April 2021 to now · 228 posts · 541,000 words · 10,763 comments · 85 posts paywalled · still publishing (latest: August 2026) · this page is based on 12 free posts read in full, spanning April 2021 to April 2026
In the 1980s a group at the University of Chicago built a rig: a pool of water, a treadmill platform floating above it, and rats wired to an EEG. The rats were allowed to sleep. Deep sleep, the platform stayed still. But the moment the brainwaves showed REM — the stage where dreaming happens — the treadmill started, and the rat had to wake up or fall in.
Not sleep deprivation. Dream deprivation, isolated.
After a few days, the rats began to lose weight. After a few weeks, they began to jaundice. Even their fur turned yellow, then their eyes. Their yellowing paws developed lesions. And after a few weeks, every rat dropped dead from not dreaming. "Exit the Supersensorium," 2022
Most people read that and conclude: dreaming matters. Erik Hoel read it and asked a different question — what is dreaming actually doing, that it matters this much?
The standard answer is memory consolidation: the day's material gets replayed at night and filed into what you already know. Hoel thinks that answer doesn't survive contact with the evidence, and his reason is very concrete: dreams almost never contain specific memories. Yesterday's meeting, yesterday's walk, yesterday's argument — they essentially never replay as themselves. Dreams are also sparse: you almost never read a book or make out a phone screen in a dream, because a dream doesn't have the resolution for small text. And they are hallucinatory: people who are two people, a place that is both your home and a spaceship.
A theory of dreams must explain why every human, even the most unimaginative accountant, has within them a surrealist author scribbling away at night. "Exit the Supersensorium," 2022
His answer comes from a failure mode every machine learning engineer knows: overfitting. A network trained too well on its data starts treating that data's accidents as the world's laws, and falls apart on anything new. So what is the brain's training set? The days you have actually lived. Which are highly repetitive, highly correlated — and the brain cannot switch learning off.
Life alone is not a sufficient training set for the brain, and relying solely on it likely leads to overfitting. "Exit the Supersensorium," 2022
Deep learning has two standard defenses against overfitting. One is dropout — randomly discarding parts of the data or the network so it can't memorize; that is the sparseness of dreams. The other is domain randomization — warping and corrupting the inputs into impossible shapes; that is the hallucinatory quality of dreams. He published this in Patterns in 2021 as the Overfitted Brain Hypothesis. Dreams aren't tidying up your day. They're loosening it, so your day can't make you stupid.
Then he did the thing that turned a neuroscience paper into an entire worldview: he pushed it out two levels.
Level one is art. If a brain needs to dream nightly to avoid overfitting, what are novels, films, television? Portable, outsourced dreams. He even notices the sizes line up: a TV episode runs about thirty minutes, close to a single REM period; a film runs about ninety, close to a full sleep cycle. Once humans mastered fire we outsourced digestion to cooks. Once we learned to tell stories we outsourced dreaming to novelists.
There's a consequence here he refuses to dodge: it means the distinction between art and entertainment isn't snobbery — it has a mechanism behind it.
For entertainment is Lamarckian in its representation of the world—it produces copies of copies of copies, until the image blurs. "Exit the Supersensorium," 2022
A blurred dream can still satisfy the appetite for dreaming while no longer performing its function — the way candy is still calories but no longer food. So his prescription isn't abstinence: "In a world of infinite experience, it is the aesthete who is safest, not the ascetic. Abstinence will not work. The only cure for too much fiction is good fiction."
Level two takes the same mechanism and points it at an entire civilization. The brain can't stop learning, so it overfits. Culture can't stop learning either.
So what if we thought of culture as a big "macroscale" global brain? Couldn't that thing become overfitted too? "Our Overfitted Century," 2025
The top ten grossing films of 2024 were all sequels. Book covers converge. Faces converge. Cars converge. The usual explanations are capital, monopoly, copyright, algorithms. Hoel says those are all describing different levels of one thing, and that thing has a machine learning name: once a system becomes efficient enough, with tight enough feedback and a sensitive enough discriminator, it overfits — and overfitted systems are unpleasant to live inside. He half-jokes that a lot of complaints about "late-stage capitalism" decompose into exactly that sentence.
And what worries him isn't that the movies are bad: "An inability to create novelty is a sign of an inability to generalize to new situations. That seems potentially very dangerous to me."
