We know about the world because we have senses.

Sight. Sound. Touch. Taste. Smell.
Balance. Movement. Body state.1

Our senses are the messengers.

This seems obvious.

Most of the time, our senses work in harmony. Your eyes see a ball bounce, and our ears hear the thud at (almost) the same moment.

But not always.

Sometimes, our senses fall out of sync. The messages don’t line up. And when that happens, perception doesn’t go quite the way we expect.

So what happens when what we see doesn’t match what we hear? Or when our brain starts ignoring input from one sense in favour of another? Or when an entire sensory channel goes offline?

This essay starts a new series about how our brains use different senses to verify reality (or good-enough reality) — and what happens when those inputs don’t go as planned. As usual, throughout the series, we’ll also consider how biological brains compare to the way machines work — and whether those differences are the kind of differences that matter.

This week, we’re starting with vision. For most of us, vision is the sense we tend to trust the most.2 Intuitively vision feels direct. Like we have a live feed from the outside world.

So what happens when we mess with our most trusted sense?

To find out let’s answer three questions:

1. What happens when we change the input our eyes get ?

2. What happens if that change lasts for days?

3. What can we these strange experiments tells us about how perception really works?

Subscribe now

Q1: What happens when we change the input our eyes get ?

You might remember this from high school biology: the image projected onto your retina is upside down.

And yet, somehow, you don’t see the world that way.

Maybe your teacher explained it by saying, “The brain flips the image.”

It’s a nice story — but that story isn’t quite right.

We’ll come back to this point in a bit. But first, let’s look at what happens when we deliberately invert the input that hits our retinas.

Imagine waking up one morning and putting on a pair of inverting goggles. These strange looking things flip the entire visual field upside down. Imagine how disorientating this would be. Imagine trying to navigate the world if the sky was down and the ground was up!

Strange as it sounds, that’s exactly what volunteers did in a series of classic experiments.3 These weren’t quick lab sessions, either. The goggles stayed on for days— sometimes weeks. In a few rare cases, even months.4

Imagine that!

Imagine reaching for your coffee pot. But your coffee pot is upside-down and somewhere near the ceiling which looks like the floor. If you reach up, you see your hand move down — away from the pot. If you reach down — away from where the pot appears — your hand moves toward it. Your eyes and your body no longer agree.

Things get especially weird when you try to pour yourself a cup of coffee. According to what you see, gravity works backwards. Liquids fall up. But of course, gravity hasn’t changed at all.

Yesterday, your vision was your most reliable sense. But today it has become your least reliable one. Your other senses that tell you about up and down — your balance, touch, proprioception, and sound — are all telling you the world is as it has always been — upright. But your vision disagrees.

Q2: So what happens when you wear these inverting goggles for days?

At first, as you might guess, it was chaos for these volunteers.5

They stumbled around, missed door handles, tripped over steps, and poured coffee onto the floor. Activities like riding a bike or climbing stairs were absurdly difficult.

But you will probably not be surprised to learn that this chaos did not last. The brain, as we know, is very good at adapting.

In those early 1950s studies, it took just five to six days of wearing the goggles before volunteers said the world looked normal again — it no longer seemed upside down.6

That alone is remarkable: the brain seemed to completely adapt to input that had been flipped on its head in just under a week!

But what’s more interesting is how the volunteers said this change happened.

They didn’t just wake up one day and suddenly see the world the right way up. The shift, they reported, was gradual.

And not only gradual — but strangely selective. Certain objects seemed to “flip” before others, as if the brain were recalibrating its sense of uprightness piece by piece.7

With vision no longer lining up with the other senses, the brain had to relearn the link between movement and sight. It could have relied on head movement, eye movement, sound, or gravity to relearn what up meant. For example, moving the head down used to reveal more of the floor. But with the goggles on, moving the head down suddenly showed more of the ceiling.

Out of all the strategies the brain might have tried, one stood out: touch.

By around day three, something strange began to happen. When the volunteers reached out and touched an object — traced its edges, felt its contours — the volunteers reported that the object they were touching, and only that object, suddenly seemed right-side up.8

The volunteers claimed that objects they were interacting with looked upright, while the rest of the world remained upside down.

Imagine how strange that must have been.

Using their hands seemed to make all the difference, too, because volunteers, who were instructed not to move around or interact with their environment, adapted much slower, and some never adapted at all.9

So what happened when these volunteers finally took the goggles off?

Well… they saw the world as upside down.

This aftereffect didn’t last long though — in some cases it was just a few minutes — and the world seemed upright again.

Subscribe now

Q3: What can we learn from these strange old experiments?

At first glance, the inverted goggle studies might seem like little more than a quirky story — some odd thing a scientist tried a long time ago. Interesting, sure, but not much more than that.

But how we interpret the results from these experiments raise deeper philosophical questions.

Some take these findings as evidence that perception requires action — that we learn to see the world correctly not just by looking, but by doing.

Others interpret the data quite differently. To them, perception is a representation of the outside world, and if you give the brain enough time, it will recalibrate and flip that internal representation back to how it’s supposed to be.

As odd as these experiments are, how to interpret their findings is still up for debate.

Take a more recent study, where volunteers wore inverting goggles for ten days.
Just like in the early research, they adapted surprisingly well. Within days, they were walking through town, shopping for groceries — even skiing down a slope — as if nothing were out of the ordinary.10

Watching them, you’d assume their brains had quickly adjusted. And if adaptation were simply a matter of flipping an internal representation, you might expect to see that flip reflected somewhere in the brain’s visual maps.

When we talk about visual maps, we’re usually referring to early visual areas like V1. These regions are retinotopically organised — meaning every point of light that hits the retina has a matching point in V1. In other words, the brain preserves the layout of the retina.

So, if the brain truly flipped the representation from upside down to right-side up, we might expect to see that flip reflected in this map. A reversal in perception should show up as a reversal in the brain’s representation.

But that’s not what the brain scans showed.

Activity in primary visual cortex remained unchanged. There was no flip in the map. No rewiring of the visual coordinates.

Some have taken this as evidence that the brain doesn’t flip a representation. Others says it provides evidence that there is no representation to flip.

But the story is more complicated than that.

Because while volunteers in the original studies claimed the world eventually looked upright again, participants in more recent experiments reported that this wasn’t their experience.

Even though they were able to function normally — walk through town, ride bikes, even ski down slopes — they didn’t say the world had flipped upright. They said it remained upside down.

So what’s going on?

To start unpacking that, let’s imagine vision were your only sense. You had no hearing, no balance, no sense of touch. No proprioception. No sense of gravity. You just had visual input.

Now suppose you put on goggles that flipped your visual field upside down. Could you say your vision had been inverted?

Probably not.

Without other senses to compare against, there’d be no baseline — no up or down to refer to. Inversion only makes sense if there’s something else to anchor it. The same goes if all your senses were inverted at once. If everything flipped — sight, balance, proprioception — how would you even notice? What would you compare your experience to?

The volunteers in the prism goggle studies weren’t in that situation. They had other senses to rely on — balance, movement, and the sense of gravity.

And they had something else too — memory. They remembered what the world looked like before the flip. Those extra anchors gave them something to recalibrate against.

Neuroscientists call this multisensory calibration. And we do it constantly.

Watch a basketball game, and your brain is syncing sight and sound to follow the ball. But light and sound travel at different speeds. They hit your brain on different timelines. But we don’t experience the sound after we see the ball hit the ground. The brain calibrates.

This is why when sight and sound are slightly out of sync (like in dubbed movies), the brain can recalibrate to reduce the mismatch.11

But this can sound like there is a little theatre in the head. And we know there is no such theatre.

Plus scans show that even after days of wearing inversion goggles, the visual cortex doesn’t flip its map.

So what, exactly, is being recalibrated when the brain recalibrates?

One idea is that the brain is recalibrating the relationship between signals.12

Your brain is constantly predicting inputs and comparing those predictions to what it actually gets from its multiple senses — vision, balance, touch, and proprioception.13

When volunteers first put on the inverting goggles the brain’s predictions are very wrong. The relationship between input from the eyes and input from other senses like the inner ear and the hands, are entirely opposite to what the brain predicts.14 The brain’s expectations about the relationship needs updating. So it gets to work — not flipping a representation, but rebuilding the relationship between inputs.

Touch becomes a powerful ally.15 When you reach out and feel the shape of an object, your brain uses that input as solid, dependable feedback. It grounds reality in what you feel and learns the new relationship.

But of course, prediction doesn’t answer all the questions we might want to ask about perception.

We might still wonder: what allows the volunteers to say the world seems upright — or inverted?

Many think we’re still left with that tricky question:

What exactly is perception?

So, how do you think we answer that question?

Leave a comment

A Huge Thank You and a Change in Gears

When I started this little newsletter about 18 months ago, I did so mostly because I love writing and thinking about the intersection of neuroscience, philosophy of mind, and artificial intelligence.

I saw it as a way to explore that intersection more deeply. But I also quietly hoped that by writing, I might find others who are just as curious about these questions as I am.

And I have. We found each other.

This community has become one of the most rewarding parts of my life. Your comments, questions, and perspectives challenge me, teach me, and often take the ideas in directions I hadn’t considered. That conversation — this conversation — is the reason I keep going.

So, thank you. Truly.

The one thing I didn’t seem to anticipate was just how limited time can be. I don’t write this newsletter full-time. I also run a company. So, the writing and the replying happen in the margins: early mornings, late evenings, or in little gaps in the day. And as many of you have noticed, replies sometimes come late — or, despite my best efforts, slip through the cracks entirely.

Until now, I’ve been prioritising the writing and squeezing in the responses when I can. But I want to try flipping that.

So, moving forward, the essays might come a little less predictably.

It’s a tradeoff, I know. But I think it’s the better one. The comments are where so many of the best ideas are discussed.

So, I’m going to try this. If it turns out terribly, I’ll just try something else.

And thanks again for being here. I mean that more than you know.

Once upon a time, there was no meaning.

And now, it seems, meaning saturates our lives.

You get a text that says, “On my way!”

That string of letters isn’t just a string of letters. It means something to you, and to the person who sent it. There’s a kind of this-stands-for-that quality to it.

But how did that happen? How did letters become not just letters? How did symbols become meaningful?1

Over the past six essays, we’ve been circling those sorts of questions. Can symbols alone be enough for meaning? Do symbols need to be grounded in bodies, our history, or the world? And what does it mean to say that brain activity has meaning?

And yet, every time we tried to pin meaning down, it seemed to slip through our fingers.

We ran into strange puzzles — circular explanations, metaphorical shortcuts, and concepts that turned out to be more confused than we first thought.

The closer we looked, the blurrier it seemed to get.

Meaning is troublesome. It seems.

So, what is the trouble with meaning?

The trouble has a lot to do with representations. So, to wrap up this series, let’s focus on representations.

We’ll ask three questions:

1. What is a representation?

2. What do representations mean in neuroscience?

3. What is the trouble with representations in the brain?

Subscribe now

The Trouble with Meaning Series

This is the essay 7, and the last essay, in the series on the trouble with meaning.

Here are the other essays in the series:

Q1: What is a representation?

I’ve mentioned a few times in this series that a representation typically involves three parts: a vehicle, a target, and a consumer.2

Let’s take a simple everyday example.

Think about the small red battery symbol that sometimes appears at the top of your phone.

• The vehicle is the red symbol itself — the battery icon.

• The target is what that symbol points to — something like: “your battery charge is low.”

• The consumer is you — the person using the phone, who sees the icon and goes to find a charger.

So we could say, a symbol (the vehicle) stands in for something (the target) for someone or something that uses it (the consumer).

Once you’ve got those three pieces — a vehicle, a target, and a consumer — and a mapping between them, you’ve got a representation.

And we can say that the red symbol at the top of your phone is about your phone’s battery level, because it means something to you.

This is how aboutness and meaning come along with representations.

A vehicle is about a target because it means something to the consumer.

That’s the everyday idea of a representation.

But what about brain representations?

Let’s see how the vehicle, target, consumer fits with neuroscience talk. Let’s say you see a dog.

• The vehicle would be some kind of brain activity — a pattern of neural firing, that happens when you see a dog.

• The target is the thing that the activity points to — like something in the world, or an idea, a thought, a belief, or a decision. In this case, it’s the dog itself.

• And the consumer is… who, exactly? You? The person with the brain? The brain itself?

There’s something off here.

A vehicle (the brain activity) is about a target (a real dog in the world) because it means something to the consumer. But who is the consumer in this situation?3

Calling the entire brain the consumer feels odd. Most neuroscientists would point to other parts or systems of the brain, maybe downstream circuits — motor, memory, or decision systems. But, as we’ll see, that move can land us in trouble too.

Q2: What do representations mean in neuroscience?

A few years ago, two philosophers set out to find out what neuroscientists actually mean when they use that word representation.4 They surveyed researchers across the cognitive sciences and concluded that…

well…

That neuroscientists use the word representation in a “confused and unclear” way.

Apparently, we’ve all been using the word representation to mean different things.

How surprising!?

The most common usage is to mean a link between brain activity and sensory inputs — like the inputs that come from our eyes, ears, and body.5

In other words, the neuroscientist looks for a causal or statistical connection between the brain activity and the sensory input. And we expect those brain responses will systematically vary with varying sensory inputs in a way that’s reliable, repeatable, and ideally causal.

So if you’re looking at a dog, the brain responds one way. If you’re looking at a person, it responds another. And the activity pattern is consistent enough that, at least in theory, you could look at the brain activity alone and make a pretty good guess about what the person was seeing.

So, we read sentences like, activity in part of the inferior temporal cortex represents face perception. Or activity in the lateral occipital complex represents object shape.

These kinds of statements are common in the neuroscience literature.

But is this really a representation?

There seems to be a few things not quite right with this picture. Some of the features of representations seem to be missing.

First

It is not clear who or what the consumer is in this situation. The neuroscientist is comfortable talking about the vehicle (the neural activity) and its link to a target (some input). But who or what is making the neural activity about the input? I’ll come back to this point in a bit.

Second

Something else seems to be missing, too. You may recall that one key feature of a representation is that it can be wrong. A symbol on a map can misrepresent the real world. A battery icon can light up when the battery’s actually fine.

But in neuroscience, there’s a noticeable reluctance to say that brain activity misrepresents anything.6

If some sensory input doesn’t line up with the expected brain activity we don’t say the brain activity misrepresented. We look for an explanation. We say there is noise in the signal, or some compensatory mechanism was involved. But we don’t usually say the brain activity was wrong.

Third

Some philosophers suggest, these so-called representations that neuroscientists talk about are not representations at all. They are natural signs.7

So what is a natural sign?

Good question, I’m glad you asked!

A natural sign is a physical or observable feature that correlates with or is caused by some state of affairs — without needing an interpreter, or a convention, or a user to give it meaning.

These philosophers want to point out a difference between something that standing in for something else — like a battery icon standing in for the actual battery level — and a link between two things.

A natural sign is not a symbol that stands in for something else, it’s a link.

The classic example is smoke. Smoke is a natural sign of fire. We can say that the presence of smoke correlates with fire or even that smoke is caused by fire, but we wouldn’t say that smoke stands in for fire.

Smoke is just a natural sign of fire.

We can distinguish natural signs from representations in three ways:

1. Natural Signs don’t require a consumer.
   Smoke still exists even if no one is around to see it.

2. Natural signs can’t misrepresent.

   • Natural signs depend on causal or statistical correlations between two things. So when the link misfires — say, there is smoke but no fire — it’s treated as an exception, not a mistake.

   • Representations work differently. Because a representation involves a symbol (the vehicle), that stands in for something else. This symbol can be wrong. A symbol on a map can mislead you. A red battery icon can flash even when the battery’s full.

3. Natural signs don’t rely on conventions.
   Language, traffic signs, battery icons, and emojis only work as representations because we agree on what those symbols mean. This is not true for natural signs.

So what does this mean for representations in the brain?

When we see a dog and there is some corresponding activity in our brain, we have found a link between the sensory input and brain activity. Does that mean we have a natural sign and not a representation?

If this true, and brain activity is simply natural signs that are caused by or correlated with inputs, then how does meaning ever enter the picture? Do we need an interpreter?

Subscribe now

Q3: What is the Trouble with Representations in the Brain?

You may have seen or heard some version of the standard textbook explanation of how the brain works: Input comes in through the senses → the brain processes that information internally → then sends an output.

This same process of input → internal process → output is used to describe lots of things.

Your desktop computer takes input from a keyboard, processes it, and shows you a result on screen.

Your digestive tract takes in food, breaks it down, and produces… well, results.

Even your toaster follows the same general flow.

This process is commonly used to explain many functions in the sciences.

Let’s take nutritional science as an example. To understand digestion a nutritional scientist might link what people eat (input) with activity in the gut (internal processes) and with output. And from these links draw conclusions about how digestion works.

Neuroscientists employ this same process. To understand what the brain does a neuroscientist might link what hits our retinas (input) with activity in the brain (internal processes) and output (behavioural responses). And from these links draw conclusions about how the brain works.

But something interesting happens in neuroscience that doesn’t happen in digestive science. In neuroscience we treat that correlation or causal link as if it explains what the brain activity is about. We say the activity in your brain means what hit your retina because it’s correlated or caused by what hit your retina.

We don’t make this move in nutritional science. We don’t say the activity in your digestive tract means a pink frosted donut with sprinkles because that activity is correlated or caused by eating a pink frosted donut with sprinkles.

In digestion, activity in the gut is a marker of what you just ate — not a message about donuts. But in visual neuroscience, spikes in the visual cortex are routinely called a representation of what you see.

It’s the same functional logic (input → process → output), but our interpretations have wildly different metaphysical load.

We want to say that the brain is not like digestion (or your toaster). We want to say the brain does something more. We want to say that what happens in the middle — those internal processes — are about things. We want to say the brain has content. That its internal states mean something. That they are about something. That they represent things.

When we label brain activity as a representation simply because it links (statistically or causally) with input, are we skating on thin conceptual ice? Is what we call an explanation of meaning, simply an added label of meaning? Have we explained meaning at all? Or have we simply snuck in meaning through the backdoor?

For example, Christof Koch and Francis Crick famously suggested that:

“A good way to begin to consider the overall behavior of the cerebral cortex is to imagine that the front of the brain is ‘looking at’ the sensory systems, most of which are at the back of the brain.”8

If we take statements like this as strictly neuro-anatomical shorthand, some might think it’s a harmless metaphor.

But others suggest there is sneaky trouble here. This sort of language where we suggest that one part of the cortex “looks at” another part invites an easy — almost automatic — interpretation:

• The back-of-brain activity is treated as the vehicle that stands in for the target (the world).

• And the front-of-brain becomes the consumer who finds meaning in that back-of-brain activity.

This is what philosophers call teleological thinking.9 It treats one brain region as if its activity is for the sake of something else — as if it’s “looking at” or “interpreting” another region with a goal in mind. The back of the brain becomes the map, the front becomes the map-reader.

It turns causal interactions between neurons into a kind of purpose-driven story. The front processes the back in order to make sense of the world.

That’s how meaning sneaks in. Not by explanation, but by implication. We apply the vehicle, target, consumer to the brain like we would for a person reading a map, or a user reacting to a battery icon, as if there’s someone in there to do the reading.

This kind of language risks bringing back what we’re supposed to have thrown out long ago: the homunculus.10 The brain activity is for the little inner observer who interprets, understands, and acts. We’re explaining brain activity by what it’s for rather than how it came to be. This is the hallmark of teleological thinking. We may not call it a little person anymore. We’ve swapped out the little man in the control room for one brain region reading another. But the logic is the same.

Just like the original little person homunculus, this frontal-lobe reader needs explaining. After all, if the front of the brain is interpreting the back, then who — or what — is interpreting the front?

If we are not careful, we find ourselves caught in the same old loop: an infinite regress of map-readers, interpreters, and inner consumers, each needing another to make sense of the last.

The Sum Up

We’ve covered a lot of ground in this series on the trouble with meaning. And we’ve been chasing one slippery question: where does meaning come from?

Perhaps one of the biggest troubles with explaining meaning in the brain is that we keep trying to explain it backwards.

We start with the end — what something is for — then reason backwards.11 It is natural to do this for built things. A plane wing is for flying. But we use this talk for living things too. A heart is for pumping blood. A brain is for thinking. And activity in the back of the brain is for the front of the brain. Before we know it, we’ve smuggled meaning into places it doesn’t belong.

Science has ways of dealing with that kind of problem. One of them is the theory of evolution by natural selection.12 It explains function without foresight. It explains how things come to look to have design and purpose, without needing a designer. It dissolves the idea that everything has to be for something.

Another classic move is to break a system into smaller and smaller parts. If the brain seems too mysterious, just break it down into smaller parts. Instead of looking for a little thinker inside the brain, break that thinker into simpler subroutines, and then break those down again — until you’re left with tiny, mechanical operations.13 No homunculus. No ghost in the machine. Just little mindless steps that add up to something smarter.

But some see trouble with this idea too.

Some thinkers want us to notice what’s underneath the need for this sort of strategy. It is especially required when we assume that minds work like typical machines that take inputs, process them, and spit out outputs. This standard model, they say, is the problem. It invites the homunculus in and then requires us to explain why there is actually no homunculus.14

For this reason, many of our modern theories of how the brain works blur the lines between input and output.15 There’s no clear beginning or end — just a continuous dynamical loop of sensing, predicting, acting, and updating.

So, what happens when we stop treating the brain as a message-passing machine and start seeing it as a dynamical system?

Does the regress finally break? Does it explain meaning? Or have we merely handed the homunculus a new kind of hamster wheel?

That, I’m afraid, is a puzzle for another day.16

Leave a comment

Next Week…

No essay next week — I’m heading to Sydney for the Asia Pacific Conference on Vision. I’ll be back on July 1st with some new puzzles to ponder.

Do you remember learning to ride a bike?

Chances are, no one gave you a manual. More likely, someone just plopped you on a bike and gave you a push. You wobbled. You overcorrected. And maybe you hit the front brakes a little too hard and flew over the handlebars.

But pretty soon, your body figured it out. You found your balance. You synced up with the bike. And now — if you’ve done it enough — you don’t have to think about how to ride a bike. You just ride it.

It’s tempting to think some things we do — like riding a bike — require having a body. While other things — like doing math in your head — don’t. Maybe those could happen even if you were just a brain in a vat.

But some philosophers and scientists say — hold on. Not so fast.

They argue that every mental act, from balancing on a bike to solving a math problem, happens in a loop: brain, body, and world all working together. This view is called enactivism. It doesn’t claim the mind lives entirely “out there” in the world — or entirely inside your skull. It says: the brain doesn’t do it alone.1

That might sound strange. It’s a shift from the more familiar view of the brain as a kind of information processor — like the brain-is-a-computer metaphor we are all too used to hearing.

So: is this a genuinely new way to understand the mind? Or is it just reminding us of something we already knew — that the brain’s activity is always tangled up with the body, and the world around it?

To find out, let’s ask four questions:

1. What is enactivism?

2. Why might someone agree with it?

3. What do the critics say?

4. Does artificial intelligence need a body?

Subscribe now

Where are we in the series…

This is Essay 6 in The Trouble with Meaning series. You don’t need to read the earlier essays to follow this one, but here’s what we have covered so far:

• Essay 1: Searle’s Chinese Room — can a system understand meaning just by following rules?

• Essay 2: Symbolic Communication — how do symbols work, and what would it take to share them with aliens?

• Essay 3: The Grounding Problem — how do words (or thoughts) get their meaning in the first place?

• Essay 4: The Rise (and Trouble) of Representationalism — why many scientists think the brain represents the world, and why that view runs into trouble.

• Essay 5: Teleosemantics — one major theories that tries to explain where meaning comes from.

This week, it’s enactivism — one of the other major theories that tries to explain where meaning comes from.

A quick note on words

Talk long enough about brains and sooner or later someone will drop mind, cognition, or consciousness — often as if they were perfect synonyms.

Philosophers, meanwhile, build entire theories on the hair-splitting differences between these words.

So, these words come with philosophical baggage.

In this essay, in an attempt to try to avoid some of this baggage, I’ll use mental activity as a broad label for things like perceiving, remembering, imagining, or planning.

No term is baggage-free, of course. But by flagging this early I hope it might keep us from unpacking dictionaries in every paragraph.

I’ll revisit whether this label really earns its keep at the end of the essay.

Q1. What is Enactivism?

Enactivism is a relatively new word in the philosophy of mind. It started to gain traction in the early 1990s, with the publication of a book called The Embodied Mind.2

In that book, the authors made a bold claim: minds aren’t things we have — they’re things we do. To have a mind, they said, is to engage in sense-making. Not by sitting back and receiving information, but by moving through the world — by acting, feeling, and responding.

Since then, the idea has caught on. And while different versions of enactivism have taken shape over the years, most of them come back to four central claims.3

1. The Mind is Embodied

For enactivism, the body isn’t just along for the ride. It matters — deeply.

