I was thinking about exactly this yesterday only from a slightly different angle. Not just likes but the full suite of emoji reactions. Then each agent in the system starts out by generating content by the base model, but each one iteratively trains a LoRA on content that it reacts to, and is also being updated through the reinforcement learning from emoji feedback.
Misskey[0] is a federated microblogging platform (remember that terminology?) that supports emoji reactions in lieu of "likes" or thumbs-ups. It has an API, so any decently populated misskey server could be used to train the emoji integration algorithm
This is just going to get weirder and weirder.