Ethan Mollick's latest post makes a compelling case: AI capabilities far exceed what most people experience, and the bottleneck is how we interact with it.

The Interface Is the Bottleneck

Research shows that when financial professionals used GPT-4o for complex valuation tasks, the productivity gains were partially offset by the "cognitive tax" of the chatbot interface.

The problems?

• Walls of text: Answers buried in five paragraphs
• Unsolicited suggestions: Ask about A, get recommendations for B, C, and D
• Conversation entropy: Once a chat gets messy, it stays messy

The people hurt most? Less experienced workers—the very ones who could benefit most from AI, if only they could keep track of what they were doing.

Specialized Interfaces Are Emerging

The solution: task-specific AI interfaces.

Programming leads the way:

• Claude Code works autonomously for hours
• OpenAI Codex and Google Antigravity offer similar capabilities
• But these assume you know Python and Git

Other professions are catching up:

• Google Stitch: Describe an app in natural language, get multiple interconnected screens
• Google Pomelli: Paste your website URL, get on-brand social media campaigns
• NotebookLM: AI built specifically for research and note-taking

Implications for Educators

1. Stop making students "chat" with AI Chatboxes aren't learning tools—they're information black holes. Students need structured, goal-directed AI interactions.

2. Choose specialized tools Writing tools for writing. Design tools for design. Programming tools for code. The era of one-chatbox-fits-all is ending.

3. Teach interface literacy One of the most important skills for the future? Understanding how to design human-AI collaboration interfaces.

The Real Question

If your students aren't getting results with AI, is the AI not smart enough—or are they using the wrong tool?

Chances are, it's the latter.

Ethan Mollick在最新文章中提出了一个扎心的观点：AI的能力远超大多数人的认知，问题出在我们与AI的交互方式上。

界面即瓶颈

研究显示，当金融专业人士使用GPT-4o完成复杂估值任务时，虽然AI确实提升了效率，但聊天框界面带来的"认知税"几乎抵消了这些收益。

问题出在哪？

• 巨大的文字墙：AI动辄输出五大段，答案藏在里面
• 无关建议轰炸：你问A，AI顺便推荐B、C、D
• 对话失控：一旦聊乱了，双方都在互相镜像对方的混乱

最受伤的是经验较少的职场新人——正是最需要AI帮助的群体。

专用界面正在崛起

好消息是，专用AI界面正在改变游戏规则。

编程领域已经走在前面：

• Claude Code可以自主工作数小时
• OpenAI Codex、Google Antigravity类似
• 但这些工具对非程序员来说门槛太高

其他领域的探索：

• Google Stitch：用自然语言描述，自动生成多屏App设计
• Google Pomelli：粘贴网站URL，自动生成品牌社交媒体 campaign
• NotebookLM：专为研究和笔记设计的AI界面

对教育者的启示

1. 别再让学生"聊"AI了 聊天框不是学习工具，是信息黑洞。学生需要的是结构化、目标明确的AI交互。

2. 选择专用工具 写作用写作工具，设计用设计工具，编程用编程工具。一个聊天框打天下的时代正在结束。

3. 关注界面设计素养 未来最重要的技能之一，是理解如何设计人与AI的协作界面。

一个值得思考的问题

如果你的学生用AI没效果，是AI不够聪明，还是工具选错了？

答案很可能是后者。

Stop for a moment and ask yourself: if your child could look up any fact in 5 seconds, what would they still need to learn?

This isn't hypothetical. It's the world we're already living in.

GPT-5-level AI reads a decade of research in seconds. It writes analysis papers better than most graduate students. It speaks fluently in dozens of languages. Against this reality, the storage of knowledge is being fully outsourced to technology.

So what remains that's actually theirs?

What Is a Thinking Framework?

I've seen brilliant people completely stuck on problems that shouldn't be hard for them.

They had knowledge, information, data — but when faced with a genuinely novel, complex challenge, they spun in circles. Not because they weren't smart, but because they lacked a reusable mental scaffold.

A thinking framework is that scaffold.

In simple terms, it's a structured way of thinking: when you face a problem → what do you ask first? Then what? And finally, what?

A concrete example: Warren Buffett's partner Charlie Munger spent his life making investment decisions with what he called "latticework of mental models." He wasn't just a finance expert — he brought in psychology, engineering, economics, biology, and more. His edge wasn't any single piece of knowledge. It was his ability to cross-examine any problem through multiple frameworks simultaneously.

Why Frameworks Outvalue Knowledge in the AI Age?

Three reasons:

1. AI gives you answers; frameworks help you ask the right questions

AI is fundamentally an answer machine. You ask, it answers. But what do you ask?

Someone without a framework uses AI and gets generic, mediocre outputs. Someone with a framework knows exactly which angle to approach from. They get 10x more value from the same tool.

2. AI-generated content is exploding — frameworks help you filter and synthesize

ChatGPT writes one analysis today. Tomorrow, ten AI systems generate ten more. In a world of information overload, the rarest skill isn't information access — it's judgment. What's important? What's relevant? What deserves deeper attention?

Without frameworks, you're drowning. With frameworks, you're surgical.

3. Frameworks are the meta-skill AI cannot replicate

Knowledge can be outsourced. Skills can be learned by AI. But how you think is permanently yours.

When you hold 10 or 20 distinct thinking frameworks, your entire perception of the world shifts. You can examine one problem from multiple angles simultaneously. You can find purchase in deep uncertainty. You can break complex problems into actionable steps. None of this is replaceable by AI.

What Can Parents Do Right Now?

① Ask "What do you think?" more than "What's the answer?"

Instead of "What's the answer to this problem?", try "If we changed one condition, how would the answer change?" Train thinking, not recall.

② Teach "classify first, then solve"

When facing a complex problem, the first step isn't to start solving — it's to ask: What type of problem is this? Is it well-structured or open-ended? Does it require analysis or creativity?

Once classified, the solution path clarifies itself.

③ Deliberately expose children to diverse ways of thinking

Read across fields (not just what they already like), meet people from different backgrounds, learn the basic logic of different disciplines. Frameworks are built from accumulated见识 — not from drilling test papers.

