News for hackers. Stuff that matters.
Re:'A lot worse than expected': AI Pac-Man clones, reviewed by Aurornis on Friday March 14, 2025 @12:21PM (#43364135)
This is a good summary of the AI app generation experience. They’re very good at producing something as long as it overlaps with what you can find in a lot of public blogs and tutorials. Pac-Man is a common intro game so there’s a lot of source material to train on.

But once you get past the initial wow factor it’s hard to move forward with the refinement. Recently I tried to use multiple LLM tools to implement a solution to a problem that I had solved before. I spent hours prompting and reprompting trying to tell it exactly what to do, but every time I’d pin down one fix it would break something else by gravitating back to a different solution that didn’t work. Finally I re-read the tutorial in the library’s documentation and realized that it was just pulling my code back toward the documentation example at every opportunity. My attempts to deviate and implement something more complex kept breaking it.

Re:'A lot worse than expected': AI Pac-Man clones, reviewed by parasti on Friday March 14, 2025 @02:27PM (#43365638)
In my experience the biggest roadblock to continuous conversation is context length. It fills up and the LLM starts forgetting parts of the conversation. Most tools don't even tell you that this has happened. But if you keep that in mind, that there is a buffer there filling up, you can massively improve the quality of the output.
Re:'A lot worse than expected': AI Pac-Man clones, reviewed by yzydserd on Friday March 14, 2025 @03:13PM (#43366139)
Many prompters don’t practice hygienic reversal by rolling the outcome of a conversation back up the thread as input. If 10 comments are spent coming to some clean code, go back up the thread and include it as the status “back then”.
Re:'A lot worse than expected': AI Pac-Man clones, reviewed by nemomarx on Friday March 14, 2025 @12:35PM (#43364294)
What you could do, I guess, is write up 5-6 implementations of it yourself first, then feed those into the training set for the model, and then it should have a decent shot at understanding that particular problem. (maybe /s?)

I guess my question is if you had a good solution you'd done before, and also a good understanding of what kinda changes it needs, why bother with hours of AI prompting?

Re:'A lot worse than expected': AI Pac-Man clones, reviewed by ehutch79 on Friday March 14, 2025 @02:32PM (#43365701)
These models are supposed to be reasoning. Why would you need to do all that work? If I can teach it to an intern, why wouldn't a model that can reason be able to do that?

Unless all the promises and hype are false...

Re:'A lot worse than expected': AI Pac-Man clones, reviewed by drdeca on Friday March 14, 2025 @02:23PM (#43365577)
Presumably the purpose of the exercise was to test the model’s capabilities / “take it for a spin” , not to obtain the working progress (as, as you mention, they already had a working program to do the task).
Re:'A lot worse than expected': AI Pac-Man clones, reviewed by junipertea on Friday March 14, 2025 @02:22PM (#43365567)
To write a blog post! "I used AI to solve this challenging problem" gets more traction than "I solved this challenging problem"
Re:'A lot worse than expected': AI Pac-Man clones, reviewed by godelski on Friday March 14, 2025 @09:46PM (#43369180)
IME vibe coding feels faster but traditional coding is faster. Every time I've tried to code with LLMs in just spending lots of time waiting.

Funny thing is it reminds me of when I'd first start coding. You write too much, compile, wait, compiler error. Read error, fix, add print statements, repeat. There's that down time with the compiler that made it feel not as bad but also make me lose concentration.

Similar to when I started coding it also results in things that technically work but are clearly spaghetti held together with duct tape. But at last that process resulted in me becoming better