You must log in or # to comment.
Claude is faster and more accurate than my junior devs, but it can’t sit in meetings for me 🙃
I’m sorry, because then your juniors must be really shit. I gave AI tools multiple chances already but they produce so much garbage. Even a very inexperienced junior will produce better code after a few weeks of training.
I’ve had to start using it a bit but, really don’t use prompts. Generally, just using it in my IDE as a fancy autocomplete that is wrong 40-60% of the time but really reduces toil when writing switch/case statements and the like in statically typed languages.
You don’t know me.
How did you use LLM for coding? IDE integration? Prompt chat? Explaning the problem and hoping to one shot it? Genuine question, no judgement.
I have only ever chatted with LLMs for code, refining the output until it does what I need it to. An example of this would be “I have a stream of bytes in Delphi, convert it to UTF-8” as starting prompt and updating with “don’t use X” or “use method Y”. It mostly works and sometimes it does not. It is very useful when I cannot find the solution on forums etc (or if my knowledge of Delphi is not good enough to extrapolate from what I can find).
I do not think I have ever asked it to create anything bigger than a function (except this one time I created a C# application mostly with ChatGPT because learning C# was too much effort for a one-time side project).
What I’m referring to is the IDE integration, like an “enhanced” auto complete. It is actually less helpful for me than the normal auto complete. Most of the stuff it generates is either not what I want to do or it doesn’t follow the code style of the rest of the software. Mainly a combination of the two.
I also briefly tried generating whole methods from prompts but that never resulted an adequate solution ever, so writing it myself is orders of magnitude faster than correcting the bullshit that Copilot vomits out.
The best use I’ve found so far is to solve usability problems, that I can’t find an easy enough solution to in the docs. For example some time ago I had to create a recusive CTE in a Postgres query. So I condensed my issue down to “I have a table A with columns X and Y and a table B…” and so on and fed that to ChatGPT and applied the result to my specific problem. Turned out I had two of my parameters swapped in previous attempted so it helped be successfully. But that’s the only kind of usage I get out of it and it happens once in 1-3 months tops.
Thanks for explaning. I have only used the IDE integration with copilot when it first came out and not since then. I was very obnoxious at times, prompting every time I stopped typing. Could be useful but also not what I needed at all.
The only thing AI was ever really good for is a slightly faster version of googling stack overflow. I’ve never got anything out of it that worked as is, but it steered me in the right direction a couple of times.
In other words, it can only solve a problem that doesn’t exist and it’s getting worse at it as time goes on (which google, sadly, is also doing).
If you use one of the LLM chats, you can add information to refine its output
This is how you get left behind being a luddite.
Same argument was had when IDEs became popularized. I remember writing Java in notepad for my Java 117 course forever ago because the professor was insistent that they were a crutch instead of a tool. (To clarify, I don’t suggest “vibe coding” is the way. At a certain point you have to take responsibility for the end product you’re producing and that includes reviewing as much code as necessary to ensure that the output is functional and quality. If I were to compare it, I’d compare it to a compiler. Another layer of abstraction. The C compilers for example used to take what is perfectly well written code and occasionally mangle the output into some unholy abomination of error ridden assembly output, until they were patched and all C standard features implemented ((which iirc some still aren’t, although that’s a bit pedantic)), but I’d bet only a slim fraction of coders review the output assembly of the C compiler nowadays, they’re basically foolproof, however a good coder should still be able to understand and review assembly because a situation may arise where it’s necessary. As for the analogy, LLM based coding tools are in that error ridden phase as well from what I can tell, so taking responsibility for the code they produce is more important now than ever.)
