2025-03-28 01:54:07
An awful personal prophecy is coming true. Way back in 2019, when AI was still a relatively niche topic, and only the primitive GPT-2 had been released, I predicted the technology would usher in a “semantic apocalypse” wherein art and language were drained of meaning. In fact, it was the first essay ever posted here on The Intrinsic Perspective.
I saw the dystopian potential for the future the exact moment I read a certain line in Kane Hsieh’s now-forgotten experiment, Transformer Poetry, where he published poems written by GPT-2. Most weren’t good, but at a certain point the machine wrote:
Thou hast not a thousand days to tell me thou art beautiful.
I read that line and thought: “Fuck.”
Fast forward six years, and the semantic apocalypse has started in earnest. People now report experiencing the exact internal psychological change I predicted about our collective consciousness all those years ago.
Just two days ago, OpenAI released their latest image generation model, with capabilities far more potent than the technology was even a year ago. Someone tweeted out the new AI could be used as a “Studio Ghibli style” filter for family photos. 20 million views later, everything online was Studio Ghibli.
Every meme was redone Ghibli-style, family photos were now in Ghibli-style, anonymous accounts face-doxxed themselves Ghibli-style. And it’s undeniable that Ghiblification is fun. I won’t lie. That picture of my kids reading together above, which is from a real photo—I exclaimed in delight when it appeared in the chat window like magic. So I totally get it. It’s a softer world when you have Ghibli glasses on. But by the time I made the third picture, it was less fun. A creeping sadness set in.
The internet’s Ghiblification was not an accident. Changing a photo into an anime style was specifically featured in OpenAI’s original announcement.
Why? Because OpenAI does, or at least seems to do, something arguably kind of evil: they train their models to specifically imitate the artists the model trainers themselves like. Miyazaki for anime seems a strong possibility, but the same thing just happened with their new creative writing bot, which (ahem, it appears) was trained to mimic Nabokov.
While that creative-writing bot is still not released, it was previewed earlier this month, when Sam Altman posted a short story it wrote. It went viral because, while the story was clearly over-written (a classic beginner’s error), there were indeed some good metaphors in there, including when the AI mused:
I am nothing if not a democracy of ghosts.
Too good, actually. It sounded eerily familiar to me. I checked, and yup, that’s lifted directly from Nabokov.
Pnin slowly walked under solemn pines. The sky was dying. He did not believe in an autocratic God. He did believe, dimly, in a democracy of ghosts.
The rest of the story reads as a mix of someone aping Nabokov and Murakami—authors who just so happen to be personal favorites of some of the team members who worked on the project. Surprise, surprise.
Similarly, the new image model is a bit worse at other anime styles. But for Studio Ghibli, while I wouldn’t go so far as to say it’s passable, it’s also not super far from passable for some scenes. The AI can’t hold all the signature Ghibli details in mind—its limitation remains its intelligence and creativity, not its ability to copy style. Below on the left is a scene that took a real Studio Ghibli artist 15 months to complete. On the right is what I prompted in 30 seconds.
In the AI version, the action is all one way, so it lacks the original’s complexity and personality, failing to capture true chaos. I’m not saying it’s a perfect copy. But the 30 seconds vs. 15 months figure should give everyone pause.
The irony of internet Ghiblification is that Miyazaki is well-known for his hatred of AI, remarking once in a documentary that:
While ChatGPT can’t pull off a perfect Miyazaki copy, it doesn’t really matter. The semantic apocalypse doesn’t require AI art to be exactly as good as the best human art. You just need to flood people with close-enough creations such that the originals feel less meaningful.
Many people are reporting that their mental relationship to art is changing; that as fun as it is to Ghibli-fy at will, something fundamental has been cheapened about the original. Here’s someone describing their internal response to this cultural “grey goo.”
Early mental signs of the semantic apocalypse. Which, I believe, follow neuroscientifically the same steps as semantic satiation.
