This becomes more true the deeper you look.

LLMs used to "fail at spatial reasoning forever". But each new generation would fail a little less. Then we got the likes of Astra, and ARC-AGI-3 straight up died.