kohllision

That should be the easiest part of the loop. Frontier models are good judges of human preference and taste and should be excellent at prioritizing proposed questions. The generation part is much harder.

I appreciate the sentiment, but this remark is far too common here. It does not address the content of the post and could apply to any submission from LessWrong. The author of the post has made none of those technical choices.

> I think the direct impact of these competitions to hiring were minimal.

I can confidently say that this is not the case in India. I have a countless list of people (including myself) who received emails from Google recruiters for advancing to later rounds of Code Jam or securing a top rank in a Kick Start round. Even if people missed out on top ranks, they would put their results on their resume/LinkedIn and would get attention from recruiters anyway. Needless to say, these were strong signals for recruiters, especially for university students. Based on the skewed demand-supply here, they were closer to necessary than sufficient.

This is comparable to the argument that programmers are learning how to eliminate compiler/linter warnings rather than learning how to program better. Immediate feedback is the best learning tool and one leads to the other. Of course, natural languages are far more complex and have no universal standard. While it is true that no easy way to check for style, be it in programming or writing, we should encourage the use of such tools (once they are secure and privacy-friendly) as writing assistants.