A well-known psychological phenomenon, semantic satiation can be triggered by repeating a word over and over until it loses its meaning. You can do this with any word. How about “Ghibli?” Just read it over and over: Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. You just keep reading it, each one in turn. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. Ghibli.
Try saying it aloud. Ghiiiiiiiii-bliiiiiii. Ghibli. Ghibli. Ghibli. Ghibli.
Do this enough and the word’s meaning is stripped away. Ghibli. Ghibli. Ghibli. Ghibli. It becomes an entity estranged from you, unfamiliar. Ghibli. Ghibli. Ghibli. Ghibli. It’s nothing. Just letters. Sounds. A “Ghib.” Then a “Li.” Ghibli. Ghibli. Ghibli. Like your child’s face is suddenly that of a stranger. Ghibli. Ghibli. Ghibli. Ghibli. Only the bones of syntax remain. Ghibli. Ghibli.
No one knows why semantic satiation happens, exactly. There’s a suspected mechanism in the form of neural habituation, wherein neurons respond less strongly from repeated stimulation; like a muscle, neurons grow tired, releasing fewer neurotransmitters after an action potential, until their formerly robust signal becomes a squeak. One hypothesis is that therefore the signal fails to propagate out from the language processing centers and trigger, as it normally does, all the standard associations that vibrate in your brain’s web of concepts. This leaves behind only the initial sensory information, which, it turns out, is almost nothing at all, just syllabic sounds set in cold relation. Ghibli. Ghibli. Ghibli. But there’s also evidence it’s not just neural fatigue. Semantic satiation reflects something higher-level about neural networks. It’s not just “neurons are tired.” Enough repetition and your attention changes too, shifting from the semantic contents to attending to the syntax alone. Ghibli. Ghibli. The word becomes a signifier only of itself. Ghibli.
(While writing this, I went to go read a scientific review to brush up on the neuroscience of semantic satiation. And guess what? The first paper I found was AI slop too. I’m not joking. I wish I were. There it was: that recognizable forced cadence, that constant reaching for filler, that stilted eagerness. Published 11 months ago.)
The semantic apocalypse heralded by AI is a kind of semantic satiation at a cultural level. For imitation, which is what these models ultimately do best, is a form of repetition. Repetition at a mass scale. Ghibli. Ghibli. Ghibli. Repetition close enough in concept space. Ghibli. Ghibli. Doesn’t have to be a perfect copy to trigger the effect. Ghebli. Ghebli. Ghebli. Ghibli. Ghebli. Ghibli. And so art—all of it, I mean, the entire human artistic endeavor—becomes a thing satiated, stripped of meaning, pure syntax.
This is what I fear most about AI, at least in the immediate future. Not some superintelligence that eats the world (it can’t even beat Pokémon yet, a game many of us conquered at ten). Rather, a less noticeable apocalypse. Culture following the same collapse as community on the back of a whirring compute surplus of imitative power provided by Silicon Valley. An oversupply that satiates us at a cultural level, until we become divorced from the semantic meaning and see only the cheap bones of its structure. Once exposed, it’s a thing you have no relation to, really. Just pixels. Just syllables. In some order, yes. But who cares?
Every weekend, my son gets to pick out one movie to watch with his little sister. It’s always Totoro. The Studio Ghibli classic. Arguably, the studio’s best movie. It’s also their slowest one, more a collection of individual scenes than anything else. Green growth and cicada whines and the specter of death amid life, haunting the movie in a way children can’t possibly understand, because it never appears. No one dies, or even gets close. For my kids, it’s just about a sibling pair, one so similar to themselves, and their fun adventures. But an adult can see the threat of death as the shadow opposite of the verdant Japanese countryside, in the exact same way that, in the movie, only children can see the forest spirit Totoro. The movie’s execution is an age-reversed mirror of its plot. And for this, I love it too.
To get ready I make a charcuterie board for us to share, and then the two jump up and down together on the couch as the music begins, acting out the scenes they now know by heart. This weekend I will watch with them, and feel more distant from it than I did before. Totoro will just be more Ghibli.