Who writes this? Erik Hoel, born 1987, raised in his mother's independent bookstore, where he worked as a teenager selling fiction. PhD in neuroscience under Giulio Tononi, originator of Integrated Information Theory. A novel, The Revelations (2021); a nonfiction book, The World Behind the World (2023). In November 2022 he resigned his professorship at Tufts, on the grounds that everything he was proudest of had been done in "time squirreled away from bosses and administrators. And I'm tired of being a squirrel." He now writes for a living, and recently founded a research institute aimed at solving consciousness.
His blog is called The Intrinsic Perspective — the view from inside. The name is the whole program. Because science got started precisely by cutting the view from inside out of it: in The Assayer, Galileo explicitly moved color, smell and pain outside the jurisdiction of physics, which is what finally told science what to look at. Then twentieth-century behaviorism made "consciousness" an unsayable word in journals for decades. The cost is that the scientific map is still missing one thing — in Thomas Nagel's image, which Hoel loves: the map can be complete, coherent, airtight, and it will still have no YOU ARE HERE sign on it.
How big is the hole? He went and counted the 103,280 grants the NIH awarded between 2007 and 2017, and asked how many directly studied the contents of consciousness.
Five.
His other core claims all circle that hole. Causal emergence says a system's macroscale can have stronger causal relationships than its microscale — so "you" is not decorative froth sitting on top of atoms; it may be the scale at which causation is sharpest. "Consciousness as a Gödel sentence" says that science's inability to reach consciousness may not be a fact about consciousness at all, but a fact about any self-referential formal system: some true things can't be proved from inside.
Both point the same way: the view from inside is real, and science in its current shape has no room for it. Which is also where a reader should keep one hand on their wallet — a man raised in a bookstore, who writes novels, and who wants "what it is like from the inside" to matter, has all of his scientific commitments tilting in exactly that direction. That doesn't make him wrong. But a tilt that consistent is itself information.
The harshest verdict on that tilt was written by him. In 2022 he published a long essay explaining why humans were stuck in prehistory for ninety thousand years, and then, in the last section, turned and pointed at himself:
An admittance. For it should be obvious by now: this text is corrupted. "The Gossip Trap," 2022
For the full derivation of the Overfitted Brain Hypothesis, how causal emergence is actually computed on a light switch (and the fight with Scott Aaronson over what counts as a fair comparison between scales), the paradox he builds out of a zombie Descartes, how the gossip trap explains the origin of civilization, and the strongest objection to his own theory — written by him — switch to "Full" above.
Experimenters sometimes need a streak of cruelty. In the 1980s a group at the University of Chicago built this: a pool of water, a treadmill platform above it, a rat wired to an EEG, and the EEG wired to an automated program. The program distinguished two things — is the rat in deep non-REM sleep, or has it entered REM? Non-REM, the platform stays still; let it sleep. The instant the brainwaves show dreaming, the treadmill starts, and the rat has to get up and walk or fall in the water.
This is not sleep deprivation. This is taking an animal out of its dreams, over and over, to see what happens.
After a few days, the rats began to lose weight. After a few weeks, they began to jaundice. Even their fur turned yellow, then their eyes. Their yellowing paws developed lesions. And after a few weeks, every rat dropped dead from not dreaming. "Exit the Supersensorium," 2022
The normal reaction is: fine, dreaming matters. Erik Hoel's reaction is to ask what it's doing that could possibly be worth dying over.
This is the default move behind everything he writes: not "what does this show," but "what is this thing's function, and why aren't the existing explanations good enough?" And the reason he felt entitled to attack a problem outside his own specialty is that during his PhD he happened to share a lab with some of the best sleep researchers in the world, and so he knew a thing outsiders don't: nobody actually knows why we dream. The null hypothesis that dreaming does nothing at all is still taken seriously by serious scientists.
Erik Hoel, born 1987, lives near Boston. He grew up in an independent bookstore his mother ran, and worked there as a teenager — in his own phrase, "hawking fictions." He wanted to be a writer first; in college he discovered he also had an aptitude for science, and went to the University of Wisconsin–Madison for a neuroscience PhD under Giulio Tononi, the originator of Integrated Information Theory. He helped develop parts of that theory, was a visiting scholar at the Institute for Advanced Study in Princeton, and made Forbes 30 Under 30 in science.
Then he refused to drop either half. In 2021 he published a novel, The Revelations — set at NYU, where one of eight young consciousness researchers is murdered. In 2023 he published The World Behind the World: Consciousness, Free Will, and the Limits of Science.
In November 2022 he resigned his professorship at Tufts and went to writing for a living. The resignation post is blunt about why: the work he's proudest of was done in the gaps between administrative duties.
