{"slug": "the-missing-equation-of-ai", "title": "The Missing Equation of AI", "summary": "Large language models and reinforcement learning agents represent two fundamentally different approaches to intelligence that mirror the centuries-old philosophical debate between rationalism and empiricism, according to an analysis of modern AI paradigms. Reinforcement learning pioneer Rich Sutton's \"Bitter Lesson\" argues that AI advances come from methods that scale with experience and computation, embodied in the Bellman equation which defines intelligence as learning through trial and error without pre-existing human knowledge. The essay suggests that a Kantian synthesis of both approaches may be the key to achieving artificial general intelligence.", "body_md": "Ask an LLM a question and it answers instantly, drawing on everything it has ever read. Ask a reinforcement learning agent the same question and it has no idea — until it has failed a thousand times and learned the hard way. Both are called intelligence. They could not be more different and yet, this divide is not new; it is a very old argument wearing new clothes!\n\nFor centuries, philosophers have debated a single question: How do humans acquire knowledge? On one side, the Rationalists (like Descartes) believed that knowledge comes from reason and innate ideas. On the other side, the Empiricists (like Locke and Hume) believed the mind is a tabula rasa, a blank slate that gets filled only through experience and the senses. Today, this debate is no longer just in books; it is happening inside the core of modern artificial intelligence.\n\nModern AI is divided into two main paradigms that mirror this ancient split. On one hand, Large Language Models (LLMs) act like digital Rationalists, holding vast structural representations of human knowledge. On the other hand, Reinforcement Learning (RL) represents pure Empiricism, learning through trial, error, and raw interaction.\n\nThis essay aims to explore how these technologies resurrect the classic philosophical debate, and why a Kantian synthesis of the two might be the ultimate key to AGI.\n\nOn the side of pure Empiricism stands **Rich Sutton**, one of the founding figures of Reinforcement Learning. Sutton is not merely one of the greatest technical researchers of our age; he represents a strong philosophical conviction about what intelligence truly is. Throughout his career, he has consistently argued that intelligence cannot be pre-written, engineered through rules, or distilled from human knowledge. It must be earned through direct interaction with the world.\n\nReinforcement Learning, as Sutton conceives it, begins in ignorance. The agent is not given concepts, explanations, or prior understanding. It acts blindly, makes mistakes, and slowly adjusts its behavior based on the consequences of its actions. This process is inefficient, slow, and often wasteful, but for Sutton, this is not a weakness. It is the price of reality. Just as animals learn by surviving, failing, and adapting, an intelligent agent must collide with the world in order to extract genuine knowledge from it.\n\nThis view is captured in Sutton’s famous “Bitter Lesson,” which argues that the greatest advances in AI have not come from embedding human insight or clever abstractions into machines, but from methods that scale with experience and computation. In philosophical terms, this is a radical form of empiricism: intelligence does not emerge from understanding the world in advance, but from repeatedly acting within it and being shaped by its constraints. Sutton’s worldview has a precise mathematical heart, captured in what is arguably the most famous equation in Reinforcement Learning: the Bellman equation:\n\nHere, V(s) is the value of being in state s , R(s,α) the immediate reward of an action, γ (gamma) how much the agent cares about the future versus the present, and P(s’|s,a) the probability of ending up in a new state after acting. At first glance, this looks intimidating, but its philosophical content is breathtakingly simple. It says: the value of being in any state of the world equals the immediate reward you can grab from your best action, plus the discounted value of all the states that action might lead you to next.\n\nNotice what’s NOT in this equation. There is no human knowledge. No instructions. No concepts inherited from teachers or textbooks. There are only states, actions, rewards, and time. The agent doesn’t need to be told what “fire” or “cliff” means but only to experience the painful reward of touching one. From a few simple ingredients (what is here now, what action to take, what will follow, and how much I care about the future) emerges a recursive recipe for behavior. Intelligence as bookkeeping over experience.\n\nThis is empiricism in its most uncompromising form: a creature that begins knowing nothing, and yet, by repeatedly bumping against the world and updating these values, eventually learns what to value and what to avoid. There is no shortcut. There is no syllabus. There is only the world, the action, and the consequence.\n\nFrom this perspective, Large Language Models appear deeply limited. For Sutton, LLMs do not **understand**; they merely **mimic** understanding. They can reproduce the patterns of human thought, but they lack its most essential feature: grounding in interaction with reality . Without friction, risk, or consequence, a system may achieve internal coherence, but not truth. It operates at the level of statistical regularities rather than grounded knowledge. What emerges is not an agent struggling within the world, but a disembodied intelligence.Of course fluent and impressive, but at the same time ultimately detached from the very reality that gives intelligence its meaning.\n\nOn the opposite side stands **Andrej Karpathy, **a founding member of OpenAI and former director of AI at Tesla whose vision of intelligence reflects a deeply rationalist tradition. While Sutton sees intelligence as something forged through struggle and interaction, Karpathy is fascinated by the emergence of abstract reasoning from language itself. In his descriptions of Large Language Models, he often portrays them not as organisms or agents, but as disembodied entities of thought — what he metaphorically calls **“Ghosts.”