Meta’s former chief AI scientist, Yann LeCun, helped lay the groundwork for the modern AI boom long before chatbots became an industry obsession. Now, after leaving Meta in late 2025, he is building a multibillion-dollar startup under the premise that the race toward AI superintelligence is starting from the wrong place. Speaking at the VivaTech conference on June 17, LeCun shared more details about why he believes his startup, Advanced Machine Intelligence (AMI Labs), is on a better path toward human-level AI. Earlier this year, the Paris-based company raised $1.03 billion at a $3.5 billion valuation, giving LeCun and his team fresh backing to pursue an alternative to today’s generative AI models developed by tech giants.

“I think there is a requirement for new paradigms to go beyond the limitations of current systems,” LeCun, 65, told Wired’s Steven Levy onstage. Building better language models, as many AI companies are doing, could be one path, LeCun said, but “it’s kind of a slow way.” According to LeCun, AMI Labs is building “world models” that learn from reality, understand consequences, predict what happens next, and choose the best actions based on those predictions. “We have good hopes that this will really be a complete change of the blueprints of intelligent systems,” LeCun said.

LeCun believes much of Silicon Valley has become “LLM-pilled” — or overly convinced that scaling large language models will eventually produce human-like intelligence. The result, he argued, is a “monoculture” of AI systems that are powerful in some areas but still fundamentally limited. While LLMs are already highly capable in areas like coding, they are built around language. Real-world data from cameras, robots, and other sensors is messier and harder to predict, which is why he believes a different approach is necessary. For AMI Labs, the result could be “considerably more intelligent systems that can anticipate the outcome of their own actions, can plan, can reason and perhaps all the way to human intelligence,” he said.

The Godfather of AI and His New Venture

Coming from LeCun, the critiques carry weight. A longtime New York University professor, he is among the trio of researchers often referred to as the “Godfathers of AI.” The others are Geoffrey Hinton, a University of Toronto professor and former Googler, and Yoshua Bengio, a University of Montreal professor who founded Quebec’s AI institute, Mila. In 2018, they collectively won the prestigious Turing Award for breakthroughs that helped push deep learning into the mainstream. LeCun’s contributions include convolutional neural networks (CNNs), which revolutionized image recognition, and early work on backpropagation. His academic pedigree and industry impact make his shift away from the dominant LLM paradigm particularly noteworthy.

In 2013, LeCun joined Facebook to launch what would become Facebook AI Research (FAIR), and later transitioned to chief AI scientist at Meta. He said AMI has its roots as an internal FAIR project and had support from company leaders, including Mark Zuckerberg and CTO Andrew Bosworth. But as Meta’s AI priorities shifted toward catching up with competitors — notably OpenAI and Google in the generative AI race — the environment became less favorable for exploratory research. At that point, it made sense to “just leave the company, try to accelerate the development, and build the things we can build with it,” LeCun said. “And I realized I could also raise enough money to keep this going.”

World Models vs. Language Models

The core idea behind AMI Labs is to build AI systems that understand the physical world, not just text. World models, as envisioned by LeCun, are neural architectures that learn a compressed representation of how the environment works — including physics, causality, and agent interactions. They can simulate possible futures, evaluate which actions lead to desired outcomes, and choose optimal behaviors. This is fundamentally different from LLMs, which predict the next token based on massive text corpora. LLMs can exhibit impressive reasoning skills in certain contexts, but they lack grounding in reality. They cannot reliably predict the consequences of physical actions or understand common-sense dynamics like object permanence or gravity without extensive fine-tuning and external tools.

LeCun’s approach draws on ideas from cognitive science and robotics. “We need models that can learn from video, from touch, from audio — not just from written language,” he explained. AMI Labs plans to train its world models using diverse sensor data, including camera feeds, robot joint states, and even wearable device inputs. This is much more challenging than training LLMs because the data is high-dimensional, noisy, and time-sensitive. But LeCun believes it is essential for achieving robust, generalizable intelligence. “If you want AI that can work in the real world — driving a car, cooking a meal, assisting in surgery — you cannot rely on text alone,” he said.

From Research to Revenue: AMI’s Application Focus

AMI Labs has already outlined several application areas where world models could deliver immediate value. Industrial process control is a prime target: factories and supply chains are full of complex, dynamic systems that require constant optimization. A world model could learn the physics of a production line, predict failures, and recommend adjustments in real time. Similarly, in robotics, world models enable robots to manipulate objects flexibly without explicit programming. AMI’s first partner is Nabla, a health AI company co-founded by CEO Alex LeBrun, who now leads AMI Labs. Healthcare applications could include surgical robots that anticipate tissue deformation, or diagnostic systems that integrate video, audio, and patient history.