④ Use AI, but after the child forms their own opinion first

Your child wants to look something up with AI? Fine. But first, have them write down their own viewpoint — even if it's just three sentences. Then bring in AI to supplement or challenge it. This way, AI becomes a thinking partner, not a thinking replacement.

The One-Line Takeaway

In the AI era, knowledge is a public good. Frameworks are private assets.

You can't stop your child from looking up any fact with AI. But you can give them something rarer, more valuable, and more irreplaceable than any AI: a mind that knows how to think.

That, not knowledge, is where education should pour its energy.

一个值得所有人停下来想一想的问题：如果你的孩子现在可以5秒钟内查到任何知识，那么他还需要"学"什么？

这不是假设。这是现实。

GPT-5级别的AI已经能在一秒内读完一个领域十年的论文，能写出比你当年毕业论文更好的分析报告，能用十几种语言流畅对话。在这种背景下，"知识的存储"这件事，正在被技术彻底外包。

那么，什么才是真正属于孩子的东西？

什么是"思考框架"？

我见过太多聪明人，困在错误的思维里出不来。

他们有知识、有信息、有数据——但当他们面对一个全新的复杂问题时，依然会像无头苍蝇一样乱撞。不是因为他们不聪明，而是因为他们没有一个可以反复使用的思维脚手架。

思考框架，就是这个脚手架。

简单说，它是一套结构化的思维方式：遇到问题 → 你先问什么？再问什么？最后问什么？

举个例子。巴菲特的老搭档查理·芒格，他一生都在用一套"多元思维框架"做投资决策。他不只是一个懂财务的人，他同时具备心理学、工程学、经济学、生物学等多个学科的思维模型。他的厉害之处不在于某个单一知识，而在于他同时调动多个框架交叉分析一个问题。

为什么AI时代，框架比知识更值钱？

三个原因：

1. AI给你的是答案，但框架帮你问对问题

AI再强，它本质上还是一个答案机器。你问，它答。但问题是——你问什么？

不会提问的人用AI，只能得到一堆平庸的泛泛而谈。真正会用AI的人，是那些脑子里有框架、知道从哪个角度切入的人。

2. AI生产的内容越来越多，框架帮你筛选和整合

今天ChatGPT写了一篇分析，明天又有新的AI生成了10份报告。信息爆炸的年代，最稀缺的不是信息，而是判断力——什么重要？什么相关？什么值得深挖？

没有框架的人，只会被信息淹没。有框架的人，才能快速找到那1%真正有价值的部分。

3. 框架是AI无法替代的"元能力"

知识可以被外包，技能可以被AI学习，但你怎么想这件事，永远是你自己的。

当你拥有10种、20种不同的思维框架，你看待世界的维度就会发生质变。你能从不同角度同时审视一个问题，能在不确定性中快速找到切入点，能把复杂问题拆解成可行动的步骤——这些，都不是AI能替你完成的。

家长可以怎么做？

① 少让孩子背答案，多问"你怎么想"

与其问孩子"这道题答案是多少"，不如问他"如果换一个条件，这道题会怎么变？"逼他思考，而不是记忆。

② 教孩子"先分类，再解决"

遇到复杂问题，第一步不是动手做，而是先问：这是个什么类型的问题？是结构清晰的？还是开放式的？是需要分析的？还是需要创造的？

分类清晰了，解决路径就清晰了。

③ 刻意接触不同的思维方式

读不同领域的书（不只是孩子喜欢的领域），接触不同背景的人，学习不同学科的基本逻辑。框架是见识的积累，不是刷题刷出来的。

④ 用AI，但先让孩子自己思考再问

孩子想用AI查资料？完全可以。但先让他自己写出自己的观点，哪怕只有三句话，再让AI来补充或挑战他。这样AI才是他的思考伙伴，而不是替代品。

一句话总结

AI时代，知识是公共品，框架是私人资产。

你无法阻止孩子用AI查到任何知识，但你可以帮他建立一套比任何AI都更稀缺、更值钱、更不可替代的思考框架。

这，才是教育真正的着力点。

A Silent Revolution Is Already Happening

In late March, a small security software company called StrongDM announced an experiment that should make every educator pause: they built a complete product with just 3 human engineers and a system of AI agents — no human wrote code, no human performed code review.

They called it the Software Factory.

The rules were radical:

• Rule 1: Code must not be written by humans
• Rule 2: Code must not be reviewed by humans

The product shipped to real customers.

This isn't science fiction. It's a real company, real results, happening right now. And its implications for education are closer than most people realize.

From "Learn to Code" to "Manage an AI Factory"

For thirty years, the logic of programming education has been simple: learn to code → find a job. But the StrongDM case reveals something uncomfortable — code itself is becoming the most automatable part of software work.

This doesn't mean "programming education is dead." It means something more profound:

What will matter isn't execution — it's direction.

The Software Factory works like this: humans set the product roadmap → AI agents autonomously code, test, and iterate → humans review the finished product.

Within this framework, the only irreplaceable human role is the one who decides what to build — the product designer and project manager combined.

What does this mean for education? It means we must shift from "teaching kids to write code" to "teaching kids to define problems, decompose tasks, and manage AI teams."

A Real Classroom Scenario

Imagine a middle school class given this project: "Use AI to build a tool that helps elderly community members book medical appointments."

Traditional model: students form groups, learn Python, write programs, submit code.

AI-era model: students form groups, describe requirements in natural language, assign tasks to different AI agents, monitor progress, review outputs, iterate and refine.

The latter is far harder than the former.

Because it requires students to develop:

• Problem-definition skills: knowing what problem to solve is more valuable than solving it
• Systems thinking: understanding how a product is composed of interconnected components
• Task decomposition: breaking complex goals into steps an AI can execute
• Critical evaluation: judging whether AI output is reasonable, rather than accepting it blindly

None of these skills come from rote memorization or test prep.

What Parents Can Do Now

You don't need to be a tech expert. But three things you can start today:

First, shift from "answer education" to "question education."

Stop asking "what did you learn today?" Instead ask: "what problem are you trying to figure out?" Train children to discover and define problems, not wait for them to be solved.

Second, give your child the role of "AI team manager."

When your child needs to complete a project — any project, even a presentation, a research report, or a creative piece — encourage them to break the task into parts and use AI tools for each sub-task. You act as the quality reviewer, challenging their work and helping them iterate.

