Yes, but LLMs are not everything. Much simpler models can be sufficient for a great recommender system.

Would it not be possible even on the client side to iterate trough global feed and mark posts based on recommended ratings and local settings? Depending on a device that could take more or less time, but should still be doable.