ChatGPT training method co-author argues text and decisions should be split, with Jev as the middle layer
Diogo Almeida's talk from six months ago argued chat models stop being reliable once something has to be decided, and the article splits the work: text stays with the large model, decisions go to a typed model that returns a probability, exact rules go in code, and Jev is the middle layer, connected in about 10 minutes.
Diogo Almeida pitched TypeSafe from a stage 6 months before Jev existed, and the problem he put in that pitch is the one everybody runs into now he co-wrote the training method behind ChatGPT, so it's worth hearing which part of it he thinks broke a chat model is reliable enough to write your emails and stops being reliable the moment something has to be decided that's the split the whole article rests on: > text stays with your big model > decisions go to a typed model that returns a probability > exact rules belong in plain code and in neither one Jev came out in September as the answer to the middle line, and connecting it takes 10 minutes the quality of every decision after that is one paragraph of text per question, and the ten rules for writing it are in the article
This post quotes another post: x.com/1450821446859624454/status/210275561781110…