Some technologies improve processes. Some open new markets. Some shift the axis of history.
LLMs belong to the third category.
Many things have been called revolutionary too early. Apps were called a revolution. Blockchain was called a revolution. The cloud was called a revolution. At some level, each of those changed the world. But none of them touched the central nerve of civilization as directly as language models do.
Human civilization was not built only with force, tools, or territory. It was built with language.
Language allowed us to name reality, transmit memory, preserve instructions, invent categories, establish laws, compose theologies, record science, draft contracts, imagine futures, and coordinate crowds around abstractions no one can hold in their hands: nation, money, law, sin, dignity, algorithm, company, property, truth.
For millennia that was an exclusively human monopoly. It no longer is.
What is happening now
Anyone who still treats AI as a lab curiosity may be looking in the wrong place. What we are seeing now is not an isolated experiment, but a new layer of infrastructure being installed on top of the economy and culture.
In 2024, most large organizations reported using AI in at least one business function, and the use of generative AI more than doubled year over year. Private investment followed the same pattern. On the consumer side, AI tools have already moved beyond the niche stage and are now used monthly by more than a billion people.
That would already matter on its own. But the decisive point is not just adoption. It is that we are beginning to delegate to machines something that used to be inseparable from human presence: competent manipulation of language, context, apparent intent, and symbolic structure.
Language: the power that made humans scalable
Human beings are not the strongest animals. We are not the fastest or the toughest. What differentiated us was our ability to operate in symbolic layers.
A wolf can react. A human being can represent.
That sounds abstract until you realize that every durable human advantage begins there. Language allowed one generation to avoid starting from zero. Knowledge could be stored outside the body. First in ritual speech, then in writing, then in libraries, then in databases, then on the web.
Language also did something deeper: it made it possible to architect knowledge.
Not just to remember facts, but to classify them.
Not just to transmit orders, but to organize concepts.
Not just to describe the world, but to build systems of interpretation around it.
From there came taxonomies, maps, contracts, liturgies, jurisprudence, formal mathematics, scientific models, and software. Civilization is, to a great extent, language compressed into institutions.
Symbols: compression of the world
Every sophisticated language performs a useful violence against reality: it reduces infinite complexity into symbols that can be manipulated.
A word is already a compression. “Tree” is not a tree. It is a symbol that lets the mind hold an entire class of objects. The same is true for numbers, icons, formulas, variables, flags, logos, laws, verses, and interfaces.
Without that compression, thinking at scale would be impossible.
Writing did this.
Mathematics did this.
Programming did this.
The web did this.
LLMs operate precisely at that level: the level where the world has already been reduced to symbols rich enough to be modeled, recombined, and projected forward.
That is why saying an LLM “just predicts the next token” is both correct and radically insufficient.
What it means to predict the next token
At the technical core, language models are systems trained to predict the next piece of a sequence given previous context. That description is valid. It is also the description that misleads the impatient.
The “next token” is not just a letter or syllable. At scale it carries regularities of grammar, style, semantics, argumentative structure, code patterns, social conventions, and correlations between concepts.
When a model learns to predict the next symbol well across billions or trillions of contexts, it learns far more than surface sequence. It learns, statistically, how meaning usually organizes itself.
It does not understand the way a human being understands, in the full embodied, moral, and existential sense of the word. But it does learn enough about the outer structure of language and its uses to produce effects that, in practice, cross into the territory we associate with understanding.
Prediction in complex systems is not a small trick. Predicting the next symbol inside a civilization that externalized nearly everything into symbols is already touching something close to general capability.
The web as the great externalization of the human mind
Before LLMs, one condition had to be met: humanity had to pour itself into the network.
The web was not merely a network of pages. It became the largest act of symbolic externalization in history. We placed there texts, forums, documentation, encyclopedias, newspapers, code, catalogs, arguments, litigation, tutorials, poetry, propaganda, improvised ontologies, and formal ontologies.
The real history of the internet followed two forces in parallel.
The elegant force: explicit semantics, linked data, carefully structured knowledge.
The brutal force: massive scale, textual chaos, indexing, scraping, redundancy, correlation, and computational brute force.
Instead of waiting for the whole world to label its knowledge perfectly, we trained machines on the raw trail of what we had already written.
We did not organize knowledge enough for machines, so we gave machines enough statistical power to infer part of that organization on their own.
The symbolic context is now shared
For a long time, socially valid words depended on human mouths, human signatures, and human institutions. Only people produced the symbols society treated as operationally relevant: diagnoses, opinions, code, lessons, articles, contracts, summaries, translations, explanations, proposals, theses.
Not anymore.
Language models now operate inside the same symbolic space where we work. They draft plausible emails, produce executable code, summarize legal material, draft essays, explain papers, restructure APIs, translate tone, organize knowledge, and alter, in practice, the flow through which operational truth circulates.
This does not mean they replaced human consciousness. It does not make them moral subjects. It does not give them wisdom, responsibility, or soul.
But it does mean that language is no longer exclusively human in its immediate execution.
And that changes history.
Where the future enters
There is a correct intuition in saying that predicting the next symbol is, in some sense, predicting the future.
All language is temporal. Each word narrows the likely space of the next. Each sentence constrains the possible futures of discourse. Every symbolic choice opens one path and closes others.
An LLM is trained exactly on that principle: given a past, estimate the most probable future of the sequence.
There is something even more interesting than that. Human text carries plans, hypotheses, diagnoses, fears, habits, norms, and projects. Predicting language is therefore also, partially, predicting human action as encoded in language.
That is why these models are so useful in programming, planning, support, technical writing, document analysis, and work interfaces. In all of those domains, the immediate future is already partially inscribed in prior symbolic structures.
Whoever reads context better predicts the next move better. Whoever predicts the next move better appears intelligent to us.
The mistake of calling it “just autocomplete”
Spreadsheets changed companies. Search engines changed the web. Smartphones changed everyday life.
LLMs touch a layer prior to all of those: the layer of general cognitive mediation.
Calling them “glorified autocomplete” is like calling a compiler “copy and paste with rules.” It describes a local mechanism while hiding the systemic phenomenon.
Yes, there is autocomplete in there. But there is also statistical compression of culture, language, form, and pattern. There is a universal interface. There is translation across domains. There is a narrowing of the gap between intent and symbolic execution.
The point is not merely that these systems generate text. The point is that they operate on the representational fabric that already organizes the world.
The real milestone
Some milestones expand the body: the wheel, the engine, antibiotics, the airplane.
Others expand the collective mind: writing, printing, the library, the web.
LLMs belong to the second lineage.
They are not powerful only because they automate tasks. They are powerful because they enter the circuit through which meaning is produced, manipulated, transmitted, and converted into action.
We built tools that can now move, with unprecedented effectiveness, inside the same symbolic universe that made humans scalable.
That is the milestone.
It is not only that machines now calculate. It is that they now participate in language.
And when a new class of agents enters language, it enters history as well.