and the seed lottery. You can see this if you try training a simple network with two inputs to learn xor. It can converge in multiple ways, and sometimes it converges to a really bad approximation. And sometimes it doesn't converge at all (or it converges so slowly that it might as well be considered not to converge). And even then it might still converge to an approximation that's more accurate on one side of the input space than the other. Tons of ways to get an undesirable result. For a simple 2-input network.

Imagine how unlikely it is for txese models to actually converge to the optimal thing. And how often the training is for nothing.