JevGPT
An experiment that makes a non-generative decision model assemble a chat reply one word at a time.
An experiment that makes a non-generative decision model assemble a chat reply one word at a time.
Jev receives the conversation transcript as state. At each step, bucketed Choices propose vocabulary words, a final Choice runs off the bucket finalists, and a Noul estimates whether the reply is complete. Local code applies repetition penalties and decodes the selected word until completion is allowed or the word limit is reached.
Is the reply complete?
Conversation transcript, latest user message, and the partial reply generated so far.
The client combines conversation history, the latest user message, and the partial reply.
Code shuffles the vocabulary into bounded Choice buckets and asks in parallel, alongside a completion Noul.
The two leading candidates from each bucket compete in a final Choice.
The client adjusts the returned distribution with repetition penalties, appends a word, and repeats until the stop rule or word limit.
It tests the boundary between typed decisions and text generation: a chain of word classifications can produce readable fragments, while bucket comparisons, stop conditions, and repetition control expose the costs and failure modes.