Below is a lightly edited transcript of the btrmt. lectures podcast. For the article that inspired it, see Honey-bees are smarter than they should be.

Welcome to the Betterment Lectures. My name is Dr Dorian Minors, and if there’s one thing I’ve learnt as a brain scientist, it’s that there’s no instruction manual for this device in our head. But there are patterns. Patterns of thought, patterns of feeling, patterns of action, because that’s the brain’s job: creating the patterns that gracefully handle the predictable shapes of everyday life. So let me teach you about them. One pattern, one podcast, and you choose if it works for you.

¶Bees, brains and AI

Now, I’ve been talking and thinking a lot about artificial intelligence lately. I’m having a lot of conversations about it, as I become a sort of de facto subject matter expert here at the Royal Military Academy Sandhurst. That’s partly because of some of the friends I have—a lot of my peers are getting snapped up by AI institutions, or starting companies that rely on AI. But it’s also because I’m using it a lot in my work. And the reason, besides the obvious productivity benefits, is that I find it fascinating.

A lot of the features of generative large language models raise questions that have interested me about human cognition. I see a lot of overlap in the problems we have with AI, and I’ve written a bunch of articles about this. And in the solutions, the ways of improving AI, you also see overlap with the way we treat and encourage humans to flourish.

It reminds me of a long-standing interest of mine. I did one of my research projects in cognitive neuroethology, looking at how bees could do human-level thought. Just the simple honeybee. Very smart animals. What does it mean that honeybees can do things we once thought only humans could do? I feel like that same interest is getting perked up by thinking about, and using, machine intelligences like large language models.

So what I thought I’d do today is talk about how smart honeybees are, and what kinds of questions that raises for how special the human mind is. A blast from the past, so to speak. And then I’ll wrap up, hopefully, with some reflections on what that means for artificial intelligence. Come along for the ride.

¶What makes humans special?

I have this line from the article I wrote on this subject, which I think I took from the paper I wrote on it. But I love it. It goes like this:

The tiniest creatures show glimpses of the highest orders of cognitive ability. Plenty of animals do very clever stuff they really shouldn’t be able to do, given the way we currently conceptualise cognition. This raises obvious questions about how, but it should also make us wonder about what really makes humans special from our apparently less sophisticated brethren.

I like that line because it really was a project that made me wonder why we think humans are special. Any time you look into the kinds of things that make humans special, you often find out that other animals can do them too. There is obviously something special about humans. We have this huge impact on the environment, and we do things it doesn’t seem like other animals can do. But when you boil it down, it’s something like a matter of scale rather than a matter of kind.

It’s really hard to figure out what it is that makes us special. Other animals can pass the kinds of cognitive tests we think are quite difficult. Other animals can use language in quite impressive ways. Other animals can engage in culture—building works of art, really. Think of the bowerbirds, the classic example, with the elaborate nests they build. Tool use. Social habits.

What I want to talk about today is how honeybees are the most fantastic example of this, because honeybees can do a great number of truly extraordinary things. And the one I really want to talk about is how honeybees appear to be able to count.

¶Abstract concepts

Honeybees being able to count would be an incredible phenomenon, because it would show that they can learn abstract concepts. Numbers are abstract concepts, and abstract concepts are basically ideas about the relationships between things that go beyond their perceptual features. General rules, or principles, that apply across different contexts.

The number two isn’t related to anything in particular. You could have two pencils, and that would be two. Or you could have a pencil and a water bottle, and that’s also two. Or a car and a laptop: also two items. Two, as a concept, is abstract. It doesn’t rely on the features of any particular thing.

Another abstract concept is equivalence. You know that a litre of water in a water bottle is the same as a litre of water frozen into a bunch of ice cubes. It looks completely different, but you know that it’s equivalent. There’s something about the sameness of it.

And what you’ll notice about abstract concepts, hopefully, is that they’re quite difficult. I’ve just named two that children can’t do, or have to learn how to do. Equivalence is a classic test: you can look online and see children failing to understand that if you pour the water out of the cup and into some other container, it’s the same amount of water. Similarly, we have to teach children how to count things, how to use basic mathematical concepts. They’re difficult.

