Launched this week
JevGPT
A chatbot built on a model that can't write
68 followers
A chatbot built on a model that can't write
68 followers
For years we've used LLMs built to write text to make choices. JevGPT returns the favor: a chat app where TypeSafe's Jev, a model built to make choices, writes every reply one multiple-choice word at a time. Does it work? Sort of. Open source.
@idlivada The “sometimes” part about strawberry having three r’s made me laugh 😂 It sounds like a fun experiment to test.
that's a really cool idea. It sounds like one of those fun challenges that someone takes on just to see: okay, can we actually pull this off? 😄
I really like this weird feature where a choice-based model somehow writes an entire conversation by entering one word at a time.
@maksym_stelmashenko Yes exactly. I just built this to see if it could be done. Actually, years ago this was effectively how text generation worked using classifiers before the deep learning era. See my paper on it here: https://aclanthology.org/N07-1056.pdf
appreciate the honesty of "does it work? sort of" more than most launch copy here today. genuinely curious what breaks first at scale - does picking the reply word-by-word from a fixed choice set start degenerating into repetitive/safe phrasing once you get past a few sentences, since it's optimizing for "best next choice" rather than anything like a coherent plan for the whole reply?