Folks, I've been looking into LLMs creativity and writing style recently, expiriencing the common fructration there. The problem is obvious, but it doesn't seem to me that there is a merit for a startup/company. Solution seems to be a new loss / protocol in RLHF / RL posttraining. And it feels like that frontier model should be capacle of that (improving comms and style), without regressing on the core metrics. So, when any frontier and sem-frontier will care - this should be solved quite easily (by them, not saying this is a simple problem) - and destroy the moat for any startup. Curious, what people here think - why this is a theme in YC request for startups? What am I missing?
in the short term, obiously frontier lab will focus on Coding. and the output distribution is different from the creative writing
if someone proves it correct, like JEV from Typesafe, frontier lab can quickly train a simple version like the gpt one. But Jev is still the best in the short term.
the proplem is whether the market is large enough so that you can be a Oligarch so they can't surpass you soon. and you can develop a totally different way to do this other than post-train based the sota Opensource LLM. which means, gpt fintuning can't beat you.
i am Vincent, and doing a creative content startup called astorie raised few million dollar. if you want explore more about creative area, and discuss more with me, welcom to email me. viz@astorie.ai
I've got an idea to train this model bypassing privacy and copyright issues (so nothing related to YouTube videos, social medias, Reddit forums, WhatsApp and Telegram chats etc.). If someone wants to discuss it, just email me at diegobarreca2004@gmail.com