Factory 联合创始人兼 CEO Matan 讨论了产出指标相对于客户痴迷等输入指标的重要性，以及 Factory 如何通过模型独立性和模块化在企业 AI 市场中脱颖而出。
Matan, co-founder and CEO of Factory, discusses the importance of output metrics over input metrics like customer obsession, and how Factory differentiates in the enterprise AI market with model independence and modularity.
要点 · TL;DR
模型独立性可避免供应商锁定，并通过开源模型实现成本节约。 Model independence prevents vendor lock-in and enables cost savings via open models.
多模型框架优于单一模型框架，更适合企业 AI。 Multi-model harnesses outperform model-specific ones for enterprise AI.
未来 12-24 个月内，令牌使用将转向 90%异步模式。 Token usage will shift to 90% asynchronous within 12-24 months.
核心观点 · Key points
模型独立性对企业避免供应商锁定至关重要。 Model independence is crucial for enterprises to avoid vendor lock-in.
构建多模型工具链优于特定模型的工具链。 Building a multi-model harness outperforms model-specific harnesses.
代币使用将在12-24个月内从同步转向90%异步。 Token usage will shift from synchronous to 90% asynchronous within 12-24 months.
开源模型正接近前沿减一的性能，实现成本节约。 Open models are approaching frontier-minus-one performance, enabling cost savings.
企业AI采用需要行为改变和实验意愿。 Enterprise AI adoption requires behavioral change and willingness to experiment.
反共识 · Contrarian takes
早两三年和做错是一样的。 Being two to three years early is the same as being wrong.
客户痴迷是输入指标；输出是创造痴迷的客户。 Customer obsession is an input metric; output is creating obsessed customers.
模型与工具链协同设计不会带来更好性能；多模型工具链才行。 Model-harness co-design does not yield better performance; multi-model harnesses do.
当产品未准备好时退还客户收入能建立长期信任。 Giving back revenue to customers when product isn't ready builds long-term trust.
定价将在2030年代从基于使用量转向基于结果。 Pricing will shift from usage-based to outcome-based in the 2030s.
本期章节 · Chapters（共 16）
客户至上vs产出指标Customer obsession vs output metrics
工厂历程与模型独立Factory's journey and model independence
先行者与沙漠之旅Being early and the desert journey
让利维护信任Giving back revenue to maintain trust
团队逆境韧性Team resilience through tough times
构建智能体框架Building a great agent harness
Token最大化与成本优化Token maxing and cost rationalization
模型路由趋势与开源采用Model routing trends and open model adoption
任务细分与验证标准Subdividing tasks and validation criteria
软件工厂概念The concept of a software factory
Token经济学与分配Tokconomics and token allocation
Token与人力成本极限Token vs headcount spending in the limit
核心能力与AI应用Core competency and AI adoption
Token与原生智能体软件未来Future of tokens and agent-native software