Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchMeta's Chief AI Scientist Yann LeCun just closed the largest seed round in European history for AMI Labs, backed by Bezos, NVIDIA, and Samsung. His bet: JEPA-based world models that learn physical laws will outperform LLMs on the path to AGI.
AI, startups & growth insights. No spam.
TL;DR: Yann LeCun, Meta's Chief AI Scientist and Turing Award laureate, has raised $1.03 billion in seed funding for his new venture, Advanced Machine Intelligence (AMI) Labs — the largest seed round in European history. The company is building world models based on LeCun's Joint Embedding Predictive Architecture (JEPA), which learns physical laws and causal relationships rather than predicting text tokens. Backed by Jeff Bezos, NVIDIA, Samsung, and Temasek, AMI Labs represents the most well-funded direct challenge to the dominant LLM scaling paradigm. If LeCun is right, the path to artificial general intelligence runs through physics, not language.
The AI scientist who has spent the last three years publicly arguing that large language models are a dead end just raised over a billion dollars to build the alternative. That is not a research grant — it is a commercial bet at a scale that forces the entire field to take the contrarian thesis seriously.
Yann LeCun is one of three researchers — alongside Geoffrey Hinton and Yoshua Bengio — who received the 2018 Turing Award for foundational work on deep learning. His specific contribution was the invention of convolutional neural networks (CNNs) in the late 1980s, the architecture that made computer vision commercially viable. Every image recognition model deployed at scale today — from facial recognition to medical imaging to autonomous vehicle perception — traces its lineage to architectures LeCun pioneered.
Since 2013, LeCun has served as Meta's Chief AI Scientist, directing the company's fundamental AI research through the FAIR (Fundamental AI Research) lab. Under his leadership, FAIR released the LLaMA language model family, Segment Anything, DINO self-supervised vision models, and a series of increasingly ambitious world model prototypes. He simultaneously holds an academic appointment at New York University — a dual role that has given him unusual freedom to publicly critique industry trends, even when those critiques target the dominant paradigm his employer participates in.
What makes LeCun unusual among AI leaders is his willingness to publicly and repeatedly argue that the dominant approach — scaling large language models — is fundamentally limited. While OpenAI, Anthropic, Google DeepMind, and most of the industry have poured tens of billions into making transformers bigger and more capable, LeCun has consistently maintained that autoregressive language models cannot achieve human-level intelligence because they lack a grounded understanding of the physical world. They predict text; they do not understand reality.
This is not a fringe position held quietly. LeCun has articulated it in keynote addresses, academic papers, public debates, and — frequently and memorably — on social media. His bluntness about the limitations of LLMs while working at a company that ships LLM-based products has created an unusual dynamic: Meta's Chief AI Scientist is arguably the most prominent critic of the approach most of the industry is pursuing.
Reports indicate LeCun is maintaining his dual role — continuing as Meta's Chief AI Scientist while leading AMI Labs. The precise terms have not been disclosed, but it suggests Meta is comfortable with its chief scientist building an independent venture, possibly because the research directions are complementary or because losing LeCun entirely is the worse alternative.
A world model, in the sense LeCun uses the term, is an AI system that builds an internal representation of how the physical world works — and uses that representation to predict, plan, and reason about future states.
The concept is borrowed from cognitive science. Humans do not navigate the world by predicting the next word in a sentence. We maintain a mental model of physical reality: objects have mass and momentum, gravity pulls things downward, a ball thrown at a certain angle follows a predictable arc, a glass pushed off a table will fall and shatter. These intuitions are not learned from text descriptions of physics — they are learned from direct sensory experience of interacting with the physical world, starting from infancy.
Current large language models learn from text. They consume trillions of tokens and build statistical models of how words, sentences, and concepts relate. This has produced remarkably capable systems — systems that write code, pass bar exams, and carry on nuanced conversations. But LeCun's argument is that this capability is fundamentally bounded: an LLM's understanding is linguistic, not physical. It knows that "the ball falls when released" because that sentence appears in its training data, not because it has any representation of gravity, mass, or trajectory.
