I don't think we're in the early ages... LLMs technology has essentially stagnated since GPT3.5, we just have bigger models that can handle more context. We're trying to cope for the lack of progress of the actual technology by coming up with contraptions of multiple models stuck together, Mixture-of-Experts, Reviewer models, PM models...