被偷走的窗口

一个男孩，一只鸟，和一个至今仍然重要的问题#

有一个关于鸟的故事。

1930 年代的某一天，一个小男孩和父亲坐在布鲁克林的院子里。一只鸟落在附近的篱笆柱上。父亲开始报出它的名字——先用意大利语，然后葡萄牙语、汉语、日语。四种语言，四个名字，都很唬人。

然后父亲说了一句男孩一辈子没忘的话。按费曼在《你管别人怎么想》里的转述，他父亲说：“你可以用世界上所有的语言知道那只鸟的名字，可等你说完，你对这只鸟仍然一无所知。你知道的只是不同地方的人，以及他们管这只鸟叫什么。所以，咱们来看看这只鸟，看看它在干什么——那才是要紧的。”¹

这个男孩后来成了理查德·费曼——物理学家，诺贝尔奖得主，那个在电视听证会上把一个 O 形圈浸进一杯冰水里、从而演示了挑战者号灾难中 O 形圈失效的人。在科学家中间，他以“追问到真实机制露面为止”而闻名。

他后来说，父亲教给他的，是“知道某样东西的名字，和真正知道某样东西，两者之间的区别”。¹

我给你讲这个故事，是因为它恰恰就是这篇文章要讲的东西。后面的一切，都是他父亲在院子里那句话的某个版本。鸟还在窗外。你可以知道它的名字——你也可以看看它在四月某个星期二的早上都在做什么：它为什么挑了那根特定的树枝，它怎么读取那些它感觉得到、而你看不见的磁场。

此刻，你的窗外多半有一只鸟。你口袋里有一个 AI，你问它什么，它都能答。这两件事并非毫无关系。

你的下午，你的 AI，你的问题#

你十二岁。或者差不多。你有一部手机，而且你今天用过它——也许是写作业，也许是掉进 YouTube 的兔子洞，也许是跟 ChatGPT 或 Copilot、或者在你那个版本的 2026 年里那个工具叫什么名字，聊了几句。

有件事你多半已经注意到了：有一种用法，用完之后你手里只剩一份写完的作业，别的什么也没有。你问了个问题，它答了，你抄下点东西，你关掉标签页。完事。还有另一种用法，四十分钟后你走出来，手里多了三个开始时并没有的新问题。

这个区别你已经注意到了。你只是可能还没给它起过名字。

这篇文章要给它起个名字。不是因为你不懂——你懂，就像你懂那些你感觉得到、却还没说出口的事情一样。我只是给你一个把手，让你握住一件你本来就已经感觉到的东西。

整篇文章就是这一个动作。每一步我都会告诉你我在干什么，包括哪里是我在猜、哪里不是。剩下的由你来判断。

两件事，按顺序来：那个窗口是什么，那个问题是什么。

学校拿你的那些年做了什么（又没能做什么）#

接下来这部分我得说说学校。我会说得快一点，因为大部分你都知道。

从幼儿园到毕业，你会在一个机构里待上大约 14000 小时，而这个机构在很多事情上是称职的。它教你事实。它教你坐得住、赶得上截止日期。那些事实里有一些确实要紧，而赶上截止日期也真的有用。我不打算告诉你学校是邪恶的。它不是。就结构而言，它只是在一件具体的事情上很糟糕。

它糟糕在：保护不了你身上那个想弄明白事情的部分——而且恰恰是在那个部分四面八方受压最重的那些年里。

证据在这儿。盖洛普对全美 2317 名 K-12 学生的调查发现，强烈同意“自己的功课重要、有趣、有挑战性，或者与自己的天赋相称”的学生，不到五分之一。² 不到五分之一。另一项盖洛普调查——只在爱荷华州，962 名五到十二年级的学生——发现只有 10% 的学生强烈同意自己喜欢上课，而大约三分之一说自己总是觉得无聊。²

关于这些数字的一句提醒：34% 和 10% 只来自爱荷华。一个州。爱荷华并不能代表所有地方。我把它们当质感用，不是当证据用。承重的是那个全国数字——不到五分之一。

现在说有意思的部分：有些学校里正在跑着一个自然实验，它告诉我们，当好奇心引擎被保护、而不是被过载时会发生什么。2023 年一项针对 32 项严格研究的系统综述发现，蒙台梭利教育有一致的正面效果。这些项目里的孩子在学业指标上——数学、语言——略好一些，而在执行功能、以及他们对自己学习的感受这类事情上，好得可测量。效应量是温和的，不是戏剧性的。（在这篇综述的所有指标里，“内在体验”那一项的不确定性最大——请轻拿轻放。）而且蒙台梭利研究有一个真实的选择偏差问题：会选蒙台梭利学校的家长本来就不寻常，他们的孩子也许无论如何都会更好。2023 年那篇综述做了敏感性分析，发现正面效果依然成立，但这个保留意见仍然有效。³

要点是：一个好奇心被保护住的孩子，学业上并没有落下。她略微往前走了一点，而且在这个过程中对学习的感受更好。你不必在好奇心和学业表现之间二选一。

初中里发生的一些事，只不过是长大。你的大脑围绕朋友、身份、什么对你要紧重新组织了一遍。这不是学校造成的。学校所做的——或者说没能做的——是保护你身上那个曾经想弄明白事情的部分，恰恰是在那个部分竞争者最多的那些年里。

那么：它没能保护的到底是什么？如果没人保护，会被侵蚀掉的又是什么？

大脑学习的两种方式#

想想一个一岁的孩子学说话。没人让她坐下来对着识字卡。没人考她。她只是在跟那些会回应她的人的真实对话中，把词听了成千上万遍，然后就这么吸收了——从里往外建起了一套语法，却一条规则都说不出来。