**\n\nThis metaphor evokes a strikingly Platonic image of intelligence. These models possess no body, no senses, and no instinct for survival. They do not touch the world directly, yet they absorb vast amounts of human language and reconstruct the structure of reality through symbols alone. In this sense, the LLM resembles a mind suspended above physical existence, operating entirely within a universe of abstraction. Like Plato’s philosopher contemplating the world of Forms, the model does not know “fire” by feeling heat or danger, but by analyzing millions of relationships surrounding the concept of fire across human discourse. Inside its latent space, concepts become structurally organized in a remarkably coherent way, detached from immediate sensory experience. The machine does not encounter the world itself, but a symbolic reflection of it.\n\nFor Karpathy, this is not necessarily a weakness. On the contrary, the power of LLMs comes precisely from their mastery of representation and language. Freed from the limitations of embodiment, these systems can manipulate ideas with extraordinary flexibility, moving through the world of symbols with a fluency that often resembles reasoning itself. They operate in a space of abstraction, logic, and internal coherence. Just as Sutton’s vision has its mathematical signature, so does Karpathy’s. The mechanical heart of every modern Large Language Model is a strikingly different equation — the attention mechanism:\n\nAgain, the symbols matter less than what they describe. At each step of “thinking,” the model looks at every word it has seen so far and asks, for each word it might produce next: how strongly should I attend to each of the words that came before? It computes a weighted sum of these relationships, purely between symbols, purely within language, and from that sum, it predicts what comes next\n\nLook at what is in this equation, and especially what isn’t. There is no world, no reward, no cliff to fall off, no fire to fear. There are only the Q, K, V queries, keys, and values, all of them vectors in an abstract space of meaning. The “Ghost” Karpathy describes is, mathematically, a probability distribution over symbols, conditioned on a history of other symbols. It contemplates language by relating language to language. Plato would have recognized this immediately: a mind moving entirely among Forms.\n\nWhere Bellman’s recursion needs a world to grind against, attention needs only a corpus of text. One learns by colliding with reality. The other learns by absorbing the residue of other minds.\n\nYet this vision also carries a philosophical cost. The “Ghost” may speak intelligently about the world, but it remains separated from the friction, risk, and immediacy of lived experience. Its knowledge is vast, but fundamentally indirect; a map constructed entirely from traces left behind by human minds.\n\nHowever, the path to Artificial General Intelligence (AGI) seems to stumble upon the inherent limits of both schools.\n\nOn one hand, the pure Empiricism of RL suffers from what Noam Chomsky identified as the* *trap of Behaviorism. An agent starting as a tabula rasa is desperately inefficient. Without innate knowledge or logical structure, it is condemned to a Sisyphean task: it must fall off a cliff a thousand times just to learn about gravity. It lacks common sense. If AGI relies solely on this, we will end up with smart animals that learn slowly and cannot generalize their knowledge.\n\nOn the other hand, the Rationalism of LLMs crashes into John Searle’s famous “Chinese Room.” Searle imagined a person locked inside a room, following an enormous book of rules for manipulating Chinese symbols. To an outside observer, the person appears perfectly fluent in Chinese, producing coherent answers to every question. Yet inside the room, there is no understanding at all, only the mechanical manipulation of symbols. According to Searle, this is precisely the limitation of systems based purely on syntax. The model can combine words with astonishing accuracy, but it does not understand the semantics. It knows the word “apple,” but lacks any direct experience, intention, or first-person perspective connected to an apple. If AGI relies solely on this approach, we may end up with “wise ghosts”: systems that speak fluently about reality while remaining fundamentally detached from it.\n\nIf a modern Kant looked at AI today, he would see the same old debate in a new form.\n\nIn Reinforcement Learning, he would see what he called “Empirical Intuition.” Sutton’s agents feel the world and react to it, but they are often “blind.” They lack the higher concepts needed to understand what they are experiencing.\n\nIn Large Language Models, he would see the opposite problem, what he might call “Pure Concepts.” Karpathy’s models reason with amazing fluency, full of rich structure, but they are “empty.” Their words are never connected to the real experience those words describe.\n\nThis brings us to Kant’s most famous line, and it now reads almost like a description of modern AI:\n\n“Thoughts without content are empty, intuitions without concepts are blind.”\n\nHere is the deepest irony of all. Each side of this debate has its own mathematical signature. Sutton’s empiricism speaks the language of Bellman, an equation built from states, actions, and rewards. Karpathy’s rationalism speaks the language of attention, an equation built from symbols relating to other symbols. Both are precise. Both can be computed. Both have produced remarkable systems.\n\nBut the synthesis that Kant pointed toward, the moment we might describe as the LLM finally gaining a body, has no equation yet. We have a mathematics of pure experience, and we have a mathematics of pure abstraction, but we still lack one system where concepts gain real meaning by touching the world they describe. The union of mind and matter that philosophers searched for across centuries is, for AI, still an unwritten line of code.\n\nMaybe the path to AGI is not simply a bigger model, or a smarter algorithm. Maybe it is the search for an equation that does not exist yet, one that finally joins Bellman and attention into something larger than either one alone.\n\n[The Missing Equation of AI](https://pub.towardsai.net/the-missing-equation-of-ai-ea969df6cc3d) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/the-missing-equation-of-ai", "canonical_source": "https://pub.towardsai.net/the-missing-equation-of-ai-ea969df6cc3d?source=rss----98111c9905da---4", "published_at": "2026-07-21 21:01:02+00:00", "updated_at": "2026-07-21 21:28:39.945241+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research"], "entities": ["Rich Sutton", "Bellman equation", "Descartes", "Locke", "Hume", "Kant"], "alternates": {"html": "https://wpnews.pro/news/the-missing-equation-of-ai", "markdown": "https://wpnews.pro/news/the-missing-equation-of-ai.md", "text": "https://wpnews.pro/news/the-missing-equation-of-ai.txt", "jsonld": "https://wpnews.pro/news/the-missing-equation-of-ai.jsonld"}}