Marc Rameaux, an AI specialist and regular contributor to The European Scientist, sees Yann LeCun as the David Hume of our time. Tracing the career of the pioneer of convolutional neural networks, from Bell Labs to Meta and on to Advanced Machine Intelligence Labs, he shows how LeCun’s empiricism led him to expose the limits of large language models and to champion “world models”. He also sees in it a way to teach ethics to machines, rather than relying on regulation alone.
The word “visionary” is overused, worn thin by stereotyped corporate communication. That is a pity, because when used properly it denotes something far deeper. And if any expert in today’s AI landscape deserves it, it is Yann LeCun. Being a visionary does not mean making predictions about the future, nor even making them successfully because one has good intuition. It means, above all, grasping the deeper meaning of an emerging innovation and projecting its future course, because one has understood what is radically new about it, but also when and where it reaches its limits. Great scientific discoveries are said to bring about a paradigm shift. Being a visionary means seeing the new paradigm when most people see only one more innovation.
For nearly forty years, LeCun has anticipated – often against prevailing opinion – the successive turning points of artificial intelligence: the return of neural networks when no one believed in them any more, the power of convolutional architectures, the decisive role of deep learning, the strategic advantage of open source, and more recently the structural limits of large language models. LeCun is not one of those techno-prophets who pass off their bets as vision. He has shown by example the power of empiricist observation: looking at what machines actually see, how they learn, what still sets them apart from human learning. His approach revisits the philosophical categories through which we discover the world and learn from it. It is this approach that gives him such sound judgement about what AI is, what it is not and what it still lacks, much as a biologist identifies our physiological and cognitive functions one by one, understanding the role each plays in the organism as a whole.
The parallel with David Hume is not a vague philosophical reference. Hume argued that our knowledge of matters of fact does not proceed from demonstrative reason but from habit formed by repeated observation, and that the idea of cause arises from the constant conjunction of phenomena rather than from logical necessity. LeCun places artificial intelligence within this tradition and revives inductivism: a truly intelligent machine does not merely recombine symbols; it builds its own model of the world through sensory experience. Only after the fact do we describe our own thinking as a rational system. In its genealogy, however, our cognition follows the biological laws by which links between concurrent perceptions are reinforced. LeCun’s vision stems above all from patient and humble observation of nature.
An engineer trained off the beaten track
Yann LeCun was born on 8 July 1960 in Soisy-sous-Montmorency, in the northern suburbs of Paris. The son of an engineer, he developed an early passion for machines and was enthralled by his first glimpse of a thinking computer, the famous HAL of 2001: A Space Odyssey. He graduated from ESIEE Paris in 1983 and in 1987 defended a doctoral thesis on connectionist learning models at Pierre and Marie Curie University. His work already proposed a formulation of gradient backpropagation, at a time when neural networks were widely regarded as a dead end.
I am well placed to know this, because in 1991 I was one of the first to implement an industrial application of neural networks, for recognising the textures of car-seat upholstery. All my fellow statisticians seemed determined to prove that neural networks, and “backprop” in particular, could be reduced to classical statistical methods of classification or regression – as if this emerging potential already provoked the mixture of aggression and fear towards AI that is now so familiar.
Then came a decisive year: a postdoctoral fellowship with Geoffrey Hinton in Toronto in 1987–88. This encounter gave rise to a scientific friendship spanning several decades, crowned in 2018 by the Turing Award, shared with Hinton and Yoshua Bengio, for the conceptual and engineering breakthroughs that made deep neural networks the main pillar of AI.
In 1988, LeCun joined Bell Labs in Holmdel, New Jersey. It was there, far from the noise of academic fashions, that he gave convolutional neural networks their modern form. The idea did not come out of nowhere: Kunihiko Fukushima’s “Neocognitron” had already sketched an architecture inspired by the visual cortex. LeCun made it trainable end to end, introduced weight sharing and subsampling, and above all brought an engineer’s discipline to the steps to be followed. The LeNet architectures, culminating in LeNet-5, recognised handwritten digits with unprecedented robustness. AT&T turned this into a cheque-reading system, one of the first lasting industrial applications of deep learning. We now know that it was this success that began to lift us out of the “AI winter” and convinced people that the connectionist approach would go far beyond all the algorithms we had known until then.
