>>106213785 (OP)
The musing that Large Language Models are merely a complex form of lossy compression is somewhat reductive. The training process of these models actively cultivates generalization, deliberately expanding the data's representation to produce a more diverse range of outputs. This process constructs fuzzy logic circuits - wherein the model learns to map input patterns to predictions(the outputs) with a degree of probabilistic similarity, thereby generalizing its understanding beyond the precise instances represented in the original training data.
