AI 时代的 “房谋杜断”：如何培养能力、品味与判断力
0x00 引言：当才华被机器追上之后
本文的出发点有两个：
- 房谋杜断是中国的一个经典成语。很多人都只是听过这个成语，却缺少实际应用的场景。很遗憾，AI 的到来，让这个成语有了最深刻的现实写照。
- 《我不得不把才华埋葬在昨天》是 DeepSeek 工程师刘胜与9月14日的文章，感慨AI快速进化将让手写算子逐渐被取代，自己也将从算子设计转向用AI生产算子。9月15日，他补充说明称，文章主要是向过去的手写算子时光告别。
本文试图回答：
- AI 时代究竟需要培养哪些能力、品味和判断力；
- “房谋杜断” 这个成语能否帮助我们理解人机协作；
- AI 是否可以被看作 “房谋”，人类是否应该成为 “杜断”；
- 杜如晦的判断力是如何形成的，今天可以借鉴什么；
- 普通人如何训练自己，而不是只等待 AI 替自己成长。
注意：AI 发展太快，也许本文的观点在发表时刻，已经被 AI 证伪。
0x01 房谋杜断
1.1 成语出处
《旧唐书·房玄龄杜如晦传论》：“世传太宗尝与文昭图事，则曰：‘非如晦莫能筹之。’及如晦至焉，竟从龄之策也。盖房知杜之能断大事，杜知房之善建嘉谋。” 后据此，总结出成语房谋杜断。
1.2 成语典故
房乔，字玄龄。杜如晦，字克明。二人是唐朝时期公认的“贤相”。“房谋杜断”更是历史上的一段佳话。
当时，唐高祖李渊封李世民为“天策上将”，允许他建天策府，广纳贤士。李世民便设置了“文学馆”，邀房玄龄、杜如晦等十八人至馆中讨论文籍。同时，还令画家阎立本为这十八人画像，令文学馆学士褚亮为他们写赞语，称号“十八学士”。就连当时的太子李建成都曾说过：“秦王府中，让我最为忌惮的就是杜如晦与房玄龄二人。”
贞观三年二月，太宗任命房玄龄为左仆射，杜如晦为右仆射，共同参与朝政大事。朝廷中，宫廷楼阁的规格和礼法规章等，都由他二人亲自制定，这些制定好的礼法也为贞观之治奠定了基础。
每次，太宗与房玄龄谋划大事时，房玄龄总是会说：“非如晦不能决断。”等杜如晦到时，原本二人犹豫不决的事，便立即有了结论。虽然依旧采用的是房玄龄的谋划，但也离不开杜如晦最后的审度。
房、杜二人私下的关系也十分亲密。在李世民还是秦王之时，房玄龄就曾力荐杜如晦：“杜如晦博学强识，是难得的人才。假如秦王想要得此天下，绝对是少不了这个人的辅佐的。”正是由于房玄龄的引荐，杜如晦才受到李世民的重视。二人相互配合，齐心协力，共同辅佐太宗。
0x03 我不得不把才华埋葬在昨天
《我不得不把才华埋葬在昨天》写出了一个很多技术工作者正在面对、却不容易说清楚的处境：
AI 先替代的，往往不是一个人的全部价值，而是他最熟练、最热爱、最能获得成就感的那部分工作。
”
作者长期从事算子设计、编写和优化。随着 AI 能够阅读 CUDA、PTX 和 SASS，分析指令停顿，提出优化方案并完成实现，原本需要专家多年积累的技能开始变成 AI 可以快速复制和扩展的能力。人的饭碗未必立即消失，但工作内容可能发生迁移：从亲手写算子，转向使用 Agent 更快地产出、评估和管理算子。 这带来的损失不只是就业风险，还有一种更隐蔽的失落：人仍然能够留在行业里，却可能失去原来热爱的工作方式。 类似的变化也会发生在数学、软件工程、产品设计、科研和管理中。AI 会越来越擅长：
- 搜索资料；
- 生成方案；
- 编写代码；
- 运行实验；
- 比较多个候选；
- 发现局部错误；
- 把想法快速变成可见结果。
当 “把想法做出来” 变得便宜，真正稀缺的能力就会逐渐转向：
- 什么问题值得解决；
- 什么方案符合真实目标；
- 哪些结果值得相信；
- 哪些风险不能接受；
- 哪些工作应该自动化；
- 哪些能力必须由人保留；
- 在不确定中如何做出承担责任的决定。
这正是 “能力、品味和判断力” 需要重新讨论的原因。
0x04 AI 带来的变化：从 “亲自完成” 转向 “组织完成”
4.1 过去的专业能力，包含大量执行工作
在传统工作中，一个专家通常需要亲自完成完整链条：
理解问题 → 查资料 → 设计方案 → 写代码或证明 →调试 / 测量 → 修改 → 交付 → 复盘
长期重复这条链条，会同时培养技术能力和判断力。一个优秀的算子工程师不仅会写代码，还会知道：
- 哪个瓶颈值得优化；
- 哪种性能提升只是测试噪音；
- 哪种设计会破坏后续扩展；
- 哪种实现虽然快，却会制造维护灾难；
- 什么结果值得继续投入。
AI 首先接管的通常是执行环节。它可以更快地生成代码、尝试方案和运行实验，但它并不会自动替人类决定问题价值，也不会自动承担错误后果。
4.2. 执行能力下降，可能反过来损害判断力
如果学生和初级工程师一开始就把所有 Lab、代码和实验交给 AI，他们可能得到更高的短期产出，却失去形成内部模型的机会：
- 不知道一个系统为什么这样组织；
- 不理解构建、依赖和部署；
- 无法预判未来需求；
- 看不出抽象边界是否合理；
- 不能判断一段生成代码何时会变成技术债；
- 只能在错误出现后被动求助 AI。
这不是说每个人都必须手写所有底层代码。真正的问题是：一个人是否理解自己正在让 AI 做什么，以及是否有能力审查结果。 可以被自动化的是重复劳动；不能轻易外包的是形成模型、识别风险和作出取舍的过程。
4.3. “转业” 不一定意味着离开原领域
文章中有一个重要判断：作者不一定会失业，但可能必须转业，从亲自写算子变成 Agent 的 “机甲驾驶员”。 这不是简单地从专家变成工具使用者。理想的转业应该是：
亲自执行一个局部环节
↓
理解整个系统和目标
↓
