Lessons from LLMs: Output More
In the wave of technological breakthroughs led by large language models, I've found some surprisingly personal lessons for how we learn.
For example — "learn from your future self," as Tiancheng Lou once put it, maps neatly onto reinforcement learning.
Here's one idea I've decided to put into practice.
Output more, let knowledge crystallize
I recently came across an idea that stuck with me: output more — don't just listen or read.
Knowledge only solidifies when you produce something with it. It's like the encoding process in a model — it requires deliberate training. This is essentially the Feynman technique: you only truly understand something when you can explain it simply enough for someone else to get it. It's a form of knowledge compression.
The "flow" that comes from output
This isn't just about deepening understanding.
Looking back at my social experiences, aside from the occasional joke that brings a moment of joy, what truly puts me in a flow state is when I can systematically articulate my understanding of something.
The structured content I've produced has almost always come from topics I explored deeply out of pure curiosity, then wrote up as a form of self-entertainment.
On many other topics, despite having consumed plenty of information, when someone actually brings them up in conversation, I find myself mumbling "I haven't really looked into that" — unable to offer what I'd consider a "high-quality" take.
It's a bit like scrolling through short videos for hours and retaining absolutely nothing.
AI conversations as "external memory"
Beyond that, I frequently have long, iterative conversations with various AIs — mostly ChatGPT and DeepSeek.
Some of these exchanges are genuinely interesting. If I take the time to write them down, they become a record of my thinking and my life — a form of context engineering for my own existence.
The barrier to writing has never been lower
Chatbots have become an indispensable part of my life. They've also drastically lowered the barrier to writing.
If writing brings me joy, why not do it?
Posting essays on WeChat Moments feels a bit pretentious (though honestly, nobody cares). So I've decided to revive my long-dormant public account and write about the questions that occupy my mind.
Below is the original Chinese draft, written by myself. The English above is a translated and polished version.
大语言模型带来的启示——多输出
在以大语言模型为代表的新一波科技浪潮中，我发现它对个人学习有一些启发。
比如——向未来的自己学习（楼天成提到过） → 类比到 reinforcement learning。
多输出，让知识沉淀
最近我听到一个观点，决定实践一下：多输出，而不是只是听或看。只有输出，知识才会沉淀下来，就像模型的 encode 过程，需要刻意训练。这很像"费曼学习法"——只有当你能简洁的讲到让别人听懂，才说明自己真正掌握了。也是一种知识的压缩。
输出带来的"心流"
这不仅仅是为了加深理解。回顾我的一些社交体验，除了偶尔的插科打诨带来些快乐之外，真正让我进入"心流"状态的，是我能系统地把自己的理解简述出来的时候。这些有条理的内容，基本都是我凭兴趣做了深入研究，然后写出来当作自我娱乐。而在很多其他话题上，虽然听过不少信息，但真有人讨论时，我只能支支吾吾地说"没怎么关注过"，很难给出自己认可的"高质量"观点。这有点像刷短视频，看了半天，脑子里却什么也没留下。
AI 对话是我的"外部记忆"
除此之外，我经常与各类 AI（主要是 ChatGPT、DeepSeek）进行长时间、反复的对话。有些交流其实挺有意思，如果能花点时间写下来，也算是对生活和思考的记录，更是我人生中某种意义上的 "context engineering"。
写作门槛已经很低了
现在各种 chatbot 已经成了我生活中不可或缺的部分，也极大降低了写作的门槛。既然写东西能取悦自己，何乐而不为？朋友圈发这种随笔略显矫情（虽然其实也没人 care），那我就找回沉寂已久的公众号，写一写生活中我关注的问题。