CHOOSEAgents & automation
Jev Performance Enhancement
Krzysztof Staroń@KrzysztofStaron
0:59This project details a method to significantly improve the performance of Jev, an AI tool, by reframing its problem-solving approach from step-by-step execution to outcome prediction.
SignalIt's very good at figuring out what needs to be done, but actually executing actions that move towards that goal is its biggest weakness.
I've been testing @typesafeai Jev.
And I found a way to make it good.
Jev is not an agent. It's very good at figuring out what needs to be done, but actually executing actions that move towards that goal is its biggest weakness.
I found a workaround and it 10x'd its performance.
> You need to "predict outcomes, not steps towards an outcome."
It really matters for many use cases.
Most problems require multi-step planning, and solving them is an interconnected graph of choices. And without this change in perspective, using it for such problems is simply impossible.
This is how this thinking is applied when trying to solve Tetris:
Asking it to "Play Tetris" won't work. It requires Jev to make a plan by itself and execute actions towards that plan.
Asking it "Which button to press" is better, but changing this question to "Where should this piece land" magically bypasses its weakness in multi-step planning.
Even better would be asking "Should you win or lose" and executing actions towards that GENIUS goal, but if we could do that for any problem, we wouldn't really need AI in the first place.
I’m yet to find a solution for more complex decision trees, where one outcome deeply changes the next. Solving that is what would actually make workflow automation with models like Jev work.