本文来源:https://mp.weixin.qq.com/s/ozqvyhPl83Ub77npvLnEAg
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

前段时间,我发现一个很严重的问题。
我每天超高强度的用 AI 输出工作、生活上的所有内容,但是我却几乎没有任何输入。
没有输入这件事,对我来说,是会让我焦虑的一个问题,因为这个代表着我只是一味的释放价值,而没有持续进步。
所以我决定,必须有计划的重启我的学习计划。
目前已经持续了一周,给大家看看我学习的内容。以上的这些领域,基本上我已经达到可以和专业的人交谈的状态。
07-16: AI外贸中市场调研与客户开发的方法
07-15: 相关性与因果关系的思维关联分析
07-14: Anthropic Skills 核心经验
07-13: RAG 所有变体及其深入研究方向
07-12: 非理性决策的心理学成因
而且经过这段时间我越来越确定,
干活省的时间是线性的,省半小时就是半小时。学东西的速度是复利的,持续会带来个人的巨大进步。
而学习速度的差距,不在于有没有用 AI,而在学习有没有结构。所以我也整理了一套用 AI 闭环(采用 Claude、Codex、Workbuddy 等等都可以)的学习方法。这里一并整理好给大家。
这套方法的来源有两个:
一个是 X 博主 Rahul(@sairahul1)的帖子:用 Claude 十倍速学习任何知识。

Rahul 帖子里最重要的是这句:
大多数人只是随意提问,感觉在学习,一周后什么都不记得。因为真正的学习需要四个要素,一条路径,一次测试,一次压缩,一个反馈循环。
另一个就是 STORM 学习法,
这是斯坦福论文提出的,发在国际顶会 NAACL 上,干的就是一件事:
论文里说一份博士级调研,人工要 40 到 60 个小时读资料、理线索,它用检索加多视角提问,把这个过程压缩成几分钟。
效果大幅超过了传统方法。

这个核心的思路比较重要,我简述一下:
我们大多数情况下问 AI“给我讲讲 X”,得到的永远是大多数人的观点,最常见的框架,最表面的信息。
可一个领域真正的样子,从来不长在大多数人的观点里。而是,天天上手干活的实践者、读遍论文的学者、觉得主流方向全错了的怀疑者、盯着钱往哪流的经济学家、见过类似剧情怎么收场的历史学家。
这五种人看同一个东西,看到的根本不是同一个东西。
而这些视角的差异,才是学习的核心。
我最近的学习,根据自己的经验融合了以上两套方法,从学习的闭环逻辑整理了从学习、深化、记忆、回顾的整个流程。
一共十步,每步一段提示词,按顺序执行。
如果你也想尝试一下,欢迎参与学习小组~
现在开始。
第 1 步 · 五视角 STORM
跑完这一步,你会拿到五份完全不同的解读。
实践者告诉你学术界脱离实际的地方,怀疑者反手拆实践者的台,经济学家把大家不好意思说的利益关系摆到桌面上,历史学家告诉你,这个问题以前上演过,后来是怎么收场的。

第 2 步 · 矛盾图谱
五个视角拿到手,下一步是让它们互相质疑。
这一步大多数人会跳过,但它恰恰是浅调研和深调研的分水岭。
一个领域里比较深的认知,往往就藏在几拨人吵得最凶的那个点上,我发现他们为什么吵搞明白了,理解一下子就下去一层。
这一步跑完有两个特别值钱的产出。
所有视角都同意的东西,大概率是真的,连对手都承认的事,你可以放心用。
所有视角都没提的东西,很可能是整个领域的盲区,有时候这反而是最大的发现。

第 3 步 · 综合简报
然后把前面所有东西,收拢成一份能用的简报。
这份简报的好处是,它不是任何一个单一视角能写出来的。角度全,矛盾会明确出来,每条结论的可信度也标了。

第 4 步 · 同行评审自检
下一步,很多人想不到:让 AI 给自己挑刺。
斯坦福团队自己也承认,STORM 有个已知短板,就是缺自我批判的环节,容易出现来源偏差、事实错配。
所以这里我增加了一步进行评审的过程。
前四步跑完,一个主题的图就开出来了。花的时间,大概也就是泡一壶茶的功夫。而它替你走过的,就是原来要在文献堆里泡几十个小时才理得出来的那些线索。

