Yann LeCun left Meta in November 2025 after twelve years serving as chief scientist. In March 2026, his new startup AMI Labs raised $1.03 billion, becoming the vehicle for the largest seed funding in the history of European technology startups. The wager behind it is radical: not the language model but the world model will become the foundation of genuine artificial intelligence.
What distinguishes a world model from a language model
Language models like GPT or Claude learn from text. They can describe how a ball falls to the ground, but they have never observed it and cannot actually predict the trajectory. This is the core of LeCun's criticism: a system learning exclusively from language does not understand physical reality, it can only describe it.
World models learn instead from observation. AMI Labs trains its V-JEPA 2 model, a system with 1.2 billion parameters, on over one million hours of video from the internet. The system does not predict pixels but rather abstract representations of future states. On physical tasks, such as deliberately grasping a cup, V-JEPA 2 achieves an 80 percent success rate according to Meta. Conventional language models fail at such tasks structurally because they lack the mental model of the physical world.
The technical foundation is LeCun's Joint Embedding Predictive Architecture (JEPA), which he proposed in 2022 in a widely noted paper. It differs from generative AI in that it does not generate pixel output or text output but rather builds an inner representation of future world states.
The industry is digging the same trench
LeCun's criticism of the industry is sharp. In a Handelsblatt interview, he phrased it thus: "The entire industry is digging the same trench." He regards language models as a dead end on the path to human-like intelligence because they cannot structurally learn fundamental capabilities like causal reasoning and reliable planning.
With this, he positions himself against the majority view in the industry. Anthropic CEO Dario Amodei predicted in early 2026 that LLM-based systems could already simulate the work of tens of thousands of programmers in 2026 and positions his company as a leader precisely in that LLM development which LeCun deems misguided. OpenAI and Google DeepMind continue to bet on scaling existing architectures and document persistent performance gains. Alexandre LeBrun, CEO of AMI Labs and former Meta employee who co-founded the health platform Nabla, offered a self-critical forecast: "In six months, every company will claim to be building world models to raise capital." He sees the hype as inevitable but also as a risk to his own research's substance.
Mathematically proven: When JEPA works
In May 2026, a preprint appeared on arXiv titled "When Does LeJEPA Learn a World Model?" The paper answers for the first time with mathematical precision under which conditions JEPA actually learns a faithful representation of reality. The result, according to arxiv.org/abs/2605.26379: LeJEPA with Gaussian regularization can linearly identify latent variables behind observations, provided these variables are Gaussian distributed and evolve according to stationary, additive-noise transitions. For this class of worlds, Gaussian distribution is the only one guaranteeing correctness.
Unusual for AI research: the proofs were formalized in the proof assistant Lean 4, making them machine-checkable. This lends the results a rigor that experimental benchmark studies cannot achieve. Limitation: the guarantee holds under specific distribution assumptions. Whether real video sequences sufficiently meet these conditions remains an open empirical question.
Nvidia and Eric Schmidt hedge their bets
The investor field behind AMI Labs is revealing. Beyond financial investors like Cathay Innovation, Greycroft, and HV Capital, and Bezos Expeditions, industrial enterprises including Nvidia and Toyota Ventures as well as individual investors like Eric Schmidt and Mark Cuban participated. Particularly noteworthy: Nvidia, which profits from LLM infrastructure more than nearly any other company, simultaneously bets on what might become the next AI paradigm. This suggests that even the prime beneficiaries of the current scaling wave view LeCun's approach at least as insurance.
AMI Labs operates with headquarters in Paris and additional offices in New York, Montreal, and Singapore. Initial application fields are medical diagnostics, where hallucinations in language models are particularly problematic, and industrial robotics. Meta FAIR, LeCun's former employer, remains a research cooperation partner for JEPA development according to company statements but does not invest in AMI Labs.
JEPA robots versus LLM assistant: decided by 2028
AMI Labs plans to produce first industrial prototypes in 2027. The real test comes after: when world models confront actual production lines and clinical diagnostic workflows. The counterproof makes the market: if LLMs continue to achieve double-digit performance leaps, LeCun's paradigm shift loses urgency. If they continue to hit structural limits on physical and causal-logical tasks, his approach gains credibility beyond research literature.
Robots working independently in unknown physical environments are considered the hardest test for world models in robotics research. If AMI Labs maintains its timeline and shows first industrial prototypes in 2027, by 2028 it should become clear whether JEPA passes this real-world test or whether the math paper's guarantee breaks on the complexity of real video sequences.