Ha! That's super specific, interesting.

It's actually at E-Line, we have some funding thru the endless side to make a game to create a lot of internships as part of the game dev education work. Which is great although sometimes I wonder if it would be easier to do it if I was 😀

That doesn't mean it's limited in the long run either - eventually as they pursue more hybrid models, adding those techniques into this kind of dataset should have bigger gains too.

Maybe a bit better than that - so far research says it gets similar performance to other models of same size (that startup is pushing a 1.2b param model, gpt3 was ~1.75b).

Yeah, it's being used for research, it's on hugging face and has downloads there), and there's startups like this one hellofuture.orange.com/en/p-c-langl...

There's Common Corpus, but that's more for people making their own llms not a service to broad users. There's some research I've seen that says it's not that different for some basic tasks... however I can't remember the specifics and the irony is I'd have to use AI to find it 😆😭

I like to think somewhere, someone, when reintroduced to the codename PQRS, Mandela Effect's it as an actual Roman game 😆

I lol'ed at this old comment
"if you rearrange the letters from the second project they were working on from PQRS to SPQR you get the Latin phrase meaning Senatus Populusque Romanus which means “The Senate and People of Rome.” Perhaps a clue as to what that mysterious game may be about…?"