I’m curious how they prompt the model or otherwise tell it what its goal is. They seem to suggest some language processing — perhaps they’re starting with a multimodal text + vision LLM?