I suspect Youtube cripples transcript usefulness because it aligns with what customers want. And customers are important to Youtube - while end users are something else.
Exactly, and the same explanation holds for any odd UI quirk you see in a for-profit SaaS. Spotify for example, has a lot of frustrating aspects to its UI, but everything makes sense in the context of $$$
I assume YouTube's core metric is watch time and so building features that reduce watch time is a no go.
Not really related, but I thought of this. I used to work on a popular search engine. Our main metric started out as number of searches and number of results clicked on. That made sense as we were working on increasing the number of relevant results. Eventually we had to pivot the opposite direction because we wanted people to find the best result the fastest, which means less searches and less clicks on results. Kind of interesting how a core metric can flip like that.
I could see YouTube potentially making lower watch time a goal if they make the experience of finding useful information so slow that people go elsewhere.
i’ve been wondering the same about twitter / earthx audio spaces. the public ones are playable to the public and are transcribed with automation in real time.
many people listen to spaces after they’re over like a podcast or other form of broadcast recording.
searching public twitter space transcripts would allow us to find more relevant things to consume perhaps.
same with youtube. i wonder whether anything is logging and indexing these with a non-internal program; cursory searches seem to imply that it’s unlikely.
> June 8, 2024 - YouTube has implemented new anti-crawling measures which hampers subtitle retrieval. Crawling is currently suspended and new data will not be loaded. Existing data is not affected. I am investigating potential options and will keep you updated. This change blocks multiple projects and players, including YouTube's own players, requiring a login to play videos.
Only somewhat related: I've been using a variant of the following LLM prompt, obtained from Raycast Prompt Explorer[1], to get summaries of YouTube videos before I watch them.
I use it with Raycast and the Raycast browser extension, but it's easily adapted for other AI browser extensions, as well as scripts and other launchers.
It's helped cut out a lot of cruft.
--BEGIN-PROMPT--
Create a summary of a YouTube video using its transcript. You will use the following template:
"""
## Summary
{Multiple sentences summarising the YouTube video}
## Notes
{Bullet points that summarize the key points or important moments from the video’s transcript with explanations}
## Quotes
{Extract the best sentences from the transcript in a list}
check out https://lectura.xyz. It’s a really powerful chatbot built to analyze YouTube videos. You can also easily search the transcript and jump to specific times in the video
16 comments
[ 5.3 ms ] story [ 68.1 ms ] threadI suspect Youtube cripples transcript usefulness because it aligns with what customers want. And customers are important to Youtube - while end users are something else.
Not really related, but I thought of this. I used to work on a popular search engine. Our main metric started out as number of searches and number of results clicked on. That made sense as we were working on increasing the number of relevant results. Eventually we had to pivot the opposite direction because we wanted people to find the best result the fastest, which means less searches and less clicks on results. Kind of interesting how a core metric can flip like that.
I could see YouTube potentially making lower watch time a goal if they make the experience of finding useful information so slow that people go elsewhere.
many people listen to spaces after they’re over like a podcast or other form of broadcast recording.
searching public twitter space transcripts would allow us to find more relevant things to consume perhaps.
same with youtube. i wonder whether anything is logging and indexing these with a non-internal program; cursory searches seem to imply that it’s unlikely.
I use it with Raycast and the Raycast browser extension, but it's easily adapted for other AI browser extensions, as well as scripts and other launchers.
It's helped cut out a lot of cruft.
--BEGIN-PROMPT--
Create a summary of a YouTube video using its transcript. You will use the following template:
"""
## Summary
{Multiple sentences summarising the YouTube video}
## Notes
{Bullet points that summarize the key points or important moments from the video’s transcript with explanations}
## Quotes
{Extract the best sentences from the transcript in a list}
"""
Transcript: {browser-tab}
--END-PROMPT--
1. https://prompts.ray.so/