World Models:The Future of AI
:format(webp))
:format(webp))
In March 2026, one startup made European tech history by raising a $1bn seed round.
The company was AMI Labs, a startup building ‘world models’ that, unlike large language models, can understand the physical world. Rather than simply generating text, these models can retain context, reason about how situations might evolve, and plan actions safely and reliably.
Cofounded by former Meta chief AI scientist and serial AI entrepreneur Alex LeBrun, AMI Labs is betting that world models are the next step on the path to artificial general intelligence (AGI).
“LLMs predict the next word, or token, based on patterns they have learned from written sources like the internet, books and research papers. They work autoregressively, meaning they use the previous words to estimate what is most likely to come next,” says Alex Joël-Carbonell, partner at HV Capital.
World models, on the other hand, learn as humans do: through vision, touch and observation. “It’s a much richer set of inputs than