Editor’s note:
This is Tabbit's "Tinsight · Insights" column, where we will share some in-depth articles worth spending time reading. We have previously shared the article "The Real Divergence in the AI Era" from offbook.press. The new series "Rebuilding Individual Learning" focuses on providing people with a series of practical solutions. The original essay is about 19,000 Chinese characters long, so we published it in six installments. This time it is Chapter 6, which is also the last part. Explore where true learning begins and how to take the first step.
▼
Author: Dawei Geng
Source: https://offbook.press/essays/rebuilding-learning/
Chapter 6 · A starting point: AI programming
The method has been explained, but many readers will be stuck one step ahead of the "method": not knowing where to start.
Or to be more precise, the things they can think of (learning a skill, developing a hobby) are all options that existed in the pre-AI era and are not suitable for this era.
There is one thing - for people with a computer and access to the Internet - the cost is close to zero, and only in the AI era is it truly open to people with zero foundation: using AI to write code to solve the troubles at hand.
This is not included in this chapter because everyone should learn to program. Because it is the most intensive place for all the arguments in the previous five chapters.
▍Wrong start
People with zero foundation want to start AI programming. The path most people take is:
Ask AI "How to get started with vibe coding". Search "How to use cursor for newbies". Go to Xiaohongshu and Station B and swipe a little round head in the lower left corner of the video to talk about how he built a website in three days. Collect a bunch of AI programming tutorials, follow a few bloggers, save dozens of prompt templates, and prepare to "study seriously for a while before starting."
Can't go more than two weeks. It's not because they are lazy, it's because this path is structurally failed.
It is the embodiment of the three modes mentioned in the first chapter in AI programming learning: asking AI "how to get started", reading other people's tutorials, and collecting templates, all used as encyclopedias; you haven't really started to do anything, there are no specific situations, no real situations; "What should I learn next"-style meta-questions, rabbit holes. No step is running the actual judgment loop.
And there is an extra trap here: those videos of "Building a Website in Three Days" look particularly exciting and efficient. The audience feels like they can do it when they watch it, but after watching it they know nothing. Just because you see others doing it, it doesn’t mean you are doing it. This is the specific manifestation of the somatosensory overlap of consumption and learning discussed in Chapter 1 in the field of AI programming.
▍The right start
Four steps.
1. Have an account that can run Claude Code or Codex
Why it must be Claude Code or Codex, because as of early 2026, they are the trinity of the strongest model + the strongest Agent + the strongest code capability you can use. Don't start with a weaker model, and don't spend time worrying about "which tool to choose?": This dilemma itself is the meta-question discussed in Chapter 1, a kind of consumption that looks like serious consumption.
I won’t go into details here about how to install and configure it.
If you really want to get started, "How to install this thing" itself is your first delta question. You ask the AI, and the AI tells you to pretend accordingly.
This step is already running the cycle of Chapter 2: if you have specific needs, ask the AI with the needs, follow the answers, and continue to ask when you encounter problems until it is installed.
If you are too lazy to even ask this step yourself, this chapter is not for you.
This step itself is a filter.
2. Pick an annoying thing that really happened to you
It's not "I want to do a grand project" or "I want to learn crawling", it's something you have really done manually recently and find it annoying every time you do it.
The exported Excel format is messy and needs to be sorted out manually every time.
Cut out certain numbers from dozens of PDFs and fill them in the form
The same message needs to be sent to twenty customers, and the names must be changed each time copy paste
A bunch of photos with messy file names. I want to rename them according to the shooting date.
The more specific, the better. Bring an example of your real data, not an abstract description, but "this is what my file actually looks like."
This is the core of Chapter 2: delta comes from real situations. You don’t need to learn programming first, you need to have something that really annoys you.
The biggest sticking point for newbies here is usually "I can't think of anything." This stuck point itself shows one thing: you may not really be observing your daily life. Any repetitive operations you do manually are candidates. Turn on desktop and mobile and see what you copy and paste, manually organize, and repeat every day. There are more than a dozen candidate tasks in there.
3. Let it write code and run it
Describe the trouble to the AI and let it write the code.
Run. There are usually three situations:
That’s right (less). It started running but the result was wrong. Can't run at all and reports an error.
The second and third situations are where learning really happens.
At this time, don’t think to yourself, “Should I learn the basics before I can understand it?” Directly send the error report, incorrect results, and the phenomenon you observed intact back to AI, and let it continue to make changes.
If you don't understand what it changed, ask it to be explained. But don’t stop just to “understand before you continue.” Let it continue to be modified until this matter is really solved.
In this loop, you will understand little by little what it is doing. It's not because you're learning grammar, it's because the same things appear repeatedly in front of you, and you naturally start to recognize them. This is what I talk about at the end of Chapter 2. The necessary basis is known after the fact, not beforehand.
4. Solved, let’s pick the next annoying thing
No need to review. There is no need to summarize "What did I learn this time?" There is no need to create a “learning roadmap”.
You solved a real problem. The next annoyance is where the next round begins.
When you do the fifth and tenth items, you will find:
You begin to vaguely understand what the code is doing
You can begin to see where there may be problems with the AI writing
You start to have your own adjustment ideas instead of just accepting the version given by AI
Your "annoyance list" has grown longer without realizing it, because you start to see more things that should have been automated.
This is the growing ability. It is not grown through "learning", but grown through "doing".
The true value of this matter
The most important thing is this sentence:
The real value of AI programming to a person is not to let him know how to program, but to have a mirror with extremely clear feedback.