Mental activity — whether we’re remembering something, imagining a scene, or thinking through a problem — isn’t just happening in the brain. It’s shaped by our body’s movements, its sensations, and its rhythms.

Our posture, our heart rate, the tension in our muscles, even signals from our gut — all of it can influence how we think.4

The body, in other words, isn’t just a delivery system for the brain. It’s part of the process.

2. The Mind is Embedded

We don’t think in a vacuum. We think in kitchens and parks and waiting rooms.

Our thoughts are shaped by what’s around us — the objects we use, the routines we follow, the people we talk to. Even the feel of a place can change the way we think.5

You’ve probably noticed this. Some tasks just feel easier in certain settings. If you’re trying to solve a problem, you might go for a walk, take a drive, or hop in the shower — because you know that’s where your best ideas tend to show up.

3. Perception is Active

If you’ve ever seen a textbook diagram of how the mind works, it probably looked like this: input goes in through the senses, gets processed in the brain, and then an output comes out — like pressing buttons on a machine.

But enactivism rejects that picture.

For them, there is no middle stage. Perception isn’t something that happens after input and before action. It is action. It’s skilled, moment-to-moment engagement with the world.6

4. Enactivism is Skeptical of Mental Representations

Enactivism doesn’t deny that the brain does complex things. But they do question whether those things are best described as internal representations — the kind you find in the brain-as-computer versions of cognitive science.

Radical enactivism argue that we don’t need representations at all.

Others take a softer view. They’re open to the idea that some representations might exist. Perhaps lean, action-oriented ones.7

Most camps agree on this point: if representations are part of the story, they show up later in both evolution and development. They emerge through culture, through language, and through shared practices. And they’re built on top of something more fundamental — our embodied, active engagement with the world.8

Q2. Why Might Someone Agree with Enactivism?

Let’s look at four reasons why someone might find enactivism appealing.

1. The Grounding Problem

We talked about this back in Essay 3. The grounding problem asks the question: how does mental activity get its meaning?

Radical enactivism responds by reject the idea that meaning has to be explained in terms of internal representations in the first place. For them, meaning isn’t something added on top of experience — it’s already there, in our ongoing engagement with the world.

Moderate enactivism take a similar line, but leave some room for representations. They argue that if representations show up, they only make sense because they’re grounded in embodied action. Our actions are already meaningfully tied to the world, they say — so any representation that develops out of those actions inherits that connection.9

2. The Behavioural Evidence

There’s a well-known experiment you may have seen. A group of people are passing a basketball. You’re asked to count the number of passes between players wearing white t-shirts.

Halfway through the clip, a person in a gorilla suit walks right into the scene — beats their chest — and strolls off.

And yet… most people don’t notice the gorilla at all.

It’s one of the most famous examples of inattentional blindness — the idea that we can miss something obvious, simply because we weren’t looking for it.

Change blindness tells a similar story. In one ad by Skoda, tiny details in a scene are constantly changing. A wall swaps colour. A teapot appears. And most viewers don’t catch a thing.

And this one is a good example, too:

Experiments like these suggest that perception isn’t a passive recording of the world. You don’t take in everything and then process it. You won’t notice the gorilla unless you’re actively looking for it. And in change blindness you won’t see a change unless you are actively looking at the right spot.

One of the most striking demonstrations of actions role in perception comes from a now-famous experiment with kittens.10

It was the 1960s. Psychologists set up a kind of carousel for cats. Each pair of kittens was harnessed into the same rotating rig — one kitten was a walker, the other as a passenger.

The walker kitten could walk with the carousel. The passenger kitten, though, just went along for the ride — carried in a little basket, seeing everything but not moving its legs.

Importantly, both kittens received identical visual input. But only one could move.

So what happened to the kittens?

The walker kittens developed normal vision. But the passenger kittens — the ones who only watched — didn’t. They walked straight toward what looked like a cliff, showing no hesitation. They had no depth cues.

Enactivism takes this as powerful evidence. Perception requires active interaction with the world.11

3. The Neuroscience Evidence

The brain doesn’t come with fully wired circuits. As babies, our circuits have to be built. And, it turns out, movement plays a key role in how that happens.

In newborn rats — and in premature human babies — early brain development involves something called spindle bursts. These are brief, bursts of brain activity that show up right around the time of spontaneous body twitches.12

If you’ve ever felt your body twitch just as you were falling asleep, you’ve felt a version of this. In adults, spindle bursts mostly show up during non-REM sleep. But in babies, they happen all the time.

The baby kicks. Or flinches. And at the same time, or just before, a spindle burst is seen in the brain. These little jolts help wire up circuits, linking neurons that fire together.

Some parts of the body work in opposition — like biceps and triceps — when one muscle contracts, the other relaxes. So, the brain learns not just what fires together, but what doesn’t. So, inhibitory patterns form too — which helps shape coordination.

We see spindle bursts in the visual system, too. Babies’ eyes twitch, even before they can focus. And those twitches help organise the visual cortex.

In other words, the brain doesn’t just wait for the world to feed it information. From the beginning, it’s building itself — through action.

4. Evidence From Robotics

Back in the 1990s, roboticist Rodney Brooks had an idea that went against the mainstream view at the time.13

Instead of programming his robots with detailed maps of its world — carefully constructed layouts and pre-defined plans — he gave them simple rules. Just enough to sense the world and react to it.

These robots didn’t have anything you’d normally call a representation. No internal model of their environment.

But they could navigate. They could avoid obstacles. They could complete tasks. And — surprisingly — they often outperformed the robots with maps.

Brooks called them robots “without representation.” He claimed intelligent behaviour doesn’t require internal representations at all. It can emerge from tight, real-time loops between sensing and acting.

Which, is exactly what enactivists were saying about living systems.

Subscribe now

Q3. What do the critics say?

Of course, not everyone is on board with enactivism. Let’s review three of the most common criticisms.

1. Enactivism is Vague About Representations

One of the most common critiques of enactivism is that it is fuzzy on where it stands with mental representations.

Even the more radical versions are hard to pin down.

Take the book Radicalizing Enactivism.14 Early on, the authors come out swinging — they argue that mental activity doesn’t involve internal contentful representations at all. No representations, full stop. But just a few pages later, they write that even radical enactivists “need not, and should not, deny the existence and importance of contentful and representationally based modes of thinking.”

Wait — what?

To many readers, that sounds contradictory.

The authors try to clear it up by drawing a distinction. They say there are two kinds of mental activity. Basic cognition — like perception, attention, and motor control — which doesn’t require representations. It’s grounded in our ongoing bodily engagement with the world. But content-involving cognition — like imagining a future vacation or solving a math problem — does involve representations. The important point, they say, is that those content-involving representations are only possible because they’re built on a foundation of non-representational activity — things built through brain-body-world interaction.

It might sound like a clever move. But not everyone’s convinced.15

Critics point out that this hybrid model may inherit the worst of both worlds. It doesn’t fully reject representations, so it doesn’t escape the problems enactivism set out to solve. But it also adds a new problem: how does the system bootstrap itself from non-representational activity to content-involving representations. And right now, the explanations that are given — well — they sound a little hand-wavy. Like saying “...and then a miracle occurs.”

In trying to keep a foot in both camps, the hybrid view risks pleasing no one: too representational for radical enactivists, and too vague for the representationalist.

2. Representationism Works in Neuroscience

Some argue — why do we need a different theory? Representation works great in neuroscience. It explains the data, makes predictions, and helps us build models of how the brain works. Isn’t that what a good scientific theory is supposed to do?

Modern neuroscience often describes the brain as an information processor. Neurons are said to encode things — like edges, faces, or places. Brain regions get mapped according to what they represent. These ideas are grounded in decades of experimental work.

So when radical enactivists say we should throw out the idea of mental representations, many neuroscientists raise an eyebrow. Are we really supposed to ignore all that data?

Enactivists say: not exactly. They argue they’re not denying that brain activity correlates with experience. What they’re questioning is how we interpret those correlations. Instead of thinking of brain activity as a kind of internal map — like a stored map in the head — they see it as part of a living, dynamic system. A system that only makes sense when you look at the whole loop: brain, body, and world, all acting together.

To some, that can sound a little vague — like replacing one metaphor with another.

But there is a growing pile of papers trying to make enactivism mathematically precise.16 The worry is: the more detailed these models become, the more they start to sound like the very representational theories they claim to oppose — just with new jargon.17

3. Enactivism is Just Behaviourism

Enactivism wants to downplay internal representations. It says mental activity is rooted in action — in how organisms engage with the world.

That focus on action sounds familiar.

In fact, it sounds a lot like behaviourism, the dominant view of the mid-20th century. And that raises a question: if we strip mental activity down to nothing but action, aren’t we just back where we started?

Enactivists say no. They’ve tried hard to draw a clear line between their view and behaviourism. Behaviourism was about inputs and outputs — what goes in and what comes out. It didn’t care what was happening in between. Enactivism, by contrast, is all about the ongoing relationship — the loop — between an organism and its environment.

For sense-making, they claim, a system has to be autonomous, it has to regulate and maintain itself, and it has to be adaptive — not just reacting, but adjusting its responses in ways that serve its continued existence. These features, they argue, are absent in purely behavioural things like thermostats or bacteria that drift toward sugar.

Q4. What About Artificial Intelligence?

If a large language model can answer your questions, write a poem, and debug your code — isn’t that mental activity?

And yet, LLMs don’t move. They don’t have bodies.

Doesn’t that suggest it’s internal representations — not embodiment — doing the real work?

Some, like neuroscientist Antonio Damasio, argue there’s a crucial difference we shouldn’t overlook.18

Biological creatures have to meet their own thermodynamic requirements. To survive, biological creatures don’t just have to perform useful behaviours — they also have to keep themselves alive. That means finding energy, repairing damage, maintaining internal order. If they fail to do either — act adaptively or maintain themselves — they cease to exist.

But LLMs are different. An LLM does not have to secure its own energy, repair itself, or maintain its own internal order.19 We do those things for it. The two roles that are together in a biological creature — performing useful behaviours and keeping itself alive — are separated in an LLM. And, some think, that separation might not be a trivial detail. It could turn out to be a very important difference.

So, what if we gave an LLM a body?20 One that could interact with the world. One that could monitor itself, regulate its internal state, manage its own energy use, and adjust its behaviour accordingly? Would that kind of interacting, self-maintaining, body-in-the-world system make a difference to the LLM? Would we expect to see something very different than we typically see from an LLM?21

The trouble with questions like these — about whether AI needs a body to have mental activity — is that the answer might depend on what we mean by mental activity. Embodiment might really matter for our kind of mental activity. But is it a universal ingredient required for any system to have mental activity?

LLMs show that representation-heavy algorithms can, at least, mimic some high-level behaviours — but enactivists reply that until a system regulates its own survival in a messy world, calling it a full-fledged mind might be premature.

Leave a comment

The Sum Up & Next Week…

Last essay we covered teleosemantics. Both teleosemantics and enactivism want to avoid magical or mysterious accounts of meaning. But they do it differently.

• Enactivism says: Meaning doesn’t come from internal representations in the brain — it comes from active, embodied engagement with the world.

• Teleosemantics says: Mental representations do exist — and they get their meaning from evolutionary history.

In some corners of philosophy and cognitive science, this difference gets framed as a showdown: representations versus enactivism. Pick a side.

But many researchers think that’s the wrong way to see it. The split isn’t as clean — or as necessary — as it’s sometimes made out to be.

Next week, I’ll wrap up this series on The Trouble with Meaning by tying up a few loose ends — including a look at where some thinkers believe the real trouble with meaning might lie.

We’re heading into the colder months here in Australia, which means the snakes are getting busy — feeding up before brumation, the reptile version of hibernation.

Just recently, not far from where I live on the east coast, a carpet python was found with an unusual meal in its belly: two golf balls.

Thanks to some quick-thinking wildlife volunteers, the snake survived. But the story made the rounds in local news outlets, mostly because of the obvious question: did the snake mistake golf balls for eggs?1

You might think there’s not much going on in a snake’s head. And fair enough — snakes aren’t known for their philosophical depth. But still — we might think that even a snake has to tell the difference between food and not-food. Presumably, something in its nervous system helps it do that.

This is what some call a representation. And, some might think this representation in the snake’s brain is about eggs.

At least, that’s how we often talk about what happens in our own brains — when we think about eggs, or dogs, or why the sky is blue.

But this raises two puzzles. First, how can any brain representation be about something? And second, if it is about something — how can it misrepresent something?

If the snake’s brain is representing egg, then what’s it doing being active for golf balls?

How do we explain that without evoking mystery or magic?

There are two approaches. One says: let’s keep the idea of representation, but try to naturalise it — explain it using nothing more than physical stuff and its relations.2 The other is more radical. It says: we should ditch the idea entirely. Because the whole concept of representation is the problem.

This week, we’ll explore the first path — keeping representations, but naturalising them. Next week, we’ll turn to the alternative view.

So here’s our question for this week: Can we find a naturalistic theory of representation that explains how representations can be about things?

To find out, we’ll ask three questions:

1. Can we naturalise representations?

2. What’s the leading proposal for how to do it?

3. And what do the critics say?

Subscribe now

First a quick note…

This is Essay 5 in The Trouble with Meaning series. You don’t need to read the earlier essays to follow this one, but if anything here feels a bit murky, Essay 4 will probably be the most helpful.

Here’s what we’ve covered so far:

• Essay 1: Searle’s Chinese Room — can a system understand meaning just by following rules?

• Essay 2: Symbolic Communication — how do symbols work, and what would it take to share them with aliens?

• Essay 3: The Grounding Problem — how do words (or thoughts) get their meaning in the first place?

• Essay 4: The Rise (and Trouble) of Representationalism — why many scientists think the brain represents the world, and why that view runs into trouble.

If we want a naturalistic theory of the mind, there are a few things it needs to do. Back in Essays 3 and Essay 4, I laid out four key criteria.

We want a theory that:

• Is naturalistic — no magic, no mysteries.

• Fits with an objective, mind-independent reality.

• Accounts for mistakes — not just when things go right.

• Avoids circularity — no infinite regresses, no hidden homunculi doing the real work.

These four won’t satisfy every philosopher’s wish-list, but it’s the bar we’ll use for this week’s essay. So, let’s see if we can find a theory that clears it?

Q1: Can We Naturalise Representations?

Let’s start with the basics. A representation is commonly thought of as something that stands in for something else. So, some might say that certain brain activity represents things — dogs, eggs, the blue sky.

The brain activity is the vehicle.3 The thing the brain activity stands in for is the target. Targets might be physical objects, like a snake, or abstract ideas, like justice.

So how do vehicles come to be about their targets?

The classic views proposed three main answers:

• Co-location: the vehicle and target are in the same place at the same time.

• Resemblance: the vehicle looks or sounds like the target.

• Simple causation: the target causes the vehicle.

Let’s take them one at a time.

Co-location

This one fails quickly. We can think about (or represent) the Eiffel Tower while we’re in Sydney. We can remember someone who’s no longer with us. And we can imagine things that don’t exist, like unicorns. Co-location can’t explain those cases — and it can’t explain misrepresentation either. So we can set it aside.

Resemblance

Back in 2015, someone opened a tub of butter and saw the face of Donald Trump.4 The shape resembled him — but does that mean the butter was about Trump? Probably not.

We might say the shape of the butter is well suited to represent Trump — but only if someone decides to treat it that way. The resemblance alone can’t be the whole story. The butter can only be about Trump because someone already has Trump in mind.

And that puts us in a loop. We’re using a representation to explain a representation. It fails our circularity criteria.

Simple Causation

Maybe a representation is about what caused it.

Take a photograph of Trump. It’s about Trump because Trump caused it — through light bouncing off his face onto the camera sensor. The photo has a causal link to its target.

But that’s not always enough. Wet grass is caused by rain, but we wouldn’t say the wet grass is about rain.

And what about misrepresentation?

Say we see a coiled garden hose and mistake it for a snake. We might want to say that your brain representation was caused by the hose.5 So should we say whatever is going on in the brain is about the hose? That feels wrong. If it is about anything, isn’t it about a snake, even though the snake isn’t there?6

Causal theories need something more.

One attempt comes from philosopher Jerry Fodor. His idea is called asymmetric dependency.7 Here’s the basic version:

Your brain is active in some way for snakes. But sometimes your brain is active in a similar way for garden hoses that look like snakes. The activity is about snakes, Fodor says, because the garden hose only causes it because snakes do. If snakes didn’t cause that brain activity, hoses wouldn’t either. But snakes would still cause it, even if hoses never did. That’s the asymmetry.

This lets us keep a causal theory and explain misrepresentation. So far, so good.

But there’s a catch.

First, this asymmetry sneaks in an assumption that the brain activity is supposed to be active for snakes, not hoses. We’ve quietly assumed what is considered successful brain activity. This is called the normative-assumption problem.

Second, it assumes we already know what the brain state is about. It fails our circularity test.8

So our search continues.

Q2: What is the Leading Proposal?

Nowadays, one of the most influential proposals for naturalising the aboutness of representations is called teleosemantics. It’s not a single theory, but a family of closely related views built around one core idea:9

The idea is that our brains represent things not because it resembles the thing it represents, or is directly caused by it, but because that is the function that was passed on.10

Take that garden hose in your backyard. You see it, and for a split second, you think snake. Something goes on in your brain. So we might say this brain’s activity functions as a detector for snakes — it’s an evolved response that reliably picks out snake-like features.11

But why does that activity go on in your brain?

There are two ways to interpret that why question.12

• One says when we ask why does that activity go on in your brain? what we are really asking is, what is the purpose or goal of this activity? This interprets the why question as a what for question. It’s a question about ends.

• The other interpretation is that the question why does that activity go on in your brain? is asking, which historical process shaped this activity? This interprets the why question as a how come question. It’s a question about means, not ends.

The first — the what for question — risks circularity. If we say, the brain activity is for representing snakes, we’ve already assumed what we were trying to explain: that the brain activity functions as a snake-detector. It end-point thinking — explaining the activity by its end goal or purpose, rather than by the process that produced it.

That’s the circularity problem again.

The second — the how come question — might give us a way out.

The philosopher Fred Dretske came to think that a physical theory of meaning needs two ingredients: information and function.

Information helps explain how the brain’s activity might correlate with something in the world. But correlation alone doesn’t explain why that activity is about the thing it correlates with — or why we can say it was wrong when the correlation fails.

We need function — and not just any function, but one grounded in history.

This is where the how come question becomes crucial.

According to teleosemantics, a representation’s function (what for) is fixed by its selected effect (how come). That is:

Its function is grounded in its evolutionary history.

So why does a certain brain activity represent snakes?

Because, over evolutionary time, animals with brains that responded in certain ways in the presence of snakes were more likely to survive. Their brains developed a detection system that was active when snakes were around — and that detection system got passed on.

It seems like a clever move. It replaces purpose-based explanations (what for) — which often sneak in circularity — with a naturalistic account of how representational functions could arise (how come). And, some argue, it does this without smuggling in the very meaning it’s supposed to explain.

Let’s bring back our golf-ball-eating snake.

Why say the snake’s brain represented the golf balls as eggs?

Because, according to teleosemantics, snakes with brains that responded to pale, egg-shaped objects were more likely to find food. Over time, because having a system like that had advantages, it was passed on, and we can now say its function is to detect eggs.

So when the brain responds in a similar way for a golf ball, it’s performing the function that was passed on — just in the wrong context. The system misrepresents, just like when we mistake a hose for a snake.

That’s the appeal of teleosemantics. It seems to explain misrepresentations.

Even when the brain misrepresents, the function stays the same. The brain response was passed on because it played an evolutionary function. And that evolutionary function determines what it’s about.

So how does teleosemantics stack up against our criteria?

At first glance, it looks promising:

• It’s naturalistic — meaning comes from evolutionary history.

• It preserves a mind-independent reality — our representations succeed because they track real features of the world.

• It accounts for misrepresentation — mistakes are part of how biological systems work.

• And it avoids circularity — functions are defined by history, not assumed meaning.

At least, that’s how it seems. But not everyone is convinced.

Subscribe now

Q3: What Do the Critics Say?

Now that we’ve seen the appeal of teleosemantics, let’s walk through the criteria again — but this time, with a skeptical lens.

One thing to keep in mind: teleosemantics isn’t a single unified theory. It’s a family of views. So not every criticism applies equally to all versions.

Does teleosemantics fit with a mind-independent world?

At first glance, yes. Teleosemantics grounds meaning in real-world history — in the evolutionary success of tracking features out there in the environment.

But some critics argue it leans too heavily on the world. If meaning depends on past success tracking real things, how do we explain thoughts about things that don’t exist?

Take novel concepts.13 Imagine you’re Freeman Dyson, the first to conceive of a shell around a star to harness its energy — a Dyson Sphere. That idea had no precedent. So how could your brain activity for Dyson Sphere have been selected for in evolutionary history?

Some versions of teleosemantics expand the idea of selection to include learning through operant conditioning, neural plasticity, and culture. Others suggest novel thoughts piggyback on older systems — for instance, the Dyson Sphere might reuse neural activity that was selected for spatial reasoning or tool use.14 Plus, Dyson-Sphere thoughts will most likely be a recombination across many circuits, not a single new state.15

But what about purely imaginary things?

You can picture unicorns, Atlantis, or an “octophant” — half octopus, half elephant. If they don’t exist, then how could any system have evolved to track them?16

Some defenders say these thoughts are compositional — they borrow from real-world concepts (like “octopus” and “elephant”) and combine them. But that raises a question: how far can this piggybacking go? At what point does recombination lose contact with biological function?

Does teleosemantics explain misrepresentations?

This is one of teleosemantics’ big selling points. If a system evolved to detect eggs, and it fires for a golf ball, that’s a misrepresentation — not because of a broken rule, but because it failed at its selected function.

But critics aren’t convinced it’s that simple.

This is the normativity or functional-indeterminacy problem.17 Our brains don’t really have distinct patterns for specific targets. What looks like one brain response often serves many roles — what we might call attention, movement, memory, threat detection. So if meaning depends on function, and function is ambiguous, how do we know when the system misrepresented?

Take the garden hose mistaken for a snake. Was that a misrepresentation of a snake? Or was it a successful case of erring on the side of caution?

Supporters of teleosemantics have suggested solutions — like looking at statistical patterns over time or granting the function to the most evolutionarily valuable one.18 But can this give us a clear, determinate content — the kind that tells us what brain activity is about, and when it’s misrepresenting? Or is it a convenient after-the-fact story that doesn’t actually avoid circularity?

Does teleosemantics avoid circularity?

Even if we view representations through a functional lens, to say that some brain activity was selected for detecting snakes, we already have to know what counts as a snake encounter. But that assumes the very content — snake — we were trying to explain.19

Likewise, calling the garden hose mistake a misrepresentation only makes sense if we’ve already decided that the brain activity was supposed to represent a snake, not a hose. But where does that judgment come from?

We’re back at the circularity problem — again.
Arguably, this is the problem that many of the other critiques collapse into.

Some defenders respond that we’re asking the wrong kind of question. Rather than fixating on narrow content like snake or hose, they suggest we take a broader view: what matters is the role the representation plays in survival.20

From this perspective, the brain activity wasn’t selected to represent snakes per se, but to help the organism avoid danger. So when the brain responds to a hose as if it was a snake, it’s still performing its general function.

But critics push back: if we blur the line between what some brain activity is about and what it merely does, we risk losing what makes representation distinct from simple reactivity. It might explain why the activity exists — but not what it is about.21

Some philosophers go further. They argue that aboutness is an explanatory dead-weight — a relic from an outdated way of thinking.22 On this view, the question of how representations come to be about things is simply the wrong question.

Some thinkers suggest we abandon the idea of aboutness altogether — and with it, the entire teleosemantic project.

Before we wrap up, there are two more critiques that come up often enough that it would feel like a glaring omission not to mention them.

Swampman

One of the most famous challenges to teleosemantics comes from Donald Davidson’s Swampman thought experiment.23

It goes like this: a lightning bolt strikes Davidson near a swamp, vaporising him — and, by sheer coincidence, assembling an atom-for-atom duplicate from the surrounding matter. This perfect copy, Swampman, gets up and walks away, indistinguishable from the original.

Since Swampman is physically identical to Davidson, it seems reasonable to think he has the same representations — that he thinks, perceives, and remembers just like Davidson did.

But here’s the problem: Swampman has no history. He wasn’t shaped by evolution or learning. So if teleosemantics is right — and aboutness comes from history — then Swampman shouldn’t have any aboutness to his representations at all.

Some defenders bite the bullet and say, if Swampman has no history, he must have no aboutness.24

But others argue the whole thought experiment should be thrown out like so many other bad thought experiments.

Why? Well, there are a number of reasons.

But one is because this thought experiment, like many thought experiments, makes claims about our world. And if we want to make claims about our world the thought experiment needs to pass two tests.

It needs to be:

• Logically possible: the scenario must not contain outright logical contradictions.

