Comment by scotty79

I think comments on such posts have bimodal distribution. On one end there are people who see the utility of AI models for programming and are generally eager to see more capable models and ways of using them. On the other there are people who see AI destroying programming and have no idea how AI could change to be a force for good.

I had idea what might be the difference between the groups. I think for the latter group the code is important part of the goal. They see software as rather ends than means. Not entirely of course.

And the first group considers artifacts that the software produces to be the goal. So as long as AI written software is capable of producing valuable artifact they are willing and eager to go with it. And AI does that.

If the result of my code is finetuning of a neural network, I don't really care how it happened. I can benchmark it afterwards and know if the code that AI made for this purpose was good or not. I can inspect the code, investigate it, pinpoint ideas I don't like, suggest some ideas to try that I believe could give better results. I can restart, or try doing same thing few times in parallel trying different harnesses and models. All in service of the result, that is not code.

If you have a program that needs to do something and are willing to try AI to write it, think foremost about how you can rephrase the problem so that the output of AI written program becomes an artifact that can be independently verified, how to turn desired behavior into an artifact to evaluate.