Yann LeCun’s billion-dollar bet against everything AI believes ( spoiler: he is right )

1.AMI Labs: a French startup designed to replace the LLM paradigm** 2. Why LeCun thinks LLMs are fundamentally broken3. The mainstream paradigm AMI is betting against 4. How LeCun earned the credibility to make this bet 5. Is a billion-dollar seed round justified for an unproven paradigm? 6. The deeper question AMI forces us to ask

The most contrarian figure in artificial intelligence just raised $1.03 billion** to prove the entire industry wrong. Yann LeCun — Turing Award winner, co-godfather of deep learning, and one of the most-cited computer scientists alive — left Meta after 12 years in November 2025 and founded Advanced Machine Intelligence Labs (AMI), a startup built on the thesis that Large Language Models are a dead end for achieving real AI. On March 10, 2026, AMI announced the largest seed round ever for a European company, at a $3.5 billion pre-money valuation, with zero revenue, zero product, and a research timeline measured in years. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with strategic backing from Nvidia, Samsung, Toyota Ventures, and Temasek. Whether this represents visionary genius or peak AI bubble will define the next era of the field.

1. AMI Labs: a French startup designed to replace the LLM paradigm

AMI Labs — pronounced “a-mee,” French for “friend” — carries the tagline “Real World. Real Intelligence.” Its founding premise is blunt: “Real intelligence does not start in language. It starts in the world.” Headquartered in Paris with offices in New York, Montreal, and Singapore, the company was formally incorporated around mid-December 2025, making it barely three months old at the time of its funding announcement.

LeCun serves as Executive Chairman. Day-to-day operations are led by CEO Alexandre LeBrun, a serial entrepreneur who previously founded Wit.ai (acquired by Facebook) and Nabla, a healthcare AI company. The founding team reads like a who’s-who of elite AI research: Saining Xie (Chief Science Officer, formerly Google DeepMind), Laurent Solly (COO, formerly Meta VP for Europe), Pascale Fung (Chief Research & Innovation Officer, formerly senior director of AI research at Meta), and Michael Rabbat (VP of World Models, formerly research science director at Meta). At announcement, AMI had roughly 12 employees and was aggressively hiring across all four hubs.

The company’s technical mission centers on building world models — AI systems that understand the physical world, maintain persistent memory, reason and plan, and remain controllable. Rather than predicting the next word in a sequence (as GPT, Claude, and Gemini do), AMI’s systems will learn abstract representations from video, audio, lidar, and other sensor data, then make predictions in a compressed “representation space” that ignores unpredictable surface-level details. The first planned product is “AMI Video,” and the first disclosed partner is Nabla, which will get early access to develop AI tools for clinical environments. LeBrun has been explicit that AMI “needs at least one year of research before launching its first practical application.”

The investor roster reflects both strategic and financial conviction. Beyond the five co-leads, the round drew capital from Bpifrance (the French public investment bank), the Dassault industrial group, and the Mulliez family office — signaling that France views AMI as a potential national champion. Individual investors include Tim Berners-Lee, Eric Schmidt, Mark Cuban, Jim Breyer, and Xavier Niel. LeBrun originally sought €500 million; the round was oversubscribed to €890 million ($1.03 billion), suggesting investors were competing for allocation.

2. Why LeCun thinks LLMs are fundamentally broken

LeCun’s critique of autoregressive language models is not new — it is a decade-long thesis that has grown sharper as LLMs have grown larger. His core argument: predicting the next token in a text sequence is a fundamentally inadequate mechanism for understanding reality. LLMs operate in what he calls the “discrete, low-dimensional space” of language. They cannot represent the continuous, high-dimensional spaces that characterize nearly everything about the physical world — physics, causality, spatial relationships, temporal dynamics.

In a November 2025 public conversation titled “Do LLMs Understand?”, LeCun stated: “LLMs perform well at the language level, but they don’t understand the world. They lack common sense and causal relationships and are just a stack of a large number of statistical correlations.” In his January 2026 MIT Technology Review interview, he was even more direct: “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.”

