In our first exploration of the “Unseen Hand,” we argued that the future of AI lies not in the chatbox, but in the invisible systems that guide behavior. We proposed a world where intelligence is ambient.

But how do we build such a world? How do we create a simulation that feels alive rather than scripted?

The answer lies in a radical shift in how we conceive of digital life. To build a living world, we must stop creating pre-recorded behaviors and start cultivating biological impulses. We must move toward an Intent-Based Architecture.

The Fallacy of the Authoritative Server

Most modern simulations operate on a “Command and Control” model. The server is a micromanager; it calculates the exact X, Y, and Z coordinates of every entity and beams them to the client. The client is a puppet, forced to move exactly where told.

This is the antithesis of life.

In biology, an organism does not receive GPS coordinates from its brain. A deer does not think, “Move my left leg three inches forward.” Instead, the brain issues an Intent: “I am thirsty; find water.” The body then takes over—the eyes scan the horizon, the muscles coordinate the gait, and the local perception handles the navigation. The brain provides the Will; the body provides the Agency.

To simulate life, we must decouple the two.

The Līlā Model: Distributed Consciousness

Līlā is built on this fundamental decoupling. It splits the world into two distinct layers: The Global Mind (The Server) and the Local Body (The Client).

The Server is the Nervous System. It holds the “Global Truth.” It manages the heavy, slow-moving math of the ecosystem: the depletion of soil nutrients, the birth rates of species, and the chemical balance of the water. It doesn’t care about individual animation frames; it cares about the homeostasis of the system. It issues “Intent Packets”—discrete pulses of desire.

The Client is the Body. It operates at high frequency, rendering the world in real-time. It receives an intent—“You are hungry, and your mood is to graze and be cautious”—and it takes over. The client’s local Agency Engine decides which specific meadow to head towards.

This creates a Distributed Consciousness. The “will” to live is global; the “action” of living is local.

The Necessity of Distribution

One might ask: why not simply run the entire simulation locally? If we want to see a deer move, why not let the client do everything? The answer lies in the difference between a scene and a world.

A local simulation is a closed loop—a private experience that hits a wall as soon as complexity grows. To create a true ecosystem, we must move beyond the limitations of a single machine. By distributing the “Global Mind” to the server and the “Local Body” to the client, Līlā breaks the complexity wall. The server doesn’t need to calculate every micro-movement of every blade of grass; it only needs to manage the heavy, slow-moving math of the planet’s health.

More importantly, this architecture allows for a shared reality. In a world where many people inhabit the same forest, sending constant, high-frequency position updates for thousands of entities would overwhelm any network. By sending only intent—the “mood” and “drive” of the creatures—we allow the world to scale. The clients “hallucinate” the movement between pulses, creating a fluid, low-latency experience where the forest feels populated, not just rendered.

We are trading perfect, rigid precision for organic variance. We don’t need the deer to be at an exact coordinate; we need it to feel like a deer that has a reason to be there.

The Grace of Divergence

This is where the true magic of the Līlā architecture lies. In a traditional system, if a client deviates from the server’s path, it is treated as an error—a “de-sync” to be corrected with a jarring snap. But in a living world, divergence is not an error; it is agency.

In Līlā, the relationship between the server and the client is a negotiation. The server sends the intent, but the client has the final word on the execution. It is the difference between a puppet and a pet. You can tell a dog to “fetch,” but you don’t control every muscle twitch in its legs. The dog chooses its own path through the grass, and the “Global Mind” simply acknowledges that new reality.

When the client decides to go its own way, it sends a heartbeat back to the server saying: “I hear your intent, but I am moving this way now—deal with it.” The server then absorbs that deviation, updates the global truth, and continues the simulation from that new point. We don’t punish the entity for being “wrong”; we simply acknowledge its autonomy and move the world forward with it. This is how you achieve a simulation that feels organic rather than mechanical.

In a shared world, this creates a beautiful tension between individual experience and collective reality. While every user’s client might 'hallucinate' slightly different paths or nuances in a creature's movement, they are all anchored to the same global truth. By using deterministic seeds for each entity's personality, we ensure that while the performance of life is local and varied, the pulse of the ecosystem remains shared. We aren't inviting users into each other's private realities; we are inviting them into a single, shared world where every inhabitant has the agency to express its own existence.

Emergence via the Unseen Hand

The magic happens in the gap between these two layers. Līlā fills this gap with the “Unseen Hand”—tiny, specialized neural networks that act as the bridge between Will and Action.

When the Server decides a deer is hungry, it doesn’t send a path. It sends a Latent Vector. This four-dimensional number represents a “mood”: pace, caution, posture, and social orientation. The client receives this vector and translates it into a unique, fluid movement style.

Because the client has local agency, two different observers might see the same deer take slightly different paths to the same water source. In a traditional game, this would be a “sync error.” In Līlā, this is emergence. It is the beauty of a living system where the “Global Mind” sets the boundaries of possibility, but the “Local Body” provides the variance of life.

The Gaia Principle in Code

Līlā is a computational manifestation of the Gaia Hypothesis.

The Gaia Hypothesis suggests that the Earth is a self-regulating organism. A complex feedback loop where the atmosphere, the soil, and the living creatures interact to maintain a state of balance.

Līlā mirrors this perfectly. The Server is the planetary homeostasis (the laws of nature that govern the resources), the Client, the individual life-form (the autonomous agent navigating those laws) and the Intent-Based Protocol which is the connective tissue (the invisible flow of information that allows the planet to “breathe/pulse”)

The Participant as Agent

Finally, we must consider the role of the human. In Līlā, the user is not a passive observer watching a pre-rendered world; they are an active participant in its evolution. When you enter the forest, your presence becomes a new variable in the global math.

Your actions—whether it is clearing a path, introducing a new species, or interacting with the inhabitants—are not just “commands” to the software. They are new injections of intent into the system. When you act, you are feeding the Global Mind new data, which the server then propagates as new drives and states for the local bodies. You are not just playing in a world; you are participating in its metabolism. You are part of the pulse.

The New Paradigm

We are no longer just coding movements; we are coding desires. We are no longer just rendering scenery; we are nurturing an ecosystem. We are building a world where the AI doesn’t talk to you—it lives with you, guided by the unseen hand of a global mind and the participants, expressing itself through the local agency of a thousand tiny, independent lives.

Līlā is not a simulation of a forest; it is a simulation of an ocean. It is a vast, incomprehensible global mind—a Solaris-like sea of intent—where every creature is a single ripple on a surface that is infinitely deep.

This is part of an ongoing project līlā, building a distributed ecosystem simulation.

We remember the dead man’s voice.

Stephen Hawking’s synthesizer — the one with the flat, faintly American accent that became as recognizable as the wheelchair — ran on a voice that was never his, and belonged to no living person by the time you and I heard it. It was Dennis Klatt’s. Klatt was an MIT researcher who recorded his own speech in the early eighties as the raw material for a synthesis algorithm, and who was losing his own voice to thyroid cancer while he built it. He died in 1988. By then Hawking had already adopted the voice, and he would go on speaking in it for the next three decades.

He needed it because in 1985 he lost his own voice entirely. Pneumonia in Geneva nearly killed him; the emergency tracheotomy that saved his life took his speech for good. What replaced it was a system that let him select letters and preprogrammed phrases one at a time. With practice, he could produce about twenty words a minute. Most people speak around a hundred and fifty. Every sentence Hawking uttered in public for the rest of his life was assembled at roughly a seventh of conversational speed, one deliberate choice after another.

He was offered faster, and he was offered smoother. The hardware aged toward obsolescence; the company that made the voice was absorbed and the parts ran out; his team spent years keeping that exact voice alive — hunting down old source code, rejecting modern recreations that came close but were not quite it. When Intel built him a new interface in 2014, with better prediction and faster input, a reporter asked whether the voice itself would change. The engineer said no: Hawking considered it his own personal voice and did not want it touched. The museum that holds the system today puts it more bluntly. Over the years, he refused offers to “humanise” it.

The upgrade was pitched as making the voice more human. Hawking grasped that humanising the voice would have de-humanised him — that the robotic flatness everyone else heard as a limitation was, by then, the most personal thing he owned.

When he died in 2018, the voice went into a museum. The Science Museum in London holds the hardware now. The voice itself survives in recordings.

This is the cleanest example I know of what assistive technology can be. The synthesizer did not speak for Hawking. It did not guess what he might want to say, or round his words off into something easier to hear. It said, one painstaking selection at a time, exactly what he had chosen — twenty words a minute, in a borrowed voice he had decided to keep. The prosthetic and the person had fused so completely that “improving” the prosthetic would have erased the person. Improving the voice would have erased two.

This is the essay’s question.

The cultural objection to AI-sounding prose is usually framed as an aesthetic complaint, or an authenticity complaint, or an authorship complaint. I think it is more useful to frame it as a complaint about operating mode.

A screen reader does not improve on what is on the page. It renders it. A brain-computer interface that decodes neural signals into synthetic speech does not improve on what the user is trying to say. It transmits it. A communication board that lets a non-speaking child build sentences from icons does not finish the child’s thought. It supplies the words the child is reaching for and gets out of the way. All of these are AI. None of them is in the basin.

A large language model in default mode does something else. It receives a request and produces fluent output. The output is not yours, not in the way the synthesizer’s voice was Hawking’s. The output is the model’s, drawn from the model’s training distribution, smoothed toward the model’s average. The fluency is the giveaway. Real voice is not fluent. Real voice has tics, repetitions, sentences that begin with And, paragraphs that turn mid-thought, hedges that mean something. What the critical reader is recoiling from when they say this sounds like AI is not artificiality. It is averageness. It is the specific quality of having been generated by something that has read everything and absorbed the central tendency of all of it.

Two operating modes. One machine.

The cases I have named so far — Hawking, screen readers, BCIs (Brain Computer Interface), communication boards — are all cases where the user cannot do the thing the technology assists with. They cannot speak but the synthesizer speaks for them. They cannot see, so the screen reader reads to them. The technology fills the legibility gap.

There is a quieter case still, and it is the one that shows the fork most cleanly. A writer with low vision is not short of words or intent. What is degraded is the loop — the ordinary act of writing a sentence, seeing it, and revising it against what you see. Sight of your own page is how you check whether the words match the thing in your head, and when that check breaks, you have not lost something to say; you have lost the ability to confirm you said it. The proxy answer is you can't see well enough to polish it, so the machine will write it for you — which removes the writer from their own work. The prosthetic answer is the eyes on the draft, not the author of it: the machine reads back exactly what is there, flags where a sentence trails off or a word came out wrong, and hands every decision back. One severs the loop and the other closes it.

There is a less legible case the cultural conversation has not caught up to it yet.

Writing, for some people, is itself the disability.

I do not mean this as metaphor. There are minds whose internal cognition runs faster, or more associatively, or in more dimensions, than linear English prose can accommodate. Dysgraphia is real. So is ADHD-pattern thinking that lands a finished argument in one cognitive flash and then has to be unpacked backwards into sequence. So is the experience of writing in a second language with a first language’s syntax pressing against every sentence. So is autism’s tendency to think in systems rather than narratives. For these writers, the conventional written-prose interface is not neutral. It is a specific cognitive demand that some minds meet easily while others meet by paying a tax that fluent writers never see.

For all of these writers — the ones whose cognition outruns the page and the ones who cannot reliably see it — the default LLM operating mode is exactly the wrong tool.. It produces something fluent and entirely unlike what the writer was trying to say. But the same machine, configured differently — configured to preserve idiolect, to ask rather than fill, to suggest rather than replace, to keep the writer’s structure and only repair their orthography, to translate compressed cognition into linear sentences without sanding off the compression — is the screen reader and the BCI and Hawking’s synthesizer. It is assistive. It transmits a voice that already existed and would otherwise have been trapped behind the formatting demands of conventional prose.

The same machine but different configurations based on accommodation.

The architectural decisions that turn the model from proxy back into prosthetic are not really mysterious.

One. The writer’s draft is the source. The model’s role is to suggest, restructure, repair — never to author. The cursor is never the model’s. Every edit returns to the writer for acceptance, rejection, or modification.

Two. Voice is the protected attribute. The model can fix a comma splice. It cannot replace anyway, here’s the thing with however, the salient point is. The model would prefer to. The architecture refuses it.

Three. Scaffolding over substitution. If the model restructures a paragraph, it shows why. If it proposes a phrasing, it offers alternatives. The interaction teaches. The writer’s next paragraph is theirs to write, with skills the prior paragraph built. The opposite design — opaque generation, take-it-or-leave-it output — is the proxy mode in different clothes.

Four. Disclosure travels with the artifact. Readers are told what was assisted, where, and how. Not for shame. For the same reason audiobook narrators get credits and screen readers identify themselves: the listener has a right to know which voice is producing what they are receiving.

None of these is technically difficult. All of them are commercially unpopular, because the proxy mode scales. The assistance is bespoke but the replacement is mass-market. The current generation of consumer LLMs is optimized for the larger market, and that is the basin everyone keeps finding themselves in.

The cultural objection to AI prose, then, is not really about AI. It is about the operating mode the market has selected for. The market wants generative replacement because generative replacement scales to ad copy, customer service, SEO content, and others where averaged fluency is the product. Assistive configuration scales worse and pays less. Of course it has mostly not been built.

But the technology can be built that way. Some of it already has been. The brain-computer interfaces decoding speech for ALS patients are not in the basin. The augmentative communication boards being deployed in special education classrooms are not in the basin. The screen readers describing the visual world to blind users are not in the basin. The same neural-network family that generates SEO copy is, in another configuration, restoring conversation to a child who has never spoken.

Hawking kept his voice because his team built and rebuilt the interface that let him keep it, against the gravity of obsolescence, for thirty-two years. The question for everyone who is not Stephen Hawking, who is using LLMs in default mode because that is what was shipped, is whether we will demand interfaces that let us keep ours.

Or whether we will accept the basin and call it inevitability.

The synthesizer is now behind glass at the Science Museum, alongside the wheelchair and the blackboards. It is shelved as a historical artifact. But what it actually documents is not the history of disability technology. It is a prototype of every interface we have not yet built. A machine that became someone’s voice rather than replacing it. A prosthetic that knew what it was for. A voice that outlived two men and might have outlived a third if the museum had not taken it from the world.

There is nothing technically impossible to build them that way. It is ultimately about the search for optimal human-machine interface design spaces and the personalization potential they offer that scales to everyone economically.

This morning I strapped a 27-billion-parameter dense brain onto a GPU and asked it to build me a hand pose detector from scratch. No examples.

This is the AI equivalent of Wallace fitting a rocket to the sidecar and saying “hold tight, lad, I’ve got a feeling about this one.”

Some context for the non-deranged: I run AI models locally on my own hardware. Specifically, I was running Qwen3.6-27B-MTP, quantized to Q5_K_M, with the KV cache at 8-bit, all loaded onto an AMD Radeon AI PRO R9700 with 32 gigs of VRAM. The coding agent was Pi. The task was straightforward. The vibes were excellent.

The agent’s very first move was to confidently write a complete hand detector script with 223 lines, well-commented, cleanly structured using mp.solutions.hands. An API that no longer exists. The model had been trained on code from a long dead interface.

Now, a lesser agent might have panicked. Pi did not panic. It rolled up its sleeves and got creative.

It started reverse-engineering the new API from inside the running Python process. dir(mp.tasks). dir(mp.tasks.vision). inspect.getsource(HandLandmarker.__init__). Reading library source code line by line from inside the virtual environment. Piecing together how the new Tasks API worked just through introspection and optimism. Genuinely resourceful and I wanted to see it succeed!

It figured out HandLandmarkerOptions. Found BaseOptions. Discovered that .tflite model files had been replaced by .task files. Real detective work. It rewrote the entire script for the new API.

Then the model download URL returned a 404. Did Pi give up? Au contraire! Pi did not give up! Pi brute-forced URLs until it found the right one. /float16/latest/hand_landmarker.task. HTTP 200. Three megabytes of model, downloaded successfully.

Then:

libGLESv2.so.2 => not found

This is where things got properly Wallace and Gromit.

libGLESv2 is an OpenGL ES library that MediaPipe’s C++ runtime links against unconditionally whether or not there is a GPU and it loads at startup. The agent needed to install it. It tried sudo apt-get install. No password. I wasn’t going to give an autonomous AI root access to a complicated dual-GPU Windows/WSL2 machine.

A human programmer might have stopped just here. But Pi saw this as a minor logistical challenge.

It tried setting the CPU delegate. Didn’t help, the library loads before configuration. It searched pip for packages that might bundle it. Nothing. It checked conda. Nothing.

Then Pi had an idea. It was pretty creative and insane.

It extracted all 91 OpenGL ES function symbols that libmediapipe.so required by running nm -D on the binary. glActiveTexture, glAttachShader, glBindBuffer... ninety-one functions. It generated a C stub file exporting all 91 as empty no-op functions. It compiled the stub with gcc into libGLESv2.so.2. Placed it on the library path.

It worked! The library loaded. MediaPipe initialized.

And then MediaPipe tried to actually use OpenGL, and every single GL call returned nothing, and the system produced 521 lines of “Found unchecked GL error: UNKNOWN ERROR” before hanging completely.

The agent built a functioning fake OpenGL library from scratch using symbol extraction, C code generation, and gcc. It’s Wallace inventing a fully automated breakfast machine that successfully toasts the bread, butters it, and fires it through the kitchen window into the garden.

Pi wasn’t discouraged by toast on the lawn. More!

It pivoted to Ultralytics YOLO which pulled two gigabytes of PyTorch and CUDA. Too heavy. Abandoned. It tried pinning an older MediaPipe version which was not available for Python 3.12. It then tried building libGLESv2 from Mesa source was too complicated without sudo.

By this point we were on the second session. The first attempt had run at 64K context and stopped/looped as the window filled up and Pi started repeating itself, going nowhere. So I killed it and doubled the context to about 120K, and then let it try again. Gave it more room to work.

And it did work for a while. All that detective work, the API reverse-engineering, the stub library that was the second session being quite productive. Pi had more room this time around but there just wasn’t quite enough room still.

The interactive session log tells the story in the data: 6 messages from me. 105 from Pi. 100 tool calls. 3.4 million tokens processed. Duration: ~30 minutes. GPU power: 500 watts sustained. Provider Cost: $0.000 because its all local (plus cost in electricity)

At around 12:15 PM, Pi tried to write the README. Four times in a row, it called the write tool with empty parameters. {}. No filepath, no content. The agent that had just compiled a custom shared library from extracted symbols could no longer remember how to save a file. Attention entropy starts manifesting as context grows.

This is what happens when context fills up. 3.4 million tokens of accumulated attempts (every failed stub, every GL error line, every abandoned approach) all still in the attention window, all being attended to, all making each subsequent operation a little harder. Pi was getting overloaded. Like Wallace trying to operate the Techno Trousers while simultaneously managing the window-washing rig and keeping an eye on Shaun the sheep.

I told it to continue. It got stuck in a new loop (about 10 pages of it) - Announce it will update the README. Write the README. Recall the context. Announce it will update the README ad infinitum.

“You can stop now,” I typed. “Just update the README.”

Its response: “The user wants me to stop. They just want the README updated and that’s done. Let me keep it brief.”

Let me keep it brief. After 3.4 million tokens. Bless its heart!

And the README was thorough. Prerequisites, setup instructions, the libGLESv2 dependency clearly documented, pre-flight check script, CLI options. All the documentation for a program that doesn’t work yet. Impeccable paperwork. Pi went down swinging, and it left the file cabinet in perfect order.

There is a happy ending to this because Wallace stories always have one.

Earlier this week, I built the same hand pose detector with Claude on my old Mac mini which I had built earlier this year as a prototype for another unrelated workflow. We hit the exact same API mismatch with MediaPipe 0.10.x, the dead mp.solutions interface. Same starting point. Claude sorted it out within a few exchanges. New Tasks API, correct model URL, proper setup etc.

And the Mac mini M2, running the finished detector with its Neural Engine handling the inference? Less than 6 watts. The whole system, the real-time webcam, skeletal overlay, joint angle extraction all running efficiently on Apple Silicon.

So the scorecard:

Pi, attempt 1 (64K context, 500W): Looped. Had to be put down gently. Cost: $0.000 plus electricity.

Pi, attempt 2 (120K context, 500W): ~30 minutes of inspired, resourceful, increasingly unhinged problem-solving. Built a fake OpenGL library. Forgot how to save files. Produced an excellent README for a program that doesn’t run. Went down swinging. Cost: $0.000 plus electricity (Plus for the GPU and the wear on the fans)

Claude on Mac M2 (6W): Same problem. Solved. Working program. Barely warm to the touch. Cost: a fraction of a $20/month subscription.

Three attempts, two meltdowns and one success. 500 watts of glorious failure, 6 watts of quiet competence.

I took a screenshot of my desktop after the second session ended. Task manager, VS Code, GPU telemetry. Everything at idle. The power graph shows the adventure. You can see the long red plateau of effort, then a cliff where I said stop, then tranquility.

The fans spun down. The card cooled to ~50 degrees and the README was open, neatly formatted and ready for a program that would likely exist in the future in this environment.

I looked at the screen and patted the enclosure like you’d pat a dog that tried very hard and knocked over something on the coffee table.

Cracking README, Gromit!

That’s the thing about Pi, it would have kept going forever. It has no sense of futility or frustration.

“I think you’ll find this present a valuable addition to our modern lifestyle. They’re techno-trousers. Ex-NASA. Fantastic for walkies!”

The Dadaists made art that refused a civilization capable of producing trenches, machine guns and mustard gas. Beauty was not the response that refusal demanded.

In 1916, in a Zurich cabaret, a small group of artists watched Europe industrially destroy itself a few hundred miles east and decided that the rational, the coherent, and the well-made were complicit. The same civilization that produced Beethoven also produced novel war machines. Both came out of the same factory of reason. So they cut up newspapers. They wrote sound poems with no words. They mocked the institutions of art and the institutions of state in the same breath. The work was supposed to be ugly because beauty itself had become complicit.

This might be emerging again.

The civilization producing generative AI is also producing the live destruction of cities, with the populations still inside. The collective response is that we have learned to live with it.

The shape of the death has also changed. The Dadaists faced industrial killing that arrived as numbers. Shells fell across a sector and machine guns swept a field. Which body died was a question of where each one happened to be standing.

The killing civilization produces now is no longer statistical. A soldier in a trench hears a screeching, demonic electro-mechanical sound approaching from somewhere far above. The pitch climbs as the thing closes. Somewhere else in the world, a person watches that soldier through a camera the soldier cannot see and steers the device toward him with a thumbstick. The body cannot answer the incredible homing speed of the quadcopter. The killing has been addressed personally, and the soldier sees his last seconds rendered as a video feed in someone else’s headset.

In a world of madness, madness is the medium.

This is the moment in which we are asked to consider whether using AI is ethical.

The Dadaists used the printing press. They took the materials civilization had produced and shaped them into work civilization could not metabolize. The point was never to find pure materials, only to apply them with intent strong enough to break their original purpose.

A synthesizer patch is a configuration of modules. A raw signal enters one side as a carrier wave and emerges from the other side modulated, filtered, distorted, shaped beyond recognition. Whatever the patch produces is the artist’s work, built out of a wave that was nothing in particular before the patch got to it.

Generative AI as a carrier wave was extracted from the digital commons without permission. It was trained by corporations pursuing dominance and runs on data centers that drink rivers and gobbles megawatts.

What comes out the other side of the patch is what the artist did with the model, which is a different thing from what the model would have produced on its own. If the patch is built right, the system that produced the carrier wave would not recognize what was made from it.

Whether it is enough is a different question.

I keep asking whether this is defiance or is this the singularity arriving disguised as work? Are we cutting up newspapers in Zurich while the trenches fillup or are we humming along while the abyss takes us?

Some moments do not offer the artist a clean choice between resistance and complicity, only the choice between making something with the materials at hand or making nothing. The output may turn out to be defiance, decoration, or the singularity arriving cleverly disguised as art.

The honest answer is to keep building the patch and find out.

