LLMs are put into a metaphorical isolation chamber, given a dataset, and are instructed to make predictions about the next token over that dataset. And this raises interesting questions about learnability, but it seems to undermine what makes UBL so interesting in the first place - the interaction.