@ryanwi @adam @js @nathan @samsethi @dave @misener Some podcast hosts will make them using an LLM. So does Apple Podcasts for playback. I make mine by hand because it’s one keystroke while editing.
(Going back to the original question - I believe that super-chapters are the way here. That chapters json file could include a lot more than it does. I tried to connect you to @theDanielJLewis earlier.)
@james @ryanwi @js @nathan @samsethi @dave @misener @theDanielJLewis
I want my podcast app to make other decisions about chapters or other relevant content and contextual data presentation for me. Based mainly on a transcript.
Once an app can use on-device "AI" the possibilities for a customized listening app are limitess
@podcastguru @james @ryanwi @js @nathan @samsethi @dave @misener @theDanielJLewis
Show me chapters that only contain content i am interested in. Summarize each segment.
Prioritize new episodes for me based on my current interest
Find all episodes of other podcasts with this guest
Generate a shareable video clip of this segment of this audio only episode
The list is endless and limited only by creativity of the user.
@adam @james @ryanwi @js @nathan @samsethi @dave @misener @theDanielJLewis So great list, but I don't grok how the proposal helps with this? I'm pretty sure I could surface these features in an app right now without any changes to the P2 spec.
They wouldn't be "free" though, all of those would require inference work, some more than others.
@podcastguru @james @ryanwi @js @nathan @samsethi @dave @misener @theDanielJLewis
I don't recall any proposal.
Indeed, this doesn't require any extra tags. And yes, all inference, but only when an LLM is running natively on device, which was what was being discussed, or so thought.
@adam @podcastguru @james @js @nathan @samsethi @dave @misener @theDanielJLewis
Hi Adam, the proposal was the one Sam called out in the original post that I published (https://www.podcastannotation.org)
This came from app/platform side work on my end to extract key information from transcripts, but in a hybrid fuzzy matcher + LLM pass processing pipeline, not on device. LLMs alone were not getting the results I was looking for in identifying entities accurately.
