What's this about?
A study on training models to be multimodal from scratch rather than bolting vision onto a text-trained model. The argument: text-only pre-training caps a model's grasp of the physical world, and starting multimodal avoids that ceiling (18 upvotes).
What exactly is "native multimodal" training?
Most vision-language models are text models with an image module attached late. "Native" means images, and other modalities are in the training diet from day one — like raising someone bilingual from birth versus teaching a second language as an adult. The claim is the native version internalizes the physical world more deeply.
Why does it matter?
As models are asked to perceive and act (see K3's native vision, PerceptionBench, computer-use agents), how deeply vision is baked in — not appended — may decide the ceiling. This is the training-recipe question under a lot of this week's perception work.