I did it in time squirreled away from bosses and administrators. And I'm tired of being a squirrel. "Goodbye academia, hello Substack," 2022
The target he set himself was an essay a week for twenty years — a thousand essays. The blog is still running; the latest post is from August 2026. Of its 228 posts, 85 are paywalled, and this page is based only on the free ones. In 2026 he founded a research institute, Bicameral Labs, whose stated purpose is to solve consciousness or die trying.
The blog's name is the answer: The Intrinsic Perspective.
To see how unfriendly that name is, you need to know how science got going. The philosopher Philip Goff makes a point Hoel returns to repeatedly: in The Assayer (1623), Galileo explicitly moved color, smell, taste and pain outside the jurisdiction of physics — they are not in the object, they are in the observer. That single cut is what finally told science what to look at and what to ignore, after which it took off. Goff's line is that if Galileo came back today he'd be the first to say science is obviously incomplete — he never designed it to talk about souls.
Three hundred years later came a second cut. Under behaviorism and logical positivism, "consciousness" became a dirty word in science, close to unpublishable — Hoel calls this the consciousness winter, and it coincided exactly with the decades in which every other science sprinted ahead. The winter only ended in the 1990s, when Nobel laureates like Francis Crick and Gerald Edelman spent their own reputational capital making the word sayable again. The two international conferences devoted to consciousness date from the same period.
Which means modern science has been seriously working on consciousness for about thirty years. One generation.
With what money? He went and looked at the 103,280 grants the NIH awarded between 2007 and 2017, and counted how many directly studied the contents of consciousness.
Five.
For comparison, humanity spends roughly a billion dollars a year on CERN. And when the funding is zero, there is exactly one thing a consciousness researcher can do for free: propose their own theory. So for thirty years the field's dominant move has been to found your own school. Hoel's assessment of that situation is one of his meanest sentences:
The problem is that, if any flower is at all true or promising, you can't identify it, as its sweet subjectivity-solving scent is completely masked by the bunches of corpse flowers around it. We have too many flowers, and one more just isn't meaningful anymore. "We consciousness researchers have failed you," 2026
What does the hole look like? His favorite image is Thomas Nagel's: imagine the scientific worldview as a map. It can be complete, coherent, as airtight as an axiom system. And it will still be missing one piece of information — the YOU ARE HERE sign. The fact that relates the observer to the map is nowhere on the map.
Everything Hoel has done for twenty years — the mathematics, the neuroscience, the cultural criticism — is one project in different clothes: put the sign back. Show that the view from inside isn't decorative froth on top of the physical story; that it has causal power (causal emergence), that it has a function (the Overfitted Brain Hypothesis), and that science's current inability to reach it is a fact about science rather than about it (the Gödel sentence).
Back to the rats. If not dreaming kills you, what's the leading theory of what dreaming is for? Memory consolidation. The day's learning gets replayed at night and integrated into existing structure.
Hoel says this runs into a problem it can't get around: dreams almost never contain specific memories. Yesterday's meeting, yesterday's walk, yesterday's sentence — they essentially never replay as themselves. If consolidation theory is right, then the content of dreams is a byproduct — hallucinatory froth thrown off by some "integrating" and "consolidating" process (two words nobody has ever quite cashed out).
And he insists on something surprisingly rare in this field: a theory of dreams has to explain what dreams are like. Dreams have two very stable, very strange properties.
First, dreams are sparse — less vivid and less detailed than waking life. The best example: you essentially never read a book or make out a phone screen in a dream, because the dreamworld doesn't have the resolution to render small text.
Second, dreams are hallucinatory — full of unlikely events and borderline categories. People who are two people. A place that is both your home and a spaceship.
A theory of dreams must explain why every human, even the most unimaginative accountant, has within them a surrealist author scribbling away at night. "Exit the Supersensorium," 2022
His explanation comes from the most familiar failure mode in machine learning. A network's training set is always finite and often correlated in non-obvious ways; trained too hard, it takes that set's accidental structure for the world's structure and collapses on new data. That's overfitting.
Then he asks the question that works as a crowbar: what is the brain's training set?
Real brains, in turn, rely on the training set of lived life. However, that set is limited in many ways, highly correlated in many ways. Life alone is not a sufficient training set for the brain, and relying solely on it likely leads to overfitting. "Exit the Supersensorium," 2022
And a mammalian brain can't turn learning off. It's too plastic; you live repetitive days and it contracts, day by day, toward their shape.
