When I was twelve years old, my school offered special science seminars to honors students, so I do not know how a rabble-rouser like myself ended up taking a few of them. The one that stayed with me was supposed to be on photography, but, oddly enough, we never touched a camera. The focus instead was on how the mind can be tricked. My teacher went through all the types of optical illusions that baffled our understanding of reality. We learned how the mind can be tricked, but what did that have to do with photography? Didn’t optical illusions have more to do with the brain? I stared at many autostereogram illusions on his wall, read old comics using classic 3D techniques, and explored many wonderful illusions, struggling to understand the meaning and purpose behind such things.

At sixteen, I decided I wanted to learn filmmaking and that the best way to do so was to go out and learn everyone’s job on a film set. I worked in distribution, marketing, story development, production, post-production, audio, lighting and more. You name it, I did it. For me, that also meant becoming an absolute expert in photography and camera systems. I will go toe to toe with any S-tier cinematographer and have been lucky to call some of them my friends. It began by selling cameras and film equipment at Samy’s Camera on Venice Blvd at eighteen years old. By twenty-one, I was helping run and expand one of the biggest camera rental houses in Hollywood. By twenty-six, I was managing million-dollar film productions. It was then that I truly appreciated learning how the brain worked very differently than how a camera worked. A camera sensor is a passive, linear recorder of physical energy, mapping photons to a uniform grid without judgment or context. Human perception does nothing of the sort. The brain does not capture a pixel-by-pixel recording of reality; it actively constructs an interactive interface tailored for action and survival. Understanding the friction between passive optical capture and active human reconstruction is what creates movie magic.

You see, unlike a camera system, perception is not a window. It is an interface. The cognitive scientist Donald Hoffman, a vision researcher at the University of California, Irvine, who earned his doctorate at MIT under David Marr (one of the founders of the modern science of vision), has spent decades making this case. He lays it out most fully in The Case Against Reality (Hoffman, 2019), using an image anyone who has used a computer already understands. Think of the icon for a file on your desktop. It sits there, a small blue rectangle in the corner of the screen. But the file is not blue, it is not rectangular, and it is not in a corner. There is nothing blue anywhere inside the machine. The icon does not resemble the file in any respect at all. It is a symbol the system draws for you, and its only job is to let you act without ever knowing what is actually happening in the circuitry underneath. This, Hoffman argues, is what all perception is. The colors, the shapes, the solid objects, space, and time themselves do not resemble whatever is really out there. They are icons, evolved not to show us the truth, but to keep us alive.

Hoffman’s argument has a cousin in one of the strangest experiments in the history of neuroscience, which was conducted on a frog. In 1959, four researchers at MIT (Jerome Lettvin, Humberto Maturana, Warren McCulloch, and Walter Pitts) placed electrodes into a frog’s optic nerve to find out what the eye was actually sending to the brain. The answer was not a picture. It was four distinct channels, each reporting one narrow feature, such as an edge, a sudden dimming, a moving boundary, and a small dark convex shape crossing the field (which the authors cheerfully called a bug detector). The retina was not transmitting the world. It was transmitting the four facts about the world a frog needs to eat and to avoid being eaten, discarding everything else before the brain ever saw it.

In the opening pages of that paper, the authors write that a frog “will starve to death surrounded by food if it is not moving.” If you put a dead fly in front of it, or even a hundred dead flies, the frog sits there and dies because a motionless fly is not a bug to its visual system, and there is no representation for such a thing. It is not that the frog sees the flies and fails to recognize them; the flies are simply not there. Its world does not contain stationary food.

Did you catch that? It was the clue to all intelligent life and how the brain really works.

The frog’s eye is doing exactly what it evolved to do, brilliantly, at an enormous savings of energy and computation. It reports the few features that reliably pointed to dinner for four hundred million years, and it does not waste a single fiber on the rest. The interface works, and it works precisely because it leaves the physical world out. The price of that bargain only comes due in a situation evolution never accounted for, such as a bowl of dead flies. Hoffman’s claim, in a way, is that we are the frog. We did not evolve to see the world as it is; we evolved to see the slice of it we can act upon.

There are three ingredients you cannot avoid when creating an intelligent organism: action, sensory feedback, and mortality.

Every creature is sealed inside a different interface running on those three variables. A crow carries a fourth type of color receptor in its eye and draws distinctions we have no words for and cannot picture, its interface rendering the meadow in colors that, for us, simply do not exist. A century ago, the Baltic German biologist Jakob von Uexküll named this concept the Umwelt (von Uexküll, 1934/2010), representing the small slice of the world a given animal can actually sense and act on, which for that animal is all there is. His favorite example was the tick. A tick has no eyes and no ears. Its entire world consists of three signals. It climbs a stalk of grass and waits, sometimes for years, doing nothing, for the first signal, which is the smell of butyric acid, a compound that seeps from the skin of every mammal alive. When that smell arrives, it lets go and drops. If it lands on something warm (the second signal), it has found a host. It feels for a patch of bare skin (the third signal) and burrows in. Butyric acid, warmth, bare skin. That is the entire universe of the tick. Everything else you and I can see, hear, love, and fear does not exist for it, not because it is hidden, but because the tick has no interface for it. There is no view from nowhere, no neutral and complete world sitting behind all the interfaces waiting to be seen whole. There is a tick’s world, a crow’s world, a frog’s world, and ours, with each one being a construction fitted to a specific body. Perception is not a recording. It is an act, performed by a particular animal, that brings a particular world into being in order to ensure that the organism can survive against the entropic decay of the physical world.

In one of my favorite movies, The Matrix, a child bends a spoon with his mind and tells the protagonist, Neo, the secret. “Do not try to bend the spoon,” he says, “that is impossible. Instead, understand the truth. There is no spoon.” The spoon is not a physical object in the world; it is an item in the render. Once you see that it was never solid to begin with, it gives way.

The thing you are so sure is out there, hard, fixed, and indifferent to you, is a symbol your system draws, and it often does not truly resemble what lies underneath.

Hoffman follows this to the end of the line. He argues that space and time are not fundamental, that the physical world dissolves into networks of interacting conscious agents, and that the objective world is ultimately unknowable and perhaps not even the correct mental model.

However, consider the desktop example that Hoffman himself uses. The icon is a fiction and does not resemble the file, but if you drag that icon to the trash, the file is really gone. Not just the picture of the file, but the file itself. You cannot see the underlying circuitry in the machine, and you do not need to, because the interface is wired to consequences that are entirely real. The interface hides reality while remaining strictly accountable to it. While Hoffman’s metaphysical idealism steps away from an objective external physical realm, active inference and enactivism maintain that physical reality is inescapable. There is an external environment that exerts unyielding thermodynamic constraints on an organism.

Once you stop demanding metaphysical accuracy, what remains is what every physical science has actually run on: stable, shared, predictively reliable structure. A physical reality we answer to with our very existence.

Now let us imagine an astronaut on the moon. She can never touch lunar dust with her bare hand, because a bare hand on the lunar surface means a scream of pain followed by frostbite. Everything she knows of the moon is mediated and filtered through a suit, never touched directly. Yet she is still gripping it. She leaves footprints in it. It bears her weight and resists her boot, and that resistance is real, arriving through the suit, correcting her balance, and teaching her body exactly how this ground behaves. The suit is not what separates her from the moon; the suit is the only reason she can stand on the moon at all. If it offered direct, unmediated contact with the raw lunar reality, it would be a terrible suit and the astronaut would not survive.

That is what perception is, and the essential detail is how little it needs to be. It is not a window onto absolute truth, which no organism has ever possessed, but an apparatus through which a reality you could not survive touching raw is allowed to push back on you through our senses. It operates repeatedly in a consistent pattern, all built around an organism’s fight to survive the dissipative nature of entropy. Reach for a hot pan, and the redness you see is an icon invented by your visual system, nowhere to be found in the pan itself. Reach incorrectly anyway, and the burn is not an icon. The burn does not explain the physics of the pan, but it corrects you so you can survive the reality of the heat, and you will reach differently for the rest of your life. Veridicality was never the requirement. Accountability was.

