Small-language models also use GPT. And I agree it's not "LLM vs RL" and that they're orthogonal, however, RL has features that more clearly resemble intelligence.