I've been doing some experimenting in the past couple of days, using "Mistral 7B" which is another open source LLM that apparently outperforms Llama in every way. It also has a much more permissive license allowing for more flexibility in commercial use-cases.
As a non-coder using its out-of-the-box solution, I can only train it based on text files that I have stored locally. So I tried feeding it the some 15 million words that has been transcribed so far, from your two 3speak channels and the cryptomaniacs podcast. Here's an example output:
This is more of a proof of concept of course. The true power of this would come from, as you also describe, live indexing and continuous training on Hive-data. It will be be very exciting to see how (and how soon) this will be implemented by various front-ends and other projects
Very interesting.
So Mistral outperforms Llama? Good to know.