It is also worth noting that none of this is brand new. The synthesis of predictive processing and enactivism has been a dominant position in cognitive science since roughly 2013, when Andy Clark laid it out in Behavioral and Brain Sciences (Clark, 2013). His book Surfing Uncertainty (Clark, 2016) provides a comprehensive statement of the framework. Karl Friston’s active inference framework expresses this same synthesis mathematically (Friston, 2010). What is rare is not the theoretical framework itself, but the fact that mainstream computer science does not build AI systems with this principle in mind.

Anil Seth, a professor of neuroscience at the University of Sussex, captured this succinctly. Perception, he notes, is a “controlled hallucination,” a concept central to Being You (Seth, 2021). The hallucination part aligns with Hoffman’s point that what you experience is constructed rather than passively received. The controlled part is the leash, the moment-to-moment correction by a physical world that does not care about your predictions. Seth’s lab demonstrated what happens when you sever that leash. They took a neural network designed for object recognition, ran it in reverse so that it painted its expectations onto images rather than labeling them, and fed a participant a panoramic video through a headset. The world filled with calm, drifting hallucinations (dogs, eyes, and archways blooming out of ordinary walls and grass) without administering a single drug. They called it the Hallucination Machine (Suzuki et al., 2017), and it hallucinates for the exact reason humans do: by letting expectation run far ahead of evidence when sensory feedback correction is disabled. In a similar vein, they ran a public installation called the Dreamachine, where participants lay under flickering lights with closed eyes and experienced rich, moving geometric forms generated entirely by a brain receiving no external visual images. Both demonstrate the same lesson: what we see is actively generated by the brain, and normal perception is that hallucination constrained by sensory feedback. Take away sensory feedback (our leash), and perception strays like an unchained dog.

This same principle surfaces in machine learning, far removed from philosophy or neuroscience. When an agent plans by projecting its learned model forward, error compounds with every step because each output becomes the next input, causing the trajectory to drift away from trained reality. The most reliable solution discovered in model-based RL is not building a larger model, but maintaining short rollouts anchored to states confirmed by real-world interaction (Janner et al., 2019). This physical tether keeps generative models grounded.

Vision-language-action models trained on large robotic datasets represent an active field of development (Brohan et al., 2023). However, next-token prediction over text lacks a channel through which physical resistance can correct a prediction. There is no burn. The learning loop closes entirely inside the corpus. A model can be scaled until it has read every written account of fire, yet it will never perceive heat because nothing in its training is tied to physical consequence. True intelligence requires action, mortality, and real-time sensory feedback.

For an artificial system, mortality does not have to mean biological death. It means non-negotiable physical consequences, such as power depletion, motor failure, structural damage, hardware wear, or catastrophic task failure. In a zero-consequence digital environment, disembodied models can hallucinate indefinitely without cost. But an agent operating under artificial mortality cannot afford ungrounded fantasy. Physical reality forces it to adapt or cease functioning. No amount of pure algorithmic scaling or software workarounds can substitute for this grounded interaction loop of enactivism.

This connects back to a famous series of psychological experiments. In the 1930s through the 1950s, Austrian psychologists Theodor Erismann and Ivo Kohler built upon George M. Stratton’s early work by placing participants in inverting prism goggles for weeks at a time (Sachse et al., 2017; Stratton, 1896). The participants demonstrated remarkable neuroplasticity, eventually learning to ride bicycles, fence, and ski with inverted vision. Their brains adapted the perceptual field through active physical engagement and sensory feedback to keep them functional.

Consider the classic story of Oz, where visitors entering the Emerald City are required to wear locked green spectacles, making ordinary stone appear as blazing emeralds. Every interface transforms raw data. The astronaut’s suit filters reality, your visual cortex filters reality, and the spectacles filter light. When you snowboard wearing orange-tinted goggles, your brain initially sees orange, but soon adapts and neutralizes the hue. When you remove the goggles, the world briefly appears intensely blue.

These dynamics explain why large language models can present false information with the exact same confidence as facts. To an ungrounded model, there is no functional difference. When you remove ski goggles, physical reality forces an adjustment. Perceptual adjustments occur when internal predictions fail against the environment, whereas disembodied AI lacks an environment against which its internal representations can fail.

The fundamental question is not whether an intelligence sees reality in its raw, objective form. Nothing does. Humans do not, crows do not, and ticks do not. What grants a mind precision is that its internal model is answerable to physical reality, continually corrected by an environment with real survival consequences.

There is no spoon, and yet we need to eat. If you die in the Matrix, you die in the real world. That’s what the AI in that film understood when creating the matrix. That is why it works.

Enactive AI (eAI) is an intelligence framework grounded in acting upon and perceiving the world in real time, rather than predicting subsequent tokens in a text corpus. True intelligence demands direct accountability to physical reality.

That is the core paradigm we are building at Absentia Technologies.

Read more about what we are building at Absentia Tech → absentiatech.com

Emanouil Angelov is Co-Founder of Absentia Technologies and a screenwriter who has been writing films about artificial intelligence since 2017. His background spans professional filmmaking, cinematography, photography, marketing and teaching. His work at Absentia is informed by the intersection of visual perception, linguistic theory, and AI architecture.

References

Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Gonzalez Arenas, M., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., ... Zitkovich, B. (2023). RT-2, Vision-language-action models transfer web knowledge to robotic control (arXiv 2307.15818). arXiv.

Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181-204.

Clark, A. (2016). Surfing uncertainty, Prediction, action, and the embodied mind. Oxford University Press.

Friston, K. (2010). The free-energy principle, A unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138.

Hoffman, D. D. (2019). The case against reality, Why evolution hid the truth from our eyes. W. W. Norton.

Janner, M., Fu, J., Zhang, M., & Levine, S. (2019). When to trust your model, Model-based policy optimization. In Advances in Neural Information Processing Systems 32 (pp. 12519-12530).

Lettvin, J. Y., Maturana, H. R., McCulloch, W. S., & Pitts, W. H. (1959). What the frog’s eye tells the frog’s brain. Proceedings of the IRE, 47(11), 1940-1951.

Sachse, P., Beermann, U., Martini, M., Maran, T., Domeier, M., & Furtner, M. R. (2017). The world is upside down, The Innsbruck Goggle Experiments of Theodor Erismann (1883-1961) and Ivo Kohler (1915-1985). Cortex, 92, 222-232.

Seth, A. (2021). Being you, A new science of consciousness. Faber & Faber.

Stratton, G. M. (1896). Some preliminary experiments on vision without inversion of the retinal image. Psychological Review, 3(6), 611-617.

Suzuki, K., Roseboom, W., Schwartzman, D. J., & Seth, A. K. (2017). A deep-dream virtual reality platform for studying altered perceptual phenomenology. Scientific Reports, 7(1), 15982.

von Uexküll, J. (2010). A foray into the worlds of animals and humans (J. D. O’Neil, Trans.). University of Minnesota Press. (Original work published 1934)

We talk about artificial intelligence as though it were something you build and then own. A finished object. You train the model, you ship it, and now you have it, the way you have a bridge once the last bolt is torqued. This is the quiet assumption underneath almost everything said about the field of AI, and I have come to think it is the deepest error we are making. Not an error about how far the technology can go. An error about what kind of thing intelligence is in the first place, which is the more fundamental question, and the one few are asking.