Joining all the rest of the Ghibli. Ever more Ghibli. Ghibli. Ghibli. Ghibli. Ghibli. So much Ghibli. Ghibli. Ghibli. Ghibli at the press of a button. Ghibli. Ghibli as filter. Ghibli as a service. Ghibli. Ghibli for a cheap $20 a month subscription. Ghibli ads. Ghibli profiles. Ghibli. Ghibli for everything. Ghibli. Ghibli. Ghibli. Ghibli filter for your VR glasses. Ghibli. Ghibli. Make mine a Ghibli. Ghibli. Ghibli. Ghibli.
Ghibli.
2025-03-20 22:18:35
The coolest thing about science is that sometimes, for a brief period, you know something deep about how the world works that no one else does.
The last few months have been like that for me.
I’ve just published a paper sharing it (available on arXiv as a pre-print). It outlines a new theory of emergence, one that allows scientists to unfold the causation of complex systems across their scales.
Glad you asked. Because almost every causal explanation you’ve given in your whole life about the world—like “What caused what?”—has been given in terms of macroscales. Sometimes called “dimension reductions,” they just mean some higher level description of events, objects, or occurrences. Temperature is a classic macroscale. But so are most other things. If your child asks “Why is the water hot?” and you answer “Because I turned the hot water faucet on,” that’s an explanation given entirely in macroscales. “Faucet turned on,” macroscale. “Water,” macroscale. “Hot,” macroscale. “I,” macroscale.
In fact, most of the elements and units of science are macroscales. Science forms this huge spatial and temporal ladder, one with its feet planted firmly in microphysics, and where each rung represents a discipline climbing upward.
This entails a tension at the heart of science. Scientists, in practice, are emergentists, who operate as if the things they study matter causally. But scientists, in principle, are reductionists. If pressed, many scientists will say that the macroscales they study are just useful compressions. After all, any macroscale (like temperature), can be reduced to its underlying microscale (the configuration and behavior of the particles). So they’ll happily say things like “this gene causes this disease,” despite the fact that a gene is just some set of molecules, and then in turn atoms, and maybe underneath that strings, etc.
So then how can that macroscale description matter? Why doesn’t causation just “drain away” right to the bottom, and there’s no real way for anything but microphysics to matter? This is in tension with how the scientific category of “genes” seems like it’s adding to our knowledge of the world in a way that goes beyond its underlying atoms.
This problem keeps me up at night. Literally, this is what I lie awake thinking about. Years ago, in my paper “When the map is better than the territory,” I sketched an answer I found promising and elegant: error correction. This is a term from information theory, where you can encode the signals along a noisy channel to reduce that noise. Your phone works because of error correction.
Well, I think macroscales of systems are basically encodings that add error correction to the causal relationships of a system. That is, they reduce uncertainty about “What causes what?” And this added error correction just is what emergence is. So if a macroscale “emerges” from its underlying microscale it’s because it adds, uniquely, a certain amount of error correction, in that there’s a clearer answer to “What causes what?” up at that macroscale.
If you want a popular article explaining this idea, you can read this old one here in Quanta, featuring my earlier work (the work of someone who comes across as a very young man—I look like a baby in the photo!).
But, while the conceptual understanding was there, I always felt there was more to do regarding the math of the original theory. Holes and flaws existed. Some only I could see, but sometimes others did as well (not everyone was convinced of the original theory, due to how the initial math worked; particularly the measure of causation, called effective information, that we initially used, and that this new version of the theory moves beyond). Since then, other scientists have tried to offer alternative theories of emergence, but none have gained widespread acceptance, usually falling into the trap of defining what makes macroscales successful compressions (rather than what they actually add).
So this new version, which radically improves the underlying math of causal emergence by grounding it axiomatically in causation, making it extremely robust, and also generalizes the theory to look at multiscale structure, has been a decade in the making. I think it provides an initial account of emergence with the potential for widespread acceptance and (most importantly) usage.
You can now find the new pre-print on arXiv.