• Physically possible in our world: it must be able to occur without violating the physical laws that govern our universe.

Swampman might pass the first test. But, the argument goes, Swampman, flunks the second.

In our world we can’t assemble anything without paying a thermodynamic price. Which means, we can’t assemble a perfect atom-by-atom copy of Davidson without an enormous amount of time, energy, and causal interaction. And that — right there — is a history.

The moment we account for how Swampman could come to exist, given our physical laws, we realise that his existence is subject to the same evolutionary or developmental processes that teleosemantics depends on.

What About Artificial Intelligence?

Modern AI systems — especially large language models — also raise questions for teleosemantics.25

These models seem to have representations that are about things. But unlike us, they weren’t shaped by biological evolution. So where does their aboutness come from?

Some theorists think the idea of teleosemantics can be extended to machines. They argue that training plays a role similar to biological evolution. A neural network trained to recognise cats might develop representations of cats because those patterns helped reduce error — a kind of computational “survival.”

The idea is that the training objective — like next-token prediction or image classification — functions like an artificial selective pressure. And if that objective was chosen by designers for a specific purpose, perhaps it can ground representational function in a way that parallels biological evolution.

But not everyone agrees.

Teleosemantic theories are based on the idea that a representation gets its aboutness from the function it was shaped to do — because they help organisms survive. So even if a representation gets reused in different ways, its core function stays the same.

Critics ask us to compare that to large language models. These systems aren’t shaped by evolution, but by training on massive amounts of (mostly) text, using a broad and open-ended goal: predict the next word. That might sound like a function, but, the argument goes, it’s not tied to any specific biological need.

For critics, that makes it difficult to say what, if anything, the model’s functions are about.

Leave a comment

Next Week…

Some philosophers respond to these problems by saying that no theory of representation will ever fully succeed. They argue we need to get rid of the whole idea of representations. This is the anti-representational view. That’s the direction we’ll head next week.

Take a look at the sentence you’re reading right now. It’s just a string of shapes on a screen.

But somehow, those shapes turn into words. And those words turn into ideas — ideas about something.

How does that happen?

Cognitive scientists have a go-to answer: your brain represents them.1 Somewhere in your head, there’s a pattern of neural activity that stands in for dog. A different one activates for tree. Another for jealousy. In other words, certain brain activity has content that is about something.2

If you read through academic papers, you’ll find sentences like, “neurons in the human medial temporal lobe represent concepts,” or “neurons in the mammalian visual cortex represent edges in the visual field.”

This way of thinking — that the brain builds internal representations — is one of the most powerful and persistent ideas in cognitive science. It’s been used to explain everything from memory to language to decision-making. It’s the core idea behind much of cognitive neuroscience.

But there’s something strange here: for a concept so widely used, it’s surprisingly hard to pin down. What exactly is a representation?

Some philosophers — and increasingly, some scientists — are starting to push back. They argue that we’ve gotten a little too comfortable with this idea.3 That, despite all the talk, representations might not actually explain how brains work at all.

So this week, let’s take a closer look. Is representational thinking a deep insight into how the brain works? Or is it a metaphor we’ve mistaken for a mechanism?

To find out, we’ll ask three questions:

1. What does it mean to say the brain represents something?

2. Why has this idea become so dominant in psychology and neuroscience?

3. What do the critics say?

Subscribe now

A quick note before we get started:

This is Essay 4 in The Trouble with Meaning series. You don’t need to read the earlier ones to follow along, but if you’re curious:

• Essay 1 looked at Searle’s Chinese Room.

• Essay 2 explored symbolic communication (with a little help from aliens).

• Essay 3 tackled the grounding problem — how words get their meaning.

Last week, I introduced a simple framework for thinking about representations: they typically have three parts — a vehicle, a target, and a consumer.

• The vehicle is what carries the message — maybe the letters d-o-g, a pattern of brain activity, or a string of numbers inside a language model.

• The target is what it’s about — a real dog, an idea of a dog, or a fictional one.

• And the consumer is whatever system uses that connection to do something useful: answer a question, guide behaviour, make a prediction.

This week, I’m focusing on one kind of vehicle in particular: brain activity.
When neuroscientists say the brain represents something — what exactly are they claiming? And how solid is that claim?

Q1. What Does it Mean to Say the Brain Represents Something?

When scientists say your brain represents something — let’s say a dog — they don’t mean your brain is literally holding a tiny photograph of a dog. There’s no little projector in your head beaming a mental slideshow.

What they mean is that some pattern in your brain — a specific configuration of neural activity — is standing in for that dog.4 It’s a signal that points to something. It’s about something. It carries content.

Those words — about and content — are doing a lot of work here. They hint at something big: that your brain activity doesn’t just happen, it means something. It refers beyond itself — to things you see, remember, or imagine.

Take vision, for example. When you look at a dog, light hits your retinas and triggers a cascade of neural signals. The information from your retina moves through the brain — from simple edge detection in the primary visual cortex to higher-level regions that process other things like its texture, shape, colour, movement, and exact place in space.

So, in this view, perception is a process of constructing an internal model of the world — a model that allows us to recognise objects, predict outcomes, and make decisions.

And representations are not limited to what our senses represent. Representations show up in theories of memory (where we store representations of the past), language (where words represent concepts), and reasoning (where we manipulate abstract representations to draw conclusions). In each case, thinking is sometimes framed as a kind of internal workspace, where symbols, patterns, or codes do the heavy lifting.5

That’s the basic picture: your brain builds internal versions of the world — and then uses those versions to perceive, remember, think, feel, and act.

But as we’ll see, this way of thinking carries some philosophical baggage. Before we unpack that, though, let’s look at why representational thinking is so popular in the cognitive sciences.

Q2. Why is Representational Thinking so Dominant in the Cognitive Sciences?

It might seem like scientists have always thought of the brain as building internal models of the world. But that hasn’t always been the case.

For much of the early 20th century, psychology took a very different approach: behaviorism.6

Behaviourists didn’t want to talk about minds at all. If you couldn’t see it or measure it, it didn’t count. Psychology, they argued, should stick to what could be observed: stimulus and response. Internal thoughts, feelings — and yes, representations — were off-limits.

But by the 1950s and ‘60s, cracks were starting to show in the behaviourist idea.7

There were just too many things it couldn’t explain. People could understand sentences they’d never heard before. They could imagine new futures, solve problems they hadn’t been trained on, and plan ahead. Clearly, something was going on inside the head — and behaviourism didn’t have the tools to deal with it.

So the field shifted.

Behaviourism was replaced by cognitivism.

Inspired by the rise of computers, psychologists began to think of the brain as an information processor — something that takes in data, stores it, transforms it, and uses it to generate behaviour. Just as a computer relies on data structures like files and variables, the brain needs its own internal information. That’s where representations come in.

Representationalism is, therefore, a central pillar of the Computational Theory of Mind.

Representational thinking offered a way to study the mind within a scientific framework. It allowed researchers to build models of memory, perception, and reasoning — using math and formal logic. Suddenly, fuzzy concepts had structure. Theories could be tested. The brain-as-representation-machine became a powerful way to make sense of behaviour.

Why? Because the idea seemed to fit the evidence.

So let’s take a quick tour of some of that evidence.

Rats in Mazes

One classic example comes from a 1996 study by Packard and McGaugh.8 The setup was a cross-shaped maze with four arms, north, south, east, and west. During training, rats were always placed in the south arm, the food was placed in the west arm, and the north arm was blocked with a clear Plexiglas barrier.

If a rat turned left at the intersection, it found the food. If it turned right, it didn’t. Over seven days, the rats learned how to find food — and they got very good at it.

Then on the eighth day, researchers flipped things: they placed the rats in the north arm and blocked the south arm.

This was the critical test: if behaviourism is correct, the rats would have learned using a stimulus-response rule. Perhaps something like “turn left to get food”. In that case, they should continue to turn left — which would now take them to the east arm.

But if the rats had formed a spatial map — an internal representation of the maze — they would turn right instead, heading for the same location in space where they previously found food.

What happened?

The rats turned right.

They weren’t just following a habit. They seemed to understand the space — to hold a mental representation of the maze.

Illusions

Representational thinking also helps explain quirks in human perception.

Take the Kanizsa triangle. Your brain sees a white triangle that isn’t really there. And the duck-rabbit illusion shows how the same image can flip between two different interpretations. In both cases, there seems to be a gap between what’s on the page and how you see it. Some take this as a sign that the brain is representing what we see.

Another striking example is called binocular rivalry. If you show one image to your left eye — say, a house — and a different one to your right — say, a face — your brain doesn’t blend them. Instead, your perception flips between seeing the face, and then seeing the house.

In a famous fMRI study, researchers found that activity in one area of the brain was stronger when people saw the face, and activity in a different area was stronger when they saw the house — even though the input to the eyes never changed.9 So if the images didn’t change, what did change? It was the internal representation, right?

The Blind Spot

Your eye has a blind spot — the place where the optic nerve exits the retina. There are no receptors there. Technically, you shouldn’t be able to see anything in that patch of space.

But you don’t see a hole in your vision.

Area V1 in your visual cortex is laid out like a map that corresponds to areas on your retina. This means that different parts of V1 correspond to specific locations on the retina.

For a long time, researchers assumed there was no region of V1 corresponding to the blind spot. Why would there be!? The brain wasn’t receiving any input from that part of the retina.

But newer studies suggest otherwise10. There is a region of V1 that responds to the blind spot. It isn’t missing. And it isn’t silent. What it seems to be doing is ‘filling in’ that area based on the surrounding visual context.11

Representational thinking is the winning framework in cognitive science for a reason.

It has allowed scientists to build models, design experiments, and explain behaviour like no other idea has before. It’s turned the mind into something we can model, measure, and manipulate.

It has won because representational thinking fits the evidence.

But hold up.

Just because something fits doesn’t mean it’s the right explanation.

Let’s turn now to the critics — and take a closer look at some of the key criticisms of representational thinking.

Subscribe now

Q3. What Do the Critics Say About Representationalism?

For such a popular idea, representational thinking has drawn a lot of criticism. One reason is that the term representation doesn’t have a single, clear meaning. It gets used in different ways by different people. So, when critics take aim, they’re often shooting at different versions of the idea.

Still, there are some common concerns. Let’s walk through four of the big ones.

1. Reverse Inference

Just because a brain pattern shows up when you see a dog doesn’t mean it represents a dog. Remember correlation does not equal causation.

When psychology shifted from behaviourism to cognitivism, it brought with it a vocabulary: attention, memory, reasoning, emotion. Neuroscientists picked up that vocabulary and went looking for where those things “live” in the brain.

Volunteers were put in fMRI machines and shown pictures. They found patches of brain that respond more strongly for some category of images — say faces — compared to another category of images — say houses. For faces, they triumphantly named the area the fusiform face area. And claimed, this is the spot where “the brain represents faces.” They did a similar thing for places.

Soon we had all sorts of areas — a word area, a fear centre, a reward pathway, a value hub, an empathy centre, a pain signature, a speech production module, and even a love centre.

But there is a problem here.

It is a classic example of reverse inference.12

Critics argue that these kind of early experiments in cognitive neuroscience — the ones that correlate brain responses to psychological terms — do not demonstrate that those brain responses are representations of those psychological terms. They don’t even demonstrate that such psychological terms exist.

One problem is that the area we’ve labeled the face area sometimes responds more strongly to things that are not faces. And, consequently, there’s currently a lively debate about what the fusiform face area actually does.13

So critics warn that when we say the brain represents something, we need to be careful we are not simply projecting our psychological categories onto patterns of activity and calling it a representational discovery.

2. Brains are Not Digital Computers

Another worry is that representational thinking borrows too heavily from computer metaphors.

Brains aren’t digital machines. They don’t run lines of code. They’re biological systems — wet, noisy, and constantly changing.

Some neuroscientists argue that the idea of neural coding is misleading. The brain doesn’t so much send messages as it participates in complex, dynamic loops.

Critics argue that we’re looking at the mind through the wrong lens. Representations might be one useful tool — but it can’t be the whole story. And if we keep trying to force everything through that framework, we are likely to miss how the brain actually works.

3. The Grounding Problem

Then there’s the problem we tackled last week: how do representations get their meaning?

What makes a particular pattern of neural activity about a dog? Why does another one refer to a tree or a memory or a future plan?

This is the grounding problem. If representations are just patterns in the brain, what ties them to things in the world? Without a link between a representation and what it refers to — between brain pattern and real-world content — we’re left with a pile of patterns and no meaning.

4. Misrepresentations

Finally, there’s the misrepresentation problem.

Imagine you glance out the window and think you see a dog in the garden. Later, you realise it was just a watering can — with a handle that looked a little too much like a tail.

Was your brain wrong?

It seems like is was wrong.

You might shrug and say, sure, misrepresentations happen. What’s the big deal?

Here’s the puzzle: what allows you to say some representations are true and some representations are mistakes?

The misrepresentation problem is related to the reverse inference problem and the grounding problem.

To see why, let’s start with something familiar: like the weather.

Let’s say, while inside your home, you check the weather forecast on your phone. There’s a bright little sun icon over your city.

But when you step outside, it’s not sunny. It’s actually pouring rain.

The forecast was wrong. And you can say that because we’ve all agreed — conventionally — that the sun icon stands for sunny skies. If that icon sometimes stands for sunshine, sometimes for snow, and sometimes for it’s Tuesday, there’d be no way to judge when it got things right or wrong.

Now, back to the brain.

Let’s say there’s a neural pattern — call it P — that reliably shows up when you see a dog. So, we say, P represents dog. What do we say about the time you thought you saw a dog, but it was actually a watering can? There was no dog, but P activated.

Did your brain misrepresent the watering can as a dog?

If P truly means dog, then yes, it misrepresented.

But here’s the problem: you only know it misrepresented if you’ve already accepted that P represents dog — and that’s the very thing we’re trying to prove.

Let’s spell out the logic:

Premise 1: If DOG is present, then P activates.

Premise 2: If P activates, then DOG is present (⇒ P represents DOG).

Conclusion 1: Therefore, P represents DOG. (from P1 & P2)

Premise 3: P activates when a watering can (¬DOG) is present.

Premise 4: If P activates without DOG, that’s a mis-representation.

Conclusion 2: Therefore, P misrepresents DOG. (from P3 & P4, but only if C1 holds)

Here’s the problem: Conclusion 2 only works if Conclusion 1 is already solid. And Conclusion 1 relies on Premise 1 — that P is active for dogs — which Premise 3 now undermines.

But to make matters worse, Premise 2 smuggles in the very content relation that we are trying to justify. So the argument assumes the very thing it’s supposed to explain.

In other words, the whole argument loops back on itself. It’s circular.

Misrepresentation only makes sense if successful representation is already grounded (P1 + P2 → C1), but that grounding itself rests on the reverse-inference move in P2 — exactly the assumption critics say is unjustified.

So where does that leave us?

Critics insist that until we have a non-circular way to say why a neural pattern means dog, all this talk of representations — of concepts, ideas, thoughts, or grounding — is just hand-waving. Calling a watering-can activation a dog error might sound scientific, but only because we’ve already baked in the very meaning we’re trying to prove.

That’s why misrepresentation is often treated as a litmus test. A convincing representational theory of mind can’t just explain how the brain gets things right; it also has to show how the same machinery can get things wrong and still count as representing the world.

So, is there a representational theory that passes the misrepresentation test?

Can we talk about representation without sneaking in assumptions? Or do we have to admit that the feeling of aboutness is an illusion?14

Next Week…

Let’s start to take a look at some of the main theory contenders — and see how they hold up.

Leave a comment

Right now, there’s a lively debate bubbling about large language models (LLMs) — systems like ChatGPT.

We’ve trained them on our language.1 And now some of us are asking whether they could actually understand it? Could these models count as having meaning in a non-trivial, non-merely-syntactical sense?

They certainly talk the talk. Ask one, “What’s a dog?” and it’ll give you a confident reply. It sounds like it knows. Like it understands.

Still, something feels off.

We might argue, a language model has never seen a dog. It’s never thrown a stick, felt soft fur, or heard a bark echo through a park. These systems don’t have eyes or ears. They don’t experience the world2 — they’re only manipulating high-dimensional vectors of numbers, right?3

We might scratch our heads and wonder, how could their outputs have meaning in their own right, rather than the meaning being entirely our interpretations?

That question, in a nutshell, is what philosophers and cognitive scientists call the Grounding Problem.

Put simply, it’s the puzzle of how representations — whether they’re symbols, words, numbers, or high-dimensional vectors — come to be about anything at all.

The grounding problem isn’t a new dilemma. It’s been talked about in academic circles for decades. But the recent rise of powerful language models has brought it back into the spotlight. Ever since Stevan Harnad gave the problem its name in 19904, thinkers have been debating what it is, how serious it is, whether it can be solved, and whether it’s not quite the problem we think it is.5

In this essay, I’m not going to try to untangle every thread of that debate (I couldn’t do it even if I tried). Instead, I’ll focus on one central question:

Do large language models have a grounding problem?

To get there, let’s ask three questions that start from the basics and build to the crux of the issue:

1. What is the grounding problem?

2. How might humans solve the problem?

3. Could a large language model do something similar?

Subscribe now

Before we get started, three quick notes:

1. This is Essay 3 in The Trouble with Meaning series. You don’t need to read the others first — but if you’re curious, Essay 1 tackled Searle’s Chinese Room, and Essay 2 explored the challenges of symbolic communication (with a little help from aliens).

2. Let’s clear up some definitions of some key terms around the grounding problem: representation, aboutness, and meaning.
   

   We could debate these terms for a long time — and philosophers often do. But for now, let’s go with a version that’s unlikely to upset anyone too badly. (Don’t worry, I’ll keep it simple.)

   First, representation.
   

   Some philosophers say a representation has three key parts: a vehicle, a target, and a consumer.6

   • The vehicle is the thing that carries the message. That could be the letters d-o-g on a page, a pattern of brain activity, or a vector of numbers inside a language model.

   • The target is whatever the vehicle is about — an actual dog, the idea of a dog, or even an imaginary dog.

   • And the consumer is some person, brain, or system that can use the link between the vehicle and the target to make sense of the world, answer a question, or make a prediction.

   If any one of those parts is missing, we don’t have a representation.7 But when all three are in place, philosophers say the vehicle has intentionality (that’s their fancy word for aboutness). So when all three line up, the word dog is about a dog.

   And meaning?
   That’s just how well the link holds up in practice.8 If the word dog helps you make the right predictions or take the right actions around actual dogs, then it has meaning. If it doesn’t, the meaning is wrong — or missing altogether.

3. The grounding problem touches on some big, messy questions in philosophy and science.9 To keep things simple, let’s focus on everyday words that we normally use to point to physical objects — dog, apple, tree, that sort of thing. Abstract terms like justice or freedom are a whole other tangle. So we’ll save those for another time.

Okay, with those notes in hand, let’s turn to the first question.

Q1: What is the Grounding Problem?

The symbol grounding problem was first coined by Stevan Harnad in 1990.10 And it starts with a simple question: Could an AI system find meaning in the words it uses?

Not just meaning because we humans interpret the words as meaningful, but a kind of meaning the system can earn in its own right.

If that sounds familiar, it’s because we brushed up against this same puzzle in Essay 1, when we explored Searle’s Chinese Room. Harnad’s version, though, doesn’t involve a room. It involves a dictionary — and two variations of a puzzle he calls the difficult problem and the impossible problem.

Let’s start with the first one.

The difficult problem

Recently, I visited Japan, and tried to learn some Japanese. Let’s imagine you and I are in the same boat — we don’t speak Japanese, but we’re determined to learn. So we head to a bookstore in Tokyo and pick up a dictionary.

The problem we soon face is that this dictionary is a Japanese–Japanese dictionary. Every definition is written using Japanese.

We flip to 犬. It says 飼い犬.
We look up 飼い—it points us to 家族の犬.

On and on it goes. A chain of symbols defined only in terms of other symbols. It’s like we’ve stepped onto a word merry-go-round we can’t get off.

It’s frustrating, yes — but it’s not hopeless. We could still grab a Duolingo subscription. Or bring in a teacher. Or reach for a bilingual dictionary.

In fact, this is exactly how cryptologists crack ancient scripts — by anchoring the unknown language to one they already understand. Given the right bridge, the symbols start to make sense.

That’s the difficult problem.

Now let’s raise the stakes with:

The impossible problem

Imagine trying to learn Japanese with no language at all. No English. No teacher or parent to teach you. And no pictures. Just that same Japanese–Japanese dictionary.

There’s no foothold. No shared experience to tether those words to the world.

You might notice that 飼い often appears alongside 犬 . But what do either of those symbols mean?

You’re just circling words, floating free of any reference.

According to Harnad, that’s the core of the grounding problem: it looks impossible to figure out what a symbol means using symbols alone.

The difficult problem could be solved by linking new symbols, or words, to something you already know. But the impossible problem? It leaves you with nothing to anchor to. No bedrock. One word simply points to another word and you never get off the word merry-go-round.

Some think that’s the exact situation large language models are in today. They’ve been trained on billions of words.11 Sure, some words might appear more frequently alongside other words, but without some link to the world how could an LLM connect words to real dogs, real apples, real trees or real people?

Q2: How Do We Ground Words?

In the years since Harnad’s paper, there has been much discussion about the grounding problem. The authors of a recent paper — The Vector Grounding Problem12 — lay out five types of grounding. That is, five different ways representations might link with the world.

Let’s walk through them. We’ll start with the most intuitive kind.

1. Sensorimotor grounding
The idea here is that we ground concrete words by linking them to our sensory and motor experiences. So the word dog becomes connected to all the sights, sounds, smells, and even the tug on a leash that go along with dogs. In this view, grounding comes from the way words are hooked into our embodied experience of the world.

2. Referential grounding
This one is all about connecting words to specific things in the world. For example, the word dog doesn’t just bring up a general image or idea of a dog — it points to an actual dog. Referential grounding happens when a word is causally connected to real-world things, in a way that allows us to be right or wrong about them.

3. Relational grounding
Here, grounding comes from a word’s place in a web of other words.

You may never have seen a platypus, but you know it’s an animal, from Australia, with a beak and an odd reputation. The word is grounded because it connects to other words in your mental lexicon.

This is how LLMs mostly work — they don’t need to see a dog or platypus to write about them. They just learn how those words relate to others like dog relates to tail and bark, and platypus relates to egg-laying mammal and Australia.

4. Communicative grounding
This one’s all about shared understanding. When we agree on what a word represents, we stabilise its definition through use. The word dog works because we all agree what it points to — at least most of the time.

5. Epistemic grounding
This is about knowledge systems. Think: dictionaries and encyclopaedias. Here, a word is grounded by how it fits into an external body of structured information.

Many assume that only the first two — sensorimotor and referential grounding — can get us off the word merry-go-round. These types of grounding connect words not just to other words, but to the world.

The others — relational, communicative, epistemic — keep us circling within language itself.13

But the authors of The Vector Grounding Problem don’t even think sensorimotor grounding will work. They make the case that only referential grounding stands a chance of getting us off the word merry-go-round. Sensorimotor grounding, they say, doesn’t solve the grounding problem — it just shifts it.14

Why?

Because, they argue, when you come across an actual dog, the light bouncing off the dog triggers a cascade of neural activity in your brain. That pattern is a representational vehicle inside the system. If you now link the word dog to that pattern, you’ve merely linked one representation (the word) to another representation (the neural signal). The merry-go-round keeps spinning with no direct link to the animal itself.15

What these authors say we really need is feedback that is causally connected to real-world things. A way for the system to check its guesses — to know when it’s getting things right or wrong. That, according to the authors, can only be achieved through referential grounding.

Critics argue that there are good reasons to think the authors are wrong about this claim. They argue that other forms of grounding do provide a causal link to the world. But let’s leave that discussion for another essay and turn our attention to LLMs.

Q3: Can LLMs Ground Their Words?

Large language models don’t have eyes. And they don’t have ears or hands. They don’t see or act in the world. So if we’re looking for sensorimotor grounding in an LLM, we’re out of luck.

The authors take this to leave only one live candidate — referential grounding.

And, they argue, some LLMs might already be inching toward referential grounding. How? Through a training process called Reinforcement Learning from Human Feedback — or RLHF.

This is the step where human evaluators score the model’s responses based on things like truthfulness, helpfulness, or appropriateness. The model doesn’t just predict the next word. It’s nudged, over and over, toward answers that align with human judgment.

That feedback loop becomes a kind of bridge to the world. A way for the model’s vectors to get tugged into alignment with the world.16

The authors call this indirect grounding. The idea is that the model doesn’t directly ground its words in the world — it borrows its grounding from us.

We might question whether this really is grounding. And if it is… whether this implies that the model actually understands the words it produces.

But there’s another, perhaps more uncomfortable, question lurking here. If LLMs gain their grounding indirectly through us and referential grounding is the only way to achieve grounding in the world, then…

How do we achieve grounding?

