Comment by lcnPylGDnU4H9OF 3 days ago
> It's like we live in different worlds.
There is the huge variance in prompt specificity as well as the subtle differences inherent to the models. People often don't give examples when they talk about their experiences with AI so it's hard to get a read on what a good prompt looks like for a given model or even what a good workflow is for getting useful code out of it.
Honestly I think LLMs really shine best when your first getting into a language.
I just recently got into JavaScript and typescript and being able to ask the llm how to do something and get some sources and link examples is really nice.
However using it in a language I'm much more familiar with really decreases the usefulness. Even more so when your code base is mid to large sized
I have scaffolded projects using LLMs in languages I don't know and I agree that it can be a great way to learn as it gives you something to iterate on. But that is only if you review/rewrite the code and read documentation alongside it. Many times LLMs will generate code that is just plain bad and confusing even if it works.
I find that LLM coding requires more in-depth understanding, because rather than just coming up with a solution you need to understand the LLMs solution and answer if the complexity is necessary, because it will add structures, defensive code and more that you wouldn't add if you coded it yourself. It's way harder to answer if some code is necessary or the correct way to do something.
This is the one place where I find real value in LLMs. I still wouldn't trust them as teachers because many details are bound to be wrong and potentially dangerous, but they're great initial points of contact for self-directed learning in all kinds of fields.
Yeah this is where I find a lot of value. Typescript is my main language, but I often use C++ and Python where my knowledge is very surface level. Being able to ask it "how do I do ____ in ____" and getting a half decent explanation is awesome.
I've found this to be not very useful in large projects or projects that are very modularized or fragment across many files.
Because sometimes it can't trace down all the data paths and by the time it does it's context window is running out.
That seems to be the biggest issue I see for my daily use anyways
I'm convinced that for coding we will have to use some sort of TDD or enhanced requirement framework to get the best code. Even on human made systems the quality is highly dependent on the specificity of the requirements and the engineer's ability to probe the edgecases. Something like writing all the tests first (even in something like cucumber) and having the LLM write code to get them to pass would likely produce better code evene though most devs hate the test-first paradigm.
Some gave. Some even recorded it, and showed it, because they thought that they are good with it. But they weren’t good at all.
They were slower than coding by hand, if you wanted to keep quality. Some were almost as quick as copy-pasting from the code just above the generated one, but their quality was worse. They even kept some bugs in the code during their reviews.
So the different world is probably what the acceptable level of quality means. I know a lot of coders who don’t give a shit whether it makes sense what they’re doing. What their bad solution will cause in the long run. They ignore everything else, just the “done” state next to their tasks in Jira. They will never solve complex bugs, they simply don’t care enough. At a lot of places, they are the majority. For them, LLM can be an improvement.
Claude Code the other day made a test for me, which mocked everything out from the live code. Everything was green, everything was good. On paper. A lot of people simply wouldn’t care to even review properly. That thing can generate a few thousands of lines of semi usable code per hour. It’s not built to review it properly. Serena MCP for example specifically built to not review what it does. It’s stated by their creators.