The New Paradigm of Knowledge Production: From "Human Writing" to "System Writing"

"The expert writing in the study" is the typical picture of the old paradigm: decades of accumulation, exchanging for a small number of carefully polished works. Its preciousness lies in the close connection between experience, expression, and responsibility; its limitation is that research progress is serially constrained by the same person's time. Research, comparison, thinking, writing, and checking all compete for one attention. The changes brought by AI should be understood within this historical continuity: more cognitive tasks that originally had to be advanced step by step by humans can now be handed to systems that can call tools and adjust paths based on feedback.

The basic unit of system writing is a traceable research process: questions enter, evidence gathers, hypotheses compete, verification occurs, conclusions form, actions provide new feedback, and rules are revised accordingly. Articles are the expression of this process to readers, and also records for the next round of research to call upon. This system must have decision gates. Gates are rules that can stop the process from continuing: if key sources are missing, return to evidence gathering; if only speculation supports it, lower the conclusion's strength; if it involves promises beyond authorization, hand it to humans for judgment; if real-world results overturn the judgment, update the rules and retain the old version. When evidence is insufficient, the system should also produce "cannot yet judge."

The software paradigm first shifts: from mainly writing code, to writing more prompts, specifications, and decision rules. In the past, many steps had to be prescribed in advance; now models can choose paths within boundaries. The operations paradigm changes accordingly: from humans continuously on duty handling every anomaly, to AI recovering within authorized scope, with humans supervising the entire operation process. The organizational paradigm also shifts toward "one-person company + AI employees." Here, AI employees are execution roles with tasks, permissions, memory, and delivery standards. Retrieval, analysis, editing, and review can be parallelized, and the ultimately responsible person maintains consistency of goals and value prioritization.

The form of valuable knowledge changes accordingly. When oriented toward business and engineering decisions, executable knowledge often has more reuse value than merely readable knowledge. "Customers are willing to pay for reliability" is a readable judgment. Further explaining which types of customers it applies to, how reliability is measured, what evidence supports it, when to adjust pricing, and when to stop promising, it becomes knowledge that can influence action. The weight of assets thus shifts from document quantity to decision models. Here, decision models are a set of rules connecting context, evidence, options, thresholds, and results, not just model weights. They tell the system when to take what action, and also explain when to stop applying them.

Economic scarcity is increasingly concentrated on attention. Simon pointed out in 1971 that information richness causes attention scarcity. Today, AI brings this problem into the production end: founders can cheaply obtain more materials, but cannot synchronously increase their ability to seriously judge them. Imagine two processes: one generates ten reports daily, requiring three hours of human review; the other delivers only two conclusions daily, requiring twenty minutes of review, and equally supports key decisions. Output rewards the former, but the attention budget might choose the latter. OPC's competitive advantage lies in shorter decision paths in specific domains and less internal resistance when changing work methods.