再想想一个三十岁的人上法语课。她能学会法语。只是运作方式不一样。她需要讲解、操练、明确的规则。她得学。她多半会一直带着一种口音，而这种口音在她小时候本来是不用费力就能避开的。这个机制是真的不同——研究者称之为自下而上的吸收式学习与自上而下的费力式加工，而且你确实能在大脑处理这项工作的方式上看到差别。⁴

值得注意的一个具体细节：研究者 Patricia Kuhl 2010 年的一项研究发现，通过真人辅导接触外语的婴儿，出现了显著的语音学习。而通过视频接触完全相同内容的婴儿，则完全没有学到。同样的输入，同样的声音——区别只在于一边有个活人在回应这个婴儿，另一边没有。⁴ 真实互动的作用，是被动观看所没有的。这是一个具体的、被重复验证过的发现，不是一个理论。

这说的是语言习得，不是一切。我接下来要把它延伸成一个类比，而我想把这一点说清楚，好让你知道我什么时候站在结实的地面上，什么时候只是在画一幅可能画错的画。

类比是这样的：面对任何一个出现在你面前的问题，似乎都有两种与它打交道的方式。一种是吸收式的——你追下去，是因为你想知道答案，就像一个孩子追着一门语言，是因为她想跟她爱的人说话。另一种是费力式的——你产出一个结果，是因为有人要拿它来评价你，就像一个成年人背单词，是因为要考试。

让这不只是一个猜想的研究是：1973 年，斯坦福的 Mark Lepper 和同事找来一批本来就喜欢画画的学龄前儿童。他们给其中一些孩子许诺，画画可以得一颗小金星。那些期待着小金星的孩子，此后画得更少了，也更不享受了——尽管他们此前一直在自己自由地画。研究者把这称为过度合理化效应：说白了就是，如果你花钱请人去做一件他本来就喜欢的事，他就会不再喜欢它。⁵ 最初的研究用的是三到五岁的学龄前儿童。这个效应此后在数百项针对学龄儿童的研究中被重复验证。

一个重要的限定条件：它最清楚地适用于那些针对“孩子本可能真心感兴趣的任务”所给出的、预期之中的、有形的奖励——比如成绩。口头表扬不会以同样的方式造成伤害；针对“本来就没人觉得有趣的任务”的奖励也不会。所以相关的条件是“针对你本可能在意的话题所打的分数”，而不是“所有奖励，永远如此”。⁵

再往后，有一项 2016 年的纵向研究，追踪了 600 名 11 到 16 岁的学生，记录下研究者所称的“青春期内在动机的显著下降”。⁶ 真正要紧的发现是：那些满足了学生基本需求——自主感、胜任感、归属感——的学校，下降幅度更小​。下降仍然发生了。这一点很重要。即便在最好的学校里，动机也在往下掉。研究者认为，这一部分只是长大而已。学校不是这场下降的原因。但学校环境预测了这个坡有多陡。

纵向数据要到 11 岁左右才变得干净。6 到 10 岁之间发生了什么，我们是从机制研究（Lepper）和后段数据的形状里推出来的。我正在跨三个研究传统做综合——语言习得、动机心理学、青少年发展。我把“我在做综合”这件事告诉你，也告诉你这三条线索中的任何一条，都可能比它看起来更弱。

现在说那个操作系统的比喻。我打算把这两种模式叫作吸收式 OS 和费力式 OS。我想说清楚这意味着什么、又不意味着什么。我用“操作系统”这个词的方式，就像物理学家用“弹簧”去描述电子在原子附近的行为——它是一幅图。这幅图做的是真活儿。但它并不拥有大脑。

这幅图所说的是：在童年和青春期早期的某个时候，你会定型出一种面对新信息的默认方式。吸收模式说的是“那是什么，它怎么运作，我把这块拽一下会怎么样”。费力模式说的是“答案是什么，什么时候交”。两种模式都是真的。两种都能学到东西。问题在于：当没人告诉你该用哪一种时，默认跑起来的是哪一种。

在感觉系统里——鸟怎么学会自己的歌，视觉皮层在幼年期怎么接线——特定窗口期内的经验会永久地塑造那条回路。⁷ 这是神经科学，不是比喻。我并不是说你的动机默认值上会发生同一件事。我要指出的是，大脑并不是一场永久开放的自由混战：窗口是存在的。动机研究的证据指向一个类似的模式，即便其机制不同、也理解得更少。

那么，吸收式 OS 跑起来的时候，究竟长什么样？

红石电路，以及你造它时大脑在做什么#

你多半在《我的世界》里造过红石电路。就算没有，你也做过类似的事——Roblox 里一条特别刁钻的障碍赛道，一个只有你自己在用的模组，一套试了十四次才成的基地防御系统。

从外面看，它是这样的：你花两个小时做了一件什么也没产出的事——没有分数，没有学分，也没人要求你做。你搞明白了比较器测的是它背后那个容器的充盈程度，也就是说，如果你把箱子朝向放错，信号就永远不会触发。你开始时并不知道“比较器”这个词。你也不知道自己在造一个与非门——你多半从没听过“NAND”这个词。但电路是通的。你是靠试、失败、再试，靠注意到改变延迟时中继器会怎么反应，把它造出来的。

那就是吸收式 OS 在跑。没人给它打分。没人叫你做。奖励是：电路做到了你想要的，或者没做到。而如果没做到，你能确切地看出为什么，然后再试一次。

研究者 Mihaly Csikszentmihalyi 花了几十年研究那种状态——四个小时感觉像二十分钟，而你从里面出来时手上有了一件你造出来的东西。他把它叫作“心流”。他的研究发现了一件奇怪的事：学校其实比你生活中大多数其他地方，更经常地创造出心流的结构性条件——相对于你的技能而言足够高的挑战。⁸ 然而你在学校里处在这种状态的时候，反而比在游戏里少。光有结构性条件并不能触发它。缺的那一块是：你的好奇心必须对准那件事。

