I completly agree, it's a diminishing return but a return nonetheles, so to have a "world class GPT" model with 10% of the data, and 10% (optimistically) of the compute that OpenAI used, you will need some very very impressive innovation elsewhere.