How AI Is Changing Enterprise

以下是将原文翻译为中文： Lightcone 播客邀请了 Y Combinator 的 Gary、Jared Harge 和 Diana，以及 Box 的 CEO Aaron Levy，共同探讨正在进行的 AI 革命及其对企业和初创公司的影响。他们深入研究了当前的 AI 格局，重点关注在基础模型之上构建应用程序的价值，而不是简单地为现有 AI 技术创建封装。 Aaron 反驳了 AI 初创公司应该害怕成为像 ChatGPT 这样的大型语言模型 (LLM) 的简单封装的观点。他强调，真正的价值在于专有的业务逻辑和围绕模型构建的软件层，这些都针对特定的客户工作流程和数据进行了定制。他认为，企业不需要模型；他们需要解决方案和结果，例如自动化客户支持或精简的合同管理，并与现有系统集成。 谈话转向了模型公司的角色，小组成员一致认为，纯粹的“模型公司”仅授权 AI token 的商业模式是不可持续的。Aaron 认为，AI 公司正日益成为软件公司，而 AI 模型为其底层软件提供支持。他以 Anthropic（面向企业的 API 业务）、OpenAI（拥有 AI 模型的软件公司）、Google (GCP) 和 Meta（开源模型）为例。Meta 的模型开源行为起到了平衡作用，将智能成本推向零，从而将重点从商品化的 token 转移到增值服务上。 讨论涉及商品化智能对初创公司的影响。Aaron 认为，AI 公司必须像传统的软件公司一样运营。初创公司需要专注于构建能够将复杂技术交付给客户以解决实际问题的软件。他强调了垂直 AI 的重要性，即构建针对特定行业和工作职能定制的 AI 驱动的解决方案。 对话还触及了像 DeepSeek 这样的开源推理模型对企业的影响。Aaron 指出，虽然推理模型显示出了前景，但它们的表现可能与非推理模型不同，并且企业可以从改进的智能中受益，因为它允许更高级的代理工作流程。他说，他目前看到银行采用通用聊天机器人的比例为 10%，采用代理工作流程的比例为 1%。 Aaron 解释了公司内部关注模型的人（CTO、AI 负责人、IT 人员）和只关心解决方案的人（业务线主管、最终用户）之间的区别。Aaron 强调了保持参与的重要性，因为最终客户可能只是切换到不同的模型。 小组成员讨论了一个初创公司如何最近在能够切换到不同的底层 AI 模型的同时提高其利润率。小组成员将当前的 AI 格局与云计算的早期阶段进行了类比，当时客户不太关心底层基础设施，而更关注最终用户体验。 他们讨论了为了让您的价值主张发挥作用，需要在 token 之上添加多少软件。然后他们深入探讨了 AI 产品的定价模式。公司应该按 AI 生成的潜在客户、合格的潜在客户还是按使用的资源获得报酬？ Aaron 分享了他与财富 500 强公司高级管理人员讨论的见解。他指出，人们的思想观念发生了重大转变，开始拥抱 AI，CEO 们公开承认 AI 具有彻底改变其业务的潜力。他说，AI 对企业来说存在竞争问题，他们需要实施 AI 战略和举措。这与云优先的心态截然不同，在云优先的心态中，云仅仅被视为一个效率故事。 Aaron 讨论了“背景 vs. 核心”框架，以确定是内部构建 AI 解决方案还是从外部供应商处购买。他强调，像人力资源和 ERP 系统这样的背景领域通常最好由第三方解决方案提供服务，而像生命科学中的药物开发或用于个性化的专有算法这样的核心领域应该在内部处理。例如，应该有一个非常强大的团队致力于药物开发，但临床试验的自动化可能应该来自供应商。他预计，随着企业将 AI 集成到其运营中，对 AI 专用开发工具和基础设施的需求将激增。 Aaron 分享说，Box 正在构建 AI 工具，以提高其员工的编码效率。该公司还使用人力资源问题和福利问题。此外，员工现在可以通过提问来查询人力资源数据。 最后，对话以对 AI 推动富足和改善生活方式的潜力的乐观展望结束。Aaron 提出了一个由杰文斯悖论驱动的革命，即 AI 导致自动化程度提高、成本降低以及获得教育和医疗保健的机会增加。他设想了一个乌托邦式的未来，在这个未来中，AI 赋予个人权力，创造一个更加繁荣和公平的社会。

The Lightcone podcast features Gary, Jared Harge, and Diana from Y Combinator, and their guest, Aaron Levy, CEO of Box, discussing the ongoing AI revolution and its implications for businesses and startups. They dive into the current AI landscape, focusing on the value of building applications on top of foundational models rather than simply creating wrappers for existing AI technology. Aaron dispels the idea that AI startups should fear being mere wrappers for large language models (LLMs) like ChatGPT. He emphasizes that the true value lies in the proprietary business logic and the software layer built around the model, tailored to specific customer workflows and data. He argues that enterprises don't want models; they want solutions and outcomes, such as automated customer support or streamlined contract management integrated with existing systems. The conversation shifts towards the role of model companies, and the panelists agree that the business model of a pure "model company" that solely licenses AI tokens is unsustainable. Aaron suggests that AI companies are increasingly becoming software companies with AI models powering their underlying software. He points to Anthropic (API business for enterprises), OpenAI (software company with AI models), Google (GCP), and Meta (open-source models) as examples. Meta's open-sourcing of models acts as a counterbalance, driving the cost of intelligence towards zero, which shifts the focus towards value add over the commoditized tokens. The discussion addresses the implications of commoditized intelligence for startups. Aaron believes that AI companies