2019 年的一项研究记录了当《我的世界》获得结构化使用时会发生什么：一个课后项目里的 118 名学生，在动机、问题解决、创造力以及读写能力上都有可测量的提升。一项研究能承载的分量有限。它是探索性的——没有对照组，自愿参与，某些方面还是自我报告的提升。研究者本人 Thierry Karsenti 也指出，这些提升需要有计划、有目的的投入——而不是漫无目的地乱按按钮。他这项研究之所以有用，不是因为它证明了《我的世界》有魔力，而是因为它记录下了你连着玩四个小时后本来就知道的事情：有目的的、在建模型的投入，会产出某种感觉像成长的东西，因为它就是成长。⁹

Karsenti 那项研究只是一项研究。它的方法有真实的局限。我用它，是因为它的发现与你本来就知道的东西吻合，而不是因为一项研究能定论什么。值得信任的，是证据与你自己对“真正投入其中、正在建造什么、正在追着机制跑时《我的世界》是什么感觉”的经验之间的那种吻合。

这就是学习在没有被翻译成服从时的样子。

你在爸妈叫你别玩了的时候所进入的那种吸收状态，练的正是这整篇文章所指向的那件事。红石电路是真的。你造它时所处的那个状态是真的。它有一个名字，而它与费曼的父亲在院子里所指向的，是同一件事。

拉远一点看：一代人，一件工具，一个正在被决定的问题#

拉远一段的距离。

你是一间卧室里的一个十二岁孩子。但此刻摆在你面前的这个问题——吸收模式还是费力模式，追机制还是收集标签——是你们整整一代人要回答的问题，一个孩子一个孩子地回答。而你回答它时所处的语境，有一个此前任何一代人都没有过的特征：你手里握着的这件工具，能把你的问题追得比你自己更远。1990 年一个想弄懂红石（或者其电学等价物）怎么运作的十二岁孩子，有的是一座图书馆、也许一套百科全书，外加一位可能同样不懂的老师。而你有一个活的伙伴，你问机制问得多快，它就能答多快。这是新的。

关于就业市场说一段，因为你多半听过某个版本的“AI 要抢走工作了”。我们实际知道的是这些。耶鲁的预算实验室一直在仔细追踪就业数据——截至他们 2025 年 10 月那份覆盖到 2025 年 7 月数据的报告，也就是 ChatGPT 发布两年半之后，他们没有发现 AI 对就业造成了可辨识的全经济范围冲击。那些头条还没有出现在数字里。耶鲁对这意味着什么持谨慎态度：历史上的技术转型往往要几十年才会重塑劳动力市场。我们确实不知道这件事会怎么演下去。但有一件事不取决于结局：无论十五年后存在什么样的工作，一个好奇心引擎还完好的人，都会比一个引擎不完好的人适应得更好。这不是一个关于 AI 的预测。这是一句关于“好奇心引擎是干什么用的”的陈述。¹⁰

这篇文章的紧迫感不来自就业市场。它来自那个窗口。

再拉回来。

这个问题——哪个 OS 在跑——正在被每一个有手机的十二岁孩子回答，一次会话接着一次会话。不是被他们的学校，不是被他们的政府，也不是被任何政策委员会。一个孩子，一次对话，一句“好的谢谢”对上一句“等等，可它为什么是那样运作的？”决定就住在那儿。趁你还在里面的时候知道这件事，是值得的。

一个问题#

一个问题。这一节就只有这个。

当你问 AI 一件事、而它给了你一个答案时，你有一个选择。接受这个答案，关掉标签页。或者，问那个机制问题——费曼的父亲在院子里问的那个问题。¹ 不是“这个答案对不对”——你才十二岁，你常常判断不了，而且你也不需要判断。就只是：问下一个。“可它为什么是那样运作的？”

你不是在审计这个 AI。你是在给它重新指向。

研究背景是这样的：Yurt 和 Kuşci 2026 年发表在《Current Psychology》上的一项研究发现，不加反思就使用 AI 的大学生，批判性思维会下降，其中介是研究者所称的“认知懒惰”——接受 AI 的答案，而不去往下推它。¹¹ 样本是大学生，不是十二岁的孩子。我在推断这个机制开始得更早。我可能是错的。我认为我没错的理由是：Ron Aboodi 2025 年发表在《Educational Theory》上的一篇论文论证，习惯性地把思考外包给 AI，会形成一种累积性的性情倾向，而不只是一次瞬时的抄近路——你十二岁时怎么对待答案，可能会塑造你三十岁时怎么对待答案。¹² Aboodi 的论证是理论性的，不是经验性的。这意味着它是一个细致的哲学论证，而不是一项对照研究。我把这个区别告诉你，好让你自己去掂量。

研究显示的是：懒惰地使用 AI 与思考能力下降相关。它还没有显示出：有策展意识地使用 AI 能保住思考能力。这是我做的一个不对称的推断。你可以这样去检验它：在一次 AI 会话之后，注意一下你走开时带走的是一个新问题，还是只是一份做完的作业。那就是你的数据。如果这个两种模式的区分与你实际经验到的东西吻合，那这个机制就抓住了某种真实的东西。如果不吻合，那我描述的就是一件并没有发生在你身上的事，你也应该据此调低我这个论证的分量。

这个动作在实践中是这样的。你问 AI 红石比较器是怎么工作的。它告诉你了。你可以说“懂了，谢谢”——于是你现在拥有一个标签。比较器会比较。你也可以说：“可它为什么是那样比较的？它在电路内部实际测量的到底是什么？”现在这个 AI 成了你所在意的那个问题里的一个活的伙伴。你不是在检索一个答案。你是在追一个机制。同一件工具，两种模式。