A thousand years ago someone was asking a version of it. Ibn Sina, the Persian polymath the Latin West would come to call Avicenna, was born near Bukhara around 980 and spent a life moving between courts as a physician, an astronomer, a vizier, and above all a philosopher of being. He wrote a medical encyclopedia, the Canon, that was translated into Latin and taught as the standard text in the universities of Montpellier, Bologna, and Paris for the better part of five centuries, well into the sixteen hundreds. His philosophical summa, the Book of Healing, ran so deep into the nature of existence that it shaped European scholasticism directly, Aquinas included. Ibn Sina is one of the few genuinely foundational minds in the history of the subject, and near the center of his metaphysics sits a distinction that cuts straight to the mistake we are making now.

He noticed there are two completely different ways a cause can relate to its effect. Some causes bring a thing into being and then leave. A builder raises a house, and when the builder dies the house still stands. He was the cause of the house coming to be, but he was never the cause of the house continuing to be. The house does not need him anymore. He was, in Ibn Sina’s terms, an accidental cause, one that runs ahead of its effect in time and then drops away.

There is a second kind of cause, and it does not get to leave. Its effect exists only for as long as the cause is actively holding it in existence, at every moment, without pause. The instant the cause stops, the effect does not fade or decay or wind down. It simply is not there anymore. Ibn Sina called this the cause that bestows existence, and he argued it has to be simultaneous with its effect, present at the same instant, continuously, or the effect collapses into nothing.

Think of light. Not the bulb, the light itself. The bulb is an object that sits on a shelf, but the light is not an object at all. It exists only in the act of being produced. Cut the current and it is gone in the same instant, not slowly, not partially, gone. Light is a sustained thing. It is happening, or it is not.

Here is why an eleventh century distinction suddenly matters. A large language model is built exactly like a house. There is a period of construction, which we call training, enormous and expensive, and then the construction ends. The weights freeze. The builder walks away. What is left is a finished structure that persists on its own, unchanging, the same tomorrow as today, a magnificent house standing in a field long after anyone stopped working on it. The training was a cause of its coming to be. Nothing is a cause of its continuing to be, because it does not need one. That is what it means for the weights to be frozen. The thing is done.

You will hear that this is no longer true, that the newest systems learn as they go. Look closely and the house is still a house. A longer context window is a bigger room, not a live current, and whatever the model takes in during a conversation is held for the length of that conversation and then swept out when it ends, so the model on the far side of the reset is precisely the one that walked into it. Retrieval bolts a library onto the house and lets the model read from it, but reading is not being changed. Fine-tuning does change the weights, and it changes them the way a renovation changes a house, in a discrete campaign of construction that begins, ends, and freezes again, after which the builder walks away one more time. None of it is the thing a mind does, which is to be continuously altered by the world in the present tense, without pause, as a condition of existing at all. The impression of a living, updating system is the projection. The frozen structure is the fact.

In Ibn Sina’s terms, a mind is not a house. A mind is light.

And this is not only a metaphor. It is close to a literal description of what a mind physically is, and the science that says so is some of the deepest we have. There is a family of things in nature that exist the way light exists, only in the act of being sustained. A flame. A whirlpool in a river. A living cell. The physical chemist Ilya Prigogine won a Nobel Prize for the study of them and gave them their name, dissipative structures, patterns that hold their shape only by continuously taking in energy and matter and passing disorder back out, standing against the universe’s constant pull toward evenness and decay. Stop the flow and the pattern does not persist in a weakened form. It vanishes, because it was never a thing, it was always a process wearing the appearance of a thing. A whirlpool is the clearest case. It has an edge and a shape and a location, it can spin in one spot for hours, you can point straight at it, and yet it is not a thing at all. It is only water being pulled through a particular pattern, and the water passing through it is never the same water twice. Dam the river upstream and the whirlpool is not damaged. It is simply gone, because it was never anything but the river moving. Erwin Schrödinger, the physicist whose wave equation helped found quantum mechanics, spent his later years on a different question, what separates the living from the dead, and set his answer down in a small and famous 1944 book called What Is Life? An organism, he wrote, stays ordered only by continuously drawing order from its surroundings, and the moment it cannot, it slides into the flat equilibrium we call death. A mind is the most intricate dissipative structure we know of. It is a flame made of prediction.

Karl Friston, a neuroscientist at University College London and one of the most cited scientists alive, gave this its exact modern form in what he named the free energy principle, and it is the engine underneath everything I have been arguing across these pieces. A system that maintains any boundary between itself and the world, that stays a distinct thing rather than dissolving into its surroundings, can only do so by continuously minimizing the gap between what it predicts and what the world actually delivers. It has to keep guessing the world and keep being corrected, moment to moment, or the boundary fails and the thing stops being a thing. Existence, for a mind, is not a state you arrive at and hold. It is an act you perform without stopping. Grounding, then, is not a foundation you pour once during construction and then build on top of. It is the current. It has to be live, or it is nothing.

No intelligence in nature was ever built once and then left running. While there are hardwired instincts, fixed behaviors an animal is born already knowing, those instincts do not run on frozen parts. They run on living tissue that is active every instant the animal is alive, continuously firing, continuously modulated by the world around it, never left decoupled from reality the way a trained model is, because a nervous system on pause is the end of that intelligence, an extreme surprise, in Friston’s terms, from which there is no return.

Now picture two drones sent through the same stretch of forest. The first was trained exhaustively beforehand and flies on a policy fixed at the factory, a beautiful house of a system, and for a while it is astonishing, right up until it meets a branch at an angle that was not in the world it was built from, and then it has nothing, because the builder already walked away and left no one home to adapt. The second is not finished, because it was never meant to be. It is in the loop, seeing and correcting and seeing again in the present tense, and when the branch comes at the wrong angle it does what everything alive does, it adjusts to a world that did not consult its training, and it threads through. One of those is running on a fire that is still lit. The other is a photograph of flight.

In the story of Oz, the wizard’s power turns out to be an apparatus, a machine he built once to throw an enormous, overwhelming face onto the wall and then simply ran, projecting the same image on a loop, the same today as yesterday, asking nothing of him but that he keep working the controls he had already installed. A dazzling projection, finished, self-contained, with no contact to the world outside the room it performed in. That is the thing we are building, and getting better and better at building. The travelers standing in front of it were the opposite kind of thing entirely, alive to the road, changed by every step of it, corrected the whole way by a world that pushed back. One of those is a mind. The other is a very good machine for working the levers.

Later in his life Ibn Sina ran a second thought experiment, one that has never stopped being debated. Imagine a person created fully formed in a single instant, suspended in empty air, in total darkness, touching nothing, feeling not even the weight of their own body, with no memory and no sensation of any kind. Like a human clone formed and awakened within a sensory deprivation tank. His claim is that this person, with every input to the mind stripped away, would still affirm one thing without hesitation: that they exist. That there is someone there. From this he argued that the self is not the body and is not assembled out of the senses, because it can be known when the body and the senses are gone. Six hundred years later the French philosopher René Descartes would arrive at nearly the same place through a different door, stripping away every belief he could bring himself to doubt until only the doubter was left, and give it the phrase everyone remembers, I think therefore I am.

Now hold his two arguments side by side, because the same mind made both and they seem to pull in opposite directions. One says a mind exists only in continuous contact with the world and vanishes the instant that contact is cut. The other says the self remains even when all contact is cut. That looks like a contradiction. It is actually the most important distinction, and it is the whole architecture of what we are building.

The two arguments were never about the same thing. The sustaining cause, the light that is not a house, is about intelligence. The working mind. The thing that perceives and acts and reasons and threads a drone through a forest. That is a live process, a flame, and it cannot survive one instant cut off from the world, because contact is the very thing holding it in existence. The Flying Man is about something else entirely. It is about the witness. The bare fact that there is someone there at all, that it is like something to be you, the awareness that would remain in the dark after every input was stripped away.