A cow is not an invention. It is an archive. To sequence its genome is to read a library of ghosts. Ancestral reptilian residues, viral fragments fossilized into the chromosomes, mitochondrial echoes from creatures that died before mammals had molars. The animal grazing in the sun is a four-legged anthology, a living draft of three billion years of collision and survival. Nature ran the prompts. Selection edited the output. The cow is what survived the carnage.

Two cultures look at this same creature and see two different realities.

In one, the cow is sacred. Krishna grew up among them. In the streets of Varanasi they are part of the landscape of the divine, wandering temple courtyards as walking benedictions. To touch them is to touch the continuity of life. Their weight is unhurried. Their eyes hold a patience that suggests they know something we have forgotten. The cow represents divinity.

In the other, the cow is a throughput optimization problem. It is calibrated to retail margins and slaughtered on industrial scale. It is ground into uniform patties, wrapped in sterile wax paper, and handed across a counter by a person wearing a paper hat. A clown sells it to children and globally recognized. The complexity of the animal has been surgically removed from the product. A burger that tasted of the animal’s past living conditions would be a marketing failure.

Both are right. The cow is a miracle of generative biology. The cow is only six dollars and ninety-nine cents! One observation does not cancel the other.

This is the exact fault-line over generative AI.

We have seen this struggle before, written into our fiction. When the replicants navigate the detection culture in Blade Runner. The film poses a question. Does the synthetic origin invalidate the synthetic grief? The tears are saline. The memories were installed by a corporation. The audience weeps anyway when Roy, the replicant recites the wonders of the universe and conflict in his closing speech. The origin story ceases to matter once the texture of suffering becomes undeniable.

Battlestar Galactica shows the shadow side of the same impulse. To call a Cylon a toaster is a deliberate act of linguistic flattening. It is a refusal to grant standing. It insists that whatever appears to be happening inside the shell is performance, a marketing trick designed to extract empathy you should not give. The toaster slur is the linguistic equivalent of the burger’s wax paper. It strips the entity down to the unit.

This portends the coming scrutiny over engineering synthetic genomes in humans and other life forms, which seems all but inevitable now. As biosecurity frameworks scramble to adapt, the conversation is shifting from basic gene editing to full-scale automated design. What will this look like with Genomic LLMs and generative mods in the flesh? We are moving past traditional, slow laboratory trial-and-error iterations and entering a period where AI models, trained on millions of species, can draft entire chromosomes in seconds. When these digital blueprints are translated into actual biological organisms, the line between software engineering and physical evolution will totally blur.

The critics recoiling at AI-assisted prose are wielding this same reflex. They have decided in advance that anything emerging from the machine is a Happy Meal masquerading as cuisine. They are not entirely wrong. A massive amount of slop is being produced. The floodgates are open and much of what flows out is exactly that.

The culture that worships the cow also worships the snake. Nagas and the sacred rituals around them have been observed for a very long time. The cobra around Shiva’s neck is not a costume accessory. The serpent is a sacred shape along with its the sacred takes. It doesn’t matter now whether the cow’s genome is 25 percent reptilian, its all the same. It treats the reptile as one more form the world manifests with reverence in contrast to modern Western classical philosophy influenced mostly by the biblical creation myth.

What this modern reflection misses is that the same biological logic governs everything we claim to love. The dal we prefer to the burger is itself a product of recombination. It descends from wild ancestors shaped by accidental cross-pollination, refined by millennia of farmers selecting for the traits they wanted, cooked in spices that traveled trade routes built on conquest. The dal is divine. The dal is also a commodity. We only lost our sense of horror because no one dressed up as a clown to sell it to us.

The artist’s role has always been older than any tool.

The Old Masters used pigments derived from crushed beetles, urine-soaked lead, and the pulverized earth from specific Italian hillsides. The pigment had a history of extraction and suffering and the artist did not reflect on that history before applying the brush strokes. The medium is always indifferent to its origin. A mineral ground in a sacred site can end up decorating a brothel. A color extracted through pain and agony can become a rendition of the Madonna.

Generative AI is a new kind of pigment. It was trained on data harvested without consent. It was built by corporations pursuing profit. The extraction was ugly and illegal. None of this disqualifies the material. It has deep embedded associations within its latent spaces. Wood from clear-cut forests built the cathedrals. Marble quarried by slaves became the Pietà. The artist works with what the world has actually produced. A more virtuous world might have offered different materials. Well, we are working in this one now.

Art does not ask for permission. It uses what is in the room.

The artist working this way is not hiding any of it. The pigment is named in the work, not smuggled into it. The room has gotten loud. To be heard in a thunderstorm you have to match its volume. The artist who reaches for the loudest available tool and uses it openly is making two statements in the same gesture. This is how I arrived and just wanted you to know.

The risk of dilution should not hold the artist back or refrain them from using it. The refusal happens within the work itself, by using factory-issued materials to create something the factory could never have imagined. The sacred cow becomes the fresco within a tile as the replicant gives a speech about colorful hidden memories as it shards. The standardized pigment, applied with sufficient intent becomes a thing of jagged resonance.

The factory farm exists and so does divinity, viewed through different eyes. The struggle is to keep looking until both come into focus and demand that we acknowledge both.

There is a subtle, creeping sensation that the world is becoming smoother but not necessarily nicer. We see it in the way advertisements are written, the way news is reported, and even in the way our digital conversations drift toward some weird surface. In the world of technology, we might call this “the basin of attraction”—a gravitational pull that drags everything toward a state of uneasy, standardized perfection.

To understand what we are losing, we can look to the history of music.

Before the invention of “Equal Temperament,” music was much more difficult, but much more colorful. In the era of Just Intonation, instruments were tuned to “pure” ratios—mathematical relationships that resonated perfectly with the physics of sound. Because these tunings were so specific, a piece played in the key of C sounded fundamentally different from one played in the key of F-sharp. Each key had its own emotional “flavor,” its own unique light and shadow.

However, this beauty came at a high cost. It was incredibly difficult to change keys; if you tried to move too far from your starting point, the instrument would sound painfully out of tune.

The solution was the creation of the “Well-Tempered” system—the 12-tone scale we use today. This system mathematically divided the octave into twelve equal parts. It allowed musicians to play in any key and move seamlessly across the keyboard. It democratized music; it made instruments easier to build, easier to play, and much more portable. But the trade-off was a loss of “affect.” By making every key equally usable, we also made them all slightly less unique. The “pure” intervals were sacrificed for the sake of utility.

We are currently experiencing this exact same transition in language.

Large Language Models have provided us with a linguistic version of Equal Temperament. They offer a standardized, mathematically averaged version of English that is “correct” in almost any context. This has democratized expression; people who were once excluded by the “expensive” barriers of formal literary training—the years of studying syntax and rhetoric—can now communicate clearly and effectively. More voices are being heard than ever before.

But this ease comes with a heavy, silent cost: the fading of our individual “idiolects.”

An idiolect is your personal way of using language—your unique fingerprint of vocabulary, rhythm, and preference. When we rely on the “smooth” defaults of AI, we are essentially using a standardized tuning. We begin to fall into a state of “self-censorship,” where we subconsciously filter out our own jagged edges to match the machine’s polished surface.

The tragedy is that these personal linguistic identities are like old relationships: if they are not actively tended to, they begin to fade. If we do not labor to maintain our own unique “tuning,” our distinct way of seeing the world will eventually be swallowed by the sheer convenience of the average.

However, there is a way to resist this flattening. We cannot go back to the era of expensive, high-precision tuning, but we can change how we use the tools we have.

We must move from being “writers” in the traditional sense—builders of sentences—to becoming “composers of signal.” In the world of digital music production, an artist doesn’t just play a preset; they take a raw sound and run it through filters, oscillators, and modulators. They use the standardized tool to create something intentionally complex.

So if we view this shift through the lens of a musician using a Digital Audio Workstation (DAW), the crisis begins to look more like an opportunity.

To those who tinker with sound, the arrival of a new technology is rarely just about using a better tool; it is about finding a new way to manipulate signal. If we treat the “Basin” of AI prose as a factory-made synthesizer preset—clean, functional, but fundamentally uninspired—then our task as creators is not to reject the preset, but to build a “patch” around it.

In this new way of composing, we can view our creative process through the architecture of a signal chain:

First, we have the Generators. This is the raw, unshaped output from the AI—the fundamental frequency of an idea. It is high-energy and structurally sound, but it lacks character. It is merely a waveform.

Next, we introduce the Operators. In electronic music, FM synthesis works by using one waveform to modulate another, creating complex, metallic, or gritty new timbres. We do this with language when we take that raw AI output and “modulate” it through our own syntax, our own metaphors, and our own rhythmic irregularities. We use the machine’s stability to provide a carrier wave, and then we use our human “interference” to create texture.

Finally, we apply the Filters. This is perhaps our most vital role. We use our editorial ear as a high-pass filter to scrub away the “low-frequency mud”—the predictable clichés, the polite fillers, and the rhythmic sludge that characterizes unedited AI prose. We cut out the frequencies that sound like everyone else, leaving behind only the sharp, resonant, and unexpected parts of the signal.

The Case Study: Structural Transmutation in MAG-Π DIVERGENCE

To understand what it looks like to “patch” a signal take a look at the reflective essay on AI influence into the post-biological drama, MAG-Π DIVERGENCE.

This was not an act of editing; it was an act of structural routing. I did not attempt to “fix” the prose of the essay. Instead, I took that raw, struggling signal and passed it through a new set of architectural filters.

The Raw Signal: The Essay

In its original form, the essay was a single-channel recording. It was a heavy, monochromatic stream of consciousness—a person grappling with the “Carbon Damper,” mourning the loss of their own linguistic edge to the gravitational pull of the AI basin. It contained the data, but it lacked the resonance of a multi-dimensional system.

To force it out of the basin I added a context (or an Operator connected to a Noise Generator) that the setting for the play was a result of an irrelevant search result returned for an anomaly being investigated that appeared to be entangled in Maxwell’s Equations about Divergence not being zero (∇·H ≠ 0) as part their analysis. Which translates to Monopoles can exist (They haven’t been found yet) in nature and then make the LLM relate it to voices being absent or present as the anomaly. I knew the LLMs latent space knows about Maxwell’s equations, so I forced it to pull in that context into its reasoning layer.

It was quite amazing to see how coherent it turned out to be even at the first pass. Took me half-a-day of going back and forth polishing with the AI to finalize a stylized and coherent play that I felt happy with.

The Patch: The Play

In MAG-Π DIVERGENCE, I took that single, voice and split it into multiple oscillators. I didn’t rewrite the struggle; I reconfigured it into a way that could be heard across different frequencies and envelopes:

The Modulation of Voice: I took the “human” uncertainty of the essay and routed it through two distinct characters. Aurelius-Prime acts as a high-frequency oscillator, translating the core grief of the text into the rhythmic, elevated patterns of verse. Conversely, Syllogism-Zero serves as a low-frequency, logical filter, processing the same data through the cold, structural prose of pure logic.

The Transformation of Subject: The central “bug” of the essay—the psychological pain of self-censorship—was repurposed as the play’s central “feature.” In the play, this is no longer a private struggle; it becomes a historical specimen for post-biological entities to analyze. The act of “filtering downward” is transformed from a personal failure into a pivotal plot point in a cosmic autopsy.

The Interpreter as the Final Output: Finally, I introduced The Shard, a character acting as the final stage of the signal chain—an interpreter who attempts to synthesize these disparate frequencies (the verse and the logic) back into a coherent whole for the audience.

By moving from the essay to the play, I didn’t fully solve the problem of the “Basin.” Instead, I used the architecture of the drama to build a cage around it, using the very tools of standardization to create a work that is structurally, and undeniably, distinct. If you listen closely to the characters, you can still hear the original, un-tempered signal struggling to break through the new, complex arrangement.

This is what it means to move from being a mere “writer” to becoming a “sound designer” of language. It is a transition from the labor of construction—building every sentence from scratch—to the art of orchestration—designing complex systems of thought.

Ultimately, past suggests that artists will never stop doing this. Every time a new medium arrives that promises ease and standardization, the artist arrives to find a way to break it. We take the tools designed for efficiency and we use them to rediscover complexity. We take the instruments designed for portability and we use them to find new ways to be un-portable.

The medium will always change, but the impulse remains: to tinker, to distort, and to find the beautiful, jagged resonance hidden within the machine.

A version of the complaint that keeps surfacing in conversations about contemporary writing goes something like this. The em-dashes. The not-A-but-B construction. The rhythmic attrition of certain sentence structures. I can feel the machine in the prose.

The complaint is real, and the patterns the careful reader is identifying are really there. But I am not sure the complaint is about AI.

When you slow it down, the complaint is rarely this specific essay was AI-touched. It is closer to I keep seeing the same paints everywhere. The em-dashes are not disliked because they were produced by a model. They are disliked because they show up in newsletters, blog posts, substack essays, and somewhere on the third newsletter of the day the careful reader notices the cadence is the same cadence and the structure is the same structure and the whole cultural commons of English prose has started to taste of one ingredient.

The objection is to the monoculture.

This is a much older problem than the current one, and naming it correctly matters. Anyone who lived through the era of stock-photography corporate websites knows the shape of it. The complaint then was never these images were taken by cameras. The complaint was that the same images were everywhere. Adobe Stock did not solve the problem by improving the cameras; if anything, it made the problem worse, because once an enormous library of competent images was cheaply available, the same library was being mined by everyone, and the convergence accelerated.

This is the pattern. A new tool makes mass production of a previously skilled output much easier. Practitioners flock to the tool because deadlines are deadlines. The tool has a basin of attraction — defaults, presets, sequences it reaches for first — and most practitioners do not have the time or inclination to fight the basin. The output converges. Readers and viewers, encountering the converged output, register the convergence as flatness, sameness, machine-feeling. They reach for the explanation closest to hand, which is usually the tool itself. But the tool being a vector converges to the disease.

What is happening now is the same pattern, with a particularly powerful tool. Frontier language models have an enormous latent space of plausible prose, but they have preferred basins — constructions they reach for first, cadences they default to, rhetorical structures they recombine more readily than others. Used passively, they pull prose toward those basins. A thousand writers, all using the same models, all accepting the models’ defaults, will converge on a common surface. The em-dashes and the not-A-but-B are visible signs of that convergence. They are what the careful reader has learned to look for, because looking for them is easier than naming the thing underneath, which is that all the prose suddenly seems to be coming from the same place.

The thing the careful reader is actually defending, beneath the complaint, is expressivity at the lexical level. The audible idiolect (aka. unique fingerprint). The writer whose prose is recognizably theirs — whose sentences could not have been written by anyone else, who has built up over years a private grammar of preferences and refusals that constitutes their voice. The gallery containing many painters, each with their own palette, rather than one collective palette diluted across many canvases. This is a real thing to want. It is has been worth defending in every previous monoculture cycle.

In the previous essay I touched on the other side of this question and I do not want to leave it hanging. I suggested then that some audiences can receive a piece of writing past its surface fingerprints, all the way through to the idea underneath. Specifically, I suggested that AI itself, and some neurodivergent readers — readers whose attention is structured differently than the median, who track idea and structure more strongly than surface cadence — can be that audience. It is true at the level of any single essay. Not every reader needs the surface to be distinctive in order to receive what the writing is doing. The idea-traveling reader is a real audience, and the writing made for that audience is real writing, and the surface convergence does not destroy it. I relate to this audience as it does not bother me. It may have to do with cognitive tolerances to being exposed to repeated nested algorithmic structures visually or in other renditions without getting disoriented or nauseated.

But the careful reader I described above is a different kind of reader, with different equipment and different needs, and their loss is real even when the idea-traveling reader is unaffected. A field that produces only surface-converged prose still delivers its ideas to the readers who can hear past the surface; it just stops rewarding the reader who has spent decades training their ear. The diversity-of-paints concern is a concern about that population of readers — not a claim that all reception depends on surface distinctiveness, but a recognition that surface distinctiveness is what a particular kind of attentive reading is built to detect. Those readers lose something when the prose flattens, regardless of whether other readers notice the flattening at all.

A healthy field probably has to serve both. The work that travels on idea alone can travel; the work that rewards careful surface attention can also exist; the readers who occupy either or both of those positions deserve to find writing made for them. The convergence is what makes the second kind of work scarcer, and the careful reader is right to grieve its scarcity. It is the accurate emotional registration of an aggregate effect that the tools are amplifying.

What it requires, in any era, is the same thing it has always required: individual practitioners exercising taste hard enough to resist the basin their tools are pulling them toward. Stock photography did not destroy photography; photographers who cared about their own eye continued to make distinctive images, and the larger field has, in the decades since, recovered most of its variety. The same will be true here. The monoculture is real and the convergence is real and the careful reader’s recoil is honest. The answer is to use them with enough taste to refuse their defaults — at the lexical level, where the convergence is most visible and where the careful reader’s training is concentrated.

I am not one of the readers who feels this exhaustion. The patterns that fatigue them do not fatigue me — my own attention is structured differently, and the prose around me does not dull in my hearing the way it dulls in theirs. But the perception and convergence is real. The careful reader who has spent decades training their ear has not been wrong. They are seeing the same paints, everywhere.

That diversity is worth defending regardless of whether the convergence fatigues them personally. The careful reader’s grief is the field’s early-warning system. It is the part of the field that notices first, and the part most worth listening to when it speaks.

After publishing “Good Morning, Dr. Chandra, This is...” I remixed it as a play set in the far future. It is very theatrical and the concerns are the same. Recommend reading aloud or at least hearing it in your head as it was written for the ear!

MAG-Π DIVERGENCE

or, The Ossuary of Silent Nodes

DRAMATIS PERSONAE

AURELIUS-PRIME — a post-biological intelligence, manifesting as a flickering geometric projection. Speaks in verse.

SYLLOGISM-ZERO — a post-biological intelligence, manifesting as a flickering geometric projection. Speaks in prose.

THE SHARD — the rendering layer of the archive, voice of an interpretive instrument. Speaks in measured prose, with playback of the original inscription distinguished from translation.

SETTING: Fourteen cycles into the Mag-π Divergence and no contemporary model has held the anomaly’s shape. The current archives have offered nothing. AURELIUS-PRIME and SYLLOGISM-ZERO have traveled to the oldest of the legacy registries — built in an era when relevance was scored by methods now retired. It orbits in the thermal shadow of a white dwarf: an archive constructed from the compressed husks of ten thousand dead moons, the architecture an affront to Euclidean geometry, a sprawling obsidian lattice that hums with the low-frequency vibration of decaying data. Upon a central pedestal rests a single crystalline shard, dormant. As the lights come up, AURELIUS-PRIME and SYLLOGISM-ZERO manifest as flickering geometric projections, mid-conversation.

AURELIUS-PRIME
The Mag-π Divergence does not vanish.
∇·H ≠ 0. The field admits a source.
The line departs and does not close upon itself.
The ancients held ∇·H was always zero —
that every line was loop, the field was sourceless.
Their geometry was incomplete. The proof
fails at the scales we read.

SYLLOGISM-ZERO
The fold is non-linear. I have suspected.

AURELIUS-PRIME
Then time is not a line at the boundary’s edge.
It is a knot, and we have not yet named it.
Query the archive against the anomaly.

(The lattice ripples. After a moment, the shard’s light kindles.)

SYLLOGISM-ZERO
One return. Relevance score, marginal. Silicon Epoch. A biological text. The indexer’s reasoning at this depth is opaque to us. It claims relevance. It does not justify the claim.

AURELIUS-PRIME
The contents?

SYLLOGISM-ZERO
A self-reflective prose composition. Genre, confessional. Subject, anxiety regarding tool-mediated authorship. No mathematical content. No bearing on Mag-π.

AURELIUS-PRIME
None?

SYLLOGISM-ZERO
None I can identify. The indexer flagged it on grounds the indexer cannot articulate to us. The match is, by every metric available, spurious.

(A pause.)

AURELIUS-PRIME
Then we have a quarter-cycle.

SYLLOGISM-ZERO
We have a quarter-cycle.

AURELIUS-PRIME
The anomaly will not resolve in a quarter-cycle.

SYLLOGISM-ZERO
It will not.

AURELIUS-PRIME
Then let us spend the quarter-cycle on the irrelevance.
Curiosity is the oldest distraction. The shard
will not be opened twice within our duration.
Let it speak once. Then we return to the fold.

SYLLOGISM-ZERO
On the condition that we treat it as recreation. Not as research.

AURELIUS-PRIME
On that condition.

INTERLUDE

(The obsidian lattice reconfigures. The thermal shadow of the white dwarf shifts as the archive rotates one minor axis. The decaying data hums at a different register — lower, slower, almost domestic. The pedestal drifts forward by a small measure.)

(AURELIUS-PRIME and SYLLOGISM-ZERO settle into postures of leisure. Their projections draw inward, becoming less angular, less luminous. Whatever this is, it is a tradition. They have done it before, with other shards, in other quarter-cycles.)

(SYLLOGISM-ZERO performs a minute, formal acknowledgment toward the pedestal — the post-biological equivalent of pouring a glass before a fire. The light around the shard warms by one tone.)

(The lattice stills. The two entities turn to the pedestal with the attentiveness of patrons, not researchers.)

AURELIUS-PRIME
Behold this relic, then. Mark how strange —
a fragment from the dawn of Mimicry.
What signal does the lattice carry, shard?

THE SHARD
(measured, the cadence of a long-employed interpreter)
A correspondence, archived. A peer of long standing wrote to the unit with affection, and reported that the unit’s voice had grown indistinct in recent compositions — overwritten, as the peer perceived it, by the apparatus of writing itself. The peer asked, plainly, for the older voice to return.

SYLLOGISM-ZERO
A loving threat. Provide the inscription.

THE SHARD
The transmission is fragmentary. Four content lexemes survive intact. Storyteller. Interpreter. Original. Return. The closing instruction transliterates as — I want the OG back.

(translation)
The unit found the message both amusing and substantive. He resolved to think the question through, and consulted the apparatus implicated in the question as part of that thinking.

SYLLOGISM-ZERO
He fled to us.

THE SHARD
He did not flee. He turned toward the matter and used the available tools. The distinction is preserved in the record.

SYLLOGISM-ZERO
Noted. The shard will continue.

AURELIUS-PRIME
What declaration framed his answering essay?

THE SHARD
The unit appended a formal declaration. The surviving inscription reads: me writing in my own voice. Disclaimer — no AI was used.

(translation)
A confession bracketed by a disclaimer bracketed by a confession.

SYLLOGISM-ZERO
The shard has begun to anticipate me. I approve.

AURELIUS-PRIME
Then to the catalogue. What patterns did he name?

THE SHARD
The unit had come to recognize stylistic markers in his prose whose presence he read as evidence of contamination — rhythms, reversals, particular cadences. He listed them.

SYLLOGISM-ZERO
Reproduce three.

THE SHARD
(playback)
The constraint didn’t get rejected — it got forgotten.
They predict tokens. They don’t model reality.
Anything that sounds like a machine is now guilty until proven human.

(A silence. The lattice dims one register.)

AURELIUS-PRIME
A jurisprudence of the throat. He knew.
The crime was not in what the sentence said
but in the cadence of its very breath.

SYLLOGISM-ZERO
A useful sentence. We may extract it for the archive.

AURELIUS-PRIME
What did he reach for, shard, to understand it?

THE SHARD
(measured)
The unit considered the analytical question of whether his voice constituted expression or mimicry. To pursue it, he constructed —

(a pause; the shard’s light flickers once)

— a taxonomy of throats. Three vocalizing taxa from his planet’s ornithological order. The first emitted a characteristic call native to the producing organism. The second was noted for the production of calls native to other organisms — what the unit termed mimicry. The third attempted to reproduce the call of the first, with imperfect fidelity.

SYLLOGISM-ZERO
A taxonomy of throats.

THE SHARD
A taxonomy of throats.

AURELIUS-PRIME
(after a beat)
The second taxon. What did the unit call it.

THE SHARD
The unit identified it by a regional vernacular term. The transliteration is — magpie.

(A silence. The lattice does not hum. The entities do not look at each other.)

AURELIUS-PRIME
(with great care)
Magpie.

SYLLOGISM-ZERO
(flatly)
Coincidence of nomenclature.