指挥AI完成多个方案
↓
评估结果、约束风险、选择路径
↓
对最终系统负责
4.4 尖锐的问题
如果只是把人从键盘前移到聊天窗口，能力并没有升级。
比如：一位匿名工程师在 X 平台上吐槽说：“Claude Code 简直榨干了我的灵魂。 ”，从产品规格、PRD、测试，到工单、代码、Bug 修复和报告，几乎所有东西都由 Claude Code 生成，而工程师们每天要工作 12～13 个小时，只为了“按回车” 。
“没有任何成就感，也没人真正解决 Bug。说到底，现在已经没人在思考了。”
”
他提出了一个简单却尖锐的问题： “如果一切都自动化了，那还剩下什么？”
也有一个网友认为：当 AI 掌握了这个世界上几乎所有的知识后，我们还需要那么多‘知识工作者’吗？未来，只需要按几个按钮就能完成的工作会越来越多。 ”
0x05 AI 时代需要的核心能力
真正的升级，是从 “完成任务的人” 变成 “定义任务、设计反馈、管理不确定性的人”。
我们来看看，AI 时代需要哪些核心能力。
5.1. 建模力：看见系统，而不是只看见任务
建模力是把复杂现实压缩成可理解结构的能力。它包括：
- 找出关键变量；
- 区分目标、约束和指标；
- 判断哪些信息重要；
- 理解模块之间的依赖；
- 预判局部改动对整体的影响。
没有建模力的人，会把 AI 当成万能搜索框；有建模力的人，会给 AI 一个清晰的系统边界和反馈标准。 以算子优化为例，问题不只是 “把这个 kernel 写快一点”，而是：
- 训练还是推理瓶颈；
- 延迟、吞吐、显存和功耗哪个优先；
- 优化是否适用于不同输入规模；
- 是否会破坏数值稳定性；
- 这项优化能否进入整个系统；
- 结果能否被后续模型和硬件复用。
5.2. 提问力：把模糊愿望变成可验证问题
高质量问题通常包含五个部分：
- 当前现象是什么；
- 真正想改善的目标是什么；
- 已知约束是什么；
- 成功标准是什么；
- 哪些范围暂时不讨论。
“帮我把它做得更好” 不是好问题；“在精度下降不超过 0.1% 的前提下，把 P99 延迟降低 30%，并在三种输入规模上复现” 才接近可执行问题。 AI 越强，提问越重要。因为 AI 能快速执行错误任务，也能把一个错误目标优化到极致。
5.3. 验证力：不把流畅输出当成可靠结论
AI 生成的内容越像专家，人越容易放弃检查。验证力要求人能够：
- 检查来源；
- 重建关键推理；
- 构造反例；
- 设计最小实验；
- 区分相关性与因果性；
- 判断结论是否超出证据；
- 要求结果可复现；
- 明确哪些部分仍然不确定。
在数学中，这意味着不能把形式化通过等同于理解；在工程中，不能把测试通过等同于系统可靠；在管理中，不能把报表达标等同于目标实现。
5.4. 取舍力：知道什么应该做，什么应该放弃
AI 可以提出很多方案，却不会自动决定哪个方案值得做。取舍力包括：
- 比较短期收益和长期代价；
- 判断复杂性是否值得；
- 识别不可逆风险；
- 保留未来选择空间；
- 接受某些目标无法同时最大化；
- 在信息不完整时作出暂时决定。
取舍力不是保守，也不是追求最优解，而是在真实约束下选择足够好的路径。
5.5. 责任力：愿意为决策承担后果
AI 可以生成建议，但不能替人承担事故、损失、信誉和伦理责任。真正的决策者必须回答：
- 如果判断错了，谁承担后果；
- 是否给系统设置了退出机制；
- 是否让相关人知道不确定性；
- 是否保留了人工复核点；
- 是否为了短期指标掩盖了长期风险。
没有责任力的人即使拥有很强的 AI，也只能制造更大规模的错误。
注：这点我实际上是存疑的。
因为如果 AI 发展很快，资本可能不会在乎是否有人负责任。它只在乎收入。
”
0x06 什么是 AI 时代的品味
能力让人 “做得到”，品味让人知道 “什么值得做”。
6.1. 品味不是审美偏好，而是价值排序
品味表现在以下判断中：
- 能否区分真正的创新与指标优化；
- 能否识别一个问题是否值得投入；
- 能否发现复杂方案其实在掩盖简单目标；
- 能否欣赏简洁、必要性和可解释性；
- 能否看到一个局部结果的长期潜力；
- 能否拒绝 “看起来很厉害但没有实际价值” 的工作。
6.2. 品味来自长期比较
品味来自不断比较：
- 为什么这个定义比另一个定义自然；
- 为什么这个系统边界更容易维护；
- 为什么某个证明值得写进教材；
- 为什么某个产品解决了真实问题；
- 为什么某个性能优化值得进入主干；
- 为什么另一个更高的指标反而不值得追求。
比较的对象越多，判断越细；亲自经历过失败，越能识别 “短期成功” 的陷阱。
6.3. 品味必须接受现实约束
品味不是拒绝效率，也不是迷恋手工。AI 时代的品味应当同时看到两件事：
- 有些手工过程确实应该被自动化；
- 有些看似低效的过程其实在培养理解和判断。
最好的问题不是 “要不要用 AI”，而是：哪些环节由 AI 完成能释放人的注意力，哪些环节如果完全交给 AI 会损害人的成长？
0x07 “房谋杜断” 能否帮助理解 AI 与人类？
7.1. 成语本义
“房谋杜断” 通常用来概括唐太宗时期房玄龄和杜如晦的互补：
- 房玄龄长于谋划；
- 杜如晦长于决断。
它不是说一个人只负责猜想、另一个人只负责签字，而是说复杂工作/决策需要不同能力：
房玄龄：搜集信息、提出方案、权衡可能性
杜如晦：识别关键、判断取舍、推动决定
”
二人真正的价值来自配合。没有谋划，决断会变成武断；没有决断，谋划会变成无休止的讨论。
7.2. AI 可以像 “房谋”，但不能简单等同于房玄龄