His technical objections are specific. First, autoregressive generation means errors compound exponentially with prediction length — the system cannot plan ahead or backtrack. Second, chain-of-thought reasoning in models like OpenAI’s o1 is, in LeCun’s view, merely “statistical sampling of the possible answer space along different trajectories, which does not connote any real understanding” — at best “System 1.1” rather than genuine System 2 deliberative reasoning. Third, LLMs cannot handle the Moravec Paradox: tasks trivial for humans (perception, navigation, physical reasoning) remain impossibly hard for text-trained systems. “This is why we don’t have a domestic robot that is as agile as a house cat, or a truly autonomous car,” he told MIT Technology Review.

LeCun’s alternative vision was laid out in a landmark 62-page position paper published in June 2022: “A Path Towards Autonomous Machine Intelligence.” The paper proposed a six-module cognitive architecture — a Configurator, Perception Module, World Model Module, Cost Module, Actor Module, and Short-term Memory — all differentiable and trainable end-to-end via gradient descent. The centerpiece is the Joint Embedding Predictive Architecture (JEPA), which learns by predicting abstract representations of future states rather than generating every detail of them. Where a generative model tries to predict every pixel of a future video frame (an impossibly complex task in an unpredictable world), JEPA predicts a compressed representation that captures only what matters, using energy-based learning to assign low energy to compatible input-output pairs and high energy to incompatible ones.

The JEPA research program has produced a family of progressively more capable models. I-JEPA (January 2023) demonstrated self-supervised image learning with 5× fewer iterations than comparable methods. V-JEPA (February 2024) extended this to video, learning to fill in masked portions of video in representation space rather than pixel space. V-JEPA 2 (June 2025) achieved state-of-the-art visual understanding and prediction after training on over one million hours of internet video — and demonstrated zero-shot robot planning, controlling Franka robot arms in novel environments for picking and placing tasks without any task-specific training. VL-JEPA (December 2025) matched or exceeded larger generative vision-language models with only 1.6 billion parameters and 50% fewer trainable parameters. These results provide the empirical foundation for AMI’s bet.

3. The mainstream paradigm AMI is betting against

The dominant approach to AI — the one backed by over $225 billion in private funding in 2025 alone — rests on a simple but powerful insight: scale transformer-based language models with more data, more parameters, and more compute, and capabilities reliably improve. This insight traces back to OpenAI’s 2020 scaling laws paper, which showed that LLM performance improves as a smooth power law across compute, model size, and data. DeepMind’s 2022 Chinchilla paper refined this, demonstrating that model size and training data should scale equally for optimal performance.

The results have been extraordinary. GPT-3’s 175 billion parameters in 2020 showed emergent zero-shot abilities that seemed almost magical. ChatGPT (November 2022) became the fastest-growing consumer application in history. GPT-4 (March 2023) demonstrated multimodal reasoning that passed bar exams and medical licensing tests. By early 2026, OpenAI claims 700 million weekly active users and has raised over $150 billion. Anthropic’s Claude, Google’s Gemini, and Meta’s own Llama models have each pushed the frontier further.

The paradigm evolved significantly in 2024–2025 with the rise of reasoning models. OpenAI’s o1 (September 2024) introduced reinforcement learning-trained “thinking” — models that generate internal chains of reasoning before answering. The o3 model (December 2024) scored 87.7% on GPQA Diamond and 88% on ARC-AGI, benchmarks previously thought years from being solved. DeepSeek-R1 (January 2025), an open-source reasoning model from China, proved that comparable reasoning capabilities could be induced through pure reinforcement learning at a fraction of the cost — reportedly trained for just $6 million versus $100 million+ for GPT-4. This represented a new “axis of scaling” — test-time compute, where models improve by thinking longer rather than being trained longer.

