Let me remind you of recipe websites as an example of how summaries can be better by ignoring all of the useless crap that has nothing to do with making the dish

Agreed, but only assuming the summary is accurate and not hallucinating, which the current state of LLMs sadly can not guarantee. Maybe next year?

Noise reduction has tons of value in many fields.

I find that if I describe an esoteric bug to a high powered LLM, I often get to my answer more quickly than if I trawl through endless search results. The synthesis itself is a valuable addition.

Frequently I cannot even find source documents which match my exact circumstances; I’m uncertain whether they actually exist.