This philosophical stance likely influenced his recent departure from Meta, where he served as the chief scientist of FAIR (Fundamental AI Research), the influential lab he himself founded. Meta’s efforts with its open-source AI model, Llama, have reportedly struggled to gain significant traction, coinciding with internal organizational shifts, including the notable acquisition of Scale AI. In an exclusive online interview with MIT Technology Review from his Paris residence, LeCun articulated his vision for his new company, his transition from Meta, his perspectives on the future of AI, and his critique of the industry’s current trajectory.

The interview, edited for conciseness and clarity, delved into the core tenets of his new endeavor. LeCun announced the formation of Advanced Machine Intelligence (AMI), a global company with its headquarters in Paris. He noted the strategic significance of this location, highlighting Europe’s rich talent pool and the increasing demand for a credible, non-US/China-centric AI frontier company. "You pronounce it ‘ami’ – it means ‘friend’ in French," LeCun explained, expressing excitement about fostering European talent and addressing sovereignty concerns prevalent among governments and industries worldwide.

LeCun elaborated on his motivation for pursuing this "third path," emphasizing the geopolitical and control imperatives surrounding AI. He posited that AI is poised to become a pervasive platform, a characteristic that historically leads to open-source adoption. He expressed concern that the American industry, driven by increasing competition, is opting for secrecy, a strategic misstep in his view. He pointed to the evolution of OpenAI from an open to a closed entity and Anthropic’s consistently closed model, while noting Google’s partial openness and Meta’s uncertain direction.

A significant concern for LeCun is China’s embrace of an open approach, leading to the dominance of Chinese open-source AI platforms. This, he argued, has led to academia and startups outside the US increasingly adopting Chinese models. While acknowledging the quality of Chinese AI and the prowess of its engineers, LeCun painted a disquieting picture of a future where our information consumption is mediated by AI. He contrasted proprietary, US-centric English-speaking models with open-source Chinese models that may require fine-tuning to address sensitive topics like the 1989 Tiananmen Square incident. He stressed the necessity of AI systems being adaptable and capable of producing diverse assistance, reflecting a wide array of linguistic capabilities, value systems, political leanings, and interests – akin to the diversity found in the press.

Venture capitalists, he reported, have responded positively to his open-source advocacy, recognizing its importance for smaller startups that cannot afford to train their own models and find proprietary models strategically risky.

Reflecting on his departure from Meta, LeCun affirmed the research success of FAIR but acknowledged Meta’s challenges in translating that research into practical applications and products. He attributed these decisions to Mark Zuckerberg’s strategic choices, stating, "I may not have agreed with all of them. For example, the robotics group at FAIR was let go, which I think was a strategic mistake. But I’m not the director of FAIR. People make decisions rationally, and there’s no reason to be upset." He emphatically denied any "bad blood" and even suggested Meta could be a potential client for AMI, as their focus on world models for the physical world is distinct from Meta’s generative AI and LLM work.

LeCun recounted his long-standing work in AI predating the LLM craze. He observed that while LLMs and chatbots have become the public face of AI, newer iterations incorporate elements beyond pure LLMs, such as perception systems and problem-solving code, positioning LLMs as orchestrators within broader systems. He underscored that substantial, behind-the-scenes AI powers much of society, from driving assistance in cars to medical imaging and social media algorithms.

Addressing the perceived overhype of LLMs, LeCun conceded their utility for tasks involving text, research, and coding. However, he refuted the notion that scaling LLMs would inevitably lead to human-level intelligence, deeming it "simply false." He highlighted the Moravec Paradox, where human-level perception and navigation are difficult for computers, contrasting it with LLMs’ limitations within the discrete world of text. "They can’t truly reason or plan, because they lack a model of the world," he stated, explaining their inability to predict consequences, which hinders the development of agile domestic robots or truly autonomous cars. LeCun believes that human-level AI will emerge from new conceptual breakthroughs, not solely from LLMs, and that AMI is focused on this next generation.

His proposed solution lies in "world models" and the "joint embedding predictive architecture" (JEPA), a framework he developed at Meta. LeCun explained JEPA as a system that learns abstract representations of the world and makes predictions in this abstract space, sidestepping unpredictable details. This approach, he argued, is the foundation for common sense and the key to building intelligent systems capable of real-world reasoning and planning, with the most promising work originating from academia.

LeCun also addressed the challenge of data limitations in understanding the physical world. JEPA’s training on videos, audio, and diverse sensor data, including robot arm positions and lidar, is crucial. He mentioned ongoing projects using JEPA to model complex physical and clinical phenomena.

Concrete applications for world models, LeCun envisions, are extensive. He cited complex industrial processes with numerous sensors (jet engines, steel mills, chemical factories) where a holistic model could predict system behavior. Smart glasses that assist users by identifying actions and predicting future movements are another example. Crucially, agentic systems, which act in the world, require a world model to reliably predict the consequences of their actions. This, he asserted, is the unlock for useful domestic robots and Level 5 autonomous driving.

Regarding the recent surge in humanoid robots, particularly from China, LeCun dismissed many demonstrations as pre-planned routines rather than genuine intelligence. He argued that the immense tele-operation data required and poor generalization in changing environments indicate a fundamental gap. He drew a parallel to a 17-year-old learning to drive, emphasizing the pre-existing understanding of the world that enables rapid learning. Achieving genuinely useful domestic robots, he reiterated, hinges on robust world models and planning capabilities.

LeCun countered the notion that foundational AI research is migrating solely to industry due to computational demands. He characterized LLMs as "technology development, not research," and noted that academia’s role has shifted from pioneering such advancements to focusing on long-term, breakthrough objectives. He advised university researchers to steer clear of LLMs and instead pursue novel techniques, asserting that significant breakthroughs will not come from scaling existing models. He highlighted that the most exciting world model research is currently in academia. While acknowledging the need for computational resources, he stressed the importance of academic focus on the "next big thing."

In his multifaceted role as professor, researcher, and public thinker, LeCun is now taking on the role of executive chairman of AMI, with Alex LeBrun as CEO. He will maintain his position at NYU, teaching one class annually and supervising PhD students and postdocs, while frequently traveling to Paris for his lab. He clarified that his hands-on involvement will be in research, not day-to-day management, which he finds unappealing. His primary mission, he stated, is to advance science and technology and inspire others.

LeCun described Alex LeBrun as an ideal partner due to his serial entrepreneurship and success in building three AI companies, two of which were acquired by Microsoft and Facebook, respectively. LeBrun’s experience leading FAIR’s engineering division in Paris and founding the successful healthcare company Nabla, makes him perfectly suited to build the company, allowing LeCun to concentrate on the scientific and technological aspects.

AMI will be a global company with offices in Paris, North America (New York being a preferred location outside the "monoculture" of Silicon Valley), and likely Asia, possibly Singapore. LeCun expressed confidence in attracting talent, noting that many in the AI research community share his belief in world models and are motivated by the technological future AMI is building. He confirmed recruitment of individuals from leading organizations like OpenAI, Google DeepMind, and xAI.

While declining to confirm specific individuals joining, LeCun expressed high regard for Saining Xie, a prominent researcher from NYU and Google DeepMind, whom he has hired twice previously. He indicated that more details about AMI Labs, including financial backing and core team members, would be shared soon, possibly in February.