I don't get this type of attitude. Surely using the source signal, before an LLM add noise would be much preferable. Besides, there seems to be heavily diminish

I don't get this type of attitude. Surely using the source signal, before an LLM add noise would be much preferable. Besides, there seems to be heavily diminishing returns to de-noising LLM output, and even a hard barrier to how much we can denoise it. Yet people claim they prefer the noisy data and don't consider the risk that they are learning the noise instead of the signal because they have by definition no way of knowing what is signal and what is noise when they ask an LLM to teach them something. Because the noise is friendly sounding and on demand? Wild.