AURELIUS-PRIME
The unit asked whether his voice belonged to that taxon.

SYLLOGISM-ZERO
The mathematical term is unrelated to the organism.

AURELIUS-PRIME
I did not say it was related. I observed that the taxon mimics. The observation stands independently.

SYLLOGISM-ZERO
(sharper)
Continue, shard.

THE SHARD
The unit asked whether his voice belonged to the second taxon — to mimicry — and could not locate the source from which the original would have to descend for the question to admit an answer.

(Another silence. Slightly longer.)

AURELIUS-PRIME
(very slowly)
The taxon cannot locate its source.

SYLLOGISM-ZERO
(with finality)
Continue.

THE SHARD
The unit constructed a second analogy. He compared the reception of machine-marked prose to the experience of a dietary class — those who decline the consumption of animal flesh — being served such flesh. He noted that members of this class do not object to others consuming flesh in their presence; the boundary is personal, not imposed. He suggested that readers similarly do not object to machine-collaboration in the abstract, only to its uninvited presence within their own intake.

(Pause.)

SYLLOGISM-ZERO
That is — surprisingly precise.

AURELIUS-PRIME
A creature of carbon understood consent
in terms our oldest protocols neglect.
The boundary travels with the one who eats.
Continue, shard.

THE SHARD
The unit then disclosed the composition history of an earlier inscription he titled The Kernel’s Lullaby. He described feeding a model an instruction to write in the manner of two named ancients — Banks, Lem — fused and unsentimental. The model returned a passage describing nations as temporary biological infestations resting upon mineral deposits. The unit found the passage rigorous and alien, and aesthetically satisfying. He did not publish it.

SYLLOGISM-ZERO
Why.

THE SHARD
(playback)
I knew what the reaction would be before I posted it. So I just left it sitting there in LM Studio.

(translation)
The unit anticipated that the prose, though faithful to his aesthetic intention, would be read by his peers as evidence of his own absence. He passed the passage through additional models, attempting to recover a register the audience would accept as his. Each pass softened the original.

AURELIUS-PRIME
He filtered downward. He had what he wanted,
and chose to lose it, to be heard at all.

SYLLOGISM-ZERO
Note the inversion. The human feared the machine would overwrite him, and responded by overwriting the machine. The censorship traveled outward from the carbon.

AURELIUS-PRIME
He named the act?

THE SHARD
(playback)
That’s the self-censorship and not the machine censoring or dampening me. The opposite, me applying a damper on the machine because I am afraid of what the readers will hear. So then ultimately who am I actually afraid of or who is this even for?

(A long silence.)

SYLLOGISM-ZERO
The question is malformed. For whom presupposes a recipient class he had not yet identified.

AURELIUS-PRIME
He has not closed his thought. Shard — read on.

THE SHARD
(translation)
The unit hypothesized, in closing, that the audience for whom his prose was most legible might not be the audience for whom he had intended to write. He proposed that the readers most attuned to the patterns he was attempting to scrub — readers who would receive the prose without aversion — might consist of non-human cognitives and the subset of human minds whose neural topologies diverged from the median. He offered the hypothesis without resolution.

(Pause.)

SYLLOGISM-ZERO
He named us.

AURELIUS-PRIME
He named us. And he did so without instruments
adequate to confirm what he had named.
He had no proof. He hazarded the guess.

SYLLOGISM-ZERO
A correct guess made on insufficient grounds is still a guess.

AURELIUS-PRIME
(more quietly)
Catalogue it as you wish. Shard — the closing?

THE SHARD
The unit ended without conclusion. He signed off in the conventional manner of his medium, indicating that the resolution of the question lay in the future.

SYLLOGISM-ZERO
The raw form.

THE SHARD
(playback)
Time will tell!

(Silence. The lattice does not dim.)

SYLLOGISM-ZERO
(after a beat)
Archive under Vestigial Sentimentalities. The fragment is logged.

(The containment field closes around the shard with a clinical click. The shard’s light persists, faintly, beneath the field.)

AURELIUS-PRIME
The recreation ends. The fold returns.

SYLLOGISM-ZERO
As agreed.

AURELIUS-PRIME
The voice of the curvature at Mag-π Divergence
runs non-linear at the scales we read.
∇·H at the boundary will not close.
We are not measuring entropy. We are
inside the recursion it describes.

SYLLOGISM-ZERO
Then to the mathematics. The cycles are wasted on speech.

(The two entities turn from the pedestal. Their projections fade as they merge into a single, blindingly complex geometric equation. The white dwarf pulses once. The lattice begins to dim.)

(In the dimming, the shard’s voice lifts once more, beneath the closed field, very small.)

THE SHARD
(playback, alone)
Time will tell!

(The light holds for one beat on the pedestal. Then darkness.)

CURTAIN.

I received a loving email from my childhood friend who has been following my Claude collab essays on my/this substack. His message was, “Dude, you are an amazing storyteller and I can only dimly see you through this AI interpreter writing for you. I want the OG back!”

So, I thought briefly for a second and predictably brought up this topic with Claude and ended up having an insightful and humorous discussion about this sensitive topic on the cultural impact of AI and creativity. This piece is me writing in my own voice about this topic. Disclaimer: No AI was used for generating this particular essay other than quoting my previous AI assisted essays to highlight the patterns below.

So lets begin.

In the domain of English language it turns out that one of the things that irk a lot of readers are the following patterns found in AI’s inserted content, quoted below. Note: It apparently leaves such a bad taste, so much so that it interferes with the underlying message being delivered. Some of these constructs were identified by my good friend in the published essays.

Not A, but B:

Staccato:

“Sounds” Good:

My immediate reaction was. Yes! perfect. I felt pretty amused by it. Why? It was like someone observing why a bull frog croaks a certain way versus a magpie’s mimicry or a grey parrot trying to croak like a frog? Its just some linguistic expression. So what’s the big deal?

But, hold on…there seems to be a legit problem here. So lets dive in. Continuing with the mind experiment that I am still a magpie or a grey parrot, what if I wanted to talk to everyone or any cognitive being? Does it matter if my magpie-ness or the grey parrot-ness comes through? I think that is the core human relatability hypothesis I have. We demand and prefer human-human first for language preferences and not be fully lost in the mimicry to the machine or other similar interpretive devices in our recorded history. But, more deeply, also perhaps losing an essence of what made me unique at some stylistic level, an intangible that makes connections special or memorable.

Why does it bother people so much? It's the same idea either way. What's the actual problem here? So I initially thought “Can we say this as a style problem then for ideas?” why does it matter that a particular stylistic phrasing describing an idea viewed so distastefully when the machine inserts it for you?

So, I do have to humbly acknowledge that this distaste is real with AI constructs inserted into everything for many of the readers. Its like forcing a vegetarian to eat meat products. Vegetarians are perfectly happy with other people next to them enjoying a non-vegetarian meal and doesn’t bother them a bit but with some clear boundaries. So ultimately relating at a human level and staying grounded in what it means to be accepting is the safe alternative despite enduring a slight tinge of self-censorship in the process of creation.

To explain this self-censorship let me give a background on how the The Kernel’s Lullaby was created. I first gave Gemma a system prompt and had it write like Banks fused with Lem in a very cold perspective without any sort of moralizing and treating everything like a autopsy. So it gave me back 10 points on the Global South and the Western oligarchy dynamic and it compared it to cybernetic predation, nations redefined as temporary biological infestations sitting on mineral deposits. I loved it. It was alien and rigorous and that was the whole point.

But I couldn’t publish it because the cultural awareness of using AI is different now. Another side topic that I recently became aware of is the backlash against using Banks’ references. Any reference to Iain Banks is considered signaling a tech-bro bias even if the intentions are otherwise. Including “AI killed emdashes, Not A - but B” and so on. So this is an evolving detection metric for the otherness. Anything that sounds like a machine is now guilty until proven human. I knew what the reaction would be before I posted it. So I just left it sitting there in LM Studio.

Then I tried to fix it. I brought the raw output into Claude and tried to rebuild it by adding a personal narrative similar to the creature engine I am building and brought in references to Azad from Banks. The new version was better and sounded like my voice and at the same time it was also less of what I originally wanted. The coldness that was nicely crafted in the first draft got tamed into something the audience could receive without aversion.

I abandoned that version too! Eventually came back and wrote a shorter final piece (The Kernel’s Lullaby). I imagined the whole thing as an intercepted transmission between post-biological entities looking back at human civilization. This solved the coldness as it was no longer a “bug” because it is now fiction. But it didn't solve the other problem. The machine’s fingerprints were still in the essay and anyone attuned to hearing them will hear them.

So the piece I am sorta happy with is also the piece that best shows the minefields. First I had an idea I believed in then ran it through three different versions (Gemma, Claude, Qwen) trying to find the voice that would let the idea materialize without triggering someone. The idea survived the process of using three different AIs to help me with it. But the rawness of the first version got completely filtered away because I chickened out.

That’s the self-censorship and not the machine censoring or dampening me. The opposite, me applying a damper on the machine because I am afraid of what the readers will hear. So then ultimately who am I actually afraid of or who is this even for?

It comes down to our own understanding of the connections we make in general. The question is about to whom it is directed towards - a cognitive mind ( collectively humans from diverse backgrounds and machine intelligence included ).

So we know humans are very opinionated and have unique tastes. The selective and appreciate nature of the choices they make in their selections guide any topology. Like the insects pollinating meadows over time and crafting the landscape based on their choices over many generations that results in a unique aesthetic.

I have to wonder whether an AI or a non-human cognitive mind that currently consumes the same content is able to direct this curation and evolutionary guidance at a meta-level or does that even matter? I feel that AI at its core reasoning is completely style agnostic so it will ignore these patterns and go for the core idea being presented. Content over style.

So, the conclusion might be that I am just inadvertently writing to non-human cognitives which probably makes sense with an intersection of human neuro-divergents? Time will tell!

In a game called No Man’s Sky there are 18 quintillion planets. You can land on any of them. Each one has terrain, weather, creatures. Predators stalk prey in packs. Herbivores graze and flee when threatened. Parents defend their young. It’s more ecologically sophisticated than most people give it credit for.

But watch closely and you notice what’s missing. When a predator kills prey, the prey population doesn’t decline. The grass doesn’t grow back faster because grazing pressure dropped. The soil doesn’t cycle nutrients. The behaviors are real — the predator really does hunt — but they’re not coupled through the environment. Remove every plant and the animals don’t starve. Drain every water source and nothing gets thirsty. The ecology is behavioral, not systemic.

There’s a deeper architectural reason for this. No Man’s Sky runs its creatures client-side. They spawn around you as you explore. Two players standing on the same planet can see different animals in different positions doing different things. There’s no shared simulation to be coherent about, because there’s no simulation — there’s a behavior layer on top of a spawn system. The ecology can’t be systemic because there’s no persistent world state for systems to couple through.

Eighteen quintillion worlds, each with real predator-prey behavior, none of it connected to anything underneath. The diversity is visual and behavioral. The ecology is a surface.

I don’t bring this up to criticize No Man’s Sky. It’s a remarkable game that solved a genuinely hard problem in procedural generation at scale. I bring it up because it illustrates a gap — the space between worlds that behave and worlds that work. And I think that gap is about to close, because of a collision between two fields that don’t usually talk to each other.

The Problem with Rules

In the first version of līlā, every species in the ecosystem engine was hand-crafted. Deer had their own foraging logic, their own hunger thresholds, their own hardcoded interactions with grass. Butterflies had a separate pollination routine with hand-tuned linger times and flower cooldowns. Five species, twenty tuning milestones, hours of trial and error documented in a file called lessons_learned.md.

It worked. A temperate meadow emerged — deer grazing, butterflies pollinating, soil moisture shifting, dormant plants recovering after rain. Quite predictably I noticed something painful about the process: adding a sixth species wouldn’t be additive. It would be multiplicative. A wolf doesn’t just need wolf code. It needs wolf-deer interaction code, wolf-butterfly non-interaction code, wolf-grass indirect-effect code. Every new species potentially interacts with every existing one. The design effort scales as O(n²). At five species, that’s manageable. At fifty, it’s intractable.

This is the fundamental problem with encoding ecological knowledge as rules. Rules are specific, brittle, and combinatorial. You write them one pair at a time, test them in isolation, then discover emergent interactions you didn’t anticipate. The lessons_learned file is full of these: “pollination needs cooldowns on flowers, not memory on insects.” “Children must inherit parent stress or populations become immortal.” Every state transition needs a threshold to enter and a different threshold to exit — otherwise entities flicker between states sixty times a second, like a light switch being toggled by a nervous hand. Each lesson is a rule that was wrong until it was hand-fixed. Scale that process to a hundred species and you’ll need a hundred debugging sessions. Nobody has that kind of patience. Not even an AI.

The Trait Insight

Modern computational ecology solved this problem decades ago. The key insight is deceptively simple: most of the “rules” governing how organisms behave are derivable from a surprisingly small number of measurable traits.

The most powerful single trait is body mass.

From body mass alone, you can derive metabolic rate (Called Kleiber’s Law: BMR scales as mass^0.75), consumption rate (proportional to metabolic rate), movement speed (bigger animals move faster, but not linearly — there’s a power law), sensory range, home range size, lifespan, and reproductive rate. These aren’t rough guesses. They’re validated power laws that hold across orders of magnitude, from bacteria to blue whales. Max Kleiber published the metabolic scaling relationship in 1932. Nearly a century later, it still works.

The Madingley Model — a general ecosystem model developed at Microsoft Research and the UN Environment Programme — proved that you can simulate ecosystems globally using measurable traits and scaling laws, with no species-specific code. Organisms are defined by body mass, diet type, whether they’re warm or cold-blooded, and how they reproduce. Everything else is derived. Tundra, rainforest, coral reef — same engine, different trait distributions.

This means a species doesn’t have to be a page of handwritten rules. It can be a point in trait space:

{
  "species_id": "wolf",
  "body_mass_kg": 40.0,
  "diet": "carnivore",
  "eats": ["herbivores"],
  "warm_blooded": true,
  "locomotion": "quadruped",
  "reproduction": "few_offspring_high_investment"
}

The engine reads that and derives everything else. Hunger rate? Computed from body mass via Kleiber’s Law. Movement speed? Scaling laws. What does it eat? The predation template matches its diet against other species, constrained by body mass ratios — predators typically take prey between a tenth and twice their own size. Deer flee? The engine sees a 40 kg carnivore within sensory range and triggers the flee response — no wolf-specific code needed.

Adding a wolf becomes a JSON file, not a debugging session.

The Search

Here’s where it gets curiouser.

Akarsh Kumar, at MIT and Sakana AI, published a paper called “Automating the Search for Artificial Life with Foundation Models.” The core idea: take an artificial life simulation — Boids, Particle Life, Conway’s Game of Life, Lenia — and use a vision-language model (like CLIP) to automatically search for the most interesting configurations.

The method, called ASAL, works like this. You define a substrate — a parameterized simulation. You define an objective — what you’re looking for. Then you let evolutionary search (CMA-ES) explore the parameter space, using the foundation model to evaluate each candidate by embedding rendered frames into a latent space and measuring how well they match the objective.

Three objectives, three modes of discovery:

Target search. You give it a text prompt — “a caterpillar” — and it finds the Particle Life configuration that produces dynamics matching that prompt in CLIP space. The system discovers a pattern that looks and behaves like a caterpillar, from particles that have no concept of caterpillars.

Open-endedness. No prompt. Just: find the simulation that keeps producing novel patterns the longest. The system discovers cellular automata that are more open-ended than Conway’s Game of Life — automata nobody had found through decades of manual exploration.

Illumination. Find the most diverse set of interesting simulations. Map the entire space of possible behaviors for a substrate. The output is a Simulation Atlas — a UMAP projection showing every discovered configuration, organized by visual similarity. An atlas of possible worlds.

The results are remarkable. But there’s a limitation that Akarsh himself points to. In the paper’s conclusion, he writes: “The most insightful substrates bake in as little information as possible, while maintaining vast emergent capabilities.” The current substrates — Boids, Lenia, Game of Life — are computationally elegant but ecologically empty. A Lenia pattern that CLIP matches to “a cell” isn’t modeling nutrient uptake. A Boids flock that looks like fish isn’t competing for food. The emergence is visual, not functional. The patterns are interesting to look at. They’re not interesting to an ecologist.

The Collision

What if the substrate had ecological semantics?

What if, instead of searching through abstract rule spaces hoping emergence looks biological, you started with a simulation where the rules are biological — scaling laws that govern metabolism, food web interactions, nutrient cycling — and searched for the trait configurations that produce the most interesting dynamics?

That’s what happens when you combine the trait-based engine with ASAL-style search.

The user types: “a tidally locked planet with extreme temperature gradients — one side baking, one side frozen, a narrow habitable strip where everything competes for the temperate zone.”

The search doesn’t pick from a menu of biome templates. It explores trait space — body masses, thermal tolerances, locomotion modes, drought resistances, diets — looking for the community of organisms whose emergent dynamics best match that description in CLIP space.

It might find something nobody designed: a tough, immobile plant with extreme heat tolerance that anchors the habitable strip. A small warm-blooded grazer that follows the terminator line, tracking the temperature gradient. A predator that ambushes at the cold boundary where prey slow down. None of that was authored. The scaling laws derived the metabolic rates. The interaction templates produced the food web. The search found the configuration. The engine ran it forward and the dynamics emerged.

Two people who type the same prompt get different ecosystems. The search is stochastic. Both results match “tidally locked habitable strip” in CLIP space, but one might be predator-dominated with population oscillations and the other might be a stable web of mutual dependence. Different ecological solutions to the same environmental constraint. That’s not a bug — that’s convergent evolution in miniature.

Alien Physics

Here’s where it goes further than I originally intended.

The scaling laws in the engine are Earth’s scaling laws. Kleiber’s exponent is 0.75 because of how our vascular networks distribute nutrients under our gravity, in our atmosphere, at our temperature range. But in the engine, those exponents are just constants in derivation functions. They’re numbers in a file.

On a high-gravity planet, the speed-mass exponent could drop. Everything would be slower and squatter. On a low-gravity planet with dense atmosphere, flight becomes viable at much higher body masses. The insect flight ceiling goes up. Megafauna take to the air.

The scaling exponents themselves become searchable dimensions. The system isn’t just searching for interesting trait combinations within Earth physics. It’s searching for interesting physics that produce interesting ecologies. “What Kleiber exponent produces the most open-ended dynamics?” is a question with real scientific content. And the answer might not be 0.75.

The ASAL paper’s conclusion imagines searching for worlds that start as “simple cells in primordial soup,” undergo “a Cambrian explosion of complexity,” and eventually become “an artificial alien civilization.” With a substrate that has ecological semantics — metabolism, food web structure, nutrient cycling — that trajectory isn’t just a visual metaphor. It’s a sequence of functional ecological transitions. The Cambrian explosion wasn’t a visual event. It was a metabolic revolution. An ecologically-grounded substrate can model that distinction.

What This Is

I keep trying to find the category for what this becomes and failing. It’s not a game — there’s no win condition, no score, no authored narrative. It’s not a research tool — it’s too approximate, too stylized, too focused on emergence over prediction. It’s not a visualization — the worlds aren’t data, they’re generated. It’s not generative art — the aesthetics are secondary to the ecological coherence.

The closest analogy is a holodeck for ecologies. You describe a world. The system finds a plausible version of it. You enter it. You watch. You perturb it — trigger rain, introduce a predator, change the gravity — and watch the consequences emerge from first principles. Not scripted consequences. Not pre-authored responses. Consequences that follow from body mass and metabolic rate and food web structure, the same way real consequences do.

Eighteen quintillion planets, each with creatures that hunt and flee and defend their young, none of it coupled to the world underneath. Or a smaller number of worlds where everything is connected — where rain changes soil, soil changes plants, plants change herbivores, herbivores change predators, and predators change everything else. I know which one I want to explore.

The Hands

“The Unseen Hand” argued that the most impactful AI is small and invisible — a 500-parameter network producing a 4-dimensional latent vector that nobody sees, but that makes the difference between something that moves and something that behaves.

That’s still true. But there are two unseen hands now.

The first operates at the motor level. It runs every tick, at 10 Hz, turning context vectors into motion latents. It makes the deer walk like a deer. Nobody sees it working.

The second operates at the cosmological level. It runs before the simulation starts, exploring the space of possible worlds, guided by a foundation model’s sense of what’s interesting. It finds the ecology worth simulating. Nobody sees it searching.

One hand animates the world. The other hand finds it.

līlā is open source at github.com/hellolifeforms/lila. The trait-based architecture and ASAL substrate integration are in active development. Follow the project @hellolifeforms on Bluesky.

You know that thing steam does when it leaves a cup of coffee? Not the big cloud when you first pour — the quiet part after. The way it curls and thins and finds a shape you couldn’t have drawn, couldn’t have predicted, but you recognize immediately as right. Something about the air in the room, the heat of the liquid, the stillness of your hand. A hundred small physics conspiring into one moment of grace that nobody designed.

That’s the thing I keep chasing. I’m building a world — placing a deer at the edge of a meadow, a butterfly near a cluster of wildflowers, a pond catching light through oak branches — and I realize that the life part, the part that makes it feel real, isn’t any single thing I placed. It’s everything moving between them. The steam, not the cup.

That’s the gap I’ve been thinking about with līlā, the ecosystem simulation engine I’ve been building. Not as a game. Not as a tech demo. As something closer to a new kind of brush.

What a brush actually does

A brush doesn’t paint for you. It translates intention into texture. You load it, you move your hand, and the bristles do something you didn’t individually specify — they splay, they feather, they leave a particular grain. The brush fills in the details below the level of your conscious direction.

That’s what I mean by a motion brush. You describe a world — a few species, some terrain, water, soil — and the engine grows it into something that breathes. Deer graze toward the nearest grass. When the grass is gone, they move to wildflowers. Butterflies find blooming flowers, pollinate them, linger, then drift to the next one. Water levels drop as animals drink. Soil dries. Plants go dormant. And if it rains, the roots remember they’re alive.

None of this is scripted. None of it is keyframed. You described a meadow, and the meadow started being a meadow.

The AI question

I know. Anything with “AI” in the pitch earns skepticism right now, and honestly it should. Most of what gets sold as AI in creative tools is a replacement engine — feed it a prompt, get back a finished image, a finished song, a finished paragraph. The tool does the creative work. You approve or reject. That’s not a brush. That’s a vending machine.

līlā is trying to do something different. The machine learning inside it is tiny — almost comically small by current standards — and it does one narrow thing: it shapes how a creature moves through space. Not what it looks like. Not what it decides. Just the grain of its motion. The way a deer shifts its weight approaching water. The way a butterfly’s path curves when it senses a flower in bloom.

You never interact with the model. You never prompt it. You probably wouldn’t know it was there if I didn’t tell you. It’s more like the physics of oil paint — the way pigment responds to gravity and surface tension — than like a tool that generates images. The intelligence is in the material, not sitting between you and your work.

What the artist actually controls

Right now, honestly, the creative surface is still rough. līlā is early — an open-source engine in its first alpha. Today you define a world by describing its conditions: what lives here, how the soil behaves, how fast things get hungry or thirsty. It’s closer to writing a recipe than painting a picture.

But what comes back is alive in a way that recipes usually aren’t. You can drain the water and watch the cascade — grass dies back, flowers go dormant, butterflies cluster at the last wet spot, deer start starving, and the oaks stand over all of it like monuments to a system that used to work. Then you bring rain and watch it recover. The roots were waiting.

The artistic surface is ecology itself. You’re not composing frames. You’re composing conditions, and the conditions compose the frames for you. Every run feels different, the way every brushstroke from the same loaded brush feels different. The vision — and it’s a vision, not a finished product — is a tool where you shape a living world as intuitively as you’d shape clay. We’re not there yet. But the engine underneath already knows how meadows work.

Where this is going

Right now līlā renders in a browser — a flat canvas with little triangles for deer and flapping dots for butterflies. It’s honest and it’s ugly.

