Yann LeCun's AMI Raises $1.03B Seed — The Biggest Bet Against LLMs

$1 Billion for a Seed Round. Read That Again.

Typical seed rounds are measured in single-digit millions. AMI Labs just closed $1.03 billion — the largest seed round in European history, and one of the largest seed rounds in the history of venture capital anywhere. Behind it stands Yann LeCun, the 65-year-old Turing Award winner who believes the entire AI industry is building on the wrong foundation.

This is not just a funding story. It's a fundamental challenge to the autoregressive paradigm that powers ChatGPT, Claude, Gemini, and every other large language model. Understanding AMI requires understanding what LeCun thinks is wrong with LLMs — and what he proposes instead.

Background: Who Is Yann LeCun?

Yann LeCun is one of the three "godfathers of deep learning," alongside Geoffrey Hinton and Yoshua Bengio. The three shared the 2018 Turing Award for their foundational work on neural networks.

Key career milestones:

• 1989: Invented convolutional neural networks (CNNs) at AT&T Bell Labs — the architecture that powers virtually all computer vision systems today
• 1998: Created LeNet-5, the first practical CNN for handwritten digit recognition, used by the US Postal Service
• 2003-present: Professor at New York University
• 2013-present: Chief AI Scientist at Meta (formerly Facebook)
• 2018: Turing Award (shared with Hinton and Bengio)
• 2026: Founded AMI Labs while retaining his Meta position

LeCun's reputation is not just academic. He has been consistently, publicly, and often controversially vocal about what he sees as fundamental limitations of current AI approaches.

What's Wrong with LLMs, According to LeCun

LeCun's criticism of autoregressive LLMs has been building for years. His core arguments:

1. Token Prediction Is Not Understanding

LLMs predict the next token based on statistical patterns in training data. They have no internal model of how the world works. When GPT-4 correctly answers a physics question, it's not because it understands physics — it's because it has seen similar question-answer patterns in training data.

LeCun's analogy: "A parrot that has memorized the physics textbook can answer physics questions, but it doesn't understand physics."

2. Text Is a Lossy Compression of Reality

Human knowledge expressed in text represents a tiny fraction of our understanding of the world. A child learns more about physics by dropping a ball 100 times than by reading 100 textbooks about gravity. LLMs are trained exclusively on text (and some images/video), missing the vast majority of human knowledge that's embodied and experiential.

3. Autoregressive Generation Can't Plan

LLMs generate text one token at a time, left to right, without the ability to go back and revise. They can't plan ahead — they can only predict what comes next based on what came before. This makes them fundamentally unsuitable for tasks requiring multi-step planning, like robotics, drug design, or complex engineering.

4. Scaling Won't Fix These Problems

Perhaps LeCun's most controversial claim: simply making LLMs bigger and training them on more data won't overcome these fundamental limitations. The architecture itself is the bottleneck, not the scale.

What Are "World Models"?

AMI's alternative is what LeCun calls "world models" — AI systems that build internal representations of how the physical world works, then use those representations for prediction, planning, and reasoning.

The Core Idea: JEPA

JEPA (Joint Embedding Predictive Architecture) is LeCun's proposed framework for building world models. It works differently from both LLMs and diffusion models.

How LLMs work: Predict the next token in a sequence.

Input: "The cat sat on the" → Predict: "mat"

How diffusion models work: Denoise a noisy image step by step.

Noise → Less noise → ... → Clean image

How JEPA works: Predict the representation of one part of the input from another part, in an abstract embedding space.

[Partial observation of world state]
  → Encoder → [Abstract representation]
    → Predictor → [Predicted representation of missing part]
      → Compare with [Actual representation of missing part]

The critical difference: JEPA doesn't predict pixels or tokens. It predicts abstract representations. This is important because:

1. Abstraction filters out irrelevant detail: Predicting the exact RGB values of every pixel in a video frame is wasteful. JEPA predicts high-level features (objects, motion, relationships).
2. Representations can capture causal structure: Instead of just statistical correlations, JEPA can learn that pushing an object causes it to move — a causal relationship.
3. Planning becomes natural: With a world model, you can simulate future states without generating text or images. "If I do X, what happens?" becomes a forward pass through the predictor.

Energy-Based Models

JEPA is implemented using energy-based models (EBMs). Instead of assigning probabilities to outputs (like LLMs do), EBMs assign an "energy" — a scalar value indicating compatibility between inputs and outputs. Low energy = compatible, high energy = incompatible.

This avoids the "curse of dimensionality" in generative models: instead of assigning probabilities to every possible output (which is exponentially expensive), EBMs only need to distinguish compatible from incompatible pairs.

The $1.03B Seed: Who's Investing and Why

Investor Roster

Investor	Type	Notable
Jeff Bezos	Individual	Amazon founder, Blue Origin
Nvidia	Strategic	GPU supplier, AI ecosystem play
Samsung	Strategic	Semiconductor + devices
Temasek	Sovereign wealth	Singapore's investment arm
Andreessen Horowitz	VC	a16z, leading AI investor
Multiple European institutions	Various	Details undisclosed

Valuation

$3.5 billion pre-product — valuing AMI purely on LeCun's reputation, the team, and the thesis.

Why Investors Are Betting

1. LeCun's track record: CNNs, which LeCun invented, became the dominant architecture for computer vision and generated hundreds of billions in economic value. If world models follow a similar trajectory, early investors win enormously.
2. Paradigm insurance: Even investors bullish on LLMs want exposure to alternative approaches. If LLMs plateau, world models could be the next paradigm.
3. Specific applications: World models have clear applications in robotics, autonomous vehicles, drug discovery, and simulation — markets worth trillions.

