I am working on an API that a (text-based) LLM can use to make paintings. It is a fascinating experiment, what an LLM determines to paint if just told to paint

I am working on an API that a (text-based) LLM can use to make paintings. It is a fascinating experiment, what an LLM determines to paint if just told to paint unprompted, what the weaknesses of an LLM are, and how solving these weaknesses also remove some soul and 'happy little accidents' from the paintings.