Factory的首席技术官兼联合创始人Eno Reyes对不断发展的AI格局提出了深刻见解，挑战了当前关于市场估值、模型经济学和战略业务决策的普遍假设。他认为，衡量AI真正价值的标准不在于基于token成本的“最便宜模型”，而在于能够高效提供最佳结果的“最便宜系统”。一个能够首次就正确执行任务的精密模型，即使其每个token的成本更高，最终也比那些更便宜、准确性更低且需要大量返工的替代方案更经济。 这种对结果的关注将推动模型的快速“物种分化”。Reyes设想，未来不会是少数几个主导的前沿模型，而是企业会基于商品化的开源模型，开发高度专业化、经过后期训练的模型，用于其独特、高流量的内部任务。这些企业将作为专有资产加以保护的内部模型，将与处理商品化任务的通用开源模型并行运作。他认为，这一趋势表明，前沿模型的潜在市场总量（TAM）被“高估了”，预测它们将扮演的角色远比普遍认为的要小，更趋向于利基市场。 Reyes批评当代AI领导者们在“最糟糕的营销工作”中利用了围绕通用人工智能（AGI）的“恐吓策略”和炒作，认为这些叙事不仅疏远了公众，也扭曲了这项技术的真正潜力和风险。他指出，萨姆·奥特曼最近承认低估了经济发展势头，是一个积极的“反思”。 Reyes哲学的核心宗旨是“主权智能”。他强调企业拥有自己的AI学习成果和工作流程至关重要，并警告说，依赖外部模型提供商存在厂商锁定和潜在竞争的风险，因为一些大型模型提供商已明确表示有意进入他们所服务的行业。这种控制需求是推动本地部署AI解决方案（on-premise AI solutions）需求的主要因素，让企业能够安心地保留对其知识资产的所有权和控制权。此外，他强调AI的持续学习主要发生在“衔接层”（harness layer）——即维护逻辑和状态的应用层，而非模型本身。 展望竞争格局，Reyes预测，未来18个月内，80-90%的现有“新实验室”（neolabs，即新兴AI初创公司）可能会“消亡”（被收购或作为独立实体失败）。而那些能够蓬勃发展的，将是专注于不易商品化、具有持久性的专有工作流程的企业，例如法律科技，而非通用知识工作。他还驳斥了“中国模型”相比美国模型天生更可怕或有风险的说法，称之为前沿实验室发动的“心理战”（psyop）。他认为，对审查、偏见和生命周期（longevity）的担忧适用于所有模型，无论其来源如何，并且必须结合具体语境进行评估。 Reyes大胆预测，三年内，99%的工作流程将由开源模型处理，而剩余的1%（高度利基、前沿科学或安全任务）将占据30-40%的经济价值。他认为微软在这种环境中处于高超的有利位置，既能从OpenAI获得优势，又能同时在Azure上培养模型独立性，他认为这一战略对于支持多样化的企业需求至关重要。 Factory独特的招聘策略体现了他们以结果为导向的方法：他们旨在收购那些已经深入参与为Factory解决问题空间构建解决方案的小公司或个人创始人。这优先考虑“使命一致性”和独立解决问题的能力，而非传统的简历或“表演式工作文化”。他建议其他创始人与企业客户一起专注于发现问题和解决问题，而不是说服，以此来创造价值。 Reyes最后预测，在五年内，无法即时为几乎任何问题动态生成定制软件的想法将显得“荒谬”。他设想，未来即使是利基服务提供商，例如伯利兹的船只操作员，都将拥有超越当今企业应用质量的复杂定制软件界面，这标志着人类与技术互动和利用方式的根本性转变。

Eno Reyes, CTO and co-founder of Factory, offers a profound perspective on the evolving AI landscape, challenging prevalent assumptions about market valuations, model economics, and strategic business decisions. He contends that the true measure of AI's value lies not in the "cheapest model" based on token cost, but in the "cheapest system" that delivers optimal outcomes efficiently. A sophisticated model that executes a task correctly the first time, even if its per-token cost is higher, ultimately proves more economical than a cheaper, less accurate alternative requiring significant rework. This focus on outcomes will drive a rapid "speciation of models." Instead of a few dominant frontier models, Reyes envisions a world where companies develop highly specialized, post-trained models for their unique, high-volume internal tasks, built upon commodity open models. These internal models, which companies will guard as proprietary assets, will operate alongside more general open models for commodity tasks. This trend, he argues, suggests that the Total Addressable Market (TAM) for frontier models is "overweighted," predicting a much smaller, niche role for them compared to widespread belief. Reyes criticizes the "worst marketing job" by contemporary AI leaders who have used "scare tactics" and hype around AGI, suggesting that such narratives not only alienate the public but also misrepresent the technology's true potential and risks. He points to Sam Altman's recent acknowledgment of underestimating the economy's momentum as a positive "reckoning." A central tenet of Reyes's philosophy is "sovereign intelligence." He stresses the critical importance for businesses to own their AI learnings and workflows, warning that reliance on external model providers risks vendor lock-in and potential competition, as some large model providers have explicitly stated intentions to enter the industries they serve. This need for control is a major driver behind the demand for on-premise AI solutions, offering businesses the peace of mind that they retain ownership and control over their intellectual assets. Furthermore, he emphasizes that continuous learning in AI happens predominantly at the "harness layer" – the application layer where logic and state are maintained – rather than within the models themselves. Looking at the competitive landscape, Reyes predicts that 80-90% of current "neolabs" (new AI startups) could "die" (be acquired or fail as independent entities) within the next 18 months. Those that will thrive are businesses focused on durable, proprietary workflows not easily commoditized, such as legal tech, as opposed to generalized knowledge work. He also dismisses the notion of "Chinese models" as inherently scary or risky compared to American models, labeling it a "psyop" by frontier labs. He argues that concerns around censorship, bias, and longevity apply to *all* models, regardless of origin, and must be evaluated contextually. Reyes boldly predicts that in three years, 99% of workflows will be handled by open models, with the remaining 1% (highly niche, frontier scientific or security tasks) holding 30-40% of the economic value. He sees Microsoft as masterfully positioned in this environment, capturing upside from OpenAI while fostering model independence on Azure, a strategy he believes is critical for supporting diverse enterprise needs. Factory's unique hiring strategy exemplifies their outcome-oriented approach: they aim to acquire small companies or individual founders who are already deeply engaged in building solutions for Factory's problem space. This prioritizes "mission alignment" and independent problem-solving over traditional resumes or a "performative work culture." He advises fellow founders to focus on discovery and problem-solving with enterprise clients, rather than persuasion, to build value. Reyes concludes by predicting that within five years, the idea of not being able to instantly generate custom software for almost any problem on the fly will seem "ludicrous." He envisions a future where even niche service providers, like a boat operator in Belize, will have sophisticated, custom-built software interfaces that surpass the quality of today's enterprise applications, marking a fundamental shift in how humanity interacts with and leverages technology.