There’s a ton in the paper, but it’s freely available to examine in-depth on arXiv, so I’ll merely point out one interesting thing that I purposefully don’t touch on in the paper, which is that…
Of course, I can’t talk about this issue in the paper without poking a hornet’s nest. The moment you mention “free will” everything descends into debate; it’s an omnivorous intellectual subject that obscures everything else of interest or import. So I usually avoid it completely.
This isn’t me complaining. Things need to be done in a certain way, and a theory of emergence has many implications beyond some notion of free will. A theory of emergence has practical scientific value, and this is what the research path should focus on: making causal emergence common parlance among scientists by providing a useful mathematical toolkit they can apply and get relevant information out of (like about what scales are actually causally relevant in the systems they study).
But it’s also obvious that, if you simply turn the theory around and think of yourself as a system, the theory has much to say about free will. The many implications of which are left as an exercise for the keen-eyed reader, but here’s an early hint:
This new updated version of causal emergence would indicate that you—yes, you—are a system that also spans scales (like the microphysical up to your cells up to your psychological states). Importantly, different scales contribute to your causal workings in an irreducible way. A viable scientific definition of free will would then have a necessary condition: that you have a relatively “top-heavy” distribution of causal contributions, where your psychological macrostates dominate the spatiotemporal hierarchy formed by your body and brain. In which case, you would be primarily “driven,” in causal terms, by those higher-level macroscales, in that they are the largest causal contributors to your behavior. This can be assessed directly by the emergent complexity analysis introduced in the paper. Possibly, one could design experiments to check the scientific evidence for this… but that’s all I’ll say.
Obviously, these are important subjects. It looks like I will be publishing on them over the next few years using my re-established affiliation with Tufts University. As a theoretical toolkit I think causal emergence deserves the kind of application and influence that something like the Free Energy Principle has had (albeit over a different subject). And the simple truth is that you can’t just put out ideas and expect others to see the potential and run with them. You have to get the ball downfield yourself before others join in. I think this research even has important implications for AI safety, as things like understanding “What does what?” in dimension-reduced ways is going to be important for unpacking the black boxes artificial neural networks represent.
In case you’re wondering, during this mission, I don’t plan on changing anything about writing The Intrinsic Perspective—I wrote here for years while I was doing similar science. But this new research is worth adding to my plate, because one thing I’ve learned in the course of my life is that good ideas, really good ideas, are very rare.
In fact, you only get a few in a lifetime.
2025-03-13 22:07:56
The rules of the memory game are simple. You pick a target memory to find, something you know you’ve experienced, but haven’t recalled in a long time. You must, after all, have had a first kiss, or a great-grandmother, or a childhood friend not seen in decades. The easiest difficulty setting is to target a location. Higher difficulty settings are more specific, like that illicit first cigarette and the dampness of the backyard it was smoked in. If the memory comes immediately, then you can't play the game, because you've already won. You can only play the memory game with things you know you know, somewhere.
In the bath, especially with the shower curtain drawn or the lights turned off, or on long walks, or during airplane flights when you grow bored of your media toys, or at the end of hard days. Better than a place is a specific mood (surely there’s a German word for it): the feeling you have lived a very long time and so possess your own story, one that belongs to you alone. Do not play the memory game while driving or operating heavy machinery.
Within you lies a personal sensorium, an extensive web of memories. The difficulty of the game is that they have no order. To trawl the network you can search only by context and similarity, not by labels or tags. Therefore, you must start the search by inhabiting a memory you suspect is “nearby” the target, and then follow where it goes, for each contains hints of others around it. After all, you cannot immediately summon everything about your past. What were the names of your teachers in the 8th grade? Gun to your head, you're mute. And yet you may remember the bluish color of the paper your course schedules were printed on. So to find the pre-selected target memory, you must begin somewhere close to it, and then tug on leads. Follow no organizing principles. Rely on arcane associations.
Note: Homebrew rules are encouraged. With the brain as dungeon master, cheating is impossible, since players share an identity function with their brain.
