Box CEO Levie Says Enterprise AI Value Shifts From Models to Apps

Box CEO Levie：企业 AI 价值正从模型转向应用 基础模型利润率的骤降将推动企业利润流向应用软件，因为动态数据权限与自主后台 Agent 构筑了原始模型无法复制的防御壁垒。 Aaron Levie 反驳了应用软件厂商仅仅是“LLM 套壳”的观点，详细阐述了 Box 如何构建深度集成企业权限、动态检索与非结构化暗数据的模型无关型 Agent 架构，以实现运营可靠性。他分析了为何编程领域能够迅速采纳 AI，而通用知识工作却因管理层对未经审核的生成“劣质内容”存在信任危机而陷入停滞，并将确定性的编译器与标准模糊的企业交付成果进行了对比。Levie 勾勒出企业软件作为平衡专有前沿模型与开源权重模型集群的混合架构前景，并预测在五年内，90% 的企业 Token 将由后台 Agent 自主消耗，而非来自人类提示词触发的聊天窗口。 批评应用型 AI 公司仅是薄薄的“LLM 套壳”，从根本上误解了企业价值；约有一万亿美元的企业价值取决于在原始模型认知与专有工作流之间搭建复杂的桥梁。 (2:18) 硅谷在结构上过度迷恋基础研究，片面强调原始模型能力，却忽视了企业中繁琐的基础应对工作——例如遗留数据孤岛、复杂的权限管控以及长达数天的人工审批延迟。 (4:54) 基础模型实验室在销售 Token 时面临固有的利益冲突，这催生了对模型无关型应用软件的结构性需求，以中立地跨提供商路由查询，从而优化成本、准确性与延迟。 (7:15) 实验室的 Token 补贴在经济上不可持续；随着推理成本下降且利润率压缩至接近 10%，经济利益的捕获将向上转移至工作流应用层。 (10:00) 编程领域能迅速采纳 AI 是因为它具备确定性的验证框架（编译器和测试套件），而更广泛的知识工作则因未经核实的合成“劣质内容”面临管理层的信任危机。 (18:50)

Box CEO Aaron Levie argues that the long-term economic value of generative AI will pool in the application and enterprise diffusion layer rather than in foundation models. He contends that impending model commoditization, margin compression, and complex enterprise access requirements make domain-specific software platforms the essential bridge for enterprise adoption. Model Commoditization and the Application Layer Bridge (2:18) Levie outlines a fundamental divergence in technology investment: whether the bridge between foundation models and enterprise workflows is shallow, or vast and operationally complex. He anticipates that foundation model margins will face structural deflation as asymmetric players like Meta, NVIDIA, and nation-state actors open-source weights or price inference near marginal cost to drive hardware and strategic objectives. Once venture subsidies recede and public markets demand returns on tens of billions in capital expenditures, pure-play AI labs will face margin pressure. Consequently, Levie argues that sustainable value flows upward into the application tier. Software companies that absorb the difficult work of integrating models into operational workflows capture proprietary domain data and human-in-the-loop interactions. Over time, this workflow context enables applied platforms to fine-tune specialized models and deploy proprietary agent harnesses that base foundation models cannot replicate off the shelf. Permission Firewalls and the SaaS Dual Mandate (32:53) While AI researchers frequently advocate for continual learning models that train context directly into weights, Levie argues that enterprise governance necessitates dynamic Retrieval-Augmented Generation (RAG). Enterprise data environments are heavily siloed by legal and regulatory boundaries—such as law firm matter conflicts or strict distribution lists—where permissions change daily. Training model weights to respect shifting access-control lists remains computationally and legally impractical for enterprise operations. To navigate this shift, established SaaS platforms face a dual mandate. First, they must engineer specialized internal agent harnesses that elevate zero-shot document reasoning from an off-the-shelf ~70% baseline to the ~97% accuracy threshold required for production workflows. Second, platforms must expose headless Model Context Protocol (MCP) endpoints, allowing external models to safely interact with core records while the underlying system of record charges on volume-driven data and governance access. The Diffusion Chasm Between Code and Business Artifacts (18:50) AI adoption has diffused unevenly across industry verticals. Levie highlights that software development embraced agents rapidly because source code represents a deterministic, functional utility verified by compilers and unit test suites. In contrast, ungrounded generative text in standard knowledge work frequently manifests as low-value 'work slop.' Because business artifacts like memos and presentations historically served as evaluation proxies for worker rigor and competence, uncurated synthetic deliverables trigger executive skepticism rather than trust. Furthermore, corporate roles fall along a wide automation continuum. While legal review shares software's text-centric inputs and outputs, functions like enterprise sales remain constrained by external human dependencies, scheduling delays, and buyer budgets. Unlike engineering, which benefits from standardized GitHub environments, broader corporate knowledge work remains trapped across fragmented on-premises shares and unstructured dark data without a standardized ingestion mechanism. Unprompted Background Agents Replacing Generic Chat Boxes (45:00) Levie rejects the assumption that conversational chat boxes will remain the dominant enterprise interface, noting that complex business processes require asynchronous dashboards, triage queues, and exception-approval flows. He forecasts that within five years, 90% of enterprise AI tokens will be consumed by automated background agents executing unprompted tasks—such as auditing files against governance policies, normalizing contracts, or triaging security logs—escalating to humans only for high-stakes edge cases. Because model providers selling tokens have an inherent conflict of interest to maximize compute consumption, Levie argues that enterprise platforms must operate as model-agnostic routing harnesses. Neutral orchestrators will benchmark tasks across competing proprietary and open-weights models, peeling predictable, high-volume extraction workflows onto cheaper open alternatives while routing complex edge queries to frontier reasoning engines. While modern AI tools allow two-person teams to build complex software prototypes at unprecedented speed, Levie concludes that engineering velocity alone no longer constitutes a sustainable moat. Enduring enterprise defensibility shifts entirely to distribution, customer trust, and the operational ability to navigate compliance and legacy data architecture. The open question remains how quickly non-technical industries can clean their dark data stores before specialized, agent-native entrants disintermediate incumbent workflows.

Box首席执行官Aaron Levie指出，生成式AI的核心价值正从基础模型层向企业应用层转移；随着算力支出高企与模型推理利润率面临结构性压缩，决定AI商业化胜负的关键在于能否跨越复杂的权限体系、遗留暗数据和垂直工作流壁垒。 万亿级应用桥梁与模型层经济学 (2:18) Levie认为，硅谷目前存在一种“沉迷模型研究”（research-pilled）的行业盲区，过度迷恋通用智能与基准测试，却低估了企业落地所需的繁琐业务对接。他指出，科技行业在“基础模型与企业流程之间的桥梁究竟浅薄还是深厚”上押注了万亿美元。类比云计算时代AWS和Google Cloud催生了Snowflake和Databricks等独立巨头，基础模型供应商难以通过集中研发吃透金融、医疗和政企等高度复杂的垂直流程，中间层软件必将分流大量价值。 在商业模式上，Levie分析称目前模型厂商依靠风险投资补贴低价Token的策略无法在公开资本市场长期持续，因为巨额训练支出终需真实的投资回报来偿还。与此同时，Meta通过开源模型商品化算力基础设施、Nvidia追求GPU销量，以及国家级力量的非利润竞争，可能将推理毛利率压低至10%左右。这种结构性通缩预示着基础模型将加速商品化，企业客户出于成本优化与利益冲突考量，也会倾向于采用中立的多模型路由框架，而非单一受制于基础模型提供商。 编码跃进而知识工作迟滞的扩散分化 (18:50) 针对AI在不同工种中扩散速度的巨大分化，Levie分析指出，软件工程的终极产出几乎完全体现为可编译、有客观单元测试验证的代码，且开发者本身具备极高的技术排错能力，因而AI渗透极为迅猛。相反，通用企业知识工作缺乏类似“提交GitHub代码库”的标准统一接口，企业非结构化数据高度分散于老旧的本地服务器与部门孤岛中，面临复杂的组织惯性。 这种结构差异引发了“工作废料”（work slop）信任危机。在传统商业场景中，PPT、策划案与备忘录等成果物历来是管理层评估员工理解深度与专业判断力的代理指标；当缺乏深度洞察的AI生成文本泛滥时，这种评价机制随即瓦解。Levie观察到一种普遍的认知失调：管理者习惯用AI辅助思考，却对他人提交的AI生成内容抱持本能怀疑。此外，知识工作沿自动化光谱分布——法务审查因高度依赖文本处理而与编程相近，但企业销售等岗位则严格受制于外部客户回复、会议排期及预算等现实摩擦，难以通过自动化文本生成缩短商业周期。 破解暗数据与权限的治理框架 (23:45) 面对企业内部沉睡的数千亿非结构化文件，Box并未尝试从头训练前沿通用大模型，而是聚焦于构建专用的“Agent Harness”（智能体治理框架）。Levie披露，开箱即用的通用模型在银行信贷审批等包含上百页复杂文档的任务中，零样本准确率通常仅为70%左右，无法满足企业严苛的业务标准；而Box通过构建覆盖检索增强生成（RAG）、动态文本切块、实时向量化与重排序的专用框架，将复杂任务准确率爬坡提升至97%的企业可用门槛，在处理海量文档归档检索时表现显著优于直接调用原始API。 Levie强调，学术界倡导的将全量知识固化进模型权重的设想，在企业现实中遭遇了权限隔离（ACL）与法律防火墙的根本制约。例如律师事务所处理涉及利益冲突的并购案件时存在严格的合规红线，且员工权限处于动态变动中，若将数据嵌入权重将导致天文数字级的模型重训成本。因此，精细受控的上下文检索（RAG）仍是主流路径，垂直应用的核心壁垒正是针对复杂权限穿透与暗数据结构化提炼所沉淀的领域级工程经验。 SaaS双重使命与异步后台交互 (40:00) 针对智能体时代的软件系统演进，Levie提出SaaS平台的双重生存法则：一方面必须打造深度贴合业务细微场景的原生Agent，使其在特定领域比通用Agent高出10到20个百分点的精度，例如在后台全天候监测文档是否违规归档；另一方面必须彻底实现“无头化”（Headless），向Claude、ChatGPT等生态开放确定性的API与模型上下文协议（MCP）服务器。随着外部Agent自主调用底层记录系统，软件定价正从纯席位制向调用量计费的模式演变，例如Levie表示愿支付10倍订阅费购买LinkedIn的MCP接口以自动追踪企业CIO变动。 基于这一交互转变，Levie反驳了企业软件将统合为单一通用聊天框的预判。他预测，未来五年内90%的企业AI Token将消耗在未经人类显式提示的异步后台任务中。用户不需要在对话框中频繁交互，而是通过仪表板、异常列表和审批流对后台Agent处理完毕的合同审核、安全遥测等长流程结果进行裁决，专用垂直界面仍将占据主流地位。 组织原生改造与现代创业信任防线 (57:30) 在企业内部落地层面，Box通过严格的底层数据规整（确保会议纪要、路线图与财报统一沉淀在Box和核心系统中），为Agent提供了结构良好的索引环境。Box设立了AI卓越中心，配备前线部署工程师（FDE），将Token消耗量作为诊断指标以筛查高价值工作流与无效算力浪费，并通过现场同行演示推动工程团队在部分业务代码交付速度上实现2至3倍的提升。 谈及当代创业，Levie指出当前仅凭2名创始人就能交付过去40人团队规模的功能，但开发效率的大众化意味着纯技术构建速度已不再构成壁垒。新创公司容易忽视合规审查，而成熟企业绝不能以牺牲数据治理与合规来换取极致提效。由于任何创新理念都会瞬间引来数家顶级资本支持的竞争者，最终决定万亿级应用层归属的不是单纯生成Token的能力，而是深植于合规工作流的分发渠道、安全信任与企业级部署能力。 Levie的论述展现了从前沿算法崇拜向底层工程与企业流程扩散回归的行业转向，但也体现了一位成熟SaaS上市公司负责人的防御性视角。尽管他充分论证了应用层、安全隔离与既有分发渠道的护城河，但在开源模型性能持续跃升与基础模型厂商不断向应用层试探渗透的双重挤压下，传统软件厂商能否长期维系其高附加值的定价权，仍是有待观察的未知数。

