Comment by stiff

This has nothing to do with simply having a "fast computer", it would be all fine if we just used 486s clocked at 10Ghz, but we don't, the changes in the architecture of CPUs do indeed make the theoretical model more diverged from reality, because it presupposes a lot of things, like that every instruction takes the same time to execute, also implying that instruction execution times are independent, that one instruction gets executed at a time, etc. If those assumptions do not hold you cannot assume the factor is really constant, and even if it were so, you cannot dismiss this difference indefinitely on this ground, if on datasets of the size commonly encountered in practice the algorithms with a larger asymptotic complexity start outperforming the ones with a smaller one. For matrix multiplication this was known for years, for linked lists it is a relatively new development that came with larger and faster CPU caches etc.