他父亲问的是一只鸟。你问的是红石比较器，或者一个数学证明，或者为什么火星的天空白天是粉色的、日落时却变蓝，而我们的天空日落时是红的。同一个问题。同一个动作。¹

“可它为什么是那样运作的？”就这样。这就是这篇文章全部的实践性提议。它称不上是一种技巧——它更像一种姿态——而这恰恰是它在十二岁管用的原因。你不需要懂任何关于 AI 如何运作的知识才能问“为什么”。你只需要在意到愿意再多问一个问题。

那么：有没有什么你一直想弄明白的、关于你真正在意的东西的问题？打开 AI。从那儿开始。然后对它给你的每一个答案都追问为什么，看看你会走到哪里。

你还在里面#

回到你身上。

这篇文章一直在描述的那个窗口，不是抽象的。你就在里面，此刻，今天。它还会开多久——没有人能给你一个精确的数字，因为这个机制是渐进的，而证据来自多个在时间线上并不完全一致的来源。我们知道的是：吸收式的默认值会在童年和青春期中逐渐移位。这个移位对每个人都会发生。问题在于这个坡有多陡，以及趁你还在宽的那一段时你做了什么。

你是唯一一个保证会在这个窗口里待到最后的人。你父母不是。你老师不是。AI 不是。是你。

那个夺回来的动作，不是某个大人替你做的。你父母能帮忙——他们可以让自己的好奇心保持可见，可以在饭桌上问“为什么”，可以在你需要多追一会儿机制的时候，忍住那股直接把答案递给你的冲动。但决定哪个模式在跑的人是你。我把这篇文章写给你、而不是写给他们，是有原因的：你是主角。我能给你一个机制和一个问题，但我给不了你那些年。那些年在你手上。

哪些是证据、哪些是推断，我说清楚。我给你看了来自动机心理学、语言研究和几项具体研究的证据。我点明了哪里证据强（Lepper 的过度合理化效应——被大量重复验证），哪里证据弱（Karsenti 的《我的世界》研究——探索性的，单一场点），以及哪里是我在做可能出错的推断（从十二岁到三十岁的习惯养成那个主张）。我把我所推荐的那个动作示范了一遍，因为另一种做法——一边叫你批判性思考，一边写一篇自己并不这样做的文章——恰恰就是费曼的父亲所警告的东西。

那么：窗外那只鸟。你还记得。它有一个名字。关于它，你还知道些什么？

鸟还在窗外#

鸟还在那儿。

你现在知道它的名字了——或者就算不知道，也不难查到。那从来都不是问题所在。问题是它在星期二早上 10 点 47 分在干什么，它为什么挑了那根特定的树枝，它在十一月怎么在没有一张你看得见的地图的情况下找到南方，它怎么学会了自己唱的那支歌。这些问题不会随着一个答案而关闭。它们随着每一个答案而打开。

这就是这篇文章一直试图命名的那个区别。

标签会停下。机制会打开成另外三个问题，而那三个又打开成另外三个。这就是吸收模式——你追机制，而不是收集标签的那个模式。学校给你的多半是标签，因为标签可以打分，而机制不行。AI 两种都能给你，取决于你怎么问。而问的那个人是你。

费曼的父亲问的是一只鸟。你问的是比较器、化学反应、语言为什么会消亡，以及红石到底是怎么带电流的。问题是同一个问题。窗口还开着。

这是你要带走的东西。每一次你问 AI 什么，每一次老师递给你一个答案，每一次你发现自己正对着一段解释点头——就再问一个。“可它为什么是那样运作的？”这就是整篇文章，压缩成一句话。这就是我想让你拿到的东西。

鸟还在那儿。你的下午还开着。

参考文献#

爱荷华州数据：Gallup / Opportunity Education Foundation. (2025 年 6 月 3 日). How student agency can boost engagement and readiness. Gallup News​. 爱荷华调查于 2024 年 9 月 23 日至 11 月 18 日进行，n=962 名学生（5–12 年级）。https://news.gallup.com/poll/660503/student-agency-boost-engagement-readiness.aspx

支撑材料：Kuhl, P. K. (2010). Brain mechanisms in early language acquisition. Neuron​, 67(5), 713–727. https://doi.org/10.1016/j.neuron.2010.08.038

另见：Smith, Z. R., et al. (2023). Academic motivation decreases across adolescence for youth with and without ADHD. Journal of Child Psychology and Psychiatry​. https://doi.org/10.1111/jcpp.13815

支撑材料：García-Álvarez, J., & Acevedo-Borrega, J. (2025). Minecraft as a pedagogical tool: A systematic literature review (2014–2024). Journal of Educational Computing Research​. https://journals.sagepub.com/doi/10.1177/15554120251341034

延伸阅读#

• Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience​. Harper & Row.——关于心流状态的原始材料：它是什么感觉，什么时候发生，以及为什么即便结构性前提具备，学校条件也无法可靠地产生它。如果《我的世界》那一节戳到你了，下一步该去这里。
• Deci, E. L., & Ryan, R. M. (2000). The "what" and "why" of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry​, 11(4), 227–268.——第 3、4 节通篇所用的“自主—胜任—归属”框架背后的研究传统。比本文更技术性，但对一个有动力的成年人来说读得下去。
• Feynman, R. P. (1985). Surely You're Joking, Mr. Feynman! W. W. Norton & Company.——那本讲鸟的书的姊妹篇。如果你想知道更多费曼的好奇心在实践中如何运作——他怎么在洛斯阿拉莫斯撬锁、怎么学画画、怎么打邦戈鼓——就是这本。
• Blakemore, S-J. (2018). Inventing Ourselves: The Secret Life of the Teenage Brain​. PublicAffairs.——本文必须吸收的那个反面论证：为什么青少年动机的下降有一部分只是长大，而不全是学校造成的。Blakemore 是写给普通读者的，而且她认真对待青少年的大脑。
• Bowen, E., et al. (2025). Building AI Literacy at Home: How Families Navigate Children's Self-Directed Learning with AI. arXiv 2510.24070.——那项发现小学生还无法独立评估 AI 输出的研究——而这恰恰是本文把门槛降到“问下一个问题”的原因。如果你想看 AI 与儿童这场争论的另一面，值得一读。

The Stolen Window

A boy, a bird, and a question that still matters#

There's a story about a bird.

