I really struggle to understand how many people seem to have "summarize this" as their main LLM use case.
And I use LLMs every day, for so many tasks - just never this one. Can you please share a couple of real-world times where you couldn't bother to read the "thing" and truly got value from having a LLM read it for you?
The only time I didn't think summaries are useful is with creative fiction or pure entertainment content.
"Summarize it. Keep only instructions and ingredients (more prompt about how to format instructions, what units to use, how to order instructions) and start with one sentence that describes the dish and it's origin"
Which is a terrible waste, really. Many recipes are blown up to make them copyrightable. Often done with AI. And then I use AI to remove all that again.
To attempt to make them copyrightable at least. My understanding is that fundamental recipes themselves as sequences of steps and ingredients are never copyrightable in the US. From having sat around a lot of recipe bloggers though it sounds like the main reason they do the long essays is to try and improve search ranking.
Then the next version of my doc becomes the summarization, and I only flesh out details where the summarization went too far and removed critical details.
Did it read the summary? Nope, I already know the material and have been using it for years. But it was a great way to communicate the key points to the model as part of project instructions.
[0]: https://jacobian.org/2021/may/25/my-estimation-technique/
So in this case, this was not you trying to ingest a condensed version of the info, but rather transforming it to build a prompt (under the assumption that shorter instructions would perform better than the full article).
Makes sense
And the HN comments!
How likely is it that the output summary is mostly made of stuff that the LLM made up in the first expansion process? (eg, you're getting summarized noise if the original signal - the prompt - was much shorter than the email)