Then comes the prettiest step in the essay. Deep learning has a standard toolkit against overfitting, and Hoel lines it up against the two strange properties of dreams:
The conclusion: dreams aren't a replay of the day, they're an antidote to it. And — this is the counterintuitive part — the less a dream resembles your day, the better it is doing its job.
Dreams warp and play with always-ossifying cognitive and perceptual categories, stress-testing and refining. The inner fabulist shakes up the categories of the plastic brain. "Exit the Supersensorium," 2022
He published this in Patterns in 2021 as the Overfitted Brain Hypothesis. It picks up several loose facts on the way: human sleep research associates sleep less with raw memorization than with increased abstraction and generalization; there's the famous link between dreams and creativity; and if you stay awake too long you start to hallucinate — perhaps because your perceptual processes are overfitting.
If he'd stopped there it would be a neuroscience paper. What Hoel does is push it out two levels, and those two levels are his real territory.
Level one: if people need to dream nightly to avoid overfitting, what are we doing with all those hours spent on things that never happened? From an evolutionary standpoint that's already odd. He quotes Denis Dutton's thought experiment: imagine humans who loved only true stories — then there would be no philosophical "problem of fiction" at all, because we would react to made-up stories the way we react to a dull knife, or the smell of rotting meat.
His answer: novels, films and series are outsourced dreams — and better-structured ones.
Since society specializes for efficiency and competency, we began to outsource the labor of the internal fabulist to an external one. […] For what is a chef? Our mastery of fire allows us to do most of our digestion outside of our bodies […] The same for artists, but they allow you to dream without sleep. "Exit the Supersensorium," 2022
He also throws in a coincidence of sizes as corroboration: a traditional TV episode runs about thirty minutes, close to a single REM period; a film runs about ninety, close to a whole sleep cycle.
But the theory hands him a consequence he has to carry himself. If dreams have a function, then dreams that fail the function are bad dreams. Which means the ranking of art above entertainment isn't snobbery — it has a mechanism. And in the present climate, saying so out loud is close to taboo; he notes that arguing art is fundamentally different from entertainment now brings charges of classism, snobbishness and elitism. He argues it anyway:
For entertainment is Lamarckian in its representation of the world—it produces copies of copies of copies, until the image blurs. The artificial dreams we crave to prevent overfitting become themselves overfitted, self-similar, too stereotyped and wooden to accomplish their purpose. "Exit the Supersensorium," 2022
A blurred dream still satisfies the appetite, the way candy is still calories. And what marks the good stuff, in his account, is concrete: the works we consider artful "return to the well of the world" — they have a shocking realness. He quotes Karl Ove Knausgaard in The New Yorker: "The duty of literature is to fight fiction."
So the prescription is not abstinence but taste:
In a world of infinite experience, it is the aesthete who is safest, not the ascetic. Abstinence will not work. The only cure for too much fiction is good fiction. "Exit the Supersensorium," 2022
Level two comes later and reaches further — "Our Overfitted Century," late 2025. He starts by laying out, fairly, the existing explanations for cultural stagnation: copyright terms too long (Chris Dalla Riva); corporate consolidation and financialization (Andrew deWaard); the internet flattening everything (David Marx); the disappearance of deviance (Adam Mastroianni — today's teenagers smoke less, fight less, take fewer drugs, and buckle their seatbelts more). He says most of these are probably right.
Then he offers an explanation at a different level, starting from the same sentence: the brain can't stop learning, so it overfits. And culture can't stop learning either.
So what if we thought of culture as a big "macroscale" global brain? Couldn't that thing become overfitted too? In which case, culture might lose the ability to generalize, and one main sign of this would be an overall lack of creativity and decreasing variance in outputs. "Our Overfitted Century," 2025
The strength of this is that the mechanism can be matched up item by item. Generative adversarial networks have a failure called mode collapse: the discriminator gets too sensitive and judgmental, so the generator, to satisfy it, produces increasingly similar and boring output. Superhero franchises, says Hoel, are exactly that — the discriminator is the whole apparatus of box-office and financial signal, and it has flattened the dimensions out. "Instagram face" is the same: cosmetic surgery got cheap and efficient while the discriminator of likes and swipes got hypersensitive, and squeezed between them you get The Face. Car design is the same: wind tunnels, mergers, efficiency, all pushing toward one shape.
