CHINA SYSTEM · OPERATING LOGIC
Read the concept through its operating structure
Foundation, general and industry models differ by use
Foundation models provide broad capabilities; general models serve many tasks; industry models target specific knowledge, rules and workflows. A tool-wrapped model is not automatically a validated industry model.
Pretraining, tuning and retrieval are different investments
Pretraining builds capability, tuning changes behavior and retrieval injects external knowledge. Prompts, parameter-efficient tuning and full tuning differ in cost and risk.
Task evaluation matters more than leaderboards
Public benchmarks can suffer contamination and task mismatch. Enterprises need evaluations covering answers, prohibited behavior, citations, permissions, tool actions and recovery.
China advances models with industrial deployment
China's AI+ manufacturing action targets 3–5 manufacturing general models, broad industry-model coverage, 1,000 industrial agents and 100 quality datasets by 2027. Targets are not results.
Korea should buy controllable outcomes, not model ownership alone
Korean firms should choose build, API, open weights or private deployment by sensitivity, latency, cost, lock-in and update duties, contracting data use, logs, changes, evaluation, incidents, IP and exit.