It is absolutely worth hiring and training juniors. The quality of your onboarding process and documentation will improve. Not only that but a junior will ask questions that senior engineers take for granted, such as "why are we doing X this way?" which can lead to improvements that your existing engineers might not have considered.
Finally, if junior engineers are joining your organisation and leaving every 9-18 months, you need to take a serious look at your career progression ladder and compensation. I have seen way too many companies that have an arbitrary "you cannot receive a promotion in the first X months" HR policy which is just asinine. You know who doesn't have this stupid policy? The company your junior just accepted an offer from.
If your organisation doesn't have the tools and processes to up skill junior engineers into seniors, then it doesn't have professional development for senior engineers and is just a career dead end.
We also have to remember that if our juniors leave 9-18 months later, everyone else's juniors are leaving too. Churn has a cost for sure, but if I can hire someone else's junior that they put 18 months into then I am still better off.
But that's fine. The junior you trained up will be more effective at their next job, and the junior your competitor trained up will be more effective when you hire them at your company. Again, that's how it's been for far longer than we've had LLMs.
I am saying that I care about them, and individually we all should pay attention to them. In this case, I will continue to push for juniors even if my higher-ups push for us to hire seniors only and augment our workforce with LLMs for more junior-level work.
This is the fundamental contradiction of LLMs. The promise today is that the tooling can largely replace juniors, and honestly that may be true.
The hope behind that promise, though, is that the tech will catch up with senior devs before the pipeline dries up and that we have found a sustainable social and economic model before humans truly aren't employable at any meaningful numbers. That hope seems ill placed to me, but I guess we'll see if we develop such skilled LLM or similar tools at all.