把 AI 比作 “房谋” 有启发性，因为 AI 擅长：
- 快速搜集资料；
- 并行提出多个方案；
- 模拟不同结果；
- 发现被忽略的组合；
- 总结已有经验；
- 生成行动计划；
- 对方案进行初步比较。
但 AI 与房玄龄仍有本质区别：
- AI 没有稳定的政治、组织和伦理责任；
- AI 不一定理解方案背后的真实利益关系；
- AI 可能把错误前提优化得很漂亮；
- AI 可以提出大量意见，却不必承担后果；
- AI 的 “谋” 依赖数据、提示和上下文质量。
因此，更准确的说法是：
AI 可以提供 “房谋式功能”，但不能自动成为真正的房玄龄。
”
7.3. 人类可以像 “杜断”，但不能只做最后签字的人
把人类比作 “杜断” 也有启发性，因为人类必须负责：
- 定义最终目标；
- 识别不可接受的风险；
- 判断不同价值之间如何排序；
- 决定何时停止搜索；
- 承担决策后果；
- 解释为什么选择这条路。
但人类不能把 “杜断” 理解成坐在 AI 输出后面按确认键。没有基础、上下文和验证能力的人，无法有效判断 AI 的谋划。 真正的 “杜断” 不是拥有最后决定权，而是拥有配得上最后决定权的能力。
7.4. 更准确的人机分工
在 AI 时代，可以把 “房谋杜断” 扩展成四层：
AI: 广泛搜索、生成方案、模拟后果
人类: 理解目标、识别约束、校验事实
团队: 讨论价值、暴露盲点、形成共识
负责人: 在不确定中决断并承担后果
这比 “AI 负责谋、人类负责断” 更完整。因为高质量决策不是两个人的机械分工，而是一个持续反馈系统。
0x08 杜如晦是如何成长的
8.1. 他不是天生的 “拍板机器”
历史叙事常用一句 “杜断” 概括杜如晦，容易让人误以为他天生具有神奇决断力。实际上，决断能力通常建立在长期积累之上：
- 对制度和现实的理解；
- 对人和组织的观察；
- 对历史经验的学习；
- 与高质量合作者长期讨论；
- 在复杂局势中区分主次；
- 在关键时刻接受责任。
“断” 不是凭直觉砍掉选项，而是把大量背景知识压缩成一个可以执行的判断。
你想和 AI 对话，想控制 AI，就得具备和 AI 接近的背景知识。
8.2. 早期识见：从细节看人的潜力
关于杜如晦早年经历，史书常强调他具有识见，能够在与人交往和任务过程中显露出非凡的判断潜力。唐初重要人物对他的评价，也不是因为他只会表达意见，而是因为他能从复杂局势中看出关键。 这对今天的借鉴是：判断力不是等到担任负责人之后才开始训练。它体现在很小的事情里：
- 能否抓住讨论的核心；
- 能否区分事实和情绪；
- 能否指出一个方案的隐含前提；
- 能否预见执行中的困难；
- 能否在意见不同的人之间建立共同语言。
8.3. 与房玄龄长期合作：在互补中形成决断
“房谋杜断” 最重要的，不是把两人的优点拆开，而是强调长期合作如何塑造判断力。 房玄龄擅长提出谋略和组织方案，杜如晦则善于在多个方案中辨别轻重、作出取舍。杜如晦的决断并不是脱离谋划的个人灵感，而是在充分听取、理解和比较之后形成的。 这说明 “杜断” 需要三个前提：
- 有足够多的高质量方案可供比较；
- 有能力理解方案的真实差异；
- 有勇气在信息不完备时作出选择。
AI 可以显著增加第一项，但第二、第三项仍需人类训练。
8.4. 在复杂政治与制度环境中形成现实感
杜如晦所处的，不是抽象考试，而是政权建立、军事行动、组织调整和利益协调交织的环境。现实中的决策往往没有完美方案，选择一个方案就意味着放弃另一些可能性。 因此，杜如晦式判断包含一种现实感：
- 不迷信单一原则；
- 不把局部正确当作整体正确；
- 知道时机和顺序同样重要；
- 能把长远目标拆成当下可执行步骤；
- 了解人的能力、利益和情绪；
- 能在理想与可行之间找到落点。
这与 AI 时代的工作高度相似。AI 可以告诉你什么方案在理想条件下最好，但人类需要判断真实团队、真实预算和真实时间下哪个方案最值得做。
8.5. “杜断” 的代价：决断意味着放弃
决断不是获得所有好处，而是接受某些损失。杜如晦式的判断力还包括：
- 知道什么不能同时拥有；
- 接受某些正确想法暂时不能实施；
- 为了整体目标牺牲局部漂亮；
- 在机会窗口出现时停止犹豫；
- 在结果不佳时承认判断需要修正。
没有放弃，就没有真正的决断；没有承担，就没有真正的领导。
0x09 如何形成 “杜断”
我们来推导下，如何形成 “杜断”。
注：毕竟我不是杜如晦的分身，所以只能推测。
9.1. 先形成自己的问题，而不是等待 AI 给意见
杜断的起点不是拍板，而是知道自己在决定什么。每次调用 AI 前，先写下：
- 我真正要解决的问题；
- 我为什么认为它重要；
- 成功和失败分别是什么；
- 哪些条件不能牺牲；
- 我最担心哪种错误。
如果连问题都没有，人只能在 AI 生成的选项中被动选择。
9.2. 建立 “谋–验–断–复盘” 循环
一个可执行的决策循环是：
谋: 让AI和人类共同提出候选方案
验: 用来源、实验、反例和小规模试运行检查
断: 明确选择、放弃和暂缓的方案
复盘: 记录结果、偏差、意外和下次规则
关键在于 “断” 之后仍然要复盘。没有复盘的决断只能积累自信，不能积累判断力。
9.3. 学会设定停止条件
AI 可以不断搜索、改写和比较。如果没有停止条件，决策会陷入 “再让模型想十个方案” 的循环。 停止条件可以是：
- 关键风险已经被验证；
- 新方案的边际收益低于成本；
- 已有方案满足最低成功标准；
- 决策窗口即将关闭；
- 继续搜索只是在逃避承担责任。
杜断不是找到绝对完美的答案，而是在足够理解风险后及时选择。