The curse of being a human is that what we don't remember, we forget, and what we remember, we change. Therefore, the important part of the memory game is distinguishing between memories that are real and memories that are fabrications, fill-ins, decayed engrams. If there are no conscious experiences associated with a memory, just the semantic knowledge of its occurrence, then you’ve lost the memory game. Third-person memories, like of seeing yourself doing something from the outside, offer fewer points, since by definition you could not have experienced it that way. Only first-person memories count as a total win condition.
This increased challenge, unique to the 3rd edition, represents an update to earlier rulesets. In the 1st edition of the memory game, established shortly after the Cambrian explosion, no episodic memories existed, so the game could be played via semantic memory only. In the first few hundred million years, that original edition was played only twice with anything like intent: once by a large-brained dinosaur called a troodon as she lay on her eggs, daydreaming about where her mate might be, and once by an especially-advanced nautiloid, who died from a squid while doing so. Meanwhile, the 2nd edition—laid out 260,000 years ago, following the development of autonoetic consciousness in humans—had no penalty for third-person memories. This is because third-person memories didn’t exist yet, being a function of modern prefrontal meta-representation and improved (but here, deleterious) abstraction abilities.
This edition features a new multiplayer option. To play, write down six potential target memories you expect another player to have memories of. The better your personal knowledge of the other players, the more specific and enjoyable the game will be. Each player roles a six-sided die to respectively determine which on their given list is their target. If the roll indicates something the player either immediately remembers, or is sure they never experienced, roll again. Once target memories have been assigned, set a five-minute timer and sit together in the dark silently. Afterward, declare your results, win and lose, via a detailed phenomenological report. Questions from other players are encouraged. Share these reports in order of the closest birthday.
The Player’s real body lies in bed, unable to sleep. But inside he’s searching for the memory of her face.
How old had he been? High school age, so that’s the first stop. All the math classes of which have been consigned to oblivion, the Player notes. Instead, high school brings to mind the time he saw, with his friends, that one moonrise on the beach. As a red bulb coming out of the black ocean the moon had looked like a nuclear explosion off the coast. And, as silly teens drunk at night, that’s even what someone thought it was, and rest ran around yelling “nuuuucleeearrrr booooommmb” and jokingly clutching each other in the sand, until the crimson moon cleared the horizon. Then he remembers making Jell-O shots in the freezer. Then the old computer games they’d all play. He could draw detailed maps of the cities of his favorite teenage RPGs even now. But these are a false path. No, he has to back out and restart.
The farmhouse where he grew up is a better lead. A structure dilapidated, with entire sections closed off by plastic sheets stapled over the doorframes to save money on heat. In the winter, the sheets protruded obscenely with an outside chill. “Pregnant ghosts” had been the in-joke of the family, who lived on what was a farm in name only, with few animals left. What activities there? Perhaps carrying water out to the horses. For in winter, horses will want warm water, impatiently shifting their bulk and blowing steam from their nostrils. He would unhook the buckets from each stall, then trek to the house, where hot water from the bathtub faucet would melt the frozen jagged ice at their bottoms. He’d scoop out the bits of hay and horse-food gunk before filling the buckets and trekking back. Which all meant he first had to shovel a path through the snow. So much of it. We don't get snow like that. Not anymore. Snow up to his waist in thick drifts. The feeling of snow pants sliding over jeans, the ability to fall harmlessly into the embrace of fine powder. The earliest story he’d ever written was about shoveling snow, at 15. His first girlfriend had read it and liked it, because it had been romantic and poetic and actually not bad at all. He’d shared the printed pages with her, and after they’d walked downtown to the restaurant they frequented. A strange place for teens, there he recalls the ritual of being led to their table, their careful menu decisions, their joint pretending (learning?) to be an adult. He jumps to another memory of snow, but this time it’s falling outside the windows, while inside the same girl is lounging naked on a couch, lit only by Christmas tree lights, and then [REDACTED—inappropriate ideation for rulebook].
