So:
- If it is rewarding, do it -> Someone will game the system
- Game the system -> Follow the gradient
- Gradient reaches the local minima - > collapse the system because proxy minima reached
- The optimizer figured out the 'surrogate intent' even if the designer's true intent was something else

Then:
- Essentially tokenomics or economics or even AI alignment in general works (unless value extraction before system collapse is your objective) only works when the 'intent' is opaque or does not have a uniquely global minima
- You need to setup a non-convex problem, where the intent is explicitly unlearnable

So the goal should be:
- Don't make the intent fully learnable, ever

creating and destroying; flying cars, roc.camera, july.rocks