solver stops early on non-convex optimization

[jeisner]

Trying lesson 9 with L2 regularization, C=-3 or C=-30. The solver seems to stop after a few iterations when the gradient is not 0. We should still climb to a local maximum (possibly with some infinite weights), no?