In practice, the real problems don't have to get very large or complicated before just typing the problem into AMPL or clicking on SOLVE won't work -- that is, solver capacities may be exceeded or the running time can become absurdly long.
Then you must exploit whatever special structure you can find in your problem and dig into the algorithms to make refined usage. So, maybe do some Lagrangian relaxation. Maybe find an embedded network flow problem and use a network flow version of the simplex algorithm.
most people get by just figuring out how to put problems in excel and hit "SOLVER" (or Matlab and fmincon) but there is some value in figuring out the algorithm works. then it's a hop skip and jump away from the other classes of optimization that you see in machine learning.