Yann LeCun, one of the foundational figures of modern deep learning and a recipient of the Turing Award, has consistently occupied a position as a profound skeptic regarding the industry’s central fixation. While the global technology sector pours trillions into scaling up Large Language Models (LLMs)—the architectural basis for systems like ChatGPT and Claude—LeCun has declared this approach a strategic dead end for achieving true human-level intelligence. He argues that the current paradigm, despite its impressive linguistic feats, is fundamentally flawed and incapable of addressing the most critical challenges involving interaction with the physical world.
This fundamental disagreement has culminated in LeCun’s departure from his role as Chief Scientist for Fundamental AI Research (FAIR) at Meta, the influential lab he founded. He has announced the formation of a new Paris-headquartered enterprise, Advanced Machine Intelligence (AMI), pronounced ami—the French word for friend. This venture is not merely a pivot away from big tech management; it is a calculated, contrarian move aimed at establishing a global, non-aligned power dedicated to building the next conceptual generation of artificial intelligence: World Models.
The Geopolitical Third Path and the Open-Source Mandate
The establishment of AMI in Paris is a statement of intent, positioning the company as an ambitious alternative to the established US-China duopoly that currently defines the AI landscape. LeCun cites a "very high concentration of talent in Europe" that often lacks the necessary environment to flourish, coupled with a palpable demand from governments and industries worldwide for a frontier AI partner that is neither American nor Chinese. This strategy taps directly into growing concerns over technological sovereignty.
LeCun views AI as an essential future platform technology, and history suggests that successful platforms ultimately gravitate toward open-source models. He sharply criticizes the American industry—specifically citing the proprietary walls erected by former open advocates like OpenAI, and the closed nature of companies such as Anthropic—as making a critical strategic error. He suggests that increasing secrecy driven by competitive pressure will ultimately slow innovation and cede ground to rivals.
In contrast, he notes that China has enthusiastically embraced the open-source approach, resulting in many leading open-source platforms originating there. While acknowledging the quality of Chinese engineering, LeCun raises a compelling geopolitical concern: a future where the choice of informational mediators is limited to proprietary US-centric models or Chinese open models that may require significant political fine-tuning to address sensitive global topics.
The AMI philosophy advocates for a robust diversity of AI assistance, mirroring the necessity of a diverse press. By championing open-source models that can be locally fine-tuned by diverse linguistic and political communities, AMI aims to ensure that AI platforms reflect a broad spectrum of value systems and interests, thereby mitigating the dangers inherent in global cognitive reliance on a handful of culturally or politically monolithic systems. For venture capitalists, this open-source stance is highly attractive, as proprietary models present strategic risks and prohibitive costs for smaller startups seeking to innovate on foundational layers.
The Delusion of Scaling: Why LLMs Cannot Achieve True Intelligence
LeCun’s core technical argument hinges on the fundamental limitations of text-based intelligence. Since the breakout success of generative AI, LLMs have become virtually synonymous with artificial intelligence in the public imagination. While conceding their undeniable utility for tasks involving text manipulation, coding, and research—demonstrating that they are certainly not overhyped in terms of practical value—he dismisses the "illusion, or delusion" that simply scaling these models up will inevitably lead to human-level general intelligence.
The critical hurdle is the Moravec Paradox, first articulated in 1988: what is easy for humans (like physical perception, common sense, and navigation) is extraordinarily difficult for computers, and what is hard for humans (complex arithmetic, symbolic logic) is trivial for computers. LLMs operate almost exclusively within the discrete, symbolic realm of text, which is an extremely high-level abstraction of reality. They are masters of statistical pattern matching within language but lack a true, predictive model of the continuous, messy, causal physical world.
This absence of a world model means LLMs cannot genuinely reason, plan, or predict the consequences of their actions in physical space. This limitation is the technical bottleneck preventing the emergence of a truly useful domestic robot, one as agile and intuitive as a house cat, or the realization of Level 5 autonomous driving. To achieve human-level intelligence, LeCun insists, conceptual breakthroughs must occur—breakthroughs that move beyond mere linguistic manipulation and into embodied understanding.
World Models and the Power of Predictive Abstraction
The alternative paradigm LeCun and AMI are pursuing is centered on World Models and the specific architecture known as Joint Embedding Predictive Architecture (JEPA). LeCun developed JEPA during his tenure at Meta, and it represents a radical departure from the prevailing philosophy of generative AI.
Generative models attempt to predict every detail of the future, often at the pixel or token level. In the real world, which is inherently chaotic and unpredictable (imagine trying to predict the exact path of every smoke particle from a chimney), this approach is doomed to failure. JEPA circumvents this issue by not attempting to be generative. Instead, it focuses on learning an abstract representation of the world.
The key innovation of JEPA is to predict the future not in the raw, high-dimensional data space (like video pixels), but within this learned, low-dimensional abstract space. By making predictions in this abstract space, the system ignores the non-essential, unpredictable details while learning the robust, underlying rules of cause and effect—the foundation of common sense. This is analogous to how a human baby learns about physical laws, such as gravity, through observation and interaction, rather than through linguistic instruction or pixel-perfect future simulation.
