Despite widespread confusionAndrej Karpathy conceived “atmospheric coding” As a kind of AI-supported coding where you “Forget that the code even exists.”
Legacy code
We already have a sense of code that nobody understands: Legacy code.
Legacy code is universally despised and for a good reason. But why? You have the code, right? Can’t you figure it out there?
Wrong. Code that nobody understands is technical debt. It takes a lot of time to understand an unknown code sufficiently to debug it, let alone introduce new functions without introducing bugs.
Programming is fundamental Theory BuildingDo not produce code rules. We know this. That is why we ridicule business people who try to measure the productivity of developers in coding rules.
When you use a code atmosphere, you use it as quickly as the LLM can spit it out. That is why the coding of atmosphere is perfect For prototypes and disposable projects: it is only legacy code if you have to maintain it!
Prototypes and disposable code
I like to have coded apps to:
I didn’t have to keep developing those apps, so it was no problem that I don’t understand their code. These apps are also very small, which means that I have not made as many debts as I have to jump and read the code. I was able to cod these apps much faster than I could have built, and it was great.
Vibe coding is a spectrum
Vibe coding is in a spectrum of how much you understand the code. The more you understand, the less you vibt.
Just by being an engineer and asking a web app with a persistent database, you already vibt less than a non-programmer who asks for an “app” without understanding the distinction between a web app and a native app, or how persistic data storage works.
Give a credit card to a child
The worst possible situation is to have a non-programmer-atmosphere code a large project that they intend to maintain. This would be the equivalent of giving a credit card to a child without first explaining the concept of debts.
As you can imagine, the first phase is ecstatic. I can wave this little piece of plastic in stores and take what I want!
That looks a lot like AI can now build everything! Nobody has to learn to cod! Look at what it just made for me!
But if you wait a month, you will receive the credit card account. Do I actually have to buy all those things? How do I get myself out of this gap?
It is similar to the atmosphere coder. My code broken. What do all these files and folders even do? How do I ever get this solved? Can I get a reimbursement for the $ 400 that I spent atmosphere coding?
If you do not understand the code, it is your only story to ask AI to repair it for you, what is when paying off credit card debt with another credit card.
Serious coding with AI in 2025
If you are building something serious that you plan to maintain in 2025, Andrej is entitled to:
[Keep] A very tight belt on this new over-gran-grazed Junior Savant with Encyclopedic knowledge of software, but which you also always have nonsense, has an upset of courage and shows little to no taste for good code. And the emphasis on slow, defensive, carefully, paranoid and on always delegating the inline learning option.
– Andrej Karpathy, Twitter
How we approach it for AI
In Val Town we have built AI in our product in dozens of ways. So Is our AI Asisstant who reads and writes in an agentic, it carries out, looks at the logs and continues to eat until it is ready.
So is a great tool for atmospheric coding. I heartily recommend it to people who understand these considerations. I sometimes use it to vibe code. Other times I stay on a tight belt because it makes surgical edits for a project I care about. Both are fun and useful.
Coding with AI changes so quickly that it is difficult to know what will bring tomorrow, but I am convinced that the theories remain central to the activity of building complex software. Our technical expertise will still be relevant! And I am optimistic that AI will continue to program in surprising ways.
But if you know that non-programmers issue thousands of dollars atmosphere that codes their idea of billion dollars today, send them this message. Vibe coding will not get them where they want to go. They will have to learn to use their human eyes to read the code 😱 and to learn that it is sometimes easier to start building a well -written code base than to repair a legacy that nobody understands.
This essay is a distillation of a lecture that I gave last month, The role of the human brain in programming. Thanks to my fiancé Emily for listening to me about these topics for months and for filming my speech. Thank you Malte and Rippel for hosting the conversation.
Thank you Geoffrey Litt, Jimmy Koppel, Max McDonnell, Tom MacWright, Charmaine Lee, Brent Jackson and then Shipper for feedback on this post. Thank you Simon Willison and Andrej Karpathy for being voices in the midst of all AI -Hype and no -sayers.
#Vibe #code #legacy #code