Yann LeCun didn\u2019t come to VivaTech to write off large language models. He uses them, finds them useful, and acknowledges their effectiveness in translation, code generation, and symbolic manipulation. But he rejects the idea that has become dominant in parts of Silicon Valley: that simply scaling up LLMs is enough to achieve human-level intelligence. <\/p>\n\n
For the French researcher\u2014winner of the 2018 Turing Award and a leading figure in deep learning\u2014this belief stems from a form of monoculture. Language models know how to manipulate symbols. They excel in fields where reasoning can be expressed as a sequence of symbols, particularly mathematics and code. But understanding the physical world, anticipating the consequences of an action, and planning in an uncertain environment\u2014all of this requires something else. <\/p>\n\n
This is the core of the project at AMI Labs (Advanced Machine Intelligence Labs), the company he founded after leaving Meta. Reuters confirmed that LeCun is leaving Meta at the end of 2025 to launch a company dedicated to this approach, after twelve years with the company and a central role in the creation of FAIR, Meta\u2019s AI research lab. <\/p>\n\n
LeCun\u2019s insight can be summed up in a single phrase: \u201cworld models.\u201d<\/p>\n\n
A truly intelligent system shouldn’t just predict the next word. It should be able to construct an abstract representation of its environment, imagine an action, and then anticipate what will happen next. This is what humans do when they avoid an obstacle, prepare a movement, drive, do DIY projects, cook, or plan a complex sequence. We don\u2019t need to reconstruct every pixel of reality. We retain the useful structures. <\/p>\n\n
LeCun has been working for years on these architectures, known as JEPA, short for Joint Embedding Predictive Architecture. Their purpose is not to generate an image or video in its entirety, but to learn an abstract representation that allows for predicting how a situation will evolve. This is the key point: intelligence does not consist of reproducing everything, but of knowing what is worth retaining. <\/p>\n\n
This criticism is directed squarely at the present day. LLMs give the impression of mastery because they are articulate. They give linguistic form to available human knowledge. But for LeCun, they remain essentially machines for disseminating knowledge, not systems capable of generating a deep understanding of reality. <\/p>\n\n
The shift is as much scientific as it is cultural.<\/p>\n\n
For more than a decade, Meta has embodied a certain vision of open AI, particularly with Llama 2, whose open-source nature has enabled many players to build applications and businesses. LeCun recalls having championed this approach internally. Today, he regrets that part of the industry has become more insular, focusing on proprietary models, short cycles, and a competitive landscape where everyone is following the same path. <\/p>\n\n
His advocacy for open source is not just technical. It is political. <\/p>\n\n
In his view, artificial intelligence is becoming a cultural medium comparable to the printing press. Restricting access to models under the pretext of danger would, in his view, amount to repeating the control mechanisms that institutions have historically employed in response to technologies for disseminating knowledge. The parallel is deliberately stark. The circulation of knowledge carries risks, but it remains a prerequisite for scientific, democratic, and cultural progress. <\/p>\n\n
This is also the purpose of its involvement in Project Tapestry, an initiative led by the AI Alliance, which brings together more than 200 organizations. The project aims to build an open-source platform for the federated development of AI models, enabling institutions, countries, universities, and companies to contribute to the training of open models while retaining control over their data. LeCun serves as Chief Science Advisor for the project. <\/p>\n\n
The issue is clear: sovereignty is not just about having a national model. It requires the ability to participate in shaping the very cognitive infrastructures of tomorrow. Today, many stakeholders download models. Few contribute to their pre-training, architectural choices, biases, or direction. <\/p>\n\n
Project Tapestry aims to shift this boundary. The data remains local. Contributions are made through federated learning. The shared model can then give rise to sovereign variants tailored to specific languages, cultures, legal frameworks, or industrial priorities. <\/p>\n\n
This issue ties into a broader concern. If, in the future, our AI assistants manage our access to information, they will become the new gatekeepers of culture. If these assistants are controlled by a handful of American or Chinese companies, the diversity of perspectives, languages, and points of reference could be diminished. <\/p>\n\n
LeCun does not promise human-level intelligence by 2027. He even rejects the idea of an imminent arrival of \u201cgeniuses in a data center.\u201d His approach is more gradual and more demanding: we need to change the architecture, build systems capable of anticipation and planning, and ensure a diversity of stakeholders in the development of AI. <\/p>\n\n
At a time when the industry sometimes confuses computing power with intelligence, this step back is a welcome one.<\/p>\n\n
The future of AI may not lie in the model that speaks best, but in the one that finally understands what it means to speak.<\/p>\n","protected":false},"excerpt":{"rendered":"
Yann LeCun didn\u2019t come to VivaTech to write off large language models. He uses them, finds them useful, and acknowledges their effectiveness in translation, code generation, and symbolic manipulation. But…<\/p>\n","protected":false},"author":2,"featured_media":2061811,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_lmt_disableupdate":"","_lmt_disable":"","footnotes":""},"categories":[77972],"tags":[78051,78110,78111,77969,78104,78105,78108,78109,78106,78107],"class_list":["post-2061830","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-luxury-and-ai","tag-ai-agents","tag-ai-alliance","tag-ai-sovereignty","tag-artificial-intelligence","tag-llm","tag-meta-ai","tag-open-source-ai","tag-project-tapestry","tag-world-models","tag-yann-lecun"],"_links":{"self":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2061830","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/comments?post=2061830"}],"version-history":[{"count":0,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/posts\/2061830\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media\/2061811"}],"wp:attachment":[{"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/media?parent=2061830"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/categories?post=2061830"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.luxsure.fr\/en\/wp-json\/wp\/v2\/tags?post=2061830"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}