Reading Jev as a next-move predictor for custom small worlds
The author argues Jev's value is letting users define the problem and solution space themselves, citing FPS games and self-driving as cases where limited inputs and outputs make prediction more reliable, and says they are eager to try it.
my take on Jev: i think Jev is filling the gap and complements what current LLMs do. it lets you define a smaller, well-defined world, and the model just needs to work and predict the next move according to the context in this environment. this could simplify the next move prediction a lot. and you are in control of defining the world the way you want. LLMs on the other hand are built for a huge world with a huge number of possible outputs. this makes them versatile, but also leaves more room for predicting the wrong next step. consider a first person shooter game. the inputs could be much more limited, like your position, health, ammo and distance to enemies. the output is also well defined: move, aim, shoot or reload. same idea for a self driving car. the world could be represented by object positions, car speed, direction, etc. and the output could be the next steering and acceleration change. obviously driving itself is still hard, but the interface can be much simpler. the ability to define the world, both the problem and solution space, feels huge. super excited to try it.
This post quotes another post: “ChatGPT co-inventor releases Jev, a model built for decisions rather than text”