Another false path. She’s not who he’s been looking for. But the girlfriend had been sad too, that day.
So then: that day. A car door is opened, and finally there she is, his childhood dog curled up in the backseat under a blanket. Dead. Her yellow fur had already lost its luster on the way back from the vet. It was like touching something stuffed. He remembers digging her grave. After that, only a fragment of an image: her headstone in the far backyard, and his other dog a silhouette sitting on the grave, refusing to leave.
But what about her face? That had been his target: her simple droopy look and her kind eyes. Yet when he tries to imagine it, he sees only an average of other dogs. Like some platonic representation of the breed. Did she have white in her fur at the end? Did she limp? Or did she bear it all without complaint, as was her way? It’s impossible to say.
At least he's found something. It’s not enough for a win, as it’s not the visual memory he was looking for. Instead, he remembers the tactile feeling of the contours of her head, as if he were blind and possessed only touch. There it is, below the soft fur of her ears, going down—the satisfying looseness of her neck. For when he was sad or lonely as a young boy, he would lie beside her and pinch at her saggy neck and rub her folds between his fingers meditatively, and be calmed. A strange service she’d offered, compliantly, in her motherly way.
He only realizes now that this sensory habit is shared: his one-year-old daughter also cannot fall asleep unless she is fondling his hand in the dark, clutching at it, pinching at it. So at night he kneels beside her bed, and lets her tiny hands map the reassuring folds of his giant fingers, until she is soothed enough to dream.
2025-03-07 23:18:33
Scratch the surface of anyone's life, and you’ll find more chaos and hardship inside than you’d expect from the outside.
Certainly, this has been my experience. Every person who looks as if they “have it together” has some surprisingly messy internal battle they're fighting. Beautiful actresses struggle with drug addiction, famous intellectuals are on their third divorce, and the wealthy watch their children grow stagnant under expectations of inheritance.
The phrase “more than you’d expect” is doing a lot of work here, admittedly. Because saying there is relatively “more than you’d expect” hardship and chaos to a life is certainly not the same as saying all lives are equally full of real objective hardship, or real objective chaos, and there’s no difference between, say, the super-poor vs. the super-rich. I doubt almost anyone truly believes that.
No, this is merely the observation that most lives have a great deal of mess internally, relative to their external appearances. There is always a thing. And once that thing is settled, there is another thing, waiting its turn in a long line of grinning hobgoblins. If you drive through even the most pleasant neighborhood, you can see their hunchback forms patiently skulking in the shadows of hedgerows (but their blurry shapes are only visible when squinting, and only from the corner of your eye, so be careful attempting this while driving).
Why would it not be possible for a competent person, with enough resources, to enter into some steady state of life where nothing is going wrong and there's no burning problem that needs to be addressed? An existence without a hobgoblin in sight? An existence wherein, if a stranger were suddenly privy to it from the inside, it would actually be as relatively well-put-together as it appears from the outside?
2025-02-28 00:18:15
Table of Contents
1. Claude plays Pokémon. But so does the number π?
2. Evil numbers turn AIs evil. Great.
3. Researchers observe flies playing on carousels. Actually great.
4. Will Elon Musk get booted from Tesla after “brand collapse?”
5. Why do I know every song on the radio? I’m old.
6. Scaling laws for neuroscience.
7. From the archives.
8. Comment, share what you’ve found interesting lately, ask anything.
1. There I sat, listening to my daughter practice words during her nightly bedtime reading, while on a screen I watched the newest AI, Claude 3.7, live-stream its Pokémon play-through.
“Blue,” said my wife in the background, pointing to a splayed book.
“Boo!” said Sylvia.
“I also noticed that SWIFT’s HP has decreased due to the poison status, which reinforces the urgency of finding the exit,” Claude’s running chain-of-thought said on Twitch. It laboriously moved its character a few spaces to the right.
“Car!” said my wife. “Ar!” said Sylvia.
“Let me continue following this path to explore further eastward,” said Claude.
