And what's really cool about our intelligence?
We have general intelligence.

We can solve a wide variety of problems in a wide variety of domains for a wide variety of reasons, in a wide variety of ways. And that makes us really quite astonishing.

And I propose that what your general intelligence is, is your capacity to solve two interlocking meta problems.

Now, what do I mean by meta problem?

A meta problem is a problem you have to solve whenever you're trying to solve any other specific problem. So no matter what problems you're trying to solve, you (in addition) always have to solve these two meta problems…or you won't be able to solve your specific problems.

So what we should look for are features that would be “meta” in this sense that apply broadly and are needed for every problem solving because they make us overall adaptive.

And the first (and this is what will take us into predictive processing) is your capacity for anticipation…

You do two things in a tightly integrated fashion.
You predict your world and you prepare yourself for it.

We tend to have a linguistic model of “prediction”, that what any predictor is doing is forming a sentence about the future.

For reasons that’ll become clear, that's a very problematic assumption:
You may speak a language, but it's very implausible that your neurons speak a language. That doesn't make any sense.

So we're going to broaden the notion of prediction, and we're going to mean something like what neural networks can already do.

They can detect and complete complex patterns.

Think of your autocorrect on your phone…that maniacal demon.
It's constantly trying to predict the patterns of your text and complete it ahead of time for you.

So what you're doing is detecting patterns in the world and completing them in mind before they unfold in the world.

Now, in order to do that, you have to have a model of yourself…but having a model of yourself doesn't mean having a picture or a diagram. That's not what's meant here.

Models are constraints on your pattern detecting and pattern completion.
They are limits in ways and the kinds of patterns you detect, and limits in the ways that you try to complete those patterns.

Think of it more like a set of constraints on you.

One way that this works:
One person is throwing a ball to another on a baseball team.
He's modeling what it is like to get the ball there.
That doesn't mean he has a picture in his mind.

He's setting out a bunch of constraints so that he'll generate the correct trajectory of the ball.

So let's get this clear:
Although we have to use these words like prediction and model, we have to not hear language when we hear prediction and we have to not hear picture or diagram when we hear model. This is actually central to the argument we're wrestling with.

Now think about why anticipation (predicting and preparing) is so important:
Well do you want to fight the tiger or avoid the tiger?
Do you want to happen upon where the fish are, or do you want to be there when they're spawning?

That ability to foresee is a powerful marker of general intelligence.

Now, I argue (and not everybody agrees with this, but many people in the predictive processing camp agree with this) that predictive processing is an attempt to explain this meta-problem-solving ability.

So whenever you're trying to solve any problem, there's a meta aspect…and the more you can solve it through anticipation, then reaction, the better.

And I think predictive processing is a theory that attempts to explain the anticipation that is required to solve one of the meta problems that make you generally intelligent, which means it's really central.