A frontier vision-language model can describe a picture of your living room but it can't tell you if your new couch will fit through the doorway. This is an important limitation in how today's multimodal models perceive the world.