Yann LeCun, the Turing Award winner who left Meta in late 2025 after roughly 12 years as its chief AI scientist, is running one of the best-capitalized contrarian bets in artificial intelligence, and in late May 2026 it received its sharpest technical test yet. His Paris-headquartered startup, AMI Labs, is building "world models" — AI that learns how physical reality behaves rather than predicting the next word in a sequence — on the argument that the large language models (LLMs) now dominating the field cannot, on their own, reach human-level intelligence. Two preprints from his research circle, posted within days of each other and covered together on May 31, now define exactly what his architecture can prove and how far today's world models still fall short.
For anyone tracking where frontier AI capital is flowing, the stakes are concrete. AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation in a seed round announced March 9, 2026 — reported as the largest seed round in European startup history — and as of early June it still has no product, a roughly dozen-person team, and a research agenda measured in years. The late-May papers are the first rigorous public signal of whether the science behind that valuation is tracking toward its goal.
AMI Labs Builds World Models to Predict Reality, Not Text
AMI stands for Advanced Machine Intelligence, and the name doubles as "ami," the French word for friend. The company confirmed its direction in late January 2026: it is developing AI systems intended to understand the real world, hold persistent memory, and reason and plan. A world model, in LeCun's framing, learns an internal representation of how an environment works so a machine can predict the consequences of its actions before taking them.
LeCun is AMI Labs' executive chairman, not its CEO. Day-to-day operations belong to Alexandre LeBrun, a serial entrepreneur who co-founded the health-AI startup Nabla and previously worked under LeCun at Meta's Fundamental AI Research (FAIR) lab. The company is headquartered in Paris, with offices planned for New York — where LeCun keeps his NYU professorship — as well as Montreal and Singapore.
How JEPA Works: Prediction in Abstract Space Instead of Pixels
The technical foundation is the Joint Embedding Predictive Architecture, or JEPA, a framework LeCun proposed in 2022, before the LLM boom. In an exclusive interview with MIT Technology Review, he described the mechanism directly: "The world is unpredictable. If you try to build a generative model that predicts every detail of the future, it will fail. JEPA is not generative AI. It is a system that learns to represent videos really well."
The engineering distinction is the source of both the bet's promise and its risk. A generative model tries to reconstruct every pixel or token, which forces it to spend capacity modeling unpredictable surface detail. JEPA instead encodes each input into an abstract representation and predicts in that latent space, discarding what it cannot predict and keeping high-level structure. That design choice — predict representations, not raw outputs — is what lets a JEPA system, in theory, learn the underlying rules of an environment "like a baby learning about gravity," as LeCun put it, and use them to plan. It is also why the architecture's guarantees depend heavily on how it is trained, a vulnerability the new research makes explicit.
👉 Read more:
Yann LeCun's World Model Earns a Formal Proof: Benchmark Finds Current Models Brittle
LeCun Argues LLMs Cannot Plan or Predict Consequences
LeCun's case against LLMs is an argued thesis, not an industry consensus, and he states it without hedging. "People have had this illusion, or delusion, that it is a matter of time until we can scale them up to having human-level intelligence, and that is simply false," he told MIT Technology Review. He invokes the Moravec Paradox — perception and navigation are easy for humans but hard for machines — and argues that LLMs "are limited to the discrete world of text," cannot truly reason or plan, and "can't predict the consequences of their actions."
He has gone further elsewhere, calling next-element generation a dead end for video and robotics and predicting the robotics field will recognize the need to move away from LLM-centric architectures by early 2027. LeBrun reached a parallel conclusion from medicine, where LLM hallucinations can carry life-threatening consequences; Nabla is AMI Labs' first disclosed partner. None of these are settled facts about AI's future — they are the wagers the $1.03 billion is funding.
What Does the LeJEPA Identifiability Proof Actually Guarantee?
The first late-May paper, "When Does LeJEPA Learn a World Model?", was submitted on May 25 by David Klindt of Cold Spring Harbor Laboratory, LeCun, and Randall Balestriero of Brown University. It proves that the LeJEPA architecture can achieve "linear identifiability" — recovering the true hidden variables behind raw observations, such as an object's position and velocity, up to a linear transformation — rather than merely latching onto whatever statistical shortcut was cheapest to find.