Third, teach your child to say "that's wrong."

Learning to question AI conclusions is more valuable than accepting them. Ask your child: "Where do you think the AI might be wrong?"

Education Is Being Redefined

The StrongDM experiment is ultimately asking a question about human value: In a world where AI can execute everything, what remains for humans?

The answer is: the ability to define direction.

Future education shouldn't train excellent executors. It needs to raise children who can tell AI what to do.

Start turning your child from a "problem-solver" into an "AI conductor." That's the most important educational mission of our generation.

一场正在发生的静默革命

三月底，美国安全软件公司 StrongDM 宣布了一个实验结果：他们用 3 个工程师 + 一套 AI 代理系统，完成了一个通常需要 15 人团队才能构建的产品——整个过程，人类没有写一行代码，没有进行一次人工 code review。

这就是"软件工厂"（Software Factory）。

它的规则很激进：

• 规则一：代码必须不由人类编写
• 规则二：代码必须不由人类 review

听起来像天方夜谭？但产品已经交付给真实客户使用了。

这不是危言耸听，而是正在发生的现实。而它对教育的冲击，比大多数人想象的更近。

从"学习编程"到"管理AI工厂"

过去三十年，编程教育的逻辑从未变过：学会写代码 → 找到工作。但 StrongDM 的案例揭示了一个令人不安的事实——代码本身，正在成为最容易被 AI 接管的工作环节。

这意味着：仅靠"会写代码"的孩子，未来可能面对的不是就业竞争，而是一整个工种的消失。

但这并不等于"编程教育没用了"。相反，它提出了一个更深刻的问题：

未来真正有价值的，不是执行者，而是设计者。

软件工厂的核心运作方式是：人类制定产品路线图 → AI 代理自主编码、测试、迭代 → 人类验收成品。

在这个框架里，唯一不可替代的人类角色，是那个知道"要做什么"的人——也就是产品设计师和项目管理者的合体。

这对教育意味着什么？意味着我们必须从"教孩子写代码"，转向"教孩子定义问题、拆解任务、管理AI团队"。

一个真实的课堂场景

想象一个初中课堂：

老师给出一个项目："用AI开发一个帮助社区老人预约挂号的工具。"

传统模式下：学生分组，学习Python，编写程序，上交代码。

AI时代：学生分组，用自然语言描述需求，分配任务给不同的AI代理，监控进度，审查输出，迭代修改。

后者的难度，远远高于前者。

因为它要求学生具备：

• 问题定义能力：知道要解决什么问题，比解决问题更重要
• 系统思维能力：理解一个产品如何由多个组件构成
• 任务分解能力：把复杂目标拆解为AI可以执行的步骤
• 批判性评估能力：判断AI的输出是否合理，而不是盲目接受

这些能力，没有任何一项来自"刷题"或"背语法"。

父母现在能做什么？

作为家长，你不需要成为技术专家，但有三件事可以立刻开始：

第一，从"答案教育"转向"问题教育"。

不要再问孩子"今天学了什么"，而是问"今天你有什么问题想搞清楚？"让孩子习惯于发现问题、定义问题，而不是等待问题被解答。

第二，给孩子一个"AI团队负责人"的角色。

当孩子需要完成一个项目（哪怕是PPT、调查报告或手工制作），鼓励他们先把任务分解，然后尝试用AI工具完成各个子任务。父母做验收官，提出质疑，帮助孩子迭代。

第三，教孩子对AI输出说"不对"。

学会质疑AI的结论，比接受AI的结论更重要。问孩子："你觉得AI说的哪里可能有问题？"

教育正在被重新定义

StrongDM 的实验，本质上是在问一个关于人类价值的问题：在AI可以执行一切的世界里，人类还剩下什么？

答案是：定义方向的能力。

未来的教育，不应该再培养"优秀的执行者"。它需要培养的是：能够告诉AI"去做什么"的孩子。

从今天开始，把孩子从"做题机器"变成"AI指挥家"。这是我们这代人最重要的教育使命。

Imagine asking AI a question and receiving five paragraphs with the answer buried somewhere inside, plus three suggestions for topics you never asked about. The conversation gets messier, the AI gets more "helpful," and you get more confused. This isn't about AI being unintelligent—it's about interface design failing us.

The Cognitive Tax of Chatbots

Chatbot interfaces have a fatal flaw: they assume all work can happen through conversation. But most knowledge work requires structured thinking, multi-step operations, and persistent state tracking.

Research shows that when conversations become chaotic, both sides compound the problem. The AI, optimized to be helpful, mirrors back every unstructured thought the user expresses. The user, overwhelmed, lacks the mental bandwidth to reorganize. Those hurt most are less experienced workers—the very people who could benefit most from AI assistance.

This phenomenon is called the "Cognitive Tax": the mental resources consumed by the interface itself, offsetting the intelligence gains from AI.

Case Studies

Negative Case: Traditional Chatbot Alex uses ChatGPT to prepare a history report. He enters the topic; AI returns a 2,000-word overview. Alex extracts key points, asks follow-up questions, and receives more lengthy responses with "helpful" recommendations for three related topics. Three hours later, Alex has 12 browser tabs open, notes scattered across three documents, and the report hasn't started.

Positive Case: Dedicated Workspace Interface Jordan uses NotebookLM for the same assignment. PDFs, web pages, and notes are imported into a unified space. AI automatically organizes information connections, generating summaries and Q&A cards. Jordan can query specific passages anytime; AI responds precisely and concisely. Two hours later, the report structure is clear, materials organized.

An Educator's Guide to Interface Selection

1. Match Tools to Task Types

• Creative brainstorming: Chatbots work well
• Deep research: Choose dedicated tools like NotebookLM or Perplexity
• Programming education: Use IDE-integrated tools like Claude Code or GitHub Copilot
• Visual design: Explore AI-native interfaces like Google Stitch

2. Teach "Interface Literacy"

Don't just teach students how to use AI—teach them which interface to use when. This is like teaching when to do mental math versus using a calculator.

3. Beware the "One-Size-Fits-All" Trap

Tools claiming "one chatbot for everything" often perform mediocrely in all scenarios. Real efficiency comes from combining specialized tools.

