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.

What makes you think that? Especially given that fact that GI (without the 'A') is evidently very much possible with only a tiny fraction of the "totality of human knowledge".