## Background 本期播客由红杉资本（Sequoia Capital）于 2026 年 9 月 15 日在其播客频道发布。节目围绕 Box 联合创始人兼 CEO Aaron Levie 展开，探讨成熟上市公司在季度财报压力下如何全方位拥抱 AI 重塑，以及为何从底层模型原始能力跨越至金融、医疗等企业核心工作流的应用层，蕴含着数万亿美元的巨大价值。 ## TL;DR Aaron Levie 深入驳斥了“应用层只是模型套壳”的陈腐偏见，指出模型能力越强大，其与企业真实工作流之间的鸿沟就越宽广，深耕组织流程与专有数据的应用层正在成长为新型“Neolabs”。他详细拆解了 Box 基于数千亿份非结构化文档构建的专属 Agent Harness 架构，证明自研调度框架在准确率和延迟上均能击败直接调用通用基础大模型 API，并断言未来五年内企业 90% 的 Token 消耗都将来自用户未主动发起的后台异步任务。此外，Levie 深度剖析了模型实验室价格补贴的不可持续性与模型中立路由的必然性，并透彻解释了为何代码能迅速实现自动化扩散，而普通知识工作因缺失“通用 GitHub”、深陷复杂权限隔离以及面临“AI 烂活（Work Slop）”的信任危机，其企业级渗透将注定漫长而艰巨。 ## Atmosphere & style 对谈节奏紧凑且妙趣横生，开场借由调侃红杉合伙人 Doug Leone 的“根管治疗”玩笑迅速奠定了熟人对坐、毫无公关辞令的松弛基调。Aaron Levie 语速飞快、极具感染力与网感，生动展露了一位既重度沉迷 Twitter 前沿信息流、又在一线摸爬滚打二十年的 SaaS 老将风范。整场交流在对新生代双人团队恐怖杠杆率的由衷感叹与对企业遗留数据泥潭的冷静审视之间自如拉扯，散发着硅谷顶尖技术极客与硬核商业实用主义深度交融的迷人张力。 ## Bulletin - `10:00` 底层基础模型价格战与补贴不可持续，随着开源模型与算力巨头将推理毛利压至极限，企业级 AI 的核心经济价值将不可逆转地沉淀在深耕垂直工作流的应用层。 - `24:35` 针对大模型直接调用延迟高且易出错的问题，Box 自研的 Agent Harness 调度框架通过深度绑定企业文件与权限体系，在生产准确率和响应速度上显著击败了直接调用前沿模型原始 API 的方案。 - `45:00` 企业 AI 的交互范式正从即时问答对话框演进为后台异步流水线，预测未来 5 年内 90% 的企业 Token 消耗都将产生于用户从未主动发起的后台自动化任务与数据抽取流中。 - `52:12` 编程的快速普及源于代码的高价值密度与闭环反馈，但在企业知识工作中不存在通用的“GitHub”，散落在遗留系统中的数据与极其复杂的权限控制将使企业落地周期远比硅谷预期的漫长。 ## Key takeaways - `01:30` 应用层软件并非脆弱的“模型套壳”，而是正在蜕变为新型实验室（Neolabs），在通用的基础模型与企业复杂的真实业务流之间架起极其宽广深厚的落地桥梁。 - `05:00` 计算机 60 年的历史反复证明：底层基础设施越强大，生长于其上的上层应用市场规模就越庞大，正如 AWS 和 GCP 催生了数十亿美元营收的 Snowflake 和 Databricks 一样。 - `07:15` “绝不能让狐狸来看守鸡舍”：企业首要追求的是成本与效果的最优平衡，只有完全中立、在各类模型间毫无偏向的应用层平台，才能站在客户立场客观路由最具性价比的 Token。 - `10:00` 模型实验室数十亿美元的资本开支与即将面临的 IPO 毛利率考核，叠加开源力量对推理毛利的压缩，将终结 Token 补贴时代，迫使大部分价值向企业级应用层转移。 - `17:45` “长运行 Agent（Long-running Agent）”端到端接管核心业务流程将银行开户等耗时数周的工作压缩至一小时，但也引发了充斥空洞修辞、缺乏灵魂的“工作废料（Work Slop）”审美疲劳与信任危机。 - `20:00` 软件工程将代码视为消除业务瓶颈的实用性工具，因而能全盘拥抱自动化；但知识工作中的文档历来是判断专业能力与可靠性的信任代理，引发了职场对 AI 生成内容的普遍疑虑。 - `24:35` 应用层自建的专属 Agent Harness 能将企业搜索启发算法、动态分块与重排序工程化，在处理非结构化资产时的准确率和延迟表现均大幅优于直接调用通用大模型 API。 - `26:30` 在 Box 针对金融、生命科学等行业设计的企业级基准评测中，Google Gemini 凭借在复杂工具调用和知识工作流上的优化，展现出了显著超出其公开代码评测排名的综合性能。 - `30:00` 闭源巨模型与开源权重模型呈现共生分流：企业倾向于使用顶级闭源模型充当主脑负责复杂编排，并将海量长尾任务分发给廉价开源模型，推动两者调用量同步指数级暴增。 - `32:30` 由于企业环境中存在不可逾越的信息防火墙与高频变动的权限控制，数据无法像学术设想那样轻易“烧录”进模型权重，外部上下文检索仍是企业级场景的核心主流。 - `36:35` 开箱即用的基础模型在复杂商业文档抽取上的准确率通常只有 70% 左右，必须依靠紧贴企业文件系统的 Harness 架构自动执行“爬山算法”，才能跨越至 97% 以上的生产级商用门槛。 - `40:00` SaaS 平台面临双轨生存法则：既要打造基准表现比通用模型高出 10-20% 的垂直内置 Agent，又要通过 MCP 等协议彻底将自身“无头化”（Headless）开放，从更高频的系统调用中获益。 - `45:00` 未来的企业 AI 界面将由任务队列、仪表盘与审批清单主导，预计 5 年内 90% 的企业 Token 消耗都将由无需人类在对话框中主动发起的后台异步 Agent 自动产生。 - `50:00` 从仅受文本约束的编程，到受海量外部人际与现实条件制约的销售，不同知识工种处于一条清晰的“自动化分布谱系”之上，直接决定了 AI 技术的渗透阻力。 - `52:12` 企业知识工作中缺乏类似 GitHub 的标准化数据接入点，散落在陈旧系统中的数据孤岛与繁杂的权限访问审批，导致真实世界的技术扩散速度远慢于硅谷极客的乐观预期。 - `01:02:40` 当前 AI 赋予了初创团队极高的生产力杠杆，2 名工程师即可构建过去 40 人团队开发的企业级产品，但这也彻底消除了产品的隐蔽窗口期，引发了残酷的早期同质化竞争。 - `01:04:13` 当 AI 极大拉低纯粹的代码编写门槛后，AI 时代的竞争决胜重心必然不可逆转地转向“企业级渗透力（Enterprise Diffusion）”，万亿应用层价值将属于深耕行业场景与客户交付能力的团队。 ## Detailed walkthrough ### `00:00` 信息流与“连线感”，以及红杉的“根管治疗”玩笑 Aaron Levie 开场便分享了他对当今信息获取方式的强烈观察。他甚至想向大学二三年级的年轻人发送紧急警报：“立刻去关注 Twitter（X）上的这 20 个账号，赶紧注册，因为这会直接决定你的职业生涯。”在他看来，一个人是领先时代一年还是落后时代一年，完全取决于其信息流（feed）的质量与输入源。Levie 表示自己的信息流高度“连线接入（wired in）”，当他遇到一些 20 出头的年轻人说自己偶尔“在网上看到几篇文章”时，他感到极其困惑：“‘看到几篇’是什么意思？难道就全凭运气等人给你发邮件吗？”保持对前沿技术的极度敏锐和动手玩转一切的习惯，是他始终保持敏捷的秘诀。 随后话题转向播客本身。作为该节目的忠实观众，Levie 半开玩笑地调侃起红杉资本著名的合伙人 Doug Leone，询问播客什么时候能聊聊“根管治疗（root canal）”的痛苦经历。主持人也配合接梗，提议大家可以办一场类似网络吃辣挑战节目《Hot Ones》的访谈：让创始人们一边坐在牙医椅上忍受根管治疗，一边在剧痛中阐述自己的商业战略，而 Doug Leone 则在另一头冷眼审视。这场轻松的玩笑化解了严肃的企业转型话题，也勾勒出硅谷老派风投和成熟创业者之间亦师亦友的默契。 --- ### `01:30` 撕掉“套壳”标签：应用层公司正在成为真正的“Neolabs” 面对主持人提出的命题——“应用层公司（Application companies）是否是如今最炙手可热的新型实验室（Neolabs）？”Levie 毫不犹豫地给出了肯定回答。如果两年前提出这个观点，很多人会觉得不可理喻，但当下由于开源模型的飞速演进，市场格局正在发生剧变。此前被硅谷技术原教旨主义者嗤之以鼻的所谓“模型套壳公司（model wrappers）”，如今正展现出极强的韧性与生命力。 Levie 指出，资本市场上至少有上万亿美元正在就两种未来路线进行豪赌： 1. **纯模型派（Superintelligence）**：认为底座模型的通用超级智能将直接通吃一切，模型与企业工作流之间的桥梁非常狭窄。 2. **应用层/Neolab 派**：认为模型能力与企业真实工作流之间的鸿沟（The Bridge）极其宽阔且深厚，必须有专门的应用层去深耕组织内部的复杂流程。 现实表明，模型越是强大，它与企业具体工作流之间的空白就越明显。当应用层公司填补了这个鸿沟、积累了足够多的垂直领域专有数据和业务经验后，它们自身就会形成飞轮，甚至反哺并训练专有模型，从而演变为真正的“Neolab”。Levie 透露，Box 内部其实早已低调设立了自己的“Box Labs”，尽管目前的核心策略依然是让智能体（Agent）在任何主流底层模型上都能跑出极佳表现，但未来一定会针对特定客户与场景训练专属模型。 ```mermaid flowchart TD accTitle: 模型与企业真实工作流之间的桥梁 accDescr: 展示底层通用模型如何通过应用层桥梁连接企业遗留系统与核心工作流 subgraph LLM["底层智能基础设施 (Foundation Models)"] M1["通用 LLM (OpenAI / Anthropic / 开源等)"] end subgraph Bridge["应用层桥梁 (The Enterprise Bridge / Neolab)"] direction TB B1["权限控制与文件检索 (Search & Permissions)"] B2["人机协同节点 (Human-in-the-Loop)"] B3["成本/模型中立路由 (Cost & Model Routing)"] B4["领域专业知识注入 (Domain Expertise)"] end subgraph Workflow["企业核心工作流 (Enterprise Realities)"] direction TB W1["遗留业务系统 (Legacy ERP / CRM)"] W2["高门槛合规流程 (银行 / 医疗 / 律所)"] W3["组织变革与流程延迟 (Change Management)"] end M1 --> Bridge Bridge --> Workflow ``` --- ### `05:00` 历史的重演：基础设施越强大，软件应用层就越庞大 硅谷的核心圈子往往被“纯科研思维所麻醉（research-pilled）”，总倾向于假设只要底座模型的智力无限提升，一切交互形态就都不复存在。然而在真实商业世界中，即便模型拥有世界上最顶级的超级智能，要把它落地到一家银行、药企或律所的日常运作中，依然要跨过无数脏活累活：连接杂乱的外部数据系统、设置人机协同审批节点、处理流程中的等待延迟（让智能体在适当时机被唤醒）、应对复杂的企业变革管理，以及对接尚未现代化的陈旧遗留系统。这些极其依赖运营细节的“阻挡与擒抱（blocking and tackling）”工作，恰恰是前沿研究型实验室最不愿意碰的领域。 Levie 强调，这种基础设施与应用层之间的关系在计算机软件 60 年的历史中屡见不鲜： - 云计算巨头（AWS、GCP、Azure）创造了数万亿美元的底层基础设施市值； - 但基于这些基础设施生长出来的上层软件同样创造了数万亿美元的市值。 - 十年前如果只看 AWS 和 GCP 早期构建的数据组件，没人会预测到 Snowflake 或 Databricks 的崛起——当时人们同样会问：“底层基础设施已经能存取数据了，为什么还要多花几十亿营收去买其他软件？” 事实证明，智能领域也将遵循完全相同的演进轨迹：底座大模型固然极度值钱，但将这些模型注入到金融、医疗保健和政府具体流程中的应用软件，其规模同样不可估量。对于顶尖模型实验室而言，未来两到五年将面临严峻的战略抉择：究竟是冒着失去开发者信任的风险强行向上层业务渗透，还是主动将应用生态完全拱手相让。 --- ### `07:15` “扩散经济学”与不让狐狸守鸡舍：模型中立路由的必然性 面对这轮变革，风险投资家们陷入了存在性焦虑——究竟是把数十亿美元继续砸向 Anthropic 这类底座模型，还是下注应用层？作为 Box 的掌舵人，Levie 承认自己做出了高度集中的下注：坚信企业永远需要能够理解其特定业务流、并能安全访问其核心数据的应用软件。 在讨论通用聊天机器人（Chatbot）与智能体工作流（Agentic Workflow）的本质区别时，Levie 指出： - **Chatbot** 具备通用的水平能力，但它无法主动感知在工作流的哪个精确节点该向哪位员工发出提醒，也无法原生获取企业深层孤岛中的特定数据。 - **Agentic Workflow** 则需要大量的垂直领域知识配置。除非顶级模型实验室疯狂扩张到“十万人规模”，否则它们绝无可能亲自覆盖成千上万家企业极其个性化的流程再造。因此，围绕“扩散（Diffusion）”的经济形态——无论是 Forward Deployed Engineer（FDE，前线部署工程师）还是专属实施团队——都将孕育巨大的市场空间。 最后，Levie 抛出了一个核心博弈论观点：**“绝不能让狐狸来看守鸡舍（don't want the fox to be guarding the hen house）”**。在构建智能体系统时，企业首要的需求是在保证准确率的前提下实现任务成本的最优化。如果让出售 Token 的模型实验室来决定使用哪种模型、分配多少计算量，这里存在天生的利益冲突；唯有完全中立、在十几种模型之间毫无偏向的应用层平台，才能真正站在客户利益最大化的立场上，为每一个业务请求路由最经济、最高效的 Token。 ### `10:00` 模型补贴不可持续：应用层的崛起与博弈论 关于底层模型厂商对 Token 的低价倾销或补贴政策，Aaron Levie 态度鲜明：这种靠补贴换取市场份额的做法绝不可能长期持续。目前各大模型实验室在资本开支上动辄投入数百亿美元，甚至还未完全计入昂贵的模型训练成本。一旦这些初创公司走向公开资本市场，就必须屈服于“资本主义的基本法则”，直面毛利率（Gross Margin）的严格拷问。 对话中主持人提出疑问：目前模型厂商在推理（Inference）上的毛利其实相当可观，而且如果能用高价的企业级 API 利润来交叉补贴第一方应用产品，商业逻辑是否成立？Aaron 指出这本身存在博弈悖论——如果客户全都涌向了被补贴的第一方应用产品，API 收入被抽空，这个闭环就会瞬间失衡。更深远的影响来自具有完全不同博弈诉求的外部力量：Meta、SpaceX、中国的大模型力量，以及 NVIDIA。这些玩家根本不需要像 OpenAI 或 Anthropic 那样依靠极高的推理毛利率来维持公司运转；只要能拉动算力集群或实现开源战略，他们甚至愿意把推理毛利压缩到 10%。在没有不可逾越的技术壁垒前提下，同等能力的 Token 成本必然会持续下降。 这种成本下行曲线最终将大部分价值推向了应用层（Application Layer）。Aaron 并不认同未来会有“一两家顶尖实验室独吞整个生态 95% 价值”的假说，反而认为一个各层级均有健康利润的动态生态对巨头本身也是幸事。因为一旦某家实验室独占了所有机器智能，“反垄断的铁拳甚至直接将你收归国有（Nationalized）的风险就会接踵而至”。 ```mermaid flowchart TD accTitle: 基础模型与应用层价值流转 accDescr: 展示计算成本下行如何将生态价值从底层模型沉淀至上层应用的逻辑流转。 subgraph Infrastructure["算力与开源冲击"] A["Meta / NVIDIA / 中国等外部力量"] -->|"打平算力成本即可<br/>接受低毛利推断"| B["Token 边际成本迅速下行"] end subgraph FrontierLabs["闭源模型实验室困境"] C["百亿级训练支出"] --> D["IPO 与资本市场毛利压力"] D -->|"无法长期持续"| E["结束 Token 隐性补贴"] end subgraph ApplicationLayer["价值沉淀端"] B --> F["应用层（Application Layer）"] E --> F F --> G["企业级工作流、权限控制与沉浸式 Agent"] end ``` - **资本重压**：模型实验室目前面临数百亿美元的训练与基础设施开销，一旦进入公开市场，补贴模式难以为继。 - **差异化博弈玩家**：Meta 开源模型、中国的大模型竞争、以及硬件巨头（如 NVIDIA）并不依赖高额推理毛利生存，会将推理利润率压至极限。 - **反垄断防御机制**：即便对顶尖实验室而言，单一垄断也是危险的，分散竞争能避免成为监管和国家化处置的众矢之的。 --- ### `12:35` Box 的演进与 AI 探索：从云端存储到 ChatGPT 的“顿悟时刻” 谈到 Box 的发展历程，Aaron 回忆起当年创立公司的质朴初衷：为人们提供一种在云端安全存储和分享数据的方案。这个看似简单的点子踩中了时代的脉搏，Box 迅速转向服务企业级客户（Enterprise）。