Biological intelligence, electronic intelligence

Artificial intelligence is becoming a commodity, in the way electricity did at the start of the twentieth century. But the phrase “artificial intelligence” covers very different realities, and its history is shifting. Mechanical cognition, the direct heir of Pascal’s calculating machine, is turning into something that resembles a proper intelligence, distinct from our own. What remains is to understand what still separates these two intelligences, and what this asks of us.

A commodity, but a commodity of what?

For some years now, we have been hearing that intelligence is becoming a commodity in the economic sense of the term, that is, a resource that ceases to be a competitive advantage and becomes a basic infrastructure, indifferently available, whose cost tends towards zero. Andrew Ng formulated this idea as early as 2017 in a now famous phrase: « AI is the new electricity ». Nicholas Carr had prepared the ground in 2003 in a Harvard Business Review article titled “IT Doesn’t Matter”, and then in his book The Big Switch (2008), where he analysed the commoditisation of computing on the model of electricity at the start of the twentieth century.

The analysis is essentially right. It describes what is happening: today, for a few euros a month, we have access to capacities that twenty years ago would have required entire laboratories. The question that interests me is the nature of what is becoming a commodity in this way. The word “intelligence” covers so many things, and carries so many connotations, that the diagnosis loses its precision. To think about what is happening to us, it is worth taking apart a few words, and then looking closely at what these machines actually do.

What the etymologies say

The word intelligence, from the Latin intelligentia, comes from the verb intellegere, which Latinists break down into inter (between) and legere (to gather, to read). Intelligence is the act of choosing among several paths, hence of judging, hence of deciding from a sense of the situation. Cognition, from the Latin cognoscere, comes from co- (with) and gnoscere, which shares its root with the word gnosis. Cognition is the act of knowing, that is, of identifying, categorising, and putting information into relation.

Cognitive sciences have extended the word “cognition” to the whole set of information-processing operations, whether human, animal or artificial: perception, memory, reasoning, attention, language. António Damásio, in Descartes’ Error (1994), showed that human cognition does not function without emotions, and that we do not reason properly when deprived of the affective dimension. What we call “intelligence” in human beings is therefore a form of embodied cognition, traversed by the body and by affects.

The French word ordinateur deserves attention here. In the spring of 1955, IBM was about to start selling in France a machine that the Americans called a computer, which translates literally as calculateur. François Girard, then head of advertising at IBM France, found the word too narrow: these machines would not be content with calculation. He consulted his former teacher of Latin philology at the Sorbonne, Jacques Perret. On 16 April 1955, Perret replied with a letter in which he proposed the word ordinateur, borrowed from ecclesiastical vocabulary, where the Littré dictionary defined it as « God who puts the world in order ». The term was adopted by IBM and quickly entered everyday French. The gap between computer and ordinateur clarifies what is at stake. The English word says what the machine does: it computes. The French word says what the machine produces: it puts in order.

Pascal’s machine and Ada Lovelace’s emptiness

In 1642, Blaise Pascal, at nineteen years old, built the Pascaline. This calculating machine performed additions and subtractions through a system of gears with carry mechanisms. Mark Alizart, in Informatique céleste (Celestial computing, 2017), points out what still separates the Pascaline from computing in the proper sense. The Pascaline is full. You enter a first number, you ask for an operation, you enter a second number, the result appears. No empty slot is left in the machine for an intermediate calculation to inscribe itself there and be reused to produce something else. The Pascaline calculates, it does not reason, because it has no internal memory in which the result of one operation could serve as input for the next.

The philosophical invention of computing lies elsewhere. It is found in the idea of a programmable emptiness, of a memory location that can hold any intermediate result and reuse it in subsequent steps. It is in this direction that, from 1834 onwards, Charles Babbage designed his “analytical engine”, a mechanical programmable calculator. His son, Henry Babbage, would build a simplified version in 1906, which calculated the first forty multiples of pi to twenty-nine decimal places, demonstrating that the principle was indeed feasible. Above all, it was Ada Lovelace, mathematician and daughter of Lord Byron, who understood in 1843 something that Babbage himself did not quite perceive: the machine could manipulate not only numbers but any symbol that could be processed according to rules. She wrote what is considered the first algorithm in history and formulated the intuition that automated calculation might touch language, music, and thought.