9.4. 训练负面判断
很多人只训练 “如何找到好方案”，却没有训练 “如何否决坏方案”。杜断尤其需要负面判断：
- 哪个方案不能做；
- 哪个指标不值得优化；
- 哪个需求不是真需求；
- 哪个证据不足以支持结论；
- 一个局部成功会制造系统性风险。
在 AI 时代，拒绝错误方向的能力可能比生成更多方向更重要。
9.5. 保留亲手做过的基本功
不是为了拒绝 AI，而是为了保留审查能力。工程师仍应亲自理解：
- 数据结构和系统边界；
- 构建、部署和调试；
- 性能、可靠性和安全；
- 代码组织与抽象；
- 一个小型系统从需求到交付的完整过程。
比如，数学学习者仍应亲自经历：
- 定义、例子和反例；
- 猜想、证明和证伪；
- 失败路径和重新建模；
- 把复杂证明讲清楚。
亲手做过不代表永远亲手做，而是让自己知道 AI 做得对不对。
9.6 具备必要的知识背景 & 体系
AI可以帮助解释概念，却不能替代学习者形成内部结构。没有基础的人可能得到一个看似漂亮的答案，却不知道答案使用了什么假设，也不知道哪些地方值得怀疑。
9.6.1 有知识，才有专业问题
提问能力不是脱离知识独立存在的技巧。同一个现象，不同专业背景的人会看到不同的问题。熟悉健身的人会追问动作形式、训练负荷和恢复周期；熟悉软件工程的人会追问系统架构、数据模型、成本和可靠性；熟悉产品的人会追问用户场景、限制条件、商业模式和反馈机制。数学也是如此。没有基础时，学习者只能问“答案是多少”；掌握定义、例子和反例后，才可能继续问
9.6.2 有上下文，才有高深问题
很多高价值问题并不需要深奥的专业知识，却需要长期合作形成的上下文。一个字段是否语义重合、一个功能是否只是临时承诺、一个“这个模块”具体指什么，往往无法从单句指令中判断，而要依赖团队共同记忆。
优秀的AI使用者不只是列出所有可能的问题，更能根据已有共识，用最短的话指出真正的分歧。未来的AI助手也许能代表个人保存权限、偏好、历史决策和项目背景，但这不能取消人类建立共同语言的责任。上下文越清楚，问题越短；问题越短，AI越可能把资源用在真正困难的部分。
9.6.3 以数学为例
以数学为例，需要长期积累的基础包括：
- 代数、几何、概率、分析、离散数学和逻辑；
- 编程、数值实验、数据处理；
- 阅读论文、理解定义和追踪引用；
- 检查证明、识别反例和发现隐含假设；
- 用清楚的语言解释“为什么”；
- 与他人合作、接受批评和持续修改。
0x10 不要只做更快的执行者
《我不得不把才华埋葬在昨天》最令人不安的地方，不是 AI 会写出更好的算子，而是一个人可能被迫离开自己最热爱的工作方式。对此没有廉价的乐观答案。技术进步确实会夺走一部分熟悉的劳动，也会改变很多人的职业身份。 但 “转业” 不必等于能力归零。一个真正理解底层硬件、模型需求、性能瓶颈和工程约束的人，即使不再亲手写每一个算子，仍然可以成为更强的系统设计者、Agent 指挥者和技术决策者。 这要求人从 “我能不能亲自完成” 逐渐转向：
- 我是否理解问题；
- 我是否能提出正确问题；
- 我是否知道什么结果值得相信；
- 我是否看见了长期代价；
- 我是否能在多个方案中作出取舍；
- 我是否愿意为决定承担后果。
“房谋杜断” 给出的不是一个简单分工，而是一种成熟协作的结构：广泛谋划与明确决断相互依赖，方案生成与责任承担不能混为一谈。 AI 可以帮助我们拥有更多 “房谋”，但如果人类不培养自己的 “杜断”，就会出现一种危险局面：选项越来越多，执行越来越快，真正的目标却越来越模糊。 最终，人类需要保留的，不只是某一项手工技能，而是能够理解世界、判断价值、形成品味并在不确定中承担选择的能力。
最好的模式不是 “AI 谋、人断”，而是共同升级。更好的协作方式是：
- AI 帮助人类扩大视野；
- 人类帮助 AI 明确目标和边界；
- 团队帮助双方暴露盲点；
- 最终负责人在可解释、可复盘的基础上决策。
人类的价值不是永远比 AI 知道更多，而是能够决定什么值得知道、什么必须验证、什么值得承担。AI时代真正值得追求的，不是让机器替人类完成越来越多的任务，也不是把一份尚待复核的公开声明包装成最终胜利，而是让人类借助机器提出更好的问题、发现更深的结构，并把这些结构转化为后来者能够学习和继续发展的知识。
TransFormer-封面
0xFF 参考
近800万人围观，连马斯克都惊了！工程师怒吐槽Claude Code“榨干灵魂”：每天工作12小时，只为不停“按回车”
本文使用 markdown.com.cn 排版
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Age “Housebreaking”: how to develop abilities, tastes and judgement
0x00 Introduction: When talent is pursued by a machine
There are two starting points for this paper:
- It's a Chinese classic. Many people have only heard the words, but there is a lack of practical applications. Unfortunately, the arrival of AI gave this phrase the most profound picture of reality.