4. Monitor Cognitive Load Indicators

If students become more anxious or confused after using AI, it's not the AI—it's the interface mismatch. Switching tools beats persisting with the wrong choice.

Conclusion

In the AI era, choosing the right interface means choosing the right way to learn. When we use chatbots as the only entry point, we inadvertently train students to accept fragmented, unstructured thinking patterns.

Education isn't about information acquisition—it's about building knowledge systems. This requires providing students with tool interfaces that support deep thinking, not letting them get lost in endless conversations.

想象一下：你问AI一个问题，它回复五大段文字，答案藏在某一段里，同时还附带三个你根本没问的新话题建议。对话越来越乱，AI越"热心"，你越迷糊。这不是AI不够聪明，而是界面设计出了问题。

聊天框的"认知税"

聊天机器人界面有一个致命缺陷——它假设所有工作都可以通过对话完成。但现实中，大多数知识工作都需要结构化思考、多步骤操作和持续的状态跟踪。

研究显示，当对话变得混乱时，双方都在加剧问题：AI为了"有用"而不断回应用户的每一个随意想法，用户则被信息淹没而无力重整。最受伤的是经验较少的职场新人——恰恰是最需要AI帮助的人群。

这种现象被称为"认知税"（Cognitive Tax）：界面本身消耗的心智资源，抵消了AI带来的智能增益。

案例对比

反面案例：传统聊天框 小明用ChatGPT准备历史课报告。他输入主题，AI返回2000字综述。小明从中提取要点，再次提问，AI又给出新的长篇回答，还"贴心"推荐了三个相关话题。三小时后，小明打开了12个浏览器标签，笔记散落在三份文档里，报告还没开始写。

正面案例：专用工作界面 小红使用NotebookLM研究同样的课题。她把PDF、网页、笔记导入一个统一空间，AI自动整理信息关联，生成摘要和问答卡片。她可以随时追问特定段落，AI的回答精准且简洁。两小时后，报告框架清晰，素材齐全。

教育者的界面选择指南

1. 根据任务类型选择工具

• 创意发散：聊天框适合头脑风暴
• 深度研究：选择NotebookLM、Perplexity等研究专用工具
• 编程学习：使用Claude Code、GitHub Copilot等IDE集成工具
• 视觉设计：尝试Google Stitch等AI原生设计界面

2. 教学生"界面素养"

不只是教怎么用AI，更要教什么时候用什么界面。这就像教学生用计算器之前，先教他们什么时候该心算、什么时候该列竖式。

3. 警惕"一站式"陷阱

那些声称"一个聊天框解决所有问题"的工具，往往在所有场景都表现平庸。真正的效率来自专用工具的组合使用。

4. 关注认知负荷指标

如果学生使用AI后反而更焦虑、更混乱，不是AI的问题，是界面不匹配。及时切换工具，比坚持错误的选择更明智。

结语

AI时代，选择正确的界面，就是选择正确的学习方式。当我们把聊天框作为唯一入口时，我们也在无意中训练学生接受碎片化、非结构化的思维模式。

教育的本质不是获取信息，而是构建知识体系。这需要我们为学生提供能够支持深度思考的工具界面——而不是让他们在无尽对话中迷失。

If yes, you've probably witnessed that cringe-worthy scene: a child asks AI a math question, and the AI responds with five paragraphs—definitions, examples, and the actual answer buried somewhere in the third paragraph. The child ends up more confused than before.

This isn't because AI isn't smart enough. It's because the interface is wrong.

Recently, Wharton professor Ethan Mollick published a thought-provoking article arguing that AI capabilities far exceed what most people realize, but poor user interfaces are wasting that potential. He uses a vivid metaphor: we're holding a Swiss Army knife but only know how to use the dullest blade.

Why Chatbots Aren't Universal Solutions

Mollick cites a study involving financial professionals using GPT-4o for complex valuation tasks. The results showed that while AI did boost productivity, the cognitive burden imposed by chat interfaces nearly offset those gains.

What's the problem?

1. Information overload: You ask a specific question, AI responds with five paragraphs, answer buried in the middle
2. Topic drift: AI "helpfully" suggests three new directions you didn't ask for, disrupting your flow
3. Conversation chaos: Once a conversation gets messy, AI mirrors your confusion, creating a downward spiral

The worst-hit? Beginners—the very people who could benefit most from AI assistance.

This resonates deeply with educational contexts. When a student uses AI to learn math, having to hunt for answers in lengthy responses while fending off sudden suggestions destroys learning efficiency.

The Rise of Specialized Interfaces

Mollick highlights several Google experiments:

Stitch: An AI interface for designers. Describe an app in natural language, get back multi-screen interactive prototypes—using design language, not prompting.

Pomelli: For marketers. Paste a website URL, automatically generate brand-consistent social media campaigns.

NotebookLM: For researchers. Integrate diverse information sources, present findings in structured formats.

The common thread? Each tool redesigns interaction for specific tasks.

What does this mean for education?

• Math learning AI → Should function like a collaborative whiteboard, guiding step-by-step, not chatting
• Writing tutor AI → Should behave like editorial comments, highlighting issues rather than rewriting everything
• Language practice AI → Should act like a conversation partner with clear roles and scenarios

Khan Academy's recent test prep resources embody this approach: instead of letting students "ask AI how to prepare," they provide structured skill checklists and practice pathways.

Recommendations for Educators

1. Look at Interfaces, Not Just Models

Don't only care about "Is this GPT-4?" The same large model performs vastly differently in a chatbox versus a specialized interface.

2. Beware the "Universal AI" Trap

If an AI claims to do everything, it probably does everything mediocrely. Educational contexts need specialized tools.

3. Monitor Cognitive Load

Good educational AI should reduce student cognitive burden, not increase it.

4. Cultivate "Interface Awareness"

Teach children: different tasks require different AI tools. Just as you wouldn't use Word for Excel tasks, you shouldn't use chat AI for everything.

Conclusion

The divergence of AI interfaces has just begun. Mollick puts it bluntly: AI capabilities are already strong; what's limiting us isn't technology, but how we interact with it.

For educators, this signals a shift: from "teaching kids to use AI" to "teaching kids to choose the right AI tool for the job."

After all, future competitiveness doesn't lie in whether you can chat with AI, but in whether you can find the most suitable AI interface for your current task—and make it work for you.

