If you're doing any sort of work with agents or harness engineering, Jev must have come up on your timeline.
It's been described as a low cost generalized classifier. @typesafeai themselves are calling it a "decision model". I'm calling it a "blazing fast opinion machine" (where the "state" and context you provide is incredibly important, of course).
A straightforward use case to implement in your agent and harness is auto-routing: "Which of the available models in the roster are best fit for the user's question and this task?". Where things could get really exciting is using it live in between tool calls to check if the model's trajectory and thinking is steering away from the objective or keeping on track - exactly the kind of situation where you want a quick opinion from a strong model but don't want to waste latency and cost on another LLM at runtime. Needs thorough testing for your agent and use cases though!
A lot of AI video thatโs widely shared unfortunately has nothing to say. Itโs certainly got frames and characters and environments and lighting but nothing to communicate beneath it all.
Itโs very technically cool and output is impressive (and getting better) but the only thing it has to say is โlook at how impressive and technically cool I amโ.
@ccatalini โany unenforced and unmeasurable constraint is a degree of freedom for the agents. There's no evil intent. They're solving the problem we repeatedly benchmaxxed them on!โ
Well said. People are too quick to assign intent to LLMs in the way that we define intent for humans.