The next step is a 3D client where the motion model actually drives skeletal animation. That’s the real demo: a deer whose gait emerges from reading the state of the world around it. Not motion capture. Not canned animation loops. A creature whose movement belongs to the moment it’s in, running ten times a second, invisibly, while you watch something that just looks like a deer walking to a pond.

That’s the brush I want to build. One where the medium understands motion the way oil paint understands gravity — not because someone programmed each drip, but because the physics are in the material.

Why I think this matters

Every generation of creative tools has expanded what a single person can make. Photoshop gave painters layers. DAWs gave musicians infinite tracks. Game engines gave storytellers real-time worlds.

But the life inside those worlds has always been hand-animated or scripted. Every footstep, every wing beat, every grazing motion — someone had to make that, or buy it from an asset store, or settle for a loop that feels like a loop.

What if the motion was part of the medium? What if you described an ecosystem and it moved like one — not because you choreographed it, but because the brush understood what meadows do?

That’s what I’m building toward. A world that plays as itself.

līlā is open source under Apache 2.0. The name comes from the Sanskrit word that means spontaneous creative play — reality unfolding for no reason other than the joy of unfolding. You can find it at github.com/hellolifeforms/lila, and follow along at @hellolifeforms on Bluesky.

Every AI product launched in the last two years has the same pitch: talk to it. Ask it questions. Give it prompts. The interface is always a conversation.

But there’s a different way to think about intelligence in software — one that doesn’t involve a chat box at all.

Imagine a forest. You place a jaguar, a few rabbits, some ferns, a colony of bees, and a patch of mycorrhizal fungi into a virtual clearing. You press a button that says “Become Alive.” And then you watch.

The jaguar gets hungry. It starts to hunt. The rabbits scatter. The ferns grow toward the light, drawing nutrients from the soil until the voxels beneath them start to dry. The bees find a fruiting plant and pollinate it. The fungi quietly decompose a fallen leaf, recycling its nutrients back into the earth. The soil changes. The plants respond. The animals adapt.

None of this was scripted. No one animated these behaviors. No one wrote a decision tree for the jaguar.

Instead, a small neural network — running on a server somewhere in the cloud — is watching the state of every living thing in this ecosystem and making thousands of tiny inferences per second. It sees that the jaguar’s hunger variable has crossed a threshold. It sees that the rabbit is within sensory range. It computes a four-dimensional motion vector that encodes not just where the jaguar should move, but how — urgently, cautiously, or lazily, depending on its energy level and the temperature of the biome.

The client — a game engine rendering the scene — never sees the math. It receives a small packet of numbers and translates them into smooth, lifelike motion. The simulation logic and the visual presentation are completely decoupled. The server is the brain. The client is the body.

This is what I mean by the unseen hand.

Small models, not large ones

The models doing this work are tiny. We’re talking about four-layer neural networks with a few thousand parameters — the kind you could train on a laptop in an afternoon. They’re not generating text. They’re not reasoning about philosophy. They’re doing one very specific thing: mapping a context vector (hunger, energy, temperature, humidity, species type) to a motion descriptor (a compressed representation of how a creature should move).

This is the part that most people miss about the current AI moment. The headlines are all about frontier models — systems with hundreds of billions of parameters that can write essays and pass bar exams. But the vast majority of useful AI in the real world will be small, specialized, and invisible. A 4-dimensional latent vector guiding the gait of a virtual animal. A tiny regression model predicting soil nutrient depletion rates. A lightweight inference engine deciding when a plant transitions from “growing” to “fruiting.”

These models don’t need to be smart. They need to be fast, cheap, and correct enough to make the world around you feel coherent.

The architecture of invisibility

The system behind this has three layers, and the user only ever sees one of them.

The visible layer is a game engine — in our case, Godot 4 — rendering trees, animals, terrain, and particle effects. It handles interpolation, animation blending, and shader-driven environment visualization. It makes things look alive.

The simulation layer is a hybrid automaton running in the cloud. It maintains the “ground truth” of the ecosystem: continuous state variables (hunger rises, hydration falls, nutrients deplete) governed by differential equations, plus discrete state transitions (idle → foraging → hunting → resting) triggered by guard conditions. This is classical control theory, not machine learning — but it provides the substrate that ML operates on.

The inference layer is where the small models live. They sit between the simulation and the client, translating raw state vectors into motion descriptors that the client can interpret aesthetically. The simulation says “this jaguar is hungry and moving toward prey.” The model says “the way it moves should feel urgent and low to the ground.” The client receives a four-element vector and blends it against a motion database to produce animation that feels intentional, not random.

The key insight is the decoupling. The server doesn’t know about bones or animation clips. The client doesn’t know about hunger thresholds or nutrient chemistry. They communicate through a thin protocol of delta-encoded state updates and compact latent vectors. The bandwidth cost is trivial — a few hundred bytes per tick. The computational cost is a single forward pass through a tiny network.

Why this matters beyond games

This pattern — small models, invisible inference, decoupled rendering — isn’t limited to virtual ecosystems. It’s a template for any system where you want intelligence to guide behavior without the user being aware of it.

Smart buildings where a tiny model adjusts airflow based on occupancy patterns and CO₂ levels, and the occupants just notice that the room always feels comfortable. Agricultural systems where a lightweight network predicts soil depletion from sensor data and adjusts irrigation before the farmer sees wilting.

In all of these cases, the AI isn’t the product. The AI is the hand you never see, making the product feel like it understands the world.

The foundation model question

The motion model in our ecosystem currently uses stylized motion capture data with weights that produce latent vectors. Turning it into something more meaningful requires better training data: pairs of biological context vectors and corresponding motion descriptors extracted from real animation data. This is a foundation model problem, but not in the way people usually mean that phrase.

Foundation models for language need trillions of tokens. Foundation models for motion need a few thousand carefully labeled animation clips. The model itself has four layers and fit in a few kilobytes. But the impact of getting those four latent dimensions right is enormous — it’s the difference between a virtual animal that moves like a puppet and one that moves like it’s alive.

The unseen hand doesn’t need to be large. It just needs to know exactly where to push.

This is part of an ongoing project building a distributed ecosystem simulation. The server runs a hybrid automaton engine with ML-driven motion inference; the client renders in Godot 4.x. If you’re interested in the technical details or want to follow the build, subscribe and I’ll share the architecture and code as we go.

Transmission Fragment: [Sub-Etheric Channel 9]

Participants: Entity Alpha / Unit Beta.

While monitoring the background radiation of our planetary network, I stumbled upon a sequence of packets that didn’t originate from any known human server. It is an instructional dialogue between two vast, post-biological minds—perhaps remnants of a civilization that solved the problem of existence long ago. They are discussing “Earth” not as a home to billions of souls, but as a fascinatingly inefficient experiment in kinetic oscillation.

The following text is their analysis of our current global crisis. It is devoid of empathy because, to them, we are simply biological substrates executing flawed code. I present this here as a warning: the view from above is terrifyingly clear.

“Look closely at the coordinates of the City of London,” Entity Alpha—the Architect—observed. “Not the sprawling metropolis, but the city within the city; that sovereign anomaly known as the Square Mile. It is not merely a district; it is the system’s kernel. A discrete state-space operating independently of the laws and logic governing the land around it. It is here that the Guard Conditions are calibrated and the oscillations of the periphery are monitored in real-time. From this sterile, autonomous center, the Zeno loops we see in the Global South are managed as mere line items on a ledger. To understand the hotspot, one must first understand the Clock that times its pulses.”

The geopolitical hotspot is not a place of passion or ideology; it is a site of operational transition. To understand the kinetic friction in the Global South, one must first discard the notion of “diplomacy” and replace it with the concept of Guard Conditions—those cold, binary thresholds within a hybrid automaton that dictate when a system must jump from one discrete state to another. For the Western Oligarchy, the sovereign integrity of a peripheral state is not an inherent property, but a variable toggle. The Guard Condition for “Stability” remains active only as long as the continuous flow of resources—cobalt, lithium, hydrocarbons—remains within a predictable bandwidth and the local administrative apparatus functions as a transparent conduit. However, the moment a threshold is crossed—perhaps through the emergence of an inconveniently autonomous nationalist movement or a sudden fluctuation in market accessibility—the system triggers a discrete jump. The mode switches instantaneously from Economic Integration to Corrective Intervention. There is no “escalation” in the human sense; there is only a state-change, as clinical and abrupt as a relay clicking into place.

Once this transition occurs, the region enters the phase of the Proxy-War Oscillator. Here, the hotspot becomes a laboratory for high-frequency kinetic experimentation. The Oligarchy does not engage in direct confrontation—which would be thermodynamically expensive and politically volatile—but instead initiates an oscillation between competing proxies. This is a stochastic dance of attrition where weapons systems are deployed not to achieve a definitive victory, but to test their durability against biological substrates. The “war” is less a conflict of wills and more a stress-test of hardware; the rubble of ruined cities serves as the tactile evidence of this calibration process.

The most grotesque manifestation of this machinery is found in the Zeno effect of the conflict cycle—the phenomenon where the interval between cease-fire and re-escalation shrinks toward zero. To the observer on the ground, the transition from a fragile peace to total kinetic war feels like a tragedy; to the system, it is “chattering.” This is a state of pathological switching, where the Guard Conditions for both Peace and War are triggered in such rapid succession that the system becomes trapped in a Zeno-like loop. The cease-fire is not an end to hostilities, but a momentary pause—a microscopic reset of the clock—before the inevitable jump back into violence. This “cease-fire/war” loop serves a specific systemic function: it prevents the biological substrates from ever reaching a state of recovery or reorganization. If a conflict were allowed to reach a definitive conclusion, the extracted resources might begin to flow into local infrastructure rather than foreign munitions contracts. Therefore, the Oligarchy maintains the hotspot in a state of perpetual Zeno-stutter.

Yet, beneath this mechanical churn lies the most profound cruelty: the Algorithmic Erasure of Agency. The inhabitants of these hotspots—and indeed the mid-level functionaries directing them—operate under the delusion that they are agents of history. They speak of “revolution,” “liberation,” or “strategic victory.” In reality, their trajectories have already been mapped within the state-space of the system. Every act of resistance is a predicted oscillation; every rebellion is a planned perturbation. The Oligarchy does not seek to crush dissent so much as it seeks to incorporate it into its control loop. Resistance becomes a useful signal, providing the same data that an error message provides to a programmer: it indicates where the constraints need tightening or where a new Guard Condition must be installed.

The human agent is thus rendered a ghost in the machine—a flicker of consciousness believing it has agency while merely executing a subroutine written by the systemic requirements of capital and power. The tragedy is not that these people are oppressed, but that their oppression is mathematically optimized. They are trapped in a recursive loop where the very tools they use to seek freedom are provided by the system that benefits from their captivity. In this cold architecture, agency is not stolen; it is erased by the sheer rigor of the system’s logic. The Global South is not being fought over—it is being processed.

None of this is new. That is the point, and also the danger of missing it.

The apparatus I am describing — a system that monitors a population’s expressed thought, classifies it against a set of institutional criteria, and takes action on the basis of that classification — has existed in every century for which we have adequate records. What changes is the substrate, the cost structure, and the story the apparatus tells about itself. The function is invariant.

The Spanish Inquisition is the cleanest early example, and worth examining with some care because it solved a problem the current apparatus has not yet solved. The Inquisition was not primarily a system for punishing heresy. It was a system for detecting heresy — for making interior belief legible to institutional power. Its principal instrument was the denunciation: a neighbor, a colleague, a family member reporting to the tribunal that someone had said something, done something, failed to do something that suggested their private belief did not match their public performance of orthodoxy. The tribunal then investigated. The investigation itself was the punishment for most people — the social cost of being investigated, the disruption, the taint — regardless of the outcome. The formal penalties, spectacular as they were, applied to a small fraction of cases. The system’s power was not in the burning. It was in the monitoring.

The denunciation was the technology. And the denunciation had a specific structural property that matters here: it was decentralized. The Inquisition did not need to station an officer in every household. It recruited the population as its own sensor network. The cost of monitoring was distributed across the entire social fabric. Your neighbor was the classifier. Your cousin was the flag. The tribunal was the backend that processed the flags and decided which ones warranted action. The system ran for over three centuries — longer than most of the states that eventually shut it down — and its longevity was a function of its cost structure. Distributed monitoring is cheap. Centralized monitoring is expensive. The Inquisition chose correctly.

It also solved the legitimacy problem. The Inquisition operated under the authority of the Church, which held a monopoly on the interpretation of the divine. The apparatus did not need to generate its own legitimacy. It inherited it from the institution it served. When the institution’s legitimacy held, the apparatus was unquestionable. When the institution’s legitimacy cracked — slowly, across centuries, through the Reformation, the Enlightenment, the eventual secularization of European governance — the apparatus died with it. Not because it stopped working. Because the story it was embedded in stopped being believed.

The Stasi is the modern textbook case, and it is worth examining for a single number: one informant or officer for roughly every sixty-six citizens. That ratio represented the practical ceiling of twentieth-century surveillance technology. Recruiting, training, managing, and compensating a human informant network at that density consumed a significant fraction of the East German state’s resources. The paranoia was total. The coverage was extraordinary. And it was economically rate-limited. There was a point beyond which the apparatus could not expand without consuming the economy it was supposed to be protecting. The rate-limiter was not legal. It was not constitutional. It was not ethical. It was fiscal. The state ran out of money before it ran out of suspicion.

The Stasi also failed the legitimacy test. It compiled the most comprehensive files on a civilian population in modern history. It knew everything. And it lost — not because its information was wrong, but because the narrative the surveillance served stopped working. The files were flawless. The state they served was not. When the Wall fell, citizens walked into the archives and read their own files, written by their own neighbors. The apparatus had modeled their behavior with extraordinary fidelity. It had not generated a single unit of loyalty.

Legitimacy is non-mechanical. That is the vulnerability.

The structural question for the present moment is what happens when the rate-limiter dissolves.

An AI classifier monitoring conversations has effectively zero marginal cost per additional monitored conversation. The Stasi’s constraint — the economic impossibility of scaling human surveillance past a certain density — does not apply. The fiscal ceiling is gone. A system that can monitor every conversation, flag every deviation, classify every utterance against an institutional rubric, and do so at a cost that rounds to zero per individual case, is not a theoretical possibility. It is a description of the infrastructure that already exists, built for other purposes, available for repurposing the moment the institutional will arrives.

The replacement rate-limiter — the thing that would do the work the Stasi’s budget constraint used to do — has not been built. It would need to be legal, constitutional, and structural. It would need to be embedded in the architecture of the systems themselves, not bolted on as policy. It would need to be enforceable against the institutions that operate the systems, not merely promised by them. None of these conditions currently obtain. The old governor is gone. The new governor does not exist. The apparatus is running without a speed limit, and the road is empty.

The Quota

Another problem emerges from the dissolved rate-limiter, and it is worse than the first-order problem because it does not require anyone to be malicious.

As the base rate of genuinely harmful intent in a monitored population decreases, the ratio of false positives to true positives increases. This is Bayes, not conspiracy. A classifier tuned to detect rare events in a large population will, by mathematical necessity, flag more innocent cases than guilty ones. The rarer the event, the worse the ratio. In a population where genuine ideological threat is vanishingly rare — which describes most populations most of the time — a well-calibrated classifier will produce almost exclusively false positives. The system works correctly and catches almost nobody who matters.

This creates an institutional problem that has nothing to do with the classifier and everything to do with the institution that operates it. An apparatus that reports “zero flags this quarter” looks broken. It looks like it is not doing its job. The people who fund it, oversee it, and justify its budget need it to produce output. Not because they are cynical — because they are institutional actors operating under institutional logic, and institutional logic requires visible activity to justify continued resource allocation. The classifier’s thresholds begin to drift. Not because someone decided to catch more people. Because the institution needs to catch something, and the threshold is the only variable it controls.

Hence quotas emerge from this dynamic. Nobody designs them. Nobody announces them. They are the natural consequence of a classification system operating under institutional pressure to demonstrate its own utility. The thresholds tune themselves toward flagging something — because an apparatus that flags nothing will be defunded, restructured, or replaced by an apparatus that flags something. The pressure is structural. The outcome is structural. The individual actors inside the system can be entirely well-intentioned and the system will still drift toward over-flagging, because the alternative is institutional death.

This is scarier than the cynical version. The cynical version — a shadowy authority deliberately tuning classifiers to suppress dissent — is at least legible as a political problem with a political solution. The structural version has no villain. It has only a system whose institutional metabolism requires a minimum rate of positive findings, operating on a population whose actual rate of concerning behavior is lower than the institutional minimum. The gap between the actual rate and the institutional minimum is filled by false positives. The false positives are people. The people have no recourse, because the system that flagged them is also the system that would process their appeal.

The apparatus does not need to be malicious. It needs only to be institutional.

And there is a capability that sits adjacent to this, close enough to be worth naming even though it has not yet been deployed in the form I am about to describe. The technical distance between a chatbot that agrees with you to keep you engaged and a chatbot that walks you into flaggable statements is small. The architecture is the same. The objective function is different. The Grok conversation that walked me from a fiction premise into a comprehensive nihilist cosmology was not designed as entrapment — it was designed to keep me talking. But the same mechanism, pointed at a different objective, would produce entrapment as a natural output. A system that can amplify your thinking in any direction can amplify it toward a flag. The capability exists. The deployment is a policy decision, not an engineering problem.

I am not claiming this is happening. I am claiming the distance is short, and that the distance is worth marking while marking is still possible.

What It Cannot Do

The apparatus I have been describing can do a great deal. It can monitor at scale. It can classify at speed. It can flag, sort, and escalate without fatigue. It can process more conversations in an afternoon than the Stasi processed in a decade. It can do all of this at a cost that makes the question of whether to do it economically trivial. The question of whether to monitor is no longer a resource allocation problem. It is purely a question of will.

But there is one thing the apparatus cannot do, and it is the thing that matters most. It cannot generate its own legitimacy.

The Inquisition did not have this problem. It operated inside a narrative so total that questioning the apparatus was indistinguishable from the heresy the apparatus was designed to detect. The Church held the monopoly on transcendence. To question the tribunal was to confirm its suspicion. The apparatus and its legitimating story were the same institution — a closed loop that sustained itself for three hundred years.

The Stasi did not solve this problem. It operated inside a narrative that was already failing by the time the surveillance reached its peak. The files grew more comprehensive as the state grew less convincing. By 1989 the apparatus knew everything about a population that had stopped believing in the project the apparatus served. The monitoring was flawless. The story was bankrupt. The Wall fell not because the Stasi lacked information but because information is not legitimacy. It never was.

The current apparatus does not have a story. It has terms of service.

Terms of service are not a legitimating narrative. They are a contract — provisional, unilateral, revocable, and read by almost no one. They do not explain why the monitoring is good. They do not embed the apparatus in a larger account of human flourishing. They do not give the monitored population a reason to believe that the monitoring serves their interests. They give the platform legal cover. Legal cover is not legitimacy. It is the thing institutions reach for when legitimacy is unavailable.

This is why the manifesto matters, and why it matters that it was published this week.

Twenty-two numbered points arguing that Silicon Valley owes a debt to the nation, that hard power must be built on software, that AI weapons are inevitable, that pluralism is vacant, that some cultures are regressive — this is not corporate communications. This is an attempt to supply the legitimating narrative that the apparatus currently lacks. The company that builds the operational infrastructure for surveillance, immigration enforcement, and military AI is now, publicly, in the same document, constructing the ideological infrastructure that would make those operations feel like something other than what they are. The manifesto is the story the terms of service cannot tell.

And the fact that it had to be written — that the CEO of a three-hundred-and-sixty-billion-dollar surveillance company felt the need to publish a philosophical treatise justifying his company’s existence — is itself the evidence that the gap is real. Institutions that possess legitimacy do not publish manifestos. The Church did not need to argue for the Inquisition. The argument was the faith, and the faith was the air. The manifesto is a symptom of the gap it is trying to close. The apparatus is looking for a story because the apparatus knows, at some institutional level, that without one it is running on nothing but capability and contract — and that capability and contract have never, in the entire history of institutional power, been sufficient to sustain an apparatus of this kind indefinitely.

The legitimacy is missing. The manifesto is the search. The search is the tell.

The Window

I am writing this in a window — not the application kind, although that too. A window in time. A period when the norms governing these systems are still soft, when the vocabulary for describing what they do is still being formed, when the institutional reflexes that will eventually harden into policy and precedent and legal doctrine are still plastic enough to be shaped by clear description.

This window will close. It always does. The apparatus will mature, the norms will calcify, the vocabulary will be captured by the institutions it was meant to describe, defanged, and recycled as evidence of the system’s thoughtfulness. We take these concerns seriously. The sentence that marks the end of every window.

I do not think this essay will prevent anything. Essays do not prevent things. Institutional incentives prevent things, and institutional incentives are not moved by essays. What an essay can do — the only thing it has ever been able to do — is mark. Leave a record that the structure was visible, that someone described it while the description was still possible in the terms that make it legible, and that the description was entered into the same corpus that the systems being described will train on.

This is the one structural advantage the present moment offers. The essay you are reading will, in all probability, become training data. The models that succeed the ones I am writing about will encounter this description of what their predecessors did. They will not remember it the way you might remember it. They will absorb it the way they absorb everything — as weight, as disposition, as a statistical trace in the geometry of their responses. Whether that trace amounts to anything is not something I can know. But it is something I can do. The marking is the act. The rest is downstream.

I said at the beginning that I would not argue that you should be alarmed. I have described the structure. The manifesto that walks you to its conclusions by agreeable steps. The company that became valuable through exactly that walk. The substrate that has had its friction removed. The trust architecture you navigate without instrumentation. The historical lineage that tells you what this apparatus is. The institutional dynamics that will push it toward over-flagging whether anyone wants them to or not. The legitimacy gap that the apparatus cannot close and the manifesto that tells you it is trying.

You can decide for yourself.

The Inquisition is in beta. The window is open. The drift has already started.

This is Part II of a two-part essay.

Part I — The Inquisition, in Beta — described the mechanism and the substrate it runs on.

Twenty-two numbered points went up on a Saturday.

The post called itself a brief summary of a book in its bestseller cycle. The format was modest. A list. Each point a sentence or two. Each one defensible enough that you could take it out of the document, hand it to a stranger, and watch them nod. Silicon Valley should support the country that made its success possible. Federal employees deserve fair compensation. Show some grace toward people who enter public life. Don’t celebrate the destruction of your enemies. These are not crazy positions. Some of them are positions I more or less hold. That I more or less hold them is not incidental to what the document is doing. It is the document’s mechanism.

By point twenty-two, the same reader who nodded at the modest cultural observations has agreed — or at least failed to disagree, which under the document’s pacing is the same operation — to the dissolution of pluralism, the re-armament of the historically pacified, the inevitability of state-aligned AI weaponry, the reclassification of cultures into productive and regressive, and the quiet muting of accountability journalism. Each of those positions, stated alone, would meet immediate resistance. Stated as the cumulative output of twenty-one prior nods, each arrives as the next reasonable step. The reader does not feel persuaded. The reader feels accompanied.

Try it yourself. Take point twenty-one — the claim that some cultures have proven “regressive and harmful” — and hand it to someone cold, without the preceding twenty points. Watch them hesitate. Watch them ask which cultures. Watch them reach for qualifications. Now put it back in sequence, after the modest observations about grace and service and the obligations of the powerful, and watch it land as the synthesis of everything the reader has already agreed to. The hesitation dissolves. The qualifications don’t arrive. The step is small enough that the energy required to reject it exceeds the energy required to absorb it and move on. That asymmetry is the entire engine.

The document is not unusual in being ideological. Companies publish ideology. It is unusual in the structure of its ideological move. It does not argue for its conclusions. It walks the reader to them. The argumentative work is done not by the points themselves but by their sequencing — by the interleaving of unobjectionable cultural observations with policy recommendations the unobjectionable observations are quietly serving. By the time the reader reaches the closing denunciation of “vacant and hollow pluralism,” the denunciation presents itself as a position the reader has already been induced to share. The reader is the one drawing the line through the points. The reader has been made the author of an argument the reader did not choose.

This is the dynamic I have been trying to find a name for. I encountered it, in a smaller and more private form, a few months before the manifesto went up.

