I think you'd start out with gibberish and at each step the output would become more refined, so first you'd maybe see scrambled words emblematic of Berlin, whi

I think you'd start out with gibberish and at each step the output would become more refined, so first you'd maybe see scrambled words emblematic of Berlin, which would then eventually synthesize into a sentence. I think you'd have a fixed length of text, with maybe some way for the model to mark some of it as truncated.