Comment by kdmytro

The dataset Z is a collection of facts. The output Y captures the patterns from Z. The LLM can generate novel insights (facts) that satisfy Y without violating the data processing inequality if we consider the patterns to be a part of the information of Z.