All software will approach "human speed," whether it's adding two integers or getting near instant predictions from a GAN. All available resources will be used until it's "good enough. Just ship it," and efficiencies will multiple as in any natural selection: defensive measures like "confusing code," and "unrelated system fails when this one is changed" allowing mostly harmless dumb code to survive and propagate.
Some intrepid learner comes to study the code one day. "Huh, I guess this is how you're supposed to do it. Seems kinda weird but I'm sure it was written this way for a reason. It's been like this for a decade!" And so they copy the procedure into their greenfield project.
Efficient code will always be sacrificed to the god of "Just Ship It" as long as the overall performance is "Good enough." And why not?