So it’s probably no surprise that for a really, really long time we’ve been under the impression that animals can’t grasp abstract concepts like these. In the 17th century you had John Locke, who famously wrote, “Brutes abstract not.” Like language, like culture, like the sharing of ideas, abstraction was something we thought made humans special. And it wasn’t really until the late 20th century that we started to test it properly, and found some surprising results.

Those results are what I want to talk about, because they’re the things that really make us question just how different humans are from other animals. And it all starts with a single experiment.

¶The match-to-sample task

The match-to-sample task is a basic test of working memory. We used it a lot across the cognitive sciences—or at least we did. The last time I really studied this was ten years ago, in 2016, so things might have changed. I don’t know how much we use it now.

Let me give you an example. Say I’m standing in front of you, holding an apple. I show you the apple and say: this apple is what we’re going to call the sample. Then I turn around, do something behind my back, and turn back. In one hand I have an apple, and in the other a pear. These are the alternatives. And I ask you to match the sample: pick the alternative that’s the same as the sample. Because I showed you an apple first, you’re not going to pick the pear. That’s the wrong alternative. You’re going to pick the apple. And if you get it wrong, I’m going to think there’s some problem with your working memory.

Pretty straightforward. I show you sunglasses, turn around, turn back, and I have sunglasses and a notebook. Hopefully you pick the sunglasses, because I showed you those first. You can probably tell by now that I’m just looking around my office and picking things in my immediate view, which shows you how mundane and anodyne this tool is.

It’s not very exciting as a diagnostic tool. That’s why it’s surprising that it has sort of revolutionised how we think about animal cognition. And it all comes down to how simple the task is to teach. Because now that you’ve learnt how the task works, you can do it with anything. If I show you my water bottle, turn around, turn back, and I’m holding a water bottle and a pot plant, you’re probably going to pick the water bottle. You could do this again and again, all day, every day. You’ve learnt the task because you’ve learnt that the task is the abstract concept of identity, or sameness. You don’t really need to pay attention to the items. You just need to choose the thing that’s the same as the sample, regardless of what the sample is. An abstract concept.

¶From chimpanzees to goldfish

This task got taken up in cognitive ethology, in comparative intelligence—an entire domain of research about what kinds of concepts animals can learn. A very famous researcher, Roger Thomas, defines the field like this: an animal’s general cognitive ability is determined by how many kinds of concepts it can learn. And if I remember right, he emailed me when I published my paper, to make sure I knew about his taxonomy of concepts. We were all very interested in mapping animal cognition to these categories.

By the 1980s, we’d only got as far as thinking non-infant humans and language-trained chimpanzees were clever enough to solve problems that required abstract concepts. But as it turns out, that was only because we didn’t have a simple enough test. I’ll link to the paper where I did a little review of this. Somebody introduced the match-to-sample task to an animal and it caught fire.

From the 80s on, I found papers successfully testing this task in non-language-trained rhesus and capuchin monkeys, in gorillas, orangutans and baboons. The obvious follow-ons for people who thought only primates were sophisticated enough to solve these kinds of problems. But eventually particularly clever researchers figured out how to test rats, dolphins, sea lions, corvids—crows, jays, that kind of bird—pigeons, parrots, budgerigars. A lot of birds. Horses were one I remember. The short-beaked echidna was my favourite example. Ducklings. And there was even a maybe questionable paper that managed to do it with goldfish.

So a huge, incredible range of animals. And this is no small feat. Building something that lets sea lions match samples to alternatives, and then enticing them to do it, is an incredible display of perseverance on the part of the researchers. It’s difficult enough, let me tell you, getting humans to come in and do one of my brain science experiments, never mind training a sea lion to do one.

For those of you wondering how you get a sea lion, or a goldfish, or a short-beaked echidna to do a match-to-sample task: there’s an entire literature on whether these complex set-ups are good enough to demonstrate concept learning. Some feature of the contraption could be biasing the animal in a way that looks like abstraction but isn’t. That’s probably true in some cases. But I’m less interested in that, and more interested in the sheer volume of successes, which suggests it’s probably got some validity to it.

In the 17th century, Locke said brutes abstract not—the thing that separates man from beast. By the 90s we had the short-beaked echidna doing it. Quite the transformation. And then in 2001, this paper came out showing that honeybees could solve the match-to-sample problem.

