, so, [50:13] so take those five or six things that [50:16] coding has as as sort of beneficial [50:18] properties to sort of automation, then [50:20] compare that to, you know, you know, [50:23] every other form of knowledge work. and [50:24] you'd probably have like a histogram and [50:26] I don't know if anybody's published [50:27] this, maybe you can like the similarity [50:29] to coding and like like what what are [50:31] the domains that like like start to sort [50:33] of sort of you know look closer like [50:35] like less and less like coding as you as [50:37] you kind of scale out. And it's like [50:39] okay well lo and behold legal is kind of [50:41] interesting because because like there's [50:43] a lot of value creation to somebody [50:45] sitting at a computer reviewing legal [50:47] documents, writing legal documents, like [50:49] processing large amounts of information. [50:51] Okay, so that's kind of blowing up and [50:52] then you kind of go down the list. Now, [50:54] let's take something like like a sales [50:56] rep. Okay, so much farther down the list [50:58] in terms of of sort of likeness. The [51:00] sales rep's value creation is basically [51:03] convincing an external customer to buy [51:06] software or technology or a caterpillar [51:08] truck, you know, uh from them. That is [51:11] the value creation to the to the economy [51:13] of the sales rep. [51:14] >> And so, let's say we brought the world's [51:16] best automation to them. Like first of [51:18] all again they'd have to like figure out [51:20] how to technically wire it up. They'd [51:21] have to make sure they give it all their [51:22] data all these kind of things. But no [51:25] matter what the they're they're still [51:27] rate limited and constrained by like did [51:28] the customer respond to them? Do they [51:30] want to meet? Can they meet next Tuesday [51:32] or can they meet today? Like does the [51:34] customer have budget? All of these other [51:36] things. So that that's maybe the entire [51:37] continuum right there is like one is [51:39] like I mean of knowledge work like [51:41] obviously this is not even you know [51:42] touching the the sort of working with [51:43] with Adam. So that's a continuum which [51:46] is on one end you have somebody rate [51:48] limited by so many external factors on [51:50] the other end you have somebody who [51:51] could sit at a computer all day long and [51:52] just type type text and that is your [51:55] ability to basically automate things is [51:56] is that continuum. So for the real [51:59] world, we have to basically bring [52:01] intelligence to these workflows in ways [52:03] that are that sort of somewhat feel like [52:06] the shape of their workflow, somewhat [52:07] feel like the shape of their work and [52:09] then find a way to deliver the change [52:12] management, deliver the implementation, [52:14] get data into a format and into an [52:17] environment that actually works with [52:18] these systems. like asterisks like one [52:20] of the other big things is like if you [52:21] go to most engineers in 2026 like maybe [52:24] like minus two months ago given the [52:26] given the latest phenomenon but like [52:28] like the code's in GitHub you just like [52:30] connected to GitHub like like remember [52:32] there's this period where like your when [52:33] you launched a coding agent like there [52:35] was no like sign up or register it was [52:37] just like give us your GitHub that [52:38] doesn't exist in knowledge work there's [52:40] no like give us your GitHub for [52:41] knowledge work there's like [52:43] >> give us your box [52:44] >> well box customers have a have a much [52:46] easier time with all this unfortunately [52:47] like We're only 1.3 billion in revenue [52:49] run rate. So like like that means [52:51] there's a lot of people not using Box. [52:52] And so what are they using? Their data [52:55] is in on premises systems, legacy [52:57] fileshares, legacy infrastructure, you [53:00] know, enterprise environments that don't [53:02] talk to agents particularly well. So [53:04] just think about that sort of [53:05] distinction between, you know, [53:07] implementing coding agents versus [53:09] everything else. Um even even something [53:11] again you'll sort of fall asleep about [53:13] is like access controls in the [53:15] enterprise are totally different. And I [53:16] actually totally forgot that that point [53:17] about coding. In coding, you get access [53:20] to basically most of the stuff ever [53:21] relevant for your job. In knowledge [53:23] work, you're like you're like, "Hey, [53:24] Sally, can you open up that that sort of [53:26] file share for me? Can you open up that [53:28] that, you know, sort of project because [53:30] I didn't get access to it. How do you [53:32] make sure the agent has access to those [53:34] set of things?" All of that work has to [53:36] get done. So, the thing I think we have [53:38] to prepare for is two things. one, [53:40] Silicon Valley has to prepare for [53:41] diffusion taking a lot longer than they [53:42] think or then I'll say we but I really [53:45] think they because because I like I know [53:48] how long it'll take. And then the second [53:49] thing is is the good news is this is [53:52] sort of all correlated to the applied [53:54] layer. Yeah. [53:55] >> Value creation like because the the [53:58] companies that will just have the [53:59] patience the the sort of the the full [54:02] sort of domain expertise the the just [54:05] the sheer work ethic because it's not [54:07] like everybody's just from the [54:08] floodgates like you have to go and and [54:10] just you know pound pavement and get out [54:12] there. That will be the applied layer. [54:14] So I think this actually represents, you [54:16] know, a trillion dollars of applied [54:18] layer AI value is is actually how do you [54:21] go get the technology to the lawyer or [54:23] to the sales rep or to the life sciences [54:25] researcher or to the person that runs [54:27] the customer support team. That's all [54:29] sort of opportunity right now that [54:31] exists. [54:32] >> Awesome. I'm going to close by asking [54:34] some advice for other founders. [54:35] >> Um maybe let's start with founder advice [54:37] and then company building advice. On the [54:39] founder side, seems like you are in [54:41] every AI cap table. you know, every cool [54:43] new company like you know, you know, [54:44] Engram, you know, um, how did you kind [54:47] of get yourself in the middle of the AI [54:49] conversation? [54:50] >> There's probably two two parts. One, I [54:53] was just very well primed for it like [54:54] working with unstructured data for 20 [54:57] years. You just like can instantly see [54:58] the the the benefit of agents on that. [55:00] So like it honestly like took longer [55:02] than I would have wanted that that we we [55:04] got to have this conversation because we [55:05] we tried to have this conversation, you [55:07] know, eight, nine, 10 years ago. Y [55:08] >> and now it's finally happening. So, so [55:10] first of all, just super well primed. [55:12] Our product sort of, you know, shape and [55:14] and what people do with our product [55:15] already lends itself extremely well to [55:17] agents. So, like obviously we had to bet [55:19] the company on that and and then lo and [55:21] behold, we had the positive feedback [55:22] loop of customers actually saying, [55:23] "Yeah, that would actually be very [55:25] powerful if I could go read every [55:26] document and answer any question." So, [55:28] that that's the first and certainly by, [55:30] you know, biggest by kind of a factor of [55:31] 10. And then the other is just like I'm [55:33] extremely fascinated by the technology. [55:35] Um, and it's just fun. It's like [55:37] >> where do you learn about it? um your [55:39] podcast. Um uh Doresh's podcast um uh [55:45] the Twitter I mean probably [55:47] unfortunately for brain cells like it's [55:49] probably 95% Twitter. Um I have a [55:52] routine where like at the end of each [55:53] night I just go through the feed and [55:55] it's just like I do like I I look like [55:58] some sad meme probably of just like I'm [56:00] just scrolling and scrolling and [56:01] scrolling and like attempting to like [56:03] triangulate all the information. That's [56:05] amazing because it is it is the global [56:06] town square for AI. [56:07] >> It is. And um I want to send an [56:09] emergency alert to like everybody who's [56:11] like a sophomore or junior in college [56:13] and just be like follow these 20 [56:15] accounts on Twitter and and also join [56:17] Twitter because like this will just help [56:19] your career. You're either like a year [56:20] ahead or a year behind simply based on [56:22] your feed. And my feed is like so so [56:25] wired in. [56:25] >> This is the number one advice I give to [56:27] people when they're asking how to get [56:28] current on AI. It's like follow follow [56:30] these hundred accounts. [56:31] >> It's it like but I mean virtually first [56:33] step join Twitter. Yeah. [56:34] >> Like you I'll still talk to to [56:35] 20-year-olds that are like, "Yeah, like [56:37] you know, I I I see some articles." I'm [56:39] like, "What do you mean? How do you see [56:41] articles? Like I don't even know what [56:42] that means." Like, [snorts] "Do you just [56:44] like get lucky that somebody emailed you [56:46] an article? Like just join Twitter. Like [56:47] what are you talking about?" So, so uh [56:50] so I would I you know, you just have to [56:52] be wired in. I I enjoy it. It's a lot of [56:54] fun. I play with everything. And [56:56] >> what's what's your favorite new AI [56:57] product? Please say Instinct. [56:59] >> Uh so um Okay. Uh full full disclaimer, [57:03] I have not done the Instinct invite code [57:04] yet. Um simply because I have a backlog [57:07] of like three other personal assistant [57:09] products and [57:10] >> you gota try instinct. It's [57:11] >> I know I know everybody's ra I I very [57:13] excited. I'm on [57:14] >> and I'm not an investor. So [57:15] >> Oh, you're not? Okay. So this is like [57:16] totally genuine. Okay. So I I 100% will [57:19] have I don't know when this is going to [57:20] run, but I'm sure I will have played [57:21] with it by the time it runs. I I'm in [57:23] pre-release in a couple right now that [57:25] I'm spending some time with. Um, I think [57:27] the personal assistant stuff is super [57:29] exciting. At least in some of my use [57:30] cases, it's still showing some of the [57:32] limits of like browser use as an [57:34] example. Like