AMI Labs Targets Physical-World AI After $1bn Seed Round

Yann LeCun's new Paris-based company, Advanced Machine Intelligence Labs, is trying to move artificial intelligence away from language-only prediction and towards systems that can reason about the physical world.

BBC reported that AMI Labs raised more than $1bn earlier this year in a seed round backed by investors including Nvidia and the fund that manages Jeff Bezos' private wealth.

BBC described the round as one of Europe's biggest early-stage startup financings.

LeCun, who worked for a decade at Meta and was its chief AI scientist before leaving in 2025, told BBC that current systems such as ChatGPT, Claude and Gemini have uses but are not built to handle complicated real-world situations.

He said they are not a path towards human-like or animal-like intelligence because they cannot deal with real-world data.

The funding backs AMI Labs' bet that robotics and industrial systems need models that understand cause, action and uncertainty, rather than only generating statistically plausible language.

JEPA Is Built Around Abstractions Rather Than Text Prediction

LeCun said AMI Labs is developing Joint Embedding Predictive Architecture, or JEPA.

BBC said the system creates abstractions of the real world so an AI can assess possible outcomes of actions without trying to predict every irrelevant detail.

BBC used the example of a pen balanced on its tip.

A person knows it will fall, but does not need to know the exact direction.

LeCun's argument is that a language model may try to predict a single next event from training patterns, while a physical-world model needs to know which uncertainty can be ignored.

The robotics link is direct.

BBC reported that billions of dollars have gone into humanoid robots, but household tasks such as ironing or stacking a dishwasher remain difficult and costly to train.

LeCun said large language models are largely hopeless for robotics and rejected the claim that scaling them alone will produce superhuman intelligence.

The difference is practical for companies building robots.

A language model can produce a plausible sentence after reading text, but a robot needs a model that can decide which physical details matter before it moves.

In the pen example, the useful knowledge is that the pen will fall, not a precise guess about direction.

Ingmar Posner, professor of Applied Artificial Intelligence at Oxford University and director of its Applied AI Lab, said the next decade will require systems that can explain what matters, what causes what and what would happen after a different action.

World Models Add A Research Race Around Robotics

Posner told BBC that his team of around 10 researchers has spent four years working on an alternative AI approach in the broad category of world models.

He said the goal is to structure knowledge so it can be recalled, combined and modified when needed.

BBC reported that an influential 2018 paper by David Ha and Jurgen Schmidhuber helped inspire the current wave of work on world models.

It also cited Google's Dreamer World Model, DeepMind's Genie model, London-based Wayve's Gaia system and Fei-Fei Li's World Labs, founded in San Francisco in 2023.

Posner said it is difficult to predict how long these models will take to mature.

BBC noted that many observers in 2017 or 2018 would have expected a ChatGPT-like system to be decades away, before the original version of ChatGPT launched in November 2022.

AMI Labs is entering a crowded research path rather than opening a finished commercial category.

The common problem is whether AI can use an internal representation of the world to plan actions before testing them in expensive or unsafe physical settings.

LeCun said AMI Labs will spend the rest of this year refining its model and hopes to put it to use next year, first in industrial settings.

AMI Labs has not named customers, contract values, deployment sites, safety certifications or a confirmed date for the first industrial rollout.