Paris, July 2026 — Yann LeCun’s new venture, Advanced Machine Intelligence Labs (AMI Labs), is pursuing a form of artificial intelligence designed to understand the physical world more deeply than current large language models (LLMs) like ChatGPT, Claude, or Gemini. Speaking on the sidelines of VivaTech, LeCun, who spent a decade as Meta’s chief AI scientist before leaving in 2025 to launch AMI Labs, argues that LLMs excel at well-defined tasks but fall short when faced with real-world unpredictability. “They’re not a path towards human level intelligence, or even animal-like intelligence, because they cannot deal with real world data,” he told BBC Business Editor Ben Morris, outlining the need for a more flexible AI paradigm.
AMI Labs has attracted significant funding: earlier this year the startup announced a seed round that topped $1 billion (£760 million), with notable backers including Nvidia and the Bezos family investment fund. LeCun emphasizes that the money signals investor appetite for alternatives to traditional LLMs, as the company develops a model architecture it calls Joint Embedding Predictive Architecture (JEPA). The aim is to build abstractions of the world that let the AI predict outcomes of actions while filtering out irrelevant information. A simple demonstration contrasts how a pen’s fall is unpredictable for a system that does not model physical causality with how JEPA would identify relevant physical cues to guide reasoning.
LICENSING THE REAL WORLD, NOT JUST TEXTUAL PATTERNS Railings of billions have already gone into humanoid robotics, but teaching machines to perform everyday chores safely remains costly and technically challenging. LeCun contends that LLMs “are largely hopeless for robotics,” arguing that scaling up language models will not deliver the general intelligence needed for non-text tasks. The discussion around JEPA centers on creating structured representations of the physical world that support planning and outcome prediction rather than mere pattern mimicking. Oxford AI expert Ingmar Posner, who directs the Applied AI Lab and is an Amazon Scholar, is among those who see value in mechanistic world models capable of answering questions like What matters? What causes what? What would happen if I did something else?
World Models: From Concept to Concrete Steps Posner and his team, about 10 researchers, have spent four years developing an AI approach that fits into the broader World Models family—a lineage inspired by a 2018 paper from David Ha and Jurgen Schmidhuber and subsequent work at Google DeepMind and elsewhere. The idea is to simulate and reason about future scenarios inside the model to guide decisions, rather than relying solely on learned textual patterns. A notable demonstration from the field involved a Dreamer variant navigating Minecraft by imagining future steps to collect diamonds, illustrating how world models can plan through mental simulation. LeCun and colleagues envision a future where AI systems use mechanistic world models to organize knowledge so it can be recalled, combined and modified when needed, enabling more reliable behavior in varied environments.
Beyond the Lab: A Path Toward Industrial Adoption Industry watchers note that the future of robotics may hinge on these flexible architectures that integrate perception with planning and causal reasoning. Experts such as Posner argue that the next decade will be defined by systems capable of explaining what matters, what causes what, and what would happen under different actions. AMI Labs intends to spend the remainder of this year refining JEPA, with early industrial deployments planned for the following year if tests prove successful. LeCun has said that long-range goals include general intelligence systems that require minimal retraining to tackle a broad array of real-world tasks, while acknowledging that human oversight remains essential for framing questions and guiding deployment.
The broader AI ecosystem continues to explore new paradigms beyond LLMs, including models from Google DeepMind and companies like Wayve that combine planning with perception. Fei-Fei Li’s World Labs, founded in San Francisco in 2023, is another example of researchers pursuing next-generation AI architectures. As AMI Labs progresses, investors and industry observers will be watching whether JEPA-style approaches can deliver deployable capabilities that translate into tangible improvements in robotics, manufacturing, and service sectors.
The human–AI partnership remains central. LeCun suggests future human–AI collaboration will resemble a captain of industry working with a staff of smarter assistants, where humans set the questions and the AI provides the reasoning tools to answer them. He emphasizes that even with more capable AI, humans will play a crucial role in guiding what questions to ask and how to apply smarter systems to real-world tasks.
