I extract text from video captions and use GPT to detect the sponsored time stamp candidates.

The upside is that it can work for any video without the initial user interaction. Manual/user corrections can be applied afterwards ofc.

I’m trying to figure out if it can be good enough to be considered useful at this stage. Sponsorblock looks great. The reputation score/leaderboard is a superior approach imho.

Sponsorblock is also integrated into piped for example https://piped.video/