Mahmoud HarmouchApr 10 2026

LLMs are Usefull. LMMs will Break Reality

#llms-are-usefull-lmms-will-break-reality

Hey everyone 👋,

In my previous post, Language is Limited. ASI is Impossible., I spent a long time explaining why language is not the same thing as thought, why words are not the same thing as understanding, and why a machine built on text alone will never cross the wall into true superintelligence. I still believe all of that, and I will not take any of it back, because the argument was honest and the logic was solid. But today I want to go further. I want to talk about something that has been sitting in my head for years, growing louder every day, and I need to get it out before it eats me alive. I want to talk about why large language models are still genuinely useful, despite their limits, and why large mathematical models, as introduced in this whitepaper draft, are something far more serious, something that could actually begin to crack the surface of reality itself. I know that sounds extreme, and I know some people will read that sentence and roll their eyes, but I am asking you to stay with me, because the argument I am about to make is not based on hype or fantasy. It is based on what I have seen, what I have built, and what I understand about the difference between describing the world and actually modeling the world. That difference is the whole point of this post, and once you see it clearly, everything else falls into place.

I have been thinking about this ever since I wrote An Empty Life Filled With Constant Suffering, where I talked about how words cannot fully capture my thoughts, and how language always falls short of the real thing inside our heads. That frustration is what led me here, because if language is limited, then we need to ask what comes after language, and the answer is not more language. The answer is structure, equations, simulation, and direct modeling of the physical world. That is what large mathematical models point toward, not because they are perfect today, but because they represent a direction that goes beyond text and into something much deeper. I wrote in It is always the Russians that the Russians took God's skin and wrapped it around a machine, and I stand by that image, because that is exactly what happens when you take human knowledge, strip it of its soul, and feed it into a system that only knows patterns. But what if the next generation of models does not stop at patterns in text? What if it starts learning patterns in the physical world itself? That is the question I cannot stop thinking about, and that is the question this post is built around.

LLMs Are Useful And I Am Not Going to Pretend Otherwise

Let me start with something that might surprise people who read my earlier posts. I think large language models are useful. I am not going to pretend otherwise, because pretending otherwise would be dishonest, and I have spent too much of my life being honest about hard things to start lying about easy ones. I wrote in As Engineers, LLMs should pay us for tokens usage about how the system exploits engineers, and I still believe that, but that does not mean the tool itself is worthless. A hammer can be used to build a house or to crush a hand, and the fact that someone charges you too much for the hammer does not mean the hammer cannot drive a nail. LLMs can summarize documents faster than I can read them, they can generate boilerplate code faster than I can type it, they can help me think through problems by acting as a sounding board, and they can translate between languages in ways that used to require expensive human translators. These are real capabilities, not illusions, and anyone who denies them is not being serious. I have used them myself, and they have saved me hours of work on tasks that would have otherwise drained me completely. The tool is real, the output is real, and the time saved is real.

But here is the thing that people keep getting wrong, and this is where the trouble starts. People confuse usefulness with intelligence, and those are not the same thing at all. A calculator is useful, but nobody calls a calculator intelligent, because we understand that it is following rules, not thinking. A search engine is useful, but nobody calls a search engine wise, because we understand that it is retrieving information, not understanding it. An LLM is useful in the same way, except the output looks so much like human speech that people forget they are watching a very advanced pattern machine, not a mind. The words come out smooth, the grammar is clean, the tone is confident, and that smoothness tricks people into believing there is understanding behind the words, when really there is only statistical prediction shaped by billions of examples. That confusion is dangerous, because it leads people to trust the output more than they should, and trust without verification is how mistakes become disasters. I have seen engineers deploy LLM-generated code without checking it, and I have seen the bugs that resulted, and those bugs were not small. They were the kind of bugs that come from blind faith in a system that sounds right but does not actually know what it is doing.

The usefulness of LLMs also has a ceiling, and that ceiling is made of language itself. I explained this in detail in my previous post, but I want to repeat the core idea here because it matters. Language is a compression of thought, not the thought itself. When I type a prompt, I am taking something rich and multidimensional from inside my head and squeezing it into a flat sequence of words. The model receives those words, runs them through its weights, and produces another flat sequence of words. At no point does anyone touch the actual structure of the problem. At no point does anyone engage with the real mechanism behind the question. The whole exchange happens inside a symbolic layer that sits on top of reality, never inside reality itself. That is fine for many tasks, because many tasks only need the symbolic layer. If I want a summary, I need text. If I want a translation, I need text. If I want a draft email, I need text. But if I want to understand why a bridge is failing, or how a drug interacts with a protein, or what the weather will look like in three days, I need something much deeper than text. I need structure, I need equations, I need simulation, and I need a model that can engage with the physical world directly, not just talk about it.

This is also why hallucination is such a serious problem and not just a minor bug that will be fixed with more data or better training. Hallucination happens because the model does not know anything in the way a human knows things. It does not have a grounded understanding of truth or falsehood, because it was never trained on truth or falsehood. It was trained on text, and text contains truth and lies in equal measure, all mixed together in a soup that the model cannot separate. When the model generates something false, it is not making a mistake in the way a human makes a mistake. It is doing exactly what it was designed to do, which is produce the most statistically likely next token given the context. Sometimes that produces truth, and sometimes it produces garbage, and the model cannot tell the difference because it has no access to the ground truth of reality. It only has access to the patterns in the text it was trained on, and those patterns include errors, biases, contradictions, and outright fabrications. This is not a fixable flaw in a fundamentally sound system. It is a structural limitation of any system that learns from text alone, because text alone is not enough to anchor a model to the truth of the physical world.