A small boy is sitting in a yard in Brooklyn with his father, sometime in the 1930s. A bird lands on a fence post nearby. The father starts naming it — in Italian, then Portuguese, then Chinese, then Japanese. Four languages, four names, all very impressive.

Then the father says something the boy never forgot. As Feynman tells it in What Do You Care What Other People Think?, his father said: "You can know the name of that bird in all the languages of the world, but when you're finished, you'll know absolutely nothing whatever about the bird. You'll only know about humans in different places, and what they call the bird. So let's look at the bird and see what it's doing — that's what counts."¹

The boy grew up to be Richard Feynman — physicist, Nobel laureate, the person who demonstrated the O-ring failure in the Challenger disaster by dunking one in a glass of ice water at a televised hearing. Famous, among scientists, for asking questions until the actual mechanism showed up.

He said later that his father's lesson was "the difference between knowing the name of something and knowing something."¹

I'm telling you this story because it's exactly what this article is about. Everything that follows is a version of what his father said in the yard. The bird is still outside the window. You can know its name — or you can watch what it does on a Tuesday morning in April: why it chose that particular branch, how it reads magnetic fields it can feel but you can't see.

There is a bird outside your window right now, probably. You have an AI in your pocket that can answer any question you ask it. Those two facts are not unrelated.

Your afternoon, your AI, your question#

You're twelve. Or close. You have a phone, and you used it today — homework, maybe, or a YouTube rabbit hole, or a conversation with ChatGPT or Copilot or whatever the tool is called in your version of 2026.

Here's something you probably already notice: there's a way of using it that leaves you with a finished assignment and not much else. You asked a question, it answered, you wrote something down, you closed the tab. Done. And there's a different way of using it where you come out forty minutes later with three new questions you didn't have when you started.

You already notice the difference. You might not have named it yet.

This article is going to name it. Not because you don't know it — you do, in the way you know things you can feel but haven't said out loud yet. I'm giving you a handle for something you already sense.

That's the whole move. I'll show you what I'm doing at each step, including where I'm guessing and where I'm not. You can decide what to make of it.

Two things, in order: what the window is, and what the question is.

What school does (and fails to do) with your years#

Here is the part where I have to talk about school. I'll make it fast, because you already know most of it.

Between kindergarten and graduation, you will spend roughly 14,000 hours inside an institution that is good at many things. It teaches you facts. It teaches you to sit still and meet deadlines. Some of those facts will matter, and meeting deadlines is genuinely useful. I'm not going to tell you school is evil. It isn't. What it is, structurally, is bad at one specific thing.

It is bad at protecting the part of you that wants to know things — during the exact years when that part is under the most pressure from everywhere else.

Here's the evidence. A Gallup survey of 2,317 K-12 students across the United States found that fewer than one in five strongly agree their schoolwork is important, interesting, challenging, or aligned with their talents.² Fewer than one in five. A separate Gallup survey — just Iowa, 962 students from fifth through twelfth grade — found that only 10% strongly agree they enjoy their classes, and about a third say they always feel bored.²

Quick note on those numbers: the 34% and 10% are Iowa-only. One state. Iowa is not exactly representative of everywhere. I'm using them as texture, not as proof. The national number — fewer than one in five — is the load-bearing one.

Now here's the interesting part: there's a natural experiment running in some schools that tells us what happens when the curiosity engine is protected instead of overloaded. A 2023 systematic review of 32 rigorous studies on Montessori education found consistent positive effects. Kids in those programs did modestly better on academic measures — math, language — and measurably better on things like executive function and how they felt about their own learning. The effect sizes are modest, not dramatic. (The inner-experience finding carries the most uncertainty of any measure in the review — hold it lightly.) And Montessori research has a real selection-bias problem: parents who choose Montessori schools are unusual, and their kids might have done better regardless. The 2023 review ran sensitivity analyses and found the positive effects held, but the caveat stands.³

The point is: a kid whose curiosity was protected didn't fall behind academically. She gained ground, slightly, and felt better about learning while doing it. You don't have to choose between curiosity and school performance.

Some of what happens in middle school is just growing up. Your brain reorganizes around friends, identity, what matters to you. School doesn't cause that. What school does — or fails to do — is protect the part of you that wanted to know things, during the years when that part has the most competition.

So: what is it failing to protect? What is the thing that erodes if nobody protects it?

Two ways a brain can learn#

Think about a one-year-old picking up language. Nobody sat her down with flash cards. Nobody quizzed her. She heard words thousands of times in live conversation with people who responded to her, and she just absorbed them — built a grammar from the inside out, without being able to state a single rule.

Now think about a thirty-year-old taking a French class. She can learn French. It just works differently. She needs instruction, drills, explicit rules. She studies. She will probably always have an accent she didn't have to work for as a child. The mechanism is genuinely different — researchers call it bottom-up, absorptive learning versus top-down, effortful processing, and you can actually see the difference in how the brain handles the work.⁴

The specific detail worth noticing: a 2010 study by researcher Patricia Kuhl found that infants exposed to a foreign language through live tutors showed significant phonetic learning. Infants exposed to the identical content via video showed no learning at all. Same input, same sounds — but one had a live person responding to the baby, and one didn't.⁴ Live interaction matters in a way that passive watching doesn't. That's a specific, replicated finding, not a theory.

This is about language acquisition, not everything. I'm about to extend it into an analogy, and I want to say that clearly so you know when I'm on solid ground and when I'm drawing a picture that might be wrong.