At Bell Labs, he understood that the first applications of AI were not limited to vision. With Léon Bottou and Patrick Haffner, he contributed to DjVu, a compression format for document images long used to distribute scanned archives. With his colleagues, he also invented the method known as “Optimal Brain Damage”, a network-pruning technique. In 1996, he became head of the image processing research department at AT&T Labs. After a brief spell as a Fellow at the NEC Research Institute in Princeton, he joined New York University in 2003, where he is still Silver Professor, affiliated with data science, neuroscience and electrical engineering.
His New York years were those of a mind that refused the easy option of following the mainstream of research. He worked on energy-based models, computer vision, mobile robotics and DARPA-funded autonomous navigation projects. In 2012, he became the founding director of NYU’s Center for Data Science. A year later, Mark Zuckerberg called on him to create Facebook AI Research in New York.
Seeing ahead of others, and sometimes against them
LeCun’s story is not that of a man who has always been feted. Throughout the 1990s and part of the 2000s, neural networks were sidelined in favour of support vector machines and other statistical methods. LeCun was not discouraged: the idea of digital avatars reproducing the laws of biology was too compelling not to win through in the end. When deep learning burst onto the scene in the early 2010s, notably with AlexNet’s success on ImageNet in 2012, the convolutional architecture he had championed for more than twenty years became the standard in computer vision. Phones, cars, factories and hospitals inherited it: the first AI revolution already carried within it the wave we are now experiencing.
On generative models, his role is more that of an intellectual architect than of a sole inventor. GANs, formalised in 2014 by Ian Goodfellow and his colleagues, are not LeCun’s creation. But they belong to a lineage he has never stopped exploring: energy-based models, unsupervised learning, the idea that a machine must learn the distribution of reality rather than merely classify labels. The convolutional variants of GANs, which made it possible to generate credible images, are moreover built on the networks he popularised. One may criticise him, here as elsewhere, for a tendency to write his own theoretical frameworks too broadly into the history of other people’s inventions. One cannot criticise him for underestimating the importance of generating, and not merely recognising.
His foresight about software infrastructure was even clearer. When Google established TensorFlow as the industry standard, LeCun bet on another path: a framework designed first and foremost for researchers – dynamic, close to Python, rooted in the Torch tradition. PyTorch, developed at FAIR largely under the impetus of Soumith Chintala, first won over the laboratories and then, contrary to the instincts of many production managers, industry itself. Mike Schroepfer, then Facebook’s CTO, later admitted that he had opted for two separate toolchains, one for prototyping and another for deployment, whereas LeCun had insisted on having just one. “Yann was right, I was wrong,” he concluded.
This victory was not only technical but also political. LeCun made openness a doctrine: publishing papers, releasing code and later model weights, at a time when part of the industry was closing up. Llama is not his personal work, but it is part of a laboratory culture that he founded and defended. Without that culture, today’s landscape of open models would be very different – far less idyllic – and research would be considerably held back. His critics point out, not without reason, that open source is not an unalloyed good, and that Meta also found in it a lever of influence. LeCun replies that the alternative – tight patents, silos, captive research – would have slowed down the entire field. On this point, the record of the past ten years largely proves him right.
The Meta moment, and the break
From 2013 to 2018, LeCun headed FAIR; he then remained its chief scientist while keeping his post at NYU. He set up a laboratory in Paris as early as 2015, convinced that the French ecosystem could carry weight: after all, is he not himself a living symbol of the “French touch”? He was a visiting professor at the Collège de France in 2016. The Legion of Honour, election to the French Academy of Sciences and to the American academies, the Princess of Asturias Award: honours followed an accomplished body of work, sometimes belatedly.
Then the world swung towards large language models. LeCun does not deny their usefulness. He denies that they are the path to human-level intelligence. An alley cat, he likes to repeat, understands the physical world, remembers, plans and reasons over short horizons better than the largest LLM. A four-year-old child has taken in, through the optic nerves alone, a volume of information comparable to the entire corpus on which a language model is trained – a corpus it would take a human hundreds of thousands of years to read. Yet the cat and the child know how to clear a table. The models excel at exams and stumble over common sense. I have often pointed out in the columns of TES that AI clearly outperforms even the best humans at every strategy game, yet is still unable – or only very clumsily able – to drive across the Place de l’Étoile.
Language, even in all its richness, remains a poor summary of reality. It carries neither sensory density, nor causal structure, nor the irreversibility of actions. You can give an LLM images, sound and tools, but that does not give it an internal model of what happens when you drop a glass, when you turn left, or when you say something that can never be taken back.