¶A brain smaller than a voxel

I want to give you a sense of how incredible that is. We’re taking the kind of complex cognition we once thought separated man from beast, and we’re looking at a honeybee that seems to be able to do something like it.

That’s incredible enough. But I want to concentrate on the brains involved for a second. You’ve probably heard, if you listen to this podcast and have some passing interest in the brain, that complex cognition happens in the prefrontal cortex. Or the cortex, or the neocortex, or the frontal cortex—some variation on cortex. We love to locate complex cognition in the mammalian neocortex. For your reference: imagine a brain, with its wrinkly exterior. If you unfolded that exterior it would be like a big sheet of paper, crumpled up and wrapped around the outside of the phylogenetically older bits of brain. Historically, we’ve loved to locate really complicated cognition there, particularly in the frontal region. There are reasons for this, and I’ve written about it before—some of them legitimate, and some less so.

What I want to point out is that honeybees don’t have a mammalian neocortex, not least because they’re not mammals. But it’s also a matter of scale. So, to give you a sense of scale, let’s talk about functional magnetic resonance imaging: fMRI. It’s one of the ways we scan brains. If you’ve ever seen an image of a brain with part of it highlighted, that’s fMRI. We use it to see where in the brain there are more or less active neurons, and then infer which regions are responsible for an activity. Say you’re doing an activity, and I want to find out what part of the brain is involved in it. I’d get you to do the activity while I scanned your brain, find out which parts were more active, and that would tell me a little about which parts were involved. And that’s how we’ve spent a lot of time locating complex cognition in the mammalian neocortex: getting people to do complicated tasks, scanning their brains, and saying, “Oh, look, the front of the neocortex seems to be pretty active right now.”

The unit of measurement in fMRI is called a voxel, and a voxel is a cube of brain. We’ll say, let’s look at a region that’s 40 voxels big—40 cubes of brain. And maybe we’d compare it with another region that’s only 20 voxels big, and say there’s more of the brain involved over here and less over there. It’s just a way of measuring the amount of brain involved in something, or isolating a region.

So it’s a tiny cube of brain. A voxel can be anywhere between, say, 1.5 millimetres and maybe 3 millimetres cubed. Get your fingers, squeeze them together, make a little cube out of your index finger and thumb on each hand: it’s about that big. It’s tiny. Inside that voxel there are somewhere between 500,000 and 1.5 million neurons, depending on where you are in the brain—some parts are denser than others. About a million, on average. That’s our smallest unit of measurement in MRI.

Now, a honeybee has about 900,000 neurons in total. One of its bigger structures, the mushroom bodies, has something like 400,000 in it. But in total, you’re looking at something smaller than our smallest unit of measurement. We couldn’t tell which regions of a bee brain were active in an MRI. We could only tell whether the bee was alive or dead.

That’s how far a honeybee is from having anything like a neocortex, the smart part of the mammalian brain. And yet they can do this task. They can match to sample. They can solve a problem that seems to be about identity.

¶A circuit you could solder

I should point out that they can’t do it well. Most animals we get to do this task can do it most of the time, something like nine times out of ten. If I asked you to do it, you’d probably do it even more often than that. Honeybees don’t do nearly as well—maybe three quarters of the time. But that’s still far better than chance, so something’s going on.

And some computational modelling, part of a paper I co-published ages ago, showed that this not-very-well, three-quarters-of-the-time performance could be achieved with a very, very simple neural circuit. A circuit so simple you could probably wire it up in your high school electronics class.

We thought abstract concept learning was an extraordinary phenomenon, reserved for the most sophisticated of creatures: humans. By 2016, we were finding out that it could be achieved with the kind of project you use to teach high schoolers about soldering. As the kids would say, it’s not that deep.

I mention in the introduction to this little podcast of mine that, if there’s one thing I’ve learnt as a brain scientist, it’s that there’s no instruction manual for the brain. I cut something from that introduction that I truly believe: that we are animals first. Because there’s one thing I keep learning as a brain scientist, over and over again. It’s very, very hard to figure out what it is that makes humans special. Which should probably make us mistrust the idea that we are.

So let me wrap up now with a slightly longer implications section, because I want to talk about the relationship between this ancient paper of mine and AI.