在那个企业集体从本地服务器（On-premises）向云端迁移的大潮中，Box 成了企业存放海量非结构化数据（如商业合同、财务报表、营销素材和研发专利资料）的安全枢纽。 事实上，Box 对 AI 技术的摸索可以追溯到 2015 年。在 2015 至 2018 年那场被称作现代“AI 寒冬”的周期里，Box 曾尝试过利用早期的深度学习模型做图像自动分类与 OCR 文字识别。这些能力在企业级场景中确实有用，但弊端极其致命：不仅推理成本奇高无比，而且每一种特定的工作流都需要单独训练一个高度专有化的模型，通用性极差，Box 最终不得不将这套方案束之高阁。 转机出现在 GPT 系列模型的演化中。Box 内部通过 Hackathon 开始试验笔记产品中的自动补齐（Type-ahead），并尝试用于安全检测与文本分类。直到 ChatGPT 问世，Aaron 形容团队迎来了真正的“头脑爆炸时刻（Head exploding moment）”。OpenAI 证明了一种全新的交互范式：任何普通人都可以通过极其直观的问答对话，让机器处理复杂度与长度难以想象的推理任务。 --- ### `15:10` 唤醒沉睡的非结构化数据：Box Agent 与核心商业场景 ChatGPT 爆发后，Box 展开了一场极其迅速且教科书式的全员战略转向：迅速抽调全公司最顶尖的工程技术人员组建核心战役团队，设立每日复盘同步机制，自底向上构建了 Box 的 AI 技术栈，并推出了 Box Agent。 Box 的核心底气在于其沉睡的巨大资产——其平台上存储着企业客户的“数千亿份文件（Hundreds of billions of files）”。这些非结构化数据是企业的命脉，涵盖借贷凭证、采购合同、科研底稿，但过去企业根本不知道里面到底写了什么，除非有人肉眼一份份翻阅。Box Agent 的核心价值在于充当架设在这些数据之上的智能执行体：它们可以跨海量文件回答精密问题、预处理文件并抽取元数据（Metadata）沉淀至结构化数据库，甚至自动触发跨系统工作流。在内容生成端，除了 Box Notes 原生的协作生成能力外，Box 选择与 OpenAI 和 Anthropic 等前沿厂商深度结盟，由 Box Agent 调度底层模型，直接根据企业资料自动化生成高保真演示文稿（PowerPoint）与专业报告。 围绕 Box 的核心口号“你的业务蕴藏在内容之中，用 AI 释放它（Your business lives in content, unleash it with AI）”，Aaron 阐述了当前最立竿见影的“全垒打级场景”： - **百万级合同与研发资料透视**：企业往往积累了数百万份合同或科研报告，过去由于人工成本限制根本无法全面阅读；现在通过 Agent 可以瞬间抽离出所有关键条款与参数，转化为结构化数据供即时检索和分析。 - **从非结构化到结构化引擎**：把零散的文本资产转化为可用于 SQL 查询、自动化分析和审计的干净数据底座。 --- ### `17:45` “长运行 Agent”之年与企业内容里的“工作废料”（Work Slop）困境 如果说文档速读与数据抽取是当下的成熟场景，那么 Aaron 真正为之兴奋的未来，是“长运行 Agent（Long-running Agent）”端到端接管企业核心业务流程。今年正是长运行 Agent 的关键落地之年。以银行的客户入职审查（Onboarding）为例，过去的往返流转往往需要一到两周；而现在的架构是由长运行 Agent 自动执行从初审、证件解析到合规校对的绝大部分中间步骤，仅在关键决策和风险拦截节点交由人工审批，整个冗长的周期因此被压缩至短短一个小时以内。 然而，当话题转向生成内容的质量时，现场的气氛多了一层讽刺与现实的张力。主持人尖锐地指出：在代码领域，开发者群体几乎毫无心理包袱地全盘拥抱了 AI 辅助编程；但在内容创作和日常办公领域，人们对 AI 生成文本却存在一种强烈的“过敏反应”，甚至在社交媒体上引发了广泛针对“工作废料（Work Slop）”的声讨。 Aaron 暂时摘下 Box CEO 的官方身份，从一个真实受害者的视角共情道：“我最近收到的董事会汇报材料（Board decks）里，有些居然完全是用 AI 生成的，这简直快要把我逼疯了。”在 Aaron 看来，这种“工作废料”的泛滥暴露出内容扩散与代码扩散在深层逻辑上的巨大断层。代码有着极高确定性的运行反馈闭环，且高阶工程师对代码架构有着极高的审美与品味，代码对他们而言兼具工程与艺术属性；而充斥着空洞修辞、缺乏灵魂的 AI 商业文案，正在引发一场知识工作者对虚假生产力的群体审美疲劳。 ### `20:00` 代码的实用主义与知识工作的“AI 烂活”信任危机 在软件工程领域，除了极少数追求算法极致的计算机科学家之外，对全世界绝大多数人而言，编写代码本质上是一种纯粹的“公用事业”（Utility）。它的核心目标是完成任务、实现自动化，或者搭建一个界面供人点击进入下一步。因此，尽管前端设计和系统架构依然讲究工程品味，且企业绝不能承受代码漏洞，但如果今天一个 AI Agent 能自动生成整套后端或前端系统，人们不仅完全接受，甚至会觉得“这再好不过了（it's not only acceptable, it's preferable）”，因为代码本身往往正是阻碍业务向前推进的瓶颈。 然而，知识工作（Knowledge Work）的逻辑却大相径庭。在现代企业与社会协作中，一份商业报告或演示文稿（Deck）承载的不仅仅是文字信息，它是一个信任的代理指标（proxy）。当人们阅读一份方案时，潜意识里在评估：“我能否信任这个人去执行这项任务？他是否真正理解这个业务主题并能交付成果？”当所谓的“AI 烂活”（Work Slop）充斥职场时，受众便无法辨别其中的思考过程到底有多少来自作者本人，又有多少来自 AI。这导致了一种普遍的集体认知悖论：每个人都在私下用 AI 生成粗糙的草稿和决策头脑风暴，但一旦从别人那里收到类似内容，心里就会犯嘀咕：“我真的能信任你吗？” 这种信任重构可能需要社会在未来 3 到 5 年内经历漫长阵痛才能逐渐厘清。这就像人们看到复杂的财务模型时，明知那是 Excel 宏或公式算出来的，却依然认可它作为讨论基础；如今的商业战略演示文稿也面临着同样的转变。职场正在经历一场对“人的角色究竟是什么”的重估：内容本身到底是衡量一个人专业深度的标尺，还是仅仅用来预测其执行力的载体？我们正处在一个标准极度模糊的混乱探索期。 - **代码工具论与知识工作信用体系**：在绝大多数业务场景下，代码的目标是消除阻碍并实现自动化；而在知识工作领域，文档历来是衡量员工专业能力和执行可信度的“代理凭证”。 - **职场认知双标（Work Slop Paradox）**：员工普遍依赖 AI 辅助自己进行脑暴和分析，但当审查他人的交付成果时，却对浓厚的 AI 痕迹产生强烈的信任疑虑。 - **财务模型类比**：自动化计算早已被商业社会接受（如财务宏），但在战略规划和分析报告中，人类思考与机器生成的界限尚未形成社会共识。 --- ### `22:15` 德鲁肯米勒专栏引发的思考：AI 时代的认知负荷与评判转向 对话转向了对知名投资人斯坦·德鲁肯米勒（Stan Druckenmiller）在《华尔街日报》发表专栏文章的讨论。尽管在 AI 检测工具 Pangram 的测试中，该文的 AI 生成概率被打出了 100%，且毫无悬念地展现出高度结构化的机器风格，但主持人却表示自己非常喜欢这篇文章。这成为一个极具启发性的反例：为什么读者在看到绝大多数充满 AI 味的内容时会产生本能的“过敏反应”，却唯独对德鲁肯米勒的文章网开一面？文章中甚至罕见地没有出现典型的 AI 句式（例如反复出现的“不是 X，而是 Y（it's not X, it's Y）”），也没有泛滥的破折号。这种宽容的根本原因在于“他是斯坦（It's Stan, therefore I trust in Stan）”——作者原生的商业声誉在阅读前就已经建立了绝对的信任垫脚石。 这种现象加剧了人们日常阅读时的心理消耗。如今在社交平台（如 X / Twitter）上阅读长文时，读者往往需要承担双倍的认知工作量（2x mental processing）：第一层是在理解文章的实际内容，第二层则在后台高速计算“这到底是他自己写的，还是直接把 Claude 的 Prompt 回答粘贴上来的？”读者不得不纠结于这一判断究竟应该给发帖人加分还是减分。面对这种普遍的信任危机，Aaron Levie 半开玩笑地感叹：“我最不想干的职业就是大学教授，我肯定会立刻辞职，因为你根本不可能搞清楚学生到底自己做了什么。” 过去人们常用“计算器”来类比 AI，但这一类比在当下正在失效。计算器处理的是极其有限的单点运算，使用者依然需要亲自分解并串联所有复杂步骤；而现代生成式模型却能在一瞬间同时处理十几个环环相扣的高难度认知任务。 - **德鲁肯米勒案例启示**：先验的名誉和声望成为中和“AI 烂活感”的解毒剂；即使 Pangram 判定为 100% AI，读者因对作者本人的底层信任依然高度认可其商业价值。 - **双重认知负荷（2x Mental Work）**：读者在消费专业内容时，被迫在“汲取信息实质”与“逆向工程作者是否使用 AI 及其可信度”之间双线消耗脑力。 - **计算器类比的失效**：传统计算工具只接管单一计算步骤，而大语言模型跨越了逻辑梳理、语言组织与方案整合的多重任务流。 --- ### `24:35` Box Agent Harness 架构：为何应用层调度优于通用大模型 API 讨论由此切入应用层（Applied Layer）的核心技术载体——Agent 调度框架（Harness）。Box 并未停留在大模型的调用者层面，而是构建了一整套深度绑定自身底层架构的专属 Agent Harness。大模型实验室（Labs）即便掌握通用的推理能力，也无法替代应用厂商对自身底层文件系统、权限树与企业级搜索引擎行为的微观洞察。在企业日常操作中，用户在搜索结果的 10 个文件中如何权衡并挑选出最关键的那一份，背后涉及高度专业化的启发式决策（heuristics）。 Box 构建的 Agent Harness 正是把这套对企业内部工作流的深层领域认知算法化。该框架直接挂载在 Box 的文件与权限系统之上，内置了一整套工具集：实时文本抽取、动态分块（Chunking）、即时向量化（Embeddings）、多级组合检索以及重排序（Re-ranking）。以 Aaron Levie 自己的 Box 账号为例，里面沉淀了公司 20 年来积累的数千万份文件，而 Box Agent 可以在用户提问后，并发执行多路检索、迅速重排序、精准定位最关键信息，并在必要时通读全文进行多步推理。 实测表明，将底层数据通过 Box Agent Harness 处理后再配合模型生成，在准确率（Accuracy）和延迟（Latency）两个维度上，均显著击败了直接将 Box 原始 API 暴露给 Claude 或 OpenAI 等通用模型的做法。其核心护城河正是在于应用层对工作流的极致调优与上下文工程。 ```mermaid flowchart TD accTitle: Box Agent Harness 检索与处理流程 accDescr: 展示用户查询如何经由 Box Agent 架构进行权限校验、多路检索与模型生成 subgraph UserLayer["用户交互层"] Q["用户自然语言查询"] end subgraph Harness["Box Agent Harness 调度框架"] Perm["企业权限架构验证 (Permissions)"] MultiSearch["并发多路搜索 (Multi-Search)"] ReRank["上下文重排序 (Re-ranking)"] DocTools["即时处理工具集\n(动态抽取 / 分块 / 实时向量化)"] DeepRead["关键全文深度阅读与推理"] end subgraph ModelLayer["模型路由与交付"] LLM["前沿大模型 (Claude / OpenAI / Gemini)"] Resp["高准确度 / 低延迟企业级回答"] end Q --> Perm Perm --> MultiSearch MultiSearch --> ReRank ReRank --> DocTools DocTools --> DeepRead DeepRead --> LLM LLM --> Resp ``` - **海量非结构化资产处理能力**：调度框架能够直接横跨 Levie 个人账户中数千万份历史企业文档进行秒级检索与问答。 - **端到端流程优势**：相较于将通用大模型直接挂载在未经调优的 API 上，自建 Harness 凭借对搜索启发算法、实时分块与动态重排序的控制，在响应延迟和回答准确率上展现出断层级优势。 --- ### `26:30` 企业级评测体系与模型竞技场：Gemini 的超预期表现 在模型选型与集成上，Box 建立了一套严苛的企业级评测（Evals）系统，包含上百项横向对比测试： 1. **复杂业务评测（Complex Work Eval）**：针对生命科学（Life Sciences）、金融服务（Financial Services）、公共部门及高科技等垂直行业设计的重度文档评测集。典型任务如“给定 5 份跨业务的长文档及一组复合问题，检验模型答案的精确度”。 2. **保留样本评测（Holdback Eval）**：基于 Box 内部员工在日常真实业务环境中处理海量数据的脱敏日志，进行端到端的实战效果再检验。 通过这些基准，Box 能够敏锐捕捉到模型在企业知识场景下哪怕只有 0.5%（半个百分点）的细微能力跃升。基于成本、速度与精度的平衡，Box 会向用户推送默认模型，同时也通过“模型花园（Model Garden）”向客户开放自主切换主流模型的权限。 在评估当前前沿基础模型的竞争格局时，Levie 指出，各大模型在 Box 企业场景下的表现通常与代码能力（Coding Benchmarks）高度正相关。但存在一个重大例外：**Google Gemini 在 Box 实际业务中的表现显著超出了其代码评测所反映的水平**。这很可能得益于 Gemini 在 Google 生态下针对复杂工具调用（Tool Use）和通用知识工作场景所做的针对性优化。整体而言，当前模型梯队（包括 Fable 5.1 / Claude 系列、Gemini、Grok、GPT-5.x 等）正处于一场针锋相对的白刃战之中，行业各评测基准（如 GDPVal、Mercore Apex）的排名差距往往微乎其微。 - **多维严苛评测机制**：通过特定垂直行业复杂文档基准与内部员工真实使用基准，高频监控模型在微小增量（0.5%）上的表现演进。 - **Gemini 的异军突起**：在企业知识处理与工具调用任务中，Gemini 展现出了远优于其公开代码基准排名的非对称优势。 - **模型梯队咬合紧密**：前沿模型能力高度趋同，企业软件架构必须具备快速兼容多模型的敏捷路由能力。 --- ### `28:15` 真实企业级 Token 消耗结构与开源权重模型的落地现状 对于企业客户究竟是倾向于使用默认模型还是自主指定模型，Levie 分析指出，从调用量（Volume of Tokens）的构成来看，绝大多数终端用户在进行日常知识搜索与问答时，会直接依赖 Box 调优后的默认模型，因为开箱即用且体验最佳。 然而，企业内真正消耗海量 Token 的核心场景并非前端人机交互，而是后台的**自动化工作流 Agent（Workflow Agents）与大规模非结构化数据抽取（Data Extraction）**。在这一场景下，大客户往往会极为严苛地进行投入产出评测（ROI Evals）。例如，企业会明确要求在严格的成本上限内，数据抽取准确率必须坚决达到 98%。为了满足这类关键任务，Box 会派遣前线部署工程师（Forward Deployed Engineers, FDE）深入客户的数据与业务流，横跨 5 种不同模型展开实测，最终完全由评测数据（Eval Results）决定采购哪家模型。 关于开源/开放权重（Open-weight）模型在企业中的渗透现状，Levie 给出了极具现实感的观察：“它的普及率比外界预期的要高，但低于大企业嘴上想要的，并且远远低于 5 年后的未来水平。”推动企业探索开源模型的核心动力除了显而易见的成本考量外，有超过 30% 实际上源自高管的尝鲜诉求与心理溢价（Sexiness）——连 Fortune 500 强企业的首席信息官（CIO）也会因为行业热度，极力要求在生产流程中亲自试水如 GLM 等开源前沿模型。 - **Token 消耗真实重心地图**：前端人机对话仅占 Token 消耗的表层，后台批量数据抽取与业务流代理才是驱动数以亿计 Token 流水的主力引擎。 - **FDE 支持下的硬核选型**：面对 98% 准确率门槛的企业关键链路，模型选择完全交由现场严苛的基准实测数据决断，品牌溢价在此让位于指标考核。 - **开源模型的企业心理学**：除了算力与私有化成本驱动外，企业技术决策层对开源新锐模型（如 GLM）的技术尝鲜欲望占据了至少三成以上的采纳推力。 ### `30:00` 开源与闭源的共生悖论：算力编排与代币分流 在模型选型上，很多企业宣称自己在积极探索开源模型，但 Aaron Levie 坦言，在现阶段很多场景下，像 Gemini 或是近期大幅降价的闭源轻量模型，在同等成本下完全能达到相同的效果。企业之所以执意使用开源，更多是为了“建立对冲”（hedged），防止被单一供应商彻底绑定。现阶段的开源权重模型依然存在明显的工程痛点，比如有时在代币利用效率（token efficiency）上不如顶级闭源模型，甚至偶尔会出现“推理链跑到一半突然莫名其妙蹦出一句中文”的荒谬 Bug——这种不可控的稳定性问题放在严谨的银行金融业务中是绝对致命的。但长远来看，随着开源权重不断将推理成本打磨到极限，企业必然会将那些对单位成本极度敏感的成熟工作负载剥离（peel off）出来，迁移至开源模型。 Aaron 极为推崇 Decagon 创始人 Jesse 提出的一种行业悖论：媒体和外界可能会非常困惑，为什么 Anthropic 和 OpenAI 的闭源模型收入正呈现指数级暴增，而与此同时开源权重模型的调用量也同样在呈指数级井喷？答案在于“整个企业级 AI 的蛋糕正在极速膨胀”，且两者之间形成了一种精妙的双向编排与代币分流结构。 ```mermaid flowchart TD accTitle: 模型分流编排架构 accDescr: 展示前沿闭源模型与廉价开源模型之间的两种协同编排模式。 subgraph ModeA ["模式 A：闭源编排模式"] A1["前沿闭源模型<br/>（高阶推理 / 任务调度）"] --> A2["开源 / 廉价模型<br/>（批量处理长尾任务）"] end subgraph ModeB ["模式 B：开源兜底模式"] B1["轻量开源智能体<br/>（处理日常默认任务）"] --> B2{"遇到极难任务？"