- "I had to bury my talent yesterday" is an article by DeepSeek Engineer Liu Sheng and September 14th, which enviouss the rapid evolution of AI that will gradually replace hand-written algorithms, and that it will shift from their design to their production. On 15 September, he added that the article was mainly a farewell to past hand-written reckoning times.
This article attempts to answer:
- What competencies, tastes and judgement are needed in the era;
- How does the phrase "housebreaking" help us understand human collaboration?
- AI can be regarded as a “house plot” and whether humans should be “during”;
- How did Duval's judgement come into being and what could be learned today;
- How ordinary people train themselves instead of just waiting for AI to grow up for themselves.
Note: AI is developing too fast, and perhaps this paper is being published at a time when the AI has been perjured.
0x01 Room breaking up.
1.1 Origins
The old Tang Book House is of the same age as the old legend: “It's not as good as it is to be able to get it.” 'and if it's too bad, it's too old. The building of the house will be an end to greatness, and the construction of the house will be good.” On this basis, it was concluded that the language room was in an attempt to break up.
1.2 Spoken
Joe, it's age. Doo-ho, Duk-min. The two men were recognized as “good minds” during the Tang Dynasty. This is even better in history.
At that time, Tang Goo Zou Li Yon Yi Shimin was a "General of Heaven" and he was allowed to build the temple of Tianxi, a great man of Kanna. Li Shi Min has set up a “Lictology Museum” to discuss the literature in 18 people, such as Shigeki and Du Xiaoqiao. At the same time, the painters were made to paint a portrait of the 18 persons, and the Bachelor of Arts was given the honour of writing them a compliment entitled “Senior 18”. Even the then Prince Lee was built to say, "The one thing I fear most from in Qin's palace is the two men of his age."
In February, three years later, the Emperor appointed the house to be of left age and Duo to be of right age and to participate in the events of the DPRK. In the court, the specifications of the palace palaces and the rules of protocol were drawn up by both of them, which also laid the foundation for the healing of chastity.
Every time he plans a big project with the age of the house, the age of the house always says, “If it's not as good as it is, it can't be decided.” When it's too bad, the two of you were hesitant to decide. Although it is still the design of the house, it is not impossible to find the final measure.
They are also very close in private. When Li Se-min was the King of Qin, the age of the house was the one who was so eager to be true: "Dou has learned so well and is a rare man. If the Qin King wanted this world, he would not have lost his support.” It was precisely because of the age of the house that Doo-hyun received attention. Together, the two worked together to support the Emperor.
I had to bury my talent yesterday.
I had to bury my talent yesterday in a situation that many technicians were facing, but it was not easy to say clearly:
AI was the first to replace, often not a human being ' s full worth, but his most skilled, loving and successful part of the job.
”
Authors have long been involved in the design, preparation and optimization of algorithms. With AI being able to read CUDA, PTX and SAFS, analytical instructions stopped, optimised and completed, skills that would have required experts over the years began to become AI ' s ability to quickly replicate and expand. People ' s jobs may not disappear immediately, but their work may be displaced: Move from hand-written to Agent to faster output, assessment and management. The loss is not just an employment risk, but a more hidden loss: people are still able to stay in the industry and may lose their favorite ways of working. Similar changes can occur in mathematics, software engineering, product design, research and management. AI will get better and better at:
- Search for information;
- Generating programmes;
- Preparation of codes;
- Operation of experiments;
- Comparison of multiple candidates;
- Discovering local errors;
- Turn ideas quickly into visible results.
When "making ideas" becomes cheap, the really scarce capabilities turn to:
- What is worth resolving;
- What programmes meet real objectives;
- What results are credible;
- Which risks are unacceptable;
- What should be automated;
- What capabilities must be retained;
- How decisions to assume responsibility are made in uncertainty.
This is why “capacity, taste and judgement” needs to be revisited.
0x04 AI: from "personally done" to "organization completed"
4.1 Past professional competence, with significant implementation
In traditional work, an expert is usually required to personally complete the full chain:
Understanding the problem → Design scheme → Write codes or prove calibration / measurement → Change delivery → Duplicate
The long-term repetition of this chain will develop both technical capacity and judgement. A good arithmetic engineer not only writes code, but also knows:
- Which bottlenecks deserve optimization;
- What kind of performance upgrade is just a noise test;
- Which design undermines subsequent expansion;
- Which, though fast, would create a disaster of maintenance;
- What results are worth continuing.
AI usually takes over the implementation chain first. It can produce code, experiment and run experiments more quickly, but it does not automatically determine the value of the problem for humanity, nor does it automatically bear the wrong consequences.
4.2. Declining enforcement capacity, which may in turn undermine judgement
If students and junior engineers hand over all Labs, codes and experiments to AI from the outset, they may have higher short-term output but lose the opportunity to form internal models:
- I wonder why a system is organized like this;
- Lack of understanding of construction, dependence and deployment;
- Cannot prejudge future needs;
- It is not clear whether abstract boundaries are reasonable;
- It cannot be judged when a generation code becomes a technical debt;
- Passive help only after an error has occurred.