we still have some work to [57:35] do there. Like probably the entire [57:37] internet needs just like a CLI for their [57:39] product. Um, so we still have like [57:41] there's still some blocking and tackling [57:42] at the infrastructure level for these [57:44] things to be totally awesome. Um, and [57:46] then and then I kind of look like your [57:48] average probably, you know, AI pill [57:51] knowledge worker on, you know, every day [57:53] I'm asking one of five different AI [57:55] systems, you know, 20 to 30 questions um [57:58] on just like doing research, looking for [58:00] talent, looking for what is a competitor [58:02] doing, what's happening in this market, [58:04] how do you expand there? Um, so so kind [58:07] of that. [58:07] >> What about on the company building side? [58:09] How do you, you know, 20-year-old [58:11] company at this point? What's your [58:12] advice for other people that are trying [58:13] to reinvent their companies to make sure [58:15] they have, you know, max adoption of AI, [58:17] not just at the individual level, but [58:19] also at the kind of company level? How [58:21] do you make your business legible for [58:23] AI? [58:24] >> Yeah, again, some of this we have as a [58:26] as a byproduct of how we've always [58:28] thought about information systems in the [58:30] company. So like like to no exaggeration [58:34] if you have a question that you would [58:36] like to ask about the the the business [58:38] that has ever been documented in a form [58:41] of unstructured data. So a meeting note, [58:44] a project plan, a road map, a [58:47] presentation, a financial document, a a [58:49] financial planning session, it's 100% [58:51] inbox. So like we benefit from a data [58:54] architecture that already is insanely [58:56] sort of tuned for because this is how [58:58] we've run the company. we didn't let [58:59] anybody use anything else and and so the [59:01] data is very easy for us to be able to [59:03] work with at scale and then of course we [59:05] have Salesforce and and all the other [59:06] kind of core canonical systems we've [59:08] been able to kind of have I think pretty [59:10] good data hygiene so agents running on [59:12] top of that makes it a little bit easier [59:14] to be kind of AI AI first in and how we [59:16] operate uh maybe a couple best practices [59:18] or or things that we've seen you know [59:20] first of all trying to figure out where [59:21] the highest leverage impact workflows [59:23] are going to be and and you know trying [59:25] to kind of target those so you know our [59:27] CIO's very AI pill. We have a little bit [59:29] of like a center of excellence on AI. [59:31] We've hired some internal AI FTEES to [59:34] kind of help with these processes. [59:35] >> Do you have leaderboards? [59:36] >> We don't token max. Um uh we do actually [59:38] have like I mean like I guess like [59:40] literally we have a list of people buy a [59:43] number of tokens but but it's usually to [59:45] like go and inspect like okay do we do [59:47] we think that's useful or is there a [59:49] learning there that we should take back [59:50] to some other function. Um, so like [59:52] probably our top, you know, in the top [59:54] three all AI users at Box, like probably [59:57] one of them is wasting, you know, half [59:59] the tokens and then two of them are [60:00] like, oh whatever they're doing, [60:02] we need to go like do like an internal [60:03] training session for everybody else. So, [60:06] you know, we had this thing like two [60:07] weeks ago where I was like, can you just [60:09] get everybody in a room, you know, on [60:11] this particular team and just like like [60:13] show them how this one person is is [60:15] using AI? And then like, you know, six [60:17] hours later, they were like in a room [60:19] and the guy was doing like a full demo [60:21] of what he was doing. Shout out to Mick. [60:23] Um and uh and so like that's the that's [60:26] the kind of stuff that we're trying to [60:27] do, which is just like how do we show [60:29] everybody what what it looks like to [60:30] work in this way. But you know, for as [60:32] fast as we're moving and and we are [60:34] shipping at some parts of the stack two [60:36] or three times more sort of, you know, [60:39] actual customerf facing products. So [60:41] like I don't care about how much code [60:42] but like did we actually deliver more [60:44] functionality the customers asking for [60:46] some parts of the stack we're doing that [60:47] then you'll go talk to a friend in [60:48] anthropic and they're just like and [60:50] you're like oh my god we still have we [60:52] still have ways to go like like I you [60:54] know the particular meta kind of [60:55] constantly changes which is like you [60:57] know two years ago you'd be like you're [60:59] just using a plugin in your IDE and [61:02] you're like man that's we're not going [61:03] to be that like we're we're not ready to [61:05] fully you know be AI first and then [61:07] you're like okay everybody roll out roll [61:08] out cursor and then you everybody rolls [61:10] out cursor And then you're like, and [61:12] then that finally happens and then [61:13] you're like, you know, you're walking [61:15] around, you're like, you're not only [61:16] working from Slack, just at