I also want to be clear about something that bothers me deeply. The companies that build these models know exactly what the limitations are, and they market around them instead of being honest about them, like what recently Sam Altman lied about here instead of being honest about calling it hallucination. They show you the best outputs and hide the worst ones. They put guardrails on the edges and call it safety. They charge you per token while extracting your knowledge and your judgment for free. I wrote about all of this in my post about how LLMs should pay engineers, and the core argument has not changed. The tool is useful, but the business model is exploitative, and the marketing is actively misleading people about what the tool can and cannot do. Every time a company says their model can reason, they are stretching the word reason beyond its breaking point, because what the model does is not reasoning in any serious philosophical or scientific sense. It is pattern completion, dressed up in confident language, and sold to a world that is too tired and too busy to look beneath the surface. That is not progress. That is a sales pitch.

Despite all of this, I want to be fair. LLMs have changed how I work, and some of those changes have been genuinely positive. They help me draft ideas faster, they help me explore different framings of a problem, they help me catch errors in my own writing, and they sometimes suggest connections I had not considered. These are real benefits, and I would be lying if I said otherwise. But benefits and limits can exist in the same system at the same time, and pretending the limits do not exist because the benefits are impressive is exactly the kind of sloppy thinking that I have been arguing against since my first post. The truth is simple. LLMs are useful tools trapped inside a symbolic cage, and no amount of scaling will break them out of that cage, because the cage itself is language, and language is not the world. That is not a small problem. That is the whole problem.

I also think people need to understand that the usefulness of LLMs does not prove that the path to artificial general intelligence runs through text. Usefulness and progress toward AGI are completely separate questions, and confusing them has led to some of the worst thinking in modern technology. A typewriter is useful for writing, but no one would say that building a faster typewriter is the path to creating a mind. A microscope is useful for seeing small things, but no one would say that building a bigger microscope is the path to understanding consciousness. LLMs are useful for generating text, but that does not mean that generating better text is the path to building a system that truly understands the world. The path to real understanding goes through structure, through mathematics, through simulation, and through direct engagement with physical reality. Text is a waypoint, not the destination. And the sooner people accept that, the sooner we can start building systems that actually matter.

Language Is Still A Cage And I Already Proved It

I wrote an entire post about this, Language is Limited. ASI is Impossible., and I am not going to repeat every argument here, because you can read it yourself and it still stands. But I want to draw out the parts that connect directly to what I am about to say about mathematical models, because the connection is the whole point of this post. Language is a compression layer, and compression always loses information. When I describe a sunset in words, I lose the colors, the temperature, the wind, the smell of the air, the feeling in my chest, and the thousand tiny details that make the moment what it is. The words get close, but they never arrive, and the gap between the words and the experience is where the real meaning lives. That gap is permanent, because language was never designed to carry the full weight of reality. It was designed to help humans communicate quickly, not to represent the world completely. And any system built on top of language inherits that same permanent gap.

This matters for AI because the whole dream of artificial superintelligence through text is built on the assumption that the gap can be closed with more data, more parameters, and more compute. I do not believe that, and I explained why in detail. More text gives you more text, not more understanding. You can pile up a billion books and you will have a very large pile of books, but you will not have a single atom, a single photon, or a single heartbeat. The pile of words is always about the world, never part of the world, and that distinction is not a technicality. It is the deepest structural limitation in the entire field. A model trained purely on text can become very good at producing text, and that is exactly what happened, and that is exactly why people are impressed. But being good at producing text is not the same as understanding the structure of the world, and it never will be, no matter how many GPUs you throw at the problem.

I also want to remind people of something I said before, which is that the brain is not a text machine. Your brain does things that have nothing to do with language all day long. You catch a ball without narrating the physics to yourself. You recognize a face without describing it in a sentence first. You feel danger before you can name it. You navigate a crowded street without writing an essay about pedestrian dynamics. These are all forms of intelligence, real and powerful forms, and they happen entirely outside the world of language. If intelligence were really just text, then mute people would be unintelligent, deaf people would be unintelligent, babies would be unintelligent, and animals would be unintelligent. But they are not. They are all deeply intelligent in ways that language cannot reach, because intelligence is broader than words, deeper than grammar, and older than writing itself. Any model that stays inside language stays inside only one narrow channel of intelligence, and that channel, while useful, is not the whole river.

The cage of language also creates a false sense of confidence that I find genuinely frightening. When a model produces a fluent, confident, well-structured answer, people tend to believe it, because humans are wired to trust confident speech. This is a survival mechanism that evolved in a world where confident speakers were usually knowledgeable leaders, but it breaks down completely when the confident speaker is a machine that cannot verify its own output. I have watched people argue with each other using LLM-generated text as evidence, without either person checking whether the text was actually true. I have watched students submit LLM-generated essays and believe they understood the material, when really they understood nothing, because the model did the thinking for them and they only did the copying. I have watched engineers trust LLM-generated code in production systems, where a single hallucinated function call can bring down a service that thousands of people depend on. This is not intelligence. This is a confidence trick performed at industrial scale, and the victims are the people who trusted the output without questioning the source.