第 5 步 · 资源筛选
这时候就需要进行广度上的操作了。
资料收集。
这个我太熟了。收集、整理、吃灰,一气呵成!
收藏那一下是真快乐,感觉知识已经进脑子了。
必须解决这个问题。
Rahul 有一段提示词就是解决这个的。
他的说法我很认同:大多数学习者手里的资源早就够了,多到根本用不过来,时间全花在收集上,从来没真用过。
让它只给你 5 个,排好顺序,告诉你哪些是被高估的,再排一条一周走完的路。够了。
5 个能真用完就已经很不错了,收藏夹里那 500 个,就先让它们躺着吧。
到这儿,进门的事说完了。资源定了,基础有了,接下来才是这套闭环真正发力的地方:吃透它。
这里说一个我的认知:学新东西,最常见的问题就是跳级学。
基础没打牢就去啃高级材料,啃不动,觉得自己笨,放弃。其实多半就是顺序错了,跟聪明不聪明没多大关系。
教育心理学里有个概念我一直很喜欢,叫最近发展区,维果茨基提的。
大意是,一个人能学会的东西,永远长在“他已经会的”旁边一小步,你踮踮脚能够着的东西才学得会,隔太远的,看一百遍也就是个眼熟。

第 6 步 · 学习阶梯
这一步的作用就一个:让你永远知道自己站在第几级,下一步该到哪。
我自己的体会是,学东西最泄气的时候,往往就是不知道自己到哪儿了、还差多远的时候,有张阶梯在手里,这个问题就基本没了。
第 7 步 · 2 小时啃下核心 20%
阶梯排出来了,别急着从第一级慢慢爬。
大多数学科都有那么一小撮概念,撬动其他所有内容。先找到那 20%。
10 次课,每次都有练习、有资源、有复习题,最后还有一个能证明你真的会用的小项目。
2 小时当然学不完一个领域,这个计划的目标,就是把“能上手干活”的那部分先拿下。
到这儿为止,都还是舒服的部分。
接下来这一步,是最不舒服的,也是我觉得最重要的。
看书、看视频、听 AI 讲,这些输入型的动作有个共同的毛病:它们让你“感觉”自己懂了。
那种懂来得太顺了,顺到你根本不会去怀疑它,到底是真懂了还是只看了个热闹,得真考一下才知道。
第 8 步 · 考到我崩溃
这一步,我劝你挑个心情好的时候用。被它考到答不上来的滋味不太好受,但也只有那一下,你才知道自己之前的“懂”里,有多少是水分。
这背后其实是认知心理学里被验证过无数遍的一个效应,叫测试效应,英文叫 retrieval practice,检索练习。
我觉得很真实的是,从脑子里往外掏知识这个动作本身,比往脑子里塞知识,更能让知识记住。(费曼学习法也是这个逻辑)
你肯定有过考前抱佛脚、第二天反而记得特别牢的经历,就是这个机制,被这么考一遍,本身就是在学,不光是在测你到底会多少。

第 9 步 · 费曼循环
考出来的窟窿,用费曼循环补。
费曼那句话大家都听过:一个东西你没法讲给 12 岁的孩子听懂,你就还没真的懂它。
费曼技巧就四步:学习,教授,回顾,简化。学一遍,讲一遍,找出讲不清的地方回炉,再用更简单的话讲一遍。
这部分就不多解释了,必须认真执行的一步。

第 10 步 · 一页速查表
最后一步最轻,但我用的频率可能反而最高。
人的脑子,记结构比记段落牢。学完的东西不压缩,就跟没打包的行李一样,要用的时候,满地找。
面试前、开会前、要给别人讲之前,掏出来花 5 分钟扫一遍。就这么个小东西,特别好使。