From a naturalist perspective, there’s no higher being handing out gold stars when we use a word correctly. And there’s no little man in our brains whispering, “Yes, that’s a dog.”

Sure, we might say that a certain neural pattern in our brain activates when a dog walks by — but what makes that pattern about dogs? Rather than about cats, or anything else? How do we sort dog from not-dog?

We could say that we learn the category dog as a child from our parents or teachers. But doesn’t that just move the mystery up a generation? How did the first teachers learn it? It seems that somewhere the chain has to meet bedrock, some causal anchor that turns bare correlations into genuine aboutness.

So what counts as that kind of causal link for us? How do we know when our thoughts are grounded? Where does that feedback come from?

Some thinkers try to escape the problem altogether.

They get to this point and decide to sidestep the whole puzzle by declaring there is no objective, independent reality. They say: all we ever have are internal representations. There’s no guarantee of a mind-independent world behind those representations.

This is the anti-realist view. And if it’s true, then grounding isn’t a problem to solve —because nothing ever needs to reach outside the system. There is no grounding problem because the merry-go-round is the whole show.

But if we go that route, we must give up quite a bit.

First, we lose the ability to talk about truth and falsehood in any ordinary sense. Within a self-contained language game, sure, we can talk about coherence. But not about whether a sentence is really true. Or whether it misrepresents the world.

And science? Well… science starts to wobble. Because scientific explanations — from perception to action, from learning to error correction — rest on the assumption that we are in causal contact with a shared, external world. Without that, we’re left either redefining science in purely internal terms, or declaring it inexplicable.

Still, most philosophers and scientists aren’t ready to give up on the real world just yet. They defend some form of realism — even if it is a pragmatic, critical, or perspectival type of realism. As such, they hunt for causal, historical, or social mechanisms that could bridge representations to the world.

The challenge for a physicalist view is to explain how representations latch onto reality — without magic, without a little homunculus, and without giving up on the idea of a real world.

Is that possible?

Some thinkers think it is. But they don’t all agree on how it is done.

I’ll get to those ideas soon, but before we get there, we need to take a closer look at representationalism because the grounding problem poses a particularly tricky challenge for some versions of this view of the mind.

Leave a comment

Next Week…

Let’s take a closer look at representational thinking in cognitive science. Why has it become such a dominant view among neuroscientists, psychologists, and cognitive scientists? And why do some philosophers smell trouble?

In 1972, when NASA launched the Pioneer 10 space probe, they sent it off with a strange looking metal plaque.

The plaque was engraved with pictures and symbols. It showed a schematic of the hyperfine transition of hydrogen, a diagram of the solar system with the planets in order of distance from the Sun, the spacecraft’s outbound path, and a starburst pattern. But the largest image was a line drawing of a naked man and woman, with the man raising his hand in greeting.

Why attach this plaque to a spaceship?

Well…

Maybe one day, far beyond our solar system, some alien life form will stumble across this plaque. Maybe the alien life form will be intelligent. And maybe they will look at the plaque. And maybe they will study the images. And maybe they will realise there is (or was) intelligent life on another planet with scientific knowledge, and peaceful intentions.

That’s the hope behind this plaque. It’s a kind of message in a bottle.

Subscribe now

The idea of pinning this message to NASA’s Pioneer 10 (and later Pioneer 11) didn’t actually come from NASA. It came from the British science writer Eric Burgess. After attending a talk by Carl Sagan, Burgess approached him and asked, in effect, “If this Pioneer 10 probe is leaving our solar system forever, shouldn’t we send it off with a postcard?”

Sagan, who had been lecturing about how we might communicate with intelligent life forms, was instantly hooked. He pitched the idea to NASA, and — astonishingly — NASA said yes.

This was just three weeks before launch. Sagan had to act fast.1

He teamed up with his wife at the time, artist Linda Salzman Sagan, and astrophysicist Frank Drake.2 Together, they sketched out a design for the plaque.3

Their thinking was simple but bold: science might be a kind of universal Rosetta Stone— a shared understanding that could unlock meaning even between totally different kinds of beings. Aliens might not share our language or culture, but if they exist, they would live in the same physical universe we do. And physics, the team hoped, could bridge the communication gap.

So they chose symbols that pointed to features of the universe they believed would be the same everywhere. Hydrogen — the most common element in the cosmos — would always behave the same way. Stars called pulsars would always blink in steady, predictable patterns.4 And the layout of our solar system — our Sun and its planets —would always have the same shape and order, no matter who was looking.5

In other words, they were betting that physics is universal — and that any intelligent beings who stumbled across this little probe would be able to recognise the patterns and figure out where it came from.

That’s a hopeful bet — because even if physics is the same everywhere, interpretation might not be.

I remember learning about this plaque when I was a kid. And I remember wondering — what the heck were they thinking!? Why would we want to tell aliens exactly how to find us? It seemed like a crazy idea to me, and a little bit scary. It felt like we were giving a stranger our home address and just hoping they turned out to be friendly.

But as an adult, a different question comes to mind:

What would an alien need to understand this plaque?

Apart from the obvious — like having the kind of body that could physically interact with the plaque — these aliens would also need a certain way of thinking.

They’d need to understand the idea of a representation.

Why? Because that’s what this plaque is full of: representations.

The sketch of the man and woman is meant to represent human beings. The line-up of circles at the bottom represents our solar system. The raised hand is meant to represent a gesture of friendliness. And the little diagram in the top left — two circles connected by a line — represents a fundamental unit of time and length.

In other words, the aliens would need to understand that the marks and shapes on this plaque stand in for something else — something that isn’t just the marks and the shapes.

The trouble with representations

To state the obvious: the naked sketch of the man and woman is not actually a man and a woman. Understanding that a few lines and marks stand in for real human beings is not a trivial thing.

But let’s say an alien did grasp the idea that some things can represent other things. The next hurdle would be realising that these particular lines and marks were representations — that they weren’t just random grooves or decorative etchings.6

And even if they understood that, they’d also need to recognise that these representations were meant to communicate something.

If they managed all that, they’d face yet another puzzle: figuring out what kind of representations they were looking at. That the drawing of the hyperfine transition of hydrogen was a scientific diagram, not an abstract doodle. That the figures of the man and woman represented living beings, not maps, or blueprints. And maybe trickiest of all, they’d need to interpret the raised hand as a friendly wave — not as a signal of impatience, aggression, or some strange anatomical quirk.7

It’s starting to sound like a complicated task, isn’t it!?

Imagine you’re an alien who stumbles across this plaque. Would you know that the shapes on it were symbols meant to represent something else? How would you know that? And even if you grasped that basic idea, how would you figure out what they were representing? The marks could be read in endless ways — as decoration, as random scratches, as technical diagrams. There’s nothing in the marks themselves that tells you what they mean.

The symbols on this plaque mean something to you and me. But why do they mean something to us? What is it about these representations that allows them to be meaningful for us? And how could they become meaningful to someone (or something) else?

This is the puzzle of representations

We humans rely on representations all the time. In fact, once you start looking for them, you realise they’re everywhere.

Words are an obvious example. We use the letters d-o-g to stand in for a real dog out in the world. Pictures work the same way — a drawing or photo of a dog represents an actual dog.

But representations don’t stop there. Road signs are another example. A red octagon almost universally means Stop, even though there’s nothing about that shape and colour that necessarily means stopping.8

Gestures are representations too. Depending on where you are in the world, a thumbs-up might mean approval, agreement, or perhaps a rude insult.

Representations are so central to how we communicate and navigate the world that it’s easy to overlook just how important they are.

One popular and influential theory of the mind — both in contemporary analytic philosophy and cognitive science — centres around the idea that the mind works by creating and using representations.9

Aptly, this theory is called representationalism. I’ll say more about representations and representationalism in an upcoming essay, but for now note that, in cognitive science, patterns of neural activity are often considered the vehicles of representation.10

When a dog comes into view or you hear the word dog, there is a certain pattern of neural activity that goes on in your brain. This pattern of activity is said to represent the dog.

Because representations are baked into how we communicate, how we make sense of the world, and how we think, we rarely stop to consider how puzzling they really are.

But it’s worth pausing to think about it.

There’s nothing about the letters d-o-g that naturally points to a furry, tail-wagging creature. Without somehow linking d-o-g to something else (like an actual dog), those letters have no meaning.

But what makes a representation about something?

Last week we bumped into this worry in Searle’s famous Chinese Room thought experiment.

Searle described the situation as syntax without semantics. To remind you: syntax is all about form — for example the shape of letters and how they’re ordered or structured. Semantics, on the other hand, is about meaning — what those letters actually represent.

Searle’s argument is that digital computers have syntax, but no way to gain semantics. For an AI, he says, symbols aren’t linked to anything in the world. In fact, he argues that calling them symbols is misleading, because for a symbol to be a symbol it must symbolise something.11 And for a representation to be a representation it must represent something.

As we discussed last week, Searle’s ideas are controversial. Critics argue that he sets up a false divide between syntax and semantics.12

But regardless of where we land on Searle’s argument, there’s a question here that is worth addressing: How do marks on a page, scratches on a plaque, and patterns of neural activity in our brain come to represent anything at all? How do these things get connected to other things in the world?13

Or, as philosophers like to ask:

How are representations grounded?

Searle’s Chinese Room might make it seem like this puzzle — how symbols get their meaning — is a problem unique to digital computers. But is that really the case? Or do we face a similar challenge ourselves?

According to representationalism, certain patterns of neural activity in our brains are just that: representations. But what gives us the right to call them that?

There’s nothing inherently dog-like about a particular pattern of neural activity. A burst of spikes in your visual cortex doesn’t come with a label that says dog.

So how does that pattern represent dog?

One tempting answer is to imagine an inner observer — something inside you that reads the spikes and interprets them. But that just pushes the problem back a level. Who — or what — is interpreting the interpreter?

The other extreme is to say the spikes are just noise. Meaningless firings. But that won’t do either. Those patterns do correlate with real things in the world. When a dog comes into your visual field, something reliable happens in your brain.

So it might seem like we’re stuck between two bad options: either there’s a little homunculus in your head reading the signals, or the signals are meaningless.

But that’s a false dichotomy. There are other possibilities.

From a physicalist perspective, somehow, without a little man inside our heads, we point to patterns in a brain and say they’re about something in the world?

But how? What gives these patterns their aboutness?

That, in a nutshell, is the grounding problem, and it’s exactly where I’ll pick things up in next week’s essay.

Leave a comment

Lately, us humans have been talking to machines.

Some of us might even have a favourite — ChatGPT, Google’s Gemini, or maybe Claude? The LLMs seem helpful, polite, articulate, endlessly patient, and sometimes, weirdly insightful. Which makes them eerily human-like. So naturally, some of us humans have started to wonder:

Is there anybody in there?

This is not a new question. Ever since the early days of computing, us humans have been wondering whether digital computers don’t just seem to understand, but whether they do (or some day could) actually understand. Like real understanding!

Back in 1980, philosopher John Searle labeled this claim strong AI: the idea that a digital computer running the right kind of program doesn’t just simulate a mind — it is a mind.1 And he introduced a thought experiment to push back against this idea. The thought experiment is known as:

The Chinese Room.

You might be thinking, haven’t we spilled enough ink on this thought experiment already? Fair point — I hear you.

But this little thought experiment raises some big questions, and it just so happens that these questions are exactly the kinds of questions I want to explore in a new series on The Trouble with Meaning. So, it’s a great place to start.

Searle’s goal with the Chinese Room was to show why strong AI can’t be right — why, no matter how good a computer’s outputs are, something will always be missing.

But what exactly is missing?

Well… meaning. Meaning is missing.

At least, that’s what Searle thinks.2

So this week, let’s explore Searle’s argument. Can his imagined room tell us anything meaningful about meaning, minds, and real understanding?

To find out, we’ll explore three questions:

1. What is Searle’s Chinese Room thought experiment?

2. Why might we find the Chinese Room argument convincing?

3. Why might we reject Searle’s argument?

Subscribe now

Before We Step Inside the Room…

Let’s step back and consider Searle’s goals here. Yes, he is attacking the idea of strong AI. But the idea of strong AI — is based on the broader philosophical theory of mind:

functionalism.

Functionalism says that what matters about the mind isn’t what it’s made of, but what it does. Just like a computer can run the same software on different hardware, a mind— according to functionalism — could be implemented in a biological brain, or in a machine, or in something else entirely. The key is the function, not the material. Searle doesn’t buy this. And the Chinese Room is his way of pushing back.

To understand that push, there’s one more distinction we need to clear up:

syntax vs. semantics.

In everyday language, syntax usually just means grammar. But Searle uses it more broadly. For him, syntax is about form — the shape of letters and symbols, the order they appear in, and the rules for moving them around.3

If you see the symbol 猫, and you don’t understand Japanese or Mandarin, then, to you, that symbol is just syntax — a shape on a page with no meaning attached.

Semantics, on the other hand, is about meaning — what the symbols and their arrangements stand for, what they refer to, what they represent. In this case, the symbol 猫 means cat in both Mandarin (māo) and Japanese (neko).

Finally, a quick word about AI.

Historically, AI research has followed two major approaches. One focuses on building symbolic systems — programs that manipulate explicit symbols using rules (like, if this, then that). The other focuses on connectionist systems, which use networks of simple units to learn patterns — this is the approach behind most modern AI systems, like large language models (LLMs).4 Searle designed the Chinese Room thought experiment with symbolic AI — the first kind — in mind.5

Q1: What is Searle’s Chinese Room thought experiment?

Imagine you’re trapped in a room with no idea what’s going on outside.

Someone keeps slipping notes through a slot in the wall. The notes have strange squiggles on them. You don’t know what these squiggles mean. Maybe they’re characters from another language — Mandarin, maybe? Or Japanese? Could be alien, for all you know. Or totally made up.

Turns out, they’re Chinese Mandarin symbols. But you have no way of knowing that.

In this room you have two things — a giant book of rules written in English (which you do understand). The rulebook is full of entries in the form of ‘if this, then do that’. The other thing in the room is a basket full of cards, each marked with different symbols.

So, when a new card is slipped into your room you flip through the rulebook and find the matching entry. You follow the rules, which asks you to find certain cards in the basket and arrange them according to the rules. Then once everything is in order, you slide those cards back out the slot.

Outside the room. The someone who has been slipping the symbols into your room, actually understands what those symbols mean. So, it seems to them that they have been chatting with someone fluent in Mandarin. But you don’t understand a word of it. You’ve just been following instructions in a book.

According to Searle, you — the person that is inside the Room — are doing everything a digital computer does when it runs a program. You’ve followed instructions. You’ve rearranged symbols. You’ve produced the right outputs. And you’ve done it so well, that you may have convinced someone outside the room that you understand Mandarin.

In other words, you might have just passed the Turing Test.6

But you don’t really understand. The symbols are just strange squiggles to you.

Searle argues, if going through a rulebook is not enough to give you an understanding of Mandarin, then going through a rulebook — or a computer program — is not enough to give any digital computer an understanding either.

There are two ways we can break down Searle’s argument. Let’s set out the intuitive gloss here, and then we’ll look at a more precise version later.

• Premise 1: If strong AI is true, then the person in the Room understands Mandarin.

• Premise 2: The person in the Room does not understand Mandarin.

• Conclusion: Therefore, strong AI is false.

Q2: Why might we find the Chinese Room argument convincing?

Let’s start with Premise 1 of The Chinese Room Argument:

This premise is a conditional statement — if P then Q

If [strong AI is true], then [the person in the Room understands Mandarin.]

It’s not meant to be a controversial claim — strong AI claims that a digital computer running the right kind of program doesn’t just simulate a mind — it is a mind. So, Searle says, if that’s true, if the person the room is executing the right kind of program, then they too should understand Mandarin.

Before publishing the Chinese Room, Searle presenting his idea in talks and lectures. And during these talks and lectures he didn’t dwell on this premise — he took it as a given. But it’s worth noting that many of the most common criticisms of the Chinese Room are a push back against this premise. So I will address these concerns and Searle’s response to them, in Q3.

For now, let’s turn to Premise 2:

The person in the Room does not understand Mandarin.

This is the part of the argument that feels, to many people, immediately obvious.

You — as the person in the Room — don’t know that you’re communicating in Mandarin. You don’t know what the characters mean. You can’t tell a question from a command, a poem from a shopping list. You’re just flipping through a rulebook and following instructions. You might get really good at this matching game, but it’s not a real conversation. It’s all syntax, and no semantics.

It’s not just that you don’t understand Mandarin right now — it’s that, on Searle’s view, there’s no possible way for you to ever understand Mandarin. There’s nothing in the system that could get you from symbol manipulation to meaning.

Remember, the instruction book doesn’t tell you what the symbols mean — it just tells you which symbols to pull out of the basket and in what order to arrange them. It’s kind of like trying to learn Mandarin with only a Chinese–Chinese dictionary. You look up one unfamiliar character, and it points you to another unfamiliar character, and then another. At no point does meaning break through — it’s all symbols pointing to other symbols, with no anchor to anything you actually grasp.

This is what Searle wants us to see. Understanding isn’t simply about producing the right responses so that we seem to understand. It’s about actually understanding.

So if all a digital computer ever does is follow instructions and manipulate symbols —then, according to Searle, there’s no way it could ever understand anything. Not now. Not later. Not even in principle.

Q3: Why might we reject Searle’s argument?

The Chinese Room thought experiment is a classic in the philosophy of mind — which means it has sparked plenty of debate. But not all objections to Searle’s argument are made for the same reasons.

There have been many critical responses over the years. Let’s review four of them —and why Searle didn’t find them persuasive.

1. The Systems Reply

This is the most common — and probably the most famous — objection. It says: sure, the person in the Room doesn’t understand Mandarin. But the person plus the room plus the rulebook plus the symbols plus the process — in other words, the whole system — does understand.

The Systems Reply is a rejection of Premise 1: If strong AI is true, then the person in the Room understands Mandarin.7

As I mentioned earlier, Premise 1 is a conditional statement — if P, then Q.

P = strong AI is true
Q = the person in the room understands Mandarin

Logically, if you want to deny the conclusion of a conditional argument, you can reject the assumption (the P), or you can reject the consequence (the Q). The Systems Reply rejects the consequence (the Q). It says: no, the person doesn’t understand. But that doesn’t mean the systems reply also rejects the condition (the P). They argue:

If strong AI is true, then it’s not the person alone who understands, it’s the whole system.

Essentially, they are suggesting that what Searle is proposing is that strong AI claims that parts of a digital computer will understand. But no one who is a proponent of strong AI would ever suggest that part of a digital computer could understand.

Searle, raises several pointed objections to the systems reply. His most serious response requires him to revise his thought experiment.

He says: okay, suppose I memorise the whole rulebook. I do all the symbol manipulation in my head. No paper, no baskets, no room. Just me. In that case, I am the system — and, he claims, he still wouldn’t understand a word of Mandarin. So adding up a bunch of non-understanding parts doesn’t somehow produce understanding.

Some critics push back by saying: hang on — memorising all those rules and symbols isn’t nothing. It’s actually a pretty serious cognitive task. Who’s to say you wouldn’t come to understand? Maybe the real problem, they suggest, is a failure of imagination on Searle’s part — an unwillingness to consider that understanding might come from the very process he’s describing.

2. The Robot Reply

This reply says the problem isn’t that a digital program couldn’t understand — it’s that a digital program that doesn’t have body can’t understand. The person in the Room is disconnected from the world. And because the words and symbols we use are tied to things in the world, if we want an AI to truely understand it would need to have access to that world. In other words, it would need to be a robot (or like a robot). It would need to do things like move through space, see, touch, listen, and respond.

In other words, understanding isn’t just about symbol manipulation — it’s about being embedded in the world. That’s where meaning comes from.

The robot reply is also a rejection of Premise 1.

Searle’s answer? He revises his thought experiment.

He asks us to imagine we put the entire Room inside the head of a robot — one with cameras and sensors that take in the outside world. But inside the room, nothing has changed. There’s still a little person sitting in the Room, flipping through rulebooks and manipulating symbols. The robot might have eyes, but the person inside the room doesn’t know that. They are still just receiving symbols and following instructions, with no idea what any of it means.

Searle also suggests, this shows that the robot reply actually concedes his point. Understanding cannot come from syntax — from simply running a syntactical program.

Philosopher Jerry Fodor thinks Searle has missed the point.8

Fodor says: sure, if there’s just a little guy in a box shuffling symbols, of course there’s no understanding. But that’s not what a real robot would be like. Real robots form connections between the symbols they process and what’s happening in the world. In other words, they build causal links between their inputs and their actions. The problem, Fodor says, is that Searle imagines the wrong kind of system — a symbolic bottleneck mediated by a human, one that can’t build or update any of those links.

3. The Wrong Model Reply

As I mentioned earlier, when Searle designed the Chinese Room thought experiment, he had symbolic AI in mind — the classic systems that manipulate symbols by following explicit rules, like if this, then do that — much like the person in the Room. But most modern AI doesn’t work that way. Instead, today’s systems are built using neural networks. Neural nets don’t use fixed rules. Instead rely on patterns learned from data, which is much closer to how we think the brain works.

This has led some critics to argue that Searle’s thought experiment is targeting the wrong kind of system. It doesn’t resemble anything like what advocates of strong AI think a real artificial mind would look like.

So, what is Searle’s reply to the Wrong Model Reply?

You guessed it — he revises the thought experiment again.

This time, instead of just one person in the Room, he asks us to imagine a whole gym full of people: dozens or even hundreds of people, each following part of the rulebook, working together to simulate the program. It’s meant to mirror the kind of distributed processing you get in a neural network.9

But for Searle, it still doesn’t change anything. Even if you scale it up and spread it out, no one in the gym understands Mandarin. And if no one understands, then — he claims — understanding just isn’t happening anywhere in the system.

But notice how much less convincing this argument is. Part of what made the original thought experiment seem so powerful was that you could imagine yourself in the Room. You could picture what it would be like to shuffle those symbols and realise that you didn’t understand a word. But once it becomes a whole gym full of people, that first-person perspective starts to break down. It’s harder to say what the system as a whole does or doesn’t understand — because there is no clear ‘you’ inside it anymore.

4. The Begging the Question Reply

Thought experiments aren’t always designed to discover something new about the world. Most of the time, they’re designed to persuade — to nudge us toward a certain feeling.

That’s how Daniel Dennett sees the Chinese Room. He thinks what Searle is really saying is: “Come on, you know this guy doesn’t understand Mandarin, right?” He’s counting on your gut reaction.

And maybe that gut reaction is right. But gut reactions aren’t arguments. They’re not grounded in logic or evidence. They’re just... intuitions.

When we read Searle’s papers, one phrase comes up again and again: syntax is not enough for semantics. It sounds right. It feels right. So it often goes unquestioned.

Searle doesn’t offer much in the way of empirical evidence to support his claim. He relies instead on logical reasoning.

So let’s take a closer look the the axiom derivation Searle uses.

• Premise 1: Computer programs are syntactic — they operate on symbols based on shape, structure, or position.

• Premise 2: Minds have semantics — real understanding is about meaning, not just structure.

• Premise 3: Syntax alone isn’t enough for semantics.

• Conclusion: Therefore, running a program — even a perfect one — can’t produce real understanding.

Here’s the criticism: the third premise — syntax alone isn’t enough for semantics — is exactly what strong AI is trying to question. It’s the core issue under debate. So, if Searle simply asserts it as one of his premises, then he’s not proving anything — he’s just assuming the conclusion he wants.

That’s what philosophers mean when they say an argument begs the question. It puts the very thing it’s supposed to prove into the setup.

Notice how Premise 1 and Premise 2 on their own don’t quite get us to the conclusion. Searle needs Premise 3. But it is precisely Premise 3 that needs defending.

The Sum Up

The point, according to many critics, isn’t that Searle is definitely wrong. He might be right. It might turn out that real understanding (whatever that turns out to be) really does require more than symbol manipulation. But this argument — the Chinese Room — doesn’t get us there. It just gives us a feeling. And hopes we don’t notice the shortcut.

If Searle thinks symbol manipulation isn’t enough to give us meaning, then we might want to ask… what is enough?

That’s the challenge raised by philosophers like Jerry Fodor. He insists that Searle must answer this question. And he points to decades of empirical research in cognitive science and work done in linguistics that supports the idea that symbol manipulation is something the brain does.

So, if Searle is right and meaning does not come from symbol manipulation, then the real question is still wide open:

Where does meaning come from?

Leave a comment

We used to think intelligence was a uniquely human trait.

Then we relaxed the criteria a little and said, well… at the very least, it’s a biological one.

And now? with our current AI?

Well… now we’re not so sure.

One popular view is that AI isn’t really intelligent. It just seems that way. ChatGPT isn’t really smart. It doesn’t really reason. It simply processes inputs and predicts the next words. It’s just a trick, right? It’s not real intelligence.

But others suspect we tell ourselves this story just to maintain a clean line between us and the bots. And as AI gets better and better, that line between human general intelligence and artificial general intelligence, is starting to look blurrier than ever.

