# AMI Labs Raises $1.03B to Build World Models

> Source: [https://icharles.com/articles/ami-labs-lecun-world-models-funding](https://icharles.com/articles/ami-labs-lecun-world-models-funding) (canonical)
> Author: iCharles News — iCharles, https://icharles.com
> Published: 2026-07-08

## TL;DR

AMI Labs, the Paris-based AI startup founded by Yann LeCun, raised $1.03 billion at a $3.5 billion pre-money valuation in March 2026. The company is building world models — AI systems that learn physical cause-and-effect from sensory data rather than text. At ICML 2026 in Seoul, AMI Labs co-founder Pascal Fung argued that LLMs cannot understand the physical world, while Anthropic separately announced it had discovered a spontaneous reasoning structure called "J-space" inside its Claude model.

## What did AMI Labs just announce?

AMI Labs raised $1.03 billion — roughly €890 million — at a $3.5 billion pre-money valuation, [according to Cathay Capital](https://www.cathaycapital.com/advanced-machine-intelligence-ami-is-enabling-the-next-ai-revolution-built-on-foundational-world-models/). The round closed on March 10, 2026. The company plans to use the capital for long-term research, global hiring, and building reliable intelligent systems that learn abstract representations of the real world.

AMI Labs is headquartered in Paris and also operates offices in New York, Montreal, and Singapore.

## Who founded AMI Labs and why?

**AMI Labs** is a frontier AI research startup founded by Yann LeCun, a Turing Award laureate and former chief AI scientist at Meta. LeCun officially launched the company in March 2026. He stated he created AMI Labs to continue the Advanced Machine Intelligence research program he had been pursuing at Meta and New York University.

The name is pronounced "ah-mee" — the French word for "friend." French President Emmanuel Macron publicly welcomed the decision to base the company in Paris and pledged to "do everything to ensure [LeCun] succeeds from France," [as reported by Built In](https://builtin.com/articles/ami-labs-yann-lecun).

The co-founders include CEO Alexandre LeBrun, Chief Science Officer Saining Xie, and Chief Research & Innovation Officer Pascal Fung.

## What are world models, and why does AMI Labs focus on them?

**A world model** is an AI system that learns the rules of the physical world — cause-and-effect, spatial logic, physical consequences — from multimodal sensory data rather than from text. This is the core technology AMI Labs is building.

LeCun has argued that [large language models](/articles/openai-inference-cost-halved-optimization) face hard limits because they only understand the world indirectly through human-written text. AMI Labs is researching world models based on **JEPA (Joint Embedding Predictive Architecture)**, a technology LeCun developed during his time at Meta. Unlike LLMs that predict the next token, JEPA focuses on predicting the next situation — understanding context rather than generating output.

## What did Pascal Fung say at ICML 2026?

Pascal Fung, AMI Labs' Chief Research & Innovation Officer, delivered a keynote at ICML 2026 at COEX in Seoul on July 7. "LLMs only understand the world indirectly through text written by humans," Fung said. "For AI agents that operate in the real world, we need a world model that directly understands the physical environment."

Fung used a soccer match as an analogy. A player instantly reads spatial relationships, physical causality, team objectives, and teammate emotions all at once. "A true world model reads both the physical and mental worlds together like this," he said.

He also flagged hallucinations as a safety risk beyond just text errors. "Hallucinations occurring at the text level may be harmless, but if a hallucination occurs in a robot, it could lead to a collision," Fung warned.

## How do current AI models perform on physical reasoning?

The gap between human and AI physical reasoning is measurable. On **DeepPhy**, a benchmark that tests physical reasoning, humans scored 64.7% accuracy. The best AI model scored just 41.2%.

Here's a snapshot of the key performance data from the sources:

| Tested entity | DeepPhy accuracy |
|---|---|
| Humans | 64.7% |
| Best AI model (top LLM) | 41.2% |

Fung listed four elements he considers essential for real-world AI agents: perception, prediction, planning, and memory. He also noted that current Vision-Language-Action (VLA) models rely heavily on imitation learning and become unreliable when they encounter situations outside their training data.

