That would be if you used a language-only model. Plenty of folks have tried multimodal or vision models on arc agi.

→ View original post on X — @jeremyphoward