以下是这段内容的中文翻译： 这期播客节目邀请了主持人哈里·斯泰宾斯与Glean创始人兼Rubrik联合创始人阿尔温德·贾恩进行了一场深刻而富有见地的讨论。贾恩从其丰富的经验出发，对不断演进的AI格局，特别是企业AI视角，提出了独到见解。 **Glean的愿景与企业AI的采用** 贾恩将Glean描述为一家企业AI公司，起初是作为企业搜索工具，好比“你工作生活中的谷歌”。它已经发展成为一个AI平台，如同ChatGPT、Claude、Gemini的“超集”，能够连接公司内部的上下文，充当员工的“同事”。 他谈到了大型企业对前沿模型提供商的疑虑，这与亚历克斯·卡普（Palantir创始人）的观点不谋而合。贾恩解释说，这种担忧源于企业对技术和运营依赖性、知识产权及数据控制权潜在丧失的顾虑，以及将核心业务转移给外部模型提供商的感知风险。他强调，如果这些模型驱动的智能代理人（agents）处理了大部分工作，企业可能会失去对其积累的机构知识和运营流程的控制权。 **前沿模型与开源：不断变化的格局** 阿尔温德认为，前沿模型公司对Glean这样的AI公司而言是“巨大的资产”，它们使得原本无法实现的产品交付成为可能。他不认同这些模型公司在应用层面构成直接竞争的观点。 然而，他指出企业内部正在显著转向采用开源模型。这一趋势主要由成本考量和对更大控制权的渴望所驱动。尽管早期对模型公司数据安全性的担忧已基本消退，但专有模型不断上涨的成本加速了开源模型的普及。贾恩声称，90%或更多的企业用例都可以由包括开源在内的各种模型处理，从而导致模型层商品化。他预测，在三年内，大多数企业工作负载将运行在开源模型上。然而，他对使用中国开源模型可能带来的地缘政治影响表示担忧，这可能涉及潜在的“后门”或竞争劣势。 **AI“荒谬”的成本与捆绑销售的挑战** 贾恩认为当前前沿模型的定价“荒谬地昂贵”，他举例说一个内部智能代理人每月成本高达100万美元，甚至比它所取代的人工团队还要贵。他认为技术应该随着时间的推移变得更便宜，而模型提供商近期提高token价格是一种反常现象。他预计推理成本将大幅下降。 关于竞争，微软与Co-pilot的捆绑销售策略构成了严峻挑战。尽管“一流”软件仍能蓬勃发展，但捆绑销售使其面临困难。然而，贾恩认为，向基于消费的AI模型转变可能会打破捆绑销售模式，因为公司将为跨各种工具的实际使用付费，这可能有利于专业的解决方案。 **AI投资回报率的衡量与工作的未来** 贾恩承认AI的投资回报率（ROI）很复杂。尽管在客户支持等领域（可衡量的生产力提升）看到了明显的价值，但在编码方面则不那么清晰。尽管代码编写速度显著加快，但整体产品交付速度并未必然提升，因为编码只是开发过程的一部分。他将这归因于需要在AI“周围进行投资”，提供正确的上下文，从而使其更快、更便宜，而不是“蛮力”地使用原始模型。 贾恩质疑“用AI取代自己”的观念，认为AI尚未准备好完全取代工作岗位。他强烈反对团队缩减的趋势，认为那些维持更大、由AI增强的团队的公司将打造出10倍更好的产品并获得竞争优势。他预见到“复合型角色”（例如，工程师/产品经理/设计师混合体）的兴起，以及数据分析师和招聘领域中的寻源专员等专业角色的过时。 **创业生态系统与创始人洞察** 贾恩发现当前的招聘比SaaS巅峰时期更容易，但顶尖的AI/机器学习人才除外，他们的薪水已大幅上涨。他建议创始人尽早尽可能多地筹集资金。他也批评了创业生态系统中“资本过剩”的现象，认为这有时会导致不可持续的结构和支出（例如过高的薪水），从而阻碍了建立伟大公司所需的纪律性。 最后，他揭开了创业光环的假象，将其描述为一份“压力巨大”的工作，需要有使命感的驱动和对现状“持续不满”，才能不断推动改进。

The podcast features an insightful and robust discussion between host Harry Stebbings and Arvind Jain, the founder of Glean and co-founder of Rubrik. Jain offers a seasoned perspective on the evolving AI landscape, particularly from an enterprise standpoint. **Glean's Vision and Enterprise AI Adoption** Jain describes Glean as an enterprise AI company that began as a search tool for businesses, analogous to "Google for your work life." It has evolved into an AI platform, acting as a "superset of ChatGPT, Claude, Gemini," connecting to a company's internal context to serve as a co-worker for employees. He addresses the skepticism of large enterprises towards frontier model providers, echoing Alex Karp's (Palantir) sentiments. Jain explains this fear stems from concerns about technological and operational dependence, potential loss of IP and data control, and the perceived risk of transferring core operations to external model providers. He highlights that if agents powered by these models handle the majority of work, enterprises could lose control over their accumulated institutional knowledge and operational processes. **Frontier Models vs. Open Source: A Shifting Landscape** Arvind views frontier model companies as a "huge asset" for AI companies like Glean, enabling product delivery that would otherwise be impossible. He rejects the notion that they are direct competition in the application layer. However, he notes a significant shift towards open-source models in the enterprise. This movement is primarily driven by cost considerations