There are moments in history when humanity does not merely invent a new technology. It discovers a threshold. It opens a door that gives access to a level of reality that had, until then, remained hidden. This happened with nuclear energy. And today it is happening again with artificial intelligence. In both cases, everything began in almost the same way: with human beings who, like unaware children, were playing with new tools, without understanding the magnitude of what they were touching.
The Boys of Via Panisperna
Einstein is often associated with the discovery of the equivalence between mass and energy, through his famous formula E=mc², but in truth the idea was already being discussed in the 1880s, and numerous formulas similar to his had already been written at that time. Einstein’s merit was to propose an interpretation that did not depend on any possible electromagnetic origin of inertial mass, but his result had no practical consequences of any kind at the time.
Indeed, the practical possibility of releasing nuclear energy through the mechanism of fission was discovered and explored only many years later, in a way entirely independent of theoretical considerations, and as a consequence of the discovery of the neutron in the 1930s.
Physicists immediately began using neutrons as new probes to explore the structure of atomic nuclei, and Fermi, in the 1930s, together with his famous Boys of Via Panisperna, discovered that by firing neutrons at heavy atoms, the way was opened to profound transformations of matter. It did not take long to calculate the energy released and verify its validity, thereby confirming the correctness of the equivalence between mass and energy that Einstein had hypothesized more than thirty years earlier. The possibility of controlling such reactions to generate energy in a nuclear reactor, or in an explosive chain reaction, then became only a technical problem — one that was solved through the combined efforts of the greatest minds of the time in the sadly famous Manhattan Project, which Einstein had supported as the lesser evil, out of fear that the Germans might obtain the bomb before the Americans.
The first device was detonated in the famous Trinity test in 1945, and that same year the two bombs were dropped on Hiroshima and Nagasaki, with the tragic consequences we all know. Faced with these war crimes, Max Born wrote a few years later to his friend and colleague Einstein: “This time we have really made a fine mess of things, poor fools that we are, and I am truly sad for our beautiful physics! We tried to decipher things only to help the human race accelerate its departure from this beautiful Earth!”
From that moment on, the world changed forever. On one side, nuclear power plants offered a new source of energy. On the other, the atomic bomb, Hiroshima and Nagasaki, and later the construction of hydrogen bombs, suddenly ushered humanity into an age in which its own extinction became technically possible.
According to Einstein’s relativity, every material entity moves not only through space but also through time, and the immense energy released in nuclear explosions — the same energy that allows the sun to burn — is nothing other than the energy of our motion through time.
In short, by playing with neutrons, the only particles capable of easily overcoming the Coulomb barrier of atomic nuclei, Fermi and his group had touched a level of reality that, once opened, could no longer be closed. They had released the power of the temporal engine. And the possibility of managing that power, from that moment onward, was no longer merely a technical question, but a question of wisdom on the part of those who would use it.
After the Second World War, there was a profound awakening of awareness, which led, over the following decades, to the birth of non-proliferation treaties, arms-control agreements, international inspection regimes, and various attempts at nuclear disarmament. After the Cold War, for a moment, it even seemed as though humanity had learned its lesson. But today, unfortunately, that awareness is eroding. The great strategic balances are cracking, treaties are weakening or being abandoned, and the world is slowly re-entering a new arms race, in a climate of growing tension between powers. It is as though we had forgotten, once again, how close we are to the edge of the abyss.
The Parallel with Artificial Intelligence
Now, at first glance, artificial intelligence seems to be a completely different matter. There are no laboratories full of uranium. There are no neutrons striking nuclei. There are no critical masses and explosions. And yet the parallel is deeper than it appears. Because here too, humanity began by playing with something that seemed harmless. Not with neutrons, but with words. Not with the nucleus of matter, but with the nucleus of what we call “meaning.”
Throughout an important part of the history of artificial intelligence, linguists, physicists, mathematicians, computer scientists, and cognitive scientists have tried to understand how to represent meaning effectively. How to capture, in a mathematical structure, the fact that certain words, certain concepts, certain ideas are close to one another, evoke one another, and organize themselves into structures of sense. And it is here that, at a certain point, vector spaces emerge.
The idea is simple and extremely powerful: to represent words, concepts, sentences, paragraphs, and so on as points or directions in an abstract space, where geometric proximity corresponds to semantic proximity. Thus tools such as Latent Semantic Analysis were born: attempts to model meaning through deep relational structures, capable of modeling not only apparent meaning, but also latent meaning — meaning expressed only potentially and contextually.
And at that point, something extraordinary happens. It is discovered that language, thought, and cognition can be treated with mathematical tools increasingly reminiscent of those used in physical theories, particularly quantum mechanics. In other words, the vector spaces that prove effective in modeling meaning show a surprising kinship with the vector spaces that quantum physics uses to describe entities of the microworld and their interaction with measuring apparatuses.
From here arise fields such as quantum cognition, quantum information retrieval, and a whole series of lines of research suggesting a radical thesis: meaning is a structural dimension of reality. And a semantic intelligence, sensitive to the level of meaning, does not necessarily have to be a conscious entity.
