RGB-as-text for image classification with LLMs


Can we use RGB (or grayscale) encoded images, represented as text, as input to an LLM to do non-trivial classification tasks?
i.e., take an MNIST image, print the sequence of grayscale values as text, and provide it in a prompt to an LLM with the question "What is this a picture of?"

If this works, it's feasible to do simple image work, inefficiently, using text-only models.