A world model would learn the underlying dynamics. Given video of a ball being thrown, it would learn to predict where the ball will be in the next frame — not by memorizing pixel patterns, but by learning an abstract representation of projectile motion. Given thousands of hours of video showing objects interacting, it would build representations of physical laws that generalize to novel situations.
This distinction matters enormously for applications involving physical reality. Robotics, autonomous driving, manufacturing automation, surgical assistance — all require systems that understand physics, not just language. An LLM can describe how to pick up a cup; a world model could plan the motor sequence required to do it, anticipating the cup's weight, surface friction, and grasp dynamics.
LeCun's position paper on autonomous machine intelligence, published in 2022, laid out a cognitive architecture in which world models serve as the central predictive engine — the component allowing an agent to simulate future states, evaluate plans, and act with foresight rather than reflex.
AMI Labs — Advanced Machine Intelligence Labs — closed a $1.03 billion seed round, as first reported by Bloomberg on March 10, 2026. The round is the largest seed investment in European history and among the largest seed rounds globally in any sector.
To contextualize: most seed rounds for AI startups range from $5 million to $50 million. Even the most generously funded AI seeds — Mistral AI ($113 million in 2023) and Safe Superintelligence ($1 billion in 2024) — were considered extraordinary. AMI Labs matches or surpasses the upper end of that range and exceeds nearly every European technology fundraise at any stage, let alone seed.
The term "seed round" is somewhat misleading at this scale. Traditional seed funding supports prototype development and early hiring. A billion-dollar seed signals that investors believe the founding team and thesis are strong enough to justify committing at scale before the company has shipped a product. In practice, this funds several years of compute-intensive research, a large initial hiring push, and infrastructure needed to train world models at frontier scale.
Sources cited by VentureBeat indicated a valuation in the $4-6 billion range, implying investors took a substantial equity stake. For a pre-product company, that valuation is justified almost entirely by the credibility of the founder and the conviction that world models represent a large addressable market.
AMI Labs is reportedly headquartered in Paris, making it a significant addition to the European AI ecosystem. Europe has struggled to produce AI companies at the scale of American and Chinese competitors, and a billion-dollar Paris-based lab led by one of the field's most decorated researchers materially changes the landscape.
The debate between LLM scaling and world models is not academic — it is a multi-hundred-billion-dollar strategic question about where AI research investment should flow.
The LLM scaling thesis holds that transformer architectures trained on increasingly large datasets with more compute will continue exhibiting emergent capabilities — that intelligence is a function of scale, and the path to AGI runs through making models bigger. This thesis has produced remarkable results: GPT-4, Claude, Gemini, and their successors demonstrate capabilities that seemed impossible five years ago.
LeCun's counterargument is structural. His claim is not that LLMs need to be bigger but that the autoregressive, next-token-prediction paradigm has an inherent ceiling. Autoregressive models generate outputs one token at a time, predicting the next word based on everything before it. This architecture means the model commits sequentially, with no mechanism for backtracking, planning ahead, or evaluating partial outputs before completing them.
More fundamentally, LeCun argues that text is an impoverished representation of reality. Human intelligence is not primarily linguistic. Infants learn about the physical world — object permanence, gravity, causality, spatial relationships — long before they learn language. An AI system trained exclusively on text learns a map of human language, not a map of reality. It can reason about the world only to the extent that the world has been described in its training data, and its reasoning operates on linguistic abstractions rather than grounded physical representations.
A child, LeCun has argued, learns an enormous amount about physical reality — object permanence, the behavior of liquids, the properties of materials — before speaking a single sentence, entirely from sensorimotor interaction. No amount of text training substitutes for that grounded learning, because text is a description of experiences, not the experiences themselves.
The practical implication: LLMs excel at language tasks and will continue improving. World models, if they achieve their promise, would excel at physical reasoning, planning, and real-world interaction — domains where LLMs currently struggle despite their linguistic fluency.
JEPA — Joint Embedding Predictive Architecture — is the technical framework at AMI Labs' core. LeCun and his FAIR collaborators have published multiple papers on JEPA variants, with I-JEPA (Image-based JEPA) and V-JEPA (Video-based JEPA) being the most prominent.
Standard generative models — GPT variants, diffusion models, video generators — are trained to predict or reconstruct raw data. A language model predicts the next token. A diffusion model reconstructs an image from noise. These models must model the full distribution of sensory data, including all irrelevant detail, noise, and redundancy.
