I largely agree, but there's a little nuance insofar as "interior-point" methods are very powerful. You can go a long way by encoding your goals as error functi

I largely agree, but there's a little nuance insofar as "interior-point" methods are very powerful. You can go a long way by encoding your goals as error functions and letting a gradient-based optimizer do the rest.