But cracks have appeared. GPT-5, launched August 7, 2025, was widely perceived as underwhelming — evolutionary rather than revolutionary, still hallucinating, still struggling with basic math and letter-counting. Users signed a petition demanding GPT-4o’s return; Sam Altman relented. Ilya Sutskever declared at NeurIPS 2024 that “pretraining as we know it will end” and that “we have achieved peak data.” Erik Hoel pointed out that “no model has made significant progress on MMLU since early 2023.” The power-law relationship between compute and performance means exponentially more spending for linear improvement — a dynamic that, on a non-logarithmic scale, looks like diminishing returns approaching a ceiling. Bain & Company identified an $800 billion annual revenue gap between AI infrastructure investment and the earnings needed to justify it.

The counter-argument remains potent: reasoning models, agentic systems, efficiency gains, and multimodal extensions open new scaling axes that may circumvent the pre-training wall. But LeCun’s position is that these are patches on a fundamentally limited architecture, not solutions. The debate has never been more consequential.

4. How LeCun earned the credibility to make this bet

To understand why investors handed $1 billion to a three-month-old company with no product, you need to understand LeCun’s place in AI history. The field has endured two devastating “winters” — the first in the 1970s, triggered by the 1973 Lighthill Report’s scathing assessment that AI had achieved “utter failure” relative to its promises, and the second in the late 1980s when brittle expert systems collapsed and the LISP machine market imploded. Neural networks were particularly marginalized: Minsky and Papert’s 1969 book “Perceptrons” had demonstrated severe limitations of single-layer networks, discouraging research for over a decade.

LeCun was one of the researchers who kept the faith. After completing his PhD at the Sorbonne in 1987 and a postdoc with Geoffrey Hinton in Toronto, he joined AT&T Bell Labs, where in 1989 he published the first convolutional neural network trained end-to-end with backpropagation — reading handwritten zip codes for the US Postal Service. His 1998 LeNet-5 architecture achieved 99.05% accuracy on handwritten digits and established the blueprint (convolutional layers, pooling layers, fully connected layers) that every modern CNN follows. But for years, this work was largely ignored by a field still in its symbolic AI phase.

The vindication came in 2012. Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton entered AlexNet into the ImageNet competition, achieving a top-5 error rate of 15.3% versus 26.2% for the runner-up — a 10.8-point margin that stunned the field. This was LeCun’s architectural vision, supercharged by GPUs and big data. The deep learning revolution had begun. Five years later, Vaswani et al. published “Attention Is All You Need” (2017), introducing the transformer architecture that would power GPT, BERT, and every frontier model since. The progression from GPT-1 (117 million parameters, 2018) through GPT-3 (175 billion, 2020) to ChatGPT (2022) transformed AI from an academic pursuit into the defining technology investment of the decade.

In 2018, LeCun, Hinton, and Yoshua Bengio received the ACM Turing Award — the “Nobel Prize of Computing” — for “conceptual and engineering breakthroughs that have made deep neural networks a critical component of computing.” LeCun had founded Meta’s FAIR lab in December 2013, building it into one of the world’s most respected AI research institutions and overseeing the creation of PyTorch (now the dominant deep learning framework, running 5+ trillion daily inferences) and the Llama family of open-weight models. He holds a Silver Professorship at NYU’s Courant Institute, has been cited over 455,000 times on Google Scholar, and has received the Queen Elizabeth Prize for Engineering, the Chevalier of the French Legion of Honour, and the Princess of Asturias Award.

His departure from Meta was driven by philosophical disagreement. After ChatGPT’s launch, Meta pivoted hard toward LLMs, sidelining LeCun’s world model research. The disbanding of FAIR’s robotics group was, in his words, “a strategic mistake.” The hiring of Alexandr Wang (the 28-year-old former Scale AI CEO) to effectively oversee LeCun’s group was the breaking point. “You don’t tell a researcher what to do,” LeCun told the Financial Times. “You certainly don’t tell a researcher like me what to do.”

