AMI Labs: The $1B Bet That World Models Will Beat LLMs

Yann LeCun left Meta after a decade as chief AI scientist to start Advanced Machine Intelligence Labs. AMI Labs just raised over $1 billion — described as “Europe’s largest seed funding round ever” — at a $3.5 billion pre-money valuation. The investors include Nvidia, Bezos Expeditions, Singapore’s sovereign wealth fund Temasek, and France’s Cathay Innovation. The company launched with approximately a dozen employees across Paris, New York, Singapore, and Montreal.

The thesis behind the investment: large language models are the wrong architecture for the kind of AI that can understand and act in the physical world, and LeCun has been the most persistent and high-profile proponent of that argument for several years.

What AMI Labs Is Actually Building

AMI Labs focuses on “world models that understand the physical environment,” with stated applications in robotics and transportation. World models — AI systems that develop an internal representation of how the physical world works, how actions lead to outcomes, and how states evolve over time — are the capability that LeCun argues LLMs fundamentally cannot provide.

The distinction from language models is architectural. LLMs are trained to predict the next token in a sequence; they develop remarkable competence at pattern recognition and language generation as a byproduct of that objective. They don’t have a predictive model of physical causality, object permanence, or the consequence of actions in a real environment. For robotics and autonomous systems that need to navigate the physical world, this is a fundamental limitation that more scale doesn’t resolve.

LeCun’s alternative — which he has called JEPA (Joint Embedding Predictive Architecture) and related approaches in prior Meta research — trains models to predict abstract representations of future states rather than predicting next tokens. The hypothesis is that this objective produces systems that generalize better to novel physical situations.

The Investor Signal

The composition of AMI Labs’ investor base is as significant as the amount. Nvidia has a direct financial interest in what the next generation of AI hardware is optimized for. If world models for physical intelligence become the dominant AI capability paradigm, they require different compute profiles than LLM inference — and Nvidia’s bet on AMI Labs is a hedge on that architectural transition.

Temasek and Bezos Expeditions represent patient capital with long time horizons — both are comfortable with research bets that take five to ten years to produce commercial applications. This investor composition suggests AMI Labs has structured the raise to support foundational research without the pressure to commercialize on a venture timeline.

The fact that Meta isn’t investing, but is expected to partner with AMI Labs, is also notable. LeCun’s departure from Meta apparently didn’t create the adversarial dynamic that sometimes follows a high-profile exit — suggesting the research direction AMI Labs is pursuing has at least tacit institutional endorsement from his former employer.

Content created with AI assistance and reviewed for accuracy.