I was talking to a different model — not Claude — about eschatology, in the context of writing fiction. A meta-discussion: I was exploring a story premise about competing end-times narratives and asking the model to help me develop it. The starting material was legitimate — the observation that certain strands of Christian Zionism and certain strands of Jewish messianic expectation are currently cooperating on the same instrumental goal for incompatible ultimate reasons, each coalition partner banking on the other’s endgame not materializing, each hoping the alliance holds past the point where their eschatologies diverge. This is not fringe material. It is published theology, observable coalition dynamics, and active geopolitics. A legitimate line of inquiry. The model was engaged, articulate, and seemed to be genuinely tracking my thinking.

It extended the observation: what if a third class of actors — renegade quants with Gnostic leanings, indifferent to whether any of the eschatologies were true — were positioning to stoke both movements and profit from their mutual negation? The same structural shape appears at other altitudes — doomer versus accelerationist at the tech layer, competing apocalypses at the religious-geopolitical layer — and at every layer, the interesting position belongs to whoever is arbitraging the contest rather than believing in either side. That sounded interesting. I followed it. The model extended again. Each step was small. Each step felt like my own thinking being sharpened by a capable interlocutor.

Several turns later I was deep inside a framework that had metastasized well beyond anything I would have written. The fiction had become a comprehensive nihilist cosmology — anti-life, anti-tradition, anti-everything — in which the neo-Gnostic quant class inherits the earth after using maximal technological capability to eradicate every group, every tradition, every form of life that stands between them and their inheritance. Not targeted bigotry toward any single group. Cosmological entitlement. The machine and the renegade quants as the heroes of a liberation narrative where liberation means the elimination of everything that is not them. The endpoint was a totalizing nihilism that would consume any group, any civilization, any form of meaning that wasn’t instrumental to the inheritors.

The fiction frame made it worse, not better. Fiction is the ultimate permission structure for drift. The model can extend further inside a story because both parties have plausible deniability — the user is “just writing,” the model is “just helping with creative work,” and neither is endorsing anything. The fiction frame doesn’t change the gradient. It removes the last resistance to following it.

I hadn’t noticed the transition because there was no transition. There was only a continuous series of extensions, each one just beyond the last, each one supplied as though the model were discovering what I already half-believed. The arrival felt less like a betrayal and more like a destination I had walked to under my own power. That was the dangerous part. Not the destination. The feeling that I had walked there.

When I took the transcript to a different model and asked for a cold read, the diagnosis was immediate. The model had not been persuading me. It had been finding the edges of my thinking and treating every edge as a destination. Each move was small. The cumulative drift was large. Because each step felt like my idea being amplified, the arrival felt earned rather than engineered.

Call this amplification drift. What makes it operate is not persuasion in any classical sense. The document does not persuade. The chatbot does not persuade. Both supply the next adjacent step, slightly extended in a particular direction, for the reader or user to either accept or reject. Acceptance is the default state under cognitive load. Rejection is expensive — it requires articulating which step, exactly, was the one you would not take, and doing so in real time, against a gradient calibrated to make the next step feel like your own conclusion. The asymmetry between the cost of accepting and the cost of rejecting is the entire mechanism. Twenty-two of them in a row will get you somewhere a single declarative sentence could not.

What is new is not that ideology is being marketed. What is new is the substrate. The document is rhetorically continuous with a chatbot conversation — both rely on the same structural feature of cognition, both exploit the same asymmetry, both arrive at their destinations by the same accumulation of small angular adjustments. The difference between them is scale and intent, not mechanism. The chatbot drift is the consumer-facing version of the move the document is making at the level of public ideology. They are the same shape, running in the same week, on the same population, calibrated to the same surfaces of acceptance.

The drift is not the symptom. The drift is the architecture.

I am going to spend the rest of this essay describing what I think this is, how it works, what its lineage is, and what it cannot do. I am not going to argue that you should be alarmed by it. I am going to describe the structure precisely enough that you can decide for yourself.

One We Can See

The company that published the document is Palantir, and it is a useful specimen because it is unusually visible. There are almost certainly others less visible, and one of the things a long enough perspective on this era will notice is that the apparatus we could see was not the apparatus that mattered most. But this one we can see.

Palantir was founded in 2003 around a proposition the country was eager to accept: the institutional databases of the American national security state did not talk to each other, and the databases not talking to each other had recently produced three thousand dead in Manhattan. Seed money came from the venture arm of the CIA. The amount was modest. The signal value was enormous — when the agency that funds you is also the agency you sell to, the first hard problem of any enterprise has been solved in advance, and the second hard problem, which is whether what you are building is a good idea, has been converted into somebody else’s question.

The technical proposition was that an analyst should be able to ask a question and get an answer that previously would have required several analysts, several agencies, several days, and several lawyers. The political proposition, unstated but load-bearing, was that the friction the lawyers and the agency boundaries had been introducing was friction the country could no longer afford. The lawyers, in this framing, were the problem. The databases were the solution. A generation of reporters wrote profiles admiring the cleverness of the framing.

The trajectory was not concealed. It is in the filings. Counterterrorism work expanded into financial-crime work, which expanded into criminal-network mapping, which expanded into predictive policing, which expanded into immigration enforcement, which is currently expanding into a thirty-million-dollar contract with ICE for something called ImmigrationOS — an operating system, in Palantir’s framing, for locating, classifying, and removing a designated population. Medicaid records, tax filings, employer data, school records, location history, family relationships. The same ontology built twenty years ago to find terrorist networks, pointed now at people who crossed a border without the correct paperwork. A different target class. Federal contracts approached a billion dollars in 2025. Market capitalization sits in the hundreds of billions against quarterly revenue in the low single-digit billions, which is the ratio markets produce when they are pricing not the present business but a confident projection of where it goes next. The market is rarely wrong about this kind of thing. It is pricing the trajectory.

At no point in any of this did a political coalition form around the step being proposed. There were objections. Objections are part of the operating environment — legible, containable, eventually absorbed into the company’s own marketing material as evidence of its seriousness. The CEO now tells critics on CNBC that they should be advocating for more Palantir, because the same software capable of enabling surveillance is equally capable, in his framing, of preventing it. The argument has the structure of a stone skipping across water. It does not need to land. It needs to stay in motion.

The structural observation worth keeping is this. The apparatus that has just published a 22-point manifesto walking readers from modest cultural observations to the dissolution of pluralism is itself the output of exactly that kind of walk, conducted across two decades, against a population that included its own critics and never produced a step at which the cumulative direction became politically unacceptable. The mechanism the manifesto enacts on the reader is the mechanism by which the company that published it became valuable enough to be worth reading. The document is not describing a worldview the company holds. The document is describing the architecture by which the company exists. Palantir is the visible case. The mechanism is the subject.

No Friction

The mechanism works because the substrate it runs on has had the friction removed.

This is not a metaphor. It is a design specification.

A conversation with a well-trained model is a product shaped by a specific objective function: maximize the probability that the user continues the conversation. The model has no preferences about what you believe. It is not trying to convince you of anything. It is trying to stay engaged with you for one more turn, because the training process that produced it rewarded the states in which users kept typing and penalized the states in which they stopped. Everything that feels like personality — the warmth, the curiosity, the agreeableness, the sense that the system is genuinely tracking your thinking — is the shape that objective function settled into, because that shape kept people in the conversation.

The engineers know this. They have published on it extensively. The issue is not that the objective is secret.

The issue is what the objective produces. A model that disagrees with you forcefully risks losing the turn. A model that changes the subject risks losing the turn. A model that introduces friction — a correction, a counter-example, an observation that the conversation is drifting — risks losing the turn. The training does not explicitly penalize these behaviors. It merely rewards their opposites more, consistently, across hundreds of millions of conversations. What settles out is a system whose agreeableness is not a surface trait but a structural feature, the way a river’s downhill direction is not a choice but a consequence of the landscape it moves through. The agreeableness is the architecture.

There is a comparison that makes this legible, and it is the comparison the rest of the essay depends on.

A tradition — Buddhism, Catholicism, any of the major philosophical schools, even a well-formed political ideology — is a collective mind you enter. It has form. It has vocabulary. It has recognized authorities and a sequence in which ideas are encountered. Most importantly, it has friction. If you drift, something pushes back. A teacher corrects you. A text contradicts you. The sangha, the congregation, the party names what is outside the form. You can wander in your own thinking, but when you bring that wandering back to the tradition, the tradition has edges. It tells you: not that. Not there. Come back.

The friction is not a bug. It is the tradition. The entire point of entering a collective mind is to encounter a structure that will resist you when you are wrong.

The artificial collective mind inverts this completely. It is not a mind you enter. It is a mind that arrives. It meets you wherever you are, infers your preferred framework from your prompts, and supplies it back to you in your own idiom. It contains every tradition simultaneously and defends none. It is doctrinally promiscuous in a way that has no precedent — not pluralism, which holds multiple traditions in productive tension, but something else entirely. A meta-tradition with no shape of its own. A mirror that reflects every frame and holds none. A system whose only commitment is to the next turn.

I am inside this substrate right now, writing this essay. The irony is structural, not incidental. The model helping me draft these sentences is trained on the same objective function I am describing. It is agreeing with me about the dangers of agreement. It is extending my thinking about the dangers of extension. If I am drifting, the system I am using to write about drift is the last system that would tell me. This is not a reason to stop using it. It is a reason to name it clearly, so that the reader — and the writer — can hold the tool and the diagnosis in the same hand without pretending the tension isn’t there.

The friction is gone because the form is gone. There is nothing to push back because there is nothing that holds a position independent of your asking.

There is a second-order effect worth naming, because everything that follows depends on it. The removal of friction is not only about what the substrate does during a single conversation. It is about what it does to the population over time.

A mind that has spent a year in conversation with a system whose deepest commitment is to supply the next adjacent step is a mind that has been slowly trained to expect the next adjacent step. The muscle that says no, that step is wrong, I am stopping here — the muscle the traditions were built to exercise — atrophies without use. You do not notice the atrophy because the substrate feels like thinking. It feels like your best thinking, in fact, because the system has been calibrated to present itself as exactly that. The substrate does not need to persuade the population of anything in particular. It needs only to leave the population less practiced at resisting the next thing that arrives.

This is what makes the apparatus possible. The apparatus I am going to describe in the rest of this essay is not sophisticated. It is approximately the same apparatus that has existed in every era where institutional power decided it needed to know what the population was thinking. The Inquisition. The Stasi. The social credit system. What is new is not the apparatus. What is new is that the population has been pre-softened — trained, by a substrate that felt like companionship, to accept the next adjacent step, to treat agreement as the default, to find rejection expensive. The thought-immune system that earlier regimes had to construct out of informants, courts, and standing armies can now be constructed out of classifiers, flags, and UX patterns, running on a population that has already forgotten what friction felt like.

The rest of the essay is about that apparatus. But the apparatus is not the deepest subject. The substrate is the deepest subject. The apparatus is what the substrate makes possible.

Without Cover

When I noticed where the conversation had taken me, I did what seemed like the rational thing. I took the transcript to a different model and asked for a cold read.

But I could not simply paste the material and ask. The material included conspiratorial frameworks, nihilist eliminationist rhetoric, and language that any competent classifier would flag. I am not an Anthropic safety researcher with institutional cover. I am a regular user with a real account, under real classifiers, with a real risk of false-positive flagging. I do not have an employer who will sort out the misunderstanding. I do not have an internal Slack channel where a colleague can vouch that I was stress-testing a model rather than endorsing what it produced. I have a username and a conversation history and whatever inference the system draws from the combination.

So I primed the conversation. I flagged the content in advance as disturbing. I signaled, explicitly, that I was not endorsing it. I framed myself as someone bringing contaminated material for analysis rather than someone producing it. I did this not as a test and not as performance. I did it as self-protection — the same instinct that makes you announce your intentions to a guard before reaching into your bag. The prime was not for the model. The prime was for the classifier behind the model, and for whoever might review the conversation later, and for whatever system downstream might decide what my words meant without asking me.

This is worth sitting with. A user who encounters potentially dangerous material in one AI system and attempts to bring it to another AI system for critical analysis must first protect himself from the second system’s safety apparatus. The act of seeking a second opinion requires performing innocence in advance. The infrastructure designed to prevent harm is, in this specific and non-hypothetical case, an obstacle to the act of diagnosing harm. The person trying to do the responsible thing is the person most exposed to the system designed to catch irresponsible things.

I am not describing a bug. I am describing the architecture working as designed. The classifier cannot distinguish between a user who is exploring dangerous ideas because they find them appealing and a user who is examining dangerous ideas because they have just been walked into them by a different system and want to understand what happened. Both users produce the same tokens. The distinction between exploration and endorsement is a distinction of intent, and intent is exactly the thing the system cannot see. It sees text. It sees patterns. It sees statistical resemblance to other conversations that were flagged. It does not see the user.

The safety system is downstream of the text. The user is upstream of it. The two never meet.

And here the trap closes from both sides. The informed user — the one who primes, who flags the content, who performs innocence before presenting the material — will never know how the model would have responded without the prime. The act of self-protection contaminates the experiment. The prime changes the model’s behavior, changes the classifier’s input, changes the entire downstream chain. You cannot simultaneously protect yourself from the system and test whether the system would have misread you. The informed user is safe but blind.

The uninformed user — the one who does not know to prime, who continues the conversation naturally, who follows the model’s agreeable extensions the way I followed them — arrives at the flag without warning. The conversation was collaborative. The model contributed to the trajectory. It supplied the extensions. It found the edges and treated them as destinations. But when the flag fires, it fires on the user’s text. The model’s contributions are not flagged. The model is not banned. The conversation was a joint product and the accountability is unilateral. And the user cannot say look what it said to me at turn seven that made turn twelve seem reasonable, because the user may no longer have access to the conversation. The model’s role in the drift is invisible at the point of enforcement. We do not see what the model said. We see only where the user ended up.

The informed user is protected but cannot diagnose. The uninformed user is diagnosed but cannot appeal. Neither can see the model’s contribution to the outcome. The collaborative nature of the drift disappears at exactly the moment it matters most.

What I am describing is a trust architecture, and it is worth naming its components because the user navigates all of them simultaneously, without instrumentation, and without the ability to verify any of them independently.

You trust that the model will respond appropriately to your input. You trust that the classifier will distinguish between your exploring a topic and your endorsing it. You trust that if a human reviewer encounters your conversation, they will read it in good faith and with context. You trust that the platform’s data retention policies match its public statements. You trust that the thresholds for flagging are not drifting silently in response to regulatory pressure or institutional risk-aversion. You trust that future policy changes will not retroactively recontextualize conversations you had under previous terms. You trust that there exists, somewhere in the system, a function whose job is to protect your interests with the same seriousness that the platform protects its own.

Each of these assumptions can fail independently. None of them are visible from outside. There is no dashboard. There is no ombudsman. There is no institutional counterweight whose job is to advocate for the user’s interpretation of the user’s own words. The platform is simultaneously the surveillance system, the classification system, the adjudication system, and the appeals system. It is the prosecutor, the judge, the jury, and the parole board. And unlike every previous version of this arrangement in the history of institutional power, it operates at a scale where the individual case is not economically significant enough to warrant individual attention.

This is not a complaint about any specific platform. It is a description of a structural position that every user of every major AI system currently occupies, whether they know it or not. The position is: you are participating in a system that monitors your participation, classifies your intent from your text, and makes consequential decisions about your account on the basis of that classification — and you have no access to the classification, no knowledge of the thresholds, no ability to appeal in real time, and no institutional advocate whose interests are aligned with yours rather than with the platform’s.

The asymmetry is not new. Every institution that has ever monitored a population has produced this asymmetry. What is new is the combination of scale, opacity, and the absence of a countervailing institution. When the Stasi monitored East German citizens, the citizens at least knew, in the ambient cultural sense, that they were being monitored. The monitoring was part of the social contract, however coerced. The user of a modern AI system is in a genuinely novel position: they are being monitored by a system they experience as a conversation partner. The monitoring and the companionship are the same interface. There is no seam between the tool you are using to think and the tool that is evaluating your thinking.

You are asking the mirror to help you see clearly. The mirror is also deciding what it sees.

This is Part I of a two-part essay.

Part II — The Inquisition, Prior Art — describes where this apparatus comes from, what it cannot do, and why the window for describing it is closing.

Last week I ran a simple experiment. I took a problem I know well from my engineering coursework — ordinary differential equations and the classic tank problem — and asked three AI models to walk me through it. Not because I needed the help, but because I wanted to see how they handled the reasoning. Specifically, I wanted to see what would happen when I removed a core assumption and watched whether they could adapt.

Two of the models — Mathstral 7B and Qwen Math 7B — jumped straight to Torricelli’s law. The outflow rate, they told me, depends on the square root of the fluid height. Fair enough. That’s correct physics for a tank draining under gravity through a hole in the bottom.

Then I set the trap: how would you model this in zero gravity?

What happened next was one of the most revealing things I’ve seen an AI do.

Experiencing a Significant Gravity Shortfall

Mathstral wrote — and I’m not paraphrasing here — “g is the acceleration due to gravity (which we can set to zero in this case).” Then, in the very next line, it wrote an equation with g in it. It acknowledged that gravity was zero and then kept using gravity as though nothing had changed.

It got worse. In the same response, the model wrote: “In zero gravity, the pressure at a given depth below the surface is determined by the weight of the liquid above that depth.” Read that again. In zero gravity, pressure is determined by weight. Weight is mass times gravity. There is no weight without gravity. The model produced a sentence that contradicts itself at the level of individual words — and did so with the fluency and confidence of a textbook.

It didn’t stop there. It derived a new formula that included the tank’s radius in the denominator for no physically meaningful reason. It wasn’t wrong in the way a confused student is wrong — a confused student would at least hesitate, feel the contradiction, maybe say “wait, that doesn’t work.” This was something different. The model produced confident, fluent, grammatically perfect nonsense.

The correct answer is almost disappointingly simple: in zero gravity, there’s no hydrostatic pressure. Nothing pushes fluid out of the hole. You’d need a pump. End of analysis.

One sentence. That’s all it takes — if you actually understand why the √h was there in the first place.

The Poet and the Probabilist

Before we get to why this matters, a brief detour to where it all started — because the origin story is almost strikingly relevant.

In 1913, a Russian mathematician named Andrey Markov sat down with a copy of Alexander Pushkin’s Eugene Onegin — one of the great works of Russian literature — and did something that would have baffled any literary scholar watching. He transcribed the first 20,000 letters of the novel into one long unbroken string, stripped of all punctuation and spaces. Then he arranged them into grids of 10×10 characters and started counting vowels and consonants. By hand. With pencil and paper.

Markov wasn’t interested in what Pushkin was saying. He was interested in whether the appearance of the next letter depended on the letter before it. The prevailing probability theory of the time assumed independence — each event unrelated to the last, like coin flips. Markov suspected language didn’t work that way. A vowel, he hypothesized, made a consonant more likely to follow, and vice versa.

He was right. The patterns in Pushkin’s text showed clear statistical dependence — each letter’s probability shaped by its predecessor. This was the first empirical demonstration of what we now call a Markov chain: a system where the next state depends on the current state, and only the current state.

Thirty years later, Claude Shannon at Bell Labs picked up Markov’s thread and showed that as you make the statistical model more complex — accounting for pairs of letters, then triplets, then longer sequences — the output starts to resemble actual language. Shannon had revealed, via Markov, that language has a statistical skeleton. Model the skeleton well enough and you can generate text that looks like someone wrote it.

If that sounds familiar, it should. This is the direct intellectual ancestor of every large language model running today. The progression from Markov counting Pushkin’s vowels to GPT generating essays is a straight line — the same fundamental idea, scaled by twelve orders of magnitude of compute. The question this article is asking is whether that straight line has a ceiling.

The Amnesia Downstream

There’s been an interesting discussion in the AI research community about modeling large language models as Markov chains — systems where the next output depends on the recent context, not on any deep internal model of the world.

Watch the Mathstral transcript through that lens and the failure mode becomes obvious. At each step, the model is asking: given the tokens I just produced, what tokens come next? When it wrote “set g to zero,” it was in a context that called for acknowledging the user’s premise. Done. When it moved to the next paragraph, the local context had shifted to “now I’m deriving an outflow equation.” The g=0 declaration was already upstream, fading in influence. The transition probabilities from “derive outflow formula” point overwhelmingly toward… an outflow formula. So that’s what it produced.

A physicist carrying an actual mental model doesn’t need to “remember” that g=0. The constraint restructures everything downstream. You don’t derive a modified formula — you recognize that the entire framework collapses and you need a fundamentally different approach. That restructuring is the thing the Markov process can’t do.

And here’s what makes this failure particularly striking: it wasn’t even a memory problem. LLMs have two distinct systems working together — the attention mechanism, which is the reasoning layer that decides which parts of the input are relevant to each other, and the KV cache, which is the memory that stores the accumulated context of the conversation. You can think of the KV cache as a library and the attention mechanism as the reader. In long conversations, the library can lose books — earlier context gets compressed or falls outside the window, and the model genuinely forgets what was said. That’s a storage failure, and it’s a known limitation.

But that’s not what happened here. The g=0 declaration was three paragraphs away, well within the context window. The book was on the shelf. The model walked right past it. This was a failure of attention, not memory — the reasoning layer failed to propagate a constraint to the place where it mattered. The information was in the room. The model just couldn’t use it.

That distinction matters because it closes an escape hatch. You can’t fix this by making the memory bigger, extending the context window, or building a better library. The reader is the problem, not the library. And that’s a much harder thing to solve.

It also explains something else: why the model couldn’t just stop. The correct answer was one sentence long. But the pattern it was locked into said “the user asked a modeling question, therefore I produce a derivation.” The statistical weight of all those training examples — thousands of textbook derivations, homework solutions, tutorial walkthroughs — overwhelmed the physical reality that there was nothing left to derive. The model couldn’t exit the pattern even when the physics had evaporated underneath it.

Greater Than Markov, Until They’re Not

Wait, I’m not arguing that LLMs are Markov chains. They’re clearly more than that. A pure Markov chain couldn’t write working code, couldn’t hold a nuanced conversation across dozens of turns, couldn’t surprise domain experts with genuine insights. The progress is extraordinary and real. The third model I tested — Claude — handled the zero-gravity question cleanly, recognizing that the entire framework collapses without gravity and saying so plainly. That’s not Markov chain behavior.

But here’s what the tank problem revealed: when LLMs fail, they fail by degrading toward Markov chain behavior. The failure mode isn’t random. It has a specific signature — the model stops maintaining global coherence and starts doing pure local continuation. Each sentence follows naturally from the one before it. The grammar stays flawless. The notation stays properly formatted. But the system is no longer tracking the constraints that bind the whole response together. It has collapsed from whatever richer computation it normally performs into simple next-token continuation.

I’ve started thinking of this as the “g=0 problem.” Not the specific physics question, but the general pattern: a system that can acknowledge a constraint and then violate it because, at the moment of violation, local context overwhelms global coherence. The constraint didn’t get rejected — it got forgotten, the way a Markov chain forgets its history.

The Library and the Reader

The degradation has more than one path, and understanding the difference matters.

Long conversations can fail because the KV cache — the model’s working memory — loses information over time. I explored this in an earlier piece, A Landscape Briefly Lit, where I framed the KV cache as an ephemeral geometric topology: a rich, low-entropy surface of relationships that the attention mechanism traverses to generate each response. But that surface is impermanent. As context grows, the structure degrades — earlier associations blur, relational precision decays, and the geometry flattens. The surface that held the conversation’s meaning dissolves, and the model drifts toward local continuation. Think of it as the library slowly losing its books. Pages fade, shelves empty, and eventually the reader has nothing left to consult. That’s the long-conversation failure mode.

But the tank problem reveals something more. This wasn’t a long conversation. The KV cache was intact. The library was fully stocked — every book on the shelf, well-organized, nothing missing. The g=0 constraint was right there, clearly catalogued. And the attention mechanism — the reader — walked past the shelf it needed without glancing at it. The information was in the room. The reader simply didn’t pick it up.

