would love this brought down to earth a little bit
Yep, this'll be a bit less than technically correct, but it'll get the gist apart. So there's a statistical concept Kullback-Leibler divergence (KLD), the divergence between two probability distributions. When we're talking about the KLD between the real statistical structure of the world, and the statistical model that people have of that world, we're talking about how in-sync a person's map is to the real territory. When there are divergences, eventually reality "bubbles" up—things that shouldn't happen, according to your model, happen, and more often than you'd expect. (This is a bit what the whole "it's 2020 and we're living in a simulation" meme is about, yeah?)
We're talking about subjective surprisal—
how hard is it to explain an even given your statistical model of possible events? If Luka heard a knock on his Londy flat right now and answered it, and it was me, that would be very surprising, because he thinks I'm in Wisconsin right now. He probably thinks the knock is the friend he invited over earlier in the afternoon, and if it's anyone but that friend he may also be mildly surprised.
When you're surprised often, it means you have a bad model of the world. Organisms with bad models die off more often than those with good models, because they cannot predict future events well, they can't pattern-match and make valuable inferences. They're cognitively inferior and that makes them less fit, evolutionarily. You can see this at every level of action: an ability to predict whether the response of a conversation partner will be positive or negative, after a given remark, is part of what we call "social intelligence," just as the famed "basketball IQ" is your ability to
read the court (i.e. size up the environment) and make decisions based on patterns of cause & effect.
The free energy principle (FEP) is a formalization of the idea that brains are dedicated to minimizing Kullback-Leibler divergence, and surprisal generally. Building better statistical models of the environment. Karl Friston, one of the authors on the paper, is the originator of the FEP, and thinks that autopoietic, statistical modeling of the environment is a prerequisite for life itself. It's taking the Good Regulator Theorem (every control system is a model of its environment), applying it to living organisms, and extrapolating it to cognitive models.