At least 30% of my queries to an LLM include deciphering something from an image. If I’m coding pretty much a hundred percent of bugs include some sort of image.