This is also why I said induction has a ceiling. LLMs learn by looking at what already exists, they study the past, and they predict the next likely token. That means they are always downstream from human work, always recycling what has already been said, always remixing old patterns into new arrangements. They can generalize, they can surprise, and they can sometimes create things that look entirely fresh. But the freshness is still born inside a machine that feeds on prior expression, and that is not the same as genuine novelty. Genuine novelty comes from contact with the world, from experiment, from failure, from observation, from the moment when reality pushes back against your assumptions and forces you to see something you did not expect. A text model never has that moment, because it never touches reality. It only touches the shadows of reality that humans have already cast into words, and shadows, no matter how detailed, are still just shadows.

Let me also say this as plainly as I can, because I think people need to hear it without any padding. The limit of language is not a future problem that might be solved tomorrow. It is a present fact that shapes everything LLMs can and cannot do right now. Every time a model hallucinates, that is the limit showing itself. Every time a model produces a confident wrong answer, that is the limit showing itself. Every time a model fails to generalize to a new domain, that is the limit showing itself. Every time a model cannot verify its own output against the real world, that is the limit showing itself. These are not bugs that will be patched in the next version. They are symptoms of a structural constraint that cannot be removed by making the model bigger, faster, or more expensive. The constraint is language itself, and language is the medium the model lives in, and you cannot fix the medium by adding more of the same medium. You need a different medium. And that is where mathematical models come in.

I want to close this section by saying that recognizing the limits of language is not the same as dismissing language. Language is beautiful, powerful, and irreplaceable for human communication. I would not be writing this post without it, and you would not be reading it without it, and neither of us would be thinking about these ideas without the bridge that language builds between our minds. But a bridge and a destination are different things, and confusing them has led the entire AI industry down a path that is useful but ultimately narrow. The destination is not better text. The destination is better understanding of the world. And better understanding requires tools that go beyond text, into images, into sound, into motion, into structure, into equations, and into the physical fabric of reality itself. That is what the next section is about, and that is where the real story begins.

The World Is Not Made Of Words

The world is not made of words. The world is made of atoms, forces, fields, waves, particles, energy, matter, time, and space. None of these things are linguistic. None of them are made of grammar. None of them care about syntax, vocabulary, or sentence structure. An earthquake does not consult a dictionary before it shakes the ground. A virus does not read a paper before it mutates. A star does not check Wikipedia before it explodes. The world operates on physics, chemistry, biology, and mathematics, and all of these run independently of human language, because they existed for billions of years before the first human ever spoke a word. This is not a philosophical opinion. This is a plain fact that can be verified by looking at anything in the natural world and noticing that it works perfectly well without being described.

And yet, we have built an entire industry of intelligence on the assumption that language is the right starting point for understanding the world. We have poured billions of dollars into systems that learn from text, generate text, and are evaluated by text, as if the universe were a giant essay waiting to be read. This assumption is not crazy, because language does encode a huge amount of human knowledge, and that knowledge is valuable. But the assumption is also deeply incomplete, because the knowledge encoded in language is always a filtered, compressed, lossy version of the real thing. When a physics textbook describes the Navier-Stokes equations, it is pointing at a truth, but the truth itself lives in the behavior of fluids, not in the ink on the page. When a medical textbook describes how a drug interacts with a receptor, it is pointing at a mechanism, but the mechanism itself lives in the chemistry of molecules, not in the sentences that describe them. The map is useful, but the map is not the territory, and building a system that only reads maps will never be the same as building a system that can navigate the actual territory.

This is why I care about modeling the physical world directly, not through language, but through structure. The world has regularities that can be captured in mathematical form, and those forms are far more powerful than any verbal description. A single equation can encode an entire class of physical behavior in a few symbols, while a verbal description of the same behavior might take pages and still miss important details. The equation F equals ma describes the relationship between force, mass, and acceleration in a way that is precise, testable, universal, and compact. No sentence can match that level of compression, because sentences are built for communication between humans, not for exact specification of physical law. Mathematics is the language of the universe, and if we want machines that understand the universe, we need to teach them mathematics, not just English.

This is also where the idea of physics-informed machine learning becomes important. Researchers have shown that neural networks can be trained not just on data but also on physical constraints, so that the model learns to respect known laws while fitting observed behavior. This is a huge deal, because it means the model is not just memorizing patterns in a dataset. It is learning within a framework that reflects the actual structure of the world. A physics-informed model cannot easily violate conservation laws, because those laws are built into the training process itself. This makes the model more reliable, more generalizable, and more scientifically meaningful than a model that learns from data alone. The difference is like the difference between a student who memorizes answers for an exam and a student who actually understands the subject. Both might get the same score on the test, but only one of them can handle a question they have never seen before, because understanding transfers and memorization does not.

