AI 的粮食,与祖传的管道 AI’s feedstock — and the ancestral pipes

模型层的叙事最响,但真正决定 AI 能否落地的,是它脚下的两块地基:数据——AI 的粮食,与遗留系统——AI 要接的老管道。标注的价值从体力众包暴力迁移到博士级判断力,大模型第一次让「读懂无人能懂的祖传代码」规模化可行——这门生意没有被 AI 消灭,而是被 AI 重新定价 The model layer makes the loudest story, but what decides whether AI lands is the ground beneath it: data — AI’s feedstock — and legacy systems — the old pipes AI must connect to. Labelling value migrated violently from crowdsourced muscle to doctoral judgment, and LLMs made «reading the ancestral code nobody understands» scalable for the first time — this business was not destroyed by AI; it was repriced by it

两个半球,两个最小单元:数据半球的「一次清洗与喂养」+ 遗留半球的「一段没人敢动的代码」——企业 AI 项目里二者串联:AI 要吃数据,也要接老系统取数、回写。主脊:双半球生命周期七环节(采集授权→清洗标注→合成与枯竭→治理确权→存储与检索→AI 考古→迁移现代化),②⑥标红;三条结构带:标注工厂的诚实层、POC 死亡谷解剖、遗留战场;判断层:数据质量与治理 ＞ 模型选型;用 AI 做「读懂」,把「敢改」留给懂业务的人。姊妹:标注劳动→gig,代码→code,财税入表→ledger,确权合规→liability,政务系统→gov。 Two hemispheres, two atoms: the data hemisphere’s «one wash-and-feed» plus the legacy hemisphere’s «one stretch of code nobody dares touch» — wired in series inside every enterprise AI project: AI must eat data, and must tap the old systems to fetch and write it back. The spine: a twin-hemisphere lifecycle in seven links (sourcing & licensing → labelling → synthesis & exhaustion → governance & rights → storage & retrieval → AI archaeology → migration), ② and ⑥ flagged; three bands: the labelling factory’s honesty layer, the POC valley of death dissected, the legacy battlefield; the judgment: data quality and governance ＞ model choice; let AI do the reading, keep the daring with people who know the business. Siblings: labelling labour → gig, code → code, accounting → ledger, rights → liability, government systems → gov.