That’s what I used to do but now I think it is a major handicap. I haven’t perfected it yet, but I’ve had good luck letting AI write entire (smallish) programs and do major updates to those.
back
1 comments
I find this fascinating because even using the last models more often than not asking the AI to generate even a test file for a simple module with some controlled state will result in an incorrect test setup that simply won't run, or nonsensical tests that clearly don't reflect what the code is doing.
I often use a coding assistant to handle boilerplate for me or write additional tests after I've given it some correct ones to work of off, but when I read accounts like these I always wonder what's different about our experiences that you feel you can give it bigger asks like that.
Wouldn’t fine-tuning make a difference? Granted, majority of HN users don’t invest the extra effort to do this.
Finetuning is as finnicky as coding it yourself directly. If you already have a toolset why bother with another that is statistically wrong as many times as it is right.