Computer feedback is the purest form of feedback you come into contact with every day. Whether the code is correct or not, you will know as soon as you run it; whether the logic is strict or not, bugs will be exposed immediately; whether your assumptions are correct, the reality will tell you within a few seconds.
This kind of feedback clarity is not available in most everyday areas. Feedback from interpersonal relationships is vague, delayed, and full of noise; feedback from workplace judgments takes months to see; feedback from business decisions is mixed with countless external factors that have nothing to do with your judgment.
Programming is different. Every time you judge, verify, get slapped in the face, and correct, the whole cycle is completed within a few minutes.
A person who is serious about AI programming, even if he only writes some small scripts, will accumulate more "hypothesis-verification-correction" cycles in a year than in most other fields. The sum of these cycles is the training volume of judgment.
So the real ROI of AI programming is not "I can write programs", but "I have a place to continuously train my judgment."
This goes back to the core of Chapter 4: the value of the mirror is not in the field itself, but in its use to illuminate other fields. The reflex of "assume first, let reality hit you in the face, and make quick corrections" that you have been trained in programming will transfer to many other areas of your life and work.
If you don't know where to start after reading Issue 1 and Issue 2 - find an environment that can run Claude Code, pick a troublesome thing, and let it write code.
Chapter 7·Finally
No matter which specific thing you start with, there is actually only one thing discussed in this article:
When all the old learning methods are failing, how should real learning happen?
The answer is not a new methodology. The answer is to go back to the oldest essence of learning.
What is learning? Learning is to continuously expose a person's judgment system to feedback from reality so that it can be continuously calibrated. This definition has been true since the time of Socrates, it has been true before the Industrial Revolution, and it has been true before the Internet. Real learning is never about "how much you read," "how much you know," or "how much you remember." It is whether your judgment is really polished by reality.
In the pre-AI era, most learning activities deviated from this essence and became the consumption of information, the hoarding of knowledge, and the pursuit of "appearing knowledgeable." These activities have certain external rewards (social status, career threshold, social capital), so they can still be maintained if they deviate from the essence.
The AI era has emptied these external rewards. Knowledge is no longer scarce, information no longer has differentiated value, and no one pays for "appearing knowledgeable". All pretense of deviating learning activities is stripped away, revealing them to be worthless.
The only things left that can maintain value are those activities that truly return to the essence of learning:
Do something real and let your judgment be calibrated
Observe the feedback on this and make the calibration really happen
Use AI to help you ask delta questions, review your judgments, organize your experience, and illuminate new areas, making this process dozens of times more efficient than in the past.
This is not methodology. This is not “learning skills in the AI era.” This is a re-understanding of the word "learning": restoring it from the pre-AI era version surrounded by various disguises to its original appearance.
Then let the AI accelerate the original thing instead of accelerating the disguise.
The last self-test is more lethal than any other method:
What new judgments have you made in the past three months that were formed as a result of your conversations with AI? Is this judgment specific, verifiable, and will it change a specific action of yours?
Yes, this system has already started working on you
Can’t answer the question? You may have used a lot of AI, but real learning hasn’t started yet.
Learning in the AI era is not about "how much AI is used", but about how much you change after using the AI.
Quotes and sources
The source of the main factual assertions involved in this article.
About curiosity, information acquisition and the brain’s reward system (Chapter 1)
Bromberg-Martin, E. S., & Hikosaka, O. (2009). "Midbrain dopamine neurons signal preference for advance information about upcoming rewards." Neuron, 63(1), 119-126. This study found in monkeys that midbrain dopamine neurons respond to "opportunities to obtain information," even if the information has no instrumental value
Gruber, M. J., Gelman, B. D., & Ranganath, C. (2014). "States of curiosity modulate hippocampus-dependent learning via the dopaminergic circuit." Neuron, 84(2), 486-496. Human fMRI study found that VTA/SN (dopamine midbrain) and NAcc (nucleus accumbens) activation pattern is consistent with anticipation of monetary rewards
Kidd, C., & Hayden, B. Y. (2015). "The psychology and neuroscience of curiosity." Neuron, 88(3), 449-460. Review, a complete theoretical framework for sorting out information as higher-order reward
A classic introductory book on thinking tools (Chapter 3)
Daniel Kahneman, Thinking, Fast and Slow (Farrar, Straus and Giroux, 2011). Chinese translation of "Thinking, Fast and Slow"
Judea Pearl & Dana Mackenzie, The Book of Why: The New Science of Cause and Effect (Basic Books, 2018). Chinese translation of "Why: The New Science of Causation"
Thomas C. Schelling, The Strategy of Conflict (Harvard University Press, 1960). Chinese translation of "The Strategy of Conflict"
Donella H. Meadows, Thinking in Systems: A Primer (Chelsea Green Publishing, 2008). Chinese translation of "The Beauty of System"
For advanced reading on these four tools (Stanovich, Cunningham, Tetlock, etc.), please see the "Appendix·Introductory Book List" of Issue 1.
About AI programming culture (Chapter 6)
The term "Vibe coding" was coined by Andrej Karpathy on the X platform on February 2, 2025. The main idea of the original post: “There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” The original post has been viewed more than 4.5 million times, and the word was later named the 2025 Word of the Year by Collins Dictionary
This article makes a conceptual paraphrase of the above source rather than a direct quote. To verify the original presentation, please refer to the original link and publication.