十步,全给你了。
回头看一眼,它其实是一条完整的线:
1. 对于一个陌生领域,让观点打一架,压成简报,再让它自查一遍;
2. 然后只留 5 个资源,排出阶梯,用 2 小时先啃下核心的 20%;
3. 学完了,让考官把自己考到崩溃,讲不清的地方用费曼循环磨平;
4. 最后,压成一页速查表,随用随取。
一个闭环。
劝人开始学习这件事,真的不讨好。
我知道评论区一定又会出现:先收藏,吃灰去吧。
不过,我在查 STORM 资料的时候,我刷到过一篇转述它的文章,结尾说接下来 18 个月是关键窗口期,先掌握的人认知优势会被无限拉大。
这种话,我从来不太买账。方法好就是好,不需要拿“再不学就晚了”来吓人。学东西这事,什么时候开始都不晚。
晚的只有一种:一直收藏,一直不开始。
输出的活,AI 替我们干得够多了。也该轮到输入了。
接下来我会让自己先坚持 21 天的周期,如果你也想,欢迎加入一起 🎉
以上,既然看到这里了,如果觉得不错,随手点个赞、在看、转发三连吧,如果想第一时间收到推送,也可以给我个星标⭐~谢谢你看我的文章
你的关注是我持续更新的动力~
我是谁
我是 AI大刘,北大毕业,大模型研究方向,腾讯犀牛鸟,先后在腾讯、百度的大模型研发部门。现在斜杠青年:给多家国企做AI顾问 \ OPC \ 研究员 \ 产品独立开发者 \ Vibe Coder
Source: https://mp.weixin.qq.com/s/ozqvyhPl83Ub77npvLnEAg If there is a copyright issue, please email: 13951431913#139.com or submit a comment on the right.
Hello, this is Da Liu's 70th original dry goods article〜
Recently, I discovered a very serious issue.
I use AI at a high intensity every day to output content for work and life, but I have almost no input.
The lack of input is a source of anxiety for me, as it means I am only releasing value without continuous improvement.
So I decided to systematically restart my learning plan.
I have been sticking to it for a week now, and here is what I have been studying. In these fields, I have reached the level where I can converse with professionals.
07-16: Methods for Market Research and Client Development in AI Foreign Trade
07-15: Analysis of the Relationship between Correlation and Causality
07-14: Core Experience of Anthropic Skills
07-13: All RAG Variants and their In-depth Research Directions
07-12: Psychological Causes of Irrational Decision-Making
And through this period, I am increasingly convinced that
Saving time on work is linear — saving half an hour means exactly half an hour. But learning is exponential — consistent study leads to massive personal progress.
The gap in learning speed is not about whether you use AI, but whether your learning has structure. So I have compiled a set of AI-powered closed-loop learning methods (using Claude, Codex, Workbuddy, etc.) to share with you.
This method is inspired by two sources:
One is X-blogger Rahul (@sairahul1)’s post: “Learn any knowledge 10x faster with Claude.”
Rahul points out something critical:
Most people just ask random questions and feel like they’re learning, but after a week, they remember nothing. True learning requires four elements: a path, a test, compression, and a feedback loop.
The second source is the STORM learning method,
proposed in a Stanford paper and published at NAACL. Its goal is to compress a 40–60 hour doctoral-level literature review into minutes using retrieval and multi-perspective questioning.
The results significantly outperform traditional methods.
When we ask AI “Tell me about X,” we always get the majority view, the most common framework, and the most superficial information.
But the true nature of a field is never captured by the majority. It’s seen by practitioners who work on it daily and know where academia diverges from reality; by skeptics who think the mainstream direction is wrong; by economists who follow the money; and by historians who know how these stories end.
The difference between these perspectives is where the real learning happens.
Based on my experience, I have integrated these two approaches into a learning workflow covering acquisition, deepening, memory, and review.
There are ten steps, each with its own prompt, executed in sequence.
If you’d like to try this, feel free to join the learning group〜
Let’s begin.
Step 1 · Five-Perspective STORM
● ● ●
After this step, you’ll have five distinct interpretations of the topic.
Practitioners will tell you where theory fails in practice; skeptics will challenge practitioners’ assumptions; economists will reveal the hidden interests; and historians will show you what happened the last time this was tried.
Step 2 · Contradiction Mapping
Most people skip this, but it’s the difference between shallow and deep research.
The deepest insights in a field usually lie in the most heated debates. Understanding why they disagree deepens your comprehension immediately.
● ● ●
This step produces two valuable outcomes:
What everyone agrees on is likely true — you can trust it.
What no one mentions may be the field’s blind spot — sometimes that’s the biggest discovery.
Step 3 · Unified Brief
● ● ●
This brief is unique because it captures the richness of multiple viewpoints while highlighting the contradictions and annotating the reliability of each conclusion.
Step 4 · Peer Review Self-Check
The Stanford team acknowledges that STORM lacks a self-critical component, making it prone to source bias or factual errors.
So I’ve added a review step to validate the output.
● ● ●
After these four steps, you have a clear visualization of the topic — achieved in the time it takes to brew a cup of tea, yet capturing insights that would take dozens of hours to extract from raw data.
Step 5 · Resource Curation