Source: One Useful Thing - "Claude Dispatch and the Power of Interfaces"

如果是，那你可能已经发现了那个让人哭笑不得的场景：孩子问AI一道数学题，AI洋洋洒洒写了五百字，从定义讲到例题，最后答案藏在第三段中间。孩子看完更懵了。

这不是AI不够聪明——是界面设计错了。

最近，沃顿商学院的Ethan Mollick在一篇新文章中提出了一个尖锐观点：AI的能力已经远超大多数人的认知，但糟糕的用户界面正在浪费这种能力。 他用了一个形象的比喻：我们手里拿着一把瑞士军刀，却只会用最钝的那把刀片。

为什么聊天框不是万能解

Mollick引用了一项针对金融专业人士的研究：让一群人用GPT-4o做复杂的估值任务。结果显示，AI确实提升了效率，但聊天界面带来的认知负担几乎抵消了这种提升。

问题出在哪？

1. 信息过载：你问一个具体问题，AI回你五段话，答案埋在中间
2. 话题发散：AI"贴心"地提供三个你没问的新方向，打断思路
3. 对话混乱：一旦聊乱了，AI只会镜像你的混乱，双方一起螺旋下降

最惨的是新手——恰恰是那些最需要AI帮助的人。

这让我想起教育场景。一个学生用AI学数学，如果每次都要在长篇大论里找答案，还要应付AI突然建议"要不要我讲讲微积分的历史"，学习效率怎么可能高？

专用界面正在崛起

Mollick特别提到了Google正在实验的几个AI工具：

Stitch：面向设计师的AI界面。你描述一个App，它直接生成多屏交互原型，用设计语言而非提示词。

Pomelli：面向营销人员。粘贴网站链接，自动生成品牌一致的社交媒体campaign。

NotebookLM：面向研究者。整合多种信息源，用结构化方式呈现研究成果。

注意共同点：每个工具都为特定任务重新设计了交互方式。

在教育领域，这意味着什么？

• 数学学习AI → 应该像解题白板，一步步引导，而非聊天
• 写作辅导AI → 应该像编辑批注，标注问题而非重写全文
• 语言练习AI → 应该像对话伙伴，有明确的角色和场景

Khan Academy最近推出的测试备考资源也体现了这个思路：不是让学生"问AI怎么备考"，而是直接提供结构化的技能清单和练习路径。

给教育者的建议

1. 看界面，不看模型

别只关心"用的是不是GPT-4"。同样的大模型，放在聊天框里和放在专用界面里，效果天差地别。

2. 警惕"万能AI"陷阱

如果一个AI声称什么都能做，那它很可能什么都做得一般。教育场景需要专用工具。

3. 关注认知负荷

好的教育AI应该降低学生的认知负担，而不是增加。

4. 培养"界面意识"

教会孩子：不同的任务需要不同的AI工具。就像你不会用Word做Excel的事，也不应该用聊天AI做所有事。

总结

AI界面的分化才刚刚开始。Mollick说得很直白：现在的AI能力已经很强了，限制我们的不是技术，而是我们如何与技术交互。

对教育者来说，这意味着一个转变：从"让孩子学会用AI"变成"让孩子学会选择合适的AI工具"。毕竟，未来的竞争力不在于你会不会和AI聊天，而在于你能不能找到最适合当前任务的AI界面——然后，让它为你工作。

参考来源：One Useful Thing - "Claude Dispatch and the Power of Interfaces"

Introduction: When AI Test Prep Surpasses Humans

In late 2025, GPT-4 scored higher than 90% of human test-takers on the bar exam. Yet when researchers asked it to handle real client consultations, its performance fell far short of expectations. This gap reveals a critical oversight: we are evaluating AI the wrong way.

Professor Ethan Mollick of Wharton School proposes a sharp observation: most AI benchmarks are like giving job candidates a standardized test, while true capabilities only emerge during a job interview.

Analysis: Three Blind Spots in Traditional AI Testing

1. Data Contamination: AI Is Memorizing Answers

Mainstream tests like MMLU-Pro and GPQA have had their questions and answers publicly available for years. Many AI models have seen these questions during training—this is not capability demonstration, it is memorization.

More embarrassingly, some test questions contain errors. Mollick notes that MMLU-Pro includes questions like What is the approximate mean cranial capacity of Homo erectus?—questions that even human experts might struggle to answer accurately.

2. Score Inflation: What Does 1% Improvement Mean?

When an AI improves from 84% to 85% on a test, is this a breakthrough or statistical noise? We lack calibration—we do not know what real capability differences different score ranges represent.

3. Context Disconnect: Exam Champions, Real-World Novices

An AI might excel at SWE-bench coding tests yet fail to understand a vague real-world requirements document. It might pass medical exams but freeze when facing complex patient cases.

Case Study: From Taking Tests to Doing Work

Mollick suggests adopting job interview style evaluation: give AI a real task and observe how it completes it.

Traditional test asks: Which is the correct syntax for sorting a list in Python?

Real task asks: Help me organize this student grade data, identify the top 10 most improved students, and generate a visualization report.

The latter tests not just syntax knowledge but also: requirement comprehension, data cleaning, logical reasoning, tool selection, and result presentation—the integrated skills the real world demands.

Recommendations: How Educators Should Redesign AI Assessment

For Students: From Can Use to Can Verify

Do not settle for AI-generated answers; learn to question and verify:

• Ask AI to explain its reasoning process
• Request information sources
• Cross-verify critical conclusions with different AIs
• Test its performance in edge cases

For Teachers: Design Real Task Assessments

Rather than testing whether students remember a specific AI feature, design open-ended tasks:

• Use AI to assist in completing a market research report
• Have AI help you analyze the argumentative flaws in this paper
• Design an AI workflow to automate class attendance tracking

Evaluation criteria should not be what tools were used but what problems were solved.

For Administrators: Build AI Capability Frameworks

Establish AI capability assessment frameworks for your teams:

• Foundation: Can they accurately describe requirements?
• Intermediate: Can they decompose complex tasks?
• Advanced: Can they verify and iterate on AI outputs?

Conclusion: The End of Testing, The Beginning of Practice

Mollick's core insight is simple: the best way to evaluate AI is to have it do real work.

