We study how minds end, not how they start
The edges of consciousness attract enormous attention. Near-death experiences, deep meditative states, psychedelics, the last flickers of activity in a dying brain — real money and real careers go into all of it. We want to know what thought looks like as it comes apart.
The other end of the spectrum gets almost nothing.
Somewhere in your house or your neighbor's, a five-year-old is running the fastest cognitive build-out that will ever happen in a human life. Children between roughly three and eight observe, reason, and generate theories about how the world works at a rate no adult can match, and they do it without much regard for which theories are permitted. Ask a six-year-old where the dark goes when you switch the light on and you may get an answer involving somewhere it is stored. It is wrong. It is also a question about the ontology of absence, asked by someone who has not been told that absence is not a substance, and adults do not arrive at it because we were trained out of the premise before we could use it.
Most of what children produce is wrong in ordinary ways. A small fraction is wrong in a way no trained adult would ever be wrong. We call it cute, we repeat it at dinner, and we let it go.
Why this matters more than it used to
The structure of human knowledge changed shape over the last few years. The internet became our collective memory, and large models became something like a shared processing layer on top of it, having absorbed most of the frameworks, patterns, and lines of argument our species has written down.
That produces a problem people in the field have been circling for a while. As model output flows back onto the web and into the next round of training data, the system begins to feed on itself. Researchers have induced real degradation in that setup, and although how much this matters in practice is still unsettled, the direction is not comforting. These systems are extraordinary at recombining what exists. Producing something genuinely outside the distribution they were trained on is a different problem, and nobody has solved it.
The obvious answer is that humans supply the novelty. But adults are a weaker source than that answer assumes. Two decades of schooling, professional training, and cultural convention leave us with remarkably similar priors. We argue loudly about conclusions and agree almost entirely on what counts as a sensible question. In a sense we have all been fine-tuned on the same dataset.
So where is the next batch of structurally different thinking supposed to come from?
Cognitive cord blood
When a child is born, some parents pay to bank the umbilical cord blood. It contains stem cells in a state the body will never produce again: undifferentiated, valuable precisely because they have not yet committed to being anything in particular.
Something similar is true of a young mind, and we discard all of it.
The metaphor can be pushed too far, so let me mark where it breaks. A child's thinking is not pure. Children absorb language, family, culture, and the assumptions of everyone around them from the first week of life — anyone who has heard a four-year-old recite a parent's opinion knows this. What they have not yet acquired is the training: the sense of which questions are serious, which combinations are absurd, which doors are not worth trying. That is what gets sanded off between eight and eighteen, and that is the part worth keeping. And unlike cord blood, this is not banked for the child's own later use. It would be pooled, which is a different proposition ethically, and I will come back to it.
There is also a practical reason this is newly possible. A child's stranger hypotheses used to be unusable, because following one up required an adult with relevant expertise and free time — a resource nobody spends on a passing remark at breakfast. That cost has fallen close to zero. A model can take an odd premise seriously, work out what follows from it, and say where it collides with what we know. That is not validation in any strict sense, and it will be confidently wrong sometimes, but it is enough to separate a strange idea from an empty one. The filter that made this impractical is gone.
Children are not raw material for improving AI, and I would not want the argument read that way. But a knowledge system that only eats its own output is in trouble, and the least-trained thinkers among us are the most underused source of anything else.
What to do, and what this proposal does not solve
The shift is from imparting to capturing. Early childhood runs almost entirely in transmission mode: teaching, correcting, aligning children to the existing map. Almost no traffic runs the other way.
For families: treat what your child says as worth recording. A notebook, voice memos, a folder of photographed drawings. The format matters less than doing it at all, and much less than resisting the urge to edit out the parts that sound wrong. The wrong parts are the point. Most of it will amount to nothing — that is also true of cord blood.
For anyone building systems: the binding design constraint is passivity. Collect and preserve without steering. The moment a system starts nudging children toward the ideas it finds interesting, it has destroyed the property that made them worth collecting.
Now the part I cannot resolve. Nobody knows how to turn a shoebox of six-year-old remarks into usable input for anything. Archiving is the cheap half; the conversion is an open problem, and I am not going to pretend otherwise. What makes me think it is still worth doing is the shape of the cost. Recording is nearly free and reversible. Not recording is not: you cannot go back and collect what a child said at five once they are thirty. Almost every archive in history was built by people who did not yet know what it was for.
And the caution belongs here, in the middle of the proposal, not appended to the end. This is a proposal to record the inner life of people who cannot consent, preserved for decades, possibly pooled. That is not an implementation detail. Anything built in this direction has to default to private, stay under family control, and give the child the right to read and delete their own archive when they are old enough to want to. A version of this run by a company harvesting children's speech at scale would be worse than not doing it at all. The value of the idea and the ease of doing it badly come from the same source.
In an age when almost any settled question is answered on demand, the scarce thing is no longer a good answer. It is a question nobody has yet been trained out of asking. Those are being produced around us constantly, by people under four feet tall, and we are writing down almost none of them.