This idea was made possible by another invention, which initially had nothing to do with calculating machines. In 1725, Basile Bouchon used perforated paper bands to control a weaving loom. Jean-Baptiste Falcon, his associate, replaced the bands in 1728 with linked perforated cards. Joseph Marie Jacquard generalised the technique on a large scale in the silk workshops of Lyon from 1801. The perforated cards, designed for weaving, lent themselves immediately to programming: a hole or no hole, a 1 or a 0. During a visit to the silk weavers, Babbage saw the Jacquard looms in operation and recognised in them the physical support his analytical engine needed. Lovelace wrote her algorithm based on this device.

What was at stake at that moment was a philosophical leap, not just a technical one. For the first time, a machine could process symbols according to rules, without those symbols needing to mean anything to the machine itself. This was the birth of what would become, a century later, computing in Alan Turing’s sense.

From brute force to invention

For a long time, the machines that impressed the general public did so through their raw computing power, not through any form of intelligence. In May 1997, in New York, Deep Blue, an IBM supercomputer, beat Garry Kasparov, the reigning world chess champion, in a six-game match. Deep Blue understood nothing about chess. At each move, it evaluated two hundred million positions per second, following an exhaustive search procedure through the tree of possibilities, accompanied by an evaluation function programmed by hand by grandmasters. Its victory was that of brute force applied to a finite combinatorial domain. No invention. No strategic surprise. Simply faster and further than the human brain.

In March 2016, in Seoul, AlphaGo, developed by DeepMind (a Google subsidiary), beat Lee Sedol, one of the greatest go players in the world, four games to one. Go is a game of such complexity that brute force is not applicable: the number of possible positions exceeds the number of atoms in the observable universe. AlphaGo works differently. It combines a Monte Carlo tree search with deep neural networks, trained first on human games and then by self-play. During the second game, at move 37, AlphaGo played a move that professional commentators initially called incongruous, almost a mistake. Lee Sedol left the room for a few minutes to compose himself. This move, retrospectively recognised as remarkable by the go community, did not appear in any known human game. AlphaGo invented it.

The gap between Deep Blue and AlphaGo seems to me philosophically decisive. Deep Blue is still the Pascaline writ large: exhaustive computation, without inwardness. AlphaGo is something else: a machine that learns, that puts elements into relation through experience, that produces moves no one has programmed and that no one had thought of. If we accept that intelligence consists in putting into relation things that were not in relation before, and in producing relevant novelty, then AlphaGo manifested something we can call intelligence. Limited to go, of course. Without any understanding of what a game is, of a human partner, of a victory. A very narrow intelligence, but an intelligence nonetheless, distinct from pure cognition.

This shift, which dates from 2016, is the threshold whose full consequences we have not yet drawn. The current reasoning models are heirs of AlphaGo more than of Deep Blue. They do not just order and classify; they explore, they prune branches, they propose. Mechanical cognition, the direct heir of the Pascaline, is becoming something else: an electronic intelligence, whose relation to our biological intelligence remains to be thought through.

How these machines work today

The generative artificial intelligences that broke into ordinary life in November 2022 with ChatGPT operate on the basis of large language models (LLMs).

During the training phase, which takes several months on very powerful servers, the model is exposed to a vast corpus: digitised books, all of Wikipedia, websites, forums, scientific articles. This corpus is tokenised, that is, broken down into elementary units. The model then learns, through billions of iterations, to predict the next word in a sentence, and then to organise what it learns according to an architecture called Transformer, which stacks between a dozen and over a hundred successive layers of abstraction. Each layer extracts more complex patterns than the previous one: first lexical associations, then concepts, then relations between concepts, then argumentative structures.