Earlier this week, I was reading a long but deeply fascinating essay in Noema magazine. It was written by two senior AI leads at Google. These authors are clearly in the ‘it’s-a-very-blurry-line’ group. In fact, at one point they suggest, if there is a line, we may have already crossed it.1 Their essay is a goldmine of provocative ideas about the future of artificial intelligence. But one idea in particular caught my attention.

The authors suggest that what we call intelligence isn’t about reasoning from first principles, or understanding in some deep, philosophical sense. It’s about prediction.

Prediction.

I paused when I read that.

Because that word does a lot of heavy lifting in this essay — so much that the authors seem to think it can explain intelligence all by itself.

LLMs predict — that’s what they’re built to do. These systems are, quite literally, prediction machines. And humans predict, too. Decades of neuroscience show that vertebrate brains are constantly predicting the sensory consequences of their own movements.

In the authors’ view, this is a meaningful parallel.

The Functionalist Move

The authors are playing the classic functionalist card: if two devices execute the same function, we should grant them the same title — even if how they perform that function couldn’t be more different.

So, if LLMs and brains perform the same function — prediction — then, in classic functionalist fashion, the authors suggest it doesn’t matter how that function is achieved. What matters is that it is achieved.

The authors argue that LLMs are intelligent — and that this isn’t a fluke. What makes both humans and LLMs intelligent, they claim, is that they both predict.

On this point, their argument can be outlined like this:

Premise 1: Prediction is sufficient for intelligence.

Premise 2: Brains predict.

Premise 3: LLMs predict.

Conclusion: LLMs are intelligent like brains are.

Before we buy this story, there are a few things we should consider.

I want to be clear: this essay is about LLMs — not AI as a whole. I’m focusing on LLMs here because they’re the systems being compared to brains in the Noema piece.

With that in mind, let’s take a closer look at the argument itself.

We could tackle the sufficiency claim in Premise 1 — prediction equals intelligence — and the circular logic it risks. But I’m going to leave that one with you.2

Instead, I want to address Premise 2 and 3. Because we use prediction to define what brains do and what LLMs do and we slide between them as if the meaning of prediction stays constant.

But maybe it doesn’t.

The authors are suggesting that when brains and LLMs predict they are performing the same function — they both predict.

So let’s stress-test the authors’ implicit claim of functional equivalence. For that claim to hold, the word prediction must refer to the same functional role in both systems—brains and LLMs. If prediction means something different in each case, then the analogy breaks down.

Subscribe now

So, What Is a Function?

In functionalist terms, a function is not defined by the material or internal makeup of a system, but by its causal structure.

If causal structure sound like jargon to you — it’s just a shorthand for saying What causes what, and in what order. Basically, the pattern of causes and effects.

Let’s imagine a toasting machine. If we put bread into a toasting machine (that’s the cause), the toasting machine does things that transforms the bread into toast (that’s the effect).

So you might have something like:

inserting bread → toaster heating up → outputting toast

The pattern which turns bread into toast is the casual structure — it is the function.

Analytical philosophers like to define functions more precisely than that. They use something called a Ramsey sentence.3 Ramsey sentences are statements that specify the causal structure in purely logical terms. In other words, they can help us not get confused about what’s actually going on.

To write a Ramsey sentence we start with a statement — usually something like this:

A toaster turns slices of bread into toast. It is activated by inserting bread into a toasting machine, and it results in a browned, crisped output.

For those who love the logic of Ramsey statements this ends up looking something like this:

Let x be a toasting machine. 
Let y be a slice of bread. 

Then: 

∃x ∃y (
  activates(y, x) ∧ 
  transforms(x, y) ∧ 
  produces browned result(x)
)

For us who prefer plain English, it reads like this:

There exists some x such that: It is activated by inserting something (y), it transforms that input, and then it produces a crispy, browned output.

Or put more simply:

insert bread (input) → machine processes → toast (output)

What a Ramsey sentence gives us is a purely functional definition of a toaster. It makes no assumptions about the physical structure of a toaster — all that matters is what a toaster does.

So whatever x is performed by — a metal box, a robot chef, or a solar heat ray — if it has the same causal structure — it follows the same pattern of causes and effects — then it is functionally a toaster.

Prediction Machines

Now that we’ve defined what a function is in functionalist terms — and how Ramsey sentences help us spell that out — we can turn to LLMs and ask. When they predict, are they doing the same thing a brain does — or something functionally different?

We start with a statement like this:

An LLM generates text. It is activated by receiving a prompt — typically a sequence of words — the model runs its algorithm, and it produces an output based on the predicted next word or token.

The Ramsey sentence might look like this:

Let x be a large language model. 
Let y be a prompt.

∃x ∃y (
  activates(y, x) ∧ 
  processes(x, y) ∧ 
  produces prediction(x)
)

In plain English:

There exists some x such that: it is activated by inputting something (y, like a text prompt), it processes that input to generate a response, and then it produces a words as output.

Simple causal form:

insert prompt (input) → model processes → words (output)

From a functionalist point of view, this is enough to say: whatever performs that causal role — whether it’s silicon circuits, mechanical gears, or a biological system — if it produces coherent, predictive text in response to a prompt, it’s functionally the same thing.

Notice how similar LLMs are to Toasters: they can be defined like this

input → internal process → output

Okay, Now It’s the Brain’s Turn

It sure does seem similar, doesn’t it? It seems that a brain receives some input, it processes that input, and out comes a response.

But is that really what the brain is doing?

Let’s take a closer look at what we know about how the brain works.

When you plan a movement — say, reaching for a glass of water — your brain does something interesting. As expected, it sends a signal to your muscles telling them how to move. But it also sends a copy of that signal — an efference copy — to other parts of the brain, including sensory areas like the visual cortex. This copy acts as a kind of internal prediction: it tells your sensory systems what to expect once the movement begins.

This idea — that our sensory systems anticipate the results of our own actions — is well-supported by decades of research in sensorimotor neuroscience.4 It helps explain how we move smoothly and adapt to unexpected changes: adjusting our grip on a glass, catching ourselves mid-step, or bracing for impact.

Beyond movement, the idea that the brain predicts gets more speculative. But one increasingly influential theory in cognitive neuroscience argues that this kind of predictive loop isn’t just for motor control — it’s central to everything the brain does.5 On this view, prediction is how we understand speech, recognise faces, anticipate reactions, or avoid danger. The brain is constantly generating hypotheses, testing them against other data, and adjusting — not just in movement, but across perception, attention, memory, and even emotion.

This broader view — often called predictive processing or the Bayesian brain hypothesis — proposes that the brain’s fundamental job is to predict. It’s a popular and increasingly influential idea, though still the subject of active debate.

But the authors of the Noēma essay invoke it, and they lean on it to support their argument. So let’s grant it in this essay — for the sake of following the argument.

Let’s do the Ramsey statement thing for the brain!

A brain generates a motor command to initiate movement. At the same time, it sends an internal signal — a prediction — to sensory areas, anticipating the expected sensory outcome. Incoming sensory input is then compared to this prediction, allowing the brain to adjust if needed.

The Ramsey statement might look like this:

Let x be a brain. 
Let y be the incoming sensory input.
Let m be a motor command.
Let p be the predicted sensory input.

Then:

∃x ∃m ∃p ∃y (
  generates motor command(x,m) ∧  
  predicts sensory input(x,p) ∧
  compares(p,y) ∧
  adjusts if mismatch (x)
)

In plain English:

There exists some x such that: it is activated by generating a command to move m, it also generates a prediction of what to expect to sense p, incoming input y is compared with p, and if there’s a mismatch, the system adjusts accordingly.

The simple causal form:

motor command and prediction → compare with sensory input → adjust if necessary

Notice how this doesn’t fit the neat input → internal process → output structure we saw with toasters and LLMs. In fact, it flips it. The brain’s prediction comes first, and the sensory input comes later.

Toasters and LLMs are reactive. They wait for input before initiating their function. Brains, on the other hand, are proactive. Prediction acts as a filter that shapes how sensory input is interpreted.

So the difference between brains and LLMs might not be just a matter of hardware — it’s a difference in causal structure. And from a functionalist perspective, that matters. If the causal structures are different, that means the functions are different. We might use the same word — prediction — but functionally speaking, they don’t seem to be the same thing.

The Trouble with Words

Let’s return to the author’s argument:

Premise 1: Prediction is sufficient for intelligence.

Premise 2: Brains predict.

Premise 3: LLMs predict.

Conclusion: LLMs are intelligent like brains are.

Premise 2 and Premise 3 might both be true. But not because brains and LLMs are doing the same thing. They might both predict, but only because we are using the same word to define different functions.

Premise 1 might also be true — maybe anything that predicts counts as intelligent. But if we’re taking a functionalist view, and we’re using the same word (prediction) to describe different functions, then we might expect that intelligence, too, will not be a single thing.

Does this mean no form of artificial intelligence, LLMs or otherwise, could ever predict in the same way a brain does?

Not at all. In fact some forms of AI — especially those trained on sensory data — already show something closer to brain-like predictive behaviour.

But if we think prediction is the key to intelligence because we think LLMs are performing the same function as brains — well… that’s an idea worth pausing over.

Leave a comment

Humans like to build things. We’re pretty good at it, too.

Give us some materials, a bit of time and effort, and the right kind of information — and we can build just about anything.

It’s that last part — the right kind of information — that might be doing more than we often give it credit for.

If we want to build a house — we find what we need in a blueprint. If we are building a bookcase — we follow the IKEA manual.

With step-by-step instructions and a little patience, we (eventually) end up with a functional-ish place to store all our unread novels.

We build shelves, sheds, bookcases, and bridges using this kind of information — the kind we find in blueprints and manuals.

But what if we want to build a brain?

Not a bookshelf — but something that thinks, learns, maybe even understands. Whether we’re building a biological brain or an artificial one, is this the kind of information we need — a blueprint or a manual?

When we think about how life is built, it can seem like a similar process. We need good materials, some time and effort, and the right kind of information.

So what’s the right kind of information for life?1 DNA, right? DNA is often described as life’s instruction manual. And if that’s true, then maybe building a brain isn’t so different from building a building.

But here’s the thing: we don’t usually say that we build life. We say — life grows. There seems to be a line — a difference between what we build and what we grow.

Is that just semantics? Are we just using different words for the animate and inanimate? Or are there real differences between building something and growing something?

And now, us humans are trying to build artificial brains — machines that can think, reason, learn, and maybe even understand the world the way we do.

And it turns out, we’ve pretty good at that, too. We’ve built algorithms that can solve problems, mimic conversation, and beat us at our own games. We talk about it like an engineering challenge — with good materials, a bunch of time and effort, and the right kind of information — we can build an artificial brain.

And whatever line there might have been between what we build and what grows —it’s starting to blur.

So we might want to know — what is the difference between building a brain and growing one? Is there a difference? And if there is, does it matter?

This week, we’re asking three questions:

1. What’s the difference between building and growing?

2. How does complexity grow?

3. What does it mean to grow a brain?

Subscribe now

Q1: What’s the difference between building and growing?

Let’s start by looking at something much simpler than a brain.

Not a living system. Not an artificial intelligence. Just a tiny computer program made of black and white squares.

If you’ve been around this newsletter for a while, you’ve probably heard me mention cellular automata before—like Rule 110. If you are new here and you want a quick overview, I explain them in more detail in the essay Having All The Information Isn’t The Same As Knowing Everything.

Cellular automata are simple computer programs that follow basic rules. They start with a straightforward set of instructions that determine whether each cell in a grid turns on (black), turns off (white), or stays the same — based on the state of its neighbours. That’s it. The set of instructions are incredibly simple.

But sometimes, those instructions — when left to run long enough — can produce astonishingly complex patterns.

The rules for Rule 110 contain some kind of information — they tell the system how to behave. But what kind of information is it?

Is it more like a blueprint? Or more like DNA? Do we build the outcome, or do we grow it?

It’s a little blueprint-like. The rules are fixed. The same starting point will always produce the same result.

But there’s a key difference.
A blueprint tells you what the final structure will be before you begin.
Rule 110 doesn’t.

With a blueprint we can read its instructions and know what will be built.
The only way to find know what a cellular automation, like Rule 110, will produce is to run the program.2

In that way, the rules for Rule 110 are more like DNA: a short sequence of code that doesn’t spell out the final form, but guides a process that unfolds over time.

So what’s the difference between building and growing?

It might come down to this:

Building is plan-driven. You know what you’re aiming for.
Growing is process-driven. You have to let it unfold.

Q2: How Does Complexity Grow?

The complexity that Rule 110 creates can’t be found in the set of instructions. For things that grow, there can be a huge gap between the complexity of the instructions and the complexity of the result those instructions produce.

A few weeks ago, in the essay Why is Complexity So Complex?, I mentioned algorithmic complexity—a way of measure complexity by measuring how much information it takes to describe something.3

The set of instructions for Rule 110 has very low algorithmic complexity — you can write those instructions in a few lines of code.

But let the instructions run for a while… and the complexity explodes.

So if the set of instructions doesn’t contain much information, and if we can start from something simple and still get something astonishingly complex —

Where does the rest of the complexity come from?

We don’t get complexity simply because the set of instructions exists. We need to run it. And to run it, two things are essential.

Time — because the pattern evolves one step at a time. Complexity doesn’t appear all at once; it unfolds.

And energy — because each computational step requires it. Without energy, the system sits still.

So the complexity in outcome of Rule 110 — or the Game of Life — doesn’t come from the code alone. It comes from code + time + energy.

This is how a cellular automaton grows. But is it how a brain grows?

Subscribe now

Q3: What Does it Mean to Grow a Brain?

We’ve seen that a system like Rule 110 grows complexity through just three things:
Code + Time + Energy.

So what about a brain?

A brain also grows though code (we call that code DNA) and time (we call that development) and energy (in the form of nutrients).

So far, the story sounds similar.

But there’s a key difference. Cellular automata like Rule 110 always produces the same outcome. If we start from the same starting point we will always get the same outcome. Every time.

Life doesn’t work that way.

If life were just code + time + energy, then every organism with the same DNA would be identical. Every identical twin would have the same brain, same personality, same preferences. But that’s not what we see. Identical twins aren’t really identical.

So something else must be going on.

Let’s focus on the code for life — DNA.

All currently known living things have DNA.4

We might imagine that DNA is a kind of program — similar to the rule for a cellular automaton. It’s code. It’s fixed. And it tells cells what to do.

But it turns out DNA is far more dynamic than people usually assume.

It’s true that DNA — the A, T, C, and G bases in your cells — is largely fixed from the moment of conception. It’s also true that most of your body’s cells carry the same core DNA, copied from the original zygote. This stability is essential — it allows cells to reliably build proteins and keep the body running.

But DNA doesn’t behave like the static code we find in cellular automata.

Let’s imagine the cells in the body as the cells in a cellular automaton.

In a cellular automaton, the same rule is applied to every cell, equally — regardless of where the cell is located or when the rule is used.

But in biology, that’s not what happens.

Each cell applies a different rule, depending on where it is, what stage of development it’s in, and what its neighbours are doing. The code doesn’t run uniformly — it adapts, locally and constantly. The rules that are applied are constantly changing.

Imagine a newly divided cell — let’s anthropomorphise it and call her Cellder. Where Cellder ends up in the body will shape her future. If she lands near the developing gut, she’ll encounter different chemical signals than if she ends up near the skin. The cells around her release proteins that influence which genes Cellder turns on or off. (Genes are small stretches of DNA that tell the cell how to make specific proteins — you can think of them as biological instructions.)

Basically, these signals tell her which instructions to follow — which genes to turn on or off. In response, Cellder begins producing her own proteins — which then influence how her neighbours behave (and which rules they apply in turn).

It’s a feedback loop.

This turning on and off of genes is known as gene regulation — and it’s shaped by everything from developmental timing to diet, stress, sleep, and exposure to toxins.5

Cells regulate genes based on local signals, and those signals ripple through the body, triggering new patterns of gene expression as the system grows and reorganises itself.

And it doesn’t stop once we’re fully grown. Gene regulation continues throughout life.

So, like a cellular automaton, DNA doesn’t give us a finished blueprint. It gives us a starting state — and a set of rules for how things interact.

But unlike cellular automata, the way the rules are applied changes.

Which genes get used — which parts of the code are read, and when — depends entirely on a cell’s environment, its location, its neighbours, and its history.

DNA is code — but it’s not code in the way most software runs code. It’s not a fixed program. It doesn’t execute sequentially. And it adapts and responds, while it grows.

Which makes it extraordinarily difficult to reverse-engineer.

This is how we grow a brain. Not by following a rigid blueprint, but by running a responsive process: one that unfolds over time, changes as it goes, adapts to its context, and learns from the inside out.

So why do we grow brains this way? Why grow them through adaptation and feedback, instead of building them following a fixed set of instructions?

Because the world is unpredictable. And life has to survive in that unpredictable world.

Building a biological brain — a real, thinking, learning brain — means building one that can deal with that unpredictability.

A rigid, unchanging set of instructions wouldn’t survive long in an environment full of variation, danger, and surprise. Life doesn’t happen in a vacuum. It unfolds in a shifting world of varying light, temperature, nutrients, stress, and noise. A one-size-fits-all solution would break the moment conditions changed.

Instead of hardcoding every outcome — or relying on a fixed set of rules like a cellular automaton — evolution found a better strategy: flexibility.

By giving cells the ability to turn genes on and off — to sense, respond, and adapt — organisms became more resilient. They can not only grow in different environments, they can adapt their behaviour based on experience. A body that can regulate its own genes is more than just a machine following a script — or an algorithm.

It’s a system that changes itself.

Is this what differentiates life from cellular automata?

Life doesn’t just grow — it adapts and changes itself. That might be the real dividing line — not between code and no-code, but between systems that are changed and systems that change themselves.

As we build more advanced AI — systems that learn, adapt, and reshape themselves in response to the world — that distinction seems to be starting to blur.

If a machine starts to change itself the way life does — adapting, responding, reorganising to survive and thrive — is it still something we built?

That’s an interesting question.

But life adapts because it has to — because survival depends on it. The drive to grow, respond, and reorganise comes from within. It’s not an optional upgrade; it’s a condition for continued existence — not just for the individual, but for the species.

Only a few years ago, we would have claimed that AI doesn’t grow like this. And for the most part that is still true. Most AI still changes because we tell it to — when we supply new data, tweak a loss function, or modify its architecture. The motivation, timing, and direction of change still come largely from the outside.

At least for now.

But some researchers are building systems that might change that — models that can learn continuously, reshape themselves, even set their own goals. And in some experimental cases, AI does begin to unfold in surprising, self-organising ways.

But there’s still a whole lot of scaffolding involved. These systems don’t grow in the wild — they grow in the lab. And even when these systems adapt it’s not quite what living systems do.6

In biology, growth and learning aren’t separate. A brain doesn’t get built and then start adapting. It adapts as it builds. The system shapes itself while figuring out how to function.

That’s because the instructions for growth — DNA — are embedded inside the very cells doing the growing. There’s no separation between the thing being changed and the thing doing the changing. The code is in the system. And the system rewrites itself.

And these changes aren’t completely random. They’re shaped by interaction — with the environment, the body, and the system’s own history. The system doesn’t just change — it changes in ways that make sense for its environment, its body, and its goals.

So maybe the more interesting questions are:
Would a neural network ever need to grow like that?
Could it grow like that?

I’ll leave those ones with you. I’d love to know what you think.

Leave a comment

It’s a tempting idea, isn’t it?: consciousness is just what happens when things get complicated enough.

In both neuroscience and pop science, we often hear that if a system — like the human brain — reaches a certain level of complexity, something clicks. The lights turn on. It’s how people sometimes explain the magic of our minds — and increasingly, how they speculate about artificial intelligence. If we keep building bigger AI models, the thinking goes, eventually they’ll get so complex that consciousness will just come along for the ride.

It sounds intuitive. After all, we’re complex, and we’re conscious.

But there is a problem with this idea: the scientific evidence to support it isn’t nearly that clean.1

So what’s really going on here?

To find out, let’s follow the science and ask three questions:

1. What does the neuroscience say about the link between consciousness and complexity?

2. Where does the data start to break down? and

3. Does consciousness come along for the ride when systems become appropriately complex?

Subscribe now

Q1: What Does the Neuroscience Say?

When it comes to consciousness, there’s no shortage of scientific theories. The last time I checked, there were 22 of them. And, as you might expect, the scientists behind these theories don’t tend to agree on much. In fact, they disagree about a lot — what consciousness is, how we should measure it, and where in the brain we might find it.

But in the middle of all this disagreement, there’s one idea almost all of them seem to agree on: Consciousness has something to do with complexity.

There’s probably good reason for the consensus. It’s backed by decades of research showing a strong relationship between brain complexity and conscious awareness. The findings come from all kinds of brain imaging tools, too. And all kinds of definitions of both consciousness and complexity.

And yet, the pattern is remarkably consistent: when the brain’s activity becomes more complex, consciousness tends to be present. When complexity drops, awareness often fades.

I know what you’re thinking… Those are some slippery words you’re using there, Suzi!
What exactly do you mean by consciousness? And how are neuroscientists measuring complexity?

Great questions. I’m glad you asked.

In neuroscience, a common way to understand consciousness is to split it into three categories:

1. States of consciousness — like being awake, asleep, in a coma, under anaesthesia, or tripping on psychedelics.

2. Contents of consciousness — the specific things you’re aware of in any given moment, like a face, a sound, a memory, or the colour red.

3. Self-consciousness — the sense that you are the one having those experiences.

In this essay, let’s just focus on the first two.

To test for states of consciousness, scientists might ask whether a person can follow simple commands, like squeeze my hand. For contents of consciousness, they might ask someone to respond to a specific stimulus. For example, press a button if you see a house.

We use tools like EEG, MEG, fMRI, or TMS to record patterns of brain activity while asking these kinds of questions. That way, we can compare what the brain looks like when someone is consciously responding — and when they’re not. These patterns are known as neural correlates of consciousness, or NCCs. (And yes, they’re controversial. But let’s leave that can of worms for another day.)

Complexity, meanwhile, is also measured in a bunch of ways. But almost all of them come from information theory. Some use Shannon entropy. Others look at how easily brain signals can be compressed — something similar to algorithmic complexity. The general idea is that more complex signals are less predictable. They’re harder to summarise, and harder to compress.2

Take sleep, for example.

When we are in deep non-REM sleep — when dreaming is rare — EEG signals tend to be slow, repetitive, and highly predictable. These are the big, rolling brain waves known as delta waves. But when we are awake, the brain’s electrical activity speeds up. It gets messier, more varied, and harder to predict.

Now imagine running both sets of EEG data — sleep and awake — through a compression algorithm. The sleep data, with its repetitive patterns, is easier to compress. The awake data, full of variation, takes up more space. In other words, the file size for the awake brain is bigger than the file size for the sleeping brain.

So, we say the brain signals we find during sleep are less complex than the ones we find when we’re awake.

Most of the research linking complexity to consciousness has focused on states of consciousness — those broad conditions like being awake, asleep, anaesthetised, or in a coma.

The findings have been so consistent that complexity measures are starting to show up in clinical settings, too. Some doctors have used them to help distinguish between patients who are in a vegetative state and those in a minimally conscious state — conditions where a person might appear unresponsive, but still retain some inner awareness.

But the link between complexity and consciousness doesn’t just show up in states of consciousness. There’s also evidence that complexity tracks the contents of consciousness, too.

In one clever study, participants watched a short film — played forward and backward. The visual input was identical in both cases: same scenes, same colours, same lighting. But only the forward version made narrative sense. The researchers found that brain activity was significantly more complex during the forward viewing than the backward one.

Using dozens of different complexity measures and a wide variety of brain imaging techniques, the same pattern keeps showing up: Complex brain activity tends to go hand-in-hand with conscious experience.

Q2: Where does the data start to break down?

If this story sounds a little too neat for neuroscience, you’re right to be suspicious.

It is true that complexity measures usually rise with consciousness and drop when it disappears. That is a robust finding — there have been hundreds of studies.

But it’s not the whole story. The story is complicated in a few ways — but let’s focus on the fact that there are exceptions to the — usual.

Sometimes complexity measures go up when consciousness measures goes down.

For example, with seizures. During a seizure, the brain doesn’t go quiet — it floods with activity. Neural signals become overactive, disorganised, and unstructured. It’s kind of like the brain goes into chaos. Most complexity measures score that activity as high complexity. But consciousness often disappears.

And then there psychedelics and certain dissociative anaesthetics, like ketamine. These are drugs that could be described as warping consciousness. People are conscious, yes, but they often report vivid hallucinations, distorted time, and out-of-body experiences. Complexity measures tend to be high, but conscious experiences are clearly altered.

So what’s going on?

It might be helpful to distinguish between two things: complexity as whatever our complexity measures are measuring, and complexity as the thing we actually think of when we talk about complexity. Let’s call the first kind measured complexity, and the second meaningful complexity. Ideally, our measured complexity would capture meaningful complexity. But that’s not always the case.