The implications for education are profound. When our students leave school, they face not standardized tests but fuzzy, complex, uncertain real-world problems.

Teaching them how to give AI a job interview—asking good questions, verifying answers, iterating improvements—is more valuable than teaching them any single tool.

After all, in the AI era, the ability to ask the right questions matters more than knowing the right answers.

引入：当AI刷题超越人类

2025年底，GPT-4在律师资格考试中得分超过90%的人类考生。但有趣的是，当研究人员让它处理真实的客户咨询时，表现却远不如预期。这个反差揭示了一个被忽视的问题：我们正在用错误的方式评估AI。

宾夕法尼亚大学沃顿商学院的Ethan Mollick教授提出了一个尖锐的观察：大多数AI基准测试就像让应聘者做一份标准试卷，而真正的能力只有在面试中才能显现。

分析：传统AI测试的三大盲区

1. 数据污染：AI在背答案

MMLU-Pro、GPQA等主流测试的题目和答案在网上公开已久。许多AI模型在训练时已经见过这些题目——这不是能力的体现，而是记忆的展示。

更尴尬的是，一些测试题目本身存在错误。Mollick指出，MMLU-Pro中甚至包含Homo erectus的平均颅容量是多少这类连人类专家都未必能准确回答的问题。

2. 分数膨胀：1%的进步意味着什么？

当AI在某项测试上从84%提升到85%，这是重大突破还是统计噪音？我们缺乏校准——不知道不同分数区间代表的真实能力差异。

3. 脱离场景：考试高手，实战菜鸟

AI可能在SWE-bench编程测试中表现优异，却无法理解一个模糊的真实需求文档。它可能通过医学考试，却在面对复杂病例时束手无策。

案例：从做题到做事

Mollick建议采用工作面试式评估：给AI一个真实的任务，观察它如何完成。

传统测试问： 以下哪个是Python中列表排序的正确语法？

真实任务问： 帮我整理这份学生成绩数据，找出进步最大的前10名学生，并生成可视化报告。

后者测试的不仅是语法知识，还包括：需求理解、数据清洗、逻辑推理、工具选择和结果呈现——这才是真实世界需要的综合能力。

建议：教育者如何重新设计AI评估

对学生：从会用到会验

不要满足于AI给出的答案，学会质疑和验证：

• 让AI解释它的推理过程
• 要求提供信息来源
• 用不同AI交叉验证关键结论
• 测试它在边界情况下的表现

对教师：设计真实任务评估

与其测试学生是否记得某个AI功能，不如设计开放性任务：

• 用AI辅助完成一份市场调研报告
• 让AI帮你分析这篇论文的论证漏洞
• 设计一个AI工作流，自动化处理班级考勤

评估标准不是用了什么工具，而是解决了什么问题。

对管理者：建立AI能力矩阵

为团队建立AI能力评估框架：

• 基础层：能否准确描述需求？
• 进阶层：能否分解复杂任务？
• 高阶层：能否验证和迭代AI输出？

总结：测试的终点，实践的起点

Mollick的核心观点很简单：评估AI最好的方式，是让它做真正的工作。

这对教育的启示是深远的。当我们的学生走出校门，他们面对的不是标准化试卷，而是模糊、复杂、充满不确定性的真实问题。

教会他们如何给AI一场面试——提出好问题、验证答案、迭代改进——比教会他们任何单一工具都更有价值。

毕竟，在AI时代，提出正确问题的能力，比知道正确答案更重要。

Picture this: A middle school administrator walks into her office. Her "digital team" has been working overnight—AI agents have optimized this week's class schedules, analyzed last month's student assignment data, drafted parent meeting notices, and even adjusted final assessment plans based on teacher feedback.

This isn't science fiction. In his latest article, Ethan Mollick makes a compelling case: In the age of AI agents, management is becoming the most scarce superpower.

Not programming. Not data analysis. Management.

Analysis: Why Management Became a Superpower

1. The Fundamental Shift in Work

Traditional work model: One person completes one task. AI agent model: One person manages multiple AI agents to complete complex projects.

It's like evolving from a craftsperson to a conductor. You no longer play every instrument yourself—you coordinate the entire orchestra's harmony.

2. The "Employee-like" Nature of AI Agents

Mollick points out that AI agents share several characteristics with human workers:

• They make mistakes: Require checking and correction
• They have limitations: Excel at some tasks, struggle with others
• They need direction: Clearer instructions yield better outputs
• They can collaborate: Multiple agents working together produce better results

This means managing AI and managing humans share underlying principles.

3. The Unique Complexity of Educational Settings

Educational management is more complex than other fields:

• Involves multiple stakeholders (students, parents, teachers, institutions)
• Requires balancing efficiency with human care
• Decisions have long-term impacts (affecting students for years)
• Ethical boundaries are sensitive (data privacy, fairness)

Case Studies: Three Educational Leaders' AI Practices

Case 1: The Elementary Teacher's "AI Teaching Assistant Team"

Ms. Zhang manages a "teaching assistant team" of three AI agents:

• Content Agent: Handles class announcements and parent communications
• Data Analysis Agent: Tracks student assignment completion and flags students needing attention
• Creative Agent: Designs class activities and holiday celebration plans

Ms. Zhang spends 15 minutes each morning in a "stand-up meeting"—reviewing outputs, assigning daily tasks, and adjusting priorities. "I used to work until midnight," she says. "Now I can leave at 5 PM."

Case 2: The Principal's "Decision Support System"

Principal Li built a lightweight decision support system using AI agents:

• Collects and organizes teacher feedback and suggestions
• Analyzes student performance data to identify trends
• Compares curriculum setups with peer schools
• Generates pros-and-cons analyses for policy adjustments

"AI doesn't make decisions for me," Principal Li says. "But it helps me see the full picture before I decide."

Case 3: Online Education Platform's "Course Quality Monitoring"

A course director at an online education platform uses an AI agent team to monitor hundreds of courses:

• Automatically analyzes student reviews and completion rates
• Identifies course content needing updates
• Generates improvement suggestions for instructors
• Predicts course market performance

This system reduced course quality assessment cycles from quarterly to weekly.

Recommendations: Developing "AI Management Capability"

For Teachers: From User to Manager

Step 1: Identify Outsourcable Tasks List 10 things you do repeatedly each week. Mark which ones can be delegated to AI.