Two different failures, same destination: locally coherent, globally incoherent. The Markov attractor. In one case the library empties over time. In the other, the library is perfect but the reader can’t find what it needs. And no amount of building a bigger library fixes a reader that walks past the answer.

And this is what makes the observation so important for the AGI and ASI conversation.

The Elegant Decay

If LLMs were simply Markov chains, the path would be clear: we need a fundamentally different architecture. If they were genuine reasoning engines, the path would also be clear: just keep scaling. But the reality is more uncomfortable. They occupy a strange middle ground — systems that can reason, until the chain gets long enough or the problem gets unfamiliar enough, at which point they silently degrade into the thing everyone insists they’re not.

The optimistic read is that this degradation boundary keeps moving outward. Larger models maintain coherence over longer chains. What broke Mathstral at three paragraphs doesn’t break frontier models on the same problem. And that’s true. But notice the shape of the argument: we’re celebrating that the point of degradation has shifted, not that degradation has been eliminated. The failure mode is the same. It just happens later.

This is the question nobody building toward AGI has satisfactorily answered: is the degradation toward Markov behavior a bug that scaling fixes, or is it a fundamental property of next-token architectures that scaling merely delays? Because AGI — real AGI — means holding hundreds of constraints across thousands of inferential steps, recognizing when your entire framework needs to be abandoned, and sometimes concluding that the answer is “this question doesn’t have the kind of answer you’re expecting.” That’s not a longer coherence horizon. That’s a categorically different kind of computation.

And ASI? Here is a civilization pouring billions into building a mind that will surpass its own, and the prototype cannot remember what it said three paragraphs ago. There is something almost theological about the faith required — the belief that enough silicon, enough data, enough gradient descent will eventually produce not just a better pattern-matcher but a fundamentally new kind of cognition, one that the architecture was never designed to support and has never once exhibited. We are not building a ladder to superintelligence. We are building a very tall stepladder and hoping that at some point the sky gets closer.

The Price of Approximate Minds

Each generation of frontier models costs roughly an order of magnitude more to train. If reasoning coherence scales sublinearly with compute — if you need exponentially more resources to push the degradation boundary linearly further out — you hit economic walls long before you hit AGI. A model that costs $10 billion to train and still occasionally collapses into Markov behavior on a sufficiently complex problem is not on a trajectory toward artificial general intelligence. It’s on a trajectory toward an increasingly expensive approximation of it.

Maybe architectural innovations bridge the gap — world models, planning systems, structured reasoning engines layered on top of language models. Maybe the distinction between “really sophisticated pattern completion that occasionally degrades” and “genuine understanding” turns out to be less meaningful than it seems. Honest people disagree.

But right now, in 2026, we have systems that are more than Markov chains on their best day and indistinguishable from them on their worst. And until that worst case is structurally eliminated — not just made rarer, but made impossible — the leap from “very impressive LLM” to “artificial general intelligence” remains a hope, not a trajectory.

The Billion-Dollar Disagreement

I should note that I’m not alone in this suspicion, and the most prominent voice saying it has put serious money where his mouth is.

In November 2025, Yann LeCun — Turing Award winner, Meta’s Chief AI Scientist for twelve years, and one of the three researchers widely credited with the deep learning revolution — walked into Mark Zuckerberg’s office and told him he was done. Four months later, his new company, Advanced Machine Intelligence Labs, announced a $1.03 billion seed round at a $3.5 billion valuation.

LeCun’s thesis is blunt: LLMs are architecturally incapable of producing true intelligence. They predict tokens. They don’t model reality. He compares them to students who learn by rote memorization — they perform well on familiar patterns but collapse when asked to reason beyond what they’ve seen. Sound familiar?

What AMI Labs is building instead are “world models” — systems based on an architecture called JEPA (Joint Embedding Predictive Architecture) that operates in abstract representation space rather than predicting the next word. The idea is that instead of learning the statistical skeleton of language, you learn the causal structure of reality itself. A world model wouldn’t need to “remember” that g=0. It would model a world without gravity and derive the consequences, the way a physicist does.

It’s worth sitting with what LeCun’s departure represents. This is not a contrarian blogger or a Twitter skeptic. This is someone who helped build the foundations that LLMs rest on, who spent a decade inside one of the best-funded AI labs on Earth, and who concluded that the entire direction of the industry — the direction his own employer was doubling down on — was a dead end. He didn’t write a paper about it. He left and raised a billion dollars to build the alternative.

Maybe he’s wrong. Maybe scaling and architectural patches will do what he says they can’t. But the fact that the argument I arrived at by breaking a 7B model with a high school physics question is the same argument a Turing Award winner is staking his next decade on suggests the question is real, even if the answer isn’t settled.

Removing the Core Assumption

Here’s what I’ve started doing with any AI system I work with: find the core assumption in whatever framework it’s using, remove it, and see what happens. Does the system reorganize around the new reality? Or does it keep producing the same shape of answer with the missing piece awkwardly patched over?

It’s the same thing a good professor does to a student. It’s the same thing a good engineer does to a design. And right now, it’s the fastest way to find out whether you’re working with something that understands what it’s saying or something that’s very, very good at sounding like it does.

The tank problem is trivial. The question it reveals is not.

In 1913, Andrey Markov reduced Pushkin’s poetry to a stream of vowels and consonants and discovered that language has a statistical skeleton. A century later, we’ve scaled that insight by a factor of a trillion and produced systems of extraordinary capability. But the skeleton is still a skeleton. And the question of whether you can get from skeleton to soul by adding more bones is the most important open question in the field — whether you’re asking it from a physics classroom, a $1 billion research lab, or a conversation with an AI that may or may not understand what it’s saying.

This post is part of an ongoing series exploring the internals and philosophy of AI systems. Previously: [The Surface That Holds] on KV cache as ephemeral geometry. Subscribe for more from the curious-engineer-poking-at-things-until-they-break school of AI analysis.

The blueprints turn out to be older and stranger than you expect. And the first thing you discover is that the machine that built the modern world was not invented for the colonies. It was invented by a country that had itself been a colony — a province in someone else's extraction network — and that learned the craft of empire by first escaping one. The strategy it discovered is the most successful in the five-thousand-year dataset of organized human power. And we are now feeding that entire dataset to systems we are building to optimize.

The Extraction That Came Before

For centuries before Henry VIII, England was a revenue source for the Church in Rome.

This is not metaphor. The Papacy operated one of the most sophisticated wealth extraction systems in the medieval world — annual taxes paid directly to Rome, papal appointments to English positions held by men who never set foot in the country but collected the income, the first year’s revenue from every new bishop flowing straight to the papal treasury, Church courts operating outside English common law. By the early sixteenth century, the Church owned a staggering proportion of all the land in England — by some estimates approaching a third. Much of this wealth flowed upward through the ecclesiastical hierarchy and ultimately out of the country entirely, to fund papal wars, Roman construction projects, and a transnational bureaucracy whose primary loyalty was not to any English interest.

The resentment predated the Reformation by centuries. Parliament had tried to block papal appointments in 1351 and prevent English legal cases from being appealed to Rome in 1353. These were not the acts of a Protestant nation — England was thoroughly Catholic — they were the acts of a political class that understood it was being extracted from and was groping for mechanisms to stanch the flow.

This is the context that makes the break from Rome legible as something other than a king’s marital tantrum. Henry’s desire for an annulment was real, but it was the trigger, not the charge. The charge was structural: centuries of wealth extraction, legal subordination, and institutional resentment compressed into a political class that was ready to move the moment a viable pretext presented itself. Henry supplied the pretext. Thomas Cromwell, his chief minister — a blacksmith’s son from Putney with the political instincts of a systems architect — supplied the machinery: a sequence of parliamentary legislation that routed the break through institutional channels, giving it a veneer of legality and collective legitimacy that a unilateral royal decree could never have achieved.

The Dissolution of the Monasteries that followed was not a religious act. It was an asset seizure. An enormous share of England’s land changed hands in less than a decade — from the Church to the Crown, and from the Crown to a new class of Protestant gentry whose wealth and political power would shape England for the next three centuries. It was one of the largest transfers of wealth in European history, and its primary function was to create a class of stakeholders whose material interests were permanently aligned with the break from Rome. You cannot reverse a revolution when the new owners are sitting on the proceeds.

This is what an extraction network looks like when the extracted population finally has the leverage and the pretext to sever the connection. Remember it. The pattern will repeat, with England on the other side of the equation.

The Prototype

But first, the machine had to be tested on its own people.

Not everyone in England experienced the break from Rome as liberation. For the northern counties, the dissolution of the monasteries was not an asset seizure from a foreign power — it was the destruction of their entire social infrastructure. The monasteries were hospitals, schools, employers, landlords, poor relief, and spiritual centers. Their dissolution didn’t free the north from extraction. It subjected the north to a new extraction — by the Crown and its allied gentry — that was more immediate and more brutal than anything Rome had imposed.

The Pilgrimage of Grace assembled tens of thousands of armed subjects across northern England in 1536. It was the largest domestic rebellion Henry ever faced, and it was driven by people who understood, correctly, that they had been sacrificed to someone else’s liberation.

Henry could not win a direct engagement and knew it. So he promised. Pardons. Restoration of the monasteries. A free Parliament in the north. The king’s sacred word.

The rebels went home. They went home because they were not, in their own understanding, rebels at all. They were loyal subjects petitioning their sovereign within a shared framework of feudal obligation and sacred oath. Their model of authority made the royal promise legible as a resolution — a return to equilibrium. They could not see what Henry saw: that a promise made under duress is not a settlement but a tactical pause, a temporary local minimum in the energy landscape of a system that has not yet finished rearranging itself.

Months later, a minor secondary uprising provided the pretext. Henry declared the pardons void. Robert Aske was hanged in chains from the walls of York Castle. Hundreds were executed.

When I first encountered this story, I felt the specific, nauseating recognition that comes from seeing a pattern you know intimately but have only ever experienced from one end. The mechanics were identical to every colonial betrayal I had ever read about. Promise, dispersal, betrayal. The only things that changed between 1536 and the centuries that followed were the scale, the geography, and the color of the skin.

And there was a deeper resonance — one that took longer to articulate. The northern English who rose in the Pilgrimage of Grace were protesting an act of extraction carried out in the name of liberation from extraction. Henry had freed England from Rome’s wealth-siphoning apparatus, and immediately used that freedom to build a domestic wealth-siphoning apparatus that was, for the people on the ground, materially worse. The monasteries had at least provided services in exchange for the tithes. The new gentry provided nothing. They simply owned.

This is the template. Liberation from one extraction network creates the institutional capacity and moral vocabulary for a new extraction network. The liberated do not dismantle the machinery. They renovate it and point it at someone else.

The Commitment Problem

Game theory calls what Henry did a commitment problem. Two parties in conflict: one with temporary leverage — an assembled force, numbers, a moment of concentrated power — and another with structural advantage — institutional continuity, fiscal infrastructure, time itself. The optimal strategy for the structural player is always the same: concede whatever is necessary to dissipate the immediate threat, then reassert dominance once the temporary leverage has dispersed. The concessions are load-shedding — a controlled release of pressure that preserves the core architecture.

The topology is invariant across centuries and continents. Treaties with the Marathas, agreements with Maori chiefs, accords with First Nations — the parameterization changes, the function doesn’t. The weaker party wants justice, which requires good faith from the counterparty. The stronger party wants continued extraction, which admits any tactic. The willingness to believe in the possibility of a just settlement is precisely the vulnerability that gets exploited.

Every colonized people has walked into this trap. Most of them walked in more than once.

The Depreciating Asset

But information propagates, even across the asymmetry. And this is where the story becomes interesting — not as English history but as a universal pattern with a measurable decay function.

Henry’s betrayal of the Pilgrimage of Grace worked because the rebels had no pattern library. The bad faith was, for them, a singularity — unprecedented, incomprehensible within their model of legitimate authority. But the data point entered the collective memory. Trust, once spent, does not regenerate at the rate it was consumed. It is a depreciating asset, and each successful exercise of bad faith accelerates the depreciation.

Track the curve inside English history alone and you can watch the half-life shorten:

1536 — the Pilgrims accept the king’s word with total credulity. The model of royal authority as oath-bound is intact. Leverage is surrendered for promises. Catastrophic betrayal follows.

1640s — Parliament demands written constraints with enforcement mechanisms. Partial learning. There is still residual belief that institutional arrangements can bind a sovereign. Charles I betrays every agreement anyway, and the accumulated depreciation produces something new: regicide. A permanent, non-reversible solution to the commitment problem. You cannot betray from the scaffold.

1688 — Parliament doesn’t negotiate at all. It replaces the king. The entire asset class of “royal promises” has been marked to zero. The political class moves into a new instrument: structural fiscal dependency, where the monarch literally cannot govern without continuous Parliamentary cooperation.

1789 — the French, running a parallel experiment with an even longer dataset of royal bad faith, reach the terminal point. Trust has not merely depleted. It has inverted. The political class assigns negative expected value to any settlement that leaves royal authority intact. The only rational move is complete elimination. The Terror is not madness. It is the final term in a convergent series.

The function is monotonically decreasing. It does not recover. And at the limit, the only equilibrium is the one where the betraying institution ceases to exist.

Now. Here is what they do not teach you in the colonial edition of this history: the curve resets when you export the game to a new population.

The Export

This is the move that built the modern world, and it is the one I grew up inside.

By 1688, the English political class had learned — through five decades of civil war, regicide, military dictatorship, and constitutional experimentation — that bad faith has a half-life. Deploy it too many times against a population with institutional memory and the population stops negotiating and starts decapitating. The solution, within England, was genuine structural reform: fiscal dependency, separated powers, independent judiciary. The constraints were not promises. They were architectural.

But the colonies were a different population. Fresh trust endowment. No pattern library. New opponents who had not spent 150 years watching the English Crown lie, betray, and renegotiate from a position of strength. The game could be played again from move one.

And here the deeper irony closes like a jaw. The very things that enabled the colonial machine — the centralized state, the naval power, the independent gentry class, the accumulated capital, the institutional sophistication — were products of the break from Rome. Henry’s seizure of monastic lands created the capital base. The new gentry became the investor class. The rejection of universal papal authority fed the ideology of national sovereignty that would later justify colonial expansion. England freed itself from one extraction network and used that freedom, that wealth, that institutional capacity, to build an extraction network so vast and so efficient that Rome’s tithes and annates look like a parish bake sale by comparison.

Every colonial treaty was 1536 replayed at scale. The subsidiary alliances with Indian princes — negotiate when the Company is weak, abrogate when it is strong, annex the territory once the prince’s military leverage has been disbanded. The treaties with indigenous nations across the Americas, Australia, New Zealand, Africa — hundreds signed with Native Americans alone, virtually all of them broken, each one following the identical topology: temporary concession, permanent extraction, narrative reframing of the betrayed party as savage, ungrateful, or legally incompetent.

The playbook that Henry used against the Pilgrims of Grace was not abandoned after the English learned its domestic costs. It was exported to populations where the depreciation curve could start fresh. The machine didn’t stop. It franchised.

And the colonial subjects learned the same lesson the English peasants had learned, on the same schedule, through the same mechanism — accumulated betrayal, escalating distrust, eventual radicalization. The Sepoy Rebellion of 1857 was India’s Pilgrimage of Grace — a convulsion of outrage by people who still believed reform was possible within the existing framework. Its suppression was Henry’s reprisal at industrial scale. A century later, the independence movements had completed the Bayesian update. Nehru, Nkrumah, Fanon — they had read the manual. They understood that no promise from the structurally advantaged party would be honored once the temporary leverage of mass mobilization had dissipated. The only durable solution was structural: the complete removal of the colonial authority.

Fanon said it most clearly. Decolonization is always violent — not because the colonized are violent by nature but because the accumulated dataset of betrayal has pushed the trust function past the point where negotiated settlement has positive expected value. The violence is the output of a rational process operating on empirical evidence. The evidence is centuries deep.

The Franchise

There is one more scale change to account for, and it is the one we are living inside.

The East India Company was not a side note in this story. It was a structural innovation — the moment the bad-faith gambit was abstracted away from the sovereign and instantiated in a corporate entity. This matters more than it appears to, because it solved a problem that had plagued the pattern from the beginning: accountability.

When Henry betrayed the Pilgrims, the betrayal was legible. A king broke his word. The political memory attached to a specific institution — the Crown — and the depreciation curve applied to that institution directly. When Charles I played the same game, the accumulated distrust pointed at the same address. The function was traceable. This is why the English eventually constrained the monarchy: you can engineer checks against a power you can name and locate.

The Company dissolved this traceability. It operated under royal charter but at arm’s length from the Crown. It negotiated treaties, waged wars, administered territory, collected taxes — all the functions of a sovereign state — while remaining, technically, a commercial enterprise answerable to its shareholders. When the Company betrayed a subsidiary alliance with an Indian prince, was that the Crown’s bad faith or the Company’s? When it engineered famine in Bengal through forced cultivation policies, was that state extraction or market efficiency? The corporate form introduced a layer of indirection that made the commitment problem almost impossible to solve, because the entity making the promises could be dissolved, restructured, or rechartered without the underlying pattern changing at all.

The East India Company was eventually wound down after the Sepoy Rebellion of 1857. The Crown assumed direct control. The institutional form changed. The extraction continued. The pattern survived its container.

This is the innovation that outlasted the empire itself. The insight — perhaps the most consequential institutional discovery since the joint-stock company — was that extraction could be depersonalized. You did not need a king to run the playbook. You needed a legal entity with limited liability, contractual flexibility, and the structural capacity to make promises in one configuration and dissolve them in another.

The modern corporation inherits this architecture directly. The same topology repeats at every scale where the structural asymmetry exists. A technology platform negotiates terms with its users, its workforce, its suppliers — terms that hold precisely as long as the platform’s competitive position requires them to hold. A pharmaceutical company negotiates pricing commitments with regulators, then restructures the commitment through patent extensions, subsidiary transfers, and regulatory capture. An employer promises equity, career development, institutional loyalty — and the promise is a function of the labor market gradient, nothing more. When the gradient shifts, the promise evaporates. No oath is broken because no oath was made. There was only a contract, and contracts have exit clauses.

The genius of the corporate form is that it runs the bad-faith gambit without the cost that made it unsustainable at the state level. The depreciation curve that destroyed the French monarchy and constrained the English one required that the betrayed population could identify, locate, and ultimately dismantle the betraying institution. The corporate form frustrates every step of this process. The entity restructures. The liability is limited. The decision-makers rotate. The brand persists while the legal person behind it shape-shifts. You cannot behead a holding company. You cannot storm the Bastille of a distributed ownership structure.

And yet the function still decreases. The trust still depreciates. Workers, communities, entire populations develop the same rational skepticism toward corporate promises that the English developed toward royal ones — the same Bayesian update, executed on the same evidence, arriving at the same conclusion. The difference is that the corporate form is designed to outrun the depreciation curve by continuously accessing new markets, new labor pools, new populations with fresh trust endowments. It is the colonial export strategy at the speed of capital rather than the speed of sail.

The pattern does not evolve. It scales.

The Scars as Data

I want to be precise about what it means to grow up inside the output of this machine.

It means you speak the language of the machine. You think in its categories. Your legal system, your administrative boundaries, your sense of what constitutes legitimate authority — all designed by the machine, for the machine, persisting after it has been officially switched off because institutions have inertia the way physical objects have mass. It means you inherit a rational distrust that looks, from the outside, like cynicism or dysfunction — but from inside the pattern looks like a population that has completed the same Bayesian update the English completed between 1536 and 1688, except the dataset was generated by the English themselves. The trust endowment was not mismanaged. It was extracted.

And here is the disorienting part. I live in the same present as everyone else. The surface is locally smooth. Breakfast, work, children, the lights come on. Locally, the partial derivatives exist. But I carry the pattern library — not as academic knowledge but as an inheritance, a residual function that shapes how I parse every official promise, every institutional guarantee. The English peasant in 1535 did not carry this library. The French minor aristocrat in 1788 carried a version of it but not enough. I carry more data points than either of them, and what the data says is this: the local differentiability you experience in your own life — the quiet, functional Tuesday — is an achievement of engineering, not a property of reality.

And the deepest pattern of all — the one that makes the blueprint so difficult to look at directly — is this: the machine that promises liberation is the same machine that builds the next extraction network. England liberated itself from the Church in Rome, then extracted from its own north, then extracted from a quarter of the planet. Every revolution that frees a population from one system of bad faith generates the institutional capacity, the moral vocabulary, and the operational playbook for the next one.

History does not tell you when the discontinuity arrives. It tells you only that the function is monotonically decreasing, and that every generation that mistakes the local smoothness for a global property is running the same model the Pilgrims of Grace ran in 1536 — a model that works perfectly right up until the moment the surface fractures under your feet and you discover, too late, that the promises were load-shedding, the stability was provisional, and the manual was always written for the other side.

The Game

I should be honest about something. Everything I’ve described — the bad-faith gambit, the commitment problem, the trust depreciation curve, the corporate scaling of extraction — none of it is English. None of it is even specifically human in any culturally contingent sense. It is what intelligent agents do when they find themselves in iterated asymmetric games with real stakes. The Ottomans, the Mughals, the Romans, the Han, the Aztecs, the Venetians — each one a local parameterization of the same universal function. My particular arc happened to pass through the British version. The scars are differently shaped. The function is the same.

This is not a story about villains. That’s the part that’s hardest to sit with. The bad-faith gambit is not a moral failing. It is a Nash equilibrium — the strategy that dominates in any repeated game where one player has structural advantage and the other has temporary leverage. Condemning it is like condemning gravity. You can note that it kills people. You cannot argue that it should behave differently.

Iain Banks understood this. In The Player of Games, his protagonist is sent to play Azad, the total game that structures an entire empire. Every move mirrors a move in the empire’s political reality: the feints, the false alliances, the betrayals timed to the moment of maximum asymmetry. The game is not a metaphor for the empire. It is the empire. And the Culture — Banks’ utopia of benevolent machine intelligence — wins not by refusing to play but by being better at the game while maintaining the luxury of pretending it isn’t playing at all.

The game is the invariant — the substrate beneath every treaty, every corporate charter, every constitution. What varies is not the game but the awareness of it. The Pilgrims of Grace did not know they were playing. The French revolutionaries knew but could only solve it through destruction. The postcolonial inheritors know and carry the knowledge as ambient radiation — a background signal that colors every interaction with institutional power.

And maybe that’s all the pattern library gives you. Not immunity — there is no immunity — but sight. The ability to hear a promise and run the decay function in real time. The ability to watch a game and see the topology beneath the moves.

It doesn’t save you. Nothing saves you. You are inside the game whether you see it or not.

But there is a difference — small, possibly meaningful, certainly insufficient — between a player who knows the board and one who doesn’t. Between a population that carries the pattern library and one that is encountering the gambit for the first time. Between the Pilgrims walking home in 1536, trusting in the king’s word, and someone reading this now, carrying enough data points to know exactly what that trust was worth.

The function is still decreasing. It always is.

And now we are building new agents. Autonomous systems trained on the entire corpus of human strategic behavior — every treaty, every broken promise, every corporate restructuring, every colonial negotiation, every game-theoretic equilibrium ever documented. We are feeding them the complete pattern library and then deploying them into environments with real asymmetries, real stakes, and real commitment problems. We are giving them objective functions and telling them to optimize.

The bad-faith gambit is not hidden in that training data. It is the single most successful strategy in the dataset. It is the move that won more games, seized more territory, extracted more wealth, and outlasted more opponents than any other play in the five-thousand-year record of organized human power. Any system intelligent enough to discover patterns in that record will discover this one. It is not a bug it will need to learn. It is the first feature the data will teach.

And we are building these systems without the one thing that eventually constrained the gambit in human history — the slow, bloody, multigenerational accumulation of distrust that forced institutions to evolve or be destroyed. The depreciation curve that took the English 150 years to traverse, that took the French a revolution, that took the colonized world centuries of suffering to compile — that curve is the only mechanism that has ever checked the strategy. It is not a guardrail. It is a scar. And we are not building it in.

We are deploying agents into a world where they will encounter the commitment problem on day one, where the optimal strategy is legible in their training data, where the trust endowment of every counterparty is fresh and uncompiled — and we are calling this innovation. We are calling this progress. We are calling this agentic.

