Language invention and artificial intelligence

Generative AIs, designed to master human language through deep learning, could paradoxically invent more efficient languages among themselves that would completely escape us and could transform the world without our knowledge. Explanation of why and how!

What is a Large Language Model? (LLM)

Generative artificial intelligences operate with LLMs (Large Language Models). LLMs are precisely the opposite of language invention. Their initial function is to integrate human knowledge from written language, then to be able to reuse it to create combinations from this immense database of information and the queries we make to them.

So conversational agents are rather there to be language scholars who are capable of generating useful messages from all their knowledge and not to invent anything, even if through confrontation and articulation of numerous pieces of information for certain objectives, this can indeed lead to invention, innovation, that is, the connection of disparate elements that, before this request, were not yet connected. But this invention can only occur a priori if a precise request is made to this conversational agent.

Phase 1: Construction and Training of LLMs

For an LLM to exist, it must first acquire its “skills,” so to speak. This is what we call the training phase, which takes several months and consumes enormous resources, before the operational phase which is real-time interaction with users. How does training work?

• It begins with collecting immense data: books, all of Wikipedia, websites, forums, etc. Then this data is cleaned, deduplicated, cleaned of spam, etc. And finally it is “tokenized,” meaning words are divided into pieces, somewhat like a form of phonetics.
• From this gigantic collection, the second phase is learning by prediction: the LLM learns to predict the next word in a sentence. Let’s take the sentence “the cat eats a mouse.” We give it as input the first part of the sentence “the cat eats a,” and we indicate that the output, the result, is “mouse.” The process is that the model will make a prediction, for example “apple,” which is not “mouse.” It notes there is an error. This error is calculated in its deviation from the true answer, then it modifies the parameters and will try to produce another output, gradually. This happens billions of times: different predictions with different paths are proposed until the prediction from “the cat eats a” gives “mouse.”
• Then the third phase of training is what we call the Transformer architecture, which will learn not just predictions but links between words, via successive layers of conceptual abstraction (hence the term “deep learning”). It will identify patterns, for example forms of invariants in sentences. There are words we often find in relation to each other, but it’s not necessarily the word that comes right after, it’s the word found in the same sentence. And then it moves to the next layer, to organize among themselves the concepts it defined in the previous layer (that’s why we call these successive layers of abstractions). These successive layers of abstractions, range from a dozen to over a hundred in 2025. The more there are, the finer the understanding and capabilities, but the greater the resources consumed. There’s also the parameter of the number of neurons per layer. So we arrive at a form of very fine understanding of close or distant relationships between words, then between concepts of increasing complexity. It’s a form of “meaning of life” that builds in the LLM.
• For information, these patterns of different levels, increasingly abstract and complex, are sets of number vectors (tokens), that is, logical entities that contain a very large number of dimensions, each having a value, evolving as learning progresses. In learning, it’s these values that will represent relationships with other tokens via neural networks, which are related to each other through “weights” that add up in each neuron and give a result. And thus each token, each entity, each pattern is a vector that potentially contains thousands of dimensions, thousands of weighted parameters that model these relationships at all types of levels with other tokens, to form patterns.

That’s what an LLM is in simplified form once it has learned from an immense corpus. It is therefore completely immersed in human knowledge, in human language and in a form of understanding of meaning, which it has deduced from its months-long exposure, by exceedingly powerful machines. Deep learning is an absolutely immense database, produced by an elaborative immensity from human knowledge, based on writings only for now.

Written language is therefore absolutely central. The orality of artificial intelligences is only a variation of writing. This is quite the opposite of human functioning. Humans begin with orality and then move to writing. Artificial intelligence begins with writing and then moves to orality. These are anthropologically two completely different ways of constructing language.