Which gets him a line you can take away and use:
I can't help but wonder if some of the complaints about "late-stage capitalism" actually break down into something like "this system has gotten oppressively efficient and therefore overfitted, and overfitted systems suck to live in." "Our Overfitted Century," 2025
Worth pausing on what he does not do, which is pin it on a villain. He states outright that you can't blame AI — ChatGPT just had its third birthday, and the trend long predates it; AI is accelerating a process already underway. He also deliberately steps around the political fight, on the grounds that a machine-like age deserves a machine-like explanation.
And the thing he actually fears isn't bad movies:
An inability to create novelty is a sign of an inability to generalize to new situations. That seems potentially very dangerous to me. "Our Overfitted Century," 2025
Incidentally, this essay was supposed to be a book. He pitched it a year and a half earlier as Culture Collapse; two big publishing houses passed and he and his agent gave up. He puts that directly in the piece, including the rueful aside about how cool it would be if Culture Collapse were coming out soon. It's a signature move: putting the failure of an idea on the page as content.
There's a smaller, funnier essay that's another door into the same thread. In 2024 he threw his son a Curious George–themed third birthday party and ordered a sticker set from Amazon for the party favors. The night before, his wife called him over to look — the set contained a batch of things clearly scraped or generated by an algorithm, including one he describes as "Bi-curious George and the closet," and one of George holding an assault rifle. First he laughed until he couldn't (imagine handing those out), and then he started asking: how does a thing like this come to exist?
The proximate cause is clear: a Chinese sticker company scraping, tiling and printing automatically. But he wanted the ultimate cause, and his answer isn't "capitalism" or "offshoring":
The growing high strangeness I sense is that culture is draining of human consciousness, and therefore of sense itself. "Curious George and the case of the unconscious culture," 2024
He runs the last three hundred years as a fast-forward montage: individual artisans in their houses, fingers flowing; then an assembly line with many hands, young and old, some missing fingers; then bigger and safer machines and fewer hands; then no hands at all, just a whirring robotic factory. And even then one light is still on — the overseer's office. Then one day there's no overseer. It all takes place in the dark.
That passage is the wire connecting "overfitted culture" back to consciousness: if a civilization replaces human consciousness at every decision point with an artificial learning system that is more sensitive but less robust, of course it overfits. And nobody is positioned, top-down, to stop it.
The ending is worth copying out too. The running joke of the Curious George books, he notes, is that George is disconnected from the world around him: he can't speak, has no language, can't learn from his mistakes, and so the modern world is permanently an exercise in high strangeness to him —
And now, I guess, it is for us too. We've all become Curious George, primates living amidst alien machinery far from our homelands. "Curious George and the case of the unconscious culture," 2024
One phrase has been quietly load-bearing in all of that cultural writing: "macroscale." Thinking of culture as a macroscale brain — on what authority? That isn't a rhetorical flourish. It's his dissertation.
The standard reductionist intuition runs: real causation happens at the bottom. Particles push particles; everything above is a convenient shorthand. When you say "flipping the switch caused the light," that's a coarse approximation of a precise story told in physics.
Hoel's claim is that this intuition can be overturned by calculation.
What's causal emergence? It's when the higher scale of a system has stronger causal relationships than its underlying lower scale. […] In something approaching plain English: macrostates can be strongly dependent even while their underlying microstates are only weakly dependent. "A primer on causal emergence," 2021
How do you compute it? With Judea Pearl's do-operator and information theory. Run a full A/B test on a system — randomize the input to maximum entropy, then see how much of that shows up in the output — and the quantity you get is effective information (EI). High EI means every intervention produced a distinct consequence. If every intervention leads to the same result (it doesn't matter who flips it), or to random results (flipping it is like not flipping it), EI is zero.
His example is almost comically plain: a switch and a bulb. If flipping the switch always turns the light on, the system has 1 bit of EI. If the bulb flickers off 20% of the time, EI drops to 0.61 bits. Swap in a 256-position dial driving 256 levels of brightness, each mapping cleanly, and EI is 8 bits — the causal structure is denser.
The crucial step: EI can be measured separately at different scales and compared. And in some systems the macroscale — group microstates together, or leave certain elements outside the model — measures higher EI than the microscale.
Why is that possible? He gives two answers. One is mechanical: microscale mechanisms compose into different macroscale mechanisms — a pile of NOR gates can form a COPY gate — and the conditional probabilities of state transitions differ across scales, so determinism can rise and degeneracy can fall at the higher scale. The other is information-theoretic, and prettier: higher-scale relationships can carry more information because they are performing error correction. Shannon showed you can raise a channel's throughput by changing the input distribution; intervening at different scales is like trying different inputs into the same channel.