must operate like traditional software companies. Startups need to focus on building software that delivers complicated technology to customers to solve real-world problems. He emphasizes the importance of vertical AI, building AI-powered solutions tailored to specific industries and job functions. The conversation also touches upon the impact of open-source reasoning models like DeepSeek on the enterprise. Aaron notes that while reasoning models have shown promise, they can perform differently than non-reasoning models, and that the enterprise can benefit from improved intelligence as it permits more advanced agent workflows. He says he is currently seeing 10% adoption of general chatbots at banks and 1% adoption of agent workflows. Aaron explains the difference between those in the company who care about the model (CTO, head of AI, IT) and those who only care about the solution (line of business executives, end-users). Aaron stresses the importance of staying in the game because end customers could just switch to a different model. The panelists discuss how a startup recently managed to increase its margins while being able to switch to different underlying AI models. The panelists draw parallels between the current AI landscape and the early days of cloud computing, when customers were less concerned with the underlying infrastructure and more focused on the end-user experience. They discuss the importance of how much software is necessary on top of the tokens for your value proposition to work. They then get into pricing models for AI products. Should a company get paid for an AI generated lead, a qualified lead, or by the resources that are utilized? Aaron shares insights from his discussions with senior executives at Fortune 500 companies. He notes a significant shift in mindset towards embracing AI, with CEOs openly acknowledging its potential to revolutionize their businesses. He says that AI has competitive issues for enterprises and that they need to implement AI strategy and initiatives. This is very different from the cloud first mentality where cloud was just seen as an efficiency story. Aaron discusses the "context vs. core" framework for determining whether to build AI solutions internally or buy them from external vendors. He emphasizes that context areas, such as HR and ERP systems, are often better served by third-party solutions, while core areas, such as drug development in life sciences or proprietary algorithms for personalization, should be handled internally. For example, one should have a very strong team working on drug development but the automation of clinical trials should probably be from a vendor. He anticipates a surge in demand for AI-specific dev tools and infrastructure as enterprises integrate AI into their operations. Aaron shares that Box is building AI tools to make coding more productive for its employees. The company also uses HR question and benefits questions. Also, employees can now interrogate HR data with questions. Finally, the conversation concludes with an optimistic view of AI's potential to drive abundance and improve lifestyles. Aaron posits a Jevons Paradox-driven revolution where AI leads to increased automation, lower costs, and improved access to education and healthcare. He envisions a utopian future where AI empowers individuals and creates a more prosperous and equitable society.