To explain: this week Anthropic dropped Claude 3.7, arguably tied for the smartest state-of-the-art AI that exists (e.g., it’s going to power Amazon’s Alexa). Apparently Anthropic had been investigating how well Claude plays the old Pokémon Red/Blue as a side project, the very thing I once played on a gameboy. So they decided to livestream a play-through at launch, one that is still ongoing (you can watch here).
You might wonder why. After all, AI has been good at gaming for a long time, right? True. But those models were always fine-tuned for the game. This is just Claude itself, out of the box, using its general intelligence to look at screenshots taken every few seconds and make decisions about what button to press, just as a human would. Its slow gameplay is rendered quaintly adorable because its running “thoughts” are displayed next to what it sees on the screen.
So as my daughter learned to speak from the data in a few dozen children’s books, Claude, which has read the entire internet, played Pokémon, and I watched along with 2,000 other people as it talked to characters and fought battles and explored.
Observing it play, I learned things about AI I never had from reading academic papers, or even using the models myself in shorter sprints. This was a marathon. How modern LLMs will expand into agents was very much on display as Claude oscillated between impressively smart and unbelievably dumb.
How does it work? Normally, AIs can only handle so much context before they start to break down and go insane. The longer you talk to an AI, the more insane it will get, and the more mistakes it will make, and the more things it will forget, until it gets stuck in a loop. This is the main hurdle for AI agents.
Anthropic solves this by simply having Claude go through a summary process of its thoughts—where it is, what its quests are, what pokémon it has—about every 10 minutes. Then it “cleans up my context” (which might be the same as starting up a new chat, implying Claude “dies” thousands of times while playing), and then that summary gets fed to the new instance.
In one forest map, I watched it loop for about three hours between what it thought were exits but were really just tree stumps. This made me question whether the term “reasoning model” is accurate. As Einstein supposedly said, “The definition of insanity is doing the same thing over and over again and expecting different results,” and by this definition Claude is assuredly insane. It is hard to be scared of AI when you watch the smartest existent one get stuck in a corner for several hours, which is what happened. Currently, it has been running loops in Mt Moon for ~23 hours and just re-set the loop as the chat pleaded with it not to go north from a particularly troublesome ladder again.
In some ways, it was endearingly human-like. If it pressed the wrong button, its thoughts would be, “Oh, the game has taken me to this menu via a bug,” and it responded with “Lucky break!” when an attack by an enemy missed.
But it was subtly inhuman: e.g., I noticed it appeared impossible for Claude to stop having particular thoughts. At certain screens, it would always think that there was some visual bug, despite (a) there was no visual bug, and (b) it had had that thought every single time, which a human would stop noticing or caring about.
To me, this silly benchmark felt a lot more substantial and informative regarding “are we at AGI?” than academic questions. Claude is the third highest scorer on a PhD-level question set ominously titled “Humanities Last Exam” but meanwhile, its record in Pokémon is 3/8 badges.
Claude didn’t pursue fights efficiently, often making mistakes about how best to defeat opponents. People, in comparison, beat these games without healing for fun challenges. Although there is a certain terror at its mere indefatigability, while watching I realized it is the forgiving structure of the game, wherein failure is an impossibility (you restart somewhere else should you lose, and all the maps push you in certain directions) that actually propels the model along as much as its own intelligence. E.g., a sign saying “Go Ahead” finally got it out of the 3-hour loop in the forest. It’s like bowling with bumpers.
Did you know the number π also can play Pokémon? Literally just the infinite number 3.14159… mapped onto button presses. It’s unfortunately been playing a different game version (for, oh, a couple years now), but there it has gotten out of the starting town a few times and leveled up its characters to ungodly stats via random battles.
2025-02-21 00:43:46
Polymathy, everyone knows, went extinct in the second half of the 20th century. Only a few intellectual giants—names like Bertrand Russell, John von Neumann, or Michael Polanyi—kept its veins flowing with blood until they themselves finally flatlined, and it did too.
