How Jev's decision model differs from an ordinary classifier
A decision model takes the state plus a schema of typed questions and scores every valid option in one pass, so output stays structured and latency stays roughly flat as fields are added.
It has been trendy to make fun of Jev but it is a real innovation and @CompleteSkeptic is going to Valhalla regardless of what happens to @typesafeai A normal classifier answers one question it was trained for and then you are done A decision model takes the state plus a schema of typed questions and scores every valid option in one pass You are not generating tokens and then hoping the parse works, and you don’t need to add a whole bunch of validation logic The bad answers are just not in the schema so the output stays structured Latency stays roughly flat when you add fields because it scores them in parallel after one prefill That’s the actual change For agents or apps that need to decide on the fly what to do next this is a different tool not a faster classifier
This post quotes another post: x.com/ritakozlov/status/2105683951595725177