It's not that everyone has to hand-written all the lower code. The real question is whether a person understands what he or she is making AI do and whether he or she is capable of reviewing the results. What can be automated is duplication of effort; what cannot be easily outsourced is the process of modelling, risk identification and trade-offs.
4.3. “Recession” does not necessarily mean departure from the original domain
One of the important findings of the article is that the author is not necessarily unemployed, but may have to move on from writing his own calculations to becoming the “aeroplane pilot” of Agent. This is not simply a shift from an expert to a tool user. The ideal transition should be:
I'm going to do a local piece myself.
↓
Understanding the system and objectives
↓
Command AI to complete multiple programmes
↓
Assessment, containment of risks, choice of path
↓
Accountability for the final system
4.4 Sharp questions
Capability is not upgraded if you simply move a person from the keyboard to the chat window.
For example, an anonymous engineer spits on the X platform and says, "Claude Code is just draining my soul., from product specifications, PRDs, tests, to worksheets, codes, bug repair and reporting, almost everything is generated by Claude Code, and engineers work 12 to 13 hours a day, just to press back.
“No sense of achievement, no real solution to Bug. After all, no one's thinking now."
”
He asked a simple but sharp question: "If everything is automated, what is left?"
And one of the netizens thought, "Do we still need so many "knowledge workers" when AI has almost all the knowledge in the world?" In the future, more work will be done with a few buttons. It's not the same.
0x05 AI Core competencies for the era
The real upgrade is from “persons who do their job” to “persons who define tasks, design feedback, manage uncertainty”.
Let's see what core competencies the AI era needs.
5.1. Modelling power: to see the system rather than only the mission
Modelling power is the ability to compress complex realities into understandable structures. It includes:
- Identification of key variables;
- Distinguishing objectives, constraints and indicators;
- Determining what information is important;
- Understanding inter-module dependence;
- Prejudicing the impact of partial changes on the whole.
If you have no modelling power, you will use AI as a search box; if you have modelling power, you will give AI a clear system boundary and feedback standard. For example, the question is not just "write this Kernel faster," but:
- Training or bottlenecks in reasoning;
- What is the priority of delay, ingestion, VRAM and utility;
- Optimization applied to different input scales;
- Whether it would undermine numerical stability;
- Whether this optimization has access to the entire system;
- Whether the results can be reused by subsequent models and hardware.
5.2. Questioning: turning vague wishes into verifiable issues
Quality issues usually consist of five parts:
- What is the current phenomenon;
- What are the real goals for improvement;
- What is known as a constraint;
- What are the success criteria;
- Which are not discussed for the time being.
"Help me do it better" is not a good thing; "With less than 0.1% of precision, reduce P99 by 30% and repeat it on three input scales" is close to an enforceable problem. The stronger AI, the more important it is. Because AI can quickly perform a wrong mission and optimize a wrong target.
5.3. Certification: Flow output is not considered a reliable conclusion
The more the AI generated content is like an expert, the easier it is to give up the inspection. Certification requires that a person:
- Examination of sources;
- Rebuilding critical reasoning;
- Constructing reverse examples;
- Design of minimum experiments;
- Distinguishing relevance from causality;
- Whether the conclusion exceeds the evidence;
- The requirement that the results be repeated;
- Clarify which parts remain uncertain.
In mathematics, this means that formalization cannot be equated with understanding; in engineering, testing cannot be equated with system reliability; in management, statements cannot be equated with achievement of objectives.
Trade-offs: know what to do and what to give up
AI could come up with many options, but it would not automatically decide which ones were worth doing. The trade-offs include:
- Compare short-term gains and long-term costs;
- Assessing whether complexity is worth it;
- Identification of irreversible risks;
- Retain space for future choices;
- Acceptance of certain objectives cannot be maximized simultaneously;
- Making a provisional decision when information is incomplete.
The trade-off is not conservatism, nor the pursuit of the best solution, but rather the choice of a path that is good enough under real constraints.
5.5. Accountability: willingness to bear the consequences of decision-making
AI could generate recommendations, but could not assume responsibility for accidents, losses, reputations and ethics. True policymakers must answer:
- If wrong, who bears the consequences;
- Whether an exit mechanism has been put in place for the system;
- Whether the uncertainty is known to the person concerned;
- Whether manual review points are retained;
- Whether long-term risks are masked for short-term indicators.
Those without responsibility, even with a strong AI, can only make bigger mistakes.
Note: I actually doubt that.
Because if AI developed quickly, capital might not care about accountability. It only cares about income.
”
0x06, what's an AI-era taste?
The ability to do it, the taste to know what is worth doing.
6.1. Taste is not aesthetic preference, but value ranking
Taste is expressed in the following judgements:
- Whether real innovation can be distinguished from indicator optimization;
- Whether it is possible to identify whether an issue deserves input;
- Whether complex programmes can be found to cover up simple objectives;
- The ability to appreciate simplicity, necessity and interpretability;
- The long-term potential for a partial outcome;
- Is it possible to refuse a job that appears to be very powerful but of no practical value?
6.2. Taste from long-term comparisons
Tastes come from constant comparisons:
- Why is this definition more natural than another?
- Why is the system border easier to maintain;
- Why should a certificate be worth writing in?
- Why a product solves a real problem;
- Why a performance optimization is worth entering the backbone;
- Why is another higher indicator not worth pursuing?
The more objects they compare, the more they judge; the more they experience their own failures, the more they identify the “short-term success” trap.
6.3. The need for taste to be realistic
Taste is not a rejection of efficiency, nor is it a obsession with handcraft. AI's taste should see two things at once:
- Some manual processes should indeed be automated;
- Some seemingly inefficient processes are actually fostering understanding and judgement.
The best question is not, "Do you want to use AI," but, "Do you want to use AI?"
0x07 "House Breaking" helps understand AI and humanity?