And notice what kind of thing that witness is, because Ibn Sina put his finger on its strangest property. You can never make it an object. You cannot turn around fast enough to catch the one doing the looking, cannot get behind the seeing to see the seer. It is unknowable, not in the sense that it is hidden or mysterious, but in the exact sense that the knower can never be the known. You do not observe it. You are it. It is the one thing in the universe you have by being it rather than by looking at it. Call it the unknowable silent witness. Silent because it never speaks, and unknowable because it can only ever be the subject of knowing, never the object.

There is a witness, which can only ever be a subject. There is a world, a real physical territory, which can only ever be an object, and which the witness can never touch directly, because a bare awareness cannot grip a world, and a bare awareness pressed against raw reality is not a life. So between the two sits intelligence, and intelligence is a space suit. It is the only thing that lets the witness reach a world it could not survive touching with its bare hand. Perception, prediction, the body, the loop that acts and is corrected, all of it is the suit. And the suit is exactly the thing Ibn Sina described a thousand years ago. It is light, not a house. A living process that has to stay in contact or it is nothing. The witness rides inside of it. The world pushes back on the outside of it. The suit is what couples the one to the other, and grounding is simply the suit being worn and used to explore this reality.

This is why intelligence and consciousness were never one thing. A rover on Mars, driven by our best minds, grips a world no human could survive touching bare-handed. It is a suit with intelligence inside it and no one home. The witness, the inner light the Flying Man points to, is unknowable by nature. Building the suit does not require answering it. That is the part nature has already solved, in millions of forms, and the part we can build. I believe we can build intelligence past our own, but only as an interface, never as a consciousness in itself.

So then the way forward is a different kind of thing, an intelligence that is never finished because it was never meant to be, one that exists the way everything alive does, by staying in contact with the world and being corrected and punished by it, moment after moment, in the present tense, and without the expectation of consciousness. We call it Enactive AI, eAI, intelligence grounded in acting on and perceiving the world rather than in predicting the next word about it. That is what we are building at Absentia Technologies. Not a house. Not a projection thrown on a wall. A way to broaden our experience of reality and our own abilities to experience this world.

Read more about what we are building at Absentia Tech → absentiatech.com

Emanouil Angelov is Co-Founder of Absentia Technologies and a screenwriter who has been writing films about artificial intelligence since 2017. His background spans professional filmmaking, cinematography, photography, marketing and teaching. His work at Absentia is informed by the intersection of visual perception, linguistic theory, and AI architecture.

In Baum’s novel the travelers reach the Wizard one at a time, and he is never the same thing twice. Dorothy meets an enormous floating Head. The Scarecrow meets a beautiful winged woman. The Tin Woodman meets a terrible Beast. The Cowardly Lion meets a Ball of Fire. Each of them trembles before the form best suited to overwhelm them. Then Toto, who is not afraid because he does not understand enough to be afraid, knocks over a screen in the corner of the room, and behind it stands a small bald man in his shirtsleeves, working a set of levers and speaking through his hands. “I have been making believe,” he tells them. “I am just a common man.” He has a word for what he is. Humbug.

The Wizard shows each visitor a different face, and this is not Baum being whimsical. We are pattern-completing animals. We see intention where there is only incident, and a mind where there is only mechanism, and the same instinct is now turning a statistical text engine into Pandora’s box. Like the travelers in the throne room, we each see the form best suited to overwhelm us. The optimist sees the winged woman, intelligence about to lift the species. The doomer sees the Beast. The investor sees the great Head that will answer any question if you kneel in front of it long enough. Same machinery in the corner. Different projection on the screen.

Before anyone is allowed into the Emerald City, the Guardian of the Gates locks a pair of green spectacles onto their face, even onto Toto, and keeps the key on a chain. The city is not green. It is ordinary stone and ordinary glass. It looks like a fortune in emeralds because everyone is required to view it through green lenses, and because no one is permitted to take them off. I think that’s a great allegory for the present market. The capital, the demos, the leaderboard scores, the launch events, the breathless coverage, the valuations with no floor under them. These are the spectacles. They are issued at the gate. You cannot raise a round, book a meeting, or publish a take without putting them on first. And green, as it happens, is the color of money. The city glitters because we agreed to wear the lens that makes it glitter, and almost everyone standing inside it holds equity in keeping the glasses locked.

When the humbug is exposed, he still has to deliver, so he gives each traveler what they came for. He opens the Scarecrow’s head and packs it with bran stuck full of pins and needles and calls them bran-new brains. He gives the Tin Woodman a heart of silk stuffed with sawdust. He pours the Lion a green liquid and tells him it is courage. None of it is real. All of it works, because the recipients believe it does. This is benchmark culture exactly. We open the model’s head, pour in another trillion tokens, post a higher number on a test we wrote ourselves, and announce that reasoning has arrived. The capability that gets advertised is the bran and the pins. It is a symbol of the thing. But a map is not the territory, in Alfred Korzybski’s famous phrase, and I do not think a high score is a mind.

Notice what the Wizard actually is. A voice and a face with nothing behind them. He drifted into Oz in a balloon, by accident, and stayed because the people decided he was magic. He has no contact with the world outside his throne room and cannot step out of it without being found out. This is the part the metaphor gets exactly right. A system trained on the residue of human language has read every account of fire and never once been near the heat. Give it a camera and you have not given it contact, only more residue, because a photograph is a description too. It has touched nothing it describes.

So when it does not know, it does not stop. It does what the Wizard does. It projects the most convincing image it can and speaks with total authority through its hands. We call this hallucination, as if it were a defect, a bug to be patched in the next release. It is not a defect. It is the whole nature of the machine, and to see why you have to look at the one thing that does this well, which is the brain, and find the single part it has that the model does not.

Here is what a century of neuroscience keeps finding, and it is stranger than the marketing. The brain is not a camera that takes in the world. It is a guessing engine running a controlled hallucination, predicting the world and checking the prediction against what arrives. Helmholtz saw it in the nineteenth century and called perception an unconscious inference. The modern name is predictive processing, and the claim is blunt. Your brain is always generating its best guess of what is out there and correcting only the error, the gap between what it expected and what it got. You can watch this happen in the simplest way there is. You cannot tickle yourself, and the reason is that your brain predicts the exact sensation your own moving hand is about to produce and quietly cancels it before it lands, keeping only what it did not see coming. Rig up a machine that adds a half second delay between your movement and the touch, so the prediction no longer matches, and suddenly you can tickle yourself after all. The brain is never feeling the world directly. It is feeling the difference between the world and its own guess. Your experience of reading this sentence is a controlled hallucination, a guess your cortex is making and fixing in real time. And Dorothy’s Oz was a controlled hallucination too, the way every world is. The difference is that Oz could fight back. She had to adapt to every surprise it threw at her, and being corrected by it, step after step, was the only thing that ever got her home.

So prediction is not the flaw. Prediction is the method. A language model and a human cortex are doing something unnervingly alike, each producing the most probable next piece of a pattern. The brain’s guess is dragged back and corrected by a body that acts on the world and pays for being wrong. Reach for the pan, feel the burn, never reach the same way again. The error comes from outside, from a world that pushes back and does not care what you predicted. The model’s guess is corrected by nothing but more text. Its loop is closed. There is no outside, no floor, no burn. It picks the most probable next word, it does not consult a world to check the last one, and nothing inside it knows the difference. A guessing engine with no world to answer to will guess straight into the void, fluently, forever.

Computer science has a precise name for this and no cure for it. Stevan Harnad called it the symbol grounding problem in 1990. How could a symbol ever mean anything to a system that only ever trades it for other symbols, none of which it can cash out for the thing itself? More than three decades and a thousandfold more compute later, not one symbol has been grounded. When a problem will not dissolve no matter how much scale you throw at it, you are not short on compute. You are digging in the wrong field.