and a desire for greater control. While early concerns about data security with model companies have largely subsided, the escalating costs of proprietary models have propelled open source adoption. Jain claims that 90% or more of enterprise use cases can be handled by various models, including open source, leading to a commoditization of the model layer. He predicts that within three years, the majority of enterprise workloads will run on open-source models. He raises concerns, however, about the geopolitical implications of using Chinese open-source models due to potential "backdoors" or competitive disadvantages. **The "Absurd" Cost of AI and Bundling Challenges** Jain believes current frontier model pricing is "absurdly expensive," citing an example of an internal agent costing $1 million per month, which was more expensive than the human team it replaced. He argues that technology should become cheaper over time, and the recent increase in token prices by model providers is an anomaly. He anticipates a significant drop in inferencing costs. Regarding competition, Microsoft's bundling strategy with Co-pilot poses a formidable challenge. While "best-of-breed" software can still thrive, bundling makes it difficult. However, Jain suggests that the shift to consumption-based AI models could disrupt bundling, as companies will pay for actual usage across various tools, potentially favoring specialized solutions. **Measuring AI ROI and the Future of Work** Jain acknowledges that AI's ROI is complex. While clear value is seen in areas like customer support (measurable productivity gains), it's less clear in coding. Despite significantly faster code writing, overall product shipping speed hasn't necessarily increased, as coding is just one part of the development process. He attributes this to the need for "investment around" AI to provide the right context, making it faster and cheaper, rather than "brute-forcing" with raw models. Jain challenges the notion of "replacing yourself with AI," arguing that AI isn't ready for full job replacement. He strongly disagrees with the trend of shrinking teams, believing that companies that maintain larger, AI-augmented teams will build 10x better products and gain a competitive edge. He envisions the rise of "composite roles" (e.g., engineer/PM/designer hybrids) and the obsolescence of specialized roles like data analysts and sourcers in recruiting. **Startup Ecosystem and Founder Insights** Jain finds current recruiting easier than during the SaaS peak, except for top-tier AI/ML talent, whose salaries have dramatically increased. He advises founders to raise as much capital as possible upfront. He also critiques the "overabundance of capital" in the startup ecosystem, suggesting it sometimes leads to unsustainable structures and spending (like inflated salaries), hindering the discipline needed to build great companies. Finally, he dispels the glamour surrounding founding, describing it as a "stressful" job requiring a mission-oriented drive and a "consistent unhappiness" with the status quo to continuously push for improvement.