4 ms·

None of these AI systems significantly outperform searching for the answer online. The AI is good because: - The programs it generates are small - The programs

None of these AI systems significantly outperform searching for the answer online. The AI is good because: - The programs it generates are small - The programs generated solve well known problems with well known answer Current AI replaces SW developers the same way StackOverflow and blog posts replace SW developers. The real question is how fast new AI is going to be developed which overcomes these challenges. I think what is pretty certain that current language models will never be a serious contender against programmers, their whole structure makes it very difficult to perform actual logical reasoning about the structure of a computer program. Another thought: How useful is an AI which is correct 99% of the time? How much development resources would be needed to fix a code base which is 1% total nonsense that does not even compile or is just plain wrong. Would the developer time (all of which have never seen the code) spent fixing that code base be actually less than a total rewrite?

steve_adams_86 4y ago
> Current AI replaces SW developers the same way StackOverflow and blog posts replace SW developers. You can ask GPT how to optimize code and it’ll actually give decent responses in a few seconds though. That’s new. It can reveal answers about questions you’re not sure how to ask, at least as a less experienced developer. All of its optimizations so far have been things I’m aware of except for one, which was mostly due to me not writing that kind of code very often. That’s becoming superior to blogs and SO if only because the response is rapid and concise. As for correctness, honestly, I come across so much human-generated stuff online with errors that I read docs and experiment unless I can find the answer from a source I have high confidence in. People with less experience don’t have that luxury, so they go down far more rabbit holes without realizing it. GPT will still parrot some of the bad ideas, but eventually it could be trained on industry standards. Tools like copilot are poised to become incredibly helpful, I think.

steve_adams_86 4y ago

> Current AI replaces SW developers the same way StackOverflow and blog posts replace SW developers. You can ask GPT how to optimize code and it’ll actually give decent responses in a few seconds though. That’s new. It can reveal answers about questions you’re not sure how to ask, at least as a less experienced developer. All of its optimizations so far have been things I’m aware of except for one, which was mostly due to me not writing that kind of code very often. That’s becoming superior to blogs and SO if only because the response is rapid and concise. As for correctness, honestly, I come across so much human-generated stuff online with errors that I read docs and experiment unless I can find the answer from a source I have high confidence in. People with less experience don’t have that luxury, so they go down far more rabbit holes without realizing it. GPT will still parrot some of the bad ideas, but eventually it could be trained on industry standards. Tools like copilot are poised to become incredibly helpful, I think.