University Is Not a Trade School
TL;DR: Eric Schmidt says universities should teach AI tools from day one. He’s confusing education with training. Universities exist to build judgment, not tool proficiency. That distinction matters more now, not less.
Eric Schmidt recently told an Abundance360 audience that universities should “teach AI tools in the first semester.” Restructure curricula around prompting, Copilot, LLMs.
Sounds practical. It’s a category error.
Training vs. Education
Training teaches you to use tools. Education teaches you which tools matter — and when to put them down.
Schmidt’s proposal is a four-year bootcamp optimized for the current stack. The problem: it’s a five-year bet on a technology that might be wrong. If the Transformer is a local maximum, students who spent four years mastering GPT-7 prompts are holding yesterday’s API docs.
This isn’t hypothetical. Universities that chased Java applets, Second Life, or blockchain curricula now have alumni with very specific, very useless skills.
Judgment Can’t Be Optimized
The irreplaceable function of a university is forming judgment — knowing which problem to solve, spotting the buried assumption, realizing your conclusion rests on a premise you never examined.
You can’t learn this from a tutorial or optimize it into a twelve-week sprint. It requires friction, boredom, and intellectual discomfort with no KPI.
Schmidt himself proved the point. In the same talk, he cited Jevons’ Paradox: AI efficiency gains will increase demand, “just like steam engines created more demand for coal.” Compelling — and wrong here. Jevons requires sufficient demand elasticity. AI capability up 100x but market needs only 10x = 90x waste, not opportunity. A real university education teaches you to catch exactly this kind of reasoning error.
Solving > Coding
Schmidt assumes coding is the skill and AI tools are the upgrade. But coding was never the point. Solving the problem was the point.
The last twenty years turned “learn to code” into a class-transition vehicle. Millions entered software for the salary, not the problem-solving. Result: a workforce bubble of tool operators — skilled at the how, disconnected from the why.
AI didn’t create this. It exposed it. When AI handles the how, people who only knew the how are naked. People who understood why? More valuable than ever.
Teaching AI tools first semester accelerates exactly the wrong thing — more tool operators, faster, right when tool operation is becoming worthless.
The Paradigm Trap
There’s a pattern worth naming: people who succeeded in one paradigm advising the next generation to optimize for it.
Schmidt’s career grew when engineering execution was the scarcest resource. In that world, “learn the tools” made sense. But we’re entering a world where execution is cheap and judgment is expensive. Teaching tools prepares students for yesterday’s bottleneck.
What Universities Should Do
Use AI the way you use calculators, not textbooks. AI is infrastructure, not curriculum.
Double down on: Problem formulation — which dataset matters and why, not “use AI to analyze this one.” Adversarial reasoning — find the assumption, break the argument, rebuild better. Cross-domain translation — the most valuable people see patterns specialists miss. Comfort with ambiguity — real problems don’t come with instructions.
One Line
AI makes tool skills cheap. Knowing which problem is worth solving was always expensive.
Universities exist to teach the expensive thing. Don’t turn them into bootcamps for the cheap thing.
一句话： Schmidt 说大学第一学期就该教 AI 工具。他搞混了教育和培训。大学是练判断力的地方，不是学工具的地方。
Schmidt 在 Abundance360 上说，大学应该”第一学期就教 AI 工具”，围绕 prompting、Copilot、LLM 重构课程。
听起来实用，但这是范畴错误。
培训 vs 教育
培训教你用工具。教育教你判断哪个工具重要——什么时候该放下工具。
Schmidt 的方案是四年制训练营，为当前技术栈优化。问题是：这是对一个可能走错方向的技术做五年优化。Transformer 如果是局部最优，花四年学 GPT-7 prompting 的学生拿的就是过期文档。
不是假设。追过 Java applets、Second Life、区块链课程的大学，现在有一批技能非常具体、也非常没用的校友。
判断力没法速成
大学不可替代的功能是形成判断力——知道该解决哪个问题，发现埋在脚注里的假设，意识到自己的结论建立在未经审视的前提上。
教程学不来，十二周冲刺优化不出来。需要摩擦、无聊、没有 KPI 的智力不适感。
Schmidt 自己就演示了这一点。同场演讲，他引用杰文斯悖论：AI 效率提升会增加需求，”像蒸汽机增加了煤的需求”。漂亮的类比——但在这里是错的。杰文斯悖论的前提是需求弹性足够大。AI 能力涨 100 倍但市场只要 10 倍 = 90 倍浪费，不是机会。真正的大学教育教你抓的就是这种推理漏洞。
Solving > Coding
Schmidt 假设编程是技能，AI 工具是升级。但编程从来不是目的，解决问题才是。
过去二十年”学编程”成了阶层跃迁通道，几百万人为薪水涌入，不是为了解决问题。结果：大量”工程师”本质是工具操作员，精通怎么做，脱离为什么做。
AI 没制造这个问题，它暴露了问题。AI 接管”怎么做”后，只会怎么做的人暴露了；理解为什么做的人比以前更值钱。
第一学期教 AI 工具，在工具操作即将一文不值的时候加速生产工具操作员。方向反了。
范式陷阱
一个值得点名的模式：在旧范式里成功的人，建议下一代为当前范式优化。
Schmidt 成长于工程执行力最稀缺的时代，”学好工具”当时合理。但我们正进入执行力廉价、判断力昂贵的世界。教工具是为昨天的瓶颈培养人才。
大学该怎么做？
AI 像计算器一样用，不像教材一样用。AI 是基础设施，不是课程。
加倍投入：问题定义——哪个数据集重要、为什么，不是”用 AI 分析这个”。对抗性推理——找假设、拆论证、重建更好的。跨领域翻译——最有价值的人看到专家看不到的模式。对模糊的耐受力——真实问题不带说明书。
一句话
AI 让工具技能变便宜了。知道哪个问题值得解决，一直都贵。
大学就是教这个贵东西的地方。别变成教便宜东西的训练营。