Built me a little "Goldilocks" zone finder. It's based on a feedback-loop with a convolutional shader that has 13 parameters - the problem is that about 95% of random parameter combinations result in a dead system that either goes to full black or white, but there is a zone where the parameters are just right and the system stays alive, but it's not a simple "stay in this range" issue. So I trained a small model which tries to predict which parameter combinations are promising in order to explore the space of interesting systems. And now the model keeps improving itself by continuously testing its assumptions and retraining itself. And its "hit" rate is much better than me pressing the space bar.