🅂🄰🄱🅁🄸🄽🄰 🄿🅈🄻🄴🅂

@rustacean_matt These vegetables aren’t gonna cut themselves, Mattsquatch.

@eggs_mayhem You're not concerned that your instinct after consuming alcohol is to go buy a weapon?

Pounding a few beers and calling an Uber to Bas Pro Ships to pick up a sick knife is self-care, right?

My daily morning pep talk: “I can do this!” “No you can’t. You’re dumb as hell.” “Right, but I’m *this stupid* but I’m *still alive*. Which means I must be a genius.”

@xeraa Ohhh nice. Thank you!!

@eggs_mayhem you should be able to add the offset (or whatever metadata you need) to the nested structure. I assume like this? gist.github.com/xeraa/c827d265… left mapping, right result of the query

Elasticsearch is more robust for my search function, which searches a huge amount of documents, but chromadb’s timestamp feature is so crispy. Sigh.

@DThompsonDev @AllThingsOpen Ohh snap just moved to the area! Gonna have to see about this one.

NORTH CAROLINA! I will be speaking at All Things Open @AllThingsOpen in October! I heard there will be around 4000-5000 people attending! So who am I going to be meeting there?

@leo_guinan I’ll have to throw all the optimizations one at a time to each speed and model combo, because why not?

@eggs_mayhem For some reason, I feel there's a non-zero chance that smaller models would perform better on faster speed audio. I don't think that makes any logical sense, but I feel a possibility that makes curiosity tingle haha

Wrote scripts today to download all the videos from my youtube channel, extract the audio, and transcribe them with local whisper. Knocking out the longest part of populating my chatbot's data.

@leo_guinan You know I never ran a formal comparison, but that’s a great idea!

@eggs_mayhem That's pretty good! Curious: did you try different speeds with different model sizes? I think that would be a cool comparison to run.

@leo_guinan Was hoping for 2x but lost fidelity after 1.4x, unfortunately. Still better than nothing!

@leo_guinan Yeah it did require some minor modifications to the pydub dependencies, but that was fun anyway.

@leo_guinan Yeah it may be good enough for a general recommendations engine. But it was definitely too riddled with mistakes for my timestamp use case 😩

@eggs_mayhem I've been using base, but I'm going to try tiny.en to see how well it does. Running over a couple hundred hours of audio with the main goal of directing to content, so if I don't have perfect matching, I think context will be enough to hit the targets.

Browser extension that reads your cookies and tells how you’re being psyopped and who’s doing it?

@leo_guinan I’ll give it another go this weekend. Getting back into the game and as I’m making a search feature for a broadcaster with over a decade of hour-long episodes, I inherently have to scale pretty fast 🫨.

@eggs_mayhem Yeah, that is weird that they got rid of it. Wonder why? 🤔

Is this cyberpunk?

Tfw when you have to spend as much time programming a fan to cool your old-ass rig while cycling through embeddings generation as you do programming the embeddings automation itself.

@xeraa Just saw you were with elastic! I love their stuff, but found implementing these to be a bit of a blocker. I wonder if I am just missing something? Wouldn’t be the first time.

@eggs_mayhem what's the timestamp feature? like date math?

@xeraa @trychroma makes it very easy to include granular timestamps with embeddings of audio transcripts. Great for enabling a user to pull up audio from a semantic search at the precise moment the relevant audio occurs.

@leo_guinan Honestly surprised that Elasticsearch removed the easy implementation of timestamps. Now seems like a great time to have them, since lots of ppl surely are thinking like us. Could be a great business case for Chroma to focus on.

@eggs_mayhem wow, that is a lot of embeddings haha. I'm pretty excited for it. Going to make my life a lot easier with some of the things that will be included

@leo_guinan I knew that was coming, but didn’t realize how soon. Great timing 😸. Too bad my next compute workload to generate the initial embeddings is two weeks at 24 hr/day 😑

@eggs_mayhem Chroma is releasing their cloud version as early release in the next week or two. Should be pretty good to handle scale I think. New, better architecture with it