As I mentioned above, most complexity measures in neuroscience come from information theory — things like Shannon entropy or compression algorithms. These tools are great at picking up on structured variation — when structure is there to be found. But they run into trouble when there’s no structure — when the signal is disordered or random.

That’s because both randomness and details are hard to compress. They both show up as high complexity on these measures — even though one is chaotic noise and the other is a meaningful pattern.

So what about meaningful complexity?

There’s an interesting relationship between disorder and meaningful complexity: as disorder increases, complexity tends to rise — up to a point. But then it flips. It hits a peak. As disorder keeps rising, complexity starts to fall. That tipping point — where complexity is at its maximum — is often called the edge of chaos.3
(See last week’s essay for a deeper dive into that idea.)

You can think of this complexity curve as having two sides.

On one side, increasing disorder leads to more complexity. This might be where most brain states sit. When we’re awake and aware, we might be closer to the top of the curve. But during deep sleep, coma, or anaesthesia, we might slide down that slope. On this side of the curve, disorder and complexity rise and fall together, and the relationship holds.

But on the other side of the curve — past the edge of chaos — disorder keeps increasing, while complexity begins to drop.

States like seizures or psychedelics might be examples of the brain operating on the far side of the complexity curve, where disorder keeps rising, but complexity starts to fall.

It might be that straying too far in either direction — too ordered or too chaotic — is what causes consciousness to fade. The brain may need to live in that narrow band near the edge of chaos to sustain awareness.

We don’t know for sure. But that explanation would fit the findings.

From this it might be tempting to conclude that the edge of chaos is the key — that if a system gets itself to the edge of chaos, consciousness will just show up.

Subscribe now

Q3: Does consciousness come along for the ride when systems become appropriately complex?

Well… not quite. And for a few reasons.

First: Not all highly complex things are conscious.

Take the cerebellum — it’s that small, wrinkly structure tucked under the back of the brain. It’s one of the most neuron-dense, intricately wired parts of the brain. By almost any measure, it would be considered highly complex.

So if consciousness simply comes along for the ride when complexity gets high enough, you’d expect the cerebellum to be one of the brain regions that contributes most to conscious experience.

But it’s not.

People with damage to the cerebellum — or those who are missing large parts of it —can still have completely intact conscious experiences. The cerebellum is complex, yes. But apparently not in the way that matters for consciousness.4

Second: Complexity and consciousness may be correlated — but that doesn’t mean one causes the other.

Let’s start with a simple example.

Say every time we saw someone opening an umbrella, we also saw wet sidewalks. If we didn’t understand weather, we might think: umbrellas cause the sidewalks to get wet.

If so, we would be mistaken. Really, there’s a third thing happening: rain. Rain causes both the umbrellas to go up and sidewalks to get wet. The two are correlated — but one doesn’t cause the other. Rain is the hidden third variable that explains both.

But our error might not stop there. If we only conducted our research on the relationship between umbrella use and wet sidewalks — when it was raining — we might think the correlation between umbrella use and wet sidewalks was rock solid — unbreakable even.

But we’d be very wrong. Because in many parts of the world people use umbrellas on sunny days too. For shade, not rain. Umbrellas go up, but sidewalks stay dry.

So what does this have to do with the brain?

Well, almost everything we understand about consciousness comes from studying humans — and to a lesser extent, other animals. We’ve mostly looked at complexity and consciousness in systems where we already assume consciousness is present. So it’s no surprise we keep finding a correlation between complexity and consciousness.

But that’s a bit like studying umbrella use only when it rains.

If we’re only looking at human brains — one specific kind of system — we might be overestimating how universal that complexity–consciousness link really is.

Just like the relationship between umbrellas and wet sidewalks, there could be a third thing happening: a particular kind of architecture or dynamic mechanism or function that is behind both consciousness and complexity.

If that’s the case, complexity and consciousness might just be co-travellers — like umbrellas and wet sidewalks.

So what could this hidden variable be?

Most theories of consciousness seem to agree that there’s a sweet spot for the brain. They don’t usually call it the edge of chaos, but the idea is similar — a balance between too much rigidity and too much randomness. Consciousness, it seems, needs just the right amount of tension in the system.

Where the theories seem to differ is in how they think the brain keeps that balance.

Integrated Information Theory (IIT) calls it integration and differentiation. Global Workspace Theory (GWT) talks about global broadcasting. Predictive processing theory describes the constant dance between expectation and surprise. And Recurrent Processing Theory (RPT) focuses on the difference between feedforward and feedback loops.5

Even the more speculative ideas — like Electromagnetic Field Theory (EMT) and quantum theories — think the system needs to stay in dynamic balance. Not too simple, not too chaotic.6

On almost every one of these mechanisms, today’s AI falls short.

Let’s take large language models, like ChatGPT.

They can simulate language, reason through problems, even hold a conversation. But they don’t have any of the balancing mechanisms that these theories of consciousness see as essential. They don’t self-regulate. They don’t respond to surprise in any adaptive way. They don’t coordinate across modules, or self-stabilise their own perceptions with feedback.

By the standards of any of these theories, they’re missing the thing that counts.7

They may be complex — but not in the way that seems to be required for consciousness.

At least according to these theories.

So where does that leave conscious AI?

According to most consciousness theories, just adding more complexity probably isn’t enough. Consciousness might need — or simply might be — specific kinds of structures, architectures, dynamics, and functions.8 So, if we build an AI like that, would it have conscious experiences?

Maybe.

Of course, it’s possible — perhaps even likely — that all of those theories are wrong. Or, at least incomplete.

Maybe consciousness isn’t simply feedback loops, or integration, or predictive models.

Maybe.

Leave a comment

Next Week…

How much information do you think it takes to build a building? A lot, right?

We need blueprints. Plans. Sequencing. The entire structure has to be laid out in advance before anything can take shape.

In that sense, a blueprint holds a lot of information — everything needed to make a building.

But what about life? How much information does it take to build that?

Well… there’s no master diagram for an oak tree. And no detailed layout for a brain. All we seem to have is DNA. And yet somehow, from this simple code, we get the most complex thing in the known universe.

What kind of information is that?

The underlying structure of the universe is simple1.

Which, frankly, makes absolutely no sense.

Because this same universe gives us apple trees, ant colonies, thunderstorms, stock markets, seven seasons of Black Mirror, democracy, Jupiter, the platform formerly known as Twitter — and physicists who say the underlying structure of the universe is simple.

This article is about complexity. Which, annoyingly — or maybe just appropriately — is a complex topic.

Newtonian and quantum physics have done a great job explaining the fundamentals.
Newton’s laws help us predict the motion of planets and projectiles. Quantum theory lets us model the strange behaviour of particles and atoms. These theories are incredibly good at predicting what happens at both large and small scales (at least in our part of the universe). The math is astonishingly precise.

But these theories haven’t yet told us much about how we get complex things like cells, weather patterns, economies, and societies, from simpler building blocks. Let alone the things we explore in this newsletter: brains, artificial intelligence, and — the mother of all complex things — consciousness.

These are the next big questions in science. Complexity science is still a young science — and right now, there are far more questions than answers. But it’s one of the most exciting frontiers in science.

In future essays, I’d like to look at what complexity science might have to say about consciousness — or at least whether it gets us anywhere interesting.

So where do we begin?

Complexity isn’t something we’re going to pin down in a single essay. Sorry. But let’s see if we can start to unpack the puzzle.

Our three guiding questions are:

1. Can We Define Complexity?

2. Where Does Complexity Come From?, and

3. Is Life a Special Kind of Complexity?

Subscribe now

Just a quick heads-up before we start:

1. There’s no single definition of complexity — and plenty of ways people have tried to measure it. In this essay, I’m going to highlight three key features of complex systems (☝️✌️🤟) and three major attempts to quantify it (1️⃣ 2️⃣ 3️⃣). Just know: there are far more ways to define and measure complexity than I’ll be able to cover here.

2. Because complexity science is still a relatively new and evolving field, many of the ideas in this essay are active areas of research. I’ve added footnotes where I think it’s helpful to flag open questions, differing views, or places where the science is still being figured out.

3. And finally, this is essay 5 in the series on Information and Complexity. I’ll be drawing on ideas from earlier essays (especially Essay 3) — but don’t worry, you don’t need to have read the earlier essays to make sense of this one.

Q1: Can We Define Complexity?

Complexity is difficult to define — which is odd, really, because it seems pretty easy to recognise.

I think (almost) all of us would agree that a brain is complex. And I think most of us would say an apple tree is complex, but probably not as complex as a brain. What about a rock? Not so much. I think most of us would probably say a rock seems fairly simple.

So what is it about a brain that makes it different from a rock? At the deepest level, they’re both made of the same fundamental particles — atoms, quarks, electrons. They follow the same physical laws. And yet one of them thinks, feels, remembers, and dreams — while the other just sits there, being a rock.

☝️ Here’s the first thing to note about complex systems…

They resist reduction. You can break a brain down into atoms, quarks, and electrons — and yes, they all obey the laws of physics. But that alone doesn’t tell you how memory works. Or why a joke makes you laugh.

Even molecular or cellular models often fall short when trying to explain things like emotion or consciousness. These aren’t just properties of the parts — they emerge from the interactions between them.

Brains — like other complex systems — behave in ways that only make sense when you zoom out and consider the whole.

In other words: complex things have emergent properties.

A Quick Note About Emergence

Not everyone loves the word emergence. And fair enough — it can sound a little magical, as if something just appears out of nowhere.

But in physics, emergence isn’t spooky. Complex systems still follow the same physical laws — nothing extra is added. What makes them emergent is that the behaviour of the whole can’t easily be predicted by analysing the parts in isolation. The key lies in how those parts are arranged and how they interact.

The more complex the system, the more of these emergent behaviours we tend to see.

Scientists like to measure things. They like equations. Pointing at a brain and saying it has more emergent properties than a rock might feel satisfying — but it’s not very rigorous. Scientists want to be able to show — mathematically — that a rock is less complex than a brain2.

That’s a big ask.

But it hasn’t stopped people from trying.

1️⃣ Algorithmic Complexity

(is our first measure of complexity)

And out of all the measures it’s the most formal attempt. It’s also known as Kolmogorov complexity.

Imagine you want to reproduce a pattern — let’s say you want to reproduce the string of letters that make up the text in your favourite novel — using a computer program. The algorithmic complexity of that pattern is the length of the shortest possible program that can generate your favourite novel.

The simpler the pattern, the shorter the program. The more complex the pattern, the longer the program has to be.

Take a repeating string like ABABABAB... That pattern is easy to generate. It’s simply a short program that says repeat AB x times. But a novel requires far more instructions — there’s no simple shortcut.

Algorithmic complexity is an elegant way to measure complexity. And it works well for things that have structure — like patterned data streams or digital photographs. But it runs into trouble when there’s no structure — like with a random string of letters.

The shortest possible program to generate a truly random string of letters is simply: print this string of random letters. And that makes the program long.

So a string of letters from a novel and a string of random letters of the same length might both have high algorithmic complexity — even though one is gibberish, and the other tells a story.

We want to be able to say the novel is more complex than the random string. There’s something about the way the letters are arranged that gives the novel meaning —something the random string doesn’t have.

So, there’s a bit of a problem algorithmic complexity. It can’t seem to differentiate between randomness and emergence.

Maybe we need a different approach. Maybe we need to ask when and where complexity arises.

Q2: Where Does Complexity Come From?

Let’s start somewhere that might feel a little unexpected: with a grid of black-and-white squares.

Have you heard of Conway’s Game of Life?

It’s a simple computer program that plays out on a grid of black and white squares. Each square on the grid follows a simple set of rules: it is born, it dies, or it hangs on for another round, depending on what its neighbours are up to.

I know — it’s a long way from brains and apple trees — but as it turns out, these simple systems, called cellular automata, have become a favourite tool for people who think a lot about complexity.

If cellular automata are new to you, I put together a short explainer earlier this year that might help. And if you’d like to see the Game of Life doing its thing, our friend has a great playlist of examples. Here’s one to start with:

One person who’s spent a lot of time thinking about these systems is Stephen Wolfram— a physicist and computer scientist who spent many years exploring cellular automata. He focused on a particular kind: one-dimensional grids with very simple rules. He gave each rule a number — like Rule 30 or Rule 110 — and then just… let them run.

The setup is always the same: you start with a single black square at the top of the grid. Each row that follows depends on the row before it. And the rules are very simple —something like: if a square is black and both its neighbours are black too, then in the next row, that square turns white.

Here’s the set of rules for Rule 110:

And this is how it would play out in the first 5 rows:

Here’s what Rule 110 looks like after 500 steps:

It’s a patterns that has structure, but it’s also not quite regular. It’s orderly, but also unpredictable.

Rules like 110 — rules that produce patterns that look complex — are rare. Most of the rules either don’t do much at all, or they spin out into what looks like chaos. That contrast — the fact that simple rules can lead to such different outcomes — is what caught the attention of Wolfram. He began grouping cellular automata into four broad classes, based on how they behave over time.

• Class 1: systems that quickly settle into a frozen, uniform state.

• Class 2: systems that repeat in neat little loops — orderly and predictable.

• Class 3: systems that spiral into chaos — disorderly and unpredictable.

• Class 4: systems that create intricate patterns. Like Rule 110.

What’s interesting about Class 4 systems is that they look complex to us. They are, in a sense, apparently complex.

So, can we measure that? Can we quantify complexity in a way that reflects how it appears to us?

2️⃣ Apparent Complexity

(is our second measure of complexity)

Earlier, we saw that algorithmic complexity doesn’t quite capture what we mean by complexity — it can’t tell the difference between something that looks richly structured and something that’s just random noise.

Complexity is hard to define, but we often feel like we know it when we see it.
Scott Aaronson, Sean Carroll, and Lauren Ouellette wanted a way to capture that kind of complexity and express it mathematically.

To do this, they looked at how cream mixes into coffee. If we start with cream-on-top coffee we have an ordered state — cream particles on top, coffee particles on the bottom. If we mix the coffee we eventually get equilibrium: where all the coffee and cream particles are evenly mixed (a disordered state). It’s between the cream-on-top state and the mixed state when coffee looks most complex. Swirls and tendrils of cream stretch through the coffee.

To try to measure this apparent complexity, they generated digital images of coffee at different stages of mixing. Then they coarse-grained the images — which is just a fancy way of saying they blurred the images (You could still tell what state the coffee was in — just with the fine grained detail was softened.) Finally, they compressed the resulting images using a standard file compressor like gzip.

They used the file size of the compressed image as a proxy for apparent complexity. Images with fewer details compress into smaller files; those with more structure or variation compress into larger ones.

This approach is similar in spirit to algorithmic complexity. But by blurring the images, they weren’t analysing every pixel of fine-grained detail — they were trying to capture the larger patterns that we associate with apparent complexity.

And they found that file size does a surprisingly good job of measuring apparent complexity.

You can try a version of this yourself. Next time you have a coffee with cream sitting on top, take a photo of it. Then, start mixing your coffee, and take another photo as the swirls begin to form. The file size of the swirled photo will be larger than the cream-on-top photo.

📌 Let’s put a pin in whether apparent complexity is a good measure — we’ll come back to that question a bit later.

But first, I want to focus on something else about Carroll and Aaronson’s experiment — the idea that complexity seems to live in the in-between. Between order and disorder.

The Relationship Between Complexity and Entropy

Let’s consider cream-on-top coffee. It’s orderly — a low-entropy state. And it’s also fairly simple. Fully mixed coffee, on the other and, is disorderly — a high-entropy state. But note that, like cream-on-top coffee, mixed coffee is also simple.

It’s that in-between stage — when the cream is stretching into tendrils and swirling through the coffee — where we see complexity.

So on an entropy scale coffee goes from order to disorder. But on a complexity scale coffee go from simple to complex to simple.

The relationship between entropy and complexity can be seen in other things too. Right after the Big Bang, the universe was in a low-entropy state—ordered, and simple.3 The stars and galaxies hadn’t formed yet.

If we jump to the end of our universe — far, far into the future — entropy will eventually reach a maximum — a state of equilibrium. Physicists call this —

The heat death of the universe. (They’re not exactly known for their charm, those physicists.)

This final state will be a disordered high entropy state, yes. But it will also be simple.

It’s now — this moment in between — when entropy is rising, that complexity happens.

Galaxies form. Stars burn. Planets spin into orbit. And life happens.

Complexity seems to exist in the in between.4

It turns out Carroll and Aaronson weren’t the first to notice that complexity seem to exist in the in between.

Back in the 1980s, around the time Wolfram was classifying his cellular automata, American computer scientist Chris Langton began wondering: why do some of Wolfram’s rules lead to simplicity, others to chaos, and only a rare few to complexity?

Each of Wolfram’s rules is made up of eight instructions. He noticed that some rules were very active — most of the eight instruction turned cells on in the next row (like Rule 254).

Others were mostly inactive — they mostly turned them off (like Rule 1).

So he figured out a way to measure how active a rule is.

3️⃣ He called this parameter lambda, written with the Greek letter λ.

(this is our third way to quantify complexity)

A rule that turns every cell off has λ = 0. A rule that turns half of them on has λ = 0.5. Langton ran simulations across a range of λ values to see what would happen.

At low values, systems fizzled out into nothing much (Class 1). Nudge λ up and you get repeating loops (Class 2). Push it further — toward 0.5 — and chaos erupts (Class 3).

But right in between — clustered around λ ≈ 0.27 — Langton found Class 4. Complexity. Rules like the Game of Life and Rule 110. .

Langton had stumbled onto a kind of Goldilocks zone. Not too rigid. Not too random.

He didn’t call it the Goldilocks zone, though. He called it —

The edge of chaos.

(again with the charming names.)

✌️ The edge of chaos is the second hallmark of complexity.

It’s a big deal in complexity science because it’s where complex, emergent behaviour tends to appear. Weather systems, traffic patterns, economies — even stock markets — all hover near that edge, where order and disorder are constantly in tension.

This all sounds perfectly reasonable — until someone leans in and asks: Wait...

Is this really what we mean by complexity?

Rule 110 looks more complex than Rule 1, and swirling cream in coffee looks more complex than fully mixed coffee. This all feels true.

But if we define complexity by saying we know it when we see it, and then we use that definition to study how and when complexity occurs… the whole approach starts to seem rather circular.

📌 Let’s return to that idea we pinned earlier — apparent complexity. Apparent complexity is one attempt at a more objective approach. It tries to quantify how complex something looks once you’ve blurred away the fine details by measuring file size.

But even apparent complexity smuggles in our assumptions about what matters. The method we choose to blur and compress images will reflect what we think is important. We tend to choose techniques that preserves the features we care about. So in the end, the measure risks reinforcing our definition of complexity.

As you can see, quantifying complexity is tricky5.

But there’s another issue, too.

Cellular automata don’t capture everything we mean when we talk about complexity.

When we say a tree is complex, or a bird is complex, or the brain is complex, we’re pointing to something that cellular automata don’t fully express.

We might say that Rule 110 and a human are both complex — they both show the first two hallmarks of complexity: they have emergent properties, and they exist at the edge of chaos. But only one of them learns, makes predictions, self-organises, and adapts to its environment.

Only one of them keeps complexity going.

🤟 Which brings us to the third hallmark of complexity…

Subscribe now

Q3: Is Life a Special Kind of Complexity?

Some systems, like weather patterns, seem to end up at the edge of chaos by accident. They drift into that zone where order and disorder balance — and then drift right back out. There’s no built-in mechanism holding them there.

But life seems to stick around. Life seems to have evolved to stay at the edge of chaos.6

If a creature becomes too rigid — too fixed in its behaviour or too specialised for a narrow niche — it can’t adapt when conditions change. A sudden drought, a new predator, or a shift in climate — and it’s going to struggle to survive.

On the other hand, if a creature is too unstable — too random, too chaotic — it might forget to eat, fly toward predators, or lay eggs in the snow. Creatures like that don’t last long either.

But right at the edge — where organisms are flexible enough to explore new possibilities, yet stable enough to preserve what works — that’s where life finds its sweet spot.

Theoretical biologist Stuart Kauffman found that systems capable of change — systems that process information, respond to feedback, and adapt over time — tend to self-organise in a way that pushes them right up to the edge of chaos and holds them there.

The edge of chaos is the perfect place for life.7 It offers the maximum diversity of behaviour without sending it into chaos.

This idea — that life doesn’t just sit at the edge of chaos, but builds scaffolds to stay there — echoes something Erwin Schrödinger hinted at in his famous book What Is Life? He saw life as a system that feeds on order to resist disorder. A structure that maintains complexity in a universe that’s always trying to pull it apart.

This might be why life, as we know it, is the most complex thing in the known universe. Not just because it emerges, and not simply because we find it at the edge of chaos — but because it has figured out how to stay there.

All these words — complexity, emergence, life, and adaptation — are things we are only just starting to figure out. There’s still a tone of work to do, but that’s what makes complexity science one of the most exciting frontiers in modern science.

Next Week…

Let’s turn to the question lurking underneath all this: is complexity enough to explain consciousness?

Leave a comment

You woke up today older. Not younger.

You remember yesterday, but tomorrow is a blank page.

In your time on this earth, you’ve seen your fair share of eggs breaking, leaves falling, and cream mixing into coffee. But you have never seen an egg unbreak, a leaf leap back onto its branch, or cream separate itself from coffee.

Of course, you say. Because time moves forward. Everything moves forward. That’s just the way it is.

Physicists call this forward march of time the arrow of time, and it’s everywhere.

Left alone, ice melts, hot coffee cools, your favourite t-shirt fades, the autumn leaves turn to red and gold, and the universe continues its accelerated expansion.

It sounds poetic, but it’s just the second law of thermodynamics.

Over time, entropy tends to increase. Over time, things tend to move from order to disorder.

On their own, with enough time, it is more probable that ice will melt, coffee will cool, your favourite t-shirt will fade, and the leaves on your backyard tree will turn gold. More probable, that is, than water freezing on its own, cold coffee warming itself on its own, your favourite t-shirt restoring its colour, or dead leaves catching the breeze and painting themselves green once more.

Time’s direction feels inevitable. There is a seemingly obvious difference between the past and the future.

But why is the direction of time inevitable? Why doesn’t time move backwards? or not at all?

For reasons we don’t yet fully understand, nearly fourteen billion years ago the universe began as a hot, dense, ordered state. Then it started expanding.

And ever since, entropy has been increasing.

This is sometimes referred to as the Past Hypothesis — the idea that the universe began in a low-entropy state.

That simple fact — that entropy was lower yesterday than it is today and it was lower a year ago than it was yesterday — is what physicists think defines the difference between past and future.

It’s why that once-crisp image on your favourite band tee has faded, much like the memory of the day you bought it.

The arrow of time — the fact that time has a direction — started with the Big Bang.

But here’s the thing about time: nothing in the fundamental laws of physics says it must move in this direction. The equations that govern motion, gravity, and quantum mechanics don’t care which way time flows. If you watched the universe in reverse, the math would still work. The laws of physics are time-symmetric. They don’t distinguish between past and future.

So, at the microscopic level — at the scale of individual atoms and molecules — individual particles don’t have a built-in direction for time. But this is not true at the macroscopic level — the everyday scale which we experience.

At the macroscopic level, time seems to only have a play button. There doesn’t seem to be a rewind. At the macroscopic level, mixed coffee does not spontaneously separate into cream-on-top coffee. And broken eggs don’t reassemble themselves.

According to our best theory of physics, time’s direction is not a fundamental property of the universe. It’s an emergent one. It is something that arises only when we look at macroscopic level.

A strict reductionist might argue that if the microscopic laws are time-symmetric, then time’s arrow is just an illusion — a kind of trick — something that vanishes if we analyse reality at a fine enough scale.

Einstein, too, questioned our intuitive sense of time’s passing. He famously wrote, “the distinction between past, present and future is only a stubbornly persistent illusion.” It sounds mystical, but in context, he was making a point about relativity: from that perspective, time doesn’t flow. All moments simply exist — like locations on a map.

But our experience feels otherwise — the direction of time seems real with real consequences.

So, we might want to ask — if it’s a trick — then what does this trick do?

What does the arrow of time give us?

To answer this question, we might want to consider what life would be like we didn’t experience time as moving forward — it simply stopped? It just held steady, locked in place from today onwards.

Well, for one, that band tee would finally stop fading. It would remain exactly as it is now — forever. Your hot coffee would never cool. The ice cubes in your iced latte would never melt. Autumn would never come. The leaves would never turn red and gold.

I’d miss autumn, but this doesn’t sound all that bad, so far!

But let’s keep going.

It’s true that things would never fall apart. But they would also never come together.

Rainclouds would never form. Seeds would never take root. Snowflakes would never crystallise.

If we didn’t experience the arrow of time, the world as we know it would be very different. Nothing would evolve. Nothing would come into existence. No growth, no decay, no difference between now, yesterday, and tomorrow.