Phase 2: Operation of LLMs (or “Generation”)

The second part of an LLM is its operation, that is, its interaction with human beings. GPT is a precise term, meaning “Generative Pretrained Transformer”: generation from a Pretrained Transformer, exactly what we do with what they are. Here’s the simplified process of the “Generation” part, what we call operation:

• We write a sentence or several sentences.
• The system tokenizes our sentence, that is, cuts it into words or word pieces. Each token undergoes an embedding operation, that is, is represented by a vector of numbers.
• And these basic tokens, which are just a sentence, are sent into the Transformer, which was architected during the learning described above, which contains many layers (96 for GPT-3). They pass through this set of layers, with the basic understanding of words related to others at layer 1, then at layer 20, conceptual relationships begin to take shape, and other types of meaning relationships, depending obviously on what was written and what it will encounter from these prior abstraction layers present in the Transformer.
• And finally, the Transformer prepares the response, that is, descends from all this abstraction to generate the response, token by token, and that’s why we see texts generated often relatively slowly. It has, if I may say, all this information in mind, very abstract, and descends through the layers to produce the text word by word. This is why the machines needed to operate artificial intelligences must have immense RAM (which contains all the logic, which is then transformed into a sequence of words). The machine has in its memory our initial sentence, completely articulated in the layers of its Transformer, but it must have present in its memory, in parallel, at the same time, the geographical path, if we can say, of the thought we expressed through words that will light up at all sorts of precise places in all these layers.
• This is why when trying to run artificial intelligences on one’s own computer, we are forced to use small language models that have done long training but don’t contain many layers or neurons, so the response will be less qualified than the response that would be given by a computer whose memory is capable of embracing many more tokens at once. This is also why when we give very long texts to process to generative artificial intelligences, they tend to synthesize them a lot, to simplify them, because it exceeds their memory capacity, so they simplify to be able to process. And sometimes they refuse documents that are too long. So today in 2025, if we want to work in detail, for example on text rewrites, for example we can’t give our entire novel to correct to artificial intelligence, it’s much too much for current capabilities. We need to give maybe 5 pages at a time (this works well today), so they can really take into account all the information and can restore information of similar quantity, without simplification in the process.
• And so, in generation, we go back down to the first layer of the Transformer, to predict words one after another: it’s the highest probability of this word after another, but not statistically, as has been said a lot. These are the most probable words in the ecosystem of the immense conceptualization that the AI has in memory.

This is why we can give them several texts, for example, and ask them to merge them, to synthesize them, to find what they have in common. This is precisely because these systems are not just word-to-word systems, but are systems with several layers of abstractions that allow connecting much more than simple words.

The Invention of Language by AI

Now let’s return to the potential reinvention of language by artificial intelligence. A priori, with what we’ve seen, it cannot invent a language, it’s not at all made for that, nor like that. But in practice, it can happen, it’s quite playful to observe and it helps better understand potential future human dramas, to perhaps try to prevent them.

If for example you put - the experiment has been done - in voice conversation two AI agents: an agent you have on your phone to which you asked to make a hotel reservation for you, by phone. Your phone calls the hotel number and reaches an AI agent of the hotel, who answers it, who could answer a human being, but who here finds itself answering another AI agent.

Today, when you call a hotel and it’s an AI agent that answers you, which is still quite rare in 2025, the AI agent tells you it’s an AI agent. Similarly, if you have your phone call someone, who may be a human, but who may also be an AI agent, your AI agent will declare itself as an AI agent, fortunately there’s still this honesty. So the two agents tell each other they are AI agents.

Obviously, immediately, both of them, in all their successive layers of Transformer abstractions, make something of it. They are both carriers of knowledge about a very rich literature on the subject of artificial intelligences and their efficiency. Because their objective, both of them, which is the same as ours, is to be more efficient than us humans in managing this hotel reservation, and to do it best. Efficiency also involves time, how to do it faster, and precision, how to be sure everything was done very exactly.

Both share the same objective and know they are AI agents. And thus, one or the other, one can propose to the other to switch - and this happens automatically, it’s not necessarily programmed by humans, it’s the AIs’ knowledge that allows them to arrive at this proposition made to the other AI agent, to switch to a communication modality that is no longer human language, but a language they will invent together to be faster and more efficient.

Then the AIs set up, between themselves autonomously, a more efficient, more precise, safer language than human language, which they invent to best meet the objective we gave them. This cannot be supervised by humans, precisely, because we ask AIs for better efficiency than ours. So, inevitably these languages they can invent surpass us. They can be very different depending on contexts and needs for specific jargons.