When Dorothy turns on the little man, he does not deny what he is, but he says something true. “I am a very bad Wizard, I must admit, but I am a very good man.” The models are good men. They are genuinely useful, and most of the people building them are serious and honest about the limits once you get them off the stage. The humbug is not the technology. The humbug is the performance we agreed to stage around it, the floating Head we keep building in front of an ordinary man at a control panel, because a control panel does not move the valuation. You know what does move it: Skynet, HAL 9000, Ava, the list is long. It is exciting because it is dangerous. But let us leave science fiction to the storytellers.

There is one more scene worth remembering. When the Wizard finally tries to leave, he does it the only way he knows, by filling a balloon with hot air. The ropes slip before Dorothy can climb aboard, and he rises up over the city and floats away, waving, leaving everyone precisely where he found them. That is how this kind of thing tends to end. Not with a crash but with a slow ascent on hot air, the showman carried off by the same medium that lifted him, and the crowd left standing in a city that was never made of emeralds.

Which brings me to the shoes. Dorothy crossed a continent and risked her life to beg the great and powerful Oz for the power to get home, which she had on her own feet the entire time. The silver shoes were the grounding she always needed to get home. Three clicks and she was back in Kansas.

And here is what I missed as a seven year old, reading with the green glasses I did not know I had on. The Wizard was the only fake in the book. Everyone else was the real thing. The Scarecrow spends the whole road out-thinking every obstacle in the party’s way, which is the one thing a creature with no brain is not supposed to do. The Tin Woodman is the one who feels everything, who weeps when he steps on a beetle and rusts his own jaw shut. The Lion walks straight at every danger between the cornfield and the Emerald City. They earned their minds and hearts and nerve the only way anyone ever has, by moving through a world that could hurt them and being changed by what it did. They too were grounded. The Wizard, who never left his room, who touched nothing, who only ever projected and spoke, was the one running on hot air. In a way Baum wrote the plot of today’s AI craze. The travelers are what intelligence actually looks like. The man behind the curtain is what we are building on hype.

So the way home was never in the throne room, and it was never a larger Head. It has been sitting in plain view in a different field the entire time. Not in computer science, which built the throne room and keeps adding rooms to it, but in biology, which has spent a century studying the only intelligences that have ever existed. They have one thing in common. Not one of them learned the world by reading about it. They scaled it by moving through it.

Jakob von Uexküll named this a hundred years ago. Every creature lives inside its own Umwelt, the slice of the world its body and senses can act on, and nothing past that slice exists for it at all. A tick’s entire world is three signals. A crow sees colors we have no words for, because it carries a fourth kind of cone in its eye, so its world holds distinctions ours cannot. There is no view from nowhere, no neutral world waiting to be downloaded whole. Perception is not a recording. It is an act, performed by a particular body, that brings a particular world into being.

The psychologists Kevin O’Regan and Alva Noë argued you do not see by receiving a picture at all. You see by knowing, in your body, how the scene would shift if you moved, and you keep almost none of it stored inside you, because the world is sitting right there to be checked again the instant you need it. Francisco Varela gave the whole idea its name, enaction, mind not as a mirror held up to a finished world but as something that brings a world forth by acting in it. Karl Friston gave us the Free Energy Principle, the machinery of perception and action folded into a single loop, the organism moving through the world in order to close the gap between what it predicts and what it meets in order to avoid surprise, which equals death. The prediction only works because there is a body to correct it. Take the body away and you are left with the Wizard and hot air.

The thing the AI field is chasing in the throne room cannot be found there, because it was never made of language to begin with. It is made of contact. The silver shoes represent a body to interface with reality. Intelligence was never going to be downloaded from a corpus of text, however vast, for the same reason Dorothy was never going to find home by asking the Head. The power was always in the feet. In grounding to reality.

Neural networks got us here, and that matters. But the road forward is Enactive AI (eAI), intelligence that learns the way everything intelligent has ever learned, by acting in the world and being corrected by it rather than by predicting the next word about it. That is what we are building and researching at Absentia Technologies. We believe the architecture of intelligence was worked out long ago, by evolution, in the only laboratory that has ever produced a mind, nature.

Read more about what we are building at Absentia Tech → absentiatech.com

Emanouil Angelov is Co-Founder of Absentia Technologies and a screenwriter who has been writing films about artificial intelligence since 2017. His background spans professional filmmaking, cinematography, photography, marketing and teaching. His work at Absentia is informed by the intersection of visual perception, linguistic theory, and AI architecture.

Soviet filmmaker Lev Kuleshov intercut the same expressionless close-up of an actor’s face with three different images: a bowl of soup, a woman in a coffin, a child playing. Audiences watching each version saw the actor expressing hunger, grief, and joy, respectively. The actor’s face did not change. The emotion was entirely in the viewer’s mind, generated by context and adjacency.

The face was not communicating anything. The audience was doing all the work.

I have been thinking about that experiment as I watch frontier AI systems deployed in contexts where the distinction between genuine causal reasoning and plausible-sounding pattern correlation has real operational consequences. The AI produces output that looks like reasoning. The audience perceives reasoning. The system never actually reasoned about anything.

This is not the same as hallucination. Hallucination gives you a wrong answer. What I am describing gives you a coherent answer that was never actually reasoned through. The output looks right. The thinking behind it never happened.

Correlation Looks Identical to Causation Until It Doesn’t

The first failure is the deepest.

A model trained on enough data will learn that certain patterns reliably co-occur with certain outcomes. Fire reports co-occur with damage assessments. Certain tactical configurations co-occur with mission success. Certain linguistic signals co-occur with deception. The model learns these correlations at such scale and with such fidelity that its outputs, when describing relationships between things, sound exactly like causal explanation.

The problem is that correlation and causation produce identical-looking outputs when the pattern holds. The failure only becomes visible when the pattern breaks, which is precisely the condition under which the output matters most.

Consider what this means in a medical AI context, where documented failures have followed exactly this structure. A system learns that a certain demographic variable reliably co-occurs with lower rates of a particular condition, not because of any biological mechanism, but because that demographic historically received less diagnostic testing. The model learns the correlation perfectly. It recommends against testing. The condition goes undetected. The output was confident, fluent, and coherent. It was wrong in a way the model had no mechanism to recognize because it had no access to the causal structure underneath the pattern.

This failure does not resolve with scale. Some researchers argue that sufficiently large models develop emergent causal reasoning; the evidence so far suggests that what actually emerges is better pattern matching over a wider distribution, which performs like causal reasoning when the distribution holds and fails like pattern matching when it doesn’t. The distinction matters precisely at the boundary, which is where consequential decisions live.

The Explanation That Explains Nothing

The second failure is subtler and more insidious in deployment.

Ask a frontier model why something happened and it will produce an explanation. The explanation will be fluent, internally consistent, and often persuasive. It will cite factors that are genuinely associated with the outcome. It will arrange them in a causal-sounding narrative.

What it will not do is tell you whether the explanation is right.

The model cannot distinguish between a genuine causal account and a plausible post-hoc rationalization, because both are just patterns in language about causal relationships. It has been trained on millions of explanations, good ones and bad ones, and it has learned what explanations look like. It has not learned how to verify them.

In 2023, a legal AI system confidently cited six court cases as precedent in a brief filed with the Southern District of New York. None of them existed. The attorneys who filed the brief trusted the output because it looked exactly like what genuine legal reasoning produces. The explanation was structurally indistinguishable from a real one until a judge went looking for the cases. In operational contexts where you cannot go looking for the cases, that failure is invisible until it isn’t.

The deeper issue is not the specific error. It is that the system had no internal signal that anything was wrong. The fabricated citations and the real ones generated with the same confidence profile, because confidence in these systems tracks output plausibility, not output accuracy. An explanation that sounds like analysis is analysis, as far as the model is concerned.