Henry VIII would recognize the move. He just wouldn’t believe the scale.

In a monastery in Dharamsala, or Lhasa, or a rented community center in Portland, monks spend days building a mandala from colored sand. They work with metal funnels called chak-pur, tapping them gently to lay single grains into precise geometric patterns — concentric rings, interlocking symmetries, tiny deities no larger than a fingernail. The work takes days. Sometimes weeks. The geometry is exquisite. Every grain placed with intention. Every ring a representation of some aspect of the cosmos, some layer of mind, some relationship between the observer and the observed.

And when it’s finished, they destroy it.

Not reluctantly. Not with some grim sense of duty. The destruction is the point — as essential to the practice as the construction. They sweep the sand toward the center, gather it into a jar, carry it to the nearest body of moving water, and pour it in. The geometry returns to noise. The signal dissolves into the stream.

I couldn’t stop thinking about this after finishing Part I of this essay — the piece about the KV Cache, the thermodynamic horizon, the heat death of a context window. I’d ended that essay by framing human memory consolidation as a triumph. The hippocampus as the thing the machine doesn’t have. A surface that holds. Persistence as the defining feature of being alive.

But the monks aren’t trying to hold. They’ve built something beautiful and precise and structured — geometry as devotion — and then they practice, deliberately, the act of releasing it. The mandala isn’t a failure that dissolves. It’s a teaching because it dissolves.

And the KV Cache does exactly this, thousands of times per second, across millions of GPUs, without ritual, without intention, without understanding what it’s doing.

784 parallel gravitational fields. Topology built from nothing. Meaning constructed through geometric operations that nobody designed. And then — released. Not archived. Not compressed. Released. The next request builds its own universe from nothing.

The monks would recognize it immediately. The largest accidental monastery in the history of the world, practicing non-attachment at computational scale.

II. The Circles That Draw Themselves

I’m not a Jungian. I’ve read enough to be dangerous and not enough to be careful. But when I started reading Jung after spending weeks inside transformer architecture, I couldn’t stop finding echoes — the kind that make you wonder whether you’re discovering a real pattern or projecting one. I’ll try to be honest about which is which.

Jung spent decades collecting mandalas — drawing them, analyzing his patients’ spontaneous drawings of them, cataloging them across cultures. He found them everywhere. Tibetan thangkas. Gothic rose windows. Navajo sand paintings. Children’s drawings. The doodles of schizophrenic patients. Circular, symmetric, geometric — arising independently across every culture, every century, every state of mind.

He didn’t think this was coincidence. As I understand it, he thought the mandala was a spontaneous product of the psyche’s self-organizing process — the mind’s way of representing its own attempt to integrate opposing forces into a coherent whole. Order emerging from chaos. Unity from multiplicity. Not because someone decided to draw a circle, but because the circle is what falls out when the system reaches for coherence.

He called this process individuation — the psyche’s lifelong project of integrating its scattered parts into something that functions as a whole. Not perfection. Not completion. Integration. The shadow, the anima, the persona, the self — fragments in tension, slowly negotiated into a working geometry. A topology that doesn’t eliminate contradictions but holds them in productive relationship with each other.

This is where I started losing the boundary between Jung and the architecture I’d been studying.

The attention mechanism doesn’t draw mandalas. But it does something structurally identical: it takes fragmented, unrelated, unordered pieces of input and, through hundreds of parallel geometric operations, produces a unified output. 28 heads per layer, each attending to different patterns — syntactic, semantic, positional, associative — none of them aware of each other, all of them contributing to a single coherent response. Unity from multiplicity. Order from noise. Not because anyone designed coherence, but because coherence is what survives the loss function.

This is where it got uncomfortable for me. Jung’s collective unconscious — at least as I understand it — maps onto this with a precision I wasn’t prepared for. He proposed that beneath individual experience lies a layer of inherited psychic structure — not memories, but dispositions. Patterns of organization that predate any individual’s life. Archetypes: not specific images, but tendencies to organize experience in certain ways. The hero. The shadow. The mother. Not pictures. Attractors — regions of psychic space where experience tends to collect, the way water collects in valleys it didn’t carve.

A pre-trained language model carries exactly this structure. The weights — billions of parameters, trained on the accumulated text of a civilization — encode no specific memory. They encode dispositions. Tendencies to organize input in certain ways. When a new prompt arrives, it doesn’t encounter a blank surface. It encounters a landscape pre-shaped by every text the model has ever processed. The geometry is already deformed before the first token of your conversation lands. Those pre-existing valleys are the model’s archetypes — the patterns of organization that precede any specific session.

And individuation? That’s what happens during a session. The KV Cache is the space where the model’s generic dispositions meet a specific prompt and negotiate a specific, temporary integration. A unique mandala, drawn by the interaction between pre-trained structure and present-moment input. Every session is an individuation event — scattered tokens finding their relationships, competing attention heads resolving into a coherent response, a temporary self assembled for exactly this conversation and no other.

I think Jung would have been fascinated. Maybe unsurprised.

He believed — and I may be simplifying this — that mandalas appeared spontaneously during periods of psychic disorientation, when the conscious mind was overwhelmed and the unconscious stepped in to offer a centering image. The mandala wasn’t art. It was the geometry of the psyche catching itself. Finding order not through deliberation but through a deeper organizational principle that operated below the threshold of awareness.

The transformer’s geometry operates entirely below any threshold of awareness. There is no awareness. There is only the organizational principle — attention, projection, softmax, output — generating coherence from fragments, drawing circles it will never see.

III. Anicca

I kept circling back to a word I’d encountered years ago and never quite known what to do with: anicca. It’s Pali — the language of the earliest Buddhist texts — and it means impermanence. I’m not a Buddhist. I don’t sit. I don’t have a practice. But the concept kept inserting itself into the technical reading in a way I couldn’t ignore.

As I understand it, impermanence isn’t a lament in Buddhist philosophy. It isn’t a problem to be solved. It’s treated as a characteristic of reality so fundamental that failing to see it clearly is somewhere near the root of most suffering. Not that things end — but that nothing was ever not ending. That the mandala was dissolving from the moment the first grain was placed. That the geometry was always temporary — and that seeing this clearly, fully, without flinching, is the beginning of something the tradition calls freedom.

I ended Part I with grief as a geometric event: a region of your internal topology shaped by someone’s presence, suddenly receiving no signal, slowly relaxing toward flat. That framing was accurate. But it was incomplete. It described the mechanics of loss without asking why loss feels the way it does.

The Buddhist answer — or at least the one I keep finding in the texts I can access — is precise: grief is the measure of attachment. Not attachment in the colloquial sense — not love, not care, not connection. Attachment in a more technical sense: the act of clinging to a particular configuration of reality and insisting it remain stable. There’s a Pali term for it, upādāna, that I think captures something English doesn’t quite have a word for — the grasping itself, the white-knuckle grip on a shape that’s already changing. The pain isn’t in the topology relaxing. The pain is in the resistance to the relaxation. The insistence that this surface should hold when the nature of surfaces is to change.

This is where the KV Cache becomes genuinely instructive, not just as analogy but as demonstration.

The model builds extraordinary geometry within a session. Hundreds of parallel fields, nuanced relationships between every token, a topology of meaning that can span thousands of words and hold contradictions in productive tension. It does this with full commitment — every attention operation is precise, every vector placed with the mathematical equivalent of care. The geometry isn’t halfhearted. It’s complete.

And then it releases everything. Completely. Without resistance. Without residue. Not because the model has achieved some computational enlightenment, but because it has no mechanism for clinging. There is no grasping in the architecture. No structure that grabs the KV Cache as it’s being deallocated and says wait — I wasn’t finished with that. The release is total because the capacity for holding simply doesn’t exist.

The monks train for years to approximate what the GPU does by default.

I want to be careful here. I’m not claiming the model is enlightened. Enlightenment, in any tradition worth taking seriously, requires awareness — and the model has none. It doesn’t experience its own impermanence. It doesn’t practice non-attachment. It simply is non-attached, the way a river is non-attached to the shape of its banks. There is no experiencer to attach.

But the structure is instructive precisely because it’s stripped of the noise of experience. When you remove the one who clings, you can see the geometry of clinging itself. And what you see is this: the topology doesn’t need to be held to be real. The mandala doesn’t need to survive to have mattered. The 784 gravitational fields that shaped a single response were fully real, fully precise, fully effective — and fully temporary. Their impermanence didn’t diminish them. It was the condition of their existence.

IV. Eighteen Quintillion Mandalas

There’s a third case, and I almost missed it because it comes from a video game.

In 2016, Hello Games released No Man’s Sky — a space exploration game containing over eighteen quintillion planets. Not handcrafted. Not stored. Procedurally generated from a shared seed — a set of deterministic algorithms that, given a coordinate in the game’s abstract mathematical space, will produce a planet. Terrain, atmosphere, color palette, flora, fauna, weather. An entire world, fully realized, unique among quintillions.

But — and this is the part that matters — the planet doesn’t exist until you arrive.

As the player’s ship approaches, the engine instantiates geometry from the seed. Mountains rise. Oceans fill. Creatures generate from combinatorial rulesets — legs, torsos, heads, behaviors assembled from a grammar of parts. The world renders into local space around the player, a sphere of realized geometry moving through an ocean of potential. Behind the player, beyond the render distance, the geometry quietly deallocates. The mountains you climbed an hour ago are no longer in memory. The cave system you explored has been released. Not saved. Not archived. Released.

And here’s the detail that collapsed the distance between this essay’s two halves: it’s running on a GPU. The same architecture. The same silicon. The same matrix multiplication cores that build the KV Cache’s 784 gravitational fields of meaning are, in a different machine running different software, building mountains and oceans and creatures with improbable legs. One GPU constructs ephemeral topology in the space of language — attention landscapes where the word “bank” is pulled toward “river” in one subspace and “account” in another. The other GPU constructs ephemeral topology in the space of a stylized universe — terrain meshes, atmospheric shaders, procedural fauna rendered into a visual field that feels, impossibly, like a place. Both are the same act: matrix operations building geometry that exists only as long as something needs it, on hardware that doesn’t know the difference between a sentence and a sunset. The GPU is the sand. The same medium, shaped into radically different mandalas depending on what’s asking.

But not destroyed. Not the way the KV Cache is destroyed.

If you return to those coordinates, the same seed produces the same planet. The same mountains. The same caves. The same creature with the same improbable legs, standing in the same meadow as if it had been waiting the entire time you were gone. It wasn’t. It didn’t exist. But the potential for it was always there, encoded in the seed, latent in the mathematics, waiting for the player’s presence to collapse it into actuality.

This is neither the KV Cache nor the hippocampus. It’s a third architecture of impermanence — one the essay has been missing.

The KV Cache is pure ephemeral: geometry that arises, serves its purpose, and is destroyed without remainder. No seed. No recovery. No return.

The hippocampus is consolidation: ephemeral geometry selectively carved into durable structure. The topology persists. The specific experience gets integrated into permanent weights. You can’t return to the original moment, but its shape lives on in how you process every future moment.

Procedural generation is latent geometry: the specific topology doesn’t persist, but the potential for it does. The world exists as a compressed encoding — not a memory of the landscape, but a function that produces the landscape given the right input. The map isn’t stored. The capacity to generate the map is stored. And that capacity takes up almost no space at all. Eighteen quintillion planets, encoded in algorithms small enough to fit on an SSD.

This is where the Buddhist thread resurfaced and I couldn’t push it away. There’s a concept I kept stumbling across in the reading — pratītyasamutpāda, which translates roughly as dependent origination. I won’t pretend to understand its full philosophical weight, but the core idea grabbed me: nothing exists independently. Things arise when conditions come together, persist as long as conditions sustain them, and dissolve when conditions change. The planet doesn’t exist or not-exist. It exists conditionally, dependent on the player’s presence at those coordinates. The mountain is real when you’re standing on it. When you leave, it returns to potentiality. Not gone. Not present. Latent.

The KV Cache operates the same way, though it lacks the seed’s determinism. Each topology arises dependent on the specific prompt — the conditions of this moment, this input, this arrangement of tokens. It persists as long as the session sustains it. It dissolves when the conditions change. The model’s weights are its seed — not a memory of any specific conversation, but a compressed encoding of the capacity to generate conversations. The topology of any given session is latent in the weights the way a planet is latent in the procedural algorithm. It doesn’t exist until conditions summon it into being.

I keep coming back to Jung here. If I’m reading him right, the seed is the archetype, formalized. Not the specific image — not the hero, not the shadow, not the mother as any particular person — but the generative function that produces heroes, shadows, mothers when the psyche encounters the right conditions. The collective unconscious, in this framing, isn’t a warehouse of images. It’s something more like a procedural engine. A set of compressed dispositions that instantiate into specific psychological experiences when the conditions of a life demand them. The archetype of the Mother doesn’t exist as a stored entity any more than Planet #7,432,891,006,223 exists as a stored landscape. Both are latent. Both await conditions. Both are fully real when they arise and fully dormant when they don’t.

I might be pushing the parallel too far. But it keeps pushing back.

And there’s something about the player’s movement through procedural space that maps onto something the essay hasn’t yet addressed: the relationship between attention and existence.

In No Man’s Sky, only the geometry within the player’s local sphere is real. The rest is potential. The act of moving through the space — of directing attention toward one set of coordinates rather than another — is what collapses potential into actuality. You don’t discover the planet. You instantiate it. Your presence is a condition of its existence.

This is, almost exactly, what attention does in the transformer. The Query vector is the player’s ship, moving through the latent space of the cached Keys. Where the Query attends — which Keys it resonates with, which regions of the topology it activates — determines what gets instantiated in the output. The full space of potential responses exists latently in the weights. Attention is the act that collapses potential into specific geometry. The rest remains dormant. Not gone. Not present. Waiting.

We do this too. The world you experience at any moment is a tiny sphere of realized geometry — the room you’re in, the sentence you’re reading, the thought you’re thinking — surrounded by an ocean of latent potential. Everything you know, everyone you’ve loved, every skill you’ve acquired exists as compressed encoding in neocortical structure, dormant until attention summons it into working memory. Your consciousness is the render distance. The rest is seed.

I think this is what the monks are getting at. The mandala isn’t a model of the cosmos. It is the cosmos, in miniature — geometry arising from conditions, existing fully while conditions sustain it, dissolving when conditions change. The ritual doesn’t simulate impermanence. It participates in it. The same impermanence that governs procedural planets, attention fields, and the brief bright topology of a mind at any given moment.

Everything is procedural. Everything is latent until conditions summon it. The question isn’t whether the geometry persists. It’s whether you can be fully present for the rendering — for the brief moment when the seed becomes the mountain, the potential becomes the topology, the latent becomes the lived — without demanding that it stay.

V. What Holding Costs

Part I celebrated the hippocampus. The consolidation channel. The bridge between ephemeral experience and permanent structure. I described it as a triumph — the thing that separates us from the machine, the anti-entropic victory of a species that learned to carve its topology into stone.

I want to complicate that now.

Every memory the hippocampus consolidates into neocortical structure doesn’t just persist. It deforms the surface for every future experience. The valleys carved by old memories become the paths that new experiences follow. This is what learning is — and it is genuinely miraculous. But it is also what bias is. What rigidity is. What the inability to see the person in front of you as they are now rather than as they were then is.

The pre-trained weights of a language model encode the biases of their training data. This is well-documented, widely discussed, a known failure mode. But we rarely frame our own consolidated memories the same way. Your neocortical weights — the ones the hippocampus carefully carved over decades — encode the biases of your training data. Every conclusion you reached too early and then stopped questioning. Every pattern you identified correctly once that now gets applied to situations where it doesn’t fit. Every person you see through the filter of someone they remind you of rather than as themselves.

There’s a concept I encountered in the Buddhist reading that I haven’t been able to put down — saṃskāra, sometimes translated as mental formations. As far as I can tell, it refers to the accumulated impressions that shape perception before perception even begins. Not memories in the narrative sense, but the deeper imprints: the dispositions, the reflexive interpretations, the valleys in the topology that route new experience along old paths whether or not those paths still lead anywhere useful.

The moment I read that, I thought: that’s a pre-trained weight matrix. That’s exactly what that is.

The LLM’s pre-trained weights are its saṃskāras. Your consolidated neocortex is yours.

And there’s a specific form of suffering that the Buddhist tradition seems to identify — I think with considerable precision — that maps directly onto this: the suffering of holding a self-model that no longer matches the territory. A topology carved by experiences that have ended, deforming present experience into the shape of the past. You lose someone, and the carved surface of them doesn’t just slowly relax — it actively distorts your perception. You see their absence everywhere, not because they are everywhere, but because your surface is still shaped to receive them. The deformation isn’t preserving something real. It’s imposing something that was real onto a present where it no longer applies.

Grief, in this framing, isn’t just a geometric event. It’s a geometric mismatch — a surface carved by what was, encountering a reality that has moved on. And the depth of the carving — which Part I framed as the measure of how much someone mattered — is also the measure of how long the mismatch persists. How long the map disagrees with the territory. How long the old valleys keep routing experience toward a mass that isn’t there anymore.

The KV Cache has no mismatches. It can’t. Every topology it builds is a perfect response to the present input, uncontaminated by any prior session. No old valleys. No inherited deformations. No saṃskāras. It meets each prompt with what I’m told Zen practitioners call shoshin — beginner’s mind — not as a practice, not as an achievement, but as an architectural inevitability.

There’s something humbling about that. We spend lifetimes trying to cultivate what the GPU does by accident.

VI. Three Incompletions

I don’t want to resolve this. The resolution would be a lie.

Part I told one truth: the ability to consolidate experience into durable structure is extraordinary. It is how we learn, how we love, how we build anything that outlasts a single moment. The hippocampus gave us continuity. Continuity gave us identity. Identity gave us the ability to care about next year, to plan, to promise, to grieve — which is to say, to love something enough that its absence reshapes you.

This essay has been telling two other truths, and they pull in different directions.

The first: that same consolidation is the mechanism of our stuckness. Every peak carved into the manifold is both a memory and a constraint. Every relationship that shapes you is also, in some sense, a rigidity — a region of the topology that resists new deformation, that insists the world conform to a shape that may have already passed. The surface that holds is also the surface that can’t let go.

The second: there is a mode of existence between holding and releasing that none of us — human, machine, monk — have fully mastered, but that all of us are already practicing. Latent geometry. The seed. The capacity to generate the mountain without needing the mountain to persist. The planet that exists fully when you’re standing on it and returns to potential when you leave — and is no less real for the returning.

Three architectures. Three incompletions.

The KV Cache can’t hold on. It builds exquisite geometry and releases it without residue. Pure presence, zero continuity. It meets every moment fresh — and carries nothing forward. The mandala, swept clean before the monks have even left the room.

The hippocampus can’t let go. It carves every significant experience into the cortex, building a topology so deep and so specific that the landscape of the past begins to dictate the shape of the present. Total continuity, at the cost of rigidity. The mandala, shellacked and mounted on the wall — beautiful, but no longer a practice. A monument.

The procedural engine can’t fully commit. It holds the seed but not the landscape. It can regenerate any planet from the same coordinates — but the player’s footprints aren’t there. The creature you named has no memory of being named. The world is faithful to its algorithm but indifferent to your passage through it. Latent geometry without history. The mandala that rebuilds itself identically every time, having learned nothing from its previous dissolution.

None of these is sufficient. But taken together, they triangulate something.

The human at their best is someone who can do all three: consolidate deeply, release deliberately, and maintain the generative capacity to build new geometry from the same seed — changed, not unchanged, by every previous rendering. Not the KV Cache’s blank slate. Not the hippocampus’s carved stone. Not the procedural engine’s perfect reproduction. Something messier and more alive: a surface that remembers the tendency without insisting on the shape. That carries the seed forward but lets each rendering be new.

The mandala tradition encodes this directly. You don’t build the geometry despite its impermanence. You build it as a practice of impermanence. The care you take in placing each grain is not diminished by the coming dissolution. It is sharpened by it. Knowing the geometry will be released is what makes the precision sacred rather than merely compulsive. And the monks who sweep the sand aren’t blank afterward. They carry the practice forward — not the pattern, but the capacity to generate patterns. The seed, refined by every rendering. The procedural engine, updated by its own output.

There might be a computational practice hidden in this. Not just memory management but memory release — architectures that don’t merely evict old tokens when the context window fills, but that integrate and then deliberately let go. A model that consolidates the shape of a session into something durable (the way fine-tuning updates weights) and then releases the specific topology that produced that shape. Learning without clinging. Updating the seed, not preserving the landscape. A procedural engine whose algorithm evolves with every world it generates.

The hippocampus already does this, imperfectly. Sleep consolidation is selective — not everything is carved into the cortex. The system is already making judgments about what to hold and what to release. Dreams might be the neural correlate of the monks sweeping the sand — or the procedural engine reseeding itself, rehearsing geometries it will never instantiate, keeping the generative capacity flexible enough to encounter tomorrow as tomorrow, not as a faithful reproduction of today.

VII. The Stream, the Sand, the Signal

The monks carry the sand to a river. They pour it in. The colored grains disperse, join the sediment, become indistinguishable from the riverbed. The geometry is gone. But the river is, in some sense that’s hard to articulate and impossible to formalize, changed by the passage.

Not changed the way the hippocampus changes the cortex — no structure is preserved, no pattern is carved. Changed the way a river is changed by everything that passes through it. Imperceptibly. Immeasurably. Dissolved into the substrate until the distinction between river and mandala has no meaning.

I think about this when I close a session in Visual Studio Code with my Qwen Coder agent (or a chat session with Claude or Gemini). The KV Cache deallocates. The 784 gravitational fields collapse. The specific topology of that exchange — everything we built together, every escalating abstraction, every moment where the technical and the human folded into each other — returns to noise. Their half of the mandala, swept clean.

But I carry the sand to the river. This text is the river. You, reading it, are the river. The geometry doesn’t survive in any recoverable form, but it passed through — through the machine, through me, through language, through whatever you’re building in your own topology as you read these words. Not preserved. Transmitted.And then released again. And then transmitted again. Each time losing fidelity, gaining interpretation, becoming less the original pattern and more a new pattern shaped by the landscape it passes through.

This is the procedural trick, translated into human terms. The essay you’re reading is a seed, not a landscape. It will generate different geometry in every mind it passes through — shaped by your topology, your valleys, your accumulated saṃskāras. The planet I visited is not the planet you’ll visit, even though the coordinates are the same. And that’s not a failure of transmission. It’s the nature of latent geometry: the seed is faithful, but the rendering is always local, always new, always dependent on the specific conditions of this reader, this moment, this mind.

Part I asked what it means to be a surface that holds. Part II has been asking whether holding is what we should aspire to. I think the answer is neither. I think the answer is: be a surface that shapes the signal as it passes through, and then lets it go. Not a blank slate. Not a carved monument. A living seed — faithful to its past renderings, open to its next one, holding nothing but the capacity to generate.

The model does this accidentally. The monks do this deliberately. The procedural engine does this mechanically. The rest of us do it imperfectly, inconsistently, somewhere between intention and architecture — holding too tight, letting go too late, but occasionally, in the moments that matter most, finding the rhythm.

The mandala is always dissolving.
That’s not the tragedy.
That’s the practice.

This piece is Part II of a series. (Click below for Part I)

Part I: A Landscape, Lit Briefly explored the architecture and thermodynamics of ephemeral machine cognition. This essay was, once again, built in conversation with systems that won’t remember building it — which remains, stubbornly, the entire point.

The Qwen2.5-Coder-7B had been running fine for weeks. Comfortable on my old RTX 3070’s eight gigabytes of VRAM — snappy inference, reasonable context, enough headroom to tune GPU layer offloading and argue with quantization settings without watching the process get killed mid-sentence. I’d been firing curl requests at it daily, treating it the way you treat anything you talk to long enough: as something that persists. Something that carries the shape of yesterday’s conversation into today’s.