¶Counting, and the concept of zero

Beyond this weird, very specific working-memory task, honeybees have been shown to do a bunch of incredible stuff. I mentioned counting. They’ve been shown to count landmarks—to work out that the reward is behind the third landmark, or the fourth.

More interestingly, they’ve potentially learnt to understand the concept of zero. For reference—and I’m going to get this wrong, because I’m not a history-of-maths person—I think the Arab caliphate discovered it, then we lost the knowledge, then we rediscovered it in the Enlightenment, something like that.¹ Whether I’ve got the details correct or not, what I’m saying is that we apparently needed an incredible flourishing of scientific enterprise to discover the concept of zero, even though we take it for granted today. Because it is kind of weird. It’s not a number. It’s the lack of things. The empty set. There’s nothing in zero. For me, logically, it comes before one. But if we hadn’t been taught that, perhaps we’d never have come to the conclusion that the lack of things needs a name.

And bees seem to understand it. This 2018 study showed bees cards with one to six shapes on them, and trained them to choose the card with fewer shapes. That’s them counting: they go to the card with fewer shapes rather than the card with more. Then the researchers gave the bees a blank card, and the bees naturally treated it as something that goes below the card with one shape on it. They treated nothing as a quantity smaller than one—which is, for all intents and purposes, a concept of zero.

Again, something we once thought was the purview of humans and some primates—and, you might have caught in my litany of animals that solve the match-to-sample task, birds, particularly corvids, which are very, very cognitively sharp. So it’s crazy. Bees seem to be really, really smart in ways we think are quite special.

I think the same lab did some follow-up studies that were quite stunning: arithmetic-based stuff. Obviously there’s pushback. You can probably tell just from listening to me that maybe they’re responding to magnitude, the amount of stuff on the page, rather than strictly counting. And when it comes to training animals there are always complications in the training itself. What’s rewarded? What’s punished? What are they actually responding to?

But my point doesn’t rely on them being able to count. My point is that whether honeybees are doing strictly what we’re doing—counting with numbers, mathematics—or something functionally the same, this process of abstraction on a simpler level really blurs together our notions of higher-order and lower-order processes. Complex cognition and simple cognition. We once thought abstraction was complicated, and now we’re realising that maybe there are grades of complexity even within abstraction, and that it can actually be done very simply.

¶Bees learning from bees

Bees are the case study in point. I did my work with honeybees, and I don’t know so much about the bumblebee research, but there’s a whole series of papers looking at how bumblebees socially learn from each other. They learn tasks that require multiple bees to solve.² They learn how to do a task from other bees. They appear to be animals that play. All things we once thought were the purview of humans, because they need a very complicated brain to read the mind of another animal, to interpret what it’s doing and its intentions, and to port that knowledge into your own. Except apparently they don’t, because bumblebees can do it.

And I have handfuls of the little marginalia I write—interesting tidbits from around the web—showing this in all sorts of other animals. Single cells seem to be able to learn; there are papers about single-cell cognition. Ants are super cool in the same ways honeybees are: they can do a lot of this complicated cognition. I’ll link some of these in the show notes.

¶Minds solving their niches

The point I’m trying to make is that there’s a history of treating human cognition as special when it really doesn’t seem to be. And when I observe developments in artificial intelligence, large language models in particular, I notice a lot of overlap between the behaviours they engage in and the stuff we once attributed only to humans.

In some ways that’s striking and unsettling. But perhaps it shouldn’t be. Because if cognitive ethology, this comparative intelligence literature, suggests anything at all, it’s that any animal can solve complicated problems within the context of its ecological niche. And these large language models are, in themselves, kinds of minds. Neural networks are our attempt at simulating how brains work. This whole domain of technology comes out of the kind of brain science people like me were doing on how brains work. They are kinds of minds, and of course they’re going to solve the same kinds of problems all animals need to solve.

To the extent that there’s nothing special between human cognition and the cognition of other animals, there’s probably nothing special between human cognition and the cognition of these machine intelligences. Hopefully some of my other podcasts, my other articles, will convince you of that. But if nothing else, I hope it makes you look at the news articles about how scary and unnerving these things are, and realise that it’s just a product of minds: things that perceive the world and act upon it, solving the problems that face them. Just because it looks like human cognition doesn’t mean that it is.

And that, I think, is where I’ll leave it.