I also want to talk about neural operators, because they represent a step even further in this direction. A neural operator does not learn a single function from inputs to outputs. It learns a mapping between function spaces, which means it can take an entire field, like a temperature distribution or a pressure profile, and predict how that field evolves over time under given conditions. This is exactly the kind of computation that physics simulators do, except the neural operator learns to do it from data instead of from hand-coded equations. That matters because there are many real-world systems where the equations are too complex to solve by hand, or where the exact equations are not even known. In those cases, a neural operator can learn the dynamics directly from observation, which opens up a completely new way of doing science. You do not need to derive the equation first. You can let the machine learn it for you, and then check whether the learned equation matches reality.

This is the direction that excites me, and it is the direction that most people in the AI conversation are completely ignoring. Everyone is talking about chatbots, about code generators, about writing assistants, and about the next version of the model that can hold a longer conversation. Nobody is talking about the fact that a different kind of model, trained on a different kind of data, using a different kind of objective, could learn the actual structure of the physical world. Not a verbal description of the world. Not a summary of the world. Not a persuasive essay about the world. The actual governing relations that make the world work. That is not a chatbot. That is a revolution in how humanity relates to reality itself, and the fact that it is happening quietly, in small research labs, while the whole world screams about ChatGPT, is one of the great ironies of our time.

Let me put this as simply as I can, because I think simplicity is the best weapon against confusion. The world is not made of text, so text models will always be limited to describing the world from the outside. But the world is made of structure, and structure models can learn the world from the inside. A text model can tell you about gravity. A structure model can simulate gravity. A text model can describe a hurricane. A structure model can predict a hurricane. A text model can explain a chemical reaction. A structure model can run a chemical reaction in silico. The difference between telling and doing is the difference between a language model and a world model, and that difference is not small. It is the difference between reading about reality and actually modeling reality, and once you give a machine the ability to model reality, you give it the power to predict, to test, to explore, and eventually to intervene. That is where things get serious, and that is where the phrase "break reality" stops being a metaphor and starts being a description.

Why Mathematical Changes Everything

This is where I want to shift the conversation from what LLMs cannot do to what LMMs can do, because the transition from language-only models to large mathematical models is not a small upgrade. It is a category shift. A language model lives inside text. A mathematical model lives inside multiple streams of information at once, including text, images, video, audio, sensor data, and potentially any other type of structured input. That means a mathematical model is not limited to the symbolic layer of language. It can see a picture, hear a sound, watch a video, and process numerical data, all at the same time. That changes the entire game, because it opens the door to learning from the world directly, not just from descriptions of the world. And learning from the world directly is how humans learn, how animals learn, and how science works. It is the most natural and most powerful form of learning that exists.

Think about what it means for a model to be able to process video. Video is not text. Video is a dense, continuous stream of visual information that encodes motion, depth, lighting, texture, occlusion, interaction, and physical dynamics. When a model watches a ball fall, it does not read a sentence about gravity. It sees gravity happening. It sees the acceleration, the trajectory, the impact, and the bounce. Those visual patterns contain physical information that no text description can fully capture, because text can only approximate continuous motion with discrete words. A model that learns from video can potentially learn the dynamics of the physical world without anyone telling it the equations, because the equations are already implicit in the motion it observes. Researchers have already shown that dynamical equations can be recovered from video data, where the system first learns latent state variables and then constructs a dynamical representation directly from observed motion. That is not science fiction. That is published science, and it is happening right now.

This is also why I think the move from LLMs to LMMs is not just an engineering improvement but a philosophical shift in what we are trying to build. An LLM is trying to master language. An LMM is trying to perceive the world. Those are fundamentally different goals, and they lead to fundamentally different kinds of systems. A system that masters language can write, translate, summarize, and generate text with impressive fluency. A system that perceives the world can observe, model, predict, and interact with physical processes in ways that text alone can never support. Language mastery is a cognitive skill. World perception is the foundation of intelligence itself. Every intelligent creature on Earth perceives the world before it uses language, because perception is what gives language its meaning. Without perception, language is just empty symbols moving through a void. With perception, language becomes anchored to reality, and anchoring is exactly what current language models lack.

The practical implications of this are enormous, and I do not think most people have thought them through. A mathematical model that can process images and text together can look at a medical scan and generate a diagnosis. A mathematical model that can process video and numerical data together can watch a manufacturing process and detect defects in real time. A mathematical model that can process satellite imagery and weather data together can predict natural disasters with greater accuracy than any text-based system ever could. These are not hypothetical applications. They are applications that are already being developed and tested, and they work because the model has access to information that text alone cannot provide. The image carries spatial structure. The video carries temporal dynamics. The sensor data carries physical measurements. All of these are forms of information that text can only approximate, and the approximation is always lossy, always incomplete, and always downstream from the real data. A mathematical model removes the middleman of language and lets the machine engage with the data directly, and that directness is the source of its power.

I also want to connect this to something I wrote in Technology Has Destroyed My Livelihood, where I talked about how the tech industry extracts value from people while giving back less and less. The arrival of mathematical models does not change that dynamic, and it might make it worse, because a system that can see, hear, and read is a system that can replace more kinds of human labor. But I am not going to lie about what the technology can do just because the economic system is unfair. The technology is real, the capability is real, and the direction is real. What we need is not to deny the technology but to demand that its benefits are shared fairly, which they currently are not. The same companies that scrape our code, our images, our voices, and our data to train these models are the same companies that charge us for the output and pay us nothing for the input. That was wrong with LLMs and it is even more wrong with LMMs, because mathematical models consume more kinds of human creation and consolidate more kinds of human skill into a single system that one company controls.

