My favorite example is implementing NEAT with keras dense layers instead of graphs. Last time I tried with claude 3.7, it wrote code to mutate the output layer (??). I tried to prevent that a few times and gave up.