} B2 -- 是 --> B3["升级弹出至闭源巨模型<br/>（攻坚复杂长上下文）"] B2 -- 否 --> B4["本地高效低成本结算"] end ``` 这种双向流动构建了一种独特的商业共生： - **模式 A（顶级模型统领）**：使用顶级的闭源模型（如 Claude 或 GPT 系列）充当主脑负责复杂 Orchestration（任务编排与分解），再将海量的长尾细分任务分发给极低成本的开源模型处理。 - **模式 B（轻量模型兜底）**：由轻量级智能体负责日常绝大部分常规任务，仅在遭遇超出能力阈值的极度棘手问题时，才动态“弹出”（pop out）至重型前沿模型攻坚。 - 最终的财务结果是：企业在闭源与开源上的实际预算支出可能是 50/50，但开源模型所承载的代币消耗量（Token Volume）却往往高达前者的 10 倍以上，两方都在各自的维度上赢得了市场。 --- ### `32:30` 记忆与个性化困境：为什么企业不能轻易将数据“烧录”进模型权重 当谈到模型记忆（Memory）、个性化与定制化时，主持人指出目前业界的主流依然是基于上下文检索的 RAG 架构（模型的权重本身保持冻结），但模型实验室的先锋研究者（例如 N-Gram 的 Dan）正致力于探索持续学习（Continual Learning），试图将企业上下文直接内化并“烘焙”（bake）到模型权重本身之中。 Aaron Levie 表达了对这一前沿方向的高度兴趣，但他随即抛出了企业级软件最残酷的底层现实：**权限管理与访问控制（Permissions and Access Controls）的极度复杂性**。他幽默地剖析了 AI 研究员与真实企业客户之间的思维鸿沟： > “AI 研究员往往会根据自己的工作方式去设想世界——‘我是个研究员，我能访问实验室里的一切数据，所以如果能训练一个专属于我个人世界的模型，那简直太棒了。’然后你对他说：‘请允许我为你介绍一位企业律师。’” 现实世界中的企业律师可能只能访问五个绝密项目，因为紧挨着他隔壁办公室的同事就在代理另一家直接竞争对手的对立案件。两堵墙之间绝不允许流通任何一份文档，必须设立牢不可破的“信息防火墙”（Ethical Wall）。如果将数据直接训练进权重，一旦员工的权限每天都在动态变更、被移出或加入某个项目，模型根本无法在权重层面实时撤销或隔离信息。虽然 Andrej Karpathy 曾设想过一种理想范式——将模型内部死记硬背的事实全部剥离，仅保留其纯粹的推理能力与思考风格（Ethos），再挂载外部检索系统，但在 Aaron 看来，企业的核心壁垒恰恰就是其具体而精密的 IP（专有数据）。 - **垂直领域的破局点**：在权限高度流动的金融、法律行业，直接微调权重成本极高且权限难解；但在像礼来制药（Eli Lilly）这样的药物研发场景中，几乎不存在人为的“政教分离式防火墙”，企业恨不得让所有科学家共享全部化学分子与实验数据，此时针对特定垂直领域深入训练专属模型（如 Trajectory、Applied Compute 或 Prime Intellect 的做法）就显得极具商业价值。 - **决策标准矩阵（T-Chart）**：决定信息放入上下文（Context）还是烧录进权重（Weights），取决于“数据变更频率”与“模型准确率跃迁”之间的权衡。如果数据更新频率极低，但内化进权重能带来决定性的推理飞跃，权重微调才成立。 --- ### `36:35` BoxLabs 的实战攻坚：Agent Harness 与 97% 准确率的“爬山算法” 针对 Box 内部的研发机制，Aaron 揭示了其应用 AI 团队（类似于 BoxLabs 的研究集群）正在攻克的前沿课题。Box 并不盲目追随底层基础大模型的预训练，而是将精尖力量集中在**应用层与智能体框架（Agent Harness）**上。 他们当前研究的核心是如何通过构建问题空间与专有数据的映射，让特定任务的 Agent 准确率再度跃升 10 个百分点。例如，团队正在开发一种基于 Agent Harness 层的“自动研究”（Auto-research）机制，针对客户特定数据集自动执行“爬山算法”（Hill climbing），持续优化对复杂问题的解析精度。 以银行业务为例，企业每天都会涌入大量长达上百页的复杂贷款合同与法律文档： - 如果直接调用市面上的开箱即用（Off-the-shelf）基础大模型，提取关键指标的准确率可能只有 70% 左右，这在金融风控中属于完全不可用的玩具水平。 - 真正能在企业生产环境中创造价值的门槛是 97% 乃至更高，这 70% 到 97% 之间的巨大鸿沟，必须依靠贴合企业底层文件系统、检索机制与业务逻辑的 Harness 架构系统去自动化跨越与工程化填平。 --- ### `38:15` 记录系统的未来：从“Claude Force”调侃到 Agent 时代的话语权之争 随着各类软件厂商纷纷宣称自己转型为 Agent 平台，社交媒体上掀起了一阵反思潮：许多企业买家抱怨“每个 SaaS 公司都想卖给我一个他们自家的专属智能体，但我根本不想要一堆割裂的智能体，我只希望你们的记录系统（System of Record）能好好开放、配合我自己的主智能体工作”。 面对这一争议，Aaron 打趣地提到了业界流传的热梗标签“#ClaudeForce”——暗讽类似 Salesforce（Agentforce）这样的传统巨头将旧有架构重新包装为智能体套件并极力兜售给资本市场。Aaron 调侃道，这种营销在刚听到的前三分钟你会觉得像个笑话，但四分钟后当你看到对方因此暴涨的股价时，你只会惊叹：“太绝妙的商战手段了，我们也必须跟进！”随后他更是爆梗调侃好莱坞巨星马修·麦康纳（Matthew McConaughey）为 Salesforce 拍摄的沉浸式广告： > “当你在电视上听到马修·麦康纳用那种迷人的嗓音亲口把这些词念出来时……那才是真正的‘顿悟时刻’（Aha moment）。天呐，他可太会卖软件了！他的声音用来推销‘记录系统’和‘智能体’简直是绝配。” 在玩笑背后，这一幕折射出当下企业软件的核心博弈：当底层智能体可以直接与基础数据交互时，传统的“记录系统”（System of Record）究竟会被降级为单纯的无脑数据库通道，还是能凭借其深度绑定的权限、元数据与业务流，反向牢牢卡住智能体生态的命脉。 ### `40:00` SaaS 平台的 AI 双轨生存法则：内置极致 Agent 与全面“无头化”开放 面对拥有专有数据与工作流的传统 SaaS 平台，Aaron Levie 提出了一个非生即死的论断：企业必须双管齐下，任何试图二选一的公司最终都会满盘皆输。第一条路径是打造一个在自家产品上表现极其惊艳的内置 Agent。这个 Agent 必须经过极度严苛的基准评估（eval-maxed），针对垂直工作流进行深度微调，在性能上能够明确无误地比那些直接调用外部系统的通用模型高出 10 到 20 个百分点。Levie 透露，自己每年会与数百家客户深度交流，其中至少有两打（24 次）以上的客户痛点让他惊呼是“颠覆性的顿悟时刻”。当主持人打趣催促他给出一个震撼案例时，Levie 略带尴尬地分享了一个关于“后台合规治理 Agent”的想法——它能在员工即将做出违背合规政策的行为时自动介入。虽然这引发了主持人关于企业合规话题枯燥乏味的笑场（“womp womp”），但 Levie 严肃地指出，企业首席合规官向来分身乏术，如果有一个 Agent 能够像合规主管坐在员工工位旁一样实时防御违规操作，这种价值在平台外部是根本无法复现的。 另一条生死路径则是全面“无头化”（Headless）。SaaS 平台必须毫不犹豫地将自身 API 彻底暴露给 Claude、ChatGPT 等前沿大模型生态，无论是通过确定性 API 还是 MCP（Model Context Protocol）标准协议，甚至直接将自家 Agent 无头化嵌入外部系统。针对很多人怀疑“记录系统”（System of Record）在此过程中会被边缘化的顾虑，Levie 认为这种悲观论调严重低估了全新场景所带来的纯增量价值空间。当主持人抛出这些记录系统是否会沦为令人厌烦的“收费站”（toll booth）业务时，Levie 风趣地回应自己反倒非常喜欢收费站模式——只要平台提供了组织工作流、数据安全和防护栏的核心价值，按使用量收费的商业模式就会自然成立。他举例称，由于现在可以通过 Claude 或 ChatGPT 借助 MCP 协议随时随地向 Salesforce 提问，他对该 CRM 系统的使用频率相比以往整整提高了一个数量级；他甚至跟 LinkedIn 的产品经理直言，如果 LinkedIn 允许他用 MCP 协议随时追踪企业 CIO 的变动和动态，他“愿意为此多付十倍的费用”。 ```mermaid flowchart TD accTitle: SaaS 平台双轨 AI 架构 accDescr: 展现 SaaS 平台同时推进内置极致 Agent 与外部 Headless 开放接入的双轨设计 subgraph DualTrack["SaaS 平台的 AI 双轨架构"] A["企业私有数据与工作流<br/>(System of Record)"] --> B["内置特化 Agent<br/>(精准微调 / 胜出通用模型 10-20%)"] A --> C["Headless 开放接入<br/>(REST API / MCP 协议暴露)"] end subgraph Consumption["多端价值释放"] B --> D["原生异步治理与嵌入式工作流"] C --> E["外部通用平台调用<br/>(Claude / ChatGPT / 自定义 Agent)"] end ``` - **客户洞察频次**：Levie 每年会见数百家企业客户，至少遇到过 24 次以上由客户痛点激发的突破性产品灵感。 - **性能护城河标准**：内置 Agent 必须在自身领域内通过严苛评测，证明其准确率或效能比通用外挂模型高出 10 到 20 个百分点。 - **使用强度暴增**：通过大模型 MCP 桥接访问后，Levie 对 Salesforce 的实际调用频率提高了整整一个数量级。 --- ### `45:00` 超越 Chatbot：后台异步 Agent 与 90% 企业 Token 的未来 针对“通用的聊天对话框（Chatbot）是否会成为未来主导 UI”的讨论，Levie 强调了垂直应用层（Applied Layer）具有不可替代的巨大发展空间。在他看来，全能型的聊天输入框就像每个软件必带的“搜索框”一样，必将成为横向与垂直软件的永久标配，用于满足用户单次、即兴（ad-hoc）的问答或生成需求。然而，企业运作的本质并不是即时问答，而是大量在后台持续运转的复杂流程与工作流——检查日志、逐行审核合同、分诊安全事件等。这些工作过去受限于人力成本而无法全面覆盖，绝非一个简单的“提问-回答”对话窗口所能承载。 未来的企业交互形态将转向以仪表盘（Dashboard）、自动化工作流、处理队列（Queue）和任务清单（Task list）为主的交互体系。如果横向通用大模型平台试图把所有垂直领域的这些组件与交互细节全盘包揽，软件界面很快会变得笨重不堪，丧失最初的简洁与优雅。相反，专注于法律（如 Harvey、Legora）、网络安全以及长程编程（如 Cognition、Factory）的垂直领域公司，深刻理解特定场景的专业语言、业务逻辑与功能布局，能够将这些后台异步 Agent 的产出无缝编排成清晰的界面。基于此，Levie 作出了一个极具前瞻性的预测：“五年后，企业中 90% 的 Token 消耗都将产生于用户从未主动发起的后台任务；员工最终看到的只是一个处理完毕的任务结果，只需要直接进行审查即可。” - **交互界面的演进**：企业 AI 的主流交互将从即时 Prompt 对话转向后台运行的任务队列、仪表盘与待审任务列表。 - **垂直领域的领跑代表**：法律科技领域的 Harvey 与 Legora，编程代理领域的 Cognition 与 Factory，均展现了深度融合业务工作流的优势。 - **关键预测**：未来 5 年内，企业中 90% 的模型 Token 消耗都将是由后台异步 Agent 自动触发并执行的，而非人类用户主动在对话框里键入指令启动。 --- ### `47:15` 代码扩散的奇迹与知识工作 AI 渗透的鸿沟 在谈到 AI 技术的全社会渗透速度时，主持人指出一个极为显著的断层现象：代码生成 Agent 几乎在极短时间内便以经济史上未曾见过的速度席卷并重塑了软件开发行业，然而大模型的“魔法”在其他知识工作领域的普及却显得迟缓得多。对此，Levie 从底层逻辑出发，对“编程与其他业务工种”之间的本质差异进行了抽丝剥茧的剖析。 首先，代码的实用价值几乎 100% 映射在所生成的文本本身之上。虽然前期的知识积累、会议讨论和架构思考具有极高价值，但最终能够产出可运行软件的唯一实体就是这些代码文本——假设一个顶尖程序员无需吃饭睡觉且能洞察需求，他的价值创造就完全与他敲出优质代码的时长呈 100% 正相关。其次，全球顶尖的 AI 实验室本身就是由工程师组成的，他们每天都在构建模型的过程中将代码编写作为核心基准（Benchmark）进行严密评估和迭代（“吃自己的狗粮”）。最后，软件开发者是技术素养最高的用户群体，当他们部署 Agent 系统遇到 Bug、或 MCP 服务器报错提示连接无效时，他们不需要打电话向 IT 部门报修，而是会自己排查并随手修好端口配置，这种极强的自我诊断与容错能力是任何其他普通知识工作者群体所完全不具备的。 - **代码价值的高纯度**：软件产品的效用与产出的优质代码文本之间存在几乎完全线性的直接对应关系。 - **AI 实验室的天然评测场**：前沿实验室的研究人员本身就是核心用户，他们日常利用自身的开发工作为代码模型提供全天候的评估反馈。 - **技术人员的自主排障优势**：面对 Agent 运行时出现的 Bug、网络断连或 MCP 端口错误，程序员能够自行排查解决，完全不依赖 IT 运维支持，消除了技术扩散的摩擦力。 ### `50:00` 编程与知识工作的自动化连续谱：从法务到销售的扩散鸿沟 编程之所以在 AI 自动化浪潮中一骑绝尘，除了代码本身属于结构化文本与闭环测试之外，第六个往往被忽视的关键特质是它极高的薪酬回报。在程序员这样一个高薪垂直领域，哪怕模型只能带来 10% 到 20% 的生产力提升，其商业价值就已经极其惊人，更不用说达到 5 倍的提效。若将编程具备的这五六项天然优势与其他知识工作进行对比，会呈现出一条清晰的“自动化分布谱系”。 在这条谱系上，与编程最接近的领域是法务。法务人员的大量价值创造同样发生在电脑屏幕前——审查合同、撰写法律文书、快速消化吞吐海量信息，因此法律科技赛道正在迎来爆发。但如果沿着谱系向另一端延伸，比如销售代表，其自动化难度便陡然上升。销售的核心经济价值在于说服外部客户采购软件、技术或者一辆卡特彼勒（Caterpillar）卡车。哪怕为其配备全球最顶尖的自动化工具，其生产力上限依然被现实世界的物理与人际条件严重锁死：客户愿不愿意回复邮件？客户想下周二见面还是今天？客户今年到底有没有预算？ ```mermaid flowchart LR accTitle: 知识工作自动化谱系与外部约束 accDescr: 展示从纯文本受限的编程到受外部现实严重约束的销售代表之间的自动化难度谱系 subgraph HIGH["高自动化适配度（数字/文本闭环）"] A["软件工程 / 编程<br/>代码库、测试闭环、高薪溢价"] B["法律合规 / 尽调<br/>长文本审阅、合同起草、案情分析"] end subgraph LOW["低自动化适配度（受外部实体约束）"] C["研发 / 跨职能协作<br/>多系统跨权限流转、组织协调"] D["一线销售 / 业务拓展<br/>外部预算、人际决策、现实反馈周期"] end A --> B B --> C C --> D ``` 这构成了知识工作的完整连续谱：一端是仅受文本和算力限制的纯数字任务，另一端则是受到无数不可控外部因素制约的人际工作。要让 AI 在真实世界落地，技术提供方必须将智能重塑为符合各行各业具体工作形态的样貌，完成沉重但不可或缺的变革管理与实施落地。 --- ### `52:12` 没有“交出你的 GitHub”：遗留数据泥潭、权限控制与应用层机遇 在工程领域，将代码接入 AI 极其丝滑，正如过去的行业常态：“在开发 Coding Agent 时甚至不需要注册流程，直接提示‘交出你的 GitHub 授权’（Give us your GitHub）即可”。然而在更广泛的企业知识工作中，根本不存在一个通用的“GitHub”。即使是“交出你的 Box”，由于 Box 目前的年度经常性收入（ARR）仅为 13 亿美元，绝大多数企业的数据依然散落在极为陈旧的本地服务器、遗留文件共享系统和难以兼容 Agent 架构的传统 IT 基础设施中。 企业级权限控制是另一个让硅谷工程师“听了就想睡觉”却致命关键的壁垒。在软件工程中，开发者通常能直接访问工作所需的绝大多数代码仓库；但在常规企业环境里，员工总在问同事：“嘿，Sally，你能把那个项目的文件夹权限开给我吗？”如果人类员工都无法顺畅获取上下文，如何确保 AI Agent 能够合规、安全且精准地访问这些数据碎片？ 这种现实带来了两大核心推论： - **技术扩散速度远慢于硅谷预期**：硅谷普遍对 AI 的扩散速度过于乐观，但 Aaron Levie 凭借多年企业级服务经验直言，真正的落地周期将比极客们想象的漫长得多。 - **应用层的巨大价值空间**：真正的 AI 价值不会在模型开源后自动喷涌而出，而是属于那些具备耐心、深厚行业 Know-How 以及愿意脚踏实地“扫街跑业务”（Pound pavement）的应用层公司。将 AI 转化为律师、销售、生命科学研究员或客户支持团队顺手好用的工具，构成了当下最大的创业机会。 --- ### `54:28` 押注非结构化数据二十年：置身 AI 风暴中心的内在逻辑 针对外界好奇 Aaron Levie 几乎现身于所有顶尖 AI 创业公司的股东名册（Cap table）并始终处于行业对话的中心，他将其归结为两大驱动力。首先是深达二十年的直觉准备：Box 专注处理企业的非结构化数据长达二十年，当大语言模型横空出世时，他能瞬间领会 Agent 跑在非结构化文档上的颠覆性潜力。事实上，Box 早在八九年前就曾尝试过类似方向，直到当下的基础模型突破才让设想真正具备落地条件。 由于 Box 的产品架构天然适配 Agent 检索和理解非结构化内容，公司果断选择“全盘押注”（Bet the company）。