在一场广泛的讨论中，Factory的联合创始人兼首席执行官Matan详细阐述了公司的发展历程、竞争战略以及对软件开发行业的未来愿景。他挑战了“客户至上”（customer obsession）的传统观念，认为那是一个投入指标。相反，Factory致力于打造卓越的产品，让客户 *对其着迷*，强调产出指标而非内部努力的重要性。 Matan回顾了Factory充满挑战的早期岁月，亲切地称之为“沙漠之旅”，始于2023年4月，当时自主智能体（autonomous agents）仍是一个新兴概念。尽管超前于时代，这段时期让Factory得以完善其为企业开发者构建产品的方法。他坦率地承认，“早了两三年，就等于错了”，突显了公司初期在那些尚未准备好接受全自主解决方案的企业中难以获得关注的困境。 Factory的一个关键时刻是做出了一个激进的决定：全额退款给所有客户，尽管已产生了近200万美元的收入。Matan解释说，产品虽然有销量，但并未真正让开发者满意。这一艰难的抉择，源于他们“创造着迷客户”的核心运营原则，凸显了他们对产品质量和长期信任的承诺，而非短期收入。他指出，这段经历在团队中培养了深刻的韧性，他们选择留下，尽管公司遭遇了财务挫折。 随着开发者心态的变化以及Factory的Droid CLI于2025年9月发布，市场发生了显著变化。这款工具满足了开发者的即时需求，提供了最先进、模型无关的性能。Matan观察到，受Andre Karpathy等人物影响的开发者开放心态，与模型进步同样关键。他认为，像Factory这样的多模型整合器（multi-model harness）优于紧密耦合的模型-整合器协同设计，因为它避免了过拟合，并受益于接触多样化的模型行为，这就像在整个互联网上训练AI，而不是仅仅使用个人数据。 Matan随后详细阐述了AI应用从“最大化token使用”到“成本合理化”的演变。最初推动AI使用导致了不加区分的token消耗。Factory通过其路由器解决这一问题，该路由器根据成本和性能动态地将任务路由到最合适的模型，避免了将昂贵的前沿模型用于微不足道的查询（例如，“天气怎么样？”）。他强调了开源模型日益突出的地位，指出由于其可与“前沿减一”模型媲美的性能、速度和成本效益，目前它们在Factory的内部token使用中占据了两位数的显著比例。 谈及商业模式，Matan表示Factory目前是基于使用量的计费模式，但他设想未来会转向基于结果的定价，这可能通过一个任务的竞争性市场来实现。他还分享了自己对“软件工厂”的愿景，即组织将低效、以人为中心的流程转变为严谨、自动化的软件开发装配线。这包括将领域知识（tribal knowledge）编码化，并建立清晰的反馈循环以衡量成果，从而让企业能够根据核心能力和预期结果，定量地分配资源（人力与token）。 展望未来，Matan预测在12-24个月内，90%的AI token将是异步的，标志着向“黑暗工厂”的转变，在其中，自主智能体无需持续的人工启动即可持续运行。尽管承认可能出现短期动荡和资源错配，但他对长远未来表达了深深的乐观。他认为AI将解放工程师，使其能够应对更广泛的全球问题，从而为政府服务和制药研究等领域创造出卓越的软件，最终为“世界带来净效益”。

In a wide-ranging discussion, Matan, co-founder and CEO of Factory, elaborated on his company's journey, competitive strategy, and a forward-looking vision for the software development industry. He challenged the conventional notion of "customer obsession," arguing that it's an input metric. Instead, Factory strives to build products so exceptional that customers *become obsessed* with them, emphasizing the importance of output metrics over internal efforts. Matan recounted Factory's challenging early years, affectionately termed a "journey in the desert," starting in April 2023 when autonomous agents were a nascent concept. Despite being ahead of its time, this period allowed Factory to refine its approach to building for developers in the enterprise. He candidly admitted that "being two or three years early is the same as being wrong," highlighting the initial struggle to gain traction with enterprises not yet ready for fully autonomous solutions. A pivotal moment for Factory involved a radical decision to refund all customers, despite generating nearly $2 million in revenue. Matan explained that the product, while selling, wasn't truly making developers happy. This difficult choice, driven by their core operating principle of "create obsessed customers," underscored their commitment to product quality and long-term trust over short-term revenue. This experience, he noted, forged deep resilience within the team, who chose to stay despite the financial setback. The market shifted significantly with developer mindset changes and the launch of Factory's Droid CLI in September 2025. This tool met developers where they were, providing state-of-the-art, model-agnostic performance. Matan observed that developer openness, influenced by figures like Andrej Karpathy, was as crucial as model advancements. He argued that a multi-model harness, like Factory's, is superior to tightly coupled model-harness co-design, as it avoids overfitting and benefits from exposure to diverse model behaviors, much like training an AI on the entire internet rather than personal data. Matan then detailed the evolution from "token maxing" to "cost rationalization" in AI adoption. Initial pushes for AI usage led to indiscriminate token consumption. Factory addresses this with its router, which dynamically routes tasks to the most appropriate model based on cost and performance, preventing the use of expensive frontier models for trivial queries (e.g., "what's the weather?"). He highlighted the increasing prominence of open models, noting that they now account for a significant double-digit percentage of Factory's internal token usage due to their comparable performance (to "frontier minus one" models), speed, and cost-effectiveness. Discussing the business model, Matan stated Factory is currently usage-based but envisions a future of outcome-based pricing, possibly through a competitive marketplace for tasks. He also shared his vision for "software factories," where organizations transform inefficient, human-centric processes into rigorous, automated assembly lines for software development. This involves codifying tribal knowledge and establishing clear feedback loops to measure outcomes, allowing businesses to quantitatively allocate resources (headcount vs. tokens) based on core competencies and desired results. Looking ahead, Matan predicted that within 12-24 months, 90% of AI tokens will be asynchronous, signaling a shift towards "dark factories" where autonomous agents continuously operate without constant human initiation. Despite acknowledging potential short-term turbulence and resource misallocation, he expressed deep optimism for the long term. He believes AI will free engineers to tackle a broader spectrum of global problems, enabling the creation of fantastic software for areas like government services and pharmaceutical research, ultimately leading to a "net good for the world."