20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

以下是内容的中文翻译： Clay Bavor 是人工智能公司 Sierra 的联合创始人，该公司估值近160亿美元，已累计融资超15亿美元。他分享了自己职业生涯和人工智能未来发展的深刻见解。在谷歌辉煌的18年职业生涯结束后，巴沃决定与布雷特·基内尔斯基（Brett Kinelsky）共同创立 Sierra，他认识到个人抱负、值得信赖的合作伙伴以及大型语言模型（LLMs）的变革潜力之间出现了独特的交汇点。他在谷歌的经历让他养成了深入投资技术栈的意愿，这一原则如今也应用于 Sierra，他们在利用开源模型的同时，构建了大量的专有解决方案。 巴沃讨论了不断发展的人工智能模型格局，并区分了“前沿（frontier）”模型和“开源（open）”模型。他认为，在科学、编码和法律工作等复杂领域，对“前沿智能水平”存在无限需求，这些领域的高风险和复杂性要求卓越的能力。他承认开源模型的强大功能日益增强，但他认为它们是互补的，适用于许多企业任务，并受益于前沿模型进步的“蒸馏”。他指出，中国公司经常将美国实验室的前沿模型进行蒸馏，以创建自己的先进开源权重模型。 对话深入探讨了token经济学。巴沃指出，尽管token成本预计会下降，但“推理模型”（AI推演过程）的兴起实际上推高了使用量。他预测，未来硬件的进步和向开源模型的迁移将降低成本，但GPU算力的根本性限制将维持价格底线。他驳斥了本地设备AI可以缓解前沿工作负载的服务器端挑战的观点，强调了所需巨大的计算能力。 Sierra 本身就是深度整合AI方法的典范。巴沃透露了他们的内部AI工具“Pinecone”和“Sierra Brain”。Pinecone 是一个内部代理，充当“单一MCP（主控制程序）网关”，整合公司系统，允许员工访问并推理所有已发布的信息，本质上赋予他们决策和任务完成的“超能力”。Sierra Brain 则作为一个战略思想伙伴，以公司文件和洞察为基础，帮助公司思考未来的方向。巴沃认为，目前开发者在token上的支出只占一小部分，但随着AI工具对生产力变得至关重要，这一比例可能会显著增长，可能达到开发者薪水的20%。 Sierra 的战略重点是大型企业，目前与财富50强中40%的公司合作。巴沃解释了他们的“前置部署（forward-deployed）”模式，这一灵感来源于 Palantir，即 Sierra 的工程师深度嵌入客户组织内部。他表示，这种方法可以加速价值实现时间并建立信任，尽管他也承认并非所有销售都严格需要这种模式。他设想 Sierra 将超越客户支持，管理整个客户生命周期，包括销售、转化和营销，Rocket 和 Next 等公司的合作案例就证明了这一点。 巴沃还阐述了 Sierra 独特的公司文化，其核心价值观是“精益求精（craftsmanship）、专注投入（intensity）和家庭至上（family）”。“精益求精”强调卓越地完成工作，通过一丝不苟的细致来建立客户信任。“专注投入”反映了AI市场竞争激烈、节奏快速的本质，要求创始人与员工保持高昂的精力，不懈追求雄心勃勃的目标。“家庭至上”则强调工作与生活平衡的重要性，认识到员工在工作之外也有自己的生活，支持个人承诺有助于培养忠诚度和幸福感。他还强烈倡导线下办公，特别是对于年轻公司而言，这有助于培养文化、指导和情谊。 在招聘方面，Sierra 彻底改变了其工程面试流程，使其成为“AI原生”的。候选人会获得一个提示，并分配一定预算使用AI编码代理，然后被要求构建一个应用程序，以展示他们熟练使用AI工具的能力。巴沃指出，他们一些最优秀的员工只有22-23岁，“完全被AI技术武装起来”，这展示了精通这些新技术的年轻人才的巨大价值。 回溯过去，巴沃分享说，桑达尔·皮查伊（Sundar Pichai）兼顾高层战略与细节的能力是他学到的一个重要经验。对于有抱负的专业人士，他推荐大卫·麦卡洛（David McCullough）的《莱特兄弟》（The Wright Brothers）一书，认为它是关于创业和发明的大师之作。作为四个孩子的父亲，他强调了专属家庭时间的重要性，建立仪式并参与孩子的兴趣，这些都是个人成就感的基石。

Clay Bavor, co-founder of Sierra, an AI company valued at nearly $16 billion with over $1.5 billion raised, shared profound insights from his career and the future of AI. After an illustrious 18-year tenure at Google, Bavor decided to co-found Sierra with Brett Kinelsky, recognizing a unique convergence of personal ambition, a trusted partner, and the transformative potential of large language models (LLMs). His time at Google instilled in him a willingness to invest deeply in the technology stack, a principle now applied at Sierra where they build extensive proprietary solutions even while leveraging open-source models. Bavor discussed the evolving landscape of AI models, distinguishing between "frontier" and "open" models. He believes there's an "unbounded demand for frontier levels of intelligence" across complex domains like science, coding, and legal work, where high stakes and complexity necessitate superior capabilities. While acknowledging the increasing power of open-source models, he sees them as complementary, suitable for many enterprise tasks and benefiting from the "distillation" of frontier model advancements. He highlighted that Chinese companies often distill frontier models from labs in the U.S. to create their own advanced open-weight models. The conversation delved into token economics, where Bavor noted that while token costs were expected to decrease, the rise of "reasoning models" (AI thinking out loud) has