The startup’s $1.03 billion seed round — the largest in European startup history — came from investors including Nvidia, Jeff Bezos and Eric Schmidt. This reflects not only faith in LeCun’s vision but also a growing appetite in the investment community for alternatives to the LLM-centric approach. “We’re not going to have a country of geniuses in a data center next year, okay?” LeCun said, tempering expectations. “It’s not going to happen next year.” He estimates it could take five to ten years for world models to reach maturity and outperform LLMs in broad domains.

Open Source and the Politics of AI

At VivaTech, the discussion also broadened to a bigger question: who should control the next generation of AI? If future assistants are built only by “a handful of companies” in Silicon Valley or China, LeCun warned, “culture is in big trouble” and “democracy is in big trouble.” Blocking open-source models in the name of safety, he argued, misses the role AI can play in spreading knowledge. That stance also informs LeCun’s work as chief science adviser to Project Tapestry, an AI Alliance-led effort to build a shared global AI model through a distributed network of contributors. The AI Alliance is a nonprofit group, co-launched in 2023 by Meta and IBM, to support open-source AI development. With Tapestry, the goal is to let countries, companies, universities and others contribute data or computing power without giving up control of that data. As AI begins to mediate what people read, watch, and search for, LeCun argues that people need alternatives to what a few dominant American or Chinese companies decide to build.

LeCun’s comments also come as Meta’s own open-source stance has grown more complicated. Llama 2, released in 2023 for free commercial use, made Meta a rare Big Tech counterweight to more closed AI rivals. But Zuckerberg has since signaled that Meta may not release all of its future “superintelligence” models openly. “We need access to a wide diversity of AI assistants for the same reason we need access to a wide diversity of the press to get multiple sources of information,” LeCun said. “The only way I can see that this can happen is if there is an open, free foundation model, on top of which anybody can build their own specialized assistant for their language or languages, their culture, their value system, their political biases, their centers of interest. And so open source has to exist.”

Historical Context and the Evolution of AI Research

LeCun’s critique of LLM monoculture is not entirely new. Since the rise of transformer-based models in 2017, the field has increasingly converged on training larger language models with more data and compute. This approach has yielded impressive results — GPT-4, Claude, Gemini, and others can pass exams, write code, and generate creative content. But critics argue that LLMs are brittle, prone to hallucination, and incapable of genuine understanding. LeCun has long advocated for a different research direction. In a 2023 position paper, he outlined a “world model” architecture called JEPA (Joint Embedding Predictive Architecture), which forms the theoretical basis for AMI’s work. The JEPA model learns to represent the structure of the world by predicting the representations of future states, rather than predicting raw pixels or text tokens. This makes it more efficient and robust.

The origins of world models trace back to robotics and control theory, but LeCun’s vision integrates ideas from probabilistic graphical models, self-supervised learning, and reinforcement learning. Unlike traditional RL, which requires explicit reward functions, JEPA-based world models can learn from passive observation, similar to how infants learn by watching and interacting with their environment. This aligns with LeCun’s long-standing interest in developmental AI — building systems that learn like children, rather than being spoon-fed vast amounts of labeled data.

The AI community is divided on whether world models can replace LLMs or simply complement them. Some researchers argue that language itself is a powerful abstraction that already captures much of the world’s structure. Others point to successes in embodied AI, such as Google DeepMind’s robotics work, which uses a combination of language and visual models. LeCun acknowledges that world models are not a panacea. “We will need to integrate multiple modalities, including language,” he said. “But language is only one component. If you build your entire AI on language, you’re missing most of what intelligence is.”

Challenges Ahead for AMI Labs

AMI Labs faces formidable challenges. Training world models that can generalize across many environments requires enormous amounts of real-world data, which is expensive and time-consuming to collect. Simulation can help, but simulations often fail to capture the nuances of the real world. The startup is building its own data pipeline, including agreements with industrial partners to access sensor streams. Another challenge is compute: while LLMs benefit from massive parallelism on GPUs, world models often require different architectures and hardware. Nvidia’s investment suggests that the company sees a market for specialized AI accelerators that can handle multimodal, time-series data.

LeCun also needs to attract top talent in a competitive AI job market. AMI’s team already includes researchers from Meta, Google DeepMind, and other leading labs, but the startup must sustain a culture of long-term research while also delivering near-term applications. The seed funding gives AMI a runway of several years, but investors will expect progress toward tangible products.

Despite these hurdles, LeCun remains optimistic. He believes that the current focus on scaling LLMs is a temporary phase, much like the earlier emphasis on symbolic AI or hand-crafted features. “Every decade, AI goes through a change of paradigm,” he said. “The current one is LLMs. The next one will be world models.” Whether AMI Labs will lead that next wave remains to be seen, but its founder’s track record and the financial backing suggest it will be a major player in shaping the future of artificial intelligence.

Source:Observer News