This position earned him harsh criticism. The laboratories that bet on scaling achieved gains that LeCun tended to play down. Some Meta researchers felt that his scepticism had contributed to the company falling behind on the latest generation of language models, even as it was releasing Llama. Others criticise his sometimes abrasive public style, a habit of declaring “dead” things that continue to advance, and an overly hasty confidence in architectures that are still young, the JEPA family first among them. Current “world models” remain fragile outside their training distribution; several recent evaluations have shown this unsparingly. LeCun knows this but does not conclude that we must return to text: what is needed is indeed a different bet.
At the end of 2025, after more than twelve years at Meta, he left to found Advanced Machine Intelligence Labs, of which he is executive chairman. Its headquarters are in Paris. The stated goal is no longer to improve the next chatbot, but to build systems that understand the physical world, have persistent memory, and can reason and plan sequences of actions. Meta, which in his words had become “LLM-pilled”, was taking a different road. The disagreement was no longer just scientific. It was institutional.
Three missing pieces, and a model of the world
What LeCun calls a “world model” is not a sensory acquisition pipeline, however sophisticated: some LLMs are already coupled to one. The architecture he advocates, JEPA (Joint Embedding Predictive Architecture), predicts the next likely events in a representation space. The idea is simple to state and hard to achieve: the world is not entirely predictable in its details – every leaf on a tree, every flicker – and forcing a machine to reconstruct those details wastes its capacity. It must learn what is stable, causal and actionable, and treat the rest as second-order.
Every human being applies an a priori filter to perceived reality, selecting the objects deemed to be in the foreground and those considered incidental. This filter has been honed by millions of years of natural selection. Building a world model means teaching a machine not only to recognise, but to undergo such an experience.
According to LeCun, three absences condemn LLMs to a ceiling:
The first is sensory perception. An embodied intelligence does not learn about the world through labels, but through sensors. Hence LeCun’s constant interest in robotics, video, touch and sensory streams. A model that has never bumped into a table has only book knowledge of tables.
The second is the arrow of time. Understanding a cause is not a matter of two words co-occurring. It means anticipating that an action makes certain outcomes possible and others impossible, and that some of them are irreversible. Breaking a glass is not the mirror image of putting it together. Without this asymmetry, there is neither genuine planning nor responsibility.
The third is a subjective filter. A human does not interpret a scene as an undifferentiated mass of pixels. We carve it into objects according to a scale of values: what matters for acting, what is background, what is a potential threat, an opportunity or a tool, and finally what is irrelevant. This scale is not the cat’s, nor the fly’s, each of which sees the world through its own filter: to us a spider is a small creature patiently spinning its web; to a fly it is the most dangerous of predators. LLMs can describe a scene with prodigious precision, at colossal cost, because they start from a statistical magma. They can be cognitively brilliant yet lack the most basic common sense as soon as a situation requires sorting, ignoring and choosing.
Hume saw that induction is not deduction in disguise. We expect fire to burn because we have experienced it, not because some internal logic dictates it. LeCun transposes this lesson to machines: an AI whose only world is a corpus of texts is a rationalist gone astray. It strings symbols together with growing virtuosity, but it has no habit of reality. In this sense, “world models” are a rebirth and a deepening of empiricism. Not a naive return to “raw observation”, but the thesis that intelligence arises from filtering, causal learning, and not from linguistic combinatorics alone. “Continental” philosophy has too often reduced empiricism to a simplistic caricature of “sensory digestion”, a matter of inputs and outputs of the senses. Yet the great lesson of empiricism, and its depth, is that we build ourselves a model of the world before exercising our thought and action, and only afterwards do we call it “reason”.
Ethical implications: LeCun’s vision may be the best way to address misalignment
It is no accident that LeCun’s journey has crossed paths with philosophy, for it hints at ethical implications.
The web has recently been abuzz with nightmare scenarios of an “AI doomsday”, raising fears that out-of-control AIs could wipe out humanity. I had the opportunity to comment on these events in an article in La Tribune. In essence, I argue that while the risks of catastrophic “misalignment” are real, the only notable feature of this recent buzz is the way the major AI players have exploited it through attention-grabbing announcements. LeCun has – rightly – been scathing about the hypocrisy of regulations presented as altruistic safeguards, when they are merely a pathetic attempt to build a protected market.