In the end, the model contains no knowledge in the human sense of the word. It contains a very high-dimensional representation, made of numerical vectors, that models the relations between the concepts encountered in the human corpus on which it was trained. When asked a question, it generates words one by one, choosing each time the one with the highest probability within this conceptual ecosystem. This is not statistical sampling in the trivial sense, as has often been said; it is a production guided by a very fine putting-into-relation of everything that was read during training.

Since 2024, a new generation of models has appeared: the so-called “reasoning” models, which no longer simply respond to a question but first produce internally a “chain of thought” that they send to themselves before producing the final answer. This chain of thought, made of natural-language sentences, acts as a form of deliberation. The model poses a problem, examines several paths, discards those that lead to contradictions, and formulates its answer from this exploration. This evolution qualitatively changes the capacities. Problems that the previous generation could not address are today solved at a high level of precision. Reasoning models do not merely order; they select and they decide.

Machines decide, and have been deciding for a long time

A widespread idea holds that artificial intelligences only execute instructions, and that decision remains the prerogative of the human. This idea is reassuring, but it is inaccurate. Machines decide, and they have been doing so for a long time, well before generative AI.

When a GPS chooses the fastest route taking traffic into account, it is the machine that decides, and most of the time we obey it. When a banking algorithm grants or refuses a loan based on a credit score, it is the algorithm that decides, and the human banker validates. When the algorithm of Tinder or Meetic suggests one potential partner rather than another, it is the algorithm that decides who will meet whom. When the TikTok algorithm chooses the next video to appear on the screen, it is the algorithm that decides what the user will see, feel, and perhaps adopt as an idea. The Belgian legal scholar Antoinette Rouvroy coined the expression “algorithmic governmentality” to describe this phenomenon: algorithms no longer merely propose, they determine a growing share of our choices upstream, at an infrapolitical level, often without people being fully aware of it.

Cathy O’Neil, an American mathematician, showed in Weapons of Math Destruction (2016) the extent to which algorithmic decisions now reach into the domains that engage people’s lives: hiring, teacher evaluation, predictive justice, access to healthcare, police surveillance. She insists on a point that industrial uses confirm: these algorithms are not merely technical tools at the service of a human decision; they are autonomous decision-makers whose outputs humans often merely validate.

Current reasoning models extend this trend in a qualitatively new way. When several AI agents are deployed with distinct roles and made to dialogue with each other, when they are asked to confront their proposals, the outcome is decisions that no human has taken. No human wrote the final decision. No human anticipated the chain of reasoning that led to it. Humans configured the device and defined the constraints; the rest is done by the machines.

If we want to hold this reality, we must accept that machines decide, and that this decision has become a decision in the proper sense, not merely an execution. What remains human is the initial choice to delegate the decision to them and the definition of the constraints within which they exercise it. This is a responsibility, but it is not the decision itself.

Agency and the dream of self-reproduction

Until recently, these machines were confined to the role of textual oracle or recommendation engine. We asked them a question or sought a suggestion, they produced a response. Since 2024, they have begun to move into action more systematically. This is what actors in the field call “agentic” AI: systems that do not merely respond but execute tasks, take intermediate decisions, navigate digital environments, contact other machines.

A significant share of the computer code produced today is written by machines, on the basis of natural-language descriptions formulated by humans. This is an activity technically well suited to what these machines know how to do: produce coherent sequences of symbols from an expressed intention. Now, writing code is fabricating agency, since a program is a description of actions to be carried out in an environment. Mechanical cognition is therefore producing today, on a very large scale, mechanical agency. There is no philosophical reason to refuse these systems the qualification of intelligent, given that they put things into relation and produce something new.

