We just did a competitive coding event and one of us based his strategy on ChatGPT, which was allowed. That got him through the qualification rounds, but then ended in the middle of the pack, which I think matched his skills as a non-ChatGPT coder. So, not a significant advantage, and none of the top performers used ChatGPT.

The thing is, most of the times, ChatGPT gives a wrong answer, so you need to look at the code, understand it, point out the mistakes and have ChatGPT output the corrected code. Sometimes manual correction is required, and sometimes, rewording of the question is required. It won't do everything for you unless you are particularly lucky, you still need to be a competent coder to work with it.

Yeah, but with the breakneck speed at which this field is moving, how long will it be before we have dedicated 'Coding Competition' models (if they don't exist already somewhere)?

I can't help but reminded of AlphaGo. When Lee Sedol saw the footage of the AI beating the then European champion, he was like 'That's neat, but it would still take a decade of training to get to world champion level'. That was about half a year before he crushingly lost to it. We just underestimate how fast those models can evolve. I know that playing a well-defined game against itself is an ideal setting for RL models, and that language models use very different architectures. Nonetheless, I still think it's reasonable to think that coding competitions may soon go the way of chess,..., where humans stand no chance against the best machines.

Not that that's a reason to cancel them immediately like Google just did (after all, chess is arguably thriving), but that's a different story.