We are living in the attention economy more than ever. While algorithms initially used human-made content to captivate our attention, they are now bombarding us with AI generated slop. What tools can we apply to cut through this and extract the essence saving precious time and attention.

AI generated slop

The other day, I was browsing YouTube and my attention was caught by one of its suggestions titled: "The Heartbreaking Tragedy of Chip Foose from Overhaulin'" with a thumbnail showing Chip in the front and a car crash in the background.

I remember watching the show when I was younger, fascinated by his drawing skills and tasteful car overhauls. He always came across as a nice person, so yes, I got emotionally engaged thinking the worst given the dramatic backdrop. The algorithm got me on its hook and started to reel me in as I clicked the video and started watching. A couple of minutes in, I felt that the content was being dragged out, the images started to repeat and the story did a lot of backtracking which made me realize that I was probably watching AI-generated content. Surprisingly, reading the description, it was actually tagged as AI generated content. With a total length of around 28 minutes, I wasn't going to spend more of my time on this slop.

But still ... What happened to the guy? ..

Forging a new axe

As it won't be the last time I'll have to deal with this nonsense I'm going to need something which allows me to extract what I need to know from a Youtube video without having to spend any time to it. Luckily, the necessary tools already exist, and thanks to a 50 year old Unix principle¹, we can create something quickly which does exactly that.

yt-dlp

yt-dlp is a feature-rich command-line audio/video downloader which allows us to download the captions of Youtube videos. The many options are a little involved so we create the following convenience function in Bash which downloads the captions, strips off all XML syntax and spits out the results to STDOUT:

get-yt-captions (){
    tmp=$(mktemp);
    trap 'rm -f "$tmp" "$tmp.en.srv1"' RETURN;
    yt-dlp --skip-download \
           --no-warnings -q \
           --write-auto-sub --sub-format srv1 --sub-lang en \
           -o "$tmp" "$1";
    xmllint --xpath '//text/text()' "$tmp.en.srv1"
}

clai

Clai is a CLI tool I wrote to interact with various LLM/AI models. It allows us to to pipe the aforementioned downloaded captions into it and process them however we desire.

$ echo "Roses are red" | clai prompt "Complete the poem's opening sentence."
Roses are red, violets are blue.

Putting it all together

$ get-yt-captions 'https://www.youtube.com/watch?v=TffKMCop8xw' | \
clai prompt "Summarize in a single sentence what happened to Chip."

After rising to TV stardom on Overhaulin’, Chip endured punishing workloads
and reputational fallout from a partner’s scandal, then quietly withdrew from
public life to build cars privately and focus on his family.

Ah, great to read that Chip is doing just fine. ✌️ 😎

Closing words

Hopefully this little workflow helps you cut through the artificial fog and keep your attention focused where it matters.

Stay sharp, stay curious, and don’t let the slop win.

Footnotes

Pipeline(Unix) ↩