Even the Little Cousins Have to Eat

I am at AABB 2026 this weekend and will have plenty of blood bank topics to write about in the coming weeks. In the meantime, here's a post I wrote awhile back about my favorite bird family, the corvids.

------------------------------------------------------------------------------------

There is a crow who knows my car.

I have been feeding this particular crow, at roughly the same time, in roughly the same spot, for long enough that something has changed. She doesn't just show up when there's food already down. She shows up when I pull in. Sometimes she shows up a few minutes early, waiting on the wire, which is its own small unsettling thing: not “food is here,” but “food is coming, because it is that time, because of that car.” That is not a reflex. That is a prediction, and it is a prediction about me, specifically, as distinct from every other car that has ever pulled into that spot.

Crows can do this. They can learn individual human faces and hold the grudge or the goodwill for years. They can pass that judgment to their offspring, so that a crow who has never personally been wronged by you will still scold you on sight, having inherited the opinion. Somewhere in a nervous system the size of a walnut, I have become a recognizable category. That is a strange thing to sit with, and it is the thing I keep sitting with.

The brain's actual job

 Here is the more clinical version of what's happening, and it's worth stating plainly because it's the whole hinge of this essay: a brain's core function is not story, or connection, or meaning. It's prediction. This isn't a fact about human brains specifically. It's closer to a design constraint that any sufficiently complex nervous system runs into. The alternative, reacting only after the world has already happened to you, is too slow to keep an animal alive. Survival does not actually require accurate perception. It requires anticipation good enough to act before the world finishes happening. Whatever the hardware, the job is the same: build a model, use it to guess what comes next, and spend your limited processing on the gap between guess and reality rather than on reality itself.

In human brains, that looks like this. Your retina, skin, and cochlea are taking in something on the order of eleven million bits of information per second. Conscious awareness processes somewhere around forty to fifty bit total. The overwhelming majority of raw signal is filtered before reaching you. What you perceive is forecast, built from a model that gets corrected only when the incoming data disagrees strongly enough to be worth the update. This is why eyewitnesses are unreliable narrators of events they genuinely believe they saw clearly. It's why optical illusions work at all; you are looking at the prediction, not the photons. Researchers working in predictive processing, Karl Friston, Andy Clark, Jakob Hohwy among them, have spent a long time formalizing a version of an old idea: that perception is mostly the model, lightly interrupted by the world.

What matters for where this essay is headed is that this compression, this bet-on-the-model architecture, isn't a fact about primates, or mammals, or cortex. It's a fact about what it takes to be a nervous system that has to act in time. Which means, if the framework is right, we should expect to find it again wherever else complex behavior evolved, running on whatever hardware happened to be available there.

The part that should not work, and does

 If prediction is the actual job, and not some incidental feature of mammalian brains, then the job should be doable by more than one kind of hardware. We tend not to act like we believe that. A neocortex, with its neat six-layered structure, gets treated less like one possible solution to the prediction problem and more like a prerequisite, as though cortical layering were the toll booth you have to pass through to get to anything resembling a mind.

Corvids did not get the memo, and they're the cleanest test case I know of for whether prediction is really the general principle it claims to be. They do not have a neocortex. Birds have a pallium instead, organized around nuclei rather than layers, structurally about as far from mammalian cortex as you can get while still being a brain. And yet look at what corvid cognition actually consists of, item by item, and it reads like a checklist of prediction put to work. Tool use is modeling physical cause and effect well enough to forecast which action gets the food out of the tube. Caching food, which crows and jays do constantly, requires not just memory of where the food is but a working model of who else is watching, so the bird can predict whether a competitor will raid the cache and re-hide accordingly. That is prediction of another mind's future behavior, which is the same skill tested more formally in theory-of-mind paradigms, where corvids also perform. Delaying gratification for a better reward later is, at bottom, a prediction that the future state is worth more than the present one, and has been identified in corvids. None of this is a separate bag of tricks sitting alongside prediction. It's the same underlying computation: model the world, model other minds, weigh future against present.

Mammals and birds split off from a common ancestor more than three hundred million years ago. Since then, two completely different architectures have independently arrived not just at similar behaviors but at the same underlying solution: build a model, use it to anticipate, spend your resources on the surprises. Nobody handed that solution down through a shared cortex, because there wasn't one to share. It got invented twice because the problem it solves, staying alive in a world that won't wait for you, is not a mammal problem. It's a nervous system problem.

I find this enormously clarifying and slightly destabilizing in the same breath. It means the thing making the crow on the wire capable of recognizing my car, and my face, and the fact that I am the source of the food, is running the same core operation my own brain is running when it recognizes her back. Not an analogous operation. Not a cute parallel. The same fundamental bet, made twice, on different hardware, because the world only ever offered brains one really good move.

What I don't get to resolve

 The relationship, if that's the word, is not symmetrical, and I don't think I should pretend otherwise. I know something about what's happening in my own head when I look forward to seeing her. I have no comparable access to what's happening in hers, only the outputs: the early arrival, the specific attention, the apparent tracking of me as me across time. I can build a model of her inner life. I cannot verify it. That gap doesn't close no matter how much predictive-processing literature I read, and I've noticed I don't especially want it to close. There's something in the not-quite-knowing that keeps the whole thing honest, and it certainly doesn't stop me from respecting this distant cousin of mine with wings.

What I keep coming back to is smaller than any of that, though. Somewhere in a walnut-sized brain built on an entirely different plan than mine, something has decided that I am worth predicting. That the world is more legible, in some very small way, for including me in it. I don't know what to do with that except keep showing up on time, and keep noticing that the little cousins, the ones who took the other three-hundred-million-year path, have to eat too.