Here is the analogy: there seem to be two ways of relating to any problem that shows up in front of you. One is absorptive — you're following it because you want to know the answer, the same way a child follows a language because she's trying to talk to people she loves. One is effortful — you're producing a result because someone is going to evaluate you on it, the same way an adult studies vocabulary because there's a test.

The research that makes this more than a guess: a 1973 study by Mark Lepper and colleagues at Stanford took preschoolers who already loved to draw. They offered some of them a gold star for drawing. The ones who expected the star drew less afterward, and enjoyed it less — even though they'd been drawing freely on their own before. Researchers call this the overjustification effect: basically, if you pay someone to do something they already liked, they stop liking it.⁵ The original study used preschoolers ages three to five. The effect has been replicated across hundreds of studies with school-age children.

Important context: it applies most clearly to expected, tangible rewards — like grades — for tasks the child could have been genuinely interested in. Verbal praise doesn't hurt the same way; neither do rewards for tasks nobody found interesting to begin with. So "grades on a topic you could have cared about" is the relevant condition, not "all rewards, always."⁵

Then there's a longitudinal study from 2016 that followed 600 students from ages 11 to 16 and documented what the researchers called a "marked decline in intrinsic motivation during adolescence."⁶ The finding that matters: schools that met students' basic needs — autonomy, feeling competent, feeling like they belonged — had smaller declines. The decline still happened. That's important. Even in the best schools, motivation dropped. The researchers think this is partly just growing up. School doesn't cause the decline. But the school environment predicts how steep it is.

The longitudinal data gets clean around age 11. What happens from 6 to 10, we're inferring from the mechanism studies (Lepper) and the shape of the later data. I'm synthesizing across three research traditions — language acquisition, motivation psychology, and adolescent development. I'm telling you I'm synthesizing, and that any one of these research threads could be weaker than it looks.

Now: the operating-system metaphor. I'm going to call these two modes the absorptive OS and the effortful OS. I want to be clear about what that means and what it doesn't. I'm using "operating system" the way a physicist uses "spring" to describe what electrons do near an atom — it's a picture. The picture does real work. It doesn't own the brain.

What the picture says: somewhere in childhood and early adolescence, you settle into a default way of meeting new information. Absorptive mode says "what is that, and how does it work, and what happens if I pull this piece here." Effortful mode says "what is the answer, and when is it due." Both modes are real. Both can learn things. The question is which one runs by default when you haven't been told which to use.

In sensory systems — how birds learn their songs, how the visual cortex wires up in early childhood — experience during specific windows shapes the circuit permanently.⁷ That's neuroscience, not metaphor. I'm not claiming the same thing happens to your motivational default. What I'm noting is that the brain is not a permanent free-for-all: windows exist. The evidence from motivation research points at a similar pattern, even if the mechanism is different and less well-understood.

The absorptive OS, running: what does it actually look like?

The redstone circuit, and what your brain is doing when you build one#

You've probably built a redstone circuit in Minecraft. If you haven't, you've done something like it — a really specific obstacle course in Roblox, a mod that only you use, a base defense system that took fourteen tries to get right.

Here's what it looks like from the outside: you spend two hours on something that produces nothing — no grade, no credit, no one asked you to. You figure out that the comparator is measuring the fullness of the container behind it, which means if you put the chest in the wrong orientation, the signal never fires. You didn't know the word "comparator" when you started. You didn't know you were building a NAND gate — you've probably never heard the word NAND. But the circuit works. You built it by trying, failing, trying again, noticing what the repeater does when you change the delay.

That's the absorptive OS running. No one graded it. Nobody told you to. The reward was: the circuit did what you wanted, or it didn't. And if it didn't, you could see exactly why and try again.

Researcher Mihaly Csikszentmihalyi spent decades studying the state where four hours feels like twenty minutes and you come out with something you made. He called it "flow." Here's the strange thing his research found: school actually creates the structural conditions for flow — high challenge relative to your skill — more often than most other places in your life.⁸ Yet you're in this state less often in school than in games. The structural conditions alone don't trigger it. The missing piece is that your curiosity has to be pointed at the thing.

A 2019 study documented what happens when Minecraft gets structured access: 118 students in an after-school program showed measurable gains in motivation, problem-solving, creativity, and reading and writing skills. One study can carry only so much. It's exploratory — no control group, voluntary participation, self-reported gains in some areas. The researcher himself, Thierry Karsenti, noted that the gains required planned, purposeful engagement — not aimless button-pressing. His study is useful not because it proves Minecraft is magic, but because it documents what you already know from playing for four hours straight: purposeful, model-building engagement produces something that feels like growth because it is.⁹

The Karsenti study is one study. Its methodology has real limitations. I'm using it because its finding matches what you already know, not because one study settles anything. The thing worth trusting is the match between the evidence and your own experience of what Minecraft feels like when you're actually in it, building something, chasing the mechanism.

This is what learning looks like when it isn't being translated into compliance.

The absorption you've been doing while your parents told you to stop playing is practicing the exact thing this whole article is pointing at. The redstone circuit is real. The state you were in when you built it is real. It has a name, and it's the same thing Feynman's father was pointing at in the yard.

The wider view: one generation, one tool, one question being decided#

Zoom out for one paragraph.

You are one twelve-year-old in one bedroom. But the question in front of you right now — absorptive mode or effortful mode, chasing the mechanism or collecting the label — is a question your entire generation is going to answer, one kid at a time. And the context you're answering it in has a feature no prior generation had: the tool you're holding can follow your question further than you can. A twelve-year-old in 1990 who wanted to understand how redstone worked (or the electrical equivalent) had a library, maybe an encyclopedia, and possibly a teacher who didn't know either. You have a live partner that can answer the mechanism question as fast as you can ask it. That is new.

