And, so far, the results of "visual data for improving general intelligence" runs were nothing but disappointments.
I think vision is just a piss poor modality to learn intelligence from? Very low value, per bit and per token both. You only ever want to tap it if you need your AI to operate based on visual data at deployment time. Otherwise, even "experience" is best gathered in text RLVR rollouts.
The secret of human sample efficiency isn't that visual data is somehow better for learning intelligence. It just isn't. Human "training data" is a hundred kinds of awful - humans are just good at scavenging it for all its worth. Evolution has tuned that very well.
Which means: AIs can get good at it too. It's not a wall - it's a skill issue.