## What did Anthropic find inside Claude?

At the same ICML 2026 conference, [Anthropic](/articles/california-anthropic-claude-state-partnership) published research titled "Global Workspace in Language Models." The paper describes a region inside Claude called **J-space** — a self-organized reasoning structure that Claude formed spontaneously during training. Anthropic did not design it intentionally and was unable to determine exactly how it emerged.

Researchers used an observation tool called the **Jacobian lens (J-lens)** to study J-space. They found that when Claude answers a question, it first activates a key concept in J-space and then builds its answer from there. When asked how many legs a web-spinning animal has, Claude first activated "spider" in J-space, then answered "eight."

Researchers also found they could alter Claude's answers by changing the concept in J-space. Swapping "spider" for "ant" caused Claude to answer "six." Changing "France" to "China" caused Claude to answer questions about France using China's capital and currency.

Anthropic mapped J-space's properties onto characteristics of "consciously accessible information" from human cognitive science: the content can be reported, it activates on demand, a single concept applies across multiple tasks, and it operates selectively.

## What is LeCun's position on AGI?

LeCun posted on X (formerly Twitter) on July 4, stating bluntly: "The 'G' in AGI is nonsense." He has argued consistently since 2023 that LLMs are a statistical interface, not true intelligence. His view directly conflicts with Anthropic's J-space findings, which suggest that human-like reasoning structures can emerge inside LLMs without being explicitly programmed.

Here's where we see the clearest tension in the sources: AMI Labs is betting that world models are the necessary path forward, while Anthropic's research suggests LLMs may be developing internal structures that look more like reasoning than LeCun's framework allows.

## What is AMI Labs building toward?

AMI Labs follows a dual-track model. It publishes open-source research and tools while also developing commercially licensable products. LeCun has suggested Meta could become one of AMI's first customers. The company positions itself as a "third path" alternative to the U.S.-China tech giant binary, with a goal of repositioning Europe as an AI hub.

AMI Labs is also focused on [robotic AI agents](/articles/deepseek-proprietary-ai-chip-development) that can learn in zero-shot settings — meaning they can operate in new environments with little or no prior training data. Fung described physical AI as "an area where many things remain unsolved."

The $1.03 billion seed round closed March 10, 2026, making it one of the largest seed rounds in AI history based on the figures reported.

## Frequently asked questions

****How much money did AMI Labs raise?****

AMI Labs raised $1.03 billion USD, approximately €890 million, in a seed funding round that closed on March 10, 2026. The round was based on a $3.5 billion pre-money valuation. The capital is earmarked for long-term research, global hiring, and development of AI systems that learn abstract representations of the real world.

****Who is Yann LeCun and what is AMI Labs?****

Yann LeCun is a Turing Award laureate and former chief AI scientist at Meta. He founded AMI Labs in March 2026 to continue the Advanced Machine Intelligence research program he had been pursuing at Meta and NYU. AMI Labs is a Paris-based startup focused on building world models — AI systems that learn physical cause-and-effect from sensory data rather than text.

****What is JEPA and how does AMI Labs use it?****

JEPA, or Joint Embedding Predictive Architecture, is a technology Yann LeCun developed during his time at Meta. Unlike large language models that predict the next token in a sequence, JEPA focuses on predicting the next situation — understanding context and physical relationships. AMI Labs is using JEPA as a foundation for its world model research.

****What is J-space in Anthropic's Claude model?****

J-space is a self-organized reasoning region that Anthropic discovered inside its Claude model. Claude formed it spontaneously during training — Anthropic did not design it. Researchers found that J-space controls Claude's reasoning: removing it drops Claude's performance to near kindergarten level. Altering a concept in J-space directly changes Claude's answers, showing it actively participates in reasoning.

****How do LLMs perform on physical reasoning benchmarks?****

On DeepPhy, a benchmark that tests physical reasoning ability, humans achieved 64.7% accuracy. The best-performing AI model scored just 41.2%. AMI Labs co-founder Pascal Fung cited this gap at ICML 2026 to argue that LLMs are fundamentally ill-suited for understanding physical causality, making world models a necessary alternative for real-world AI agents.