One paragraph on the labor market, because you've probably heard some version of "AI is going to take the jobs." Here's what we actually know. Yale's Budget Lab has been tracking employment data carefully — as of their October 2025 report covering data through July 2025, more than two and a half years after ChatGPT's release, they find no discernible economy-wide disruption from AI in employment. The headlines haven't shown up in the numbers yet. Yale is cautious about what that means: historical technological transitions take decades to reshape labor markets. We genuinely don't know how this plays out. But here's what doesn't depend on the outcome: whatever work exists in fifteen years, a person whose curiosity engine is intact will adapt to it better than a person whose isn't. That's not a prediction about AI. It's a statement about what the curiosity engine is for.¹⁰

The urgency of this article isn't the labor market. It's the window.

Zoom back in.

This question — which OS runs — is being answered by every twelve-year-old with a phone, one session at a time. Not by their school, not by their government, not by any policy committee. One kid, one chat, one "thanks got it" versus one "but wait, why does it work that way?" That is where the decision lives. It's worth knowing about while you're still inside it.

One question#

One question. That's all this section is.

When you ask an AI something and it gives you an answer, you have a choice. Accept the answer, close the tab. Or ask the mechanism question — the question Feynman's father was asking in the yard.¹ Not "is this answer correct" — you're twelve, you often can't tell, and you don't need to. Just: ask the next one. "But WHY does it work that way?"

You're not auditing the AI. You're redirecting it.

Here's the research context: a 2026 study in Current Psychology by Yurt and Kuşci found that university students who use AI without reflection show reduced critical thinking, mediated through what the researchers call "epistemic laziness" — accepting the AI's answer without pushing on it.¹¹ The sample is university students, not twelve-year-olds. I'm inferring the mechanism starts earlier. I could be wrong. Here's why I think I'm not: a 2025 paper in Educational Theory by Ron Aboodi argues that habitual outsourcing of thinking to AI creates a cumulative disposition, not just a momentary shortcut — how you treat answers when you're twelve may shape how you treat answers when you're thirty.¹² Aboodi's argument is theoretical, not empirical. That means it's a careful philosophical argument, not a controlled study. I'm telling you the difference so you can weigh it yourself.

The research shows that lazy AI use correlates with reduced thinking. It does not yet show that curatorial AI use preserves thinking. That's an asymmetric inference I'm making. Here's how you'd check it: after an AI session, notice whether you walked away with a new question or just a completed assignment. That's your data. If the two-modes distinction matches what you actually experience, the mechanism has some grip on something real. If it doesn't, I'm describing something that isn't happening to you, and you should weight my argument accordingly.

Here's what the move looks like in practice. You ask the AI how a redstone comparator works. It tells you. You could say "Got it, thanks" — and you now own a label. The comparator compares. You could also say: "But why does it compare that way? What is it actually measuring inside the circuit?" Now the AI is a live partner in the question you care about. You're not retrieving an answer. You're chasing the mechanism. Same tool, two modes.

His father was asking it about a bird. You ask it about a redstone comparator, or a math proof, or why the sky on Mars is pink during the day but turns blue at sunset where ours turns red. Same question. Same move.¹

"But why does it work that way?" That's it. That's the whole practical offer of this article. It's barely a technique — it's more like a posture — and that's exactly why it works at twelve. You don't have to know anything about how AI works to ask why. You just have to care enough to ask one more question.

So: what's a question you've been wanting to understand about something you actually care about? Open the AI. Start there. Then ask why about every answer it gives you, and see where you end up.

You are still inside#

Back to you.

The window the article has been describing is not abstract. You are inside it, right now, today. How long it stays open — nobody can give you a precise number, because the mechanism is gradual and the evidence comes from multiple sources that don't perfectly agree on the timeline. What we know: the absorptive default shifts through childhood and adolescence. The shift happens to everyone. The question is how steep the slope is, and what you do while you're still in the wide part.

You are the only person guaranteed to be inside the window for the whole rest of it. Your parents aren't. Your teachers aren't. The AI isn't. You.

The reclaim move isn't something an adult does for you. Your parents can help — they can keep their own curiosity visible, they can ask why questions at the dinner table, they can resist the pull to hand you the answer when what you need is to chase the mechanism a little longer. But the person who decides which mode runs is you. I've been writing this article to you, not to them, for a reason: you're the protagonist. I can give you a mechanism and a question, but I can't give you the years. You have those.

Here's what's evidence and what's inference. I've shown you evidence from motivation psychology and language research and a few specific studies. I've named where the evidence is strong (Lepper's overjustification effect — extensively replicated), where it's weaker (Karsenti's Minecraft study — exploratory, one site), and where I'm making inferences that could be wrong (the habit-formation claim from age twelve to thirty). I've modeled the move I'm recommending, because the alternative — telling you to think critically while writing an article that doesn't — would be exactly what Feynman's father was warning against.

So: the bird outside the window. You remember. It had a name. What else do you know about it?

The bird is still outside the window#

The bird is still there.

You know its name now — or if you don't, it doesn't take much to find it. That was never the question. The question is what it's doing on Tuesday morning at 10:47, why it chose that particular branch, how it finds its way south in November without a map you can see, how it learned the song it sings. Those questions don't close with an answer. They open with every answer.

That's the difference this article has been trying to name.

A label stops. A mechanism opens into three more questions, which open into three more. That's what the absorptive mode is — the mode where you follow the mechanism instead of collecting the label. School gives you labels, mostly, because labels are gradeable and mechanisms are not. The AI can give you either, depending on how you ask. And you are the one asking.

Feynman's father asked it about a bird. You ask it about comparators and chemical reactions and why languages die and how redstone actually carries current. The question is the same question. The window is still open.

