Turbowarp solved that one good setup is bias at 2.64, neural weights -0.28, 0.34, 0.3, and -0.4. I made this at https://scratch.mit.edu/projects/1321759554/