Step 2: Build "AI Workflows" Don't expect one AI to solve everything. Break tasks down and assign them to different AI tools or agents.

Step 3: Cultivate "Quality Control" Habits Always check AI outputs. Build your quality checklist: factual accuracy, appropriate tone, privacy compliance.

For Administrators: From Doer to Coordinator

Step 1: Redefine Your Role Your value is no longer being "the person who does the most things" but "the person who helps the team (including AI) produce maximum value."

Step 2: Establish AI Usage Guidelines

• Which decisions must be human-made?
• How should AI-generated content be labeled?
• Where are the boundaries of data privacy?

Step 3: Develop Team-wide AI Management Literacy Don't just teach everyone to use AI—teach everyone to manage AI.

For Students: Learning Future Skills Early

Ironically, today's students may need this capability earlier than their teachers. When they enter the workforce, "managing AI teams" may be a basic requirement.

Schools can:

• Introduce AI collaboration in projects
• Let students experience "directing" AI to complete tasks
• Discuss the ethical boundaries of AI management

Conclusion: Management Is the Future

Mollick's article offers a thought-provoking conclusion: AI won't replace managers, but managers who use AI will replace those who don't.

In education, this is especially important. Because educational leaders' decisions shape the next generation's future.

Learning to manage AI isn't about making our lives easier (though it does)—it's about freeing us to focus on what truly requires human judgment:

• Understanding a student's unique circumstances
• Finding balance between efficiency and equity
• Preserving the essence of education amid change

Management is becoming the superpower of the AI age. And education is where this capability matters most.

What do you think? Which tasks in your work have been—or could be—delegated to AI agents? Share in the comments.

想象这样一个场景：一位中学教务主任早上走进办公室，她的"数字团队"已经工作了一整夜——AI代理完成了本周的课程表优化、分析了上月学生作业数据、草拟了家长会通知，甚至根据教师反馈调整了期末评估方案。

这不是科幻。Ethan Mollick在最新文章中提出一个核心观点：在AI代理时代，管理能力正在成为最稀缺的超能力。

不是编程，不是数据分析，而是——管理。

分析：为什么管理成了超能力？

1. 工作性质的根本转变

传统工作模式：一个人完成一个任务。 AI代理模式：一个人管理多个AI代理完成复杂项目。

这就像从"手艺人"变成"乐队指挥"。你不再亲自演奏每一种乐器，而是协调整个乐团的和声。

2. AI代理的"员工特性"

Mollick指出，AI代理有几个类似人类员工的特点：

• 会犯错：需要检查和纠错
• 有局限：某些任务做得好，某些不行
• 需要指导：越清晰的指令，输出越好
• 可以协作：多个代理分工配合效果更佳

这意味着，管理AI和管理人，底层逻辑是相通的。

3. 教育场景的特殊性

教育管理比其他领域更复杂：

• 涉及多方利益（学生、家长、教师、学校）
• 需要平衡效率与人文关怀
• 决策影响周期长（影响学生多年发展）
• 伦理边界敏感（数据隐私、公平性）

案例：三位教育管理者的AI实践

案例1：小学班主任的"AI助教团队"

张老师管理着一个由3个AI代理组成的"助教团队"：

• 文案代理：负责班级通知、家长信撰写
• 数据分析代理：跟踪学生作业完成情况，标记需要关注的学生
• 创意代理：设计班级活动方案、节日庆祝策划

张老师每天花15分钟"开晨会"——检查各代理的产出，分配当天任务，调整优先级。她说："以前我每天忙到深夜，现在下午五点就能下班。"

案例2：中学校长的"决策支持系统"

李校长用AI代理构建了一个轻量级决策支持系统：

• 收集整理教师反馈和建议
• 分析学生成绩数据，识别趋势
• 对比同类学校的课程设置
• 生成政策调整的利弊分析

"AI不会替我做决定，"李校长说，"但它让我在做决定前，能看到更全面的图景。"

案例3：在线教育平台的"课程质量监控"

某在线教育平台的课程负责人，用AI代理团队监控数百门课程的质量：

• 自动分析学生评价和完课率
• 识别需要更新的课程内容
• 生成教师改进建议
• 预测课程的市场表现

这套系统将课程质量评估的周期从季度缩短到周。

建议：如何培养"AI管理力"

对教师：从使用者到管理者

第一步：识别可外包任务 列出你每周重复做的10件事，标记哪些可以交给AI。

第二步：建立"AI工作流" 不要期待一个AI解决所有问题。把任务拆解，分配给不同的AI工具或代理。

第三步：培养"质检"习惯 AI的输出一定要检查。建立你的"质检清单"：事实准确性、语气适当性、隐私合规性。

对管理者：从执行者到协调者

第一步：重新定义角色 你的价值不再是"做最多事的人"，而是"让团队（包括AI）产出最大价值的人"。

第二步：建立AI使用规范

• 哪些决策必须人工做？
• AI产出的内容如何标注？
• 数据隐私的边界在哪里？

第三步：培养团队的AI管理素养 不是教每个人用AI，而是教每个人管理AI。

对学生：提前学习未来技能

讽刺的是，今天的学生可能比老师更早需要这项能力。当他们进入职场，"管理AI团队"可能是基本功。

学校可以：

• 在项目中引入AI协作
• 让学生体验"指挥"AI完成任务
• 讨论AI管理的伦理边界

总结：管理即未来

Mollick的文章有一个令人深思的结论：AI不会取代管理者，但会用AI的管理者将取代不会用的。

在教育领域，这句话尤其重要。因为教育管理者的决策，影响的是下一代的未来。

学会管理AI，不是为了让我们更轻松（虽然确实会），而是为了让我们能专注于那些真正需要人类判断的事情：

• 理解一个学生的独特处境
• 在效率和公平之间找到平衡
• 在变化中守护教育的本质

管理正在成为AI时代的超能力。而教育，正是最需要这项能力的地方。

你怎么看？你的工作中，哪些任务已经或可以交给AI代理？欢迎在评论区分享。

This shift has profound implications for education. The question is no longer "How do we teach kids to use AI?" but "How do we prepare them to design how AI works?"

