As an ML engineer, I improve models through trial and error. Sometimes a change improves my metrics, and other times it makes them worse, and I just revert it and try something else. It's scary to think about neurological surgery making progress in a similar way. We may not have a good theoretical understanding of why a procedure works, but we observe positive outcomes in a small number of patients and then start looking into expanding its use. It reminds me of procedures where doctors sever connections between the two hemispheres of the brain to reduce seizures. I've heard they can reduce symptoms, but disrupting connections in the brain can also have unexpected consequences. I hope we can do better than essentially throwing spaghetti at the wall.