I don't know what to save from this article. Maybe only "[LLMs are] very bad at spatial reasoning. Which shouldn’t be surprising, because that’s also not in the training data."
Frankly, Claude has been unbelievably proficient at spatial design since Opus 4.6. I think a lot of the people commenting here are relying on outdated assumptions. Simply put, LLMs have crossed a threshold and can now produce professional, shippable visual designs, similar to the way they became good enough to produce shippable code in 2024.