A co-inventor of RLHF started a new company a few years ago known as TypeSafe AI. They recently opened up a wait list and it has gained a lot of attention. It doesn't generate tokens, it doesn't hallucinate, and it's much faster than an LLM, up to hundreds of times faster. The TypeSafe name likely comes to it's guarantee that you are essentially "typing" your query and it cannot produce hallucinations.
It's a decision maker.
Give it a prompt and a predefined list of options and it will pick the best fitting one. It can still be wrong of course, it's only going to be as smart as the model is. It could be useful for a ton of things really. I think it would be great at deciding what models to route to, but the most popular examples are around things like credit card fraud. A payment processor needs to determine if a charge is potentially fraud quickly, so you don't spend a lot of time waiting at checkout. The industry uses specialized and performant models of course, but with Jev they can make the decision up to 200 times faster, and for much much cheaper.
It's very early days and many are already rushing to make their own competitors, but there is a ton of promise. I personally believe AI will end up being more modular. Smaller models that are specialized for specific tasks, but share parts like the encoder, Lego's if you will. Jev-type-systems seem like a great way to facilitate the routing in a scenario like that.
Something to keep a close eye on, it could be a game changer for the industry!