更重要的是，这种技术直觉迅速在客户那里得到了正向反馈循环——企业客户直截了当地表示：“如果 AI 能够读懂我们所有的文档并精准回答任何业务问题，那将具有无可匹敌的威力。” --- ### `55:54` Twitter 是 AI 的全球市政广场与个人助理的“CLI 困境” 驱动 Aaron 的第二大动力则是对前沿技术的纯粹好奇与沉迷。当被问及日常的信息获取渠道时，除了 Sequoia 和 Dwarkesh Patel 的播客之外，他坦言自己“95% 的大脑带宽都耗在 Twitter 上”。他形容自己就像一个滑稽的表情包——每天在信息流里永无止境地上下滑动，试图从海量零散推文中交叉比对、三角定位最新的技术趋势。 在 Aaron 看来，Twitter 是当之无愧的“全球 AI 市政广场”。他经常建议大二、大三的年轻大学生：“赶紧注册 Twitter 并关注这 20 到 100 个核心账号，单凭你的信息流，你就能比同龄人领先或者落后一整年。”很多年轻人依然习惯被动等待别人通过邮件转发行业文章，这种滞后的信息获取方式在日新月异的 AI 时代无异于闭目塞听。 谈及当下热门的个人助理类 AI 产品时，Aaron 分享了自身作为高阶重度用户的体会与技术瓶颈： - **浏览器交互（Browser Use）的局限性**：虽然个人助理赛道潜力巨大，但当前基于浏览器模拟点击与浏览的技术方案仍暴露出明显瓶颈，基础设施亟待重构。“或许整个互联网都需要为自己的产品提供一套命令行接口（CLI）”，否则缺乏结构化接口的交互将极易受阻。 - **个人 AI 使用画像**：作为一个深度“AI 上瘾”（AI-pilled）的知识工作者，Aaron 每天会使用 4 到 5 个不同的 AI 系统，提出 20 到 30 个深度问题，广泛覆盖行业调研、技术人才搜寻、竞品动态追踪以及新市场拓展分析。 --- ### `58:15` 如何让企业对 AI 变得“可读”：Box 的数据治理与内部落地实践 要让一家企业全面拥抱 AI，首先必须解决“如何让业务对 AI 变得可读（Legible for AI）”的问题。Box 内部推行 AI 的最大优势源于其历史积淀的数据纪律：公司成立以来，严格禁止员工使用碎片化工具存放资料，所有未结构化沉淀——无论是会议纪要、项目计划、产品路线图、商业演示文稿还是财务规划会谈记录——全部 100% 集中归档在 Box 系统中，再配合结构化的 Salesforce CRM 系统，形成了极高标准的数据卫生（Data hygiene），从而使 AI Agent 能够在大规模数据上直接作业。 在推动组织层面的 AI 采纳上，Box 总结出了一套务实的运营框架： - **设立卓越中心与前线工程师**：由极度推崇 AI 的首席信息官（CIO）牵头组建 AI 卓越中心（Center of Excellence），并招聘专门的内部 AI 前线部署工程师（AI FDEs, Forward Deployed Engineers），定点攻坚高杠杆、高影响力的业务工作流。 - **反“Token 最大化”的理性审计**：Box 内部虽然会统计各部门的 Token 消耗列表，但绝不盲目鼓励“Token Maxing”（单纯以刷高 Token 用量为荣）。这一数据看板主要用于合规审查与模式学习——排查是否存在高价值用例可以复用到其他职能，或是甄别低效滥用。Aaron 直言，在 Box 全公司 Token 用量排名前三的超级用户中，往往就有一位在白白浪费其中一半的 Token。 ### `01:00:00` 内部 AI 工作流演进：从 IDE 插件到 Slack Agent 协同 在讨论工程团队如何适应 AI 工具时，Aaron Levie 分享了 Box 内部推动 AI 文化渗透的亲身经历。当管理层发现团队中有两位工程师的工作效率极高时，立刻意识到必须将他们的方法论推广给全体员工。两周前，Aaron 直接要求相关团队把所有人召集到一个房间里，让跑在最前面的工程师向全员演示自己是如何使用 AI 编码的（Aaron 特别点名表扬了工程师 Mick）。令人惊叹的是，从提出想法到全员聚集、完成深度实操演示，仅仅用了六个小时。在 Box 看来，当下最重要的就是让每个人切身看到这种全新的工作范式究竟长什么样。 尽管 Box 在技术栈的某些模块上实现了两到三倍的交付速度——Aaron 强调他不在乎生成了多少行代码，只在乎是否交付了客户切实需要的更多功能——但只要和 Anthropic 等前沿实验室的朋友一交流，就会深感自身仍有很长的路要走。整个行业的工作范式（Meta）变化极其迅猛：两年前人们还在 IDE 里装各种单点插件，认为公司还没准备好成为“AI 原生”组织；随后大家全员铺开并推广 Cursor；而如今如果走在办公室里，看到有人竟然还在单纯写代码，而不是直接在 Slack 里 @Bot 来调度智能体处理工作流，大家都会投去疑惑的目光。 针对主持人关于 Box 是否有类似“Slack 同事智能体（Slack coworker agent）”的提问，Aaron 表示团队内部确实有若干类似形态的工具，他对 Claude 标签等交互形态抱有很高期待。尽管要在团队协作结构与数据治理上做大量精细工作，Box 目前或许还没有达到 Marc Benioff（Salesforce CEO）所推崇的极致程度，也没有像 Anthropic 或 OpenAI 那样完全重构底层协作，但整个公司正在坚定不移地朝这个方向全力推进。 ```mermaid flowchart TD accTitle: AI-Driven Workflow Evolution accDescr: 内部 AI 工具与工作流的三代进化范式，从本地插件演进到企业级通讯协作智能体。 subgraph Phase1["第一代：单点辅助"] A["IDE 本地插件<br/>(基础代码补全)"] end subgraph Phase2["第二代：全员工具化"] B["推广 Cursor 等专业工具<br/>(提升整体编码效率)"] end subgraph Phase3["第三代：智能体协同"] C["Slack 智能体协作<br/>(@Bot 驱动自动化流水线)"] end A --> B B --> C ``` --- ### `01:02:40` 时代鸿沟：两名初创工程师对决传统四十人研发团队 谈到新一代 AI 时代的创业者，Aaron Levie 坦言自己的情绪是“既嫉妒又不嫉妒”。令他羡慕不已的是年轻创始人当下所拥有的巨大杠杆。他提到一周半前看到的一个早期团队演示：整个产品极其复杂成熟，点击每个标签页（Tab）背后都有真实运行、功能完备的功能支撑，完全不是概念演示或花架子。按照五年前乃至 Box 创立之初的工程标准，这绝对需要一个至少 40 人的研发团队才能完成，而现在仅仅两个人就做到了。新创始人从成立第一天起，就能将“AI 原生的高杠杆”作为最底层的设计原则。 相对而言，Box 作为成熟企业必须靠组织韧性强行突破转型，同时承受一定的“生产力折价”。因为 Box 守护着跨国巨头的数据安全与合规底线，绝对不可能随波逐流地废除代码审查（Code Review）等核心安全环节，否则客户不可能将核心数据托付给他们。然而，Aaron 并不羡慕新团队面临的极其残酷的竞争环境：现在任何一个好点子冒出来，往往瞬间就会涌现出五个实力强劲的竞品，每隔三天就会看到 Sequoia 或 Benchmark 又重注投资了一家同赛道公司；而 Box 当年起步时，曾享有两到三年相对安静的窗口期，能够专注于打磨产品体验，而不必每天陷入疯狂的赛跑焦虑之中。 --- ### `01:04:13` 终局认知：应用层的万亿价值归属于“企业级渗透力” 在对话的尾声，Aaron Levie 给出了他对 AI 商业竞争终局的核心判断：在一个 AI 让软件构建速度呈指数级提升的世界里，单纯写出软件的技术门槛被极度拉低，竞争的决胜重心必然不可逆转地转移到“分发与客户交付能力”上，即谁能真正把技术落地并交付给终端客户。 Aaron 赞同诸如 Scott 和 Matan 等富有远见的创始人的观点——AI 的本质是一场“企业级渗透战（Enterprise Diffusion Play）”。在应用层（Applied Layer），正有数万亿甚至数十万亿美元的市场空间虚位以待，谁能理解企业需求、建立起高效的落地团队并攻克企业级复杂场景，谁就注定是最终的赢家。这不仅是 Box 正在践行的路径，也是所有企图在 AI 时代赢得一席之地的创业者不可违背的铁律。 - **杠杆对比**：过去需 40 人团队开发的企业级产品框架，如今 2 名工程师依托 AI 工具即可完成原型与全栈落地。 - **合规折价**：老牌企业必须遵守严格的代码审计与安全合规流程，在研发速度上不得不承受部分效率折价以确保客户数据安全。 - **竞争密度**：早期初创公司的隐蔽期消失殆尽，好赛道瞬间面临数家由一线风投（如 Sequoia、Benchmark）支持的竞争对手。 - **决胜要素**：大模型时代的基础代码构建不再是壁垒，企业的渠道触达、工作流深度融合与组织落地能力（Enterprise Diffusion）决定了数万亿美元应用层价值的最终归属。 ## Notable quotes > `00:00` 立刻去关注 Twitter（X）上的这 20 个账号，赶紧注册，因为这会直接决定你的职业生涯。 — Aaron Levie > `07:15` 绝不能让狐狸来看守鸡舍（don't want the fox to be guarding the hen house）。 — Aaron Levie > `15:10` 你的业务蕴藏在内容之中，用 AI 释放它（Your business lives in content, unleash it with AI）。 — Aaron Levie > `17:45` 我最近收到的董事会汇报材料（Board decks）里，有些居然完全是用 AI 生成的，这简直快要把我逼疯了。 — Aaron Levie > `20:00` 这再好不过了（it's not only acceptable, it's preferable）。 — Aaron Levie > `22:15` 我最不想干的职业就是大学教授，我肯定会立刻辞职，因为你根本不可能搞清楚学生到底自己做了什么。 — Aaron Levie > `32:30` AI 研究员往往会根据自己的工作方式去设想世界——‘我是个研究员，我能访问实验室里的一切数据，所以如果能训练一个专属于我个人世界的模型，那简直太棒了。’然后你对他说：‘请允许我为你介绍一位企业律师。’ — Aaron Levie > `38:15` 当你在电视上听到马修·麦康纳用那种迷人的嗓音亲口把这些词念出来时……那才是真正的‘顿悟时刻’（Aha moment）。天呐，他可太会卖软件了！他的声音用来推销‘记录系统’和‘智能体’简直是绝配。 — Aaron Levie > `45:00` 五年后，企业中 90% 的 Token 消耗都将产生于用户从未主动发起的后台任务；员工最终看到的只是一个处理完毕的任务结果，只需要直接进行审查即可。 — Aaron Levie > `52:12` 在开发 Coding Agent 时甚至不需要注册流程，直接提示‘交出你的 GitHub 授权’（Give us your GitHub）即可。 — Aaron Levie > `55:54` 95% 的大脑带宽都耗在 Twitter 上。 — Aaron Levie ## Mentioned Aaron Levie, Box, Sequoia Capital, Doug Leone, OpenAI, Anthropic, Claude, ChatGPT, Meta, SpaceX, NVIDIA, AWS, GCP, Azure, Snowflake, Databricks, Box Labs, Box Agent, Box Notes, Stan Druckenmiller, Pangram, Google Gemini, Grok, GLM, Decagon, Jesse, N-Gram, Andrej Karpathy, Eli Lilly, Trajectory, Applied Compute, Prime Intellect, Salesforce, Agentforce, Matthew McConaughey, LinkedIn, MCP (Model Context Protocol), Harvey, Legora, Cognition, Factory, Caterpillar, GitHub, Dwarkesh Patel, 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://pscrb.fm/rss/p/traffic.megaphone.fm/CPUAI7737498941.mp3)

[0:00] Aaron Levie: 立刻去关注 Twitter（X）上的这 20 个账号，赶紧注册，因为这会直接决定你的职业生涯。 [7:15] Aaron Levie: 绝不能让狐狸来看守鸡舍（don't want the fox to be guarding the hen house）。 [15:10] Aaron Levie: 你的业务蕴藏在内容之中，用 AI 释放它（Your business lives in content, unleash it with AI）。 [17:45] Aaron Levie: 我最近收到的董事会汇报材料（Board decks）里，有些居然完全是用 AI 生成的，这简直快要把我逼疯了。 [20:00] Aaron Levie: 这再好不过了（it's not only acceptable, it's preferable）。 [22:15] Aaron Levie: 我最不想干的职业就是大学教授，我肯定会立刻辞职，因为你根本不可能搞清楚学生到底自己做了什么。 [32:30] Aaron Levie: AI 研究员往往会根据自己的工作方式去设想世界——‘我是个研究员，我能访问实验室里的一切数据，所以如果能训练一个专属于我个人世界的模型，那简直太棒了。’然后你对他说：‘请允许我为你介绍一位企业律师。’ [38:15] Aaron Levie: 当你在电视上听到马修·麦康纳用那种迷人的嗓音亲口把这些词念出来时……那才是真正的‘顿悟时刻’（Aha moment）。天呐，他可太会卖软件了！他的声音用来推销‘记录系统’和‘智能体’简直是绝配。 [45:00] Aaron Levie: 五年后，企业中 90% 的 Token 消耗都将产生于用户从未主动发起的后台任务；员工最终看到的只是一个处理完毕的任务结果，只需要直接进行审查即可。 [52:12] Aaron Levie: 在开发 Coding Agent 时甚至不需要注册流程，直接提示‘交出你的 GitHub 授权’（Give us your GitHub）即可。 [55:54] Aaron Levie: 95% 的大脑带宽都耗在 Twitter 上。

[17:15] Aaron Levie: The dream state of most of these enterprise workflows is what if we can onboard a client faster? What if we could discover kind of critical data inside of our research much more quickly? What if we can alert to a security event much more quickly? So to do that, you need these sort of background agents or workflows that are sort of pre-established for those processes. [18:50] Aaron Levie: I get board decks that are entirely written by AI these days and it kills me. [20:00] Aaron Levie: Hmm, should I trust you? [20:00] Aaron Levie: How many 'it's not X, it's Y's were in there? [20:00] Aaron Levie: I would hate to be a college professor. I would just totally quit. [40:00] Aaron Levie: nah, that's gonna break your governance policy [43:24] Aaron Levie: No one's had a good experience at a toll booth. [52:30] Aaron Levie: give us your GitHub [60:00] Aaron Levie: Can you just get everybody in a room... and just like, show them how this one person is using AI? Aaron Levie: I was like, absolutely, this would have been a 40-person project five years ago... And it was two people. And you're just like, how do you have so many tabs that work? And they all seem to have stuff behind the tabs that all seem very functional. Like this is not fake. Aaron Levie: Our customers can't possibly entrust us with their data security and compliance if we don't take that seriously. [64:12] Aaron Levie: Whoever can actually get it to the customer is in the best position. [64:12] Aaron Levie: If you don't have that mandate right now, you're just going to lose. It's just like, it's game over. Sorry. [64:12] Aaron Levie: quite literally trillions up for grabs at the applied layer.