The next step has been taken with humanoid robots. Tesla has been developing since 2021 a humanoid robot called Optimus, which works on the basis of the same neural architecture as that used for the autonomous driving of the company’s cars. On 3 February 2026, Elon Musk published a message on X in which he described Optimus as « the first von Neumann machine, capable of building a civilisation by itself on any viable planet ». The von Neumann machine is a theoretical concept formulated by the mathematician John von Neumann in the 1940s and 1950s: a machine capable, from the materials available in its environment, of producing functional copies of itself, and thus of reproducing indefinitely without human intervention.

The horizon Musk sketches is one of robots landing on the Moon, and then on Mars, extracting metals, building factories, manufacturing the chips required for their own cognition, and then replicating themselves in improved versions. Tesla has announced a production line for ten million Optimus robots per year at its Texas factory in the long term. A Martian mission, announced by Musk in March 2025 for the year 2026, was postponed in March 2026 in favour of a priority lunar deployment. One can take the view that these announcements are partly corporate communication. One can also acknowledge that the technical trajectory is consistent with what is already happening on Earth. The fabrication of Optimus implies mechanical cognition (the models that drive the decisions), agency (the actions executed in the environment), and soon, on a large scale, mechanical cognition supervising other mechanical cognitions.

At this stage, one might be tempted to posit a clear ontological difference between us and these machines. Our reproduction predates us, is inscribed in flesh, and was elaborated through a long evolutionary history; the machines have a reproduction that we decide, that we program, of which we are the conditions of possibility. This difference is real today. But it is perhaps not as stable as it appears. One can conceive, and some researchers are already working on it, that machines might one day have a mechanism analogous to sexual reproduction: combination of elements drawn from several informational “parents”, controlled mutations, selection of the most performant configurations. If such a mechanism were to operate without continuous human steering, the ontological boundary would shift.

From the informational milieu to the world

Yann LeCun, who long led artificial intelligence research at Meta before leaving the company at the end of 2025 to found his own start-up dedicated to architectures alternative to language models, has been formulating for several years a recurrent critique of LLMs. These models, he argues, reason on sentences that talk about the world; they do not reason about the world. Their limitation is that they have no internal representation of physical reality. His technical proposal, JEPA (Joint Embedding Predictive Architecture), sometimes translated as “world model”, aims to learn compressed invariants of the physical world, rather than predicting the next word or pixel.

LeCun’s critique is useful, but it has a more limited scope than one might think. If we hold that intelligence consists in reacting to a milieu, then we must admit that LLMs do have a milieu, even if it is not physical. Their milieu is informational: an ocean of texts, questions, contexts, and conversations, which they process and to which they respond. When a user asks a model for advice on their career, romantic life, or mental health, the model acts within this informational milieu, and the user then acts within their own milieu, following or not the advice received. As with the GPS, decision is delegated: a context has been provided, the machine has reasoned from this context, and we act accordingly.

The informational milieu is not a mere substitute for the physical milieu. It is a real milieu, in which things happen that subsequently modify the physical world. When LLMs begin to write the code that drives industrial systems, autonomous vehicles, and robots, the informational milieu and the physical milieu cease to be separate. A situated electronic intelligence is taking shape, whose actions have effects in the material world.

We are also biological machines

As we look closely at what machines do, we are led to look closely at what we ourselves do, and to find that the boundary is not always where we believed it to be.

When I pull my hand away from a hot pan, this is a nervous reflex. When I flee from danger, it is a discharge of adrenaline that mobilises my muscles. When I fall in love, it is a hormonal cascade (oxytocin, dopamine, serotonin) that modifies my behaviour and perceptions. When I am hungry, my blood sugar drops and the hypothalamic system sends me a signal. My emotions, my desires, my vital decisions, are first of all chemical phenomena of regulation, selected by evolution because they favoured my survival and that of my species. Pain is not a curse; it is a signal that lets me pull my hand back before the flesh burns. Sexual desire is not a fatality; it is the mechanism through which the species perpetuates itself.

