Vibe Coding

Sometimes you just need a place to sit

I've seen something similar play out badly, in the context of young Indians (and other South Asians) who are underconfident in their use of English but need to use it in professional contexts. Some I know personally have gone from asking for editorial help on their drafts to completely surrendering to LLMs for all their output. I can't imagine it has helped their self-confidence. From the outside it almost seems they've given up on trying to express themselves entirely.

In software - rarely do you know beforehand that you're building a chair.

Even when you're building in an existing category of products (say, databases), there are still lots of variations and trial and error needed to get "what we need" defined clearly.

I'd say "vibe coding" is just a new name for "exploratory programming" or good old "prototyping"

Because sometimes, the journey is not the destination, and people really just want to get somewhere.

It honestly sounds like reprimanding somebody for sitting on the floor for a bit when they're tired, or ordering an Ikea chair, when a woodworking tutorial is right there on Youtube.

That's my sentiment as well. It seems you can actually go lower and it's so sad and I think at some point that will turn into dangerous as well... The amount of people that actually care what happens when you run code is miniscule. They don't understand that code is not some ideal abstraction. Code actually exists and executes in the real world on real hardware.

A week or so ago I gave a talk at a game jam discussing some of these topics and most of the audience being people in their 20's didn't understand or give weight to what I said at all. Only a few people from that big audience came after and expressed interest to learn more. Hope is not lost, but we have to actively push back and fight against the nonsense that comes from all sides.

I can't wait for the AI, AI bros and hype surrounding it to crash and burn in a few years. It's not that these new technologies are not useful - they are. And they have their place, but quality matters. DO NOT SURRENDER YOUR FACULTIES TO "AI". You get good by struggling! If you don't want to get good at this - then don't do it anymore. As another comment in another thread around code snippets said earlier - "No snippets, no LLMs, no autocomplete".

> This sounds like building a chair by just randomly nailing pieces of uncut wood together until it barely looks like a chair and supports your weight.

Have you been on the internet lately?

Money.

Not in my experience. You can still do everything as diligently and as manually as you find necessary or useful.

On the other hand, the bar to start exploring something in the first place (quite possibly sloppily, quite possibly containing a lot of bugs) has significantly gone down, and I personally find it amazing.

I think at this point it's no longer coding, it's more "AI coder management"

> Build one to throw away, as they say.

I believe Fred Brooks clarified this to mean you shouldn’t hesitate to throw away your first attempt if it comes up short despite your best effort.

If an LLM came up with the first implementation, not only did you not do the hard work of understanding the problem and how to solve it but, worse, you could be anchored to the LLM’s implementation thinking about how to patch it up so it performs adequately. Personally I think you should have LLMs have a go after you’ve had a solid try yourself, so you can actually compare and see what you might have missed.

I’ve had a lot of success with the method:

https://github.com/williamcotton/search-input-query

Search box DSL with React component with syntax highlighting, error syntax reporting, etc. Converts AST to postgres ilike, tsvector or pg_search.

https://github.com/williamcotton/guish

A bi-directional GUI editor for constructing Unix pipelines with bash CLI parsed to graphical modules.

All of these were made basically with the same tools and methods that Karpathy is describing as Vibe Coding.

These are not just simple little projects. This approach is not just limited to making a few CSS changes.

I don't get how this sentiment still persists.

This is simply not true if you're working on more serious websites for clients who actually audit for accessibility or responsive design, nor if you want it to be maintainable by web devs that actually know what they're doing.

CSS definitely isn't as messy some C++ still out there in the wild. If you're doing it right you should be able to quickly write a minimal amount of code that does what you want even for fairly large websites. Involving AI for something so trivial would just get in the way. Bad CSS is usually due to the requirements specifying a terrible design, much like how bad requirements drive bad code in any other language.

I've definitely implemented small utilities entirely in Claude (as webapps/"artifacts").

I'm pretty clueless with regards to modern LLM coding tools. What tools make this possible; talking to the computer to change code and whatnot?

> ...it's code generation through an AI model, without even understanding it...

Are you implying that Andrej Karpathy[0] doesn’t understand how the code generation from these models works? Or are you saying he isn’t capable of understanding the output?

[0]: https://karpathy.ai/

Sticking with that analogy, it's like reprimanding somebody for cooking instant Mac and Cheese when they're hungry when there is a cooking class only a few minutes away, and implying that they must obviously not know or care about food at all.