「AI构文」AI 时代不太可能呈现「广场」单向战胜「高塔」的周期，而更可能形成三层结构：底层高塔化、应用层广场化、入口层再高塔化。晶片、算力、能源、云端、资本与监管会集中；开源模型、低成本工具、个人创作者、微型企业与垂直社群会扩散；少数 AI assistant、agent platform、身份、支付与企业软体则会重新掌握分发权。到 2025 年，全球组织 AI 使用率已达 88%，但 agent 在具体业务职能中的部署仍多为个位数。2025 年逾九成重要 frontier models 由企业推出，International Energy Agency 预计全球资料中心用电到 2030 年可能接近 945 TWh，约为当前两倍。

AI 的主要影响不是立即消灭多数职业，而是把职业拆成任务、判断与责任。客服实验显示 AI 平均提高生产率约 14%，且对低经验员工帮助更大；顾问在能力边界内更快且质量更高，但边界外错误率上升；METR 对资深开源开发者的随机试验则显示，使用当时 AI 工具的组别平均多花约 19% 时间。企业可能更小，个人可获得近似小团队的研究、文案、程式、设计与营运能力，但模型、流量、支付、云端、身份与法规依赖会加深，所以生产单位变小不等于权力分散。

未来十年的稀缺性会从普通内容生成转向第一手资料、工作流嵌入、评估系统、身份信任、客户关系、法规资格与物理执行。Pew Research Center 观察到，Google 搜索页有 AI summary 时，传统结果点击率约 8%，无 summary 时约 15%，显示流量会由公共网站转向私人化代理环境。EU AI Act 多数条款将于 2026 年 8 月 2 日适用，监管会更深入模型训练、资料、算力、部署、风险评估与内容标识。个体的较佳策略是以 evaluation-first 建立 20—50 个代表性案例，采用最小权限 agent，保持模型可替换，控制真实资料、评估集、决策记录与直接客户关系。

The AI era is unlikely to show a one-way cycle in which the “square” defeats the “tower”; it is more likely to form a three-layer structure: towerization at the base, squareization at the application layer, and re-towerization at the entry layer. Chips, compute, energy, cloud, capital, and regulation will concentrate; open-source models, low-cost tools, individual creators, micro-firms, and vertical communities will proliferate; a few AI assistants, agent platforms, identity, payment, and enterprise software systems will regain distribution power. (Key numbers: 2025, 88, 2030, 945)

By 2025, global organizational AI adoption had reached 88%, but agent deployment in specific business functions was still mostly in single digits. (Key numbers: 14, 19)

Over the next decade, scarcity will shift from ordinary content generation toward first-party data, workflow embedding, evaluation systems, identity trust, customer relationships, regulatory credentials, and physical execution. Pew Research Center observed that when Google search pages include an AI summary, clicks on traditional results are about 8%, versus about 15% without a summary, showing traffic moving from public websites toward private agent-mediated environments. Most EU AI Act provisions will apply on August 2, 2026, and regulation will reach deeper into model training, data, compute, deployment, risk assessment, and content labeling. A better individual strategy is to build 20–50 representative cases through evaluation-first practice, use least-privilege agents, keep models replaceable, and control real data, evaluation sets, decision records, and direct customer relationships.