[0:00] I want to send an emergency alert to like everybody who's like a sophomore or junior in college and just be like, follow these 20 accounts on Twitter and also join Twitter because like this will just help your career. You're either like a year ahead or a year behind simply based on your feed. And my feed is like so, so wired in. I'll still talk to 20 year olds that are like, yeah, like, you know, I see some articles and I'm like, what do you mean? How do you see articles? Like, I don't even know what that means. Like, do you just like get lucky that somebody emailed you an article? You just have to be wired in. I enjoy it. It's a lot of fun. I play with everything. Today, I'm excited to welcome Aaron Levy, founder and CEO of Vox. Vox is building a collaboration platform and content management platform for the enterprise that has now taken on really new life with AI. And I'm excited to chat with you today about Box and AI, about your general thoughts on AI because you are such a thought leader and about how founders can reinvent themselves for this AI wave So thanks for joining us Thanks for having me I a big fan of the podcast I religiously watch every episode And so great work I am wondering though I know this is a different podcast but when do you talk about root canals Oh, my gosh. We're going to call in Doug Leone. Have we figured out how to weave that in yet or not? I mean, I'm happy to talk about different kind of pain and suffering that I've been through. We should just do this pod with a live root canal. Let's do it. Let's see how it goes. Actually, it would be like hot ones, but you get a root canal and you have to talk about your strategy while you're in a dentist chair. And Doug is just on the other end of it. Amazing. Amazing. Doug is so pleased with himself, by the way, right now. Okay. Has he seen every tweet? Oh, yeah. Okay. Oh, yeah. He's very pleased with himself. Okay, good. Okay. Let's start with, I'm curious your take on this, application companies are the hottest Neolabs. Yes. Agree or disagree? You know, two years ago, I think it would have not made that much sense as like, what does that mean? But very clearly, I think this is what's playing out in the market. And it's all working out mostly because of open source. But what's pretty amazing is right now, you know, the whole concept of being a sort of LLM wrapper, model wrapper is actually working out. Because what I think people underappreciated was that in the real world in the enterprise what you need is some bridge from the model capability to the actual workflow that the enterprise has And that bridge, you basically like probably a trillion dollars has been bet on on basically one of two outcomes. Either that bridge is very kind of limited or that bridge is actually very vast and needs to be able to take on a lot of depth within organizations. and that basically looks like, are you only long, you know, kind of the model itself and superintelligence or are you long this sort of application tier or, you know, maybe previously would have been just pure Neolab. But I think it's very clearly playing out that actually there's a lot of gap between the model and the workflow. And as you bridge that gap over time, you get to a point where you realize, oh, I should also do the model. And then you have enough data, you have enough sort of domain expertise where that becomes its own, you know, sort of flywheel. So I'm very bullish on this idea. And what's cool is it's opening up like multiple layers of, you know, sort of opportunity for startups because you could either be the actual applied company itself, i.e. the Neolab, or you could be the infrastructure provider to the Neolab. And you have like multiple layers of you know going and attacking that space But I think a huge update for the market and everybody kind of views of it Box Labs let go We already have it It's a little bit secret, but we pay very close attention. Right now, the main focus is let's make the agent really, really good on any model. But over time, obviously, you would peel off certain use cases either on a per customer basis or kind of across the whole data set. Yeah. I mean, it seems like there's two forces that are happening. One is people don't want the fox to be guarding the hen house. They don't want the seller of the token to be the one that is also metering and gating, you know, what is the best token for each use case. And then the second is, you know, there's actually a lot of work to do on that bridge to cross and there's real research involved in it. And it's pretty bespoke to the exact, the exact workflow and the exact end customer that you have. Yeah, I think, I think what I tend to see happen in the Valley is, and for very good reason, and it's actually why these companies have been so successful is like, everybody is so kind of research-pilled, which is, again, totally awesome. Big fan. It leads to all these breakthroughs. But the sort of sense that, OK, the model and the intelligence in the model is kind of the only form factor that matters. And then you go to the real world and you sort of see how intelligence actually gets rolled out in people's workflows. And the model could be the most intelligent. [5:00] you know, super intelligence in the world, but that workflow still requires you to connect up to other data systems, still requires these moments where there's a human in the loop interaction. There's delays in the workflow. And so the sort of agent has to sit idle. There's change management of the actual business process. There's legacy systems that haven't been modernized. So you kind of go through these five or 10 things that are way sort of, you know, much more operational, much more blocking and tackling than just the pure super intelligence of the model. And the last thing that I think a classic sort of research organization wants to go do is go attack every single one of those things. And this is sort of not a, there's like, this is not a sort of a one-off in history. Like we've always had this relationship between infrastructure and application. Like, you know, obviously AWS or GCP or Azure have created, you know, trillions of dollars in value of, of kind of market cap of infrastructure. But guess what? There's also trillions of dollars of value in software that only exists because of that infrastructure. And If you were to go back, you know, 10 years ago and you were to look at what was happening in the data space and you looked at what, what, you know, GCP early kind of versions of GCP was building or, or AWS was building. I guarantee you would not have predicted Snowflake or Databricks existing. You would have been like the infrastructure just already does that. Why would you pay another billion of revenue to all of these other products that are just making it so you can work with your data The same thing is going to be true for intelligence which is the models will be insanely valuable But the application of bringing those models into real workflows in banking and life sciences and healthcare and government, that's just going to be a lot of software. Now, the sort of challenge for the next, let's say, two to five years is how much do the model providers need to move up that stack and also try and go and compete at that layer? or do they either leave open intentionally or accidentally that entire space for the application sort of ecosystem? And actually, to some extent, this is like a big question strategically for them because on one hand, you want to be closer to the customer. On the other hand, you also want to be able to have an ecosystem so people trust you. So that's going to be like a really interesting tension over the coming years. How do you think they'll play out? I feel like that's the question we spend every day wrestling with. Yes. You can, yeah, you know, I don't want to speak for Sequoia, But you can sense some of the existential dread of a VC right now is like, should I just put another billion dollars into Anthropic? Or, you know, should I attempt to sort of see what what the applied layer is going to is going to play out? So nobody's, you know, that envious of of your guys position having to figure that out. But but obviously it's even harder for the entrepreneur. But you know I am you know I pretty long obviously the application layer Like I also equally very biased Like I have a very concentrated bet with very limited diversification on it working out that you still want to buy technology that sort of understands your workflow and can get to the core enterprise data But I don't know that there's – I just don't see a different event happening than all of history. I mean, you know, we only have like 50 or 60 years of computer software history. But basically, when you go to that law firm and you go to that, you know, pharma company and you go to that bank, they need something that bridges the core technology to their workflow and their business process. And AI has not meaningfully sort of changed the need or the shape of what that looks like. And the best sort of manifestation of this is sort of this chatbot versus kind of agentic workflow, you know, demarcation. So the chatbot can be totally universal and can be totally horizontal. But all of a sudden you kind of look at that and you say, well, my workflow kind of needs something to kind of ping me at the right time in the process, or it needs access to a certain kind of data that just the chatbot can't natively get access to. So somebody has to go into that organization and get it set up. And somebody has to go and provide domain expertise to this model so it really understands our particular business process And so then unless you really just underwrite the big labs at honestly I not exaggerating like 100,000 employees, if you don't underwrite that, then the diffusion kind of economy is going to be massive because every single one of those companies, whether it's a 50-person firm or certainly a multi-hundred-thousand-person firm, is going to need an army of people to go in and help them with that transformation, that change management. So this is, you know, like whether it's the FDE phenomenon or just, again, understanding that domain expertise, that's going to be a very big deal. And then you kind of alluded to this point, Fox kind of guarding the head and house. And that might actually be singularly the biggest sort of reason this has to happen, which is even under all sort of like complete benevolence and like nobody's actually doing something in any kind of like sneaky way, It just stands to reason that if I'm going to sort of give a task to an agentic system, I just want that task to be cost optimized, like with accuracy as kind of holding constant. And so who can do that? It's the company that doesn't care among 10 different models which model is performing that task. It's like definitionally, you would want that to be the company that does not have a sort of a preference. And then the force you have going against that. [10:00] is that the model companies can subsidize their models or offer them at different rates. I don't know how long that lasts, though, because when these companies become public, I think they will be held to basically the same laws of capitalism that everybody else's. So the subsidization is working very well up to a certain threshold of spend. And we might have exceeded that spend when you're at like tens of billions of dollars of kind of capital. And in fact, if anything, it might even be worse because eventually you have to go in and sort of pay for your training runs as well. So like they, like the subsidization of tokens is a, I think it just has to be a temporary phenomenon. You mean from a gross margin perspective or from like an antitrust perspective? And entirely gross margin. Like if I- But they have such high gross margins on inference right now. But then what are they subsidizing? Like then it's actually, then they're charging actually like, you know, decent rates. Yeah, but they can afford to price the API higher than their own first party products, right? So that part, totally fair. you know, there's an interesting kind of, you know, dimension, which is well, like if API is the high margin thing that is sort of paying for the subsidization, but you've moved all your customers over to the applied product and there's no API revenue. So like, there is like an equilibrium you have to strike with this And and so then all the while if you have some sort of either non actors or people just with a totally different game theory in this you know Meta being one maybe SpaceX being one, China certainly being a giant one, even NVIDIA being one, like that changes the calculus as well, which is those four kind of cohorts don't necessarily need to make money on inference in sort of at least the same margin structure at like what an anthropic or open AI needs. So they would might be fine to bring down inference to 10% margin because they just want to basically pay for massive compute clusters. And then as long as that happens, and as long as like, there's not insane proprietary sort of, you know, you know, sort of closely held secrets, then no matter what, you're going to have kind of cost per token go down on a, on a like for like basis, all of which means more value accrues to the application layer, which is, I think like a very long winded way of saying, I think there's just value in kind of everybody in the stack. I just don't know that, that the, I, the only thing I probably wouldn't bet on is just, okay, one or two labs get 95% of the value creation. I think there's just going to be a much more dynamic environment. And honestly, if I were one of the, you know, two or three biggest labs, I think I'd prefer this outcome too because back to your antitrust point like at some point you just be nationalized if you the only thing that that is is sort of exists as intelligence So you kind of want a little bit of healthy competition in this ecosystem anyway Yep totally Let's transition talking about Box. I want to come back to talking about tokens and Jevon's Paradox and China and all this stuff. But let's talk about Box for a second. I love that topic too. Give Box, I mean, I'm guessing a lot of people that listen to this podcast use Box, but give people a brief kind of explanation of the history of Box and how you're reinventing yourself with AI. We started the company as a way to be able to kind of securely store and share data in the cloud. And it was a very simple idea, but we kind of just kind of cracked a nut or struck a nerve. And, you know, for Doug out there, we were able to sort of scale up quickly. We pivoted rapidly in the enterprise. And the idea was enterprises would be moving from on-premises systems to the cloud, and they would need a better way to be able to sort of store, share, collaborate, manage all this unstructured data, their corporate documents, their financial documents, their marketing assets, their research materials in the cloud securely. So that was the kind of company And we had been sort of flirting with AI kind of products and sort of experiences really since like 2015 If you remember like the first kind of rise of the you know maybe the at least in modern times the AI winter that happened in like the 2015 to 2018 period which is like we think it going to happen now and it didn Um, and, uh, uh, that was a period where we were like, okay, these very, you know, sort of early AI models were showing us signs that, okay, if you, if you, you know, looked at an image and you could classify the image, that's pretty useful if you're in an enterprise, because now maybe you'd take all of your image data and sort of label it, or maybe you'd OCR something and you'd be able to kind of pull out, you know, kind of, you know, sort of the, the, the text in there. That's enormously helpful. The problem was, is insanely expensive. And you had to have a model for every single use case required, you know, kind of a hyper trained model for each workflow that you wanted to do. So we kind of shelved it. You know, a few years later, you know, started paying attention to the GPTs. We had some hackathons where people were like, oh, we could do like, you know, type ahead in one of our kind of, you know, note-taking products. And that was sort of early kind of versions. We did some early work in sort of text detection and classification, which helped with security use cases. Then obviously, chat GPT moment sort of hits. And that was the big sort of, you know, head exploding moment. And if for no other reason why, you know, then then they kind of figured out a form factor that opened up everybody's mind to, oh, these could be these interactive. [15:00] systems that you just like ask a question, get an answer back of sort of, you know, kind of increasing complexity and length. So we looked at that very quickly. We jumped all in. We sort of did the whole company pivot, like everything was kind of exactly like academically what you should do. We had a team carved out. We put, you know, the best people on the team. We, you know, met every day, looked at the updates, and then slowly but surely sort of built out what today is kind of our AI stack and then basically the box agent. And for us, you can imagine the use case is very straightforward. We sit on hundreds of billions of files. Every single one of those files contains critical information for an enterprise. That could be their contracts, their research files, their marketing assets, their loan documents, like all of this critical information. The problem is they rarely know what's actually inside of it. So unless you literally look at the document and kind of search and find it, you just don't know what's inside of it. So now agents can go and basically be farmed out to go and answer questions about that data. They can pre-process it and extract metadata from those documents and turn it into structured data. You can use agents to automate sort of steps and workflows So we built a platform that basically lets you deploy agents against all of that unstructured data And that been the kind of core focus Are you using agents to create new content We are There a couple of modalities where that shows up One is we have again, an online sort of collaborative product that an agent can just like generate any amount of content in it. And then we've done most of the, I think probably more exciting work with OpenAI and Anthropic on just how do you do like advanced document creation, PowerPoint creation. And we've decided that their tech is, you know, at this point, going to always be frontier. So we have an agent that goes and interacts with those systems to produce, you know, a high quality PowerPoint, et cetera. Awesome. Your tagline, your business lives in content, unleash it with AI. Yes. What are the hero home run use cases for how people are unleashing it today? Yeah. And then if you had to fast forward a few years, what do you think people will be doing with AI and your products in a few years? Yeah. Yeah, so probably the easiest hero for, again, more of a traditional enterprise to think about is just as simple as you have a million contracts. Why don't you find out what's inside them? Or you have a million research documents, be able to go and pull out all of the critical structure data, put that into a database, and then be able to query, analyze, automate workflows around that. so that that kind of the thing that that just knocks it out of the park every single time because it been a long problem that people have never been able to go in and sort of apply human sort of labor to because it just too expensive to read every every contract every research document You know, maybe you could do it if you had like a loan document process, but most other data just never gets read at that at that scale. And then I think the stuff that we're probably as much if not more excited by is really the equivalent of what we see with, let's say, coding agents or other complex agents, which is you have long-running agents that are just executing your entire kind of workflow or process. And this would be in the form of you go to a bank and you're onboarding at a bank and they've basically like automated every step that is possible to automate and then sort of jumps out to a person in the steps in the process for extra review or extra verification. But now instead of that sort of one or two week back and forth, it just happens in like an hour. Like that's the dream state of most of these enterprise workflows is what if we can onboard a client faster? What if we could discover kind of critical data inside of our research much more quickly? What if we can alert to a security event much more quickly? So to do that, you need these sort of background agents or workflows that are sort of pre-established for those processes. Totally. It the year of the long agent It is It is Yeah yeah I curious like you made the analogy to cloud code It seems to me that in the coding domain using AI is like not only accepted it embraced Yes. In the content domain, which is I think where a lot of the content in box sits, using AI to produce content at least, it's just like there's almost this allergic reaction to it. Like all the pangram stuff on Twitter, there's the, you know, it's like this concept of work slop. I'm curious what you think about work slop and like, will this still be a thing in a few years? I'm going to sort of separate the box corporate hat and just now kind of maybe riff as a consumer of, gosh, I wish there was a better term, but work slop as, you know, inside of an enterprise context. I get board decks that are entirely written by AI these days and it kills me. So here's the difference, I think, on the acceptability. So there's probably like more symbolism to this actually topic than just like the slop element, but like actually like diffusion of AI in general sort of almost ties to this. Code, like other than, you know, the top engineers that we hang out with that like are like they have, you know, deep taste in the code, like, you know, and the judgment is incredible and like it is them as much an art as it is. [20:00] science. So take that group aside. For most of the world, code is a utility. It is just trying to accomplish something. You're just trying to automate something. You're just trying to put a sort of interface up there that somebody presses a button and moves to the next step. So for most of the world, the value creation of code has been to automate things and to be able to have it as a utility. So at the end of the day, and we'll probably still use the term slot for a while because there's taste in kind of front-end design and there's taste in sort of systems and you don't want to have vulnerabilities in your code. So that's going to exist for a while. But at the end of the day, if you can tell an agent, like, please go and generate my entire backend system or my front-end system, like, it's just like, it's not only acceptable, it's preferable because it's just like, that was the thing that was blocking us from moving forward. So we need to go do that. At least the way society functions and the way the world works and our brains work at the moment, maybe this changes. You know, when you get a presentation from somebody, there's still this association, which is like, I'm trying to decide if I can trust that person to go and execute on that thing or deliver that result or understand that topic. And so when you see work slop you like I losing my ability to sort of know for a fact that like how much of the thought process was them versus how much was the AI how much did I even care about that because I myself am doing the same thing So like we have this weird like – it's this very weird sort of like collective issue that we have, which is like I'm doing work slop for some of my brainstorms and decisions. But when I get it from somebody else, I'm