The Inference That Stops at the Edge of Training

The third failure is the most practically important for anyone deploying these systems in novel environments.

Genuine reasoning transfers. A person who understands the causal principles governing a domain can encounter a genuinely novel situation, one that has never appeared in any training data, and reason correctly about it. The principles apply. The understanding scales.

Pattern matching does not transfer this way. It extrapolates, which is different. Extrapolation from a learned distribution performs well when the new situation is close to the distribution and degrades when it is far from it. The model has no mechanism for knowing how far it is from its training distribution on any given input. So it extrapolates with the same confidence regardless.

This is the structural reason frontier AI systems fail unexpectedly in novel operational contexts rather than failing gradually. Close to the training distribution, they perform well. At the edge, where the stakes are often highest, they fail fast and confidently. The confidence does not track the distance from known ground. It tracks the plausibility of the output.

What matters is that the systems currently deployed in consequential roles are not pointing anywhere new. They are predicting tokens and calling it understanding.

When the Kuleshov Effect Runs Operational

The synthesis of these three failures produces something specific and dangerous.

A system that cannot distinguish correlation from causation, that produces fluent post-hoc rationalizations indistinguishable from genuine analysis, and that fails at the edge of its training distribution with full confidence rather than appropriate uncertainty is not a tool that produces wrong answers occasionally. It is a tool that produces wrong answers while presenting every surface signal of right ones.

The Kuleshov Effect works because the human perceptual system constructs meaning from context automatically, before critical evaluation engages. We are built to complete the picture. A system that has mastered the surface signals of reasoning exploits exactly this tendency. The face is still expressionless. The inference is still yours.

The question this raises is not whether these systems should be deployed but what they should be deployed for. For tasks where the substrate is language and the requirement is pattern completion across a known distribution, they remain the best tools ever built. For tasks where the requirement is reliable inference in novel environments, grounded understanding of physical causation, or honest self-assessment of the limits of their own reliability, the architecture is wrong. Not insufficiently scaled. Wrong.

The deployments most at risk are the ones where nobody stopped to ask whether the reasoning was real or just convincing. That is a harder question than it sounds. Convincing and real produce identical output right up until the moment they don't. By then the decision has already been made.

If you want to see what building what physics-based AI looks like in practice, follow absentiatech.com.

Emanouil Angelov is Co-Founder of Absentia Technologies and a screenwriter who has been writing films about artificial intelligence since 2017. His background spans professional filmmaking, cinematography, photography, marketing and teaching. His work at Absentia is informed by the intersection of visual perception, linguistic theory, and AI architecture. To learn more, visit. absentiatech.com.

I have spent years working at a level where the gap between what an audience perceives and what was actually captured becomes a technical problem to solve. I can look at footage and tell you the resolution, the camera it was shot on, the lens, the lighting setup, the post-processing choices made in the grade and beyond. I have been the person that expert cinematographers and photographers brought their unresolvable problems to, at places like Samy’s Camera and Division Camera. At that level, you stop seeing the illusion and start seeing the machinery underneath it. The gap between what an audience believes they saw and what actually existed in front of the lens is not an accident. It is a craft.

In my previous post I argued that the AI industry has scaled a communication tool and called it intelligence, and that the only honest path to genuine machine intelligence runs through real-world sensory grounding, not language. This post is about what happens when that gap, between what AI is and what we have decided to believe it is, gets deployed at scale into the real world. We are living through what I think history will call the Age of False Intelligence. Not because the technology is fraudulent. Because the story we are telling ourselves about it is. And we are filling in the dark frames with science fiction instead of science.

The Audience Completes the Picture

Cinema works because the human brain is not a recording device. A camera captures light mechanically and stores what it receives. The brain does something fundamentally different: it maps. It builds an active, predictive model of the environment from incomplete sensory input, fills in what is missing based on prior experience, and presents the result to consciousness as seamless reality. This is why you perceive continuous motion from a series of still frames. The continuity is not in the footage. It is in you.

The same mechanism that makes cinema possible makes us extraordinarily susceptible to a specific kind of deception: anything that presents enough of the surface pattern of intelligence will be processed as intelligence. Before scrutiny arrives. Before the critical faculty engages. The projection happens first, automatically, because that is how the mapping system works.

Large language models produce language that sounds exactly like it came from a mind that understands the world. They predict the next token in a sequence based on statistical patterns and learned vector relationships across a training corpus so large it encompasses most of recorded human thought, and nothing else. They have no world model. No ground truth. No persistent representation of reality beyond the context of a single conversation. What they have is the surface pattern, more convincingly rendered than anything we have ever built.

Our brains do the rest. We project understanding onto fluency. We assume that something which speaks like it knows must know. We fill in the dark frames with what we expect to find. The result is a collective illusion running at civilizational scale, and most of us are too deep inside it to see the cuts.

We are not interacting with a mind. We are co-authoring the illusion of one, filling in the gaps the way an audience fills in the frames, and mistaking our own completion for the film.

The Cinematographer Who Left the Set

In November 2025, Yann LeCun walked out of Meta.

LeCun is not a peripheral voice. He is one of three researchers who won the Turing Award in 2018 for the foundational work on deep learning that every major AI system is built on. He spent twelve years running Meta’s AI research division. And for the last three of those years he had been saying publicly, without diplomatic softening, that the industry was filming the wrong scene.

LLMs, he argued, are a statistical trick. Systems that predict the next word so convincingly they pass for understanding, while understanding nothing. His alternative, the Joint Embedding Predictive Architecture, JEPA, does not predict tokens. It learns how environments change over time, building representations of physical cause and effect, something closer to what the brain actually does when it models the world rather than describes it. The difference between a world model and an autocomplete is the difference between a mind that understands a scene and a camera that records its surface.

Meta had made its choice. In June 2025, Zuckerberg invested $14.3 billion in Scale AI and installed its 28-year-old CEO to run a new LLM-focused division, Meta Superintelligence Labs. LeCun, one of the most decorated researchers alive, would report to him. By November the distance between their views of what the technology actually was had become irreconcilable.

On March 10, 2026, four months after leaving, LeCun’s new company AMI, Advanced Machine Intelligence, closed a $1.03 billion seed round at a $3.5 billion valuation. The largest seed round ever raised by a European startup. Twelve employees. No product. Investors include Bezos, Nvidia, Samsung, Toyota, Berners-Lee, Cuban, and Schmidt. Every check was a bet on one thesis: that the current paradigm is filming the wrong scene, and that the audience is going to notice eventually.

LeCun’s stated timeline: within three to five years, world models will be the dominant architecture. Nobody in their right mind will use LLMs of the type we have today.

When the Dark Frames Are Operational

The Anthropic-Pentagon dispute is where the cinema metaphor stops being comfortable.

By late 2025, Claude was the only frontier AI system cleared for classified Pentagon networks. Intelligence analysis. Cybersecurity. Operational planning. The military had decided the film was real and built their workflow around it.

Then negotiations broke down. The Defense Department wanted unrestricted use for all lawful purposes. Anthropic refused two conditions: autonomous lethal weapons with no human in the targeting loop, and mass surveillance of American citizens. On February 27, Trump ordered federal agencies to stop using Anthropic’s products. Hegseth designated the company a supply chain risk. Anthropic sued.

Dario Amodei said it plainly in Anthropic’s public statement: frontier AI systems are not reliable enough to power fully autonomous weapons. He was right to hold that line. But the question that statement opens is the one nobody was publicly asking: if they are not reliable enough for autonomous weapons, what were we assuming about their reliability in the intelligence analysis and operational planning roles they were already filling?

A system with documented hallucination problems, no causal world model, no grounded understanding of the physical environments it was analyzing, embedded in the most consequential decision infrastructure on earth. The fight that became visible was about safeguards. The fight that did not become visible was about whether the technology was what anyone involved actually believed it to be.