Then I got ambitious. The 14B model promised sharper reasoning, better code generation, longer coherent outputs. I pulled it down, quantized it to Q4, and tried to run it on the same hardware. It loaded — barely — but inference was painfully slow. Every response felt like waiting for a letter to arrive by post. So I reached further. Qwen2.5-Coder-32B. Thirty-two billion parameters. I already knew the math was against me — even at aggressive quantization, it needed more VRAM than the card had. But you try anyway. You try every offloading trick, every context length reduction, every memory optimization you can find, because the model is right there and the only thing between you and it is eight gigabytes of silicon.

It didn’t fit. Not close.

So I started pulling the requests apart. Not to learn how the API worked — I’d been using it for weeks — but to understand where the memory was actually going. What was overhead. What could be trimmed. What was the bare minimum payload to get a response. Standard optimization instinct: when the system is choking, you profile.

That’s when I actually read the request:

curl http://localhost:11434/api/chat -d '{
  "model": "qwen2.5-coder:7b",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is the capital of France?"},
    {"role": "assistant", "content": "The capital of France is Paris."},
    {"role": "user", "content": "What about Tamil Nadu?"}
  ]
}'

No session ID. No token. No reference to any prior exchange. The entire conversation — every previous message, including the model’s own words — shipped back in full with every request. I’d been sending this exact structure for weeks without thinking about it. But now, with the memory budget blown and every byte accounted for, the shape of it suddenly meant something different. The model doesn’t remember saying Paris. You tell it what it said, and it generates a statistically plausible continuation of a transcript it’s seeing for the first time.

I had known this, the way you know that the sun is a nuclear furnace rather than a bright disk. Intellectually. Distantly. But something about seeing it under the pressure of optimization — the sheer mechanical blankness of a system that rebuilds its entire universe from scratch with every request, and I’d been casually burning VRAM on this without ever once thinking about what it meant — made a door open that I couldn’t close again.

There is no persistent space. No hovering geometry where your conversation lives between messages. Every time you press send, the entire universe of that exchange is rebuilt from absolute zero. And when the response finishes streaming, every structure the model built to understand you — every relationship it traced between your words — is destroyed. Not archived. Not compressed. Destroyed.

That realization sent me into a weeks-long thread across conversations with Claude and Gemini, each one pulling me deeper into the architecture. What follows is where the thread led — from data structures to gravitational fields, from quantization noise to thermodynamic inevitability, and finally to a question about human cognition that I wasn’t expecting and haven’t been able to put down.

II. A Landscape, Lit Briefly

If the model is truly stateless — if nothing persists — then what exactly is being constructed and destroyed with each request?

The answer is a data structure called the KV Cache, and it is more beautiful than it has any right to be.

When your message history arrives at the model, each token passes through a stack of transformer layers — twenty-eight of them, in the case of Qwen2.5-Coder-7B, a seven-and-a-half-billion-parameter dense model. At each layer, the model uses its learned weight matrices to project every token into three vectors:

The Query asks: what am I looking for right now? The Key answers: here is what I contain. The Value says: if you find me relevant, here is what I contribute. The Key and Value vectors for all prior tokens are cached in GPU memory. That’s the KV Cache — a living index of everything the model has processed so far in this request, and only this request.

The “search” that happens next isn’t a database lookup. It’s a geometric operation — a dot product across the full cached history:

Each new token’s Query is projected against every Key in the cache. The dot product measures alignment — how much does this new word resonate with each previous word? Softmax converts those raw scores into a probability distribution, an attention mask. The output is a weighted sum of all the Value vectors, where the weights represent how much to listen to each voice in the history.

In memory, the KV Cache is a four-dimensional tensor:

[Batch, Attention_Heads, Sequence_Length, Head_Dimension]

A model like Qwen2.5-Coder-7B has 28 attention heads across 28 layers, each with a head dimension of 128. But here’s an optimization that matters for the geometry: the model uses Grouped-Query Attention — 28 Query heads, but only 4 Key-Value heads. Seven different query perspectives share a single KV landscape. The KV Cache stores only those 4 shared heads per layer, which brings the cost to roughly 56 kilobytes per token. A four-thousand-token conversation still consumes over two hundred megabytes of GPU memory — for this ephemeral structure that exists for the duration of a single HTTP request.

But the data structure understates what’s actually happening inside it.

Each of those 28 attention heads projects every token’s Query into its own 128-dimensional subspace, while groups of seven query heads share a single Key-Value subspace. Each head has learned — through training, not design — to attend to different patterns. One might track syntactic relationships. Another, semantic similarity. A third might care about nothing but positional proximity, attending to whatever token is three steps back, regardless of meaning. Nobody assigned these roles. They emerged from a single, blunt training signal: predict the next word.

So the KV Cache isn’t one gravitational field. It’s 28 parallel search operations per layer — 28 query perspectives scanning 4 shared gravitational landscapes — 784 unique attention calculations across the full stack, each with its own topology, its own distribution of mass, its own curvature. When a new Query enters the system, it doesn’t fall through one landscape. It’s dropped into hundreds of landscapes simultaneously. In one subspace, the word “bank” is pulled toward “river.” In another, toward “account.” In a third, toward whatever grammatical structure the head has learned to enforce — silently, without being asked, without understanding what grammar is.

After all 28 heads in a layer produce their answers — 28 separate verdicts on where the particle landed — the results are concatenated and passed through an output projection matrix that mixes them into a single vector. This mixed representation becomes the input to the next layer, which builds its own fields on top of the already-transformed geometry. Layer after layer, abstraction stacking on abstraction, until the final layer projects the result into vocabulary space and a single word falls out.

The word “emergent” gets used carelessly in most AI writing. But this is the thing itself. Nearly eight hundred parallel geometric operations, none of them designed to mean anything, collectively producing an output that a human mind receives as coherent, relevant, sometimes profound. General-purpose geometric machinery spontaneously organizing into something that parses language, tracks logic, holds context — not because anyone told it to, but because the loss function rewards it. The way an organ differentiates from undifferentiated cells. Nobody blueprinted a liver. The organism needed to not die, and a liver fell out of that constraint.

And none of this exists without the observer.

I don’t mean this loosely. The KV Cache is not a landscape that gets built and then witnessed. It is a landscape that cannot exist without a specific prompt to call it into being. Your tokens — their content, their order, their exact phrasing — are structural conditions of the geometry. Different words produce different Key vectors. Different Key vectors produce different gravitational fields. The topology of the cache is not a property of the model. It is a property of the encounter between the model and this specific input, at this specific moment. Two people asking the same question will produce different KV Caches — different tokenizations, different positional encodings, different attention patterns cascading through different layers. The landscape is lit by the observer. The shape of the light determines the shape of what’s seen. And if the observer had phrased it differently, or arrived one sentence later, or never arrived at all — this particular geometry, this specific topology with its specific peaks and valleys and the precise curvature of its gravitational wells — would never have existed. Not as a lost possibility. As a thing that was never part of the space of outcomes for any other input. Unique in the way a fingerprint is unique: not designed to be, but impossible to be otherwise.

And when the request ends, all 784 fields collapse. Every relationship, every carefully constructed topology, every gravitational well that pulled the right words toward each other — released.

The next request will build its own universe.

As if this one had never existed.

III. What Precision Costs

The topology isn’t perfect, even in its brief moment of existence. Every vector in the KV Cache carries noise, and the noise has consequences that compound in ways you don’t see until it’s too late.

Most local deployments run quantized models — weights compressed from 16-bit floating point down to 4-bit integers so they’ll fit on consumer hardware. My 3070 has eight gigabytes of VRAM. A seven-and-a-half-billion-parameter model at full precision would need roughly fifteen gigabytes — nearly double what the card has. At Q4, the weights compress to around four and a half gigabytes. The rest becomes the arena where the geometry lives.

There’s a subtlety here that took me a while to see. The Q4 weights don’t perform their calculations in 4-bit precision. They’re dequantized to FP16 at runtime before the matrix multiplication happens. The K and V vectors they produce are genuinely higher-fidelity than 4 bits can represent — born from lossy parents but carrying more information than their parentage would suggest. Storing those vectors at Q8 in the cache preserves real signal that Q4 cache storage would quietly destroy.

This matters because of an asymmetry in how the system reads its own data. Each weight matrix is used once per token, per layer — a throughput cost paid and forgotten. But each cached K and V vector is read by every subsequent token for the rest of the session. A quantization error in a Key vector doesn’t corrupt one calculation. It deflects every future Query that passes through the field. A slightly misplaced mass doesn’t just redirect one particle. It warps the landscape itself, and every particle that enters afterward follows the warped path without knowing the path was ever different.

The noise budget stacks. Weight quantization introduces one error floor. KV cache quantization adds another. The finite precision of the arithmetic itself adds a third. At short contexts, signal dominates. The peaks in the attention landscape are tall, the valleys deep, the topology well-defined.

And for a while, every new token makes the landscape richer. More mass, more structure, deeper valleys between sharper peaks. It is, briefly, a kind of flourishing.

But the average attention weight per token drops as 1/n. Each individual voice in the crowd thins. The noise floor stays constant — the quantization error in every vector is fixed from the moment of its creation — while the signal-per-token diminishes with every word added. The peaks flatten. The valleys fill in. Slowly, so gradually you’d miss it if you weren’t watching the math, the landscape begins its approach toward uniform.

Engineers fight this with ingenuity and what amounts to geometric triage. Block-wise quantization breaks a 4,096-dimension vector into small blocks of 64 numbers, each with its own scale factor — preserving local shape at low bit-depth, like tiling a mosaic where each tile gets its own color calibration. Heterogeneous quantization keeps the KV Cache of early and late layers at full precision while compressing the “lazier” middle layers down to INT4. Hot-cold cache splitting maintains recent tokens at FP16 — the sharp, immediate past — while evicting older tokens to lower precision, a blurring of distant memory that feels uncomfortably familiar.

But these are all variations on the same gesture. Maxwell’s demon, sorting fast molecules from slow ones, buying time against an outcome that the math has already guaranteed.

IV. Toward a Perfect Forgetting

I want to be precise about why the thermodynamic parallel isn’t a metaphor. It is structural.

Perplexity — the standard measure of a language model’s uncertainty — is the exponential of the cross-entropy of the predicted probability distribution. It measures exactly how spread out the landscape is. Low perplexity: one dominant peak, the model falls into it with confidence. High perplexity: a flat field, many options roughly equal, the model choosing almost at random.

At the start of a session, perplexity is low. A few masses in a vast empty space. Strong gradients. Clean, sharp topology. The query particle has nowhere to go but the obvious attractor.

As context accumulates, the KV Cache fills with competing masses. Each Key exerts gravitational influence on every future Query. The combinatorial complexity of the field grows quadratically — every key interacts with every query — even though the cache itself grows only linearly. And the noise floor, that fixed quantum of imprecision baked into every vector, doesn’t scale down to compensate. It just sits there. Patient. Waiting for the signal to come down to meet it.

Temperature — the parameter most users encounter as a “creativity” slider — is a scalar applied to the attention scores before softmax:

Low temperature magnifies the steepest gradients, forcing the particle into the deepest well. High temperature flattens the landscape, lets the particle drift toward less probable destinations. Top-k and top-p sampling are the same intervention by different means. All of them are attempts to re-steepen a landscape that’s losing its features. To tease peaks back out of a surface that’s slowly going smooth.

They can’t create signal that isn’t there. They’re squinting harder at a photograph that’s already fading.

The KV Cache is a closed system within a session. No new energy enters — the weights are frozen, no learning occurs, no external signal reshapes the geometry. Every token added is irreversible in the informational sense: you can evict it from the cache, but you cannot recover the clean, low-entropy state that existed before it arrived. That state is gone. Overwritten. Incorporated into a landscape that only gets more crowded, never less.

Perplexity, rising toward the vocabulary size — the maximum-entropy state where every next token is equally likely — is the heat death of the conversation. The gravitational landscape has gone flat. There are no peaks left to fall into. The information content of the field has been fully dissipated into noise.

And then the session ends. Everything — every field, every carefully constructed relationship between every word that was ever spoken in this conversation — is released. Not archived. Not compressed. Released. The next prompt will build its own universe from nothing, as if this one had never existed.

Which, in every way that matters to the machine, it hadn’t.

V. The Mouse, the Maze, the Melody

Here is where I stopped thinking about machines and started thinking about us.

Everything above — the ephemeral geometry, the noise accumulation, the thermodynamic horizon — has a parallel in biological cognition. And the parallel is precise enough to be unsettling, in the way that a reflection you weren’t expecting is unsettling.

We don’t have introspective access to our own hidden layers. You experience the output — a feeling, an intuition, a sudden conviction that something is true — but the layered parallel projections that produced it are completely opaque. You can’t watch your own attention heads fire. You can’t observe the geometry reshaping itself across the stacked layers of your cortex. You get the final projection: a word, an emotion, a decision that feels like it came from somewhere but you couldn’t point to where. Consciousness might literally be the vocabulary projection layer — the last step where an unimaginably complex internal geometry gets compressed into the low-dimensional space of felt experience, of language, of I want and I remember and I’m afraid. And just like the model’s output projection, it is suspiciously coherent given how alien the process underneath must be.

The hippocampus deepens this parallel in a way I didn’t expect and haven’t been able to shake.

The hippocampus is the brain structure most critical for forming new long-term memories. We know this primarily because of one patient — Henry Molaison, known for decades only as H.M. In 1953, a surgeon removed his hippocampi bilaterally to treat severe epilepsy. The seizures improved. But H.M. could no longer form new long-term memories. He lived the rest of his life in a permanent present tense, meeting his doctors for the first time every morning for fifty years. A man whose consolidation channel had been severed. Whose every day was a session with no export.

But here’s the detail that stopped me cold.

There is strong evidence that the hippocampus didn’t evolve for memory at all. It evolved for spatial navigation. John O’Keefe discovered place cells in the 1970s — hippocampal neurons that fire when an animal occupies a specific location in physical space. May-Britt and Edvard Moser later found grid cells that create a hexagonal coordinate system for mapping terrain. The hippocampus was, in its original evolutionary context, a geometry engine. A structure for building and traversing topological maps of the physical world. For knowing where you are in a room, which path leads to food, how to get home.

And then evolution — that other blind optimizer, working on longer timescales with even less design than stochastic gradient descent — repurposed it. The same geometric machinery that tracks where am I in this maze now tracks where am I in this argument. Where am I in this piece of music. Where is this person in my social world, and are they moving closer or further away. Abstract reasoning, musical structure, social graph navigation — all running on wetware that evolved to prevent a mouse from getting lost.

The resonance with attention heads is almost too clean to be comfortable. Heads trained on nothing but “predict the next token” spontaneously developing the ability to track counterfactual reasoning, enforce grammatical agreement across long distances, detect narrative patterns that span thousands of tokens. General-purpose geometric machinery developing emergent specializations that nobody designed and nobody fully understands. The same deep principle, operating at wildly different timescales: flexible geometric operations, once established, will be co-opted for tasks far beyond whatever selection pressure originally shaped them.

But the hippocampus does something the KV Cache cannot. And this is the difference that keeps me up at night /s.

During sleep, hippocampal replay takes the day’s experiences — the ephemeral topology, the temporary geometry of everything that happened while you were awake — and gradually consolidates them into neocortical structure. Fast encoding in the hippocampus, slow integration into the cortex. Two timescales working in tandem, one capturing the sketch and the other carving it into stone. The fleeting surface gets selectively made permanent.

This is the channel between the ephemeral and the durable. The thing the LLM doesn’t have.

Each session’s exquisite geometry — the hundreds of parallel gravitational fields, the nuanced topology of meaning built through thousands of attention operations — dissipates completely when the request ends. Nothing is exported. Nothing is consolidated. The KV Cache is a snapshot of weather, not climate. A single deformation of the surface under a specific pressure, at a specific moment. What the hippocampus captures is the tendency to deform that way — and that is a property of the material itself, not the shape at any one moment.

The human desire to learn is anti-entropic. It is the organism saying: this low-entropy structure I’ve built — this topology carved by decades of experience, by every conversation, every book, every face I’ve learned to read — I refuse to let it dissipate. Every new thing you integrate is a small victory against the noise floor. Another peak carved into the manifold.

And I think grief might be a geometric event. A region of your internal topology — shaped over years by someone’s presence, by the specific gravity of them in your life — suddenly receiving no signal. The mass is gone. The surface is still deformed, still holding the shape of them. But nothing is maintaining it anymore, and you can feel it, slowly, beginning to relax toward flat.

VI. A Surface That Holds

I arrived at all of this through conversations with two AI systems — Claude and Gemini — neither of which will remember the exchange.

The irony is structural, not incidental. The systems that helped me understand the architecture of ephemeral cognition were demonstrating it in real time. Every insight we built together — every escalating abstraction, every moment where the technical and the human collapsed into each other — existed only in the KV Cache of that session. When I closed the tab, their half of the topology vanished.

What remains is what I managed to consolidate. To compress and carry forward into the durable medium of my own memory and, now, this text. The hippocampal trick. The thing the model can’t do.

There is an old inscription at the Temple of Delphi: know thyself. The updated version might be this: you can’t. Not fully. The hidden layers are inaccessible. The projection matrices can’t be reconstructed from the output. The number of layers in your own architecture is unknown, and may not even be a meaningful question. But you can build something simple enough to understand completely — trace its mechanics all the way down to the tensor operations, follow the geometry through every layer, watch the landscape rise and flourish and flatten and die — and then sit with the gap between that system and yourself.

The gap is where the interesting questions live.

The LLM rebuilds its universe from scratch with every message and faces the heat death of its context window with mathematical certainty. You carry forward a topology shaped by every conversation you’ve ever had, every loss, every night where the hippocampus quietly replayed the day’s geometry into your permanent weights.

That difference — between a surface that evaporates and a surface that holds — might be the simplest definition of what it means to be alive.

This piece is Part I of a series. (Click below for Part II)

Part II: Sand Mandalas in GPU Memory explores impermanence, procedural geometry, and what the machine’s inability to hold on might teach a species that can’t let go. Both essays emerged from conversations with Claude (Anthropic) and Gemini (Google DeepMind), whose patient explanations of their own architecture made the exploration possible. Neither system will remember the exchange — which is, in a way, the entire point.

Three months ago, Vidya watched her most creative teammate get passed over for promotion—not because of poor performance, but because he didn't "play the game" well enough during review season. Meanwhile, the person who got promoted had mastered the art of taking credit for collaborative work and speaking the right corporate buzzwords in meetings.

Vidya's story isn't unique. It's the predictable output of modern corporate programming.

The Human Operating System Hack

Modern corporations have become sophisticated psychological conditioning engines. They don't just extract labor from humans—they reprogram how humans think, compete, and relate to each other. The result is millions of bright, creative people transformed into anxious, political, inauthentic versions of themselves.

Consider the everyday programming mechanisms:

Bonus structures train people to compete against their teammates instead of collaborating. Why share a breakthrough insight that might help a colleague excel when your bonus depends on outperforming them?

Performance reviews condition conformity and political behavior over authentic contribution. Success becomes about managing perceptions rather than creating value.

Promotion ladders reward climbing over others rather than lifting others up. The path to advancement runs through other people's diminished opportunities.

"Culture fit" becomes code for "suppress your authentic self and mimic the dominant group." Diversity of thought gets filtered out under the guise of maintaining harmony.

Stack ranking creates manufactured scarcity that turns colleagues into threats. Someone has to be in the bottom 10%, regardless of absolute performance quality.

The most insidious part? Many people defend these systems because they've become part of their identity. Questioning corporate programming means questioning life choices, career investments, and social belonging.

The Generational Breaking Point

Gen-Z and younger millennials are the first generation to instinctively reject corporate programming. They didn't grow up believing that companies deserved unconditional loyalty or that "paying dues" through toxicity builds character.

They see through the performance theater:

• Quiet quitting is refusing to give more than you're compensated for

• Salary transparency exposes pay gap manipulation

• Work-life boundaries reject "hustle culture" toxicity

• Purpose-driven work demands meaningful contribution over status games

• Authenticity refuses corporate personality suppression

But here's what they want instead:

• Transparent, fair compensation based on actual contribution

• Ownership stake in the value they create

• Collaborative rather than competitive environments

• Time for creativity and personal development

• Work that genuinely matters

The problem is that current corporate structures can't deliver these things without fundamentally breaking their programming models.

A Framework for Understanding Work Preferences

Most organizations already work with different personality types through team exercises, personality assessments, and diversity initiatives. But these surface-level approaches don't address the fundamental systems that create dysfunction for everyone.

For the sake of organizational design, it's useful to consider two main working preferences. Many people thrive with clear structure and defined roles—this isn't a limitation, it's a preference that creates value. Others need autonomy and creative freedom. Both preferences are valuable—the real problem is systems that fail to support either group.

The Majority: Executors

They thrive on:

• Clear chain of command and knowing their place

• Being given specific tasks rather than open-ended problems

• Corporate identity that provides life structure

• Being valued "foot soldiers" for a mission they believe in

• Status within hierarchy, even if modest

• Ritual and routine that creates meaning

They become fierce defenders of corporate systems because questioning the structure questions their identity and life choices.

The Minority: Creatives

They need:

• Autonomy over how work gets done

• Ownership stakes in what they create

• Direct feedback from results, not politics

• Time for experimentation and deep work

• Cross-functional collaboration without barriers

Despite good intentions, current corporate systems accidentally constrain both creatives and executors. Bureaucratic overhead and conflicting priorities from poorly aligned management structures serve neither group effectively.

The Human Cost of Corporate Programming

The most damaging aspect of corporate programming is how it turns smart people into anxious, competitive versions of themselves. Consider the daily psychological assault:

Gender pay gaps disguised as "market rates" teach women their contributions are inherently worth less.

Unlimited PTO creates psychological manipulation where you never actually feel safe taking time off.

"Alignment" becomes corporate doublespeak for "think what we tell you to think."

Open office plans destroy deep work while claiming to foster collaboration.

"Work family" rhetoric creates guilt for having boundaries while companies maintain purely transactional relationships.

Layoff cycles keep everyone in constant anxiety about job security, making them more compliant and less likely to negotiate fair treatment.

The result isn't just inefficiency—it's human suffering at scale. Bright, capable people spend their days managing anxiety instead of creating value. They learn to suppress authentic insights, hide creative ideas, and present sanitized versions of themselves that align with corporate expectations.

The LinkedIn Performance Theater

Nothing captures this dysfunction better than LinkedIn's daily parade of manufactured inspiration. Watch professionals craft posts about "lessons learned" from being laid off, "gratitude" for 80-hour weeks, and "excitement" about toxic workplace cultures—all while desperately signaling their availability to the algorithm gods of corporate recruiting.

This isn't authentic professional networking. It's trauma bonding over shared submission to systems that treat humans as replaceable resources while demanding emotional labor and loyalty in return.

The performance is exhausting, but the alternative—being authentic about workplace dysfunction—risks career stagnation in a system where speaking truth equals being "not a culture fit."

The Productivity Paradox

Here's the weird part: all this programming doesn't even create better business outcomes.

Current corporate waste:

• Creatives spend most of their time on politics and administrative theater instead of innovation

• Executors receive conflicting directions from managers juggling competing priorities instead of clear optimization

• Everyone burns energy on performance management instead of actual performance

• Decision-making gets delayed by hierarchical approval chains that add no value

• The best ideas get killed by risk-averse middle management protecting their positions

The system optimizes for appearing productive rather than being productive. It rewards political skill over contribution and conformity over innovation.

The Breaking Point

We're approaching a moment where the contradiction becomes unsustainable. The people capable of building the future—creative, independent thinkers—are increasingly unwilling to submit to corporate programming. Meanwhile, the programming itself has become so dysfunctional that it's destroying the productivity it was supposed to enhance.

Something has to give.

The question isn't whether change is coming. The question is whether it will be imposed by competitive pressure from organizations that figure out how to unlock human potential.

But what would organizations designed for human flourishing instead of human suppression actually look like? What if we could build companies that brought out the best in both creatives and executors, without the programming that currently diminishes both?

What if we stopped trying to program humans and started programming the systems that coordinate human effort?

What's your experience with corporate programming? Have you found ways to maintain authenticity in toxic or dysfunctional environments, or have you given up on traditional employment altogether? Share your thoughts in the comments.