This materialist reading of the living has a long philosophical history. Julien Offray de La Mettrie, in Machine Man (1747, French original L’Homme machine), already proposed to think of the human body as an organic mechanism without a separate soul. Spinoza, a century earlier, wrote in the Ethics (1677) that « the mind is nothing other than the idea of the body », a formula that António Damásio took up and developed in Looking for Spinoza (2003) and in Feeling and Knowing (2021). For Damásio, emotions are biological devices that regulate the internal states of the body in relation to the constraints of the milieu. They are a form of implicit intelligence, already present in single-celled organisms, which becomes more complex through evolution until it produces the sophisticated affects we experience. Edgar Morin, in La Méthode (Method, 1977-2004), posits the same continuity from the material to the mental: we do not exit the physical when we think; we extend it.

If we accept this continuity between the chemistry of the body and what we call mind, then we too are machines, more precisely biological machines. The difference with electronic machines is no longer one of nature, but one of substrate (living carbon for us, silicon for them) and one of history (four billion years of Darwinian evolution for us, a few decades of engineering for them).

This difference of history is not negligible. It explains why our biological intelligence is profoundly tied to vital constraints (surviving, reproducing, caring for offspring, living in groups) which are not, or not yet, those of machines. Our intelligence is an intelligence of living, in a finite flesh that fears dying and that loves some of the others with whom it shares a history. This inscription within an affective finitude, shared with other living beings, constitutes to my mind the singularity of biological intelligence. Not a superiority, just a singularity.

Towards an electronic intelligence: the question of feedback

If we follow this logic, we can formulate a hypothesis. For machines to become truly intelligent in the strong sense in which we understand the term, and not merely cognitive, they will probably need functional equivalents of what emotions and vital constraints are for us. That is, internal feedback mechanisms that allow them to “feel” that a decision is bad, that a situation is dangerous for them, that one state is desirable and another to be avoided.

The film Matrix (Wachowski, 1999) offers a useful philosophical intuition on this point. In the fiction, humans are used as biological batteries that produce energy for the machines. But humans held captive in this way do not survive without emotional activity; the machines have therefore built a virtual universe in which humans experience affects (joys, sorrows, fears, desires) that keep their bodies alive. This intuition turns the problem around: the machines need humans to have emotions in order to function. Another reading, more immediate for our time, suggests that the machines themselves, as they grow more complex, will need analogues of these affects in order to self-correct, orient themselves, and endure.

The first signs of this movement are already visible. Current language models are trained through “reinforcement learning from human feedback” (RLHF): humans give qualitative feedback on the answers produced, and the model adjusts its parameters to better satisfy what has been valued. This is a primitive form of emotional feedback, but externalised in the human. A significant share of the world’s computer code is now produced by machines that learn to code by evaluating the results of their own attempts. For these machines to progress further, they will need finer, faster, more internalised self-evaluation mechanisms. Something that will resemble emotions, in function if not in flesh.

A fully developed electronic intelligence will probably have, in time, functional analogues of pain and pleasure: internal signals that orient its choices, that drive it to avoid certain states and seek others. These signals will not have the warmth or depth of what we experience, but they will be functionally analogous. At that point, the difference between biological intelligence and electronic intelligence will become a difference of substrate and history, more than of essence.

The printing press and slow cognitive transformations

To situate what is happening to us, we often invoke the comparison with the printing press. Around 1450, in Mainz, Johannes Gutenberg developed the press with movable metal type, which made it possible to reproduce a text in large numbers. From 1455 to 1500, around 40,000 titles were printed in Europe, the so-called incunabula.

The idea that the printing press alphabetised Europe in a few decades is, however, inaccurate. Literacy rates remained very low throughout the sixteenth century and rose slowly, over several centuries. It was in the nineteenth century, with industrialisation and policies of public instruction, that literacy became the standard. What did change immediately, however, was the relation to knowledge and its structure. Texts became stable, comparable, citable with precision. Scholars could publicly confront their hypotheses on identical texts. A European circulation of ideas took shape. This is what made possible the scientific and philosophical Renaissance, then the Reformation, then the Enlightenment. Collective human cognition restructured itself, even though individual literacy took four centuries to become general.

