hey have u guys heard about this new model called Jev from TypeSafe AI
it's not an LLM at all. it doesn't output text, it outputs probabilities and calibrated decisions instead
the founder is Diogo Almeida, one of the original OpenAI researchers who literally invented RLHF. he left cause he thought LLMs were too focused on language and not useful enough for actual automation
the wild part is it literally cannot hallucinate cause you define the outputs in advance. and the cost is insane, output tokens are free and input tokens are billed by the billion not the million
Vercel already swapped out an OpenAI model for Jev on a safety classifier and got 5x to 18x faster with better accuracy
been reading the docs and honestly this feels like a different category of AI entirely. not a chatbot, just a decision engine
anyone played with it yet? curious if the accuracy holds up outside the demos
Top comments (0)