Comment by navvyeanand

This is very true. However, I wonder how much of this can be mitigated by using training data from other open-source models like Olmo3 for textual data, Emu3.5 for vision?