AMI’s systems will train on a vastly richer dataset than just text, incorporating video, audio, lidar, and diverse sensor data from robotic systems and industrial processes. This multimodal approach is essential for building a holistic understanding of physics and causality.
Unlocking the Age of Reliable Agentic Systems
The applications for robust world models extend far beyond consumer chatbots. LeCun envisions their immediate application in highly complex industrial environments. Consider a modern steel mill, a jet engine, or a sophisticated chemical factory, where thousands of sensors generate continuous, heterogeneous data. Currently, no technique exists to build a complete, holistic predictive model of these systems. A world model, trained on this sensor data, could accurately forecast system behavior, predict maintenance needs, and optimize complex processes in ways current predictive analytics cannot.
Crucially, world models are the necessary prerequisite for reliable agentic systems. An agent—an AI designed to take action in the physical world, whether as a robot or a digital assistant—cannot operate safely or reliably without the ability to model the environment and accurately predict the consequences of its own actions. Without this predictive capability, the agent will inevitably make mistakes that are often catastrophic in physical domains.
LeCun is particularly skeptical of the current hype surrounding humanoid robotics, often spearheaded by Chinese firms. He argues that current robotic demonstrations, such as dancing or martial arts routines, rely on massive amounts of pre-planned motion capture or tele-operation data for narrow, specific tasks. These systems lack the ability to generalize because they lack common sense derived from an internal world model. The reason a 17-year-old can learn to drive in 20 hours is that they already possess a lifetime of intuitive knowledge about physics; they don’t need millions of driving hours to learn gravity or object permanence. World models provide the missing piece required to imbue robots with this generalized understanding.
From Meta to AMI: Strategy and Leadership
LeCun’s transition from Meta signals a desire to return to pure, foundational research unencumbered by the pressures of productization within a massive corporate structure. While acknowledging the immense success of FAIR in research, he points to Meta’s subsequent struggles in converting that research into practical products. He suggests strategic missteps by leadership, including the decision to disband the robotics group at FAIR, as examples of product-driven choices that hindered long-term foundational work.
LeCun, who describes himself as a visionary and a scientist rather than a manager, has structured AMI to maximize his research impact. He will serve as Executive Chairman, focusing on scientific direction and technology advancement. The company will be led by CEO Alex LeBrun, a serial entrepreneur with a proven track record, including successful exits to Microsoft and Meta, and the creation of the healthcare AI company Nabla. LeBrun’s expertise in building successful companies will allow LeCun to maintain his academic position at NYU and focus his efforts on groundbreaking research projects, fulfilling his mission to "make science and technology progress."
AMI’s global ambition is reflected in its planned expansion, with offices slated for North America (likely New York, intentionally avoiding the "monoculture" of Silicon Valley) and potential expansion into Asia (Singapore being a likely candidate). The company reports no difficulty in attracting top-tier talent, successfully recruiting researchers from major labs like OpenAI, Google DeepMind, and xAI. This reflects a clear gravitational pull toward a mission that many high-level researchers believe represents the true path to next-generation AI, even if it runs contrary to current market trends. The potential addition of Saining Xie, a brilliant researcher from NYU and Google DeepMind, as a Chief Scientist would further cement AMI’s status as a formidable research hub.
Reclaiming the Research Mandate for Academia
The ongoing AI arms race, characterized by massive compute requirements and data access challenges, has led to a perception that foundational AI research is increasingly restricted to the industrial sphere. LeCun challenges this narrative, arguing that while the development of ever-larger LLMs requires industry resources, LLMs themselves are now a matter of technology development and engineering, not foundational research.
He draws a parallel to the state of speech recognition technology in the early 2010s: once the conceptual breakthrough was made, the subsequent progress became an industrial optimization problem. He advises academics to abandon the pursuit of incremental LLM improvements, where they cannot possibly compete with the resources of Google or OpenAI.
Instead, LeCun insists that the role of academia is to focus on long-term objectives that surpass current system capabilities—precisely the domain of World Models. He reminds the industry that key architectural breakthroughs, such as the attention mechanism that underlies the transformer architecture, originated not in corporate labs but in university settings. By focusing on the "next big thing" rather than refining the last one, and by leveraging non-industrial open platforms, academia remains critical.
AMI’s venture is thus more than a startup; it is a declaration that the road to general artificial intelligence requires a fundamental conceptual leap that current generative models cannot deliver. By combining deep theoretical research (World Models/JEPA) with a principled, open-source approach and a global geopolitical strategy, LeCun is attempting to carve out a new, foundational center of gravity for AI research that promises to unlock true intelligence in the physical world. The market will soon learn whether this contrarian technical bet, focused on common sense and physical causality, can deliver the agentic reliability that the LLM era has yet to achieve.