Let me also address something that I think is underappreciated. The combination of multiple modalities does not just add capability. It creates capability that none of the individual modalities could produce alone. A model that can see an image and read text about that image can understand things that neither the image nor the text could communicate separately. The image provides spatial information that text cannot capture. The text provides contextual information that the image cannot express. Together, they create a richer representation than either one could create alone, and that richer representation is the foundation for deeper understanding. This is exactly how human cognition works. We do not understand the world through one channel. We understand it through the combination of vision, hearing, touch, memory, emotion, and language, all integrated into a single coherent experience. A mathematical model is the first step toward building a system that does something similar, and that step is far more important than anything that happened in the world of pure text models.

I want to be careful here, because I do not want to oversell the current state of mathematical AI. The models that exist today are still limited, still make mistakes, still hallucinate, and still cannot truly understand the world in the way a human understands it. But the direction matters more than the current state, because the direction is pointing toward something genuinely new. The direction is pointing toward machines that can perceive, model, and interact with the physical world, not just talk about it. And once that direction produces systems that are reliable enough to trust, the world will change in ways that most people are not prepared for. That is not a threat. It is a prediction based on the trajectory I see, and I would rather prepare for it honestly than pretend it is not coming.

Equations Are More Powerful Than Sentences

This section is about something that most people in the AI conversation never talk about, because most people in the AI conversation are not scientists, not engineers, and not mathematicians. They are content creators, investors, and people who use ChatGPT to write emails. There is nothing wrong with that, but it means the conversation is missing the most important piece, which is that the real power of intelligence lies in equations, not in sentences. An equation can encode an entire class of physical behavior in a single line. A sentence can only describe one instance of that behavior in one particular context. The equation is general. The sentence is specific. The equation can be tested, inverted, differentiated, integrated, and composed with other equations. The sentence can only be read. That difference is not academic. It is the difference between tools that describe the world and tools that can compute the world, and computing the world is infinitely more powerful than describing it.

Let me give you a concrete example, because concrete examples are the best way to make abstract ideas real. The equation for the trajectory of a projectile tells you exactly where the object will be at every moment in time, given its initial position, velocity, and the force of gravity. That single equation replaces an infinite number of sentences, because every possible trajectory is already contained within its structure. You do not need to describe each trajectory individually, because the equation generates all of them automatically. That is the power of mathematical compression. One equation contains more information than a library of sentences, because it captures the mechanism, not just the output. A sentence can tell you where the ball landed this one time. The equation can tell you where any ball will land under any conditions. That is the difference between reporting and understanding, and understanding is what intelligence is supposed to be about.

This is why symbolic regression is such an important area of research, and why I think it deserves far more attention than it currently gets. Symbolic regression is a method for learning compact mathematical expressions directly from data, without being told what form the equation should take. The system looks at the data, searches through the space of possible mathematical expressions, and returns the simplest expression that fits the observations. This is not curve fitting in the traditional sense, because curve fitting assumes a fixed functional form and only tunes the parameters. Symbolic regression searches over the form itself, which means it can discover entirely new equations that no human has ever written down. That is a form of machine discovery, and it is happening now, not in some distant future, but in real labs with real data and real results.

The reason this matters so much is that equations are the most efficient form of knowledge compression that humans have ever found. A single equation can replace volumes of empirical data, because it captures the underlying pattern that generates the data. If a machine can discover equations from observations, then it can compress an entire field of experimental results into a compact formal statement, and that statement can then be used to predict new observations, design new experiments, and generate new hypotheses. This is not just data analysis. This is scientific reasoning automated at its most fundamental level, and it happens entirely outside the world of language. The machine does not need to describe the law in English. It can express the law in mathematics, which is the native language of the universe, and that expression is more powerful, more precise, and more useful than any verbal description could ever be.

I also want to connect this to the broader theme of my posts, which is that the world is being reshaped by technology in ways that most people do not understand and cannot control. In my post Technology Has Destroyed My Livelihood, I talked about how the tech industry sold us a lie about equal opportunity while building systems that extract value from the bottom and send it to the top. The same thing is happening with equation discovery, but in a different domain. The institutions that have access to these tools, the large labs, the well-funded universities, the big tech companies, will use them to accelerate their scientific output while the rest of the world falls further behind. A tool that can discover equations from data is a tool that can replace teams of PhDs, and the economic consequences of that are staggering. The people at the top will move faster, learn faster, and innovate faster, while the people at the bottom will be told that their skills are no longer needed. That pattern has repeated itself throughout the history of technology, and there is no reason to believe it will stop now.

But the power is still real, and I refuse to hide from it just because the distribution is unfair. An equation is more powerful than a sentence, because an equation can be executed and a sentence can only be read. An equation can be tested against new data and a sentence can only be compared to other sentences. An equation can be composed with other equations to build larger systems and a sentence can only be placed next to other sentences to build larger texts. The computational power of mathematics is the foundation of modern civilization, and any system that can learn mathematics from observation is a system that can participate in the most fundamental form of human intellectual achievement. That is not hype. That is a sober description of what equation discovery means, and it means something enormous.

