I think the intent is important. Using LLMs to do well on the public leaderboard is like using cheats/aim assist. But learning how to use LLMs to solve complex puzzles independent of any sense of "competition" is more like when people train neural networks to drive a car in GTA or something - it's not hurting anyone and it can be a real learning experience that leads to other interesting byproducts.

But, yeah, don't use LLMs to try and get 9 second solve times on the public leaderboard, it's not in the spirit of the thing and is more like taking a dictionary to a spelling bee.