I’m curious to know how well top tier AIs help you with your add-on development? Do you find yourself using just as an advanced search engine, do you use it to generate parts of logic that you unit-test and refactor it before including it in your project, or do you just vibe-code stuff? And which LLM’s do you use?
I happen to use the mix of the first two. Personally, I have found that DeepSeek and ChatGPT in most cases can’t even produced decent logic for most basic tasks, although there are some occasional outlying instances, however, for the most part either the code is too complex and long due to not using the already existing modules such as math_utils or it introduces unnecessary conditional statements, try’s and excerpt’s that only slow everything down even when instructed not to. But this is all with using free AI chat not IDE coding agents which makes a big difference I hear.
Just couple of days ago, when I tried testing to see how well DeepSeek performs, I asked if it would be possible to delete a dictionary member in one line that may or may not exist. To my surprise it started giving me couple of multi-line approaches, none of which were generic_dict.pop(key, None). These hiccups as well as many others I have experienced, do not instill a lot of confidence that it has the ability to produce quality code, especially not on larger scale.
On the side note, is it me or does superhive as of some time now have an influx of add-ons that seem vibe-coded with intent of just being a quick crash-grab?
I think it is what it is. I think what you describe is accurate. You just cannot trust it at all. And you have to really understand the situation and the context fully to be able to double check what it does, which is very tricky when coding, since coding usually involves learning new things and figuring stuff out and it is very tempting to think AI can do it for you, but it actually cannot. You still have to understand all the things you probably would like to avoid understanding in the first place when using it. That’s not intuitive. Vibe coding is absolutely a terrible idea and it doesn’t work in very nasty ways, but I do believe AI can be useful when coding. You can give it clear examples and instructions and supervise it closely to make it do some things for you faster. It can lead you to discover some things to actually look into and then do properly.
It’s not you. Every corner of the internet is full of that crap. It also has many problems. A lot of stuff that actually makes absolutely no sense.
It’s very tempting to trust AI in all sorts of ways that lead into all sorts of problems. It should be up to everybody using it to take personal responsibility for the stuff they produce, but that’s not the case. Of course it isn’t - we are all people, that’s just not what we want to do And it’s hard. For everyone, not just for the people who cannot code(but them especially ). I code. I coded before AI was a thing. I have made a fool in front of myself quite a couple of times with AI as well. I thought it’s fine, and it wasn’t. And I keep doing it. It’s hard for me to get it just how much you cannot trust it with anything at all as well. It’s a process. Hopefully we will all eventually get it and learn to use it as a tool.
And it’s hard. For everyone, not just for the people who cannot code(but them especially