@Vrimj@mastodon.sandwich.net In that case, how about chain of thought prompting? It's true that LMs only model an input/output distribution, but they can mimic and reproduce as many convincing artifacts along the way as necessary.