One example of an ’experiment’ would be to explore the latent space with random/procedurally generated prompts and do semantic analysis on the results to look for topics or sentiments to emerge.

My guess is that the current language models don’t have enough information in the training data to do this usefully today, but over time it seems potentially viable.