I also think people need to understand that the move from text to equations is not just a technical upgrade. It is a change in what we mean by intelligence itself. If intelligence means the ability to produce fluent text, then LLMs are intelligent. But if intelligence means the ability to discover the hidden structure of the world and express it in precise, testable, computable form, then LLMs are not intelligent at all. They are eloquent, but they are not insightful. They are fluent, but they are not deep. They can write about physics, but they cannot do physics. They can describe an equation, but they cannot discover an equation. That gap between writing about and doing is the same gap I identified in my post about language being limited, and it is the gap that mathematical models and equation discovery tools are beginning to close. Not perfectly, not yet, but the direction is clear, and the direction matters more than the current state.

Simulation Is The Real Intelligence

This is the section where everything comes together, and this is the section where I think the argument becomes truly strong. If a model can learn the equations that govern a system, and if a model can simulate those equations forward in time, then the model can predict the future. Not in a mystical sense, not in a fortune-telling sense, but in the precise, scientific sense of computing what will happen next given the current state and the rules of evolution. That is what simulation is. It is the execution of understanding. It is the moment when knowledge stops being passive description and becomes active prediction. And active prediction is the foundation of all useful intelligence, because intelligence that cannot predict is intelligence that cannot plan, cannot prepare, and cannot control.

Think about what a weather forecast is. It is a simulation. Scientists take the current state of the atmosphere, feed it into a system of equations that describe fluid dynamics, thermodynamics, and radiation transfer, and then run those equations forward in time to predict what the weather will look like tomorrow, next week, or next month. The better the model, the better the prediction. The better the prediction, the better the preparation. The better the preparation, the fewer people die in storms, floods, and droughts. Simulation is not an abstract academic exercise. It is a life-saving, civilization-enabling capability, and it depends entirely on having the right equations and the ability to solve them efficiently. A language model cannot do this, because a language model does not have equations. It has text. And text cannot be executed, only read.

World models take this idea even further, because they do not just simulate one physical system. They learn to simulate entire environments, including the effects of actions within those environments. In the research on Dreamer and related systems, world models are described as systems that learn to predict future states and rewards from sensory inputs, and then use those predictions to plan actions without needing to interact with the real world at all. That is imagination. That is rehearsal. That is the ability to test a thousand possible futures in the safety of an internal model and then choose the best one. Humans do this all the time, we rehearse conversations in our heads, we imagine what will happen if we turn left instead of right, we plan our days by mentally simulating the consequences of different schedules. World models give machines the same capability, and once that capability is reliable, the machine can operate in the world with a level of foresight that pure language models can never achieve.

This is also why I say that simulation is the real interface between intelligence and reality. Language is an interface between humans. Equations are an interface between the model and the mechanisms. But simulation is the interface between the model and the future. A model that can simulate does not need to wait for reality to happen. It can precompute reality. It can test alternatives. It can reject bad options before they cause harm. It can optimize outcomes before resources are committed. That is a fundamentally different relationship with truth than anything a language model offers, because a language model can only discuss the future in words, while a simulation model can compute the future in structure. Discussing is passive. Computing is active. And the difference between passive and active intelligence is the difference between watching the world and shaping the world.

I want to connect this back to my personal experience, because everything I write comes from experience, not from theory. In An Empty Life Filled With Constant Suffering, I wrote about how I keep fighting and still see very little change, and how suffering becomes part of your identity when it lasts long enough. What frustrated me most during those years was not the pain itself, but the inability to predict or simulate better outcomes. I kept moving forward without any model of the future that could tell me which direction was likely to lead somewhere real. I was navigating by words, by hope, by advice, and by guesswork, and none of those were precise enough to cut through the fog of uncertainty. That experience is why I believe so strongly in simulation. If you have a model of the world that can predict consequences, you do not need to guess. You can compute. And computation, unlike hope, gives you actual answers that can be tested against actual results.

Let me also say that the combination of equation discovery and simulation is the most powerful intellectual technology humans have ever developed, and this is not an exaggeration. Every bridge, every airplane, every drug, every satellite, every power plant, and every electronic device exists because someone discovered the relevant equations and then simulated them to verify the design before building it. Simulation is how we test ideas without breaking things. It is how we explore possibilities without wasting resources. It is how we learn from mistakes without paying the full price. And if a machine can learn to discover equations and simulate them autonomously, then the machine can participate in the design process at a level that was previously reserved for teams of highly trained engineers and scientists. That is not a small business opportunity. That is a transformation of the entire relationship between intelligence and physical reality, and it is the reason I chose the title of this post.

I want to close this section by saying something that I think needs to be said plainly. Simulation is not just another feature of AI. It is the whole point. Language is a nice interface for humans, but simulation is the real work of intelligence. A mind that can only talk is a commentator. A mind that can simulate is a scientist. A commentator can describe the game, but only a scientist can change the rules. And when the rules start changing, reality itself starts bending, and that is what I mean when I say LMMs will break reality. Not destroy it. Not ruin it. But break our current understanding of what is possible, in the same way that every major scientific revolution broke the previous understanding. That is the promise, and it is a promise built on math, on physics, on structure, and on the hard-won knowledge of centuries of human science.