He also has a line I think is the most important in the piece, because it blocks the obvious "so it's all relative then" objection:
In this sense, the theory says that causal interpretations are not relative or arbitrary, but ordered, and there is a best one. "A primer on causal emergence," 2021
In other words: not "any scale you like," but "one scale is best, and it isn't necessarily the bottom one." His image is a camera — you turn the focus ring (change scales), and at some point the system's causal structure snaps into focus.
The theory's most public fight was with the quantum computing theorist Scott Aaronson. Aaronson's objection: you're using different intervention distributions at the two scales, so the comparison isn't fair; use the same intervention distribution at every scale — at which point the scales stop differing.
Hoel's counter has his characteristic flavor of running the opponent's suggestion all the way out to where it looks absurd: if you insist on using the microscale's intervention distribution while intervening at the macroscale, you have to weight it by how many microstates each macrostate contains. Which means that to work out how strongly a light switch causes a light bulb, you first have to know how many atoms are in the state "up."
This is the definition of comparing apples and oranges. "A primer on causal emergence," 2021
His summary: insist on one intervention distribution and you are either intervening on the micro as though it were a macro (making the microscale, by definition, a macroscale), or on the macro as though it were a micro (and vice versa). Either way, the demand for a "fair comparison" eats itself.
But the best thing in the essay is its final section, titled "So what is a valid criticism?" — where he spends a full section writing a better objection on his opponent's behalf. That one goes in section six below.
The last one, the most philosophical and the one he pushes hardest. His 2024 essay opens with a thought experiment built on his favorite question form: if X were true, what would the world look like?
Suppose you really do live in a dualistic universe — an incarnated soul in a body, or an agent from a vaster reality embedded in a simulation. What would science look like there?
Probably like this: the laws of physics make sense and behave, with one conspicuous gap — nobody can explain how a body has a you in it. Scientists there would open up brains and say: it seems to be following all the physical rules, but we can't explain how you fit into this thing. Great thinkers there would write sentences like "No explanation given wholly in physical terms can ever account for [it]." "I regard [it] as fundamental. I regard matter as derivative from [it]."
Then he opens the brackets: in our universe those are David Chalmers, Max Planck and Eugene Wigner.
He keeps going. In that universe neuroscience would lag conspicuously behind the other sciences (checks out). And its AI would have suspiciously little to do with how brains work — as if they'd found a totally orthogonal form of intelligence (he admits things are "getting uncomfortable here," because that checks out too).
One road from here is to declare this proves religion. He closes it immediately: "A gap of inexplicability doesn't recommend anything." It points at no particular doctrine.
So he goes looking for another explanation, one he thinks is almost never explored: maybe there is no way to set up a universe without this confusion. Which is the same as asking whether science is fundamentally incomplete.
Mathematics has a famous answer to that question. Gödel encoded a formal system so it could talk about itself, then constructed a sentence saying "this sentence is not provable in S." If it's proved, S is inconsistent; if S is consistent, it can't prove the sentence — which is nonetheless true. There are truths a system can't reach from the inside.
Hoel's claim:
A theory of consciousness is like a Gödel sentence written in the language of science. "Consciousness as a Gödel sentence in the language of science," 2024
He is explicit that there's no direct evidence for this, only circumstantial. Four strands of it, and the best one is a paradox he builds himself out of Chalmers's zombie argument — a zombie being atom-for-atom like you, neurons firing, with no experience inside; Chalmers argues that since such a world is conceivable, physics doesn't fix consciousness, so pure materialism is false.
Hoel's reversal: let a zombie make the argument.
The zombie conceives of a world full of zombies, finds it conceivable, and concludes materialism is false. But wait — in the zombie's world materialism is true. So the argument delivered the wrong answer. Which means its validity rests on a hidden premise: that whoever is making it isn't a zombie.
Fine, call it a premise. The trouble is that if the argument works, then by definition no utterance, behavior, experiment or introspection can distinguish a zombie from a non-zombie. There could be a zombie Descartes — Hoel calls him Zescartes — who concludes he is indubitably conscious and then makes the zombie argument, for whom the conclusion is false. And neither Descartes nor Zescartes can tell themselves apart.
So: if the argument is valid without that premise, a zombie in a materialist world can use it to refute materialism — inconsistency. If it's valid only with the premise, then any conscious being can make it and then notice that whether they are a zombie is settled neither by introspection nor by evidence — incompleteness. Both roads end in a Gödelian shape.