5. Is a billion-dollar seed round justified for an unproven paradigm?

The case for AMI’s valuation rests on three pillars: LeCun’s singular credibility, the growing evidence that LLMs are plateauing, and the competitive dynamics of frontier AI research. The JEPA family of models has produced genuine results — V-JEPA 2’s zero-shot robot control is a concrete demonstration, not a slide deck. The investor roster (Nvidia, Samsung, Toyota, Temasek, Bezos Expeditions) suggests strategic buyers who see world models as infrastructure for physical-world AI applications where LLMs fundamentally cannot operate: manufacturing, robotics, autonomous driving, healthcare. And the European sovereignty angle is real — France’s Bpifrance and the Dassault group are investing partly because AMI represents a credible frontier AI lab that is “neither Chinese nor American.”

The case against is equally compelling. JEPA has not been demonstrated at frontier scale in any production system. LeCun himself acknowledges that “major conceptual breakthroughs” must still occur. The company is asking investors to fund what CEO LeBrun explicitly describes as a fundamental research program — “not your typical applied AI startup that can release a product in three months.” The first year will be pure R&D. If world models don’t produce breakthrough demonstrations within 18–24 months, pressure will mount. And the competitive landscape is formidable: Google DeepMind is developing world models (Genie family) with vastly more resources, Fei-Fei Li’s World Labs has $1.23 billion and has already launched a product called Marble, and Meta itself continues the research LeCun started.

The cautionary tale is Inflection AI, which raised $1.52 billion at a $4 billion valuation, only to collapse in March 2024 when Microsoft hired its co-founders and most of its 70-person team. Capital alone does not guarantee technical breakthroughs or product-market fit. But the precedent for researcher-led mega-rounds is now established: Ilya Sutskever’s Safe Superintelligence (SSI) has raised $3 billion at a $32 billion valuation with roughly 20 employees and no product. Mira Murati’s Thinking Machines Lab raised a $2 billion seed — the largest ever — in July 2025. David Silver’s Ineffable Intelligence raised $1 billion in a seed round in February 2026. In this context, AMI’s round is large but not extreme.

What would justify the valuation? In the near term (2026), AMI needs published research breakthroughs demonstrating JEPA at meaningful scale, plus open-source releases that build a developer ecosystem analogous to Meta’s credibility-building Llama releases. By 2027–2028, enterprise pilot deployments and initial revenue through technology licensing become critical. The long-term thesis requires AMI’s world models to become infrastructure for physical-world AI — the foundation layer that robotics companies, manufacturers, automakers, and healthcare systems build upon. If that happens, a $3.5 billion seed valuation will look like a bargain. If JEPA remains elegant theory without scaled deployment, it will look like the most expensive research grant in history.

6. The deeper question AMI forces us to ask

AMI Labs is not merely a startup. It is a referendum on whether the current AI paradigm has a ceiling. The mainstream bet — represented by $225 billion in 2025 investment — says that transformer-based autoregressive models, enhanced with reasoning via reinforcement learning, will continue to improve and will eventually solve physical-world understanding through multimodal scaling. LeCun’s bet says this is a fundamental category error: you cannot build real intelligence by predicting text, no matter how much text you train on, because intelligence emerges from embodied interaction with the physical world.

Both positions have historical precedent. The symbolic AI community spent decades insisting that neural networks could not work — and was proven catastrophically wrong by AlexNet in 2012. But the early neural network community also spent decades in the wilderness, vindicated only after computational power caught up with theoretical insights. LeCun was on the right side of that earlier revolution. Whether he is right again — or whether this time the contrarian position is simply wrong — will determine not just AMI’s fate, but the trajectory of artificial intelligence itself. The $1.03 billion buys him the chance to find out.

Fun Fact: Of course he is right. Nature had Mio of years to experiment with (the human) brain. If “old school” was right, we would have a single type of cells and structure between our ears, but we don´t — Guess why?