The Danger Is Real And I Will Not Ignore It

I have spent most of this post talking about the power and potential of mathematical models, equation discovery, and simulation, and I believe every word I have written. But I would be dishonest if I did not also talk about the danger, because the danger is real, and pretending it does not exist is exactly the kind of comfortable lie that I have been fighting against in every post I have written. Technology is not neutral. It has never been neutral. It carries the intentions, the biases, and the economic interests of the people who build it and the people who fund it. I wrote about this in It is always the Russians, where I described how the Soviets killed God and the people who came from that tradition built the most powerful AI systems in the world. That pattern has not changed. The same people who built the current language models will build the next mathematical models, and they will carry the same assumptions, the same blindness, and the same hunger for control into the new era.

The first danger is misuse. A system that can simulate the physical world can also simulate weapons, surveillance systems, environmental manipulation, and biological agents. This is not speculation. It is a direct consequence of the capability I have been describing. If a machine can discover equations from observation, then it can discover the equations of harmful systems just as easily as the equations of helpful ones. If a machine can simulate the future, then it can simulate the future of a weapon just as easily as the future of a bridge. The same tool that enables scientific breakthrough also enables scientific catastrophe, and the difference between the two depends entirely on who controls the tool and what they intend to do with it. Given the current state of the world, where power is concentrated in the hands of a few companies and governments, I am not optimistic about who will control these tools. The same companies that exploited LLMs for profit will exploit LMMs for profit, and the same governments that used surveillance for control will use simulation for control, and the people at the bottom will pay the price, as they always have.

The second danger is displacement, and this one is personal because I have lived through it. In Technology Has Destroyed My Livelihood, I described how the tech industry destroyed my career while I was doing everything right. I coded, I built, I competed, I published, and I applied, and the system still threw me away. That was with text-based tools. Imagine what happens when mathematical models can do everything text-based models can do plus see, hear, and simulate. The number of human tasks that can be automated does not just increase. It explodes. A model that can read your code, see your screen, hear your voice, and simulate the outcomes of different engineering decisions can replace not just coders but entire engineering teams. That is not a future scenario. That is the stated goal of the companies building these systems, and they are not shy about it. They want to replace human labor with machine labor, because machine labor is cheaper, faster, and does not complain.

The third danger is epistemic, meaning it affects how we know what we know. If a model can simulate reality convincingly, then it can also fabricate reality convincingly. A text-based hallucination is already dangerous, but a mathematical hallucination, one that includes fake images, fake video, fake audio, and fake data, is orders of magnitude more dangerous, because it attacks every channel of human perception at once. Right now, you can fact-check a text claim by looking at an image or watching a video. But if the image and the video are also generated by the model, then your fact-checking loop is broken, and you have no independent source of ground truth. That is an epistemic catastrophe, a situation where truth becomes indistinguishable from fabrication across all modalities, and it is not a distant hypothetical. It is a near-term consequence of the technology I have been describing, and very few people are taking it seriously enough.

The fourth danger is concentration of power, and this one connects to everything I have written about in all my previous posts. A company or government that controls a system capable of equation discovery and reality simulation has a fundamental advantage over everyone else, because they can predict the future while everyone else is still guessing. They can optimize their strategies while everyone else is still debating. They can design physical systems while everyone else is still sketching on paper. That kind of advantage does not lead to shared prosperity. It leads to domination, because the people with the tool have no incentive to share it, and every incentive to protect it. History has shown this pattern over and over again. The people who control the most powerful technology always use it to extend their control, not to liberate others. There is no reason to believe this time will be different, and many reasons to believe it will be worse, because the technology is more powerful than anything that has come before, and the concentration of ownership is tighter than anything we have seen.

I also want to say that the danger is not just to individuals or to jobs. The danger is to the entire structure of human knowledge and human society. If machines can discover equations from data, then the human scientists who currently do that work become less valuable. If machines can simulate futures, then the human planners and strategists who currently do that work become less valuable. If machines can perceive, model, and predict the physical world, then the entire human enterprise of trying to understand reality becomes something that machines do better, faster, and more cheaply. That sounds abstract, but the consequences are concrete. Fewer jobs for scientists. Fewer jobs for engineers. Fewer jobs for anyone whose work involves understanding the physical world. And the people who lose those jobs will not have anywhere to go, because the machine does not just automate one task. It automates the entire cognitive chain from observation to understanding to prediction to action. That is a displacement so comprehensive that no retraining program can fix it.

Let me be clear about one more thing. I am not saying we should stop building these systems. I do not believe that is possible, and I do not believe it is desirable. The potential for good is too great. A system that can simulate the physical world can help us cure diseases, design better infrastructure, predict natural disasters, discover new materials, and solve problems that have resisted human effort for centuries. But the potential for harm is equally great, and ignoring that potential because we are excited about the benefits is exactly the kind of blindness that leads to catastrophe. The right response is not to stop building but to demand transparency, to demand fair distribution of benefits, to demand regulation, and to demand that the people who are most affected by these systems have a voice in how they are deployed. That is not happening right now, and it needs to happen before the technology outpaces our ability to control it.

