I ran across this video the other day that I thought did a great job of summing up all the reasons why I intend never to touch GenAI bullshit.
And it’s from a former “vibe coder”. I’ll definitely be keeping this in my back pocket to share with people who ask me why I eschew GenAI.
I consider myself a mediocre developer, I have a security+ and a year of experience, before getting my bachelor’s. I am by no means an expert in coding or such, but even then, I can’t stress just the number of completely abysmal experiences I’ve had with AI coding tools, there are dozens of individual examples I can list but it’s easier to just talk about the central problems instead.
it writes incredibly inefficient and/or broken code, following almost no coding standards or guidelines, it will also gladly take a broken foundation and build on it.
context window is either lost or slowly poisoned, the “best” models(up to 1m tokens) can barely keep up with re-writing single files, since it needs to keep the whole file in-context before editing, let alone whole project (20+ files cohesive changes are basically still a fantasy)
“suggestions” during development are just hyper idealistic google fetch results, it retrieves anything in the first page mildly related to some keywords and tells you “so, those are basically your options” otherwise it just flat-out refuses or massively overcomplicates certain implementations
incapable of noticing errors or debugging proactively, it is essentially blind to all coding errors unless you keep referencing or correcting the results, you must now sit there and QA and babysit
incapable of learning or improving, you can tell it in detail about a mistake it made, it will sincerely apologize, then moments later repeat the exact same mistake
The list can go on forever but the last one alone would be a fireable offense for any employee in a company, I’m not sure why this shitbox is getting preferential treatment, but it’s whatever I guess
It 100% does all of the things you’re saying. Using these tools to produce things that are not 100% rubbish requires: (1) for you to be a babysitter, (2) for you to set standards early on using automated checks, and (3) for you to generally understand on a large scale how a software project typically runs.
That’s where I just don’t see software developers being obsolete even with these tools. Managers and directors will not be able to produce anything but garbage that doesn’t work using them.
A big issue IMO is that entry level programmers are likely also not going to be capable of running them with much success.
Just fyi, the wording of this particular line is confusing (to me). It could be interpreted as you disagreeing, when the rest of your comment suggests otherwise
It’s weird to me how inconsistent people’s reported experience with AI coding tools can be. Someone else in this thread said it “feels like developing software with a cheat code enabled using claude code”.
I’m glad for my own sake to be able to say that I don’t have enough experience with LLMs to know which would be closer to my experience. (Though if I had to guess, I can’t imagine I’d ever feel subjectively like I was having a “good” experience with it.)
But I do wonder what the difference is between the different experiences. I suppose it could be some tools are “better” than others, but I doubt it. Maybe some people are more skilled at bending LLMs to their will. But if I had to bet, (and not to say this is necessarily the case about aesthelete specifically), there’s a certain amount of evidence that people who have “good experiences” with LLMs may be self-deluded. The existence of “AI psychosis” points that direction. But I’ve heard “self-delusion” used as an explanation for why GenAI tends to make a lot of people less productive while making them feel more productive.