Data collection is my weak point. I gather, organize, and never use.
To fix this, I use a prompt from Rahul that forces the system to pick only 5 essential resources.
● ● ●
You get 5 resources, ranked, with notes on what’s overrated and a 1-week reading plan. That’s enough.
Finishing 5 resources is a win; keeping 500 in your library is procrastinating.
Now that the door is opened, the real work begins: mastering the material.
A common pitfall in learning is “skipping grades.”
We jump to advanced materials without solidifying the basics, fail, and assume we’re dumb. In reality, it’s often just poor sequencing.
Vygotsky’s concept of the “zone of proximal development” is key here: people learn best at the edge of their current abilities — not far beyond them.
Step 6 · Learning Ladder
● ● ●
This step ensures you always know where you are and what’s next.
There’s nothing more demoralizing than feeling lost in a field. A clear ladder eliminates that uncertainty.
Step 7 · 2 Hours to Core 20%
Once the ladder is built, don’t start at the bottom.
● ● ●
10 lessons, with exercises, resources, and a final project to prove mastery — all within 2 hours.
This isn’t about mastering the whole field, but about reaching the point where you can actually start working.
So far, it’s all been smooth sailing. But the next step is the most uncomfortable — and critical.
Reading and listening make us feel smart. But to know if we’re actually learning, we need to be tested.
Step 8 · Test Until Failure
● ● ●
Pick a bright day for this one. Being wrong hurts. But it’s the only way to separate true understanding from shallow awareness.
This leverages the “testing effect” (retrieval practice), which shows that recalling information strengthens memory far more than rereading does.
Step 9 · Feynman Cycle
When the test reveals gaps, use the Feynman technique to fill them.
Feynman’s mantra — “If you can’t explain it to a 12-year-old, you don’t understand it” — is the cornerstone here.
● ● ●
Study, teach, review, simplify. Repeat until the concept sinks in.
Step 10 · One-Page Cheat Sheet
The final step is the lightest, but likely the most used.
Uncompressed knowledge is like unorganized baggage — impossible to find what you need.
● ● ●
That’s the full loop.
Looking back, it’s a coherent journey:
1. Gather diverse views, synthesize them, self-check for errors;
2. Curate 5 resources, build a ladder, tackle the core 20%;
3. Test relentlessly, then loop back with Feynman to fix weaknesses;
This is a full feedback loop designed to transform reading into mastery.
Encouraging people to start learning is tough. The immediate response is often, “I’ll bookmark this and read it later.”
I recently read that the next 18 months are critical — those who master these tools now will gain an insurmountable cognitive advantage. But that’s fear-mongering. The truth is, you can start learning at any time — the only thing holding you back is inaction.
AI has taken over our output. Now it’s time to feed the brain.
I’ll stick to this for 21 days. If you want to join, let’s do it together.
Keep learning — it’s the only way to stay ahead. Cheers!
Looking back, it's actually a complete pipeline:
1. For an unfamiliar field, let the viewpoints battle it out, compress them into a brief, and then make them self-check once;
2. Then keep only 5 resources, arrange them in a ladder, and use 2 hours to first chew through the core 20%;
3. After learning, let the examiner test you until you collapse, and use the Feynman cycle to smooth out the unclear parts;
4. Finally, compress it into a one-page quick reference sheet for instant retrieval.
一个闭环。
Persuading people to start learning is really not flattering.
I know the comment section will definitely say again: "Collect it first, let it gather dust."
However, while checking the STORM data, I came across an article that quoted it, saying that the next 18 months are a critical window period, and those who master it first will have an infinite cognitive advantage.
I've never bought into such talk. If the method is good, it's good; there's no need to scare people with "it's too late if you don't learn now." Learning is never too late — no matter when you start.
The only thing that is truly "late" is collecting forever and never starting.
AI has done enough of the output work for us. Now it's our turn to focus on input.
In the next step, I'll commit to a 21-day cycle. If you're interested, feel free to join me 🎉
Since you've read this far, if you like it, a quick like, share, or star ⭐ would be appreciated. Thanks for reading my article!
Your attention is the driving force behind my continuous updates.
我是谁
I'm AI 大刘 (Daliu), graduated from Peking University, specializing in large model research. I worked at Tencent and Baidu in their large model R&D departments. Now, I'm a slashie: an AI consultant for several state-owned enterprises, an OPC, a researcher, an independent product developer, and a Vibe Coder.