20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

以下是内容的中文翻译： Fireworks.ai的创始人兼首席执行官林国在20VC节目上，讨论了公司快速增长的历程以及对人工智能未来的展望。Fireworks.ai在四年内将年经常性收入（ARR）提升至10亿美元，公司专注于推理层业务，介于英伟达（Nvidia）等芯片供应商和模型开发者之间。林国的核心理念围绕着“专门化智能”，从根本上挑战了由单一通用人工智能（AGI）主导的未来。 林国认为，绝大多数数据，尤其是企业私有数据，尚未被当前的通用智能模型所充分利用。他将OpenAI和Anthropic等公司视为正在构建重要的“电力线”或基础性基础设施。然而，他认为真正的价值和多样性体现在那些“应用终端”（appliances）中——即基于专有数据和独特工作流程构建的专门化应用。林国呼应英伟达（Nvidia）的黄仁勋（Jensen Huang），强调“每家公司都建立在一种独特的信念之上”，创造出独特的东西来证明其存在的价值，因此必然需要专门化的人工智能。他反对“由一个标准统治”或由一家公司决定品味的观念，强调了人类的创造力和多样性。 Fireworks.ai的战略严重依赖于开源模型，林国推崇它们，因为它们能为用户提供完全控制权、更大的定制能力以及显著降低的成本。他指出许多使用昂贵前沿模型的公司面临“扩展至破产”的问题，并预测未来三年内token的成本将“降低10倍”，这将反过来推动“100倍的使用量”。这种成本效率，再加上Fireworks.ai对质量的坚定不移的关注（确保训练和推理之间的“位等效”），对于企业的采用至关重要。他指出，开源模型的权重获取成本为零，能够提供深度定制能力，在解决特定问题上往往优于通用模型。 Fireworks.ai目前每天处理超过40万亿个token，主要来自定制模型，林国预计到年底，其当前的ARR将至少翻一番。尽管承认目前的利润率低于传统SaaS，但他将这归因于公司正处于“超高速增长阶段”，在此阶段，积极扩张和创新优先于眼前的利润优化。他强调了敏捷性和专注于Fireworks.ai核心优势的重要性，强调应利用合作伙伴，而不是试图拥有从芯片到应用程序的整个堆栈。尽管未来可能会考虑建设数据中心，一旦公司达到一定规模（就像Meta那样），但考虑到AI工作负载的动态性，芯片开发被认为是时机未到的。 林国认为，当前的一个主要瓶颈是缺乏“针对超大型模型”（例如10万亿参数）的优秀系统设计，这需要整个AI堆栈的复杂协同设计。他还强调了硬件折旧速度的加快，新模型总是倾向于最新的硬件，这使得长期硬件投资策略变得复杂。 回顾自己的创业历程，现年48岁的创始人林国分享说，他推迟了创业，直到他对“人”的理解磨练成熟。他最近聘请了曾任Salesforce总裁的乔治•胡（George Hu），称赞胡兼具经验和好奇心。林国认为，AI行业需要具备“矛盾特质”的人才，例如深厚的经验与强烈的好奇心相结合。他从黄仁勋那里学到的最重要一课是，对于有效、快节奏的领导力而言，深入、持续了解背景信息至关重要。展望未来，林国坚信，在三年内，“每家公司都将必须拥有自己的智能，而非可选项”，这与当前拥有软件堆栈的必要性如出一辙。

Lin Kuo, founder and CEO of Fireworks.ai, discussed his company's rapid growth and vision for the future of AI on 20VC. Fireworks.ai, which has scaled to $1 billion ARR in four years, operates in the inference layer, situated between chip providers like Nvidia and model developers. Kuo's central philosophy revolves around "specialized intelligence," fundamentally challenging a future dominated by a single, general Artificial General Intelligence (AGI). Kuo argues that the vast majority of data, particularly private corporate data, remains untapped by current general intelligence models. He sees companies like OpenAI and Anthropic as building essential "power lines" or foundational infrastructure. However, he believes the true value and diversity lie in the "appliances"—specialized applications built on proprietary data and unique workflows. Echoing Nvidia's Jensen Huang, Kuo emphasized that "every company is built on a special belief," creating something unique that justifies its existence, thus necessitating specialized AI. He rejects the notion of a world "ruled by one standard" or taste dictated by a single company, highlighting human creativity and diversity. Fireworks.ai's strategy heavily relies on open models, which Kuo champions for offering users full control, greater customization, and significantly lower costs. He points out the "scaling to bankruptcy" problem for many companies using expensive frontier models and predicts a "10x cost reduction in the next three years" for tokens, which will, in turn, drive "100x usage." This cost efficiency, combined with Fireworks.ai's unwavering focus on quality (ensuring "bit equivalence" between training and inference), is crucial for enterprise adoption. Open models, he notes, have zero acquisition cost for their weights, providing deep customization capabilities often superior to general models for specific problems. Fireworks.ai currently processes over 40 trillion tokens daily, predominantly from customized models, and Kuo expects to at least double their current ARR by year-end. While acknowledging that current margins are lower than traditional SaaS, he attributes this to the company being in a "hyper-growth phase" where aggressive expansion and innovation take precedence over immediate margin optimization. He stresses the importance of agility and focus on Fireworks.ai's core strength, emphasizing leveraging partners rather than attempting to own the entire stack from chips to applications. While building data centers might be considered in the future, once the company reaches a certain scale (much like Meta did), chip development is viewed as premature given the dynamic nature of AI workloads. Kuo identifies a significant current bottleneck as the lack of a "great system design for very large models" (e.g., 10 trillion parameters), requiring complex co-design across the entire AI stack. He also highlights the accelerating pace of hardware depreciation, where new models consistently prefer the newest hardware, complicating long-term hardware investment strategies. Reflecting on his journey, Kuo, a 48-year-old founder, shared that he delayed entrepreneurship until he honed his understanding of "people." He recently hired George Hu, former Salesforce president, praising Hu's blend of experience and curiosity. Kuo believes the AI industry demands individuals with "contradictory characteristics" like deep experience coupled with high curiosity. His biggest lesson from Jensen Huang is the importance of deep, constant context for effective, fast-paced leadership. Looking ahead, Kuo firmly believes that within three years, "every single company will own their own intelligence as a must-have, not optional," drawing parallels to the current necessity of owning a software stack.