7.1. Textual meanings
"The House of Discovery" is usually used to describe the age of the house and the complementarities of Du as well:
- The age of the house is older than the planning period;
- Dullo's longer than judgment.
It is not that one person is responsible only for conjecture, another for signing, but that complex work/decision-making requires different capabilities:
Age: gathering information, proposing programmes, weighing possibilities
Duly: identifying key, judging, driving decisions
”
The real value of both men comes from cooperation. Without a plan, the decision becomes arbitrary; without a decision, the plan becomes an endless discussion.
AI may be like a “house plot”, but it cannot simply be equated with the age of the house
It's enlightening to compare AI to the "House Planning" because AI is good at:
- Rapid collection of information;
- Multiple parallel proposals;
- Simulation of different results;
- The discovery of neglected combinations;
- Lessons learned;
- Generating action plans;
- Preliminary comparison of programmes.
However, AI is still fundamentally different from the age of the house:
- AI lacks stable political, organizational and ethical responsibilities;
- AI does not necessarily understand the real interests behind the programme;
- AI may have perfected the wrong premise;
- AI could make a large number of comments without having to bear the consequences;
- AI's scheme relies on data, tips and context quality.
Thus, the more precise statement is:
AI can provide a home-search function, but it cannot automatically become a real house-age.
”
7.3. Humans can act like "Duchon" but not just as final signatories. People
It's also instructive to compare human beings to "during" because human beings are responsible for:
- Define the ultimate objective;
- Identification of unacceptable risks;
- How to rank different values;
- Decide when to stop searching;
- To assume the consequences of decision-making;
- Explain why you chose this way.
But humans can't understand "dough" as sitting behind the AI output by pressing the confirmation key. Those who do not have a foundation, context and certification capability cannot effectively judge AI ' s designs. The real "dout" is not the final decision, but the ability to match it.
7.4. A more accurate division of labour
In the AI era, it was possible to expand the House Break to four levels:
AI: Extensive search, generation of programs, simulation of consequences
Human: Understanding the target, identifying constraints, verifying facts
Team: discuss values, expose blind spots, build consensus
Responsible: To decide in uncertainty and bear the consequences
This is more complete than "AI is responsible, man is responsible." Because high-quality decision-making is not a mechanical division of labour between two people, but a continuous feedback system.
0x08, how Duane grew up.
He's not a natural "board machine."
It's a common phrase in historical narratives, "Doo-Doo-Doo-Doo-Doo-Doo-Doo-Doo-Doo-Doo-Doo." In practice, decisive capacity is usually built on long-term accumulation:
- Understanding systems and realities;
- Observation of people and organizations;
- Learning from historical experience;
- Long-term discussions with high-quality collaborators;
- Distinction of priorities in complex situations;
- Accept responsibility at a critical time.
“Several” is not an intuitive cutting off of the option, but a condensed version of the vast body of background knowledge into an enforceable judgement.
If you want to talk to AI, if you want to control AI, you need to have background knowledge close to AI.
Early recognition: human potential in detail
With regard to Duval's early experience, Historic Books have often emphasized his insight and ability to reveal extraordinary judgmental potential in dealing with people and in the course of his mission. He was not judged by the leading figures, not because he only expressed his views, but because he was able to see the key from a complex situation. The lesson of today is that judgement does not wait until the person in charge to begin training. It's in the small things:
- The ability to capture the core of the discussion;
- Whether facts and emotions can be distinguished;
- Could an implicit premise of a programme be identified;
- Whether difficulties in implementation can be foreseen;
- The possibility of establishing a common language between people with different views.
8.3. Long-term cooperation with the housing age: decisions in complementarity
The most important thing is not to tear apart the merits of the two, but to emphasize how long-term cooperation shapes judgement. The age of the house is good at proposing strategies and organizing programmes, and so is the good at identifying and choosing options in multiple programmes. This decision is not a departure from the personal inspiration of the scheme, but rather a full hearing, understanding and comparison. This means that "Douth" requires three premises:
- A sufficient number of high-quality programmes for comparison;
- The ability to understand real differences in programmes;
- Have the courage to choose when information is incomplete.
AI can significantly increase the first, but the second and third are still in need of human training.
8.4. Building a sense of realism in a complex political and institutional environment
It is not an abstract examination, but an environment in which regimes are created, military operations, organizational adjustments and the harmonization of interests. Decision-making in reality is often incomplete, and choosing one means abandoning other possibilities. Thus, Du Psychic judgment contains a sense of reality:
- Not superstitious to the single principle;
- Not to treat local correctness as whole;
- Knowledge of timing and sequencing is equally important;
- Dismantling long-term objectives into immediate implementable steps;
- Understanding human capabilities, interests and emotions;
- to find a place between ideal and practical.
This is highly similar to the work of the AI era. AI can tell you which options are best under ideal conditions, but human beings need to judge which ones are most worthwhile under real teams, real budgets and real times.
8.5. The cost of “during”: adjudication means giving up
The determination is not to obtain all the benefits, but to accept certain losses. The following are among other things:
- Know what cannot be owned at the same time;
- Acceptance of certain correct ideas that cannot be implemented for the time being;
- The sacrifice of local beauty for the overall purpose;
- Stop hesitation when the window of opportunity appears;
- Recognition of the need for revision of judgements in cases of poor results.
Without giving up, there can be no real determination; without responsibility, there can be no real leadership.
0x09 How to Form "Doo-Doo-Doo"
Let's figure out how to make a "dough."
Note: After all, I'm not a double, so I can only speculate.
9.1. Develop their own problems rather than wait for comments from AI
The starting point is not a board, but a decision. Before each call to AI, write:
- I'm really going to solve the problem;
- Why do I think it matters?
- What is the difference between success and failure;
- What conditions cannot be sacrificed;
- What kind of mistake do I worry about most?