中文版
我们研究心智如何终结，却不研究它如何开始
意识的边缘吸引了巨大的注意力。濒死体验、深度冥想、致幻状态、大脑熄灭前最后的电活动——这些方向上投入了真金白银，也投入了许多严肃的职业生涯。我们很想知道，思维在瓦解时是什么样子。
而光谱的另一端，几乎无人问津。
此刻在你家里，或者邻居家里，有一个五岁的孩子正在进行人一生中速度最快的一轮认知搭建。三到八岁的孩子观察、推理、生成关于世界如何运转的理论，速率是任何成年人都追不上的，而且他们并不太在意哪些理论"被允许"。你问一个六岁的孩子，开灯之后黑暗去哪儿了，你可能会得到一个关于它被收在某处的回答。这个答案是错的。但它同时是一个关于"缺席是否是一种实体"的本体论问题，提出者只是还没有被告知"黑暗不是一种物质"；成年人问不出这个问题，因为我们在能用上这个前提之前，就已经被训练得不再持有它了。
孩子说的大部分东西，错得很普通。但其中有一小部分，错得是任何受过训练的成年人都不可能错的那种方式。我们说一句"好可爱"，晚饭时当段子讲一遍，然后就任它流失。
为什么这件事现在比过去更要紧
过去几年，人类知识的结构发生了形变。互联网成了我们的集体记忆，大模型则像是叠在它之上的一层共享处理层——我们这个物种写下来的思维范式、模式与论证路径，它基本上都吸收了。
由此产生了一个业内已经绕了一段时间的问题。当模型的输出重新流回网络、进入下一轮训练数据，这个系统就开始以自身为食。研究者已经能在这种设定下诱发出真实的性能退化；它在现实中究竟有多要紧仍无定论，但方向并不让人安心。这些系统重组既有事物的能力极其出色，而要产生真正落在训练分布之外的东西，是另一个问题，且无人解决。
顺理成章的回答是：新意由人类提供。但成年人作为来源，比这个回答假定的要弱。二十年的学校教育、专业训练与文化惯例，让我们拥有了高度相似的先验。我们在结论上吵得很凶，在"什么算是一个像样的问题"上却几乎完全一致。某种意义上，我们都是在同一份数据集上微调出来的。
那么，下一批结构上真正不同的思维，该从哪里来？
认知"脐带血"
孩子出生时，一些父母会付费保存脐带血。它含有的干细胞处于身体此后再也不会产生的状态：未分化，而其价值恰恰在于它尚未承诺要成为任何特定的东西。
幼小的心智具有类似的性质，而我们把它全部丢掉了。
这个比喻容易被推得过头，所以我想标出它断裂的地方。孩子的思维并不"纯净"。从出生第一周起，他们就在吸收语言、家庭、文化以及周围所有人的默认假设——听过四岁孩子原样复述父母观点的人都懂。他们尚未获得的是训练：那种关于哪些问题严肃、哪些组合荒谬、哪些门不值得推的判断力。这部分会在八岁到十八岁之间被一点点磨掉，而它正是值得留存的部分。另外，与脐带血不同，这些东西并不是存给孩子本人日后使用的。它会被汇集起来，这在伦理上是完全不同的一件事，后面我会再说。
还有一个务实的理由，说明为什么这件事现在才具备可行性。孩子那些更古怪的假设过去无法处理：跟进任何一个，都需要一位有相关专业能力又恰好有空的成年人——没有人会为早餐桌上的一句随口之言支付这种资源。今天这个成本已接近于零。模型可以认真对待一个古怪的前提，推演它会导向什么，并指出它在哪里与已知的东西相撞。这算不上严格意义上的"验证"，它也会时不时自信地出错，但足以把一个奇特的想法与一个空洞的想法区分开。过去让这件事不切实际的那道过滤器，已经不在了。
孩子不是用来改良 AI 的原材料，我也不希望这篇文章被这样读。但一个只吃自己产出的知识系统是有麻烦的，而我们当中受训练最少的那群思考者，是最被闲置的另一种来源。
具体该做什么，以及这个提议解决不了什么
方向上的转变，是从"灌输"到"捕捉"。幼儿期几乎全部运行在传输模式：教、纠正、把孩子对齐到已有的地图上。反方向的流量几乎为零。
对家庭而言： 把孩子说的话当作值得记录的东西。一个本子、几段语音、一个存着翻拍画作的文件夹。用什么形式不重要，重要的是真的开始记，更重要的是忍住那股把"听起来不对"的部分删掉的冲动。不对的部分才是重点。这些记录里绝大部分最终什么都不是——脐带血也一样。
对做产品和系统的人而言： 最关键的设计约束是被动。只收集和保存，不引导。一旦系统开始把孩子往它觉得有意思的方向上推，它就已经摧毁了让这些内容值得收集的那个属性。
接下来是我解决不了的那部分。没有人知道该怎么把一鞋盒六岁小孩的话，转化成任何东西可用的输入。归档是便宜的那一半；转化是一个开放问题，我不打算假装不是。让我认为仍然值得做的，是成本的形状：记录几乎是免费的，而且可逆；不记录则不然——孩子三十岁时，你无法回头去采集他五岁时说过的话。历史上几乎每一份档案，都是由还不知道它将来有什么用的人建立起来的。
而那条警告属于这里，属于提议的中间，而不是被附在结尾。这是在提议记录一群无法给出同意的人的内心活动，保存数十年，并且可能被汇集。这不是一个实现细节。任何在这个方向上做出来的东西，都必须默认私有、由家庭掌控，并在孩子长大到想要行使权利时，让他能够读取和删除属于自己的那份档案。一个由公司大规模采集儿童语音来运行的版本，比完全不做还要糟。这个想法的价值，和把它做坏的容易程度，来自同一个地方。
在一个几乎任何已有问题都能被随时解答的时代，稀缺的已经不是好答案，而是一个还没有被训练得问不出来的问题。这样的问题每天都在我们周围被大量生产，生产者身高不到一米二，而我们几乎一个都没有记下来。