亚伦·利维 (Aaron Levy)，Box 首席执行官，分享了关于人工智能对企业软件的变革性影响、Box 的战略以及更广泛技术格局的见解。他首先强调了在 AI 讨论中保持“紧密联系”的重要性，尤其是在推特 (Twitter) 等平台上，他认为这对大学生的职业发展至关重要。 利维坚信，在 AI 时代，应用公司，即“新实验室”（neolabs），至关重要。尽管基础模型极具价值，但模型的能力与企业具体工作流程之间存在着“万亿美元的鸿沟”。弥合这一“鸿沟”需要大量工作，而纯研究机构不太愿意承担这些工作。他将此比作基础设施提供商（AWS、GCP）和应用公司（Snowflake、Databricks），并预测智能模型将创造数万亿美元的价值，但将这些模型引入各行业具体工作流程的应用层也将创造巨大价值。 Box 最初是一个云内容管理平台，现正围绕 AI 进行自我重塑。自 2015 年以来，Box 曾涉足 AI，而“ChatGPT 时刻”促使公司进行了全面转型。Box 拥有数千亿个包含关键非结构化数据（合同、研究资料、营销资产）的文件，现在其 AI 代理能够回答有关这些数据的问题、提取元数据并自动化工作流程。这些“长期运行的代理”可以将银行客户入职等复杂流程从数周缩短到数小时，从根本上改变企业的运营方式。 在讨论 AI 生成内容时，利维区分了编程和其他知识工作。在编程领域，AI 被视为文本生成的实用工具而受到欢迎。然而，在内容创作（例如董事会演示文稿）中，存在一种“工作粗糙”（work slop）现象，用户会质疑 AI 生成内容背后的真实性和人类思维过程。他承认这是一个混乱的时期，但预计会有一个演变过程，类似于社会接受宏生成的财务模型。 Box 开发了一个“代理框架”（agentic harness），利用其对文件系统权限、搜索和用户行为的深入理解。该框架能够对 Box 内部的海量数据集进行高精度、低延迟的查询，由于领域特定调优，其性能优于原始的 LLM API 调用。Box 使用内部基准评估模型，包括针对不同行业的“复杂工作评估”（complex work eval）和基于 Box 员工数据的内部“留存评估”（hold back eval），并维护一个“模型花园”（model garden）供客户选择。 利维指出，模型性能竞赛不相上下，Gemini 在 Box 的某些用例中表现出特别的优势。他观察到，开放权重模型的采用程度超出了普遍认知，这主要是受成本驱动。他认同这样一个观点：随着用例的成熟，它们可以被“剥离”给开源模型，从而形成一种混合方法，即专有模型处理复杂的编排，而更便宜的开放权重模型管理长尾任务。这种动态将导致专有模型和开放权重模型的使用量都呈指数级增长。 关于记忆和定制化，利维认为持续学习引人入胜，但他强调了企业访问控制的挑战。用个性化、自适应的权重训练模型需要处理严格的权限，这目前有利于基于 RAG（检索增强生成）的方法，即通过查找获取上下文，而不是将其固化到权重中。他设想未来模型将封装特定的推理能力，而实际内容则通过查找系统进行访问。 关于记录系统（SoR）在 AI 驱动世界中的作用，利维认为 SoR 必须做到两点：为其自身产品构建一个极其有效的代理（比通用代理性能高 10-20 个百分点），并为外部代理开放强大的 API。这将创造新的收入来源和机会，将 SoR 转型为“以交易量为导向的商业模式”，通过数据洞察增加巨大价值。 他将 AI 在编程领域的快速普及（得益于其基于文本的特性、高质量训练数据、技术用户和明显的生产力提升）与在其他知识工作领域较慢的普及进行了对比。例如，对于销售代表而言，外部因素严重限制了生产力，使得自动化难以实现。在编程之外的这种较慢普及，对那些愿意深入一线耕耘，将 AI 整合到复杂现实工作流程中的应用型 AI 公司来说，代表着巨大的机遇。 最后，在为创始人提供建议时，利维强调了他对 AI 的个人迷恋，这源于他与非结构化数据打交道的 20 年经验。对于公司建设，他强调了通过良好的数据卫生和架构，使业务“便于 AI 理解”的重要性。Box 得益于其所有数据都存储在自己的平台中。他建议关注高杠杆影响力的工作流程，建立卓越中心，并培养“AI 优先”的文化。他指出，新 AI 公司具有非凡的杠杆效应（两个人就能完成过去 40 人的工作），但也存在激烈的竞争。最终，在 AI 时代获得成功的将是那些能够通过深厚的领域专业知识和在应用层不懈执行，有效“将其交付给客户”的公司。

Aaron Levy, CEO of Box, shared insights on AI's transformative impact on enterprise software, Box's strategy, and the broader tech landscape. He began by emphasizing the importance of staying "wired in" to AI discussions, particularly on platforms like Twitter, which he sees as crucial for career development for college students. Levy strongly believes that application companies, or "neolabs," are paramount in the AI era. While foundational models are incredibly valuable, there's a "trillion-dollar gap" between a model's capabilities and an enterprise's specific workflows. This "bridge" requires significant work, which pure research organizations are less inclined to do. He draws an analogy to infrastructure providers (AWS, GCP) and application companies (Snowflake, Databricks), predicting that intelligence models will create trillions in value, but the application layer bringing these models into specific workflows across industries will also create immense value. Box, originally a cloud content management platform, is reinventing itself around AI. Having flirted with AI since 2015, the "ChatGPT moment" spurred a full company pivot. With hundreds of billions of files containing critical unstructured data (contracts, research, marketing assets), Box's AI agents can now answer questions about this data, extract metadata, and automate workflows. These "long-running agents" can streamline complex processes, like bank onboarding, from weeks to hours, fundamentally changing how enterprises operate. Discussing AI-generated content, Levy distinguishes between coding and other knowledge work. In coding, AI is embraced as a utility for text