If there is no problem, you can only choose passively in the AI-generated option.
9.2. Creation of a `guilty-check-break-drive' cycle
An enforceable decision-making cycle is:
Get AI and humanity together to put forward a candidate.
Verification: Inspections using sources, experiments, contrivances and small-scale test operations
Break: Clear selection, waiver and suspension
Repeat: Record results, deviations, surprises and next rule
It's about breaking it and still having to reset it. No decision without a resuscitation can only accumulate confidence and judgement.
9.3. Learning to set conditions for cessation
AI can constantly search, rewrite and compare. If there are no conditions to stop, decision-making will fall into a cycle of "let the model think of ten scenarios." Conditions for cessation may be:
- Critical risks have been validated;
- The marginal gains of the new programme are below cost;
- Programmes already in place to meet minimum success criteria;
- The decision window is about to close;
- Keep searching just to avoid responsibility.
Duqqun did not find absolutely perfect answers, but chose them in a timely manner when the risks were sufficiently understood.
9.4. Training in negative judgement
Many people train only “how to find a good solution”, but not “how to reject bad solutions”. In particular, Duqun needs a negative judgment:
- Which option cannot be done;
- Which indicators are not worth optimizing;
- Which needs are not real;
- Which evidence is insufficient to support the conclusion;
- A partial success creates systemic risks.
In the AI era, the ability to reject the wrong direction may be more important than generating more direction.
9.5. Retention of basic work done by hands
Not to reject AI, but to retain the capacity to review. Engineers should still understand:
- Data structure and system boundaries;
- Construction, deployment and commissioning;
- Performance, reliability and safety;
- Code organization and abstraction;
- A complete process of a small system from demand to delivery.
Mathematicians, for example, should continue to experience:
- Definitions, examples and reverse examples;
- Conjecture, proof and perjury;
- Failed pathways and remodelling;
- Make it clear.
It doesn't mean you'll always do it, but let yourself know that AI did it right.
9.6 Availability of the necessary knowledge background & systems
AI can help interpret concepts, but it cannot be a substitute for learners to form internal structures. People without a basis may have a seemingly beautiful answer, but they do not know what assumptions it uses or where it is doubtful.
9.6.1 Knowledge for professional problems
The ability to ask questions is not the technique to exist independently of knowledge. In the same phenomenon, people from different professional backgrounds see different problems. Those who are familiar with fitness will ask about action forms, training loads and recovery cycles; those who are familiar with software engineering will ask about system architecture, data models, costs and reliability; and those who are familiar with products will ask about user scenes, constraints, business models and feedback mechanisms. The same goes for math. When there is no basis, the learner can only ask “what is the answer”; once a definition, example and a contrario are available, it is possible to continue.
9.6.2 context is the reason for the problem.
Many high-value issues did not require deep expertise, but rather long-term collaborative context. Whether a field is semantic, whether a function is a temporary commitment, and what a “module” refers to is often not judged by a single sentence directive, but relies on team memory.
Good AI users are not just listing all possible issues, but are better able to point out real differences in the shortest possible terms on the basis of existing consensus. Future AI assistants may be able to preserve permission, preferences, historical decision-making and project contexts on behalf of individuals, but this does not absolve humanity of its responsibility to create a common language. The clearer the context, the shorter the problem; the shorter the problem, the more likely AI is to devote its resources to a truly difficult part.
9.6.3 Take mathematics as an example
In mathematics, for example, the foundations that need to be built up over time include:
- Algebra, geometry, probability, analysis, discrete mathematics and logic;
- Programming, numerical experiments, data processing;
- Reading papers, understanding definitions and tracking references;
- Examination of certificates, identification of counterclaims and detection of implicit assumptions;
- Explain “why” in clear language;
- Cooperation with others, acceptance of criticism and continuous revision.
0x10, don't just be faster implementers.
I had to bury my talent in yesterday's most disturbing place, not that AI would write better calculations, but that a man might be forced out of his favorite way of working. There is no cheap optimistic answer to this. Technological advances do take away part of the familiar work and change the professional status of many. But “re-employment” does not have to be equal to zero capacity. A person who truly understands bottom hardware, model needs, performance bottlenecks and engineering constraints, even if he or she does not write every calculator, can still be a stronger system designer, Agent commander and technical decision-maker. This requires a gradual shift from "can I do it myself" to:
- Do I understand the question?
- Can I ask the right question?
- Do I know what to believe?
- Did I see the long-term cost;
- Can I make trade-offs in multiple programmes;
- I'm willing to take the consequences for the decision.
It is not a simple division of labour, but a structure of mature collaboration: extensive planning and clear determination are interdependent, and programme generation and responsibility cannot be confused. AI can help us to have more "houses," but if humans don't cultivate their "breaks," there's a dangerous situation: more options, faster implementation, less real targets. Ultimately, human beings need to retain not just a certain skill, but the ability to understand the world, to judge values, to develop tastes and to take choices in uncertainty.
The best model is not "AI", but a joint escalation. Better ways of collaborating are:
- AI to help expand human horizons;
- Human Help AI to clarify objectives and boundaries;
- Teams help both sides to expose blind spots;
- The final responsible person decides on an interpretable and revolving basis.
Human values do not always know more than AI, but can decide what is worth knowing, what must be proven, what is worth bearing. What is truly worth pursuing in the AI era is not to allow the machine to carry out more and more tasks for humanity, nor to package a public statement that is yet to be reviewed as a final victory, but rather to use the machine to raise better questions, discover deeper structures and translate them into knowledge that later people can learn and continue to develop.
TransFormer - Cover
0xFF Reference
Nearly 8 million people are watching, even Mask. Engineers are pissed off about Claude Code's "drilling the soul": working 12 hours a day just to keep pressing back.
This article is available in markdown.com.cn
Auto-fetched and translated by AI Atlas. All rights belong to the original author.