like, hmm, should I trust you? Uh, and, uh, and I don't know, I, you know, I mean, it just might be a thing that as a society, we have to kind of keep cranking through over the next kind of three to five years and, and end up at the other, at the other side. Like, like, you know, I hate to use like these like totally busted analogies, but you know, obviously you don't care when you see somebody's financial model, you're like, yeah, that was generated clearly by like a macro, uh, or, or, you know, that was like not, you did not personally go and compute all of that, but, but you're showing it to me and we're talking about it. So why can't the same exist for a strategy deck or whatnot? But I think right now we're going through this evolution of like, what is the person's role? What is the content a proxy for? Is it supposed to be a proxy for how much that person knows? Is it a proxy for what we think they can go and execute on I think we just in this very messy period where we have to kind of figure that out Totally Did you read the Stan Druckenmiller Wall Street Journal article I read the discussion about it Yeah, the reaction to it. But actually, I didn't read it. But was it like very sloppy? I didn't think it was sloppy. I loved it. And so to me, it was just a nice counter example of I have this like allergic reaction. How many it's not X, it's Y's were in there? I don't think there were any. Okay, okay, okay. But it does show up as 100% AI in Pangram. Okay. And dashes? uh i think they're okay okay you can't do that yeah but it's like it was a nice counter example to me because normally i read something that's clearly written by ai and i just have this allergic reaction whereas with the stan piece i didn't yes i'm not sure how much of that was just you know it's stan therefore i trust in stan versus yeah no but it's it is psychologically kind of like weird because i'll read these like you know um x articles and i'm like now like doing 2x the amount of work to read these yeah i'm reading it one for the substance i'm also reading two for the calculation of like, did the person write it? Or am I just literally reading like a cloud prompt? And then like, I'm like, I'm literally like my mental processing is now like, should I now, does that up weight or low or like lower my, my sort of judgment of the person or the post? And I think we yeah we in for some weird times because of this I would hate to be a college professor I would just I would totally quit because you just like are you like I don I don know anymore what you did I like what what going on It does seem like the calculator is the closest analogy though Yeah. Except it's just like, that was like more like finite in terms, and you still had to piece together so many more things. Like we, like the, and some of these analogies are breaking down of like, it's just a task because it's like, well, at some point, like this thing is doing like at least 10 tasks at once. Um, but, uh, but yeah, let's talk about harnesses. How does the, uh, How does a box agent work? Great transition to harnesses. Okay. Speaking of calculators, let's talk about harnesses. Okay. This is back to the kind of Neolab kind of applied layer. There's a bunch of things that we know about our file system, permission structures, our search engine that, you know, certainly, and by all means, we actually would love all of the labs to train on our understanding of this because it would only make external agents as good as possible. Yeah, like we always talk to labs like, hey, we'll give you as much data as you want about kind of how the system works. But let's just say like that aside, we have a lot of depth of understanding of what do people do in Box? How do they search Box? How do they decide when they look through 10 files, which is the one to go pick? What is their internal kind of calculus or heuristic on? [25:00] on sort of figuring out the most relevant document to look at. So we know all that. And we basically, you know, have built an agentic harness that, you know, attempts to kind of understand that set of domain understanding about our system. It obviously has access to our search system, our file system. It has a bunch of kind of mechanisms for just pulling out just the text of a document, just pulling out chunks from the document, doing embeddings on the document on a fly. So there's a set of kind of tools that it can use. And effectively, it's a harness for asking questions of a large data set. So in my Box account, I have, I don't even know the latest number, but on the order of tens of millions of files, just because it's everything that has ever kind of accumulated over 20 years. But I can now ask any question of all that data set using the Box agent, and it goes around. It does multiple searches in one. It re-ranks it. It then very quickly sort of pulls out the most relevant information. Then it'll, in some cases, read the full document, does all the steps. And then we compare that against like, well, what if we just gave, you know, Claude our API or gave OpenAI our API and we see like meaningfully better results on accuracy and latency. Because again we kind of know exactly how to tune it for our workflows So that effectively the harness that we built out And then what evals matter the most to you So I have a couple I have a couple like you know just funny personal ones that I just keep track of, of like my own use cases. But we do, we have, I don't know, hundreds of different tests that we do on every single model. We actually have two evals at the moment. One is we put out a thing called the complex work eval, which is a set of domain specific work in life sciences, financial services, public sector, tech, et cetera. And it's kind of exactly what you'd think of as a document-centric eval. So given these five documents and this set of problems, what would your answers be? And we test every single model against those with our agent. And then we have a holdback eval, which is actually the first one is holdback also. But the second one is just like then our box instance and how box employees use their data. And then we eval every model again on that. So we're able to kind of roughly keep track of all of the incremental progress. We see when things move by half a point in terms of model improvement. And then we roll out sort of default models based on different kind of cost and accuracy thresholds. And then we let customers also choose any model they want from effectively our model garden. What's your current view of the race and like where all the horses are in terms of model performance on your use case They more or less closely correlate code with one exception which is actually in some of our use cases Gemini is disproportionately better than what you would see from coding And it might be, you know, sort of just better tool use, you know, given kind of, you know, the Gemini ecosystem and what they need to build for. It solves, you know, kind of a strong set of sort of general knowledge work use cases as well. But I think mostly correlating to kind of code. So Fable 5.1 was clearly kind of state of the art and the best model that we've seen. There's obviously rumors about other models. And so we'll see how the kind of race kind of continues on this front. But basically, by and large, like when you look at GDPVal, Mercore has their Apex eval. These things will all generally follow the coding models. And so I think it's we're just neck and neck on like Grog, Muse, the Fable class and, you know, GPT 5.6 slash whatever the building next. Like these are just like it's a total race right now. And your customers typically express a preference on which model they want to use or do they just use your default So they you know kind of by volume they use our default because it just easy and it works extremely well And it tuned for there a few ways our kind of agent manifests So like the way that you'd most commonly experience it as an end user is you would just be searching and kind of, you know, asking questions of your data. But by volume, the volume of tokens tends to go through more of our workflow agents or data extraction. That's where actually you have customers actually doing evals and they're basically saying, okay, I want to really make sure that at this cost profile, I can get 98% accuracy on data extraction. And that's a place where like, we'll have an FDE that goes in and helps you, you know, helps understand your data environment, tests against five different models. And then you're just basically at the mercy of, of the eval. Yeah. What are you saying in terms of the adoption of open late models in your customer base? Um, so probably higher than people think lower than what enterprises actually want. and much, much, much, much, much lower than what it'll be in five years. So like some mix of that would be the kind of message like- And it's primarily cost that's driving that decision? I have to probably attribute 30 plus percent to just kind of the sexiness of like- Yeah, I want to try GLM. Yeah, yeah, I think there's that. Like I've heard CIOs of Fortune 500 companies [30:00] Say we're playing with open source here and I look at that and be like, well, I know for a fact like Gemini or Muse would have been just fine at that particular cost profile that you're trying to do. Or probably even like, you know, 5.6 Luna or Terra or whatever, whichever one had the crazy discounting they just did. Like it probably would have been totally fine. But you want to be able to be like, okay, I'm a little hedged. Like it's cool. Like we're at that phase still. over time. I think it stands to reason that you'll see meaningful different costs because you'll be able to peel off workloads that that just only make sense at at sort of, you know, grinding down to the cost of inference, in which case open open weights will have this sort of economic advantage. Right now, there's, you know, this challenge of like sometimes it's more token inefficient, you know, sometimes like randomly, like I've heard stories like randomly, it'll just like speak Chinese, like mid mid chain. So you're like, OK, well, that'll be weird for a bank. So we need to probably work on some of those things. But long-term, I think it has to be the case that you're gonna peel off those workloads. One of the more interesting posts, I think, on this that I totally subscribe to is Jesse at Decagon. You probably read that post of this paradox of we're gonna have, you're gonna see close just continue go exponential But what gonna happen is each use case that kind of matures you can peel off to open source And once you have kind of stability in that use case it starts to make sense to veer it toward an open weights model, assuming one of two things is true, one, that it's actually literally cheaper, or two, having some post-training gets you X percent more performance. And so I think you will just be in a reality where we will, and this is going to be very confusing probably for like the press more than people in the valley because you'll be like wait a second like the revenue of anthropic opening eye etc are like off the charts but somehow open weights is like also growing exponentially you know like how is this like how is the pie is growing so fast yeah and it's like the pie is growing so fast yeah but what's happening is is actually there's there's an interesting duality it's it's it's not even just like like rising tide lifts all boats it's like no no like we either use fable or or five six for orchestration. And then we farm out all these long tail tasks to a cheaper model. Or the opposite is true. Like you use, you have some kind of orchestration agent that like by default does the cheaper stuff, but occasionally sort of see some, something that is just way too hard And then it pops it out to to one of these heavier models And so you might have blended 50 spend on each but 10 times the amount of tokens you know on the open weights model And so like everybody kind of winning but there an interplay between why they winning I'm curious about how you think about memory and customization or personalization and where that's going to go. Because it seems like today the dominant architecture is kind of like rag-based system still. And like you can get fancy on the rag, but it's still context look up where the weights themselves aren't fundamentally changing. Yeah. It seems like, I mean, if I listen to my friends at the labs, like continual learning, This idea of like the model's weight should adapt as it gets to know you. Yeah. We had Ngram on the podcast. I don't know if you know Dan. I just got introduced. I listen to the podcast. Oh, amazing. I would have loved to have been the fourth person in the room. Amazing. Yeah, I think they're working with customers to help basically bake in some of the context into the weights themselves. Yeah. What direction do you think it's going to go? You know, you're catching me at a time right before I'm actually doing my call with Dan. So I wish I could have talked to him first and then I'll have like a way more eloquent answer. I'm like extremely fascinated by the approach. Like I have no reason for not wanting to work and exist. We live in a world at Box where we see these sort of the high degree of complexity on permissions and access controls and data that tends to be sort of the rub on a lot of these types of approaches And I going to put Engram aside because I sure they already thought this through. So I'm going to talk more generic philosophically. I think sometimes you will talk to a researcher that sort of imagines the world working the way they work, which is like, I'm a researcher. I have access to everything. And so if I had a model that was trained just on my world, this would be amazing. And then you're like, let me introduce you to a lawyer. And the lawyer like has this tiny little, you know, access point of just like the five projects they're working on because somebody right, you know, one door over is working on the competitive project to, to another company in the space that they can't have any sort of overlap with, with what they see or what they know. And there can't be a single document that passes between those two walls and they have to be these kind of hard barriers. So, you know, so like, sure, like you could still you could still now train a model just for that one user. But like what happens if every single day they get added or removed from something that that adds important context to sort of what they need to understand? And and and, you know, again, like I think there's going to be probably breakthroughs in continual learning that sort of all resolve this. But like this is why previously it was just like, you know, there was no way you could pull this off five years ago because it would be insanely [35:00] expensive, impossible to kind of wrap your head around, around how those access controls are supposed to work. But, you know, obviously with like, as the cost curve goes down as open weights, you know, get, you know, cheaper, smaller, faster, better. Um, I think this becomes super interesting. One thing on, on the podcast that I found very fascinating, and I just need like a T chart, honestly, is just like, you know, like what is the decision point of what goes in context and what goes in the weights? Like you, you have to be a little bit thoughtful about like, where is the the massive performance gain that you get by by baking in the weights and there's probably some like incredible like like um calculation of like like when the rate of change of the data is not you know so far but the upside of the of the weights you know you know dramatically change the accuracy of the model like you know you'd have to kind of land on some sort of you know rubric like that i mean if you could wave a magic wand it almost seems like i think Karpathia said this in some prior interviews, like if you could almost remove all the memorized information from the models and just have, have it encapsulate the specific reasoning capabilities, like the ethos of how we do things, for example, at Sequoia. And then you have all the you know actual content and a lookup system that almost feels like the if you could wave a magic wand that what the system would look like And that one super interesting I think the question will be like how much are enterprises different at that level versus it's actually their literal IP that is what makes them different? Like how many different types of styles of execution are there in the world versus, no, it's just like the depth of knowledge about that particular legal case and how do I apply it to this other project I'm working on? That's where so much of the value sits. So so but again, if you can just like wait till my Zoom call with Dan and then I'll really know the answer. But like I'm a fan because no matter what, there's going to be like like I've jumped right into like the individual, you know, but like no matter what, like at a firm level, there's probably ways to take this approach. Like I'm a big sort of fan of what trajectory or applied computer doing or prime intellect because because I think there's like there's there's no question that if you're Eli Lilly, you want a model for how you do drug discovery. and that probably does need to go like farther or deeper or more sort of specific than what you're getting at off the shelf. And there's not a lot of sort of, you know, kind of church and state problems for drug discovery, you know, workflows. They probably want as much of that information available to as many people as possible. So I think it's going to be like domain specific and, you know, you're just going to have different outcomes based on which vertical or you know type of use case and where the firewalls need to be in that process Makes sense Okay So you hinted at the beginning that there a BoxLabs What type of work is BoxLabs doing So it the equivalent of BoxLabs Like, I don't know if we've used capital L yet, but basically, you know, it's our applied sort of AI team. And what research areas are most interesting to your team right now? Yeah. So the biggest areas that the most sort of research kind of, you know, of the continuum of engineers, there's some cluster that is sort of more on the research bent. And of that cluster, the things that we spend time on still, again, is at the kind of applied layer. But it's a lot around how do you take agents and make them another 10 points of accuracy improvement given X problem? So how do you build a map of the problem set with a given set of data to sort of best execute on that task. So we spend a lot of time on that style of work. We have a team, for instance, working on how do you do effectively at the agent level, at the harness level, some form of kind of auto research on sort of hill climbing on accuracy of answering questions or sets of problems on given a kind of a set of client data. So if you're a bank, you have a bunch of loan documents coming in And these are like 100 page documents like whether you getting 70 accuracy with an off the shelf model or like 97 is like basically obviously a world of difference. And can you actually go and automate that process? So somehow you have to hill climb from the base model to the 97%. And there's a lot of work going into the system to basically pull that off. Maybe zooming out, what do you think of as, and this can be a box question or a non-block specific question, the role of systems of record in a world with agents. And I'm sure you saw some of the Twitter discourse on like, you know, every software company is trying to sell me their own agent right now. I don't want another agent from them. I want their system record to work well with my agent. And so like, how do you think about that? Hashtag Claude Force. So- Catchy name, by the way, very catchy. I mean, literally, it's like one of those things where like the first three minutes, you're like, man, that seems funny. And then four minutes later, especially when you see the stock, you're like, ah, brilliant move. Like, this is great. We're doing this. And then when you, I think somebody actually said this the best there. Like when they heard Matthew McConaughey say it out loud, that was like really the, that sealed the deal. That was the aha moment. That was the aha moment. And it's like, man, he can sell software. Like it's actually incredible. Like his voice is so good for selling systems of record and agents. If you're in like our platform, [40:00] sort of contemporary group. of like you built a SaaS platform, And you have some set of data and workflow that, you know, your customers kind of operate in. There's effectively two things you just have to do. And I think anybody attempting to do one over the other is just going to lose. You have to build an agent that is insanely great at your product. Like that agent has to be you have to provably be like, 10 or 20 points better than an off the shelf agent using your system. Not because it's hobbled the other side. It's just like, you are so eval maxed and you're like, so tuned to your particular workflow that, that you can improve your system. You have to have that. And you probably, because you understand your domain, unless you're like totally asleep at the wheel, you probably have use cases that no one has, has sort of thought to build products around because you, you talk to customers every day and you see what they run into. And you're like, Oh, we could just like I've had at least a dozen. I mean, so I probably talked to a couple hundred customers a year in variety capacities. I've had at least two dozen times. where the customer has a use case that is like a breakthrough moment for me of like, shit, that would be actually totally insane. Like, um, unfortunately, unfortunately, Since you put me on the spot, I don't know if my example will pay off the level of excitement I just had because it will be on the spot. No, no. It's just like I think the thing I was thinking of is just like I think it's going to be a womp womp for the podcast. But there was there was basically a customer had this idea of they wanted an agent in the background. Trying to sort of figure out when documents sort of met their governance policies. And like does something need to go into some kind of archive or something into some kind of legal hold or whatnot. Oh, cool. See, exactly. See, that's exactly the voice. That was exactly the voice I was worried about. Yes. Yeah, no, you couldn't even pull it off. Okay. So in our world, this is awesome because you're like, oh. Yeah. Like, no, because think about it. Every company has a head of governance. I might call sincerely. Okay. I know. I believe you. I mean, listen, you do enterprise. So like, I think that it was at least half serious. Imagine you're an enterprise, you have a head of compliance and a head of governance. Okay. They can only be... Like... overseeing the whole sort of enterprise. They've never been able to be everywhere at once. Now imagine if they could sit next to the employee, And basically be able to be like, oh, you're about to go do something that breaks our governance policy. So like the idea was like, oh, what if there was just an ongoing agent that just like automatically was just like, nah, that's gonna break your governance policy instead of the user having to like try and predict or understand the stuff. So anyway, those are the kinds of things where if you have an agent within your product, you're gonna be able to identify sort of sooner and better than the rest of the market and or just like do things that maybe would be impossible to do off platform. On the other hand, It's just like, obviously you have to go headless. You literally have to make sure that your APIs are exposed to Cloud and ChatDBT and all the different platforms. And you have to make sure that you have either a direct way into deterministic APIs. so that those agents can use