this post was submitted on 28 Jun 2026
78 points (93.3% liked)
Programming
27916 readers
205 users here now
Welcome to the main community in programming.dev! Feel free to post anything relating to programming here!
Cross posting is strongly encouraged in the instance. If you feel your post or another person's post makes sense in another community cross post into it.
Hope you enjoy the instance!
Rules
Rules
- Follow the programming.dev instance rules
- Keep content related to programming in some way
- If you're posting long videos try to add in some form of tldr for those who don't want to watch videos
Wormhole
Follow the wormhole through a path of communities !webdev@programming.dev
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Even if it does write code that works, it usually (about 50% of the time in my experience) has bugs, and sometimes those bugs can be really difficult to spot. For me, it has never saved me any time. I’m either fixing something it doesn’t know how to do correctly, or going over its code with a fine tooth comb because when it says, “this is production ready code, with no bugs,” it’s usually wrong. That takes a lot of time. It’s easier for me to just write the code correctly myself.
Admittedly, I haven’t used that new model that Anthropic revoked access to the public to recently. Maybe that one is good enough for government work.
When I first tried it I felt lost, but after watching a couple videos about writing good prompts I had no trouble getting it to produce perfectly good code that did what I wanted. Your mileage may vary.
Honestly, I think what you consider „good code“ is just shifted from what the previous commenter considers „good code“. Prompting is about giving enough information so the AI can solve the task without needing to reconstruct a lot of context. Most people using AI somewhat regularly will have figured out to write good enough prompts. I‘ve never seen AI generate perfectly good code beyond hello world and the fibonacci sequence. And by perfectly good I mean I wouldn’t change beyond 30% of what it produces, which is not a high bar.
The last thing I had it do was create a filesystem using Discord as the storage medium. Don't ask me to justify the approach or tell me it breaks Discord's rules - that's irrelevant to my point. It was a fun idea to explore a way to host a bot for public use without also hosting other people's data. The data is pretty minimal but does involve several small tables with relational links. I thought it would be interesting to store it as a few Discord messages in a private channel.
As sample data I gave the AI the JSON file I had been using locally, and discussed various aspects of the design with it. It came up with the scheme of splitting up the data using a channel for each table and a message for each row. I don't remember the interaction in detail but it asked me a bunch of questions - for example, it commented that Discord's limit of 2000 characters per message didn't seem to be a problem given the sample data, but wanting me to confirm. I had it generate the code as a node module with complete CRUD functionality, and refactor my existing code to use the module instead of the JSON file. What it gave me worked perfectly right away.
Now of course this isn't "Write me an accounting system," but it's far more complex than Hello World or Fibonnacci. Whether or not you code with AI is your choice, but deciding my notion of quality isn't up to someone else's because my results aren't "AI sucks" is just pure denialism.
Honestly, I would argue about the complexity of it. The discord API is straightforward, and what you're describing doesn't sound that complicated.
And we have seen fairly complex stuff that AI did, but if it does pull it off somehow, the next question is quality.
I just read a bit of AI code a few minutes ago, and it didn't take long to step over a chain of if / else if statements where only the first if statement if reachable, the rest is not. The code should've been one if statement followed by a match statement. Not sure why the LLM thought "combine" two logically completely separate statements into one. And that was in 1.5kloc it generated, so not much. There's probably even more, I just read a fraction of it.
Fable was "just ask and get it done" quality level, but really I don't get THAT much bugs - about the same amount that I see irl developers do: get a new feature, find 5 problems, get them fixed, find one more, done. As a recommendation - try to alleviate the biggest problems that ai models have:
I’m glad it works for you, but it simply does not work for me. Maybe you could try yourself on some of my libraries, because I have never gotten it to save me any time. It’s just spending money and making the work less fun for no reason. Oh, also not having the copyrights to the things that go in my code base, don’t forget that.