HPC where you try to wring the last ounce of performance from machines insufficiently powerful for the task. Weather forecasting, weapons simulation, protein folding...

That's interesting. I don't think as optimized memory manipulation via convoluted casting and such as "complex stuff". I think of that as optimization.

I thought what the author meant by complex was a gigantic system that needs to use many design patterns for good reasons; the code models a very complex, convoluted domain.

When you prototype. But once it reaches production, you should hopefuly have managed to reduce the complexity to its bare minimum