That’s what the arrow of time is. It’s the unraveling of what was into what is, and what is into what will be. It’s our experience of entropy at work.

So, what does the arrow of time give us?

Everything interesting.

Not just fading band tees and melting ice cubes, but growth, structure, complexity, memories, and life.

And those are all very interesting things.

But what is more interesting — what seems to be more shaped by time’s arrow — than experience itself?

William James famously described experience as a stream of consciousness — a continuous, unfolding flow. Our memories, perceptions, and thoughts evolve from moment to moment.

But what is a moment without a direction in time? What is a stream that doesn’t change?

Some suggest that consciousness might be possible without time’s direction. Those who meditate sometimes describe something like this — a state of simply being, without thoughts or any perception of time — but for many of us, it’s difficult to imagine what that would feel like.

For many, it’s difficult to imagine what a sense of self would be like — or a conscious thought existing at all — without a direction in time.

What do you think?

If we didn’t experiencing a direction in time, could we simply be — without context, without narrative?

Leave a comment

Next week…

Sometimes, we hear that complex things like life and consciousness are exceptions to entropy. That complexity pushes back against the universe’s natural drift toward disorder.

But what if complexity — things like life and consciousness — are not exceptions to entropy, but products of it?

Your brain is an information processor.

That’s how we tend to think about it, anyway.

It takes in signals, sorts them, stores them, recalls them when needed. Right now, as you’re reading this, your brain is translating shapes on a screen into words, connecting them to ideas, predicting where this is going.

We navigate the world by processing information.

But hold on. What do we actually mean by information? We use the word all the time, but do we ever stop to think about what it is?

Given how central information is to the topics we explore in this newsletter — artificial intelligence, neuroscience, and consciousness — getting a clearer picture of what we really mean by information is necessary.

In the last essay, I talked about Information Theory — the mathematical framework pioneered by Claude Shannon. This is the kind of information that powers the internet, telecommunications, and even AI.

Information Theory is certainly one way to define information. But it’s not the only way.

Physics talks about information too. And when physicists talk about information, they mean something — physical.

Information in Information Theory and information in physics seem to have little to do with each other. One belongs to the world of communication and computation, the other to the fundamental laws. They seem to be speaking different languages.

But appearances can be deceiving.

There’s a famous problem in physics that hints at a deeper connection between the two.

(This is where things get weird.)

In 1867, physicist James Clerk Maxwell dreamed up a tiny demon. A creature that could seemingly break the rules of physics. And for over a century, some of the brightest minds in science tried to outsmart it — and failed.

The puzzle — now known as Maxwell’s Demon — was a problem in physics. But the answer came from — Information Theory.

Why?

What does information, a concept born in the age of telegraphs and computers, have to do with the fundamental laws of the universe?

The answer lies in entropy.

Entropy?

Wait. What?

Isn’t entropy just another word for disorder? What does that have to do with information?

Turns out, quite a lot, actually.

To find out why, let’s explore three questions:

1. What does physics have to say about entropy?

2. What is the puzzle that stumped physicists for over a century?

3. How was the puzzle solved — and what does that tell us about information?

Subscribe now

Before we get to Maxwell’s Demon, we need to take a quick pit stop at —Information Theory.

Why? Because it will help us understand why the demon puzzled physicists for so long.

In Information Theory, information is all about surprise. The more unpredictable something is, the more information it carries.

Think of it this way: If I tell you the sun will rise in the east tomorrow, you haven’t learned much. But if I tell you the sun will rise in the west… well, that’s a lot of information.

The cool thing about information in Information Theory is that we can actually measure information.

The unit of measurement? It’s the humble bit.

A single bit is the amount of information you get from a yes-or-no question. Flip a fair coin — heads or tails? That’s one bit of information.

But what if I pick a random card from a shuffled deck? Now, instead of just two options, there are 52. That means when you flip over one card, you gain about 5.7 bits of information.

In information theory entropy is defined as the average amount of information or uncertainty associated with a random variable.

That’s really all you need to know about Information Theory for this essay.

But if you’re the kind of person who actually wants to calculate the information content of a coin toss in excruciating detail (no judgment — I’m one of those people), I’ve put together a primer on the math.

Subscribe now

Q1: What Does Physics Have to Say about Entropy?

In Information Theory, entropy is fairly easy. It’s a number. A value you can calculate. No ambiguity.

But in physics? Entropy is… slipperier.

Entropy is one of the most fundamental ideas in science. It explains a lot — from why coffee cools to why time only moves forward.

But it’s also one of the most misunderstood.

To understand the puzzle Maxwell posed, we don’t need to master everything about entropy. We just need to focus on two key definitions of entropy — and one fundamental law of physics.

Two Definitions of Entropy

Definition 1: Entropy as Disorder

If you remember entropy from high school, you probably heard it described as disorder.

You might have been told that a messy room has high entropy. And a tidy room has low entropy. Seems simple enough, right?

But although disorder is a helpful shorthand for thinking about entropy, it’s not quite right. Physicists prefer a different definition — one that explains entropy with more precision.

Definition 2: Entropy as the Number of Possible Arrangements

Ludwig Boltzmann, one of the great pioneers of statistical mechanics, gave us a better way to think about entropy.

Forget disorder. Instead, think of entropy as

Let’s break that down with a cup of coffee.

Imagine you pour cream into your coffee. At first, it sits in a neat layer on top. That’s the macrostate — the big-picture description.

But if we zoom in, we see the microstate — the countless tiny molecules of coffee and cream.

Now, think about all the different ways those molecules could be arranged.

Some of those arrangements will give us a nice, separated layer of cream-on-top coffee. But many, many, many more will give us something else: fully mixed coffee.

So, we say that mixed coffee has higher entropy than cream-on-top coffee — because compared to cream-on-top coffee, there are way more ways for mixed coffee to exist.

This brings us to one of the most famous laws in physics:

The Second Law of Thermodynamics

You’ve probably heard it before — entropy always increases.

But why does it increase?

If you understand the second definition of entropy, the answer is fairly simple — high-entropy states are overwhelmingly more likely than low-entropy ones.

Imagine you take a brand-new deck of cards, perfectly arranged in order.

Now shuffle it.

What are the odds that, by sheer chance, you’ll shuffle it back into perfect order?

Virtually zero.

There’s only one way for the deck to be ordered, but countless ways for it to be unordered (or random). A randomly shuffled deck is simply much more likely than a perfectly ordered one.

The second law explains many everyday phenomena: logs burn to ash but ash never reassembles back into a log, sugar dissolves in water but never crystallises back into a cube, and eggs break but never unbreak.

Technically, the laws of physics don’t forbid an egg from unbreaking.

It’s just… breathtakingly unlikely. If we did the math it would be something in the order of 1 in 10 to the power of 150 billion trillion.

But could there be a loophole?

Could there be a system that — against all odds — seems to defy the Second Law?

Maxwell wondered that too.

Q2: What is the Puzzle That Stumped Physicists for Over a Century?

By 1867, physicists had made peace with the Second Law of Thermodynamics. Entropy always increases.

But then, James Clerk Maxwell had an idea. A provocative idea.

What if… there was an exception?

He imagined a tiny creature that, in theory, could cheat the Second Law. This seemingly simple idea was so perplexing, it haunted physicists for over a century.

We now call it Maxwell’s Demon.

Here’s how it works

Imagine a box filled with gas molecules bouncing around randomly. Over time, those molecules spread out evenly. No part of the box will be hotter or colder than any other. All good so far — that’s exactly what the Second Law predicts.

But now, let’s mess with that.

Divide the box into two chambers, separated by a tiny, frictionless door. And put our demon in charge of that door.

The demon watches the gas molecules very carefully. It knows that heat is just molecules in motion — hot molecules move fast, cold molecules move slow.

So, the demon gets to work.

• Whenever a fast-moving (hot) molecule approaches from the right, the demon quickly opens the door, letting it pass into the left chamber.

• Whenever a slow-moving (cold) molecule approaches from the left, the demon opens the door to let it into the right chamber.

Over time, the left chamber fills with fast molecules — hotter gas. And the right chamber fills with slow molecules — colder gas.

The demon has separated the hot and cold molecules.

It’s created order from disorder.

It’s reversed entropy.

It’s defied the Second Law of Thermodynamics.

That… should be impossible.

Where’s the Catch?

In the real world, reducing entropy always comes at a cost.

A refrigerator keeps food cold — but it does so by consuming electricity and releasing heat. A steam engine converts heat into motion — but only by burning fuel and releasing waste energy.

If you want to decrease entropy in one place, you must increase it somewhere else.

So… if the demon is actually lowering the entropy of the gas, where is the cost?

It can’t be from opening and closing the door — Maxwell specifically imagined it as frictionless, requiring no energy at all.

For over a century, no one could answer this paradox.

Physicists believed the Second Law was… well law. Unbreakable.

And yet, Maxwell’s Demon was suggesting he could.

Eggs break, but don’t unbreak.
Hot coffee cools, but doesn’t reheat itself.
And gas molecules don’t separate on their own — unless there’s a cost.

So where was entropy sneaking back in?

Where was the hidden price?

Q3: How Was the Puzzle Solved — And What Does That Tell Us About Information?

Physicists were convinced that Maxwell’s thought experiment had to be wrong.

But for decades, no one could say why.

Then, in 1929, a Hungarian physicist named Leó Szilárd took a new approach. Like others, he suspected the demon must be paying a thermodynamic cost. But he had a hunch the cost wasn’t obvious.

So he ran the numbers. And he found something surprising. Each time the demon makes a decision—choosing whether to let a molecule through—it gains one bit of information about the system.

One bit.

As in Shannon information.

The same fundamental unit that underpins modern computing. A simple choice between two possibilities: fast or slow, left or right.

But in 1929, acquiring information wasn’t considered work. Information was just an abstract concept—mathematical, not physical.

And that was the key. Szilárd’s insight suggested something radical: information isn’t free.

Somehow, gathering information has a hidden energy cost.

Let’s take a moment to appreciate just how strange this is.

We started with the idea that thermodynamic entropy and Shannon entropy belonged to completely different worlds.

One describes heat and disorder in physical systems. The other describes messages, bits, and uncertainty in communication. But this century-old problem in thermodynamics seemed be about bits — Shannon information.

What Szilárd was suggesting was profound: information has physical consequences.

But how?

That mystery remained unsolved for another 30 years.

In 1961, a physicist at IBM named Rolf Landauer made a crucial realisation. For the demon to keep functioning — to sort molecules forever — it must eventually erase old information to make room for new measurements.

And erasing information, Landauer discovered, is a physical process. A process that releases heat and increases entropy. This was the missing entropy.

It wasn’t in the demon’s movements.

And it wasn’t in the frictionless door.

It was hiding in the cost of erasing information.

The demon wasn’t cheating the Second Law.

To tie this all together, we need one last definition of entropy.

Definition 3: Entropy and Energy

Entropy is sometimes described in terms of energy, but it’s not actually energy.

Rather,

Physicists love closed systems — systems where no energy is added or removed.
Maxwell’s box of gas molecules is one of these systems.

The key thing about closed systems is that their total energy never changes.
That’s the First Law of Thermodynamics.

But while the total energy stays the same, what does change is how much of that energy is useful.

A charged battery is an example of low-entropy energy—energy that’s ready to do work. As the battery is used, that energy isn’t lost, but it becomes less useful, spreading out as heat. Entropy increases.

And that’s exactly what happens when the demon erases information.

To keep sorting molecules, the demon has to gather information about the movement of the molecules. And eventually, it has to erase old information to make room for new information. And that’s when the price is paid.

Each time the demon erases a bit of information, it releases heat — converting useful, low-entropy energy into less useful, high-entropy energy.

That’s where the missing entropy was hiding all along.

The second law wins.

Landauer had solved the puzzle.

Information processing isn’t free — it has real, physical consequences.

At least, that’s the mainstream view.

As always, the conversation doesn’t end.

Maxwell’s Demon may have been defeated, but some lingering questions remain. And in the spirit of intellectual honesty, let’s take a moment to acknowledge some of them.

Are Information and Thermodynamic Entropy Really the Same Thing?

The math is strikingly similar. Both Shannon entropy (from communication theory) and thermodynamic entropy (from physics) follow the same fundamental equations. They both describe uncertainty, randomness, and constraints on what’s possible.

But are they fundamentally connected? Or is the overlap just a coincidence—a mathematical trick that happens to work in both worlds?

Most physicists believe the connection runs deep. But some remain unconvinced.

Did We Really Defeat the Demon?

For most scientists, Szilárd and Landauer gave us the final answer. Maxwell’s Demon isn’t a loophole in thermodynamics — it just hides the cost of information processing.

But some critics, like John Earman and John D. Norton, argue that this explanation cheats. They claim Szilárd and Landauer’s arguments assume from the start that the Second Law of Thermodynamics must hold. Then, based on that assumption, they derive constraints — like the energy cost of erasing information — that justify why entropy always increases.

If that’s true, then the whole argument is circular reasoning.

They argue that instead of proving the Second Law, the argument just restates it in a more complicated way.

This isn’t a mainstream view in physics or information theory. But in the philosophy of physics, the debate isn’t entirely closed.

The Bigger Question: What Is Information, Really?

So far, we’ve treated information as something quantifiable, measurable, and tied to entropy and computation. Bits and data.

But… is that all information is?

When we talk about meaning, interpretation, and understanding, are we still talking about the same thing? Does the physics definition of information capture everything we mean by information?

Or is there something more?

Leave a comment

Next Week...

Even with lingering uncertainties, the mainstream position in modern physics is clear: Information isn’t just an abstract concept — it has real, measurable effects in the physical world.

And if that’s true, it raises some big questions.

Think about it.

Every time you permanently delete a file — you’re not just managing information. You’re increasing entropy. If erasing information has a physical cost, what does that mean for systems that process information?

Computers process information.

So do brains.

Entropy might help us understand memory and thinking in machines — like AI, neural networks, and computation.

But what about us?

Next week, let’s get speculative.

If entropy is at the heart of information processing… then what else might it be responsible for? And what does this tell us about systems that think?

Hello Friends!

I was all set to continue our series on information and complexity this week — specifically, the physics of information. But then physics decided to stage a live demonstration instead.

As many of you know, I live on the east coast of Australia. This week, Cyclone Alfred made an unwelcome visit. He was a category 2 — not the kind of monster storm you see making headlines as a category 5 hurricane in the U.S. or a super typhoon in the Pacific, but he stuck around long enough to make life interesting.

Cyclones usually stick to the northern parts of the Australian coast, so having one swing by our area was a bit of a surprise. And since we live near the coastline in a flood-prone area, we decided it was a good time for an impromptu evacuation.

We lost power, internet, and cell coverage for a couple of days — which was a little inconvenient. But we’re back now, in cleanup mode. There’s some flooding in neighboring areas, but we’re okay here. And our community is okay.

So, no essay this week — just a quick note to say we’re safe, we’re cleaning up, and if you’re in the area, I hope you’re doing okay too.

Barring any more unexpected meteorological events, I’ll be back next week with more on information and complexity.

Take care,

Suzi

What is the mind?

If you’ve ever thought about this question, you’ve probably run into René Descartes. He’s the guy who famously argued that the mind and body are made of completely different stuff — one physical, one immaterial. His ideas shaped the conversation for generations.

But, of course, the story doesn’t end with Descartes. Ideas about the mind aren’t set in stone. They shift with the times, like intellectual fashion trends. What we think the mind is is influenced by what’s happening in philosophy, science, and even technology.

Today, if you tune into a podcast or pick up a popular book on the mind, you won’t hear much about what the mind is made of. Instead, the focus is on what it does — thinking, deciding, remembering, feeling. It’s all about function. And one of the key functions of the mind, it seems, is to process information.

So how did we get here? How did we go from philosophers arguing about evil demons tricking our senses to a world where the mind is framed in terms of information processing?

To answer that, we need to take a little trip through history. Along the way, we’ll meet some contemplative philosophers, some revolutionary psychologists, and a surprising theory borrowed from the world of codebreakers.

Let’s take our trip in three key stages:

1. The move from philosophy to science

2. The rise and fall of behaviourism

3. And the sneaky influence of information theory

Subscribe now

1. The Move from Philosophy to Science

If you asked someone today, who studies the mind?, they’d probably point to a cognitive psychologist — maybe someone in a lab, running experiments or scanning brains. But not long ago, the mind wasn’t something you studied with machines and measurements. It was something you thought about.

The mind was the philosopher’s territory. And one of the biggest questions these philosophers wrestled with was knowledge. How do we know what we know? How can we be sure we know anything at all?

On this question, philosophers split into two camps: the rationalists and the empiricists.

Rationalists believed that reason was our most reliable tool for understanding the world. If you want to uncover truth, they argued, the best place to start is with your own thoughts — carefully working through ideas, step by step.

Empiricists, on the other hand, weren’t having any of that armchair speculation. For them, knowledge came from the senses. If you want to know how the world works, you have to get out there and observe it.

Like most intellectual divides, the reality was messier than the labels suggest. Few thinkers were purely on one side or the other. Even Descartes — the poster child for rationalism — borrowed from both camps. His famous, I think, therefore I am is a classic rationalist move, but elsewhere he wrote about the importance of studying nature through observation.

Still, during Descartes’s time, rationalism was in fashion. But by the late 17th century, the tide was turning — especially in the English-speaking world.

What changed?

Science.

While philosophers debated knowledge in the halls of academia, something seismic was happening in the world of science. Astronomers were charting the heavens with telescopes. Biologists were peering into the hidden world of cells with microscopes. The scientific method — based on observation, measurement, and experimentation — was transforming how we understood the universe.

With that shift came a new way of thinking about the mind. If the best way to understand the natural world was through observation and measurement, maybe the same was true for the mind itself.

Enter John Locke — one of the first philosophers to really run with this idea. Locke argued that the mind starts out as a blank slate — tabula rasa. Everything we know, he claimed, comes from experience. No built-in truths, no innate knowledge — just sensory impressions slowly building up over time.

David Hume took this further. He wasn’t just interested in what we know — he wanted to understand how thoughts stick together in our minds. He came up with the Laws of Association — a set of rules that explain why certain ideas seem to naturally follow others. If two things happen close together in time, we assume one caused the other. If two things look alike, we link them in our minds.

Locke and Hume were still working with ideas, not experiments. But their influence planted a seed. If experience shapes the mind... and if thought follows predictable patterns... then maybe the mind could be studied in the same way as the natural world — systematically, scientifically, even experimentally.

Philosophers had laid the foundations. But the study of the mind was about to break free from philosophy. It was becoming a science.

2. The Rise and Fall of Behaviourism

In the late 19th century, a new discipline began to emerge — one that aimed to study the mind not through speculation or introspection, but through careful observation and experiment.

Psychology.

At first, psychology’s ambitions were broad. Early pioneers like Wilhelm Wundt and William James explored everything from perception and memory to attention and emotion. They believed the mind could be studied scientifically — without abandoning the richness of inner experience.

But by the early 20th century, psychology took a dramatic turn — and, in some ways, narrowed its focus.

Enter John B. Watson.

Watson wanted to turn psychology into a ‘proper’ science — one that could rival physics or chemistry. But there was a problem: thoughts, feelings, and mental states weren’t observable. You couldn’t see them, measure them, or put them under a microscope. And for Watson, that made them unscientific.

If psychology wanted to be taken seriously, he argued, it had to ditch the whole messy business of the mind altogether. Instead, psychologists should focus only on what they could observe directly: behaviour.

And just like that, behaviourism was the new trend.

The behaviourists took Locke’s blank slate idea as a starting point. If the mind begins empty and all knowledge comes from experience, then psychology should focus on how experiences shape behaviour.

In other words — they should focus on learning.

And if Hume’s Laws of Association explained how ideas stick together, maybe Hume’s Laws could explain how we learn everything — from walking to language to complex problem-solving.

Watson’s idea was not new — philosophers had been arguing about it for centuries. But now, behaviourists wanted to take it a step further: if learning was just a matter of experience and association, could psychology predict and control learning?

This was the era of the famous conditioning experiments. Pavlov’s dogs, salivating at the sound of a bell. B.F. Skinner’s pigeons, pecking at levers for food. Watson himself made a bold claim: give him any infant, and with the right conditioning, he could turn them into a doctor, a lawyer — or even a tightrope walker.

Bahviourism was an idea that fit neatly with the American dream. If everyone starts with the same blank slate, then success isn’t about talent or innate ability — it’s about learning. Given the right opportunity and hard work, every person can succeed and attain a better life.

By the mid-20th century, behaviourism was psychology. Nearly every major department in the U.S. was run by behaviourists. The study of the mind — once the realm of philosophers — had been stripped down to stimulus and response.

But not everyone was convinced.

What’s surprising isn’t that behaviourism rose to dominance — it’s how long it managed to hold on, despite mounting evidence that something was missing.

Philosophers like Descartes had argued centuries earlier that the mind must have innate structures that shape how we think. Kant suggested that certain ways of understanding the world — like space, time, and causality — might be hardwired into our minds. Even Darwin’s theory of evolution hinted that mental capacities, just like physical traits, could be products of natural selection.

Yet behaviourism pushed all of that aside. The idea that learning was everything became dogma. For decades, psychology largely ignored the role of biology in shaping the mind.

But outside the U.S., the first cracks in behaviourism’s hold were starting to appear.

In Montreal, Donald Hebb proposed what, at the time, seemed like a radical idea: learning wasn’t just about external behaviour — it had a biological basis. His theory, now summed up by the phrase ‘cells that fire together, wire together,’ suggested that learning happened through changes in the brain’s wiring.

But in the U.S., the challenge to behaviourism didn’t come from biology.

It came from a completely different way of thinking — it redefined what the mind is.

This new way of thinking wasn’t about behaviour.

It wasn’t even about biology.

It was about information.

Subscribe now

3. The Sneaky Influence of Information Theory

Right now, if you were to name the most influential figures in psychology, you’d probably rattle off the big ones — James, Pavlov, Skinner, Piaget.

But one of the most quietly influential figures in how we think about the mind today isn’t someone most people recognise.

His name was George Miller.

And in 1956, everything he thought he believed about the mind was about to be turned upside down.

In the early 1950s, Miller was like most American psychologists of his time: a committed behaviourist. He was interested in language and communication, and —unsurprisingly — he believed their foundations were rooted in stimulus, response, and reinforcement.

But then came 1956.

Miller attended a symposium.

And everything changed.

Before we get to that moment — the moment Miller’s entire perspective changed —let’s zoom out for a second.

Because something big was brewing outside of psychology.

Something that had nothing to do with rats in mazes or conditioned responses.

Something that wasn’t supposed to be about the mind at all.

The hot topic was information.

Just a few years earlier, Claude Shannon had published A Mathematical Theory of Communication, and the new field — Information Theory — was causing a stir.

Shannon wasn’t a psychologist. He wasn’t trying to crack the mystery of the human mind. He was a mathematician who wanted to figure out how to send messages more efficiently.

But, in doing so, he accidentally gave us a whole new way to think about thinking — though not exactly in the way he intended.

Shannon’s Information Theory at its core is about quantifying information. Information was defined as:

Let’s say our event is turning over to the top card in a deck of playing cards.

Take an unshuffled deck so you know the order of the cards. When you turn over the top card, do you gain information?

No.

You don’t gain information because you know the order of the cards before you flip over the top card — you know exactly what the top card will be. There is no uncertainty. No surprise. No Learning. No information.

But what if the deck was shuffled?

With a well-shuffled deck you don’t know what the top card will be. When you flip over that top card, your uncertainty drops. You learn something new. That, according to Shannon’s theory, is information.

If you want a deeper dive into Shannon’s Information Theory, I wrote a more detailed explanation in my essay, What is Information.

But let’s get back to 1956.

In September of 1956, at MIT’s Second Symposium on Information Theory, our friend, George Miller sat in the audience.

Miller listened as engineers, mathematicians, and computer scientists described how machines could store memories, solve problems, and process information.

And something clicked.

If computers do this... couldn’t the mind?

Maybe learning wasn’t just stimulus, response and reinforcement, like behaviourists claimed. Maybe the mind was processing information.

That was the moment behaviourism lost its grip on Miller.

Back in 1951, Miller had written a book called Language and Communication, and right there in the preface, he made his stance clear:

"The bias is behavioristic."

At the time, that wasn’t surprising — behaviourism was psychology’s dominant framework, especially in the U.S.

But fast forward eleven years, and Miller’s writing had a very different message.

His new book was called Psychology: The Science of Mental Life.

That title alone said it all.

This wasn’t just a slight shift in perspective. This was a complete break from the idea that psychology should only study behaviour. With that book, Miller wasn’t just changing his own mind — he was helping to change the entire field of psychology.

The cognitive revolution was in full swing.

Psychology had moved on from behaviourism.

Psychologists weren’t the only ones breaking away from behaviourism.

Philosophy was going through its own transformation.

