Multi-modal models will be exciting only when each modality supports both analysis and synthesis. What makes LLMs exciting is feedback and recursion and conditi

Multi-modal models will be exciting only when each modality supports both analysis and synthesis. What makes LLMs exciting is feedback and recursion and conditional sampling: natural language is a cartesian closed category. Text + Vision models will only become exciting once we can conditionally sample images given text and text given images (and all other combinations).