I tell it to write something only someone with typoglycemia can read and remove random letters from words longer than three letters. This usually gets around wo

I tell it to write something only someone with typoglycemia can read and remove random letters from words longer than three letters. This usually gets around word filters and semantic analysis which will give me more interesting results. I also just find it interesting that the model can keep coherent despite outputting something that is, language wise, complete nonsense. Though because of however my brain processes I can read it with only a little extra effort. I further wonder if this reveals some as of yet unexplored method of shrinking models or increasing output speed. You could train it on a set of shorthand words that another model could translate over to which would make it effectively faster at generating characters per second and increase window size by representing words as something smaller.