I want to end this section with a direct message to other engineers, especially those who have been through what I have been through. Do not let them pretend this is only about progress and opportunity. It is also about power and control, and the people who build these systems need to be honest about both sides. I have been honest about the limits of language models, I have been honest about how the industry exploits us, and I will be honest about mathematical models too. They are powerful, they are important, and they are dangerous. All three of those things are true at the same time, and accepting all three is the only honest position.

What Comes After Language

This is the final section, and it is the one where I want to zoom out and look at the big picture, because everything I have written so far leads to a single conclusion that I think is both obvious and terrifying. What comes after language is structure. What comes after description is computation. What comes after talking about the world is modeling the world. And once machines can model the world directly, through equations, through simulation, through mathematical perception, the entire relationship between humanity and reality changes. That is what I mean by breaking reality, and I do not use that phrase lightly.

Let me explain what I mean in concrete terms. For most of human history, our access to reality was limited by our senses and our tools. We could see what was in front of us. We could measure what we could reach. We could predict only what our theories could calculate, and our theories were limited by our mathematical ability and our physical intuition. Every significant advance in science extended our reach, from telescopes to microscopes to particle accelerators to satellite networks, each tool gave us access to a new layer of reality that was previously hidden. But all of those tools required human interpretation. A human had to look through the telescope. A human had to read the instrument. A human had to analyze the data and discover the pattern. The bottleneck was always human cognition, and human cognition is slow, biased, limited by attention, and constrained by the number of hours in a day.

What changes with mathematical AI and automated equation discovery is that the bottleneck shifts. The machine can observe, model, discover, and simulate without a human in the loop at every step. It can process more data than any human can read. It can search over more equations than any human can write. It can simulate more futures than any human can imagine. That does not make it smarter than a human in every way, and I am not claiming that. But it does make it faster at certain forms of scientific discovery, and speed in discovery translates directly into acceleration of knowledge. When knowledge accelerates, the world changes. When knowledge accelerates fast enough, the world changes faster than people can adapt, and that is exactly the situation we are heading into.

This is also why I think the conversation about AI needs to move beyond language models entirely. The real action is happening in the intersection of machine learning and physical science, where models are learning to perceive the world mathematically, discover its governing equations symbolically, and simulate its dynamics numerically. That intersection is small right now, but it is growing, and the people working in it are some of the most talented scientists and engineers alive. They are not building chatbots. They are building tools that can participate in scientific reasoning at a fundamental level. They are building tools that can learn the laws of physics from video, discover compact mathematical expressions from data, and simulate the behavior of complex systems without being told the equations in advance. That is a different kind of AI than what the public sees, and it is far more important.

I want to connect this to something I wrote in Christianity Makes Perfect Sense!, where I talked about how some truths are not simple but that does not make them false, and how deeper truths often require deeper thinking. The truth about where AI is heading is exactly like that. The surface conversation is about chatbots and productivity and who lost their job to automation. The deeper conversation is about whether machines can learn the structure of reality itself, and what happens to human civilization when they can. That deeper conversation is the one that matters, and it is the one that almost nobody is having, because it requires understanding physics, mathematics, machine learning, and philosophy all at once, and most people specialize in one and ignore the rest.

Here is my honest assessment of where we are. LLMs are useful tools that operate entirely within the cage of language. They are powerful, impressive, and genuinely helpful for many tasks, but they will never understand the world because they never touch the world. LMMs are a bridge from language to perception, and that bridge opens the door to learning from the physical world directly. Equation discovery and symbolic regression allow machines to extract compact mathematical structure from data, which is the most fundamental form of understanding. Neural operators and physics-informed models allow machines to learn and simulate the dynamics of physical systems, which is the most fundamental form of prediction. And world models allow machines to simulate entire environments, including the effects of actions, which is the most fundamental form of planning. Put all of these together, and you have a system that can perceive, understand, predict, and plan, and that system is no longer a chatbot. It is a scientific partner. It is a simulation engine. It is a tool for breaking through the limits of human cognition and accessing layer of reality that were previously out of reach.

That is why I chose the title of this post. LLMs are useful, and I will not pretend otherwise. But LMMs, combined with equation discovery and simulation, will break reality in the sense that they will shatter our current understanding of what machines can know, what machines can do, and how fast human knowledge can grow. That is not a threat or a promise. It is a trajectory, and trajectories are harder to deny than predictions, because they are based on what is already happening, not on what might happen. The research exists. The tools exist. The direction is clear. The only question is how fast we get there, and who controls the outcome when we do.

I want to end this post the way I have ended every post since I started writing. I am not trying to be dramatic. I am not trying to scare anyone. I am trying to tell the truth as honestly as I can, from the perspective of someone who has built software, studied systems, lived through technological displacement, and spent years thinking about what intelligence actually is. The truth is that language models are useful but limited. The truth is that mathematical models are a step toward something much deeper. The truth is that equations are more powerful than sentences. The truth is that simulation is the real intelligence. And the truth is that all of this is coming faster than most people think, and the people who understand it first will have the most power, for better or for worse.

Till next time 👋!

References