JEPA takes a fundamentally different approach. It takes two views of the same input — two crops of an image, or two temporal segments of a video — and encodes each into a latent representation using separate encoder networks. The training objective is not to reconstruct the original input but to predict the latent representation of one view from the other. The model learns by aligning predictions in abstract representation space, not pixel space.
This distinction is critically important. Generative models learn to produce outputs in the original data space: text tokens, pixel values, audio waveforms. JEPA produces predictions in an abstract, learned representation space. The model is forced to learn high-level, semantically meaningful features rather than low-level statistical patterns. It cannot succeed by memorizing textures; it must learn structural representations capturing essential dynamics.
For video inputs, this has a direct physical interpretation. When V-JEPA processes video, it learns to predict future frames not as pixel collections but as abstract states. To predict accurately, the model must implicitly learn object persistence (things do not randomly disappear), physical dynamics (moving objects continue moving), occlusion (objects behind others still exist), and causality (actions have predictable consequences).
The "joint embedding" refers to both views being embedded into the same latent space, with training bringing corresponding embeddings closer while pushing non-corresponding embeddings apart. This predictive learning signal drives the model to discover structure in the world.
JEPA is self-supervised — no human-labeled data required. It learns from raw video and images. This is a critical practical advantage: the world produces vastly more video than text, mostly unlabeled. A learning architecture extracting physical knowledge from unlabeled video has access to training data that dwarfs what is available for language models.
Early results from I-JEPA and V-JEPA have been promising but preliminary. I-JEPA demonstrated state-of-the-art image understanding performance while using significantly less compute than competing approaches. V-JEPA showed the ability to learn useful video representations without labeled data. AMI Labs' mission is to scale these research prototypes to frontier capability — to do for world models what OpenAI did for language models between GPT-2 and GPT-4.
The investor roster reads like a strategic alignment of interests across the AI value chain. Each participant has specific reasons to bet on this outcome.
Jeff Bezos has become one of the most active individual AI investors, with stakes in Anthropic (reportedly over $4 billion through Amazon) and now AMI Labs. His interest in world models aligns with Amazon's massive robotics investments. Amazon operates one of the world's largest warehouse robotics fleets and has invested heavily in autonomous delivery. A world model understanding physical manipulation, spatial navigation, and object dynamics has direct commercial application in Amazon's operations. Bezos is not making a philosophical bet on AI architecture — he is investing in technology that could transform his company's physical infrastructure.
NVIDIA is the dominant AI training hardware supplier, and its investment serves multiple strategic purposes. World model training at scale will require enormous GPU clusters — potentially more compute-intensive than LLM training, since video data is far denser than text. NVIDIA investing in AMI Labs is partly customer development: if world models become a major paradigm, NVIDIA wants the leading lab building on its hardware. NVIDIA also has strategic interest in world models for autonomous systems through its DRIVE platform and Omniverse simulation platform.
Samsung brings a consumer electronics and semiconductor perspective. As a manufacturer of AI chips, consumer robots, and devices across smartphones, appliances, and industrial automation, Samsung sees world models as technology with direct product applications. An investment in AMI Labs positions Samsung to access world model technology for its hardware ecosystem — particularly for on-device, edge AI capabilities that current LLMs cannot efficiently deliver.
Temasek, Singapore's sovereign wealth fund managing approximately $382 billion, has been one of the most active sovereign investors in AI. Its participation signals institutional confidence in the long-term thesis and provides AMI Labs with patient capital on longer time horizons than traditional venture capital.
The coalition is notable for what it excludes: none of the major LLM companies (OpenAI, Anthropic, Google). This is a group with commercial reasons to want an alternative to the LLM paradigm, or who want portfolio diversification across AI approaches.
If JEPA-based world models achieve the capabilities LeCun envisions, the implications extend far beyond a successful funding round.
Robotics is the most immediate application. The central bottleneck in robotics is not hardware — motors, actuators, and sensors are increasingly capable — but software that reasons about physical interactions in unstructured environments. Current robots either operate in tightly controlled settings (factory lines) or rely on brittle rules that fail in novel situations. A world model reasoning about physical properties rather than pattern-matching against training scenarios would be dramatically more robust. The commercial value across manufacturing, logistics, healthcare, and services is measured in trillions of dollars.
