Claude is multimodal, it has been trained on images

Multimodal is a farce. It still can’t see anything, it just generates a as list of descriptors that the LLM part can LLM about.

Humans got by for hundreds of thousands of years without language. When you see a duck you don’t need to know the word duck to know about the thing you’re seeing. That’s not true for “multimodal” models.