actually driven usage up. He predicts a future where hardware advancements and migration to open-source models will reduce costs, but the fundamental constraint of GPU capacity will maintain a floor on pricing. He dismisses the idea of local device AI alleviating server-side challenges for frontier workloads, emphasizing the sheer computational power required. Sierra itself exemplifies a deeply integrated AI approach. Bavor revealed their internal AI tools, "Pinecone" and "Sierra Brain." Pinecone, an internal agent, acts as a "single MCP gateway" aggregating company systems, allowing employees to access and reason over all published information, essentially giving them "superpowers" for decision-making and task completion. Sierra Brain serves as a strategic thought partner, grounding itself in company documents and insights to help reason about future directions. Bavor believes developer spending on tokens, currently a small fraction, will likely grow significantly, potentially reaching 20% of a developer's salary as AI tools become integral to productivity. Sierra’s strategy focuses on large enterprises, working with 40% of the Fortune 50. Bavor explained their "forward-deployed" model, inspired by Palantir, where Sierra engineers are deeply embedded within customer organizations. This approach, he states, accelerates time-to-value and builds trust, though he acknowledges it's not strictly necessary for every sale. He envisions Sierra expanding beyond customer support to manage the entire customer lifecycle, including sales, conversion, and marketing, as demonstrated by their work with companies like Rocket and Next. Bavor also shed light on Sierra's distinct company culture, built on values of "craftsmanship, intensity, and family." Craftsmanship emphasizes doing things with excellence, building trust with customers by demonstrating meticulous care. Intensity reflects the competitive and fast-paced nature of the AI market, requiring founders and employees to maintain high energy and a relentless pursuit of ambitious goals. "Family" highlights the importance of work-life balance, recognizing that employees have lives outside of work, and supporting personal commitments fosters loyalty and well-being. He also strongly advocates for in-person work, especially for young companies, to cultivate culture, mentorship, and camaraderie. In terms of hiring, Sierra has radically altered its engineering interview process to be "AI-native." Candidates are given a prompt, provided a budget for AI coding agents, and then asked to build an application, demonstrating their proficiency with AI tools. Bavor notes that some of their most effective employees are 22-23 years old, "completely AI-pilled," showcasing the immense value of young talent proficient in these new technologies. Reflecting on his past, Bavor shared that Sundar Pichai's ability to span high-level strategy and minute details was a key lesson. For aspiring professionals, he recommended David McCullough's "The Wright Brothers" as a masterclass in entrepreneurship and invention. As a father of four, he stressed the importance of dedicated family time, establishing rituals, and engaging with his children's interests as a cornerstone of personal fulfillment.