I don't know if anyone else shares my experience, but for me, vibe coding has been a complete miss.
I'm not sure if it's the way I've been prompting, but for anything beyond the banal, LLMs have completely failed to generate what I would consider to be acceptable code.
For anything with WebGL, high-performance networking, or generally anything requiring taste, I think there's still quite a way to go...
Maybe ironically, but it seems to me that we are still _so far_ from being able to generate something like this: https://www.stripe.press/scaling purely from prompting.
> LLMs have completely failed to generate what I would consider to be acceptable code
Totally agree. Recently I asked ChatGPT to write a Z-80 assembler function to multiply two 8-bit numbers. I specified Z-80 assembler to get away from the mainstream. I tried the same prompt three times and got three different responses, two of which were obviously wrong and the third had no obvious flaw but was rather clumsy.
I'm not sure if it's the way I've been prompting, but for anything beyond the banal, LLMs have completely failed to generate what I would consider to be acceptable code.
For anything with WebGL, high-performance networking, or generally anything requiring taste, I think there's still quite a way to go...
Maybe ironically, but it seems to me that we are still _so far_ from being able to generate something like this: https://www.stripe.press/scaling purely from prompting.