Yann LeCun’s AMI Labs Raises $1bn For AI Beyond Language Models

Yann LeCun's Paris-based startup, Advanced Machine Intelligence Labs, is using more than $1bn in seed funding to develop AI systems that can reason about the physical world, rather than mainly predict language.

Investors include Nvidia and the fund that manages Amazon founder Jeff Bezos' private wealth.

The round was one of Europe's biggest early-stage startup financings, according to the BBC.

LeCun spent a decade at Meta, where he was chief AI scientist, before leaving in 2025 to found AMI Labs.

The company's target is a limitation he sees in current systems such as ChatGPT, Claude and Gemini: they can code, solve mathematical problems and generate text, but they struggle with situations in which an action can produce many possible outcomes.

"They're not a path towards human level or human-like intelligence, or even animal-like intelligence, because they cannot deal with real world data, they just are not built for that," LeCun said at VivaTech, France's leading technology conference.

The startup is developing a system called Joint Embedding Predictive Architecture, or JEPA.

Instead of trying to predict every detail of a scene, JEPA creates abstractions of the real world and uses them to assess the likely results of actions.

Difficult mathematics is involved, but the basic idea is to filter out information that does not matter to the decision.

LeCun illustrated the distinction with a pen balanced upright on its tip.

A person, even a toddler, knows that the pen will fall once released.

There is no reason to predict the exact direction because it cannot be known in advance.

A language model might instead generate one statistically plausible prediction based on patterns in its training data.

That difference matters in robotics.

Billions of dollars have been invested in humanoid robots, yet training them to perform household jobs such as ironing or stacking a dishwasher remains difficult and expensive.

A robot must decide which physical details matter before it moves; a plausible verbal answer is not enough.

"LLMs are largely hopeless for robotics," LeCun said.

He also rejected the idea that scaling up large language models alone will produce superhuman intelligence.

Ingmar Posner, professor of Applied Artificial Intelligence at Oxford University and director of its Applied AI Lab, shares the view that future systems will need to explain relationships between actions and outcomes.

His team of around 10 researchers has spent four years working on an alternative approach known broadly as world models.

"You need models that can answer questions like: What matters? What causes what? What would happen if I did something else - like if I took a different action?" Posner said.

World models have been discussed for decades, but advances in machine learning and computing helped renew interest.

A 2018 paper by David Ha and Jurgen Schmidhuber argued that an AI could learn to act through a learned mental simulation of its surroundings.

Since then, related work has included Google's Dreamer World Model.

Last year, a Dreamer variant collected diamonds in the video game Minecraft by imagining possible future scenarios before making decisions.

Other projects include DeepMind's Genie model, Alphabet's London-based Wayve and its Gaia system, and World Labs, founded in San Francisco by AI pioneer Fei-Fei Li in 2023.

Posner calls his team's approach a "mechanistic world model" designed to organise knowledge so it can be recalled, combined and modified when it matters.

The timetable remains uncertain.

Posner noted that people in 2017 or 2018 might have expected a ChatGPT-like system to be decades away; the original version of ChatGPT launched in November 2022.

AMI Labs plans to spend the rest of this year refining its model.

LeCun said the company hopes to put it to work next year, initially in industrial settings.

If that succeeds, he said, the systems could eventually be adapted to a much wider range of tasks with minimal training or fine-tuning.