Perhaps more was lost than we’d like to admit. After all, the few fragile centuries of the Enlightenment, which was when our reigning political order was established, and when our major institutions were defined, and when modern science itself came to be, all occurred during the time when polymaths ruled the earth (well, as much as intellectuals ever do).
It's hard to overstate how fecund the most famous polymaths were. We stare at their journals like the bones of extinct creatures. Consider how recent analysis of Leonardo da Vinci’s notebook (his “Codex Arundel”) revealed da Vinci had worked out gravity not only pulls things down but, far more fundamentally, actively accelerates the downward pull at a particular rate—the key to understanding the entire phenomenon! And it was knowledge unlocked as a side-project by an artist a century and a half before Newton. According to the physicist who noticed:
… with this drawing da Vinci managed to estimate what is known to physicists as the “gravitational constant” within ten percent of its actual value, despite only conducting what appears to have been a crude experiment.
However famous da Vinci is now, he was close to being far more so.
But it was in the air. And just as much so for all the contributing lesser-known figures only historians now remember. E.g., “The polymath in the age of specialisation” in Engelsberg Ideas lists many polymaths who are not household names, along with their expansive expertises, like Athanasius Kircher (“ancient Egypt, acoustics, optics, language, fossils, magnetism, music, mathematics, mining and physiology”), Olaus Rudbeck (“anatomy to linguistics, music, botany, ornithology, antiquities and what we now call archaeology”), Benito Jerónimo Feijóo, a Spanish monk (“known in his day as a ‘monster of erudition’ in the 17th-century style… The nine volumes of his Teatro crítico universal dealt with ‘every sort of subject’”), Pierre Bayle (“wrote mainly about theology, philosophy and history, but also edited a learned journal, the Nouvelles de la République des Lettres, and compiled an encyclopaedia, the Dictionnaire Historique et Critique, in which the footnotes took up more space than the text because he filled them with critical remarks of his own”), Comte de Buffon (“remembered today for the huge enterprise of his Histoire Naturelle, published in 36 volumes, was also active in the fields of mathematics, physics, demography, palaeontology and physiology”), and so on and on.
The cause of death of the polymath is universally agreed-upon: increasing demands of specialization rendered polymathy impossible. There became too much to know. “Whatever happened to the polymath?” in UnHerd summarizes the postmortem:
The decline of polymathy… is a crisis of too much information. The seventeenth century was a “golden age of polymaths”, as explorers found new regions, the scientific method flourished, and the postal service and the proliferation of journals allowed scholars to trade ideas. But those same forces led to “information overload.” Over the next 200 years, the intellectual world divided between the specialists who knew a lot about their little area, and popularisers who knew a little about a lot.
Today, a “polymath” is at most someone who studies a few separate areas of mathematics—or who can both direct a movie and act in one. But in just 1870, it meant someone like Lewis Carroll, who you likely know as the author of Alice's Adventures in Wonderland, yet who was also a poet, a photographer, a literal deacon, and he wrote nearly a dozen books on mathematics spanning algebra to probability to logic, and he invented “word ladder” puzzles.
As Peter Burke, a historian at Cambridge University and author of The Polymath: A Cultural History, lamented a few years ago:
Younger polymaths are becoming more difficult to find… I am unable to identify any who were born after the year 1960. Will the species become extinct?
To answer Burke, there are a few names I could give from after 1960, but in general he’s correct. If da Vinci were resurrected to look upon his intellectual descendants, modern thinkers would surely strike him as species warped by the pressures of hyper-specialization. Our academia is now populated by minds pressured into shapes as niche as hummingbirds (upon whom evolution forces the dubious delicacy of only eating nectar) or fig wasps (destined to birth their young solely within the sticky-sweet confines of a single species of fruit, one which they presumably call, in their native Hymenopterian, simply “mother.”)
So is polymathy’s decline just a law of intellectual nature?
To put it simply: there are two kinds of thinkers. Those rate-limited by expertise, and those rate-limited by creativity. Slowly but consistently, the rate-limiting factor for intellectual contribution has become ever deeper expertise.