The guarantee is conditional, and the conditions are the news. It holds when the latent variables follow a Gaussian distribution and evolve under stationary, additive-noise dynamics, and when training data approximates broad, roughly uniform exploration of the state space. The paper's signature result takes an "if and only if" form: among that class of worlds, the Gaussian is the unique latent distribution for which the guarantee holds. The proofs were formalized in the Lean 4 proof assistant, giving them machine-checkable rigor beyond standard paper convention. For engineers, the practical edge is blunt: goal-directed training data, the kind most robotic pipelines rely on, can quietly push observations into a regime where the guarantee no longer applies.
AI World Model Benchmark Shows Current Systems Collapse Under Small Shifts
If the theory paper maps the destination, the second measures the distance to it. The stable-worldmodel benchmark, posted May 20 by a team led by Lucas Maes of Mila and Université de Montréal that also includes LeCun and Balestriero, is an open-source platform built because the field had fragmented to the point of unreliability — one common planning algorithm had been independently reimplemented in at least five recent papers.
The verdict on current systems is direct: they remain brittle. On a standard task that requires pushing an object into a target position, one tested model succeeded about 50 percent of the time under clean conditions; success fell to roughly 12 percent when the agent's color changed and to about 6 percent when the background color shifted, with added visual distractors producing a collapse across every baseline. A subtler finding cuts deeper — prediction accuracy proved a poor proxy for planning success, meaning a model can forecast the next frame correctly while having latched onto a background color rather than the task's geometry. All figures come from the preprint and have not yet been independently replicated.
World Model Race Includes Fei-Fei Li's World Labs and Google DeepMind
AMI Labs is not alone in betting that the next leap requires modeling the physical world. World Labs, founded by Stanford's Fei-Fei Li, became a unicorn after leaving stealth and shipped Marble, a system that generates physically coherent 3D worlds; it has been reported in talks at a $5 billion valuation. Google DeepMind's Genie line of generative interactive environments marks another front. The approaches differ — World Labs centers on spatial intelligence and 3D generation, AMI Labs on JEPA-based prediction and planning — but they share the conviction that language alone is not enough.
The money behind AMI Labs reflects how much capital that conviction now commands. The March round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with individual backers including Tim and Rosemary Berners-Lee, Jim Breyer, Mark Cuban, Mark Leslie, Xavier Niel, and Eric Schmidt, and corporate investors including NVIDIA, Samsung, Sea, Temasek, and Toyota Ventures. The startup had reportedly sought roughly €500 million and ended up raising about €890 million, the figure several European outlets cite for the same round.
👉 Read more:
Fei-Fei Li's ESI-Bench Catches Frontier AI Failing 3D Space: Seeing and Acting Diverge
What the Late-May Papers Mean for the $1.03 Billion Bet
Neither paper proves AMI Labs can ship deployable world models on its stated schedule. The identifiability result is formal but conditional; the benchmark result is empirical but limited to simulated environments and existing baselines. What the two do together is sharpen the target: the theorem identifies the data-collection conditions under which faithful learning becomes mathematically attainable, and the benchmark specifies the robustness failures that must be solved before attainability becomes reliability. LeBrun has been candid that the horizon is long, telling TechCrunch the work "starts with fundamental research" and is "not your typical applied AI startup that can release a product in three months." The company also plans to publish papers and open-source much of its code, consistent with LeCun's long-standing advocacy for open research.
Frequently Asked Questions
What is a world model in AI?
A world model is an AI system that learns an internal representation of how an environment behaves so it can predict the consequences of actions and plan, rather than only predicting the next word or pixel. Yann LeCun argues this capability is necessary for reasoning, robotics, and reliable autonomous systems. His startup AMI Labs is building world models on the JEPA architecture.
Why does Yann LeCun think large language models are not enough?
LeCun argues that LLMs are confined to the discrete world of text, cannot truly reason or plan, and cannot predict the physical consequences of actions because they lack a model of how the world works. He calls the belief that scaling LLMs will reach human-level intelligence false. This is his argued position, not an industry consensus.
How much did AMI Labs raise and who invested?
AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation in a round announced March 9, 2026, reported as Europe's largest seed round. Backers include Bezos Expeditions, NVIDIA, Samsung, Temasek, Toyota Ventures, and individuals such as Eric Schmidt, Mark Cuban, Xavier Niel, and Tim and Rosemary Berners-Lee.
What did the May 2026 world model papers find?
One preprint proved that LeCun's LeJEPA architecture can recover the true hidden structure of an environment, but only under specific conditions including Gaussian latent variables and broad data exploration. A second, the stable-worldmodel benchmark, found that current world models stay brittle, with success rates collapsing when colors or backgrounds change. Both are preprints that have not yet been peer-reviewed or independently replicated.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
