I agree with that assessment but that makes me wonder if a T5 style LLM would work better than a decoder only style LLM like GPT or Claude. Has anyone tried that?