The AMI Team

AMI is based in Paris, with additional offices in New York. LeCun retains his position at Meta while serving as AMI's chief scientist. Key team details:

• Headquarters: Paris, France (making this the largest European AI seed round)
• Team size: ~50 researchers at founding, with plans to scale to 200+ within 18 months
• Recruiting from: Meta FAIR, DeepMind, Google Brain, top European universities

LeCun has been building the intellectual framework for world models at Meta FAIR for years. AMI allows him to pursue this vision with dedicated resources and without the constraints of a large corporation's priorities.

Competitive Landscape

World Labs (Fei-Fei Li)

Stanford professor Fei-Fei Li's World Labs raised $230M in September 2024 to build "Large World Models" for 3D visual intelligence. While both companies use the term "world models," their approaches differ:

• World Labs: Focused on 3D visual understanding and generation
• AMI: Focused on abstract representation learning and planning

Runway

Runway's Gen-3 and upcoming Gen-4 models learn visual representations of the world, but through a generative (diffusion-based) approach rather than JEPA's predictive approach.

Meta FAIR

LeCun's former (and current) employer continues world model research. The V-JEPA model, released in 2024, demonstrated early success in learning physical world representations from video. AMI can be seen as the commercialization path for this research.

Why AMI Might Succeed

1. LeCun has been right before — CNNs were dismissed for a decade before becoming dominant
2. The timing may be right — LLM scaling returns are showing diminishing improvements, creating an opening for alternative approaches
3. Hardware is ready — Modern GPUs have enough compute for the massive self-supervised learning world models require

Why AMI Might Fail

1. World models are unproven at scale — No one has demonstrated a world model that generalizes across domains
2. LLMs keep getting better — Each new model generation partially addresses LeCun's criticisms
3. The "just add data" approach is hard to beat — LLMs' simplicity (predict the next token) has proven remarkably powerful

Healthcare: The First Proving Ground

AMI's technology is being applied through Nabla, a separate health AI startup where world models could first prove their value:

• Understanding complex biological systems
• Modeling drug interactions and side effects
• Predicting patient trajectories based on medical history
• Simulating treatment outcomes

Healthcare is a natural first application because:

1. Biological systems follow causal rules (unlike text generation)
2. The cost of errors is high (making LLM hallucinations unacceptable)
3. Data is multimodal (vitals, images, lab results, clinical notes)
4. Planning is essential (treatment plans must account for future states)

What This Means for the AI Industry

Short Term (2026-2027)

• No immediate disruption: AMI is a research-stage company. Don't expect usable products soon.
• Talent competition: AMI's $1B war chest will intensify the competition for top AI researchers.
• Narrative shift: The "LLMs are all you need" narrative will face more scrutiny.

Medium Term (2027-2029)

• First demonstrations: Expect published results showing world models outperforming LLMs on specific tasks (robotics, simulation, planning).
• Hybrid approaches: Other labs will likely combine world model concepts with LLM architectures.
• Industry diversification: More investment in non-LLM AI approaches.

Long Term (2029+)

• If AMI succeeds: Paradigm shift — AI systems that truly understand the physical world, enabling breakthroughs in robotics, scientific discovery, and engineering.
• If AMI fails: The $1B investment still advances fundamental research and trains hundreds of researchers who will carry world model concepts into other organizations.

Key Quotes

"Current AI systems are basically fancy autocomplete. They don't understand the world." — Yann LeCun, multiple public statements

"We need machines that can learn like babies — by observing the world, not by reading the internet." — Yann LeCun, Meta AI Day 2025

"The most appropriate image for an unveiling" (referring to his own photo of the Veil Nebula used in the announcement) — Yann LeCun, X/Twitter announcement

Related Projects and Background

The Lineage of World Model Research

LeCun didn't invent the World Model concept. In 2018, David Ha and Jürgen Schmidhuber published "World Models" — agents that learn environments using VAE + RNN. Google DeepMind's Dreamer series (2020–2025) follows the same thread. LeCun's differentiator: predicting without generating. Prior World Models tried to generate future frames pixel-by-pixel, which was computationally prohibitive. JEPA predicts in abstract space — far more efficient.

Difference from Video Generation Models

Sora (OpenAI), Runway Gen-3, and Kling (Kuaishou) "generate" video. AMI aims to "understand" video. Generative models produce plausible video but can violate physics (water flowing upward, objects disappearing). World Models understand physical causation, structurally preventing such errors.

Contrastive Learning vs JEPA

Mainstream self-supervised methods like SimCLR and DINO use contrastive learning — distinguishing positive/negative pairs. JEPA doesn't explicitly use contrastive learning. Instead, it predicts one view from another while learning to ignore unpredictable information. LeCun argues this is closer to human learning — we don't memorize every detail of the world, we learn important abstract structures.

Expectations in Robotics

Robotics is the most anticipated application for AMI's World Model — and why Toyota Ventures invested. Current robots require thousands of trial-and-error attempts to learn new tasks. With a complete World Model, robots could simulate action outcomes in their heads before choosing the optimal action. "A robot that learns new tasks from a single demonstration" is LeCun's vision.

References

• TechCrunch: AMI Labs Raises $1.03B
• Bloomberg: LeCun's New AI Startup
• AMI Labs Official Website
• LeCun's JEPA Paper (2022)
• V-JEPA: Video Joint Embedding Predictive Architecture
• Built In: Yann LeCun Launches AMI Labs