generation. However, in content creation (e.g., board decks), there's a "work slop" phenomenon, where users question the authenticity and human thought process behind AI-generated output. He acknowledges this messy period but anticipates an evolution, similar to how society accepted financial models generated by macros. Box has developed an "agentic harness" that leverages its deep understanding of file system permissions, search, and user behavior. This harness allows for highly accurate and low-latency queries over massive datasets within Box, outperforming raw LLM API calls due to domain-specific tuning. Box evaluates models using internal benchmarks, including a "complex work eval" for various industries and an internal "hold back eval" based on Box employee data, maintaining a "model garden" for customer choice. Levy notes the model performance race is neck and neck, with Gemini showing particular strength in some Box use cases. He observes that open-weight models are seeing more adoption than often perceived, driven by cost. He subscribes to the idea that as use cases mature, they can be "peeled off" to open-source models, creating a blended approach where proprietary models handle complex orchestration while cheaper open weights manage long-tail tasks. This dynamic will lead to an exponential growth in both proprietary and open-weight usage. Regarding memory and customization, Levy finds continual learning fascinating but highlights the challenge of enterprise access controls. Training models with personalized, adapting weights needs to navigate strict permissions, which currently favors a RAG-based (retrieval-augmented generation) approach where context is looked up rather than baked into weights. He envisions a future where models encapsulate specific reasoning capabilities, while actual content is accessed via lookup systems. On the role of systems of record (SoRs) in an AI-driven world, Levy argues SoRs must do two things: build an incredibly effective agent specifically for their own product (outperforming generic agents by 10-20 points) and expose robust APIs for external agents. This creates new revenue streams and opportunities, transforming SoRs into "volume-oriented business models" that add tremendous value through data insights. He contrasts the rapid diffusion of AI in coding (due to its text-based nature, high-quality training data, technical users, and clear productivity gains) with the slower diffusion in other knowledge work. For sales reps, for example, external factors heavily constrain productivity, making automation harder to implement. This slower diffusion outside of coding represents a massive opportunity for applied AI companies willing to do the "pounding pavement" to integrate AI into complex, real-world workflows. Finally, offering advice for founders, Levy emphasizes his personal fascination with AI, fueled by his 20 years with unstructured data. For company building, he stresses the importance of making businesses "legible for AI" through good data hygiene and architecture. Box benefits from having all its data in its own platform. He advises focusing on high-leverage impact workflows, establishing centers of excellence, and fostering an "AI-first" culture. He notes the extraordinary leverage of new AI companies (two people building what once took 40), but also the intense competition. Ultimately, success in the AI era will come to those who can effectively "get it to the customer" through deep domain expertise and relentless execution at the applied layer.