In 1963, Hilary Putnam introduced a now-famous thought experiment that exposed a major flaw in behaviourism.

Imagine a society of Super-Spartans — people who feel pain but never show it. No grimacing. No wincing. No crying out.

If you saw them, you’d never know they were in pain.

As we know, behaviourists thought pain was just a pattern of behaviour — grimacing, wincing, crying out. But Putnam’s thought experiment made that idea look absurd.

If pain were only behaviour, then these Super-Spartans wouldn’t actually be in pain.

But, according to the thought experiment, they were in pain.

Putnam argued that this meant pain had to be more than just behaviour.

Putnam was making a case for functionalism — the idea that mental states are defined by what they do, not by what they’re made of.

Just as psychology was shifting from behaviourism to cognitivism, philosophy was shifting from behaviourism to functionalism.

These shifts weren’t identical, but they shared a crucial commonality: both abandoned the idea that mental states could be fully explained in terms of external behavior.

Cognitivism embraced information processing directly, while functionalism set up a framework that would later click perfectly with computational models of the mind. In their own ways, both cognitivism and functionalism were stepping stones to an era in which thinking about the mind meant thinking about information.

Nowadays, it is common to think of the mind as an information processor. Before this shift in thinking, mental processes — things like memory, problem-solving, even consciousness — were seen as off-limits to science. Too abstract. Too mysterious. Too unobservable.

But thinking of the mind as an information-processing system changed the rules.

Suddenly, things that once seemed unknowable could be measured, modeled, and tested. The mind wasn’t a black box anymore — it was something you could study.

But there’s a catch.

What we mean by information and information processing has expanded far beyond what Shannon originally meant — especially in psychology and philosophy.

When we say computers process information, we mean they process something close to Shannon-style information — signals, statistical patterns, bits that flow from one place to another.

But when we talk about the mind processing information, we often mean something more — the information has meaning.

Shannon’s Information Theory was never about meaning. As he put it, “The semantic aspects of communication are irrelevant to the engineering problem.” He wasn’t concerned with what a message meant — only with how efficiently it could be transmitted.

But cognitive psychologists and philosophers? They care about meaning.

When we say the brain processes information, most of us aren’t thinking about that information as Shannon’s bits and bytes. We’re thinking about information as something that means something to the system (the system being us).

When we talk about the brain storing information (aka memory), we don’t just mean raw data. Memories for us have significance, interpretation, and meaning.

Yet, despite these different concepts of information, we still draw parallels between the information processed by a mind and the information processed by a computer.

But how different are these kinds of information, really? Is information just one thing, or are we lumping together fundamentally distinct ideas under the same word? And if they are different, does that difference matter?

If we imagine artificial general intelligence (AGI), we are potentially mixing together multiple meanings of information — blending Shannon-style information processing with other mathematical frameworks, as well as cognitive theories about meaning and representation. The boundaries between these definitions of information aren’t always clear.

And with this slippery concept of information precariously in hand, we like to ask questions like…

If the mind is an information processor… why couldn’t a machine that processes information have a mind?

Leave a comment

Next Week…

Earlier this year, just for fun, I wrote some code to analyse the words I use most frequently in my essays (excluding all the usual suspects like ‘the,’ ‘is,’ and ‘and’).

The top two words? Brain and consciousness. No surprises there.

But the third most common word? Information.

We lean on the word information all the time. And not just when we talk about the mind or computers. Weather forecasts, medical diagnoses, and even restaurant reviews all rely on the careful interpretation of information.

But do we even know what we mean by information?

Next week, let’s start pulling on that thread.

From DNA to black holes, information seems to take on different meanings depending on the context. We have to start somewhere — so let’s start by pulling on the physics thread.

What happens when information theory meets physics?

Science comes in many flavours.

There’s physics, biology, chemistry, neuroscience, psychology, geology, astronomy… the list goes on… and on.

Each of these sciences has its own specialty — from uncovering the fundamental laws of the universe to studying complex human behaviours and social systems.

Some sciences — looking at you, physics — aim to uncover the basic rules that govern everything. Others focus on what we might call emergent properties: patterns and phenomena that we get from the interactions of simpler components. Think of how biology builds from chemistry or how economics might build on principles of psychology.

These special sciences, as philosopher Jerry Fodor calls them, each have their own methods, theories, and tools. And while they might not always call their subjects emergent properties (many dislike the term, but that’s a discussion for another day), they all operate on one shared assumption: physics is fundamental. Everything these sciences study — from chemical reactions to social networks — ultimately happens in the physical world, following the laws of physics.

If you’re a physicalist (as most scientists are), this idea probably feels intuitive. According to the physicalist, everything we observe, from a lightning storm to a market crash, boils down to physical matter and energy interacting according to physical principles.

And yet, this creates quite the puzzle.

If the laws of physics can, in theory, explain everything, why do we need all these other sciences? Why do we turn to chemistry to understand molecules, biology to understand life, or psychology to explain human behaviour? What’s left for these other sciences to discover if physics already covers everything?

This is what I’m calling the puzzle of the special sciences (borrowing the term from Fodor’s famous paper). It’s not just a thorny puzzle about how science works — the answer to this puzzle has implications for how we think about complexity, reality, and even the possibility of artificial minds.

To start unravelling this puzzle, we’ll focus on three key questions:

1. How do we typically try to resolve the puzzle of the special sciences?

2. What does it mean if the special sciences discover things that physics alone can’t explain? and,

3. What does the puzzle of the special sciences mean for consciousness?

This essay kicks off a new series on Information and Complexity. It builds on ideas we’ve explored before — like Reductionism and Functionalism — and raises new questions we started exploring in the Little Series about Big Questions.

Subscribe now

Q1: How do we typically resolve the puzzle of the Special Sciences?

One common way to tackle this puzzle is by appealing to the concept of different levels of abstraction. The idea is simple: sure, everything is ultimately made up of physical objects and events, but that doesn’t mean the sciences are in conflict — they’re just describing the same physical reality from different perspectives.

We can think of it a bit like zooming in and out with a camera. A sociologist studying group dynamics, a psychologist examining individual behaviour, and a neuroscientist analysing brain activity are all looking at the same underlying reality. The sociologist might focus on how societal norms shape behavior, while the psychologist looks at an individual’s emotions, and the neuroscientist measures the electrical signals in the brain that correlate with those emotions. Different zoom levels, same scene.

The same applies across the natural sciences: a biologist studying cells, a chemist looking at molecules, and a physicist probing atoms are all exploring physical systems — just at different scales. From this perspective, the special sciences don’t discover new laws of nature. Instead, they provide more accessible ways to understand physical principles in context.

At first glance, this seems like a tidy solution to our puzzle. All the sciences fit neatly together, with physics at the foundation, tying everything back to the same universal laws.

But is it really that simple?

Q2: What does it mean if the special sciences discover things that physics can’t?

Think about a psychologist studying memory or an economist exploring markets. When the psychologist uncovers a new principle about how we recall memories, or the economist identifies a pattern in market behaviour, what exactly are they discovering? Are they just describing physics in a different way — like zooming out to a broader perspective? Or are they finding something genuinely new — a principle or law about the world that physics alone could never reveal?

This question pulls us into the long-standing debate between the reductionists and the non-reductionists.

Reductionism is the idea that complex phenomena — like memories or markets — can ultimately be explained in terms of simpler components. A reductionist might say that memory will eventually be fully explained by neural activity, and market behaviour will be reduced to the actions of individual agents. If you follow this view to its conclusion, all sciences will eventually point back to the same fundamental principles — the laws of physics (or, perhaps, mathematical laws).

Non-reductionism, on the other hand, takes a very different view. It argues that some phenomena simply can’t be reduced to the fundamental. Memories, markets, and other complex systems may involve principles or patterns that emerge at higher levels of abstraction, which physics alone could never capture.

This tension between reductionists and non-reductionists isn’t just an abstract debate. It has major implications for how we do science and how we understand reality itself.

To be scientifically meaningful, a concept needs to support valid generalisations. In other words, a scientific concept should allow us to identify principles or laws that apply across many instances and help us make reliable predictions.

Take gravity, for example. Gravity isn’t just a concept tied to a single event — it applies universally. Whether we’re studying the fall of an apple, the orbit of a planet, or the motion of an entire galaxy, the concept of gravity gives us a framework for understanding all these phenomena. And it lets us predict how objects will behave in different contexts — the law of gravity is a law that makes generalisable predictions.

The key question here is whether the special sciences — like psychology or economics — discover concepts that meet this standard. Do they uncover valid scientific principles and laws that make generalisable predictions that physics alone could never fully describe? (this would be the non-reductive view). Or are the concepts in the special sciences merely convenient shortcuts, tools we use to get a handle on extremely complex systems — but ultimately, these concepts can be explained by physics? (the reductive view).

This might all sound a bit abstract. So, let’s look at two examples to try to make this question more concrete. We’ll look at money and the gemstone jade. The money example comes from philosopher Jerry Fodor (a champion of the non-reductive view), and the jade example is from philosopher Jaegwon Kim.

Money

Economists like to study money. They like to uncover principles and patterns that make predictions about the effect money has on the economy.

For example, one well-known principle is the quantity theory of money. This principle states that when there is more money circulating in an economy, prices tend to rise. Crucially, this principle holds true no matter what form money takes — whether the money is gold coins, paper bills, digital bits, or even seashells.

Fodor points to this as a perfect example of why some concepts in the special sciences are irreducible.

If we tried to define money in purely physical terms, we’d end up with what philosophers call a heterogeneous disjunction: ‘money is gold coins OR paper bills OR digital bits OR seashells OR...’

Fodor isn’t claiming that disjunctions can’t be scientific. Rather, he’s claiming that a list of things that vary widely in their physical makeup couldn’t possibly form a natural kind of thing — the type of thing we might study in physics.

Despite the fact that money is (in physical terms) a disjunction, economics can formulate meaningful laws about money. These principles and laws generalise across money’s different physical forms. Economics is able to do this because it recognises money as something that is more than its physical forms. Money, it seems, is a scientifically meaningful term even though to reduce it to its physical forms results in a disjunction.

What About Jade?

What we call jade is actually two entirely different minerals: jadeite and nephrite. Jadeite and nephrite look similar, but they have very different physical properties.

In other words, jadeite and nephrite form a disjunction. Observations about jadeite do not generalise to nephrite and vice versa. For example, if a geologist studied 1,000 pieces of jade and found they were all green but later discovered all her samples were actually jadeite, she couldn’t say anything about the colour of nephrite.

We can’t form meaningful principles and laws about jade that generalise across jadeite and nephrite.

Let’s return to our original question: Can the special sciences identify principles or laws that generalise and make reliable predictions?

It seems that economics can identify such principles for money, but geology can’t do the same for jade.

Why not?

The reason has a lot to do with multiple realisability. You may recall that something is multiply realisable if it can exist in many different forms while still fulfilling the same function in each form.

For example, a toaster is multiply realisable. Toasters can come in many shapes and sizes, but all toasters serve the same function: to toast.

Money is also multiply realisable. Gold coins, paper bills, digital bits, and seashells can all serve the same function: as a medium of exchange in an economy.

Jade, on the other hand, is not multiply realisable. While jadeite and nephrite can both function as gemstones, we don’t define jade by its function. Instead, we define it by what it is made of.

Things that are multiply realisable — toasters and money — are defined by what it does. Things that are not multiply realisable — like jade — are often defined by what they are made of.

A short side note before we move on: While I’ve presented Kim’s jade example as a contrast to Fodor’s money example, it’s worth noting that Kim uses the jade analogy to argue against Fodor. He questions whether multiply realised kinds — like toasters and money — can be proper scientific kinds at all. But let’s not get into that debate yet.

Q3: What does this mean for consciousness?

In the introduction to this essay, I hinted that the puzzle of the special sciences might have implications for understanding consciousness.

Functionalism — the idea that consciousness is defined by what it does, not by what it is made of — is currently the most popular theory of consciousness. It suggests that consciousness is more like money than jade. Consciousness is multiply realisable.

Just as money can exist as gold coins, paper bills, or digital bits, consciousness, according to functionalism, could theoretically be realised in human brains, alien biology, or advanced AI systems. The whole idea of functionalism hinges on whether consciousness is, in fact, multiply realisable.

How confident can we be in this claim? What evidence supports the idea that consciousness is independent of its physical composition?

When it comes to consciousness, does the matter really not matter?

To try to get at this, let’s return to Kim’s jade example.

Remember the geologist who collected 1,000 pieces of jade? She found that every single piece was green. From this, the geologist might declare, “all jade is green.” But let’s suppose that later we discover that the geologist only collected pieces of jadeite . None of the samples were nephrite. Nephrite, it turns out, is not always green. I can come in other colours, too, like white or grey.

What can we conclude?

At best, we might confidently assert that all jadeite is green. But we have no basis to conclude that all jade is green, since the geologist didn’t examine any nephrite. Her initial claim that “all jade is green.” was wrong. She overgeneralised from incomplete data.

This brings us back to functionalism’s claim about consciousness. When functionalists argue that consciousness is multiply realisable, could they be making a similar mistake?

Our understanding of consciousness comes almost entirely from studying biological systems — primarily humans and, to a lesser extent, other animals. If we only know of biological consciousness, can we confidently generalise to non-biological systems, like AI or silicon-based lifeforms? Or are we overgeneralising from incomplete data?

Many believe we are inching closer to the development of artificial general intelligence (AGI). With this possibility, we might want to ask: Is consciousness more like money — are we inching toward something functionally equivalent to human consciousness? Or is consciousness like jade — are we inching closer to something that might merely imitate the outward appearance of consciousness but actually be something entirely different?

The Sum Up

I’ve presented a dichotomy here: consciousness is either like money or like jade. But, of course, it’s possible that consciousness is like neither. Consciousness has a way of resisting easy comparisons, which is part of what makes it such a fascinating puzzle.

The idea of the special sciences raises all kinds of big questions that I didn’t get time to get into in this essay. For example: What does it really mean for something to be an emergent property? How connected are our conscious experiences to the physical stuff we’re made of? Is there something special about biology that gives rise to consciousness? And can we ever truly bridge the gap between physics and the messy, complex world of the mind? These are all things I’m looking forward to exploring in future essays.

Next Week...

Earlier, I mentioned that functionalism is the popular kid right now. That wasn’t always the case. Theories of consciousness go in and out of style, just like most other things. Next week, let’s take a brief journey through history to explore how we got here. Why is functionalism so popular today? And where might the tide be turning next?

Have you ever played catch? The ball flies toward you — fast — and you snatch it out of the air without a second thought. Catch seems like a simple, almost effortless game. A way to relax, even. But there’s something strange about this seemingly easy pastime.

It’s common to think that we perceive the world like a camera — our senses passively receive signals from out there and relay them to our brains.

This common view (which philosopher John Locke also held) suggests that our perceptual experience is built directly from incoming sensory signals. We feel like we’re simply seeing what’s out there.

But if this is true — if our conscious experience is simply a simulation of the world built from the signals coming from our senses — then we have a problem.

Let’s say you’re playing catch with your nephew. He throws you the ball. Photoreceptors in your eyes register the movement and send those signals to your brain.

But that takes time.

At least 100 milliseconds.

Even longer if you count the time it takes to process those signals into a coherent perception.

The problem is that the brain is too slow.

You’ve probably heard about this problem before — you may have been told that because it takes time for sensory information to be processed by the brain, we’re always living in the past. But if that were true, in the way it’s often described, it would create some strange effects.

To state the obvious, if we want to catch a ball, we must track the movement of that ball. If our experience were simply a delayed simulation of what happened 100 ms or more ago, by the time we consciously experienced the ball’s location, that ball would already be somewhere else.

If we had to wait for our brain to fully process the visual signals before we act, we’d always be moving too late — we’d move our hands to where we perceive the ball is rather than where it actually is. And, as a consequence, we’d probably get hit in the face — a lot.

But, typically, we don’t get hit in the face when playing catch. Somehow, we can catch balls in real-time. This suggests that our brain is doing something else — perhaps something more sophisticated than simply trying to simulate the world.

This has led to a radical alternative theory: conscious experience is not a simulation of what is happening now — it’s a prediction of what is happening now.

This is an extraordinary claim. And extraordinary claims require extraordinary evidence.

It’s early days. But yes, this is an active area of investigation, and yes, there is evidence to back it up.

Here’s what we know so far.

When we track a moving object (like a ball in midair) that follows a predictable path, our visual system responds as if it’s seeing it in real-time.

Think about that for a second.

Before the sensory signals even arrive, the brain acts as if those signals already have.

But this only works for movements that follow familiar patterns — like a ball travelling in a straight line at a steady speed. In these cases, the brain doesn’t simply react to incoming signals; it anticipates them, accurately predicting what it expects to see next.

That’s why practice helps us catch a ball. Over time, we learn how objects typically move, picking up on the basic rules of motion.

But if the ball suddenly zigzags or comes at us from an unexpected angle, the brain’s prediction fails. It expected something else. And when the brain’s guess is wrong, we’re caught off guard — and much more likely to miss the catch.

The idea that our conscious experience is a prediction rather than a real-time simulation raises some pretty wild questions about consciousness.

What, exactly, are we experiencing?

If consciousness isn’t a direct simulation of the present but a prediction of what’s about to happen, then…

When exactly does conscious experience happen?

Perception isn’t passive; it’s an active construction. We don’t simply see the world as it is — we see what the brain expects to be there. Most of the time, these expectations align with reality well enough that we don’t notice. But when they don’t — when a ball moves unpredictably or an illusion deceives us — we catch a rare glimpse of how much our experience is built on inference rather than direct awareness.

So, if consciousness is a prediction, perhaps we aren’t living in the past. Perhaps we’re living in the brain’s best guess of now.

And if that’s true — if our experience is always a step ahead — then when, exactly, are we?

Leave a comment

If you’re interested in exploring this research further, I recommend these papers by my friend and former colleague, Hinze Hogendoorn:

• Johnson, P. A., Blom, T., van Gaal, S., Feuerriegel, D., Bode, S., & Hogendoorn, H. (2023). Position representations of moving objects align with real-time position in the early visual response. eLife, 12, Article e82424. https://doi.org/10.7554/eLife.82424

• Blom, T., Feuerriegel, D., Johnson, P., Bode, S., & Hogendoorn, H. (2020). Predictions drive neural representations of visual events ahead of incoming sensory information. Proceedings of the National Academy of Sciences, 117(13), 7510-7515. https://doi.org/10.1073/pnas.1917777117

Next Week…

Over the past six weeks, we’ve explored six big questions in six mini-essays. These were just introductions — throughout 2025, these questions will keep coming up in the topics I want to explore.

The six questions were:

1. What is information, and where does it come from?

2. How do simple, deterministic rules lead to unpredictable complexity?

3. As we try to understand complex phenomena — like how brains work, how to build intelligent systems, or even how consciousness happens — what details can we safely ignore? And what essential insights might be lost when we do?

4. Can general intelligence be built, or does it need to grow?

5. What (the heck) is physical?

6. When does conscious experience happen?

I’ll be taking a short break next week, but I’ll be back on February 25th with the first essay in our first series.

Thank you so much for your support — I appreciate this little corner of the internet more than I can put into words. And all of you make that possible. So, thank you!

I’m really excited about the year ahead, and I can’t wait to explore these ideas and learn from all of you.

Subscribe now

We tend to think of physical things as things we can touch and measure. Atoms and molecules are physical. Forces and energy are physical. A rock is physical. Light is physical. Even empty space is physical.

But what about information? Can information be considered physical?

Physicist Rolf Landauer thought so. He even wrote a famous paper titled Information is Physical arguing that information cannot exist without some physical embodiment.

Whether encoded as marks on a stone tablet, holes in a punched card, or the spin of a quantum particle, information is always tied to a physical process. And crucially, changing or erasing information isn’t just a conceptual shift — it has real physical consequences, consuming energy and increasing entropy.

The idea that information has physical consequences has pretty big implications. It means that information and energy are fundamentally linked — it takes energy to create, change, move, or erase information. (Whether information can truly be erased is a fascinating question we’ll explore another time.)

But the idea that information is physical raises a bit of a puzzle. If information is physical, what exactly is it made of?

Consider the number 5. Not the symbol 5 displayed on your screen, but the abstract concept of five-ness. Can it exist independently of a physical system that encodes or represents it? Neural activity, ink marks, and digital states all seem to carry the concept, but do they constitute it? Or is the number 5 something that exists beyond physical representation?

After all, these physical patterns seem to merely represent the number 5 rather than be the number itself. But if that’s the case, then what kind of existence does 5 have? Could something as abstract as a mathematical concept be physical?

Questions like these echo debates that go back centuries.

In the 17th century, René Descartes defined physical things in terms of spatial extension — anything physical, he argued, must take up space. For Descartes, the primary characteristic of physical matter was that it had length, breadth, and depth.

Descartes’s definition works well for objects like rocks or bodies but not so much for the number 5.

What about other things we like to label physical?

These days, modern physics has moved far beyond Descartes’ simple definition. Now, what we call physical encompasses things that don’t fit neatly into his idea of extension.

For example, general relativity describes gravity as a curvature in spacetime. We can measure its effects with precision, yet we don’t think of gravity as having length, breadth, and depth in the way physical objects do.

Nowdays, we define what is physical by referring to scientific theory — what is physical is whatever our best physics theories explain. But this raises troubling questions. Which theories should we trust? Our current physical theories are incomplete, and future theories are unknown.

For all we know, future theories may radically change our understanding of the physical.

This brings me to the 5th question on my list of questions for 2025:

What (the heck) is physical?

One possible answer comes from theoretical physicist John Wheeler. He offered a radical idea: it from bit. In his view, physical reality (it) arises from information (bit). The universe, he argued, isn’t fundamentally made of extended things like particles—or even physical entities like energy. Instead, reality emerges from information rather than the other way around.

We usually assume that information is something that comes from the physical world. What if, instead of information emerging from the physical, the physical emerges from information?

Some might argue that if Wheeler is right, then our confusion about defining the physical isn’t a mistake — it’s a clue.

This is indeed a radical idea! But sometimes (even if ultimately we disagree), radical ideas are worth trying to understand.

Leave a comment

Subscribe now

When we build something — let’s say a car — there is a time when we build the car and a time when we use the car. This is how engineering traditionally works: first, develop something, then use it. We see this process of building (development) and then use (function) everywhere in the human-made world.

It’s even the process used to build our current artificial intelligence systems like ChatGPT. These models are first built (trained) and then deployed for functional use, with a clear separation between the development and use phases. Like a car rolling off the production line, once training is complete, the model is considered built and ready for use.

For decades, neuroscientists thought this was how the brain worked, too. We viewed brain development and brain function as separate processes. Development was seen as the period when neurons grow, form connections, and establish basic circuits — essentially building the brain’s hardware. What the brain does (its functions), on the other hand, was considered to be what happens after development is complete. You’ve probably heard the idea that what the brain does is like software that runs on the hardware of the brain.

This clean separation between the development of the brain and the function of the brain has been highly influential in how we think biological neural networks work.

But this view is not accurate. In biology, the distinction between development and function is not so tidy.

The same processes that help build our brains during childhood — things like strengthening some connections while pruning away others — keep working throughout our lives. Even as adults, our brains are constantly being remodelled. New connections form while others are eliminated, all based on our experiences.

This blurry line between development and function has got me thinking about the differences between biological intelligence and artificial intelligence.

When we learn something new — like riding a bike or solving math problems — our brains physically change. New connections form between neurons, while others are eliminated or weakened. These physical changes are how our intelligence develops, so the development of intelligence and the changes in our brains are not easily separated. This continuous physical reorganisation allows our brains to flexibly learn and adapt to new situations while maintaining existing capabilities — what we call general intelligence.

So what about artificial general intelligence (AGI)?

Current artificial neural networks, like ChatGPT, excel at specific tasks like image recognition or language processing, but they don’t exhibit the kind of general intelligence we see in biological brains — the ability to flexibly learn and adapt to new situations while maintaining previous capabilities. So, we don’t yet consider our LLMs to have AGI.

Today’s AI systems are built like cars — first, we build them, then we use them. But what if that’s not how general intelligence is built? What if general intelligence requires constant development and change?

Let’s say our goal was to build AGI (which might not be the best idea, but this is just a thought experiment). Would we need to rethink how we build that AI system? Would we need to move away from the development of static architectures toward networks that develop — continually growing and self-organising?

Would artificial neural networks achieve higher forms of intelligence if they underwent a developmental growth process similar to biological brains?

So, the fourth question on my list of questions for 2025 is:

Can general intelligence be built, or does it need to build itself?

Leave a comment

This question raises many others: What do we mean by intelligence? Are there different types of intelligence, each with unique characteristics and developmental needs? When we talk about general intelligence, are we referring to a single thing, or could there be many forms of general intelligence we’ve yet to imagine?

We often think of AGI as simply human intelligence — but faster. But human intelligence might be just one of many possible forms. Perhaps we will need to broaden our understanding of what intelligence could be.