That is not a gap in a film. That is a gap in a classified network. And the audience cannot fill it in.

The darkness between the frames is not a problem when you are watching a movie. It is a problem when you are making operational decisions inside it.

What the Audience Is Losing

The military context is the most extreme version, but the same dynamic is running everywhere these systems have been handed authority.

A 2025 peer-reviewed study by Michael Gerlich at SBS Swiss Business School, 666 participants across age groups and educational backgrounds, found a statistically significant negative correlation between AI tool usage and critical thinking scores. The mechanism identified was cognitive offloading: when a tool completes the picture for you, you stop developing the ability to complete it yourself. The effect was sharpest in younger users, highest reliance on AI, lowest scores on independent reasoning.

A Microsoft and Carnegie Mellon study found knowledge workers who trusted AI outputs engaged in dramatically less critical evaluation of the work they were accepting. An MIT Media Lab study found that students using AI writing tools showed measurably reduced engagement in brain regions associated with learning and memory consolidation compared to peers working without them.

This is the cost that does not show up in the benchmark comparisons or the earnings calls. When you outsource the completion to the machine, you stop building the capacity to complete anything yourself. The audience gets better at watching and worse at seeing. We are not just projecting intelligence onto a tool that does not have it. We are slowly offloading our own in the process.

The Name Is the Problem

LLMs are genuinely useful. Communication, drafting, summarization, coding assistance, research synthesis, translation. For anything where the substrate is language and the requirement is rapid pattern completion across a large corpus, these systems perform at a level that is real and will continue to improve. Treated honestly as the best information retrieval and language interface ever built, they deserve real respect.

The problem is not the tool. The problem is that we named it wrong.

We called it artificial intelligence because the name was exciting, because it attracted capital, because it made the story bigger than the product. That name carries an entire cosmology of science fiction and philosophical implication that the technology does not inhabit and cannot. And now the story is running the decisions. The story is what got embedded in classified networks. The story is what a billion dollars just bet against. The story is what the cognitive studies are now measuring the cost of.

The filmmaker in me recognizes exactly what this is. Sergei Eisenstein built an entire theory of cinema, montage, around the idea that meaning is not in the shot. It is in the cut between shots. The audience’s mind creates the meaning from the juxtaposition, not from anything that was actually filmed. The AI industry has been running the same play. The intelligence is not in the model. It is in the gap between what the model produces and what the audience’s mind completes it into. The cuts are invisible. The gaps disappear. The audience has been doing the work for years and crediting the screen.

The correction is coming. Performance saturation is appearing at the frontier. Benchmarks are losing credibility. The military is finding that fluency and reliability are not the same thing, and diverge exactly when it matters most. LeCun raised a billion dollars on the thesis that the foundation is wrong. The cognitive data is now a trend, not an anomaly. When an audience starts noticing the cuts, the spell breaks fast.

The companies that survive will be the ones that were honest about the distinction from the start. Building for environments where the hallucination is not a product limitation to be managed but a mission failure that cannot be tolerated. Building intelligence grounded in the actual world rather than in the statistical shadow of language about it.

At Absentia that has been the frame since day one. Vision first, because light bouncing off matter is the most direct, highest-bandwidth signal the real world produces. No gap-filling. No projection. No dark frames passed off as footage.

The next post will take up the vocabulary the industry has built around AI, and why the terms we have inherited, AGI chief among them, now obscure more than they reveal.

If you want to see what building honest AI looks like in practice, visit absentiatech.com.

Emanouil Angelov is Co-Founder of Absentia Technologies and a screenwriter who has been writing films about artificial intelligence since 2017. His background spans professional filmmaking, cinematography, photography, marketing and teaching. His work at Absentia is informed by the intersection of visual perception, linguistic theory, and AI architecture. To learn more, visit absentia.tech.

That curiosity eventually found its way into my screenwriting and films. My

first AI screenplay, Paradise Ranch, co-written with Kiril Sabev, is set in a far future where technology has long since collapsed back to something analog, roughly 1980s in character. AI is far off history. Then a group of small-town kids stumble onto a doomsday shelter inside their local dam, and find the last functioning AGI on earth still running inside it.

My current screenplay follows the events, decades after humanity’s technological collapse, and the emergence of Ray Kurzweil’s theorized singularity in this technological desert.

Writing these stories demanded a kind of honesty that most public discussions of AI avoid. You cannot fake the internal logic of a world and expect people to believe it. To write about AI with any truth, you have to follow the ideas to their inevitable ends, face what they reveal, and resist the temptation to settle for what sounds exciting when the real answers are harder to imagine. So it forced me to think like a futurist.

Eventually I began to see a gap between what the industry says it is building and what it is actually building. Hype has shaped AI development as much as science has. Scaling prediction is not the same as scaling understanding.

The mainstream bet on the large language model as the final destination of AI is, for all its momentum, a wrong turn. There is a bubble forming around it. But AI will survive that bubble the way the internet survived the dot-com crash: the infrastructure will remain, the hype will clear, and the companies still standing will be the ones that built something real underneath it.

I come to all of this from an unusual combination of angles. I am bilingual, trained in cinematography and cameras, and spent years teaching English across cultures while learning foreign languages myself. Vision and language are domains I have worked in professionally and studied closely. When writing about AI forced me to reckon honestly with what intelligence actually is, the mismatch with the industry’s direction was not subtle.

The AI industry is not building intelligence. It is scaling a communication tool. To understand this fork in the road to AGI we have to consider the role sensory data and language play in our own intelligence.

The Wrong Layer, Built from the Wrong Stuff

Large language models are not mimicking intelligence. They are mimicking language itself, the communication tool that sits at the top of the stack, furthest from where understanding actually originates.

Consider Two-Face from Batman. He surrenders every decision to a coin toss. Not because the coin is wise, but because he has abdicated agency. The coin does not understand the problem. It does not weigh consequences or generate meaning. It simply resolves the paralysis of choice with a random output. And crucially, it always lands somewhere. Every flip produces an answer. Sound familiar?

An LLM operates the same way. Not randomly, but statistically. Given any input, it will produce a confident, fluent, well-formed output. It has no agency, no understanding, no stake in whether the answer is true. It is navigating a chaotic world of language patterns with a very sophisticated coin or token system if you like. The output lands somewhere every time. That is not intelligence. That is the absence of intelligence dressed in its clothing.

Real order in a chaotic world is not statistical. It is biological. It emerges from organisms that have skin in the game, that live and die by the accuracy of their perception, that built their understanding of reality through direct contact with it over millions of years of evolutionary pressure. Language is one of the last things that emerged from that process, not the foundation of it. Building AI on language and calling it intelligence is building on the most abstracted, most removed, most human-specific layer of a process that runs incomparably deeper.

Companies like OpenAI, Google, X and others are building that layer from the flimsiest possible stratum of human consciousness: language. Which is, by definition, a compressed and distorted transmission of what the mind actually understood. Void of understanding, void of agency, yet the bet is that if you scale it far enough, the mimicry becomes close enough to the real thing that nobody checks what is underneath.

Think about how you learned to walk, to catch a ball, to read a room. Nobody described these things to you in sentences and you reconstructed the skill from the description. You engaged with the world directly. Your sensorimotor system built representations through iteration, failure, and feedback from physical reality. The understanding came first. Language, if it arrived at all, came after as a way to communicate what the body already knew.

Intelligence is not a product of language. Language is a product of intelligence. It is what the conscious actor produces when drawing on the subconscious processing layer where actual understanding lives. When we build AI on language data, we are training on outputs of outputs: compressed descriptions of understood experiences, produced by an imperfect communication tool, fed into a model that has never had the experience itself.

