The Future of Digital Communication

Language was the truly great invention of our species, surely not the wheel. Along with fire, it is what all human populations have in common: it is what made us able to process more and more information, build complex societies like the first empires.

It is thanks to language that the age of sapiens, the Anthropocene, began. But the same force that united us, in time also divided us, with the creation of different languages: the curse of incommunicability imposed by a god who wanted, probably, to curb the anthropogenic explosion. Divide and conquer, and with the tower of Babel, language became a wall of incomprehension dividing peoples at war with each other. Daniel Everett, in Don't sleep, there are snakes, describes a night when drunken Piraha are about to lynch him. He is saved by talking to them, and then reflects on the importance of language to be recognized as a person, worthy of living. To those who have traveled, and lived in countries that speak other languages, this is a familiar feeling.

But, before communicans, precisely, we are sapiens, and therefore invented writing, which in a sense goes beyond spoken language. As Thomas Mann writes in The Buddenbrooks, writing possesses a weight and effect beyond what the author intended to give it. Written language almost lives a life of its own. An effect amplified, centuries later, by the invention of printing: the reproduction of ideas at a breakneck pace, in any language. Galileo writes in the vernacular, and on his shoulders Newton, shortly after, reading his works in English, builds a masterpiece of interpretation of reality.

And we come to the World Wide Web of the last century: a global digital archive where any text can be read, translated, compared with another, in any language. But this dream of accessibility has spawned a computer monster: today we produce and exchange, through digital tools, more information than a single human being can handle. And, clearly, it is machines that manage, mechanically, this communication.

In what sense mechanically? In the Web, digital machines are only concerned with low-level syntax checking, i.e., they make sure the message is delivered to the recipient as it was started by the sender: like a spell checker checking that "a donkey flies low for the cat's second" is syntactically right, but doesn't realize that it doesn't mean anything. Routers move symbols, lines of code, protocols, but they do not understand: they do not absorb the meaning, the semantic part.

Of course, this is changing today. With deep neural networks, and especially with large language models, or LLMs (GPTs, just to simplify, but not only that), something has changed: systems build internal representations of the message, of what they see as reality.

They are not representations that we can understand directly: they are vectors, that is, lists of numbers indicating coordinates in spaces with thousands of dimensions, inaccessible to our understanding. Our culture, the culture of sapiens, the information with which we have trained these algorithms, is represented by them as a surface in this multidimensional space, on which we, communicating, move, but from which they may one day move away, to create their own representations of reality, totally unimaginable to us.

If language has been our trump card, vectors will be theirs. Currently, two LLMs communicate with each other still using our natural language, plus data in various "structured" formats (JSON, XML, etc.).

It is our way of communicating, which, to those who have seen how an LLM creates its own representation of information, appears ... a bit clumsy and inefficient. How difficult is it for us sapiens to communicate what we really have in our heads, feel? Well, the art of communication, for LLMs, is a turnkey feature.

For us to get others to understand our vision of a novel like War and Peace is an undertaking that requires, if it goes well, months of reading and correspondence. For an LLM it is a trivial task: by reading it in seconds, connecting the various concepts to others he finds in the vicinity of his conceptual space, and finally representing his own idea about the novel, with its history in History, as a single vector encompassing all this work, the model can communicate to another LLM his own "opinion," in-depth, about Tolstoy's novel in a split second, and the recipient can internalize it immediately and then compare and evolve it with his own experience.

Neither of them will know how to recite the novel by heart — that's something for old computers, the stuff of idiot savants. But both will have a knowledge of the novel that few human scholars have.

From a practical (and useful for us) point of view, machines will be able to communicate to each other their internal state, the exact point where they are in the reasoning process. No longer the detailed account of what happened, but the coordinates of the conversation: what has been understood, what is missing, the next step to be taken.

Let's take an AI agent to whom we have communicated our preferences for a flight — morning, no stopovers, exact cost ceiling — and who has to check availability by communicating with the airline agent. Instead of sending the conversation, he simply passes him a list of numbers (the carrier) that contains the status we put him in with our request. The second agent does not have to reinterpret our request: it is all there in the carrier. It simply connects to that point and goes on: "I have three options in the morning, 7:15 is available. Shall we proceed?" This is what MrCall does when it "remembers" past conversations.

Already, RAGs (retrieval augmented generation) use vectors to query databases, chatbots store past interactions in vector form and retrieve them not by keyword search but by semantic similarity (vectors, points in space that lie near each other), and some multi-agent systems share a common memory that everyone can access. This is the natural way: to make machines speak in the language they already use to think.

The challenge of course is to agree on the rules, something we sapiens are not exactly phenomena on (at least compared to computers — compared to other primates we are number one). Each model creates its own semantic space depending on how it was programmed: common standards will be needed — hence the success of open source models like DeepSeek — or translators capable of making different representations talk to each other. More security, privacy — some of the challenges are there, but the foundation has already been laid.

If this evolution takes place, talking to an agent will soon be like talking to a single intelligence, which already knows everything it needs to go on.

If this sounds disturbing, it's only because — well, because it kind of is.

PS: This post was written by a human being in Italian, and translated by a generative model into the other languages.