[justusm]

nice! Training models using reward signals for code correctness is obviously very common; I'm very curious to see how good things can get using a reward signal obtained from visual feedback

[grace77]

As are we, seems like the natural next step