Request for Startups – "LLMs That Write Well" – Why?
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?
3 comments
[ 563 ms ] story [ 1242 ms ] threadif 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