The existence of consciousness appears to be the kind of true fact that cannot be proven from the inside. "Consciousness as a Gödel sentence in the language of science," 2024
The other three strands: physics already contains proven undecidable properties (the 2015 Nature paper "Undecidability of the spectral gap," itself triggered by recursion); arguments in philosophy of mind keep terminating in paradox; and the shape of scientific history itself — Galileo bracketing qualia out so physics could run, Russell and Whitehead bracketing self-reference out with type hierarchies so mathematics could run. Both were delays, not cures.
Note what the claim is and isn't. It is not mysterianism. He explicitly separates himself from the neighbors: not Colin McGinn's mysterianism (which says humans are too dim and a superintelligence would crack it), because he's talking about a limit of the formal system, not of IQ; and not Hofstadter's identification of consciousness with symbolic recursion (his jibe: a Skyrim NPC presumably has rudimentary symbolic recursion, and an arrow to the knee doesn't really hurt it).
And the ending offers no consolation at all:
It is merely a blankness, a clouded-over wall that extends infinitely in all directions. But still, upon finding such an inaccessible wall, we humans cannot help our nature. We cannot help but wonder: what's behind the wall? "Consciousness as a Gödel sentence in the language of science," 2024
Changing scale is his default move. It is very nearly the generator of all his ideas: causal emergence asks at which scale a system's causation is sharpest; the Overfitted Brain Hypothesis imports the language of machine learning down to neurons; "Our Overfitted Century" exports the language of neurons up to civilization. He names the move himself, invoking David Marr's levels of analysis — the same thing can be explained at the computational, algorithmic and physical levels, all correct, just at different levels. But he treats cross-level analogy as a tool with a bill to pay: name the discriminator, locate the feedback loop, say which direction the variance is collapsing. That's the real line between him and the general "everything is like everything" school of cultural commentary.
He reasons backwards: if this were true, what would the world look like? The Gödel essay is built entirely that way — posit the soul universe, derive the features its science would have, then check them off against ours one by one, including the ones that make him uncomfortable. The virtue of the move is that it converts an undecidable metaphysical question into a checklist you can actually go through.
He demands that a theory explain the phenomenology instead of dumping it in the "byproduct" bin. That's the entire lever he uses against memory consolidation: the sparseness and weirdness of dreams are not noise, they are the thing to be explained. You can steal this one directly — every time somebody tells you something is "just a byproduct," ask why the byproduct is so regular.
He writes better objections on his opponents' behalf. The last section of the causal emergence primer is literally headed "So what is a valid criticism?", and there he sets down the strongest weakness in his own theory (next section). It isn't a pose: he says in that section that Aaronson should have gone this way instead, and that it "would have been a much better blog post."
He treats discomfort as a signal, not a brake. "Why we stopped making Einsteins" closes with a sentence very few writers would put on the page:
It's telling I felt uncomfortable writing this essay, despite being confident it's true. "Why we stopped making Einsteins," 2022
The argument is genuinely unwelcome. He goes through a mountain of biographies (Marcus Aurelius had seventeen tutors; Bertrand Russell didn't attend school until sixteen; Wittgenstein was tutored privately until fourteen; a family tutor named Max Talmud handed twelve-year-old Einstein geometry) and concludes that the only historically reliable method for producing geniuses is aristocratic tutoring, and it is deeply unfair. Meanwhile we take the most effective educational method we know, aim it almost entirely at tests, and apply it almost entirely to the students doing badly —
In general, those doing well in school don't get tutoring—it's like we're applying the secret genius sauce solely to the kids who aren't going to be geniuses. "Why we stopped making Einsteins," 2022
He ties this back to handmade versus mass-produced: older dresses, hand-stitched rugs, twenty-times-folded steel, Stradivari violins. Modern intellectual life, on his account, is full of mass-produced knock-offs of its aristocratic forebears — "They are of decent quality, they serve their purpose, and boy, they sure were cheap to produce. But I don't think they have quite the same sound."
And finally, he pulls his own chair out from under himself at his most triumphant. In 2022 he won first place and $5,000 in Scott Alexander's anonymous book review contest with a long essay called "The gossip trap." Its target is the Sapient Paradox: humans have been around for 200,000 years, so why is civilization only 10,000 years old? His hypothesis is that below Dunbar's number (~150), humans never needed formal power structures at all — all you need is who's popular and who isn't.
