Are there any packs you would specifically recommend?

It's so good! Imagine your grilled cheeses, but way sweeter - and tasting very vaguely of tomato adjacent chemicals.

With TOMATO SOUP?! That's some bougie grilled cheese! We just do ketchup!

It's clear it uses the tokens it's leaving in its own wake to predict even better tokens - and that there is some very excellent training to encourage it to not predict that eos token right away - but I'm missing the final jump to "more than just predicting the next token".

Great series of posts! However I still can't resolve this kind of thinking with: "But it's literally 'just' predicting the next token", or maybe series of tokens (depending)

Hell yeah!