That is only true if your path to AGI is to take models similar to current models, and feed them with tons of data. Advances in architecture and training proto

That is only true if your path to AGI is to take models similar to current models, and feed them with tons of data. Advances in architecture and training protocols can and will easily dwarf "more data". I think that is quite obvious from the fact that humans learn to be quite intelligent using only a fraction of the data available to current LLMs. Our advantage is a very good pre-baked model, and feedback-based training.