"A good cyclist using training wheels will go circles around a good cyclist who doesn't." No, training wheels only help bad cyclists. You can't generalize that and assume they will make a good cyclist even better.
AIs generate deeply mediocre code. This is better than anything a person who can't code on their own would produce, but an experienced developer will have to spend all their time babysitting the AI to get it to behave properly.
Not really. Today's mediocre code is tomorrow's technical debt. LLMs often inject subtle bugs or misunderstand the context they're being used in and have to be badgered into respecting the parameters of requests.
Would you rather write code yourself, or ask a first-year student to write it for you while you watch over their shoulder and tell them to go back and try again every time you notice a mistake? Which of these do you think is faster and better in the long run for the quality of your codebase?
AIs generate deeply mediocre code. This is better than anything a person who can't code on their own would produce, but an experienced developer will have to spend all their time babysitting the AI to get it to behave properly.