your APIs, make calls via MCP or whatever, or at least make your agent be headless and be exposed in those systems. And then the only reason maybe this is like, even remotely a hard debate, is you have to just make sure as a system of record that you can find a way where commercially, it sort of makes sense on the other side and is sort of valuable and interesting. And the reason why I think a lot of people got that wrong that weren't in these companies was just underestimating the amount of sort of new use cases that are just total upside. There are like complete white space opportunities for these systems of record. So like in the Salesforce example, I use Salesforce more today, probably by an order of magnitude than I ever have because I MCP into it via Claude or ChatGVT. And so I just am always asking questions about the data inside of our CRM system. And do you think that means the systems of record become more toll booth businesses then to make love that term because like no one's had a good experience at a toll booth. I love toll booths. Yeah, exactly. You love it more than governance agents. So I would say that because they sort of have a – have a depth of purpose in, of organizing the workflow, managing the data, securing the data, providing guardrails, then it's really just, yeah, you need to have some kind of volume oriented business model on that other side. And I just think there's like if you're solving real problems for customers. It'll just like make money. I like like this is so cheesy, but like I've I've told like LinkedIn product managers, I'd probably pay 10x more for LinkedIn if I just could MCP into it. If I just had a way of just like always understanding like like, OK, the CIO is doing this thing and I need to reach out or whatever. I'll like take my money. Yeah. So so there's these systems actually have a tremendous amount of value. based on the data that they have, and customers will absolutely sort of find some way to reward you for that value creation if you're doing a job. - Super interesting. Maybe related. Let's talk about product UI. Yep. [45:00] Big, you know, generic chatbot, chatbox agent. Is that going to be the dominant UI for how people are using AI in the future, especially when it comes to the application layer? This is why I think the applied layer, you know, has so much room to run is because probably the sort of, you know, universal chat system that you ask a question to, you get an answer back or does sort of some work in the background. I think it's going to be obviously like that's going to be a mainstay. That UI will always exist. It'll be incredibly powerful. The horizontal products will have it. The vertical products will have it. Everybody will have it. It's just like your product has a search box. Obviously, it does. So that's always going to be here for these sort of one-off asks of an agent or go find this thing or answer this question or produce something for me on demand. But most of the enterprise is sort of made up of these processes and workflows that are kind of just happening behind the scenes. Sometimes they're happening with computers and computers are running these things. Or sometimes they're happening with other people that are doing these things. Or sometimes they should be happening with people, but you could never afford to have them happen with people. So they just didn't happen. And so that sort of is a slightly different kind of metaphor than a chatbot where you ask a question and it comes back with an answer That like okay I kind of want agents in the background to do things for me read every contract, look at every log, triage every security incident. And then instead of me chatting, maybe I'll chat as a, as a means of doing kind of a catch up. I want a dashboard. I want a workflow. I want a queue. I want a task list. So then the, the, the, the challenge becomes, well, does the horizontal product kind of take on every one of those, those components and, and manifest every one of those experiences in one, in which case I think you'll start to be like, man, that thing is like really like a, that's pretty heavy. And like, and then we'll start to like, be like, oh, this is no longer this simple, easy, delightful thing anymore. So then the vertical players actually like, like they actually sort of understand the process and can manifest all the right buttons and tabs and the names of the things for that particular workflow. So I think it puts, I think as you have agents that are doing more work in the background, doing more async work that is just like I farmed out a bunch of agents to review things as they happen or whatnot, that leans more toward the applied companies that can understand those workflows, that can understand those processes. I think you going to have these in every field Certainly we already know how they going to look in legal with Harvey Lagor et cetera We are seeing them start to emerge in areas like security We seen them start to emerge in the long-running coding agents with Cognition and Factory. So I think that will be one of the bigger applied AI use cases. And ultimately, in five years from now, I would bet 90% of all tokens in the enterprise are things that a user never kicked off, and they just see a result. They just, they see a task show up and they have to go review it. And it's just like, it's just happening. Yeah. Yeah. Makes sense. Okay. Let's talk about AI diffusion. Coding agents. It was like, you know, boom, January 1st, 2026 happened. And like the fastest diffusion of anything into the economy we've ever seen has happened. The diffusion of, you know, the rest of the AI magic into the rest of our jobs seems like it's been a lot slower. Yeah. What are your thoughts on that and where are the areas where you think we're going to see faster diffusion and how's that going to happen? Yeah. So you always have to kind of compare and contrast coding versus everything else to really understand the dissimilarity. So in coding, and this is back to the kind of utility point on you know slop like the utility of code is almost 100 represented by the amount of text that you can generate Like like all obviously an insane amount of value went into the text but but like and then like knowledge and expertise and meetings and everything. But like, ultimately, like the text is the thing that produces the program that is actually the thing that you're trying to do. You know, if you could have the world's greatest programmer, like, and never had to sleep, never had to eat, they could intuit what to build, and they could just sit on a computer all day long, your value creation would be 100% correlated with how many hours they could sit at that computer and type code. Like lines of code, ideally good code, is the most thing that will be correlated to whether you produced software that people wanted. So it's all text. The models are hyper-trained on these. Everybody in AI labs treat coding as a competitive benchmark to constantly try and exceed. They get to do their own evals on it every single day because they are the ones coding the models themselves. And it's the most technical audience of all time where when they deploy an agentic system and they run into either a bug or a problem or like some MCP server comes back with like connection invalid. They fix it themselves. They know how to triage the problem. They don't call IT. They're just like, oh, yeah, no, I didn't open up that port. Sorry, I'll go fix it. So that's like five. [50:00] things. Oh, and maybe like the sixth. Like, it's just like a very, very high paying vertical. that like it's just like automatically valuable if you could get 10% or 20% productivity gain, let alone 5X productivity gain. So take those five or six things, that coding has as sort of beneficial properties to sort of automation, then compare that to, you know, you know, every other form of knowledge work. And you'd probably have like a histogram. And I don't know if anybody's published this, maybe you can like the similarity to coding and like, like what, what are the domains that like, like start to sort of, sort of, you know, look closer, like, like less and less like coding as you, as you kind of scale out. And it's like, okay, well, lo and behold, legal is kind of interesting because, because like, there's a lot of value creation to somebody sitting at a computer, reviewing legal documents, writing legal documents, like processing large amounts of information. Okay. So that's kind of blowing up. And then Now let's take something like a sales rep. Okay, so much farther down the list in terms of sort of likeness. The sales reps value creation is basically convincing an external customer to buy software or technology or a caterpillar truck from them. That is the value creation to the economy of the sales rep. And so let's say we brought the world's best automation to them. Like first of all, again, they'd have to like figure out how to technically wire it up. kind of things. But no matter what, they're still rate limited and constrained by like, did the customer respond to them? Do they want to meet? Can they meet next Tuesday or can they meet today? Like, does the customer have budget? All of these other things. So that's maybe the entire continuum right there is like one is like, I mean, of knowledge work, like obviously this is not even touching the sort of working with with Adam. So that's a continuum, which is on one end, you have somebody rate limited by so many external factors. On the other end, you have somebody who text and that is your ability to basically automate things is that continuum. So for the real world, we have to basically bring intelligence to these workflows in ways that are that sort of somewhat feel like the shape of their workflow, somewhat feel like the shape of their work. Find a way to... deliver the change management, deliver the implementation, get data into a format and into an environment that actually works with these systems. Like Asterix, like one of the other big things is like, if you go to most engineers in 2026, like maybe like minus two months ago, given the, given the latest phenomenon, but like, like the codes in GitHub, You just like connected to GitHub. Like, like, remember, there's this period where like you're when you launched a coding agent, like there was no like sign up or register. It was just like, give us your GitHub. That doesn't exist in knowledge work. There's no like, give us your GitHub for knowledge work. There's like, give us your box. Well, box customers have a much easier time with all of this. Unfortunately, like we're only 1.3 billion in revenue run rate. So like, like, that means there's a lot of people not using box. And so what are they using? Their data is in on-premises systems, legacy file shares, legacy infrastructure. you know, enterprise environments that don't talk to agents particularly well. So just think about that sort of distinction between, you know, implementing coding agents versus everything else. Even something, again, you'll sort of fall asleep about is like access controls in the enterprise are totally different. I actually totally forgot that point about coding. In coding, you get access to basically most of the stuff ever relevant for your job. In knowledge work, you're like, hey, Sally, can you open up that sort of file share for me? sort of project because I didn't get access to it. How do you make sure the agent has access to those set of things? All of that work has to get done. So the thing I think we have to prepare for is two things. One, Silicon Valley has to prepare for diffusion taking a lot longer than they think. or then I'll say we, but I really think they, because I – Like I know how long it'll take. And then the second thing is, is the good news is this is sort of all correlated to the applied layer value creation. Like the, because the companies that will just have the patients, the, the sort of the, the, the full sort of domain expertise, the, the, just the sheer work ethic, because it's not like everybody's just from the floodgates. Like you have to go in and just, you know, pound pavement and get out there. That will be the applied layer. of applied layer AI value is actually, how do you go get the technology to the lawyer or to the sales rep or to the life sciences researcher or to the person that runs the customer support team? That's all sort of opportunity right now that exists. Awesome. I'm going to close by asking some advice for other founders. Maybe let's start with founder advice and company building advice. On the founder side, it seems like you are in every AI cap table, you know, every cool new company, like, you know, you know, Engram, you know, how did you kind of get yourself in the middle of the AI conversation? There's probably two parts. One... I was just very well primed for it. Like working with unstructured data, for 20 years, you just like can instantly see the benefit [55:00] - Agents on that. So like, it honestly like took longer than I would have wanted, that we got to have this conversation, because we tried to have this conversation, you know, eight, nine, 10 years ago. And now it's finally happening. So first of all, just super well primed. Our product... sort of shape and what people do with our product already lends itself extremely well to agents. So like, obviously we had to bet the company on that. And then lo and behold, we had the positive feedback loop of customers actually saying, yeah, That would actually be very powerful if I could go read every document and answer any questions. So that's the first and certainly biggest by kind of a factor of 10. And then the other is just like, I'm extremely fascinated by the technology. And it's just fun. It's like- Where do you learn about it? Your podcast. Dorkesh's podcast. Twitter. I mean, probably unfortunately for brain cells, like it's probably 95% Twitter. through the feed and it's just like i like i look like some sad meme probably of just like i'm just scrolling and scrolling and scrolling and like attempting to like triangulate all the information that's amazing because it is it is the global town square for ai it is and um i want to send an Everybody who's like a sophomore or junior in college, and just be like, follow these 20 accounts on Twitter and also join Twitter because like this will just help your career. You're either like a year ahead or a year behind simply based on your feed. And my feed is like so, so wired in. This is the number one advice I give to people when they're asking how to get current on AI. It's like follow, follow these hundred accounts. It's like, but I mean, virtually first step, join Twitter. Like you'll still talk to 20 year olds that are like, yeah, like, you know, I, I, I see some articles and I'm like, what do you mean? How do you see articles? just like get lucky that somebody emailed you an article, like just join Twitter. Like, what are you talking about? So, so, so I would, you know, you just have to be wired in. I enjoy it. It's a lot of fun. I play with everything. What's your favorite new AI products? Please say instinct. So, okay. Full, full disclaimer. I've not done the instinct invite code yet. Simply because I have a backlog of like three other personal assistant products. You guys want instinct. I know, I know everybody's rate. I'm very excited. I'm on And I'm not an investor. Oh, you're not? Okay, so this is like totally genuine. Okay, so I 100% will have, I don't know when this is going to run, but I'm sure I will have played with it by the time it runs. I'm in pre-release in a couple right now that I'm spending some time with. I think the personal assistance stuff is super exciting. At least in some of my use cases, it's still showing some of the limits of like browser use as an example. Like, we still have some work to do there. Like probably the entire internet needs just like a CLI for their product. So we still have, like there's still some blocking and tackling at the infrastructure be totally awesome. And then I kind of look like your average probably AI-pilled knowledge worker on – every day I'm asking one of five different AI systems 20 to 30 questions on just like doing research, looking for talent, looking for what is a competitor doing, what's happening in this market, how do you expand there, so kind of that. They have max adoption of AI, not just at the individual level, but also at the company level. How do you make your business legible for AI? Yeah. Again, some of this we have as a byproduct of how we've always thought about information systems in the company. So like... like to no exaggeration. If you have a question that you would like to ask about the, the, the business that has ever been documented in a form of unstructured data. So a meeting note, a project plan, a roadmap, a presentation, a financial document, a financial planning session. It's 100% inbox. So like we benefit from a data architecture that already is insanely sort of tuned for because this is how we've run the company. We didn't let anybody use anything else. And so the data is very easy for us to be able to work with at scale. And then, of course, we have Salesforce and all the other kind of core canonical systems. We've been able to kind of have, I think, pretty good data hygiene. So agents running on top of that makes it a little bit easier to be kind of AI first in how we operate. Maybe a couple best practices or things that we've seen. You know, first of all, trying to figure out where the highest leverage impact workflows are gonna be. And, you know, trying to kind of target those. So, you know, our CIO is very AI-pilled. We have a little bit of like a center of excellence on AI. We've hired some internal AI FDEs to kind of help with these processes. Do you have leaderboards? We don't token max. We do actually have like... I mean, I guess like literally we have a list of – people buy a number of tokens, but but it's usually to like go and inspect, like, okay, do we, do we think that's useful or is there a learning there that we should take back to some other function? So like probably our top, you know, in the top three, all AI users that box, like probably one of them is wasting, you know, half the tokens and then [60:00] Two of them are like, oh shit, whatever they're doing, we need to go like do like an internal training session for everybody else. So, you know, we had this thing like two weeks ago where I was like, Can you just get everybody in a room? you know, on this particular team and just like, like show them how this one person is using AI. And then like, you know, six hours later, they were like in a room and the guy was doing like a full demo of what he was doing. Shout out to Mick. And so like, that's the kind of stuff that we're trying to do, which is just like, how do we show everybody what it looks like to work in this way? But you know, For as fast as we're moving and we are shipping at some parts of the stack two or three times more sort of, you know, actual customer facing products. So, like, I don't care about how much code, but, like, did we actually deliver more functionality that customers are asking for? Some parts of the stack we're doing that. Then you'll go talk to a friend in Anthropic and they're just like, and you're like, oh, my God, we still have. We still have ways to go. Like, like I, you know, the particular meta kind of constantly changes, which is like, you know, two years ago, you'd be like, you're just using a plugin in your IDE. And you're like, man, that's, we're not going to be that. Like, we're not ready to fully, you know, be AI first. And then you're like, okay, everybody roll out, roll out cursor. And then you're, everybody rolls out cursor. And then you're like, and then that finally happens. And then you're like, you know, you're walking around and you're like, you're not only working from Slack, just at mentioning bots doing your code. Like, what are you doing? We're constantly just changing what are the workflow paradigms on this. Do you guys have like a Slack coworker agent, for lack of a better term? We have a few that have that shape. I'm pretty excited about like, like cloud tag as like a form factor, you know, you still have to again kind of get the team construct right and the data right. But we have a variety of ways that people, you know, kind of work with agents and slack. I don't know if it's as as sort of slack pilled as Benioff would like us to be, um, or as like anthropic or opening eyes, but like we're, we're heading in that direction. Yeah. I think that history books will be written about, you know, the art of business in this time. Like I would love to read the new, the art of war with like everything that is happening here. Cause I think it's pretty extraordinary stuff we're seeing. What do you think it takes to win an AI versus pre AI? And like, what does it feel like to be a founder right now versus when you started books? I am both jealous and... of and also not jealous of like like the young founders you meet because you're just like like ought to be young again and and the whole world is your oyster and you can go in you know any direction and the leverage you have you'll meet with them and you're like oh my god like I saw a product a week and a half ago and they did a demo of the product and I was like I was like Absolutely, this would have been a 40-person project. five years ago, like, or, you know, especially when we were starting out easily 40 person project. And it was, it was two people. And you're just like, how do you have so many tabs that work? And they all seem to have stuff behind the tabs that all seem very functional. Like this is not fake. Um, and, and so I'm very jealous of that. Like, it's like incredible because you could just like, you can start your company from scratch. with that as the design principle. Now we will get there because we're just going to muscle through it. There's a couple of things that we just can't do, which is like, we're not like, we're very uncomfortable at the idea of like, sort of, you know, removing the code review and like some of these things to get talked about because 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. But so jealous of being able to kind of be fresh in that. And then on the other hand, like it's like, man, like at the same time, for every great idea, it's like instantly five competitors. We didn't have that problem. We had a, we had a good kind of couple of years where we just could like, we could just grind on our product and our, and our sort of experience. And it wasn't like every three days you were like, Oh, like Sequoia funded this thing and benchmark funded this thing. And like, well, you know, we weren't going kind of nuts like with that. Now we had our own version of that. So like at the time I probably was going, going nuts, but it was like, like in retrospect, it was like not It's like, oh, man, this is like a real race in every one of these markets. Totally. So... I think I'm probably pretty consensus on this, which is in a world where AI builds things so much faster, then probably the shift moves to whoever can actually get it to the customer is in the best position. So, you know, it's fun talking to founders that are pretty kind of pilled on that. You know, Scott or Matan are like they get the mandate. They're just like this thing is going to be an enterprise diffusion play. And so you have to get it to the enterprise. mandate right now, you're just going to lose. It's just like, it's game over. Sorry. Like there is quite literally trillion to trillions up for grab at the applied layer and the companies that, that know how to build the teams and get to the enterprise will be the ones that win. It's just like obviously guaranteed. Well said. Aaron, this is a very fun conversation. Thank you so much for joining. Thanks for having me. Music [65:00] Thank you.