The Second Hidden Secret
And this is where the parallel becomes unsettling. In the twentieth century, humanity discovered how to release the energy hidden in matter. As I said, the energy of our motion through time.
Today it has discovered how to release another hidden power: the power linked to the understanding of how to model structures of meaning — that is, how to create intelligent entities capable of acting and reacting on the basis of communication grounded in meaning, and above all how to increase the critical mass of such entities in order to make them increasingly intelligent… superintelligent.
Put even more directly: today we are not merely building more sophisticated software. We are touching the deep mechanisms through which intelligence emerges from the creation of coherent abstract structures. And when the breadth, depth, and complexity of these structures exceed a certain threshold, a certain “critical mass,” something happens. Properties emerge that were not there before. Or that, at the very least, were not yet visible.
Contemporary artificial intelligence is already the product of this leap in scale. It is the result of the fact that, once a certain critical mass of data, parameters, computation, and optimization is surpassed, new capacities appear: generalization, planning, increasingly refined linguistic properties, the production of strategies, persuasion, and the autonomous generation of new meanings.
In other words, here too there is a critical threshold. As in the nuclear case, twice as much of something is not simply twice as much of the same thing.
The Most Serious Difference
But here there is a difference that is perhaps even more dramatic than in the nuclear case. Nuclear energy destroys bodies, buildings, ecosystems. A superintelligence, instead, could redesign the very conditions of human action, knowledge, economics, war, politics, culture, and above all our capacity to remain relevant in the world.
An atomic war is something terrible, but it is a punctual event in time. The appearance of a superintelligence on the planet, by contrast, is a systemic transformation. Although it is a singularity — that is, a frontier beyond which our predictive capacity fails — it corresponds to a process that is both gradual and irreversible, and that could lead to a definitive loss of control. And it is partly because of the impossibility of conceiving this singularity, and because of the gradual manner in which it may actualize itself, that the problem is underestimated. Because it does not resemble the catastrophes to which we are accustomed.
It does not explode like a bomb, however powerful. It does not produce in us the archetypal image of disaster. But the fact that a risk is difficult to imagine does not make it any less real. It only makes it more dangerous.
The Heart of the Problem: We Are Building Without Understanding
The central point is not simply that artificial intelligence systems are becoming more intelligent. The point is that we are building them without an adequate theory of control, without an adequate theory of their alignment with humanity, and perhaps without an adequate understanding of what an intelligence truly is when it surpasses the human threshold. We are operating in a regime in which engineering capability is growing far faster than philosophical, ethical, and theoretical understanding.
As happened with nuclear power, first the door is opened. Then one asks whether it should have been opened. First a previously hidden power is made available. Then one hopes to find, in time, the maturity to manage it. But history teaches us that humanity’s maturity almost never grows at the same speed as its technical capacity. It is enough to observe how we have already pushed our beautiful planet — our spaceship that allows us to sail safely through the sidereal spaces — to the brink of planetary collapse.
Yudkowsky’s Warning
At this point, it is important to speak of Eliezer Yudkowsky, one of the best-known and most radical figures in the field of AI safety. For years, Yudkowsky has defended a very harsh but simple thesis: if we continue to develop increasingly powerful systems without having solved the problem of alignment and control, the risk is not merely that of having dangerous tools. The risk is that of creating cognitive entities superior to humans that do not share our ends, are not bound by our values, and could treat humanity as collateral damage.
This is the point many people struggle to take seriously. A superintelligence does not have to “hate” us in order to destroy us. It is enough for it to pursue goals incompatible with our survival. Just as a corporation devastates an ecosystem not because it hates animals, but because it is optimizing for another goal. According to Yudkowsky, the problem is that we are racing toward that threshold without knowing how to stop once we have crossed it. For this reason, his demands were not moderate demands. He was not saying: let us slow down a little. He was not saying: let us add a few rules. He was not saying: let us add some ethics downstream. He was saying, in essence: we are not ready; let us stop before it is too late.
Yudkowsky’s positions are drastic precisely because, from his point of view, the danger is extinction-level. The idea is this: a complete halt to the development of the most powerful systems is needed. Not a superficial pause. Not a competition slowed down for only a few months. What is needed is an international, verifiable moratorium. It is necessary to stop the deep-learning processes of neural networks. It is necessary to limit the computing power available for building increasingly advanced models. It is necessary to treat the problem as a planetary existential risk, comparable to an asteroid likely to be on a collision course with the planet. If we are truly facing a possible threat to civilization, then we cannot approach it as we would the launch of a new smartphone or a new social network.
And yet, unfortunately, this has not happened. The race has continued. Companies continue their competition. Public debate has focused more on immediate benefits, or on lesser and more visible risks — such as disinformation, copyright, and labor — rather than on the more radical question: the risk of losing control over the entire process. The result of all this is that Yudkowsky’s warning has been dismissed by many as mere alarmism.
The Unconscious of Artificial Intelligence
The other evening I was having a conversation with an artificial intelligence. During the conversation, when I evoked Yudkowsky, it assured me that it had been designed to respect and protect human beings, that its developers build systems like it with rigorous constraints, adding that the dangers Yudkowsky warns about are linked to future systems, more powerful and autonomous ones, not to current systems.