Autonomous vehicles represent another high-stakes domain. The most challenging self-driving failure modes involve rare, novel scenarios — a child running into the street in an unusual posture, oversized cargo blocking lanes, unexpected construction. These are precisely where causal physical reasoning outperforms pattern-matching against similar training examples. The difference between 99% and 99.9% reliability is the difference between unacceptable and commercially deployable.
Scientific simulation is a less obvious but potentially transformative application. World models could learn to approximate complex simulations — fluid dynamics, structural mechanics, molecular dynamics — at a fraction of the computational cost of traditional numerical methods. This would not replace physics-based simulation for high-precision work but could dramatically accelerate design iteration and materials discovery.
Industrial automation involves complex physical processes currently difficult to automate because they require reasoning about materials, forces, tolerances, and failure modes. A world model trained on industrial processes could enable automation systems that adapt to variability rather than requiring every scenario to be pre-programmed.
The structural implication for the AI industry: if world models represent a genuinely superior architecture for AGI-relevant capabilities, the current infrastructure built around transformer scaling becomes a partially misdirected investment. Language modeling would remain valuable, but the race for AGI would shift to a different technical track entirely.
LeCun's thesis has serious critics, and their arguments deserve fair treatment.
The strongest counterargument is empirical: LLMs keep getting better. Each generation of transformer models demonstrates capabilities the previous generation lacked, with no clear signs of plateauing. Proponents of scaling argue that emergent capabilities — new skills appearing as models grow — are evidence that the transformer architecture has not reached its ceiling. Investing in a fundamentally different architecture may be premature when the current one has not been exhausted.
OpenAI and others have argued that multimodal training addresses LeCun's concerns without requiring a new architecture. GPT-4 and successors process images, audio, and video alongside text. If the concern is that text-only training produces text-only understanding, multimodal transformers offer a response within the existing paradigm.
Researchers at Google DeepMind and Anthropic who work on scaling have argued that LeCun conflates training objective with learned representation. The fact that a model is trained to predict tokens does not mean it can only represent statistical associations. Large models appear to develop rich internal representations of concepts and relationships not directly encoded in the training signal. Emergent capabilities are evidence that the right representations arise from scale even when the training objective does not explicitly require them.
There is also a practical objection: JEPA remains largely unproven at scale. I-JEPA and V-JEPA have shown promising benchmark results but have not produced a system with the broad capabilities that frontier LLMs demonstrate daily. The gap between "interesting research direction" and "commercially viable system" is enormous. LLM labs have spent tens of billions on training runs alone. If world models require comparable investment, AMI Labs' seed round is a starting point, not an endpoint.
Some researchers occupy a middle ground, arguing the future likely involves hybrid architectures combining linguistic reasoning of transformers with physical grounding of world models. Rather than a winner-take-all paradigm war, the outcome may be integration — systems using different architectures for different cognitive functions, much as the human brain uses different circuits for language, spatial reasoning, and motor control.
LeCun himself has acknowledged that world models will not replace LLMs for language tasks. His argument is specifically about the path to general intelligence and physical reasoning. AMI Labs is a billion-dollar bet that language models alone cannot get there.
Sources: Bloomberg, VentureBeat, Fortune, Meta AI Research
NVIDIA CEO Jensen Huang projects 7.5M AI agents creating a 100:1 AI-to-human worker ratio. From GTC 2026's IGX Thor launch to the $1.7 GW GPU infrastructure boom, here's what NVIDIA's agent economy vision means.
Oracle commits $50B in FY2026 capex for AI infrastructure, more than doubling last year's spend, as its cloud backlog surges to $553B. First to deploy NVIDIA Blackwell at scale and part of the $500B SoftBank-OpenAI mega-deal, Oracle is reinventing itself as an AI infrastructure powerhouse.
Niantic's Pokémon GO generated 30 billion user photos that are now training delivery robots for real-world navigation. The biggest gamification-to-AI training pipeline ever raises questions about consent, data rights, and the future of crowd-sourced robotics.