2025-06-29

Theseus devlog

/projects/youtube-of-theseus

  • I was able to make it control the youtube UI (tested "hide shorts" button)

  • Language is the barrior. It has pretty hard time understanding video's language based on its title. Some are even auto-translated making it even harder to guess. But I found open-source LLMs are just not capable of understanding the context. It cannot even distinguish Korean and Japanese

Seems like my expectation was too high. I think it's nearly impossible for LLM to simply guess the video's language from its title.

  • Youtube tab snapshot is pretty big. I think I need some pre-processer that can summarize the current screen (and only the current screen. ecluding elements that aren't shown)