mentioning [61:18] bots doing your code. Like, what are you [61:20] doing? Like, like it's like we're [61:22] constantly just changing what are the uh [61:24] what are the workflow paradigms on this? [61:25] >> Do you guys have like a Slack co-worker [61:27] agent for lack of a better term? [61:29] >> We have a a few that that have that [61:31] shape. Um, I'm pretty excited about like [61:33] like cloud tag as like a form factor. [61:35] You know, you still have to again kind [61:36] of get the the team construct right and [61:38] the data right. Um, but we have a [61:40] variety of ways that people, you know, [61:42] kind of work with agents in Slack. I [61:43] don't know if it's as as sort of [61:45] slackpilled as Benny Off would like us [61:46] to be, um, or as like anthropic or [61:48] opening eyes, but like we're we're [61:50] heading in that direction. [61:51] >> Yeah, I think that history books will be [61:53] written about, you know, the art of [61:54] business in this time. Like I would love [61:56] to read the new The Art of War with like [61:59] everything that is happening here [62:00] because I think it's pretty [62:01] extraordinary stuff we're seeing. What [62:03] do you think it takes to win in AI [62:05] versus preAI? And like what does it feel [62:07] like to be a founder right now versus [62:08] when you started books? I am both [62:11] jealous of and also not jealous of like [62:14] like the young founders you meet because [62:16] you're just like like oh to be young [62:18] again and and the whole world is your [62:20] oyster and you can go in you know any [62:22] direction and the leverage you have [62:23] you'll meet with them and the and you're [62:25] like oh my god like I I saw a product a [62:28] week and a half ago and they they did a [62:30] demo of the product and I was like I was [62:32] like absolutely this would have been a [62:34] 40 person project five years ago like or [62:36] or they you know especially when we were [62:38] starting out easily 40 person and it was [62:40] it was two people and you're just like [62:42] how do you have so many tabs that work [62:44] and they all seem to have stuff behind [62:45] the tabs that all seem very functional [62:47] like this is not fake. [62:48] >> Um and and so I'm very jealous of that. [62:51] Like it's like incredible because you [62:52] could just like you can start your [62:53] company from scratch with that as the [62:56] design principle. Now we will we will [62:57] get there because we're just going to [62:58] muscle through it. There's a couple [63:00] things that we just can't do which is [63:01] like we're not like we're very [63:02] uncomfortable at the idea of like sort [63:05] of you know removing the code review and [63:06] like some of these things that get [63:07] talked about because our customers can't [63:09] possibly entrust us with their data [63:11] security and compliance if we don't take [63:12] that seriously. So we're always going to [63:13] have a little bit of a discount on the [63:15] productivity because of of where we are [63:17] in the stack what we what we do as a [63:19] business. But so jealous of of being [63:21] able to kind of be fresh in that. And [63:22] then on the other hand like it's like [63:23] man like at the same time for every [63:26] great idea it's like instantly five [63:28] competitors. We didn't have that, you [63:29] know, problem. We had a we had a good [63:31] kind of couple years where we just could [63:33] like we could just grind on our product [63:35] and our and our sort of experience. And [63:37] it wasn't like every 3 days you were [63:39] like, "Oh, like Sequoia funded this [63:42] thing and benchmark funded this thing [63:44] and like you know, we weren't going kind [63:45] of nuts like with that." Now we we had [63:47] our own version of that. So like at the [63:49] time I probably was going going nuts, [63:51] but it was like like in retrospect it [63:53] was like not worthy of going nuts. Now [63:55] it's like, oh man, this is like a real [63:57] race in every one of these markets. [63:58] >> Totally. [63:59] >> So I think I'm probably pretty consensus [64:01] on this, which is in a world where AI [64:04] builds things so much faster. Then [64:06] probably the shift moves to whoever can [64:07] actually get it to the customer is in [64:09] the best position. So um you know, it's [64:12] it's fun talking to founders that are [64:14] pretty kind of pled on that. Um you [64:16] know, Scott or Matan are like they they [64:20] get the mandate. They're just like this [64:22] thing is going to be an enterprise [64:23] diffusion play. and and so you have to [64:25] get it to the enterprise. Um and and [64:28] anybody who kind of mistakes the the [64:29] mandate right now, you're just going to [64:31] lose. It's just like it's game over. [64:32] Sorry. Like there is quite literally [64:35] trillion to trillions up for grab at the [64:38] applied layer and the companies that [64:40] that know how to build the teams and and [64:42] get to the enterprise will be the ones [64:43] that win. It's just like obviously [64:44] guaranteed. [64:45] >> Well said, Aaron. This is a very fun [64:47] conversation. Thank you so much for [64:48] joining. [64:49] >> Thanks for having me. [64:52] >> [music] [65:05] [music] [65:14] [music]