There is definitely a divide in users - those for which it works and those for which it doesn't. I suspect it comes down to what language and what tooling you use. People doing web-related or python work seem to be doing much better than people doing embedded C or C++. Similarly doing C++ in a popular framework like QT also yields better results. When the system design is not pre-defined or rigid like in QT, then you get completely unmaintainable code as a result.

If you are writing code that is/can be "heavily borrowed" - things that have complete examples on Github, then an LLM is perfect.

While I agree that AI assisted coding probably works much better for languages and use cases that have a lot more relevant training data, when I read comments from people who like LLM assisted coding vs. those that don't, I strongly get the impression that the difference has a lot more to do with the programmers than their programming language.

The primary difference I see in people who get the most value from AI tools is that they expect it to make mistakes: they always carefully review the code and are fine with acting, in some cases, more like an editor than an author. They also seem to have a good sense of where AI can add a lot of value (implementing well-defined functions, writing tests, etc.) vs. where it tends to fall over (e.g. tasks where large scale context is required). Those who can't seem to get value from AI tools seem (at least to me) less tolerant of AI mistakes, and less willing to iterate with AI agents, and they seem more willing to "throw the baby out with the bathwater", i.e. fixate on some of the failure cases but then not willing to just limit usage to cases where AI does a better job.

To be clear, I'm not saying one is necessarily "better" than the other, just that the reason for the dichotomy has a lot more to do with the programmers than the domain. For me personally, while I get a lot of value in AI coding, I also find that I don't enjoy the "editing" aspect as much as the "authoring" aspect.

> If you are writing code that is/can be "heavily borrowed" - things that have complete examples on Github, then an LLM is perfect.

I agree with the general premise. There is however more to it than "heavily borrowed". The degree to which a code base is organized and structured and curated plays as big of a role as what framework you use.

If your project is a huge pile of unmaintainable and buggy spaghetti code then don't expect a LLM to do well. If your codebase is well structured, clear, and follows patterns systematically the of course a glorified pattern matching service will do far better in outputting acceptable results.

There is a reason why one of the most basic vibecoding guidelines is to include a prompt cycle to clean up and refactor code between introducing new features. LLMs fare much better when the project in their context is in line with their training. If you refactor your project to align it with what a LLM is trained to handle, it will do much better when prompted to fill in the gaps. This goes way beyond being "heavily borrowed".

I don't expect your average developer struggling with LLMs to acknowledge this fact, because then they would need to explain why their work is unintelligible to a system trained on vast volumes of code. Garbage in, garbage out. But who exactly created all the garbage going in?

I suspect it comes down to how novel the code you are writing is and how tolerant of bugs you are.

People who use it to create a proof of concept of something that is in the LLM training set will have a wildly different experience to somebody writing novel production code.

Even there the people who rave the most rave about how well it does boilerplate.

> When the system design is not pre-defined or rigid like

Why would a LLM be any worse building from language fundamentals (which it knows, in ~every language)? Given how new this paradigm is the far more obvious and likely explanation seems to be: LLM powered coding requires somewhat different skills and strategies. The success of each user heavily depends on their learning rate.

I think there are still lots of code “artisans” who are completely dogmatic about what code should look like, once the tunnel vision goes and you realise the code just enables the business it all of a sudden becomes a velocity God send.