The Equation of Meaning

Every time you chat with an AI, you’re talking to one of the most expensive machines ever built, and neither you nor its creators can fully explain how it does what it does. Artificial intelligence keeps getting more capable, but only by becoming dramatically more expensive. Every major leap in performance has required more data, more parameters, more specialized hardware, and more energy. The pattern is familiar: larger models, trained on more text, for longer periods, at higher cost.

The approach has worked remarkably well. It has also created a quiet problem. Efficiency keeps improving, and yesterday’s capability gets cheaper every year. But the frontier is different. Each new increment of capability at the leading edge costs more than the last, and nothing about the paradigm suggests that changes on its own.

There is another path.

If we understood more about how meaning itself works, we might represent it directly instead of approximating it through ever-greater scale. That possibility changes the economics of intelligence.

Inside the Black Box

Humans use meaning every day without knowing how we actually do it. We understand sarcasm, implication, and context instantly, but we can’t explain the inner mechanics.

AI faces the same problem from the other side. It turns every word into a long list of numbers and learns to predict the next word by studying billions of sentences. Through massive amounts of training data, it becomes extremely good at predicting what comes next. Whether that counts as understanding is a debate I’ll leave alone. The simpler point is that nobody can open the model and point to where meaning lives or how it works.

Both humans and today’s AI are using meaning without truly understanding how it works.

There is one telling difference, though. When researchers pry AI models open, they keep finding hints of geometry inside: concepts that behave like directions, meanings that add and subtract. The structure of meaning seems to be in there. The models just had to rediscover it from scratch, at enormous cost, because nobody has written it down.

Words are meaning at rest. Sentences are meaning in motion. Meaning is composed, not merely retrieved: words mark coordinates in semantic space, and sentences trace paths through it, with each word reshaping the context of the others.

Meaning Does Not Need to Become Transparent

The idea that meaning might eventually be expressed mathematically is easy to misunderstand.

It does not require that every thought, feeling, or interpretation become fully explainable. Meaning may never become completely transparent. Consciousness, personal history, cultural context, and subjective experience will likely continue to introduce elements that resist full formalization.

But meaning does not need to become transparent to become useful. It only needs to become translucent: understood well enough that we can identify the deeper, more stable patterns that form the foundation of meaning. While language itself is often messy and ambiguous at the surface, these foundational patterns remain consistent across contexts.

We already accept this kind of partial formalization in other domains. We do not need a complete philosophical account of life to build useful models of genetics. We do not need to fully understand consciousness to predict how certain chemicals affect the brain. We do not need perfect weather models to produce forecasts that are accurate enough to be valuable.

A mathematical framework would not need to capture every nuance. It would only need to capture enough structure that some of the current reliance on brute-force approximation becomes unnecessary.

The Brute-Force Tax

Current AI systems depend heavily on scale. They require enormous datasets and large numbers of parameters in part because they lack an explicit framework for how semantic relationships work.

Without that framework, the model must infer related patterns across vast numbers of examples. It encounters variations of the same underlying idea in different contexts and gradually builds an internal representation through repetition and statistical association.

This is expensive.

Much of what the model learns could, in principle, be derived from a smaller set of foundational relationships if those relationships were already known. Instead, the system must rediscover them indirectly through volume.

A mathematical approach to meaning would not eliminate the need for learned representations. It would organize and constrain them. The model would still need to learn about the world, but it would no longer need to re-learn basic patterns of how meaning transforms across contexts at the same scale.

Some of what is currently stored as parameters could be replaced by more compact rules and transformations.

This is the brute-force tax: the extra scale required because we lack a more direct way to represent what meaning is and how it changes.

The Obvious Objection

Anyone who has followed AI for a while will recognize this idea, and will also remember what happened to it. For decades, researchers tried to build knowledge into machines by hand: expert systems, rule collections, giant hand-crafted ontologies. Scale beat all of them. The lesson was so consistent it earned a name, the Bitter Lesson: general methods plus more compute win, and built-in human knowledge loses.

But look closely at what actually lost. What died was hand-coded content, humans typing facts and rules into machines. What kept winning was structure derived from math. Convolution is the famous example: someone wrote down the simple fact that an object is still the same object when it shifts across an image, and computer vision became dramatically cheaper almost overnight. Attention, the mechanism inside today’s models, is another. In each case, a compact statement about the structure of a domain replaced an enormous amount of learning the machine no longer had to do.

A mathematical framework for meaning belongs to the second category, not the first. It is not a rulebook of facts about the world. It is a statement about how meaning itself is shaped, the same kind of move that made vision cheap.

An equation of meaning would not replace neural networks. It would provide the inductive bias that could make them far more efficient at what they do best: navigating geometry instead of memorizing coordinates.

Compressing Meaning, Not Just Data

Traditional compression removes redundancy from data. A compressed image does not store every pixel independently. It identifies patterns and stores the information needed to reconstruct them. The same principle applies to video, audio, and text.

A mathematical understanding of meaning would enable a deeper form of compression: semantic compression. Instead of compressing the surface form of language, it would compress the underlying structure of what the language is trying to convey.

Two passages that use entirely different wording can express nearly the same idea. Today we store both as separate sequences of tokens. A semantic system could recognize that the underlying structures are closely related and represent them more efficiently. A long explanation could be reduced to a compact set of relationships, assumptions, confidence levels, and contextual dependencies from which an AI could later reconstruct appropriate versions for different audiences.

This is the mechanism that could make future AI systems dramatically smaller. The reduction would not come from making models dumber. It would come from removing redundancy that currently exists because we lack a better representation of meaning itself.

What AI Could Become After Scale

If meaning becomes partially formalizable, the architecture of AI could shift.

Instead of relying primarily on ever-larger monolithic models, future systems could combine a smaller universal semantic framework with lighter specialized components. The universal part would handle the structural relationships that current models must repeatedly approximate. The specialized parts would handle domain knowledge, cultural variation, and real-time context.

The result would not be a purely symbolic system replacing neural networks. It would more likely be a hybrid in which learned representations are organized and compressed by a deeper mathematical theory of meaning. Intelligence would become more compressible than today’s scaling paradigm assumes.

This does not mean AI would suddenly become trivial or that all problems would be solved. It means the marginal cost of additional capability could fall significantly once we stop paying the full brute-force tax on meaning.

Beyond Knowledge

Compressing knowledge is one thing.

But once meaning itself becomes mathematically tractable, we may be able to compress something far more complicated than facts or language: the recurring patterns of an individual human mind.

That possibility changes what intelligence can be.

Pieces of this puzzle are already being explored, some inside major labs and some by independent researchers working in the open. The timeline is uncertain, but the direction is not. Sooner or later, someone is going to make meaning mathematically usable.

The only real question is who gets there first, and what we choose to build with it.