Here's the thing you take with you. Every time you ask an AI something, every time a teacher hands you an answer, every time you catch yourself nodding at an explanation — ask one more. "But why does it work that way?" That is the whole article, compressed into five words. That is what I wanted you to have.

The bird is still there. Your afternoon is still open.

References#

1. Feynman, R. P. (as told to R. Leighton). (1988). What Do You Care What Other People Think?: Further Adventures of a Curious Character​. W. W. Norton & Company. "The Making of a Scientist," pp. 13–14 (per secondary source). ↩︎ ↩︎ ↩︎ ↩︎

2. Gallup. (2024). K-12 schools struggle to engage Gen Z students. Gallup News​. National survey conducted April 26–May 9, 2024, with n=2,317 K-12 students. https://news.gallup.com/poll/648896/schools-struggle-engage-gen-students.aspx ↩︎ ↩︎

3. Randolph, J. J., et al. (2023). Montessori education's impact on academic and nonacademic outcomes: A systematic review. Campbell Systematic Reviews​, 19(3), e1330. https://doi.org/10.1002/cl2.1330 https://pmc.ncbi.nlm.nih.gov/articles/PMC10406168/ ↩︎

4. White, E. J., Hutka, S. A., Williams, L. J., & Moreno, S. (2013). Learning, neural plasticity and sensitive periods: implications for language acquisition, music training and transfer across the lifespan. Frontiers in Systems Neuroscience​, 7, 90. https://pmc.ncbi.nlm.nih.gov/articles/PMC3834520/ ↩︎ ↩︎

5. Lepper, M. R., Greene, D., & Nisbett, R. E. (1973). Undermining children's intrinsic interest with extrinsic reward: A test of the "overjustification" hypothesis. Journal of Personality and Social Psychology​, 28(1), 129–137. https://psycnet.apa.org/record/1974-10497-001 ↩︎ ↩︎

6. Gnambs, T., & Hanfstingl, B. (2016). The decline of academic motivation during adolescence: An accelerated longitudinal cohort analysis on the effect of psychological need satisfaction. Educational Psychology​, 36(9), 1691–1705. https://doi.org/10.1080/01443410.2015.1113236 ↩︎

7. Knudsen, E. I. (2004). Sensitive periods in the development of the brain and behavior. Journal of Cognitive Neuroscience​, 16(8), 1412–1425. https://doi.org/10.1162/0898929042304796 ↩︎

8. Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience​. Harper & Row. See also: Csikszentmihalyi, M., & Larson, R. (1984). Being Adolescent​. Basic Books. ↩︎

9. Karsenti, T. (2019). Minecraft can increase problem solving, collaboration and learning — yes, at school. The Conversation​. https://theconversation.com/minecraft-can-increase-problem-solving-collaboration-and-learning-yes-at-school-113335 ↩︎

10. The Budget Lab at Yale. (2025, October 1). Evaluating the impact of AI on the labor market: Current state of affairs. https://budgetlab.yale.edu/research/evaluating-impact-ai-labor-market-current-state-affairs ↩︎

11. Yurt, E., & Kuşci, I. (2026). Factors influencing critical thinking during AI use among university students: the mediating effects of epistemic laziness and metacognitive weakness. Current Psychology​, 45. https://doi.org/10.1007/s12144-025-08800-0 ↩︎

12. Aboodi, R. (2025). The worrisome potential of outsourcing critical thinking to artificial intelligence. Educational Theory​, 75, 626–645. https://doi.org/10.1111/edth.70037 ↩︎

Iowa figures: Gallup / Opportunity Education Foundation. (2025, June 3). How student agency can boost engagement and readiness. Gallup News​. Iowa survey conducted September 23–November 18, 2024, with n=962 students (grades 5–12). https://news.gallup.com/poll/660503/student-agency-boost-engagement-readiness.aspx

Supporting: Kuhl, P. K. (2010). Brain mechanisms in early language acquisition. Neuron​, 67(5), 713–727. https://doi.org/10.1016/j.neuron.2010.08.038

See also: Smith, Z. R., et al. (2023). Academic motivation decreases across adolescence for youth with and without ADHD. Journal of Child Psychology and Psychiatry​. https://doi.org/10.1111/jcpp.13815

Supporting: García-Álvarez, J., & Acevedo-Borrega, J. (2025). Minecraft as a pedagogical tool: A systematic literature review (2014–2024). Journal of Educational Computing Research​. https://journals.sagepub.com/doi/10.1177/15554120251341034

Further Reading#

• Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience​. Harper & Row. — The source material on flow state: what it feels like, when it happens, why school conditions don't reliably produce it even when the structural prerequisites are there. If the Minecraft section resonated, this is where to go next.
• Deci, E. L., & Ryan, R. M. (2000). The "what" and "why" of goal pursuits: Human needs and the self-determination of behavior. Psychological Inquiry​, 11(4), 227–268. — The research tradition behind the autonomy-competence-relatedness framework used throughout §3 and §4. More technical than this article but readable for a motivated adult.
• Feynman, R. P. (1985). Surely You're Joking, Mr. Feynman! W. W. Norton & Company. — The companion volume to the bird-story book. If you want more of how Feynman's curiosity operated in practice — how he picked locks at Los Alamos, learned to draw, played bongo drums — this is the one.
• Blakemore, S-J. (2018). Inventing Ourselves: The Secret Life of the Teenage Brain​. PublicAffairs. — The counterargument this article had to absorb: why adolescent motivation decline is partly just growing up, not only school. Blakemore writes for the general reader and takes the teenage brain seriously.
• Bowen, E., et al. (2025). Building AI Literacy at Home: How Families Navigate Children's Self-Directed Learning with AI. arXiv 2510.24070. — The study that found primary-school children can't yet independently evaluate AI outputs — which is exactly why this article lowered the bar to "ask the next question." Worth reading if you want the other side of the AI-and-children argument.