Regulations in this field have perverse effects that make the cure far worse than the disease they are meant to fight. They look more like the profits of war profiteers than an effective remedy against war itself.
Rather than regulating AI for its ethical excesses, would the best approach not be to teach it ethics, just as it learns everything else? This is where LeCun’s journey takes on a particular depth.
If an AI were endowed with the two dimensions it lacks – the arrow of time, with its notion of irreversibility, and a scale of values that determines what is worth preserving and what is not – ethics could become internalised.
But is a scale of values not subjective? Does it not lead us back to an endless philosophical debate about whether a universal ethics exists?
Entropy and complexity: a scale of values
There is, however, a simple representation that could be used to assess the value of anything. Draw a plane with two axes. On the x-axis, the measure of entropy, with the axis running in inverse proportion to entropy: low entropy has more value than high entropy – or, in more everyday terms, high entropy characterises systems whose energy is no longer available and can no longer be transmitted. On the y-axis, Kolmogorov complexity, that is, the length of the shortest program capable of generating the entity in question. Low Kolmogorov complexity denotes a simplistic, repetitive system; high Kolmogorov complexity denotes a highly complex one.
On a plane with these two axes, four quadrants naturally appear:
Bottom left: systems of low Kolmogorov complexity and high entropy. Typically, a rubbish dump: there is no structure, the dump comes down to a simple inventory of the objects piled up in it, and it cannot produce useful energy by itself.
Bottom right: systems of low entropy but also low Kolmogorov complexity. These are well-ordered, structured entities, but following a highly predictable, tidy order: regular tiling, a building façade repeating symmetrical patterns. Such an entity has a certain value by virtue of its structure, but it follows a fairly simple construction plan.
Top left: systems of high Kolmogorov complexity but also very high entropy. What does this correspond to? Typically, randomness. A random series cannot be summarised by any program other than its own enumeration: its Kolmogorov complexity is therefore maximal. On the other hand, it has no useful energy, nothing salient that can be put to use: its entropy is also maximal.
Finally, top right are systems with low entropy and high Kolmogorov complexity. They are thus both structured and complex. They follow an order, but one so rich that it sometimes becomes unpredictable. Prigogine’s dissipative systems fall into this category, with life itself at the very top right of the diagram.
What has value could thus be taught using these two dimensions. Somewhat provocatively, one might say that ethics may be nothing more than an emergent property of thermodynamics. I would wager that an AI endowed with the arrow of time and a scale of values about the world would arrive on its own at notions such as the preservation of life, reciprocity between individuals, and even respect for the dead, as one of the combinations in the universe of possibilities whose memory and trace must be preserved.
Regulations will remain necessary, as they are for any powerful technology. They are also, by their very nature, slow, clumsy and open to abuse. LeCun’s bet is that the challenge posed by our own intelligence cannot be met by law alone, but by better-designed intelligence: able to perceive, to anticipate, to prioritise. “World models” do not guarantee alignment; a model of the world can serve abject ends as well as just ones. Yet they offer something that text scaling does not: a place where prudence can take root, because the agent finally knows that it is acting in a world that pushes back. As with humans, well-taught ethics becomes internalised: the biblical figure of Jacob gradually becomes an accomplished sage by learning ethics through experience – including when he transgresses it in his youth – until he internalises the notion of justice instead of treating it as a mere external rule.
An intelligence that grapples with the world
Yann LeCun is not Hume. He did not write A Treatise of Human Nature, and Hume did not train LeNet on postcodes. But both refuse to let reason pass itself off as a faculty separate from experience. Both make habit, observation and the conjunction of facts the true engine of knowledge. At a time when the AI industry confuses eloquence with understanding, LeCun reminds us of an old and uncomfortable truth: an intelligence that has never grappled with the world has only its vocabulary. The next milestone is not a bigger book. It is a way of looking at the world.
At a time when our leaders are flailing feverishly in the face of a technological revolution that overwhelms them and frightens them more than it excites them, French decision-makers would do well to remember Yann LeCun and to draw on his unique perspective on the world.
AI policy circles are crowded with “experts” who are in fact merely journalists, communicators or outright impostors. France’s current and future leaders should know that a genuine French AI expert is very much present – not to mention the equally important voice of Arthur Mensch – and that the vision of these two great figures could be the game changer that puts us back in the race.
This post is also available in: FR