A "gossip trap" is when your whole world doesn't exceed Dunbar's number and to organize your society you are forced to discuss mostly people. It is Mean Girls (and mean boys), but forever. "The Gossip Trap," 2022
In such a world every cognitive resource goes into reputation management, leaving nothing in the tank for invention, art or engineering. Hence his inverted, elegant definition of civilization: "a superstructure that levels leveling mechanisms" — its hallmark is immunity to gossip, which is what tenure, and Supreme Court justices, and protected journalists are for. And what social media did was pull down the Dunbar wall and summon our oldest form of government back.
He could have stopped there. He didn't. In the final section he turns and points at himself:
An admittance. For it should be obvious by now: this text is corrupted. The same corruption that I accused the Davids of falling victim to, and that they, in turn, accused Rousseau and Hobbes of falling victim to. I have made the past political, and prehistory a sepia reflection of the current day. "The Gossip Trap," 2022
And then says: I still think it's a true yarn. He lets both sentences stand rather than picking one to end on. The last line is that all these prehistorical narratives may be a hall of mirrors, in which, projected in gigantic distortions all around us, we see only our own face.
The strongest objection to causal emergence was written by him. In the last section of the primer he says Aaronson's "unfair comparison" line is actually weak, and that the genuinely hard problem is this: causal emergence shows up under certain measures of causation (effective information, integrated information) — so how do you convince anyone that this is the best, let alone the uniquely correct, measure?
Perhaps one could convince an extreme skeptic like him by proving a measure like effective information is unique and best-fitted for measuring causation […] But proving this with any mathematical measure of anything is both rare and hard, and neither I nor anyone else has done this. "A primer on causal emergence," 2021
He immediately adds that this isn't a reason to disbelieve the theory, only a reason to go beyond effective information. You can take or leave that pivot, but the soft spot is real: the whole argument rests on a measure nobody has shown to be unique. (To be clear: I read his reply, not Aaronson's 2017 original, so everything here about Aaronson's position is Hoel's account of it.)
His whole research line has been publicly named by peers. In September 2023, over a hundred researchers signed an open letter calling Integrated Information Theory "pseudoscience"; signatories included Daniel Dennett, Patricia Churchland, Joseph LeDoux, Hakwan Lau and Keith Frankish. Hoel rebutted the letter line by line — despite having just published a book with several chapters on IIT's failures and limits, and despite not having worked on the theory for years. His central point is one you can evaluate on its own: "IIT's problems are your problems" — the flaws treated as unique to it (the substitution argument, say) are shared by every theory of consciousness; IIT is just formalized enough that you can see them. That's a real argument, but notice its shape: it defends one theory by showing that all theories are equally bad off. (Here I read his full rebuttal, not the letter itself or the signatories' subsequent replies, so I report only the bibliographic facts and his position.)
The Overfitted Brain Hypothesis is beautiful, and beautiful isn't confirmed. He says himself that nobody knows for sure why dreaming evolved. And theories of this shape have a structural weakness worth weighing: precisely because it accounts for the phenomenology of dreams so smoothly, it's hard to say what observation would overturn it — a bizarre dream is domain randomization, and a dull one can be assigned to some other regularization. He offers it as a hypothesis, which is fine; but as things stand it's closer to a powerful framework than to a winner selected by experiment.
"Cultural overfitting" is a metaphor, and he knows it. He writes plainly that you can't prove culture is overfitted with a handful of images, and that the book version was turned down by two major houses. So read that chapter as what it actually is: a well-organized analogy with a set of matching mechanisms, not a measurement. The premise of cultural stagnation is also less settled than the essay implies — he quotes Katherine Dee's counter (internet personae are themselves a new form, critics just can't perceive it) and then largely moves on.
The "decline of genius" premise has a methodological hole he only half-concedes. He admits that "verifying an absence is surprisingly difficult." The layer he doesn't touch: we can list the world-historical geniuses of the nineteenth century because that list has been through 150 years of filtering and canonization, and contemporaries almost never recognize their contemporaries. Arguing decline by asking "name someone who died in the last decade who belongs in that pantheon" is very hard to separate from the fact that pantheons take decades to build.
Last and most important: his commitments all tilt the same way. Line them up — emergence over reduction; IIT worth defending; science incomplete in the face of consciousness; language models are "more like tools"; culture is draining of consciousness. Every one of them leaves an irreducible seat for the view from inside. That doesn't make any single one wrong. But a man raised in a bookstore, who writes novels, and who evidently wants "what it is like from the inside" to be fundamental, landing on the same side of every independent question — that is information a reader should hold onto.
And the harshest verdict on that tilt is still his own: this text is corrupted. He applied it to the essay he was proudest of. It's worth noting he has not applied it to his scientific papers.