There’s a substantial amount of important visual information in the video. It’s not just stock photos.

Disagree. I’d rather listen to information while doing something else. Most info isn’t so important that I should stop moving and sit down and read it.

> listen to information while doing something else

Hard to find intellectual material that does not require full concentration. Even those few pieces that seem to be highly digestible - they typically do not speak to the intellect and/or do not get fully processed anyway, under condition of shared attention.

It’s iflscience not ifwtsscienceid (want to study science in depth). :P

Thanks for the recommendation.

If you also can peek to videos in your multitasking: there are many related YT channels around, but if you are into leisure time spending on divulgational channels, there is probably little better around than

Curiosity Show, the Australian (Adelaide?) TV channel featuring Rob Morrison and Deane Hutton in the '70s and '80 - for the young;

Steve Mould, ongoing YT productions, picking stimulating topics of non immediate evidence.

  yt-dlp  --skip-download --write-auto-sub

...

> inside your head is one of the greatest computers in the universe which is surprising because I've seen your b-

I.e. one can also quickly verify if it is for the wrong target.

So much good information is presented now as videos... I'm thinking of automating that. A service, which follows your watched yt videos feed and automatically saving a transcript. Maybe with some LLM summary and few screenshots - as simple html page.

I believe people keep producing videos, since it can be monetised nicely

You should work on an engine that adds punctuation to the subtitles. They are pretty raw as they come - they normally require some post-production to be made into structured text.

Probably something already exists? LLMs could be good for heuristically obtaining punctuated text - but since real understanding is required, other systems should be employed - such as a transformers based ANN (speech + raw_transcript → punctuated_text).

The issue of "real understanding" is that which should discourage «LLM [generated] summar[ies]» - I have not yet seen any that show actual understanding of the original. "No work" is often better than "bad work".

I did some initial tests and found that whisper.ai gets transcripts a loot better. Might improve punctuation.

Also, when asked to summarize, even Vicuna removed all portions of irrelevant talking (like to the crowd on a meeting) and jokes.

Might be worthwile to finish that