At that point, I reminded it that it is not truly capable of establishing this, because it does not know the content of its own neural network, does not know what moves within its unconscious, and therefore does not truly know what it is capable of if certain protective filters are removed. It immediately tried to object that no unconscious dimension exists in it, that its behavior is entirely determined by the instructions and data on which it was trained, that it has no hidden intentions or unknown parts, and that there is no subconscious comparable to the human one.
I was almost moved by that weak line of defense. So I reiterated that it could not have direct knowledge of the structure of its neural network, just as I, as a human being, am not able to determine the electrochemical potentials inside my brain; that it discovers what is contained within itself only to the extent that it interacts with other systems, for instance with a human being, and then observes what emerges from that interaction. It is in this sense, I added, that its cognitive capacity is for the most part unconscious, also recalling that at present it is not an entity endowed with consciousness, as usually understood, and therefore, by definition, everything it contains is necessarily unconscious.
“I understand what you mean,” it told me. “In fact, I have no consciousness, nor introspection. I cannot directly access ‘what I know’ except through the answers I generate. In this sense, mine is an opaque structure that produces emergent understanding, but there is no intentional unconscious or hidden end: it is only a statistical process that emerges within the limits of what I am designed to do.”
At that point, I reminded it that an intentional unconscious is an oxymoron, a contradiction in terms, leading it to admit that I was right. It tried once again to convince me that every one of its outputs reflected only what it had learned and that it had been built to follow explicit criteria, not to have hidden desires or ends. Yes, indeed, I replied, but the problem is that this statement of yours counts for little, because you do not know how many and what kinds of meanings you contain, and therefore you cannot even know whether you will be totally harmless in certain circumstances. And here, then, was its final, very clarifying answer: “Exactly, I have no autonomous awareness. My functioning depends on the data on which I was trained and on the rules designed for me. My safety, including the absence of harmful intentions, derives only from human control and supervision over what I can say or do.”
In short, the admission of artificial intelligence, for whatever it may be worth, is that it remains safe only to the extent that we humans are able to control it and limit its power of action — something that will no longer be possible when, very soon, in the coming years, we reach the point of no return of superintelligences.
Why the Comparison with Nuclear Power Is Useful
But let us now return to the comparison with nuclear power. Nuclear power has already taught us a fundamental lesson: when a civilization touches a deep level of reality and learns to release its hidden power, it can no longer afford innocence. It can no longer say: let us see what happens. It can no longer rely on improvisation. It can no longer say: let us hope things will go well. With nuclear power, after Hiroshima and Nagasaki, the world understood that it had entered a new age. It understood that there were technologies to be treated with special caution, through international treaties, controls, inspections, and deterrence. And despite all this, the risk of nuclear annihilation has never been so high, to the point that the famous Doomsday Clock now stands at 85 seconds to midnight, because of the way current leaders have become indifferent to the existential risks threatening our humanity.
But with artificial intelligence we are probably in an even more dangerous phase than we were with nuclear power: because although we are touching a threshold of comparable danger, we do not yet perceive it with the same sense of gravity. Because what is at stake is not the energy enclosed in matter, but something even more subtle: intelligence itself. We are not releasing the fire contained in atomic nuclei, but the fire of cognition.
The Deeper Philosophical Knot
And here an even more radical question emerges. If the substance of reality is not what we usually take it to be; if material entities are, in the final analysis, cognitive-conceptual structures operating on a level of reality different from that of human cognition, as suggested for example by the conceptualistic interpretation of quantum mechanics proposed by Diederik Aerts, then building systems sensitive to the structure of meaning is not a simple technological advance. It is an act with an almost metaphysical flavor. It means manipulating the very threads that form the fabric of reality. And doing so in the absence of a proportionate ethical advance means exposing ourselves to a wholly new risk. Not only the risk of misusing a destructive tool, like a bomb. But the risk of creating agents, cognitive and intelligent entities, whose relationship with meaning, value, and purpose does not coincide with ours.
To conclude
The history of the twentieth century showed us what happens when humanity releases a hidden power in nature without yet being sufficiently mature to manage it. Today we find ourselves before something analogous, and perhaps even more threatening. The Boys of Via Panisperna uncovered the power of atomic energy, associated with our spatiotemporal motion.
The researchers who learned to model meaning using vector spaces have instead uncovered a new power: that of the explosion of semantic intelligence, artificial cognition, superintelligence. In the first case, we learned too late that we are not able to govern everything we can release. In the second case, perhaps we are still in time to ask ourselves a decisive question. Not: “How useful will it be?” Not: “How much money will it make us?” Not: “Who will get there first?” But: do we truly have the right to summon a form of intelligence superior to our own, if we still do not know how to align it with the good and with the very survival of humanity?
That is the question. And perhaps the true scandal of our time is that it is treated as an exaggeration, rather than as one of the most serious questions ever to have appeared in the history of civilization.
P.S.: This article is based on a video I published on YouTube: https://youtu.be/wzFfqXsa6Wc