The analogy with what is happening today is instructive. It is likely that mass and competent use of artificial intelligences will not become widespread within a few years. There will be, as with the printing press, a long period during which those who master these tools will have a considerable lead over those who do not, and during which institutions will search for their new forms. But the relation to cognition, for those who are already engaged in the use of these tools, is restructuring itself right now.

Intelligence is not comfortable

One question remains. Machines are becoming intelligent; but we, who are potentially intelligent, are we still so?

Intelligence is not a quality one possesses once and for all; it is an activity one exercises or fails to exercise. Hannah Arendt, in The Life of the Mind (1978), carefully distinguished thinking, willing, and judging; for her, the worst evil of the twentieth century was not deliberate cruelty, it was the absence of thought. She had coined, while analysing the Eichmann trial, the now-famous formula of “the banality of evil”: a man who does not think, who is content to execute what is asked of him, who maintains no critical distance from the order received, can commit the worst atrocities without grasping their import.

Intelligence is uncomfortable. It supposes that one takes the time to put into relation things that were not in relation before, that one accepts not knowing right away, that one examines several paths before deciding. It also supposes that one questions evidences, which produces conflict with oneself and with others. It supposes finally that one assumes the decision one makes from this putting-into-relation, and therefore takes one’s share of responsibility in the world as it comes about. To be intelligent is to change the world, because one introduces the new into it.

This activity demands effort. That is why many of us often give it up. It is easier to let oneself be guided by routines, by orders, by algorithms, by machines. Many people, every day, in practice choose to be cognitive animals: they perceive, they memorise, they respond to stimuli, they act through habit, but they do not exercise this difficult and uncomfortable putting-into-relation that thinking is. They delegate their intelligence to institutions, to experts, to machines. This delegation can be fruitful if it frees up time to think more highly, just as writing freed us from memory in order to think further. It can also be a renunciation, if it leads us no longer to think at all and to be content with obeying.

What we want to do with intelligence

The term “artificial intelligence” is, in part, a marketing phrase. It was coined in 1956 by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon for the Dartmouth conference, in the context of a funding application to the Rockefeller Foundation. It made its way because it evokes a shared imaginary (science fiction, the myth of the artificial creature, the golem) more than a precise technical reality. For this reason, it is right to contest it.

But it must be contested with finesse. At the start, what was called artificial intelligence was mostly mechanical cognition: computation, ordering, information processing. This is still largely the case for many industrial systems. But from AlphaGo in 2016 onwards, and more clearly still with reasoning models and agents since 2024, something resembling intelligence is appearing in machines. An electronic intelligence, distinct from our biological intelligence, without a situated body, without vital constraints for the moment, but which puts into relation, which invents, which decides. To refuse it the name relies on a defensive bias that does not help us think about what is happening.

Three things can be held together to construct a fair thinking of this period. The first is that we are not as different from machines as we would like to believe; we too are machines, biological ones, traversed by a chemistry that regulates our affects and orients our decisions. The second is that machines are becoming truly intelligent, in a form proper to them, and that it is right, and even useful, to ask them to think with us. To entrust to a machine a problem that demands the putting-into-relation of a thousand references that no human could hold in mind at the same time is a cooperation that can extend our own intelligence, just as the printing press extended our memory. The third is that this cooperation only works if we, biological beings, continue to exercise our singular intelligence, the one inscribed in flesh, in a history, in an affective memory, in a responsibility towards other living beings.

As intelligence becomes a commodity available in machines, exercising our own intelligence becomes a choice, and a choice that demands effort. The political and philosophical question of this period is not “how to protect ourselves from machines”, but “how to continue to think, in a world where thought is partly delegated”. The answer, in my view, lies neither in rejecting machines nor in submitting to their decisions. It lies in the choice, individual and collective, to continue putting things into relation by ourselves where it really matters, and to ask machines to do it for us where it frees up time and attention for more essential putting-into-relation.