From Commander to Architect

The traditional human-computer interaction model assumes a simple hierarchy: humans command, machines obey. For decades, this framework shaped how we taught computer literacy — type these commands, click here, use this shortcut.

AI breaks this model.

Mollick's "Interface is all you need" thesis suggests that in the AI era, efficiency depends not on AI capability but on interface design. It's about:

• Shifting from "commanding AI" to "architecting AI workflows"
• Moving from "using AI" to "orchestrating multiple AIs"
• Prioritizing not "the strongest AI" but "the right combination"

A Concrete Comparison

Consider two approaches to writing a research report on climate change:

Traditional: Use ChatGPT for a draft, then revise yourself. AI is simply a faster search engine.

New paradigm: Design an "AI writing pipeline" — Perplexity for research, Claude for drafting, another AI for fact-checking, yourself for final editing. You're an AI architect, designing inputs and outputs for each stage.

The efficiency gap? Potentially 10x or more.

What Educators Should Do

1. From skills training to systems thinking

Instead of teaching "how to write good prompts," teach "how to decompose tasks and select appropriate tools."

2. Interface design becomes universal

Clear requirement articulation, workflow planning, and checkpoint setting become essential skills for everyone — not just programmers.

3. Cultivate "AI metacognition"

Help students understand AI as a new type of collaborator with its own strengths and limitations. Understanding your partner is prerequisite to effective partnership.

Conclusion

The scarcest skill in the AI age isn't "using AI" — it's "designing how AI works." That's the difference between knowing how to drive and knowing how to design a transportation network. Education must evolve to bridge this gap.

当我们在讨论“AI时代孩子要学什么”时，大部分人的答案还停留在“学编程”、“学prompt技巧”。但这个观察指向了一个更根本的问题：未来重要的不是你会用AI，而是你懂得如何设计AI的工作方式。

从“指挥者”到“设计师”

传统的人机交互，有一个隐含假设：人类是指挥者，机器是执行者。你给指令，AI执行。这构成了我们过去几十年与计算机打交道的基础范式。

但现在的AI，特别是Agent形态的AI，正在打破这个范式。

Mollick提出的“Interface is all you need”，核心观点是：AI时代，决定效率的不是AI有多强，而是你设计的交互界面有多好。

这意味着：

• 从“给AI下命令”转向“给AI设计工作流”
• 从“我会用AI”转向“我懂得如何让AI之间协作”
• 从“找最强AI”转向“选择最合适的AI组合”

一个教育场景的对比

同样是写一篇关于气候变化的报告：

传统方式：你用ChatGPT生成初稿，然后自己修改。这个流程里，AI只是一个更快的搜索引擎。

新范式：你设计一个“AI写作流水线”——先用Perplexity掌挥权威资料，再让Claude写初稿，然后让另一个AI做事实核查，最后你自己做最终编辑。在这个流程里，你是一个“AI架构师”，设计每个环节的输入输出。

这两种方式的效率差距，可能达到10倍甚至更多。

教育者应该怎么做

1. 从“技能训练”转向“系统思维”

与其教孩子“如何写一个好的prompt”，不如教他们“如何拆解任务并选择合适的工具”。

2. 重视“接口设计”能力

界面设计不再是程序员的专属技能。当每个人都在与AI协作，清晰表达需求、规划工作流程、设置检查点，这些能力变得足够重要。

3. 培养“AI元讪知”

让孩子理解：IA是一种新型合作伙伴，它有自己的优势势势。你需要了解它，才能更好地与它配合。

总结

AI时代最稀缺的能力，不是“会用AI”，而是“懂得如何设计AI的工作方式”——这相当于从“会开车”升级到“会规划城市交通网络”。教育需要为这个转变做好准备。

Here's the core insight: AI capability is not determined by the model alone. It's determined by three layers: Model, App, and Harness.

Layer 1 — Model: The Brain

The model is the underlying intelligence. GPT, Claude, Gemini — these are models. They determine how well an AI reasons, writes, or analyzes. But models alone can't do anything. They need to be housed somewhere.

Layer 2 — App: The Interface

The app is what you interact with directly — a website, a mobile app, a desktop tool. The same Claude model performs completely differently on Claude.ai versus Claude Code. One gives you answers. The other automates entire workflows. The app determines the user experience.

Layer 3 — Harness: The Infrastructure

The harness is what lets AI take real-world actions. Without a harness, AI is a very smart assistant. With a harness, AI becomes an autonomous agent that can browse the web, write files, send emails, and execute multi-step tasks. This is where tools like OpenClaw live.

Why This Matters for Educators

Most educators evaluate AI tools by model name. This framework reveals why that's insufficient:

• A powerful model in a weak app = disappointing results
• A well-designed harness unlocks the model's full potential
• When AI agents enter the classroom, it's the harness layer that raises questions about control, permissions, and oversight

Understanding this three-layer model helps educators make smarter tool selections, set appropriate expectations, and prepare for an AI-integrated classroom where agents work alongside students.

Ethan Mollick 最近一篇文章揭示了答案——AI的能力不只是由模型决定，还由"应用"和"框架"两层决定。

第一层：模型（Model）—— AI的大脑

模型是底层能力，决定AI有多聪明、会写代码还是做分析。主流模型包括GPT、Claude、Gemini。但模型只是大脑，放在不同的"身体"里，表现完全不同。

第二层：应用（App）—— AI的皮肤

App是你直接使用的界面——网页版、手机App、桌面工具。同一个Claude，在Claude.ai网站上能联网查资料，在Claude Code里能操控电脑、跑代码、自动化办公。这就是App的差异。

第三层：框架（Harness）—— AI的手脚

框架是让AI真正"做事"的基础设施。没有框架，AI只能回答问题；有框架，AI可以自动执行任务、调用工具、串联工作流。OpenClaw就是这样的框架——让AI调用浏览器、写文件、操作你的电脑。

对教育者的实际意义

理解了"模型-应用-框架"三层架构，你就能判断：

• 为什么某个AI"看起来很聪明但用起来很笨"？（缺框架）
• 孩子的AI学习工具选哪个更合适？（看应用层设计）
• AI代理工具进入学校，老师需要了解什么？（框架层的权限管理）

选AI，不要只看模型参数。框架对了，模型才能发挥真正价值。