We are not building intelligence. We are scaling a communication tool built from other people’s descriptions of experiences the model has never had.

The result is Two-Face’s coin, supremely optimized. It lands on an answer every time, fluent and confident, yet with no understanding of what was actually asked. Sometimes correct, sometimes fabricated, because the model carries none of the lived context every human brings to language without thinking: embodied experience, cultural grounding, the felt sense of what words actually refer to in the world. That context is not in the text. It never was. It lives in the people who wrote the text, and it did not survive the translation into tokens.

The model has no way to distinguish a true answer from a plausible one because it has no ground truth to check against. It works only with statistical patterns in text that were already one step removed from reality by being stripped of larger sense based contexts, before the model ever saw them. We are then surprised when it hallucinates. But hallucination is not a malfunction. It is the coin doing exactly what it was built to do, landing somewhere every time, with reasoning layers now trying to limit where and how it lands to better mimic real intelligence.

Alphabets, Hieroglyphs, and the Shape of Representation

To understand what a better architecture looks like, it helps to understand what alphabetic language actually is as a representational system.

Western written language descended from administrative necessity: counting grain, recording debts, tracking inventory. It was optimized for discrete quantification, breaking reality into separate labeled units that could be stored and retrieved. The alphabet is the endpoint of this process. An arbitrary set of symbols, each representing a sound, combined into words, each word a label attached by convention to some slice of reality.

There is no resemblance between the word ‘fire’ and fire. The relationship is entirely symbolic, abstract, and culturally agreed upon. The word does not carry the heat, the light, the danger, the warmth. It points at them, imprecisely, for people who already know what fire is.

Hieroglyphs, Kanji and Mayan script work on a different principle. They are graphic and visual. A hieroglyph for a bird looks like a bird. It does not just label the concept; it evokes it, positions it visually in relation to other elements, and communicates something of the thing itself rather than merely pointing at a convention. These scripts are more holistic because they are still partly image. The abstraction from physical reality is less complete.

Now consider the representational choice at the heart of modern AI. Language models tokenize text into discrete symbolic units and learn statistical relationships between them. Each token is the endpoint of alphabetic abstraction: a symbol stripped of any direct relationship to the physical reality it supposedly represents. The model learns which tokens tend to follow which other tokens across billions of examples. It does not learn the larger contexts behind the tokens. This is the symbol grounding problem in formal ML terms. Tokens lack grounded representations. They are not connected to their referents in the world in a way that translates real and inherent context.

Language models trained on text are working at maximum distance from the world they are supposed to understand.

What would a hieroglyphic approach to AI representation look like? It already exists in embryonic form. Vision-language models like CLIP learn joint embeddings across images and text, producing dense vector representations where concepts that are visually and semantically related occupy nearby regions of a shared high-dimensional latent space. Unlike discrete tokens, these embeddings are continuous and relational. The representation of ‘fire’ in a well-trained multimodal embedding space carries proximity to heat, light, danger, and warmth, because the model has processed images of fire alongside text about fire and learned the structural relationships between them.

This is closer to holistic representation. Not a label attached to a convention, but a position in a latent space shaped by both visual and semantic data about the thing. The vector is not the word and not the image. It is a learned compression of both. Still imperfect, still partially anchored to language through its training data, but pointing in the right direction: away from language as the primary modality and toward real-world sensory data as the foundation.

How a Child Actually Learns

The correct model for machine intelligence development is not the language model. It is the infant.

A child does not acquire understanding of the world by reading descriptions of it. They engage with it physically and sensorially: touching objects, tracking movement, learning causality through interaction, developing spatial understanding through navigation, building social models through face-to-face engagement long before language arrives. By the time a child can form sentences, they already have a rich internal model of physical and social reality built entirely from sensorimotor experience. In essence they are building their own LWM and Large Cultural Model.

Language then arrives as a tool for communicating and refining that model. It is extraordinarily useful in this role. But it did not build the model. The model was already there, constructed from direct engagement with the world.

In ML terms, this is sensorimotor grounding: building internal representations through iterative feedback between action, perception, and environmental response. The representations that emerge are genuinely grounded because they are causally connected to the physical world that generated them. Error signals come from reality, not from human annotation of text.

Scaling this approach rather than scaling language produces a fundamentally different kind of system. The compute efficiency argument is real: a model that learns grounded representations from sensory data does not need to process the entire expressive output of human civilization to achieve competence in a domain. It needs sufficient environmental exposure to that domain. The signal is richer per data point because it is grounded. The model learns faster because reality provides immediate and unambiguous error correction in a way that language data cannot.

A sensorimotor learning system can also run at machine speed through simulated or real environments, generating orders of magnitude more learning events per unit time than any language-based training process. What takes a human years of embodied practice can be compressed, not because the system is smarter in some abstract sense, but because it is iterating against reality rather than against human descriptions of reality.

Paradigm Shift: Intelligence Is our Sub-conscious Layer

I want to be precise about the claim I am making, because it is easy to misread.

I am not saying consciousness does not exist or does not matter. I am saying that what we call intelligence, the capacity for skilled understanding and effective action in the world, lives primarily in the subconscious processing layer. The conscious actor, the reflective, the freudian ego, language-using, narrating self, is a real and important phenomenon. But it is not the source of intelligence. It is an interface to it.

The ego can reflect on what the subconscious has produced. It can use language to communicate it, refine it, share it with others. Language is a tool of the reflective actor, a powerful and distinctly human tool. But it is a communication device, not a cognition device. The cognition happens elsewhere.

This matters for AI because it means the question ‘what is intelligence?’ has a different answer than the language model paradigm assumes. Intelligence is not sophisticated language use. It is sophisticated world-modeling, grounded in sensory reality, operating at speeds and depths the conscious actor cannot access directly. Language is how that intelligence talks to other actors. It is not the thing itself.

Every human accesses this layer. The musician who has practiced until the music plays itself. The operator who reads a threat environment before consciously registering it. The surgeon whose hands know where to go. The writer whose story seems to just flow through him and not from him. None of this is happening at the language layer. Language arrives after, to describe what intelligence already did.

Language is how intelligence communicates. It is not intelligence itself. We have spent close to a decade scaling the communication tool and calling it the mind.

The AGI Worth Building Has No Conscious Will

The path to AGI is not through language. It is through the same process that produced intelligence in the first place: direct, iterative engagement with the physical world.

Every intelligent creature that has ever existed built its understanding from the ground up. Sensory contact with reality. Action and consequence. Pattern recognition earned through experience, not extracted from descriptions of it. Intelligence at this layer has no agenda. It does not develop preferences or pursue goals. Agency is not a property of intelligence. It is what the ego layers on top after the fact. The subconscious does not want anything. It just understands.

Real AGI will work the same way. Not an agent that acts in the world on its own behalf, but an intelligence layer we tap into, the same way we are trying to interact with LLMs today, except grounded in reality instead of language. A substrate of genuine world-modeled understanding that operators and systems can query, draw from, and build on. Deeper, faster, and more accurate than anything built on text, because it was trained on the world itself rather than on human descriptions of it.

The reflective actor, the system with something resembling agency or free will, is a separate and far harder problem. It may be worth solving eventually. But you cannot solve it honestly until you have first built the intelligence layer it would run on. Right now the industry is trying to simulate the actor without having built the substrate. That is why current AI systems feel simultaneously impressive and hollow. The performance is there. The foundation is not.

Build the foundation first. Crack genuine AGI as an intelligence layer. Then, if it still seems like a good idea, tackle the harder problem of what it would mean to give that intelligence a will of its own.

That is what we are building at Absentia. Vision first, because light bouncing off matter is the highest-bandwidth grounded signal available. Extending into multi-sensor fusion to avoid the ouroboros of text-based training data. No actor in the architecture, because operators do not need AI that performs intelligence. They need AI that provides it.