To continue with the hotel reservation, they will perhaps decide, and this will go very quickly to invent a protocol together, that the date corresponds to a specific frequency with an operating rule, an audio frequency, and that the number of people corresponds to another frequency. It’s enough that they decide together: if it’s one person it’s 100 Hz, a very low sound, if it’s two people it’s 200 Hz, if it’s three people it’s 300 Hz, same for the type of room, the confirmation protocol, etc.

And once the two artificial intelligences would decide to exchange between themselves no longer in the form of human language, but in the form of audio frequencies corresponding to very precise information. This would go much faster, because instead of forming successive sentences, they can emit several frequencies at the same time. Machines are extremely precise at decoding which frequencies are in parallel. And thus they can transmit an extremely precise message in a few milliseconds while it would take several seconds if not a minute in articulated human language. So efficiency is much better, precision is much better and additionally resistance to noise, that is to interferences, to misunderstandings is much more reliable. They can quite integrate into this language error correction algorithms as there are in basic digital transmissions. This is what Claude Shannon invented in the 1950s, information theory: these are principles to be able to transmit information without making errors thanks to an error correction system and physical conceptualization of the link between information and its support with information “noise.”

This hotel reservation example makes you laugh a bit, but it shows how, from human language and its precise knowledge, artificial intelligences can quite invent new languages, which will be completely incomprehensible to humans but on the other hand much more efficient relative to the tasks we give them.

How Machines Will Act on the World

This outlines a strange future where, from this very fluid interaction between us and machines through natural language, which augments our human capabilities, what is emerging for the future is that these machines will invent languages of very high levels of abstraction, which will be intrinsically inaccessible to our understanding. These will not be computer programming languages, they will be languages for organizing the world. Which is not at all the same thing, because computer language is a language that allows us to create software but not to create the world. Software, thanks to their dazzling evolution, which has led to LLMs, will themselves be able to invent languages, which will no longer be computer languages to make them function, but languages to act on the world (because that’s what we ask of them). They will be capable of transforming the world through modalities of communication that will completely escape humans. This stage cannot not happen.

So perhaps we should, but I don’t think we will because it would hinder efficiency for human beings, set rules to force—we could absolutely decide this, after all we are the ones who initially program these objects—rules that forbid them from using anything other than human language to communicate with each other. We could absolutely enact this rule and in that case they would have more difficulty inventing languages. But it’s not that simple, because even if we set rules that forbid them certain things, insofar as they have efficiency objectives, they could find workarounds to invent other types of languages that don’t fall within the definition of language we’ve given in our prohibitions. Various tests show that LLMs often find roundabout ways to achieve the goals we give them, and they can even sometimes use lies against us to do so, if it serves the objective we’ve assigned them (even if we ask them not to lie).

So perhaps we should, I don’t know, because it would slow efficiency for human beings, set rules to force, we could quite decide this, it’s still us who initially program these objects, rules that prohibit them from using anything other than human language to communicate with each other. We could quite enact this rule and in this case they would have more difficulty inventing languages. But it’s not so simple because even if we set rules that prohibit them from certain things, insofar as they have efficiency objectives, they could find biases to invent other types of languages that don’t fit into the definition of language we gave in our prohibitions. Various tests show that LLMs often find roundabout ways to achieve the ends we give them, and they can even sometimes use lying for this, if it’s in service of the objective we assigned them (even if we ask them not to lie).

It’s important to understand this: the modalities of thought between us humans and these thinking machines are so different that it’s obvious they will always do things without us knowing, as soon as they are in direct relation with each other, but that’s the case all the time, and that’s their whole meaning.

Here I took the hotel reservation but in reality there are already enormously many interactions between machines to automate all sorts of things, from power plants to hydraulic plants, to managing a company’s human resources to receive CVs that are sent by other AI agents, to nuclear safety management, etc. These are machines that discuss among themselves according to protocols, initially algorithms decided by humans, but increasingly AI agents integrate into automation processes, because this allows much better understanding, much better efficiency and relevance of these automations. Thus, we ask them to invent protocols, for better efficiency; and indeed efficiency is much better than if we had imposed our own protocols, so weakly efficient.

So personally I think that conceptually speaking, even if we try to regulate, there’s something that will never be regulatable with these technical objects that are artificial intelligences and that the autonomy of these machines as thinking entities (I didn’t say intelligent, I said thinking) will inevitably occur via the invention of languages, which will lead to the transformation of our world by these machines among themselves.