I took time to try all four examples in the feature request.
BlinkNotes didn't work for a technical reason, TLDR; it did not have AI capabilities - only crowdsourced summaries.
TLDR This was good but seemed to use an NLP algorithm they call "AI." So instead of summarizing articles, it seems to produce a bullet list of the essential text fragments. This is in contrast to the much easier-to-read conversational style summaries GPT-3 can provide. TLDR This also did poorly with some of the technical articles I have authored, sometimes extracting pieces of C++ code as meaningful text.
Summari worked very well for me. It summarized even quite technical articles well because it uses a conversational AI. Thanks for the recommendation.
If I had to name one key criterion that differentiates good summaries from bad, in my opinion, it would be the effective use of language. Conversational/language models like GPT-3 are proficient at absorbing much contextual information and synthesizing a short and effective summary. NLP algorithms are good at throwing away superfluous context, which is common in casual writing, but they do not seem to work well for technical writing or texts whose purpose is to explain concepts and where there is little superfluous context to throw away.
Perhaps for something like Unclutter, if the users mainly read news sites, then an NLP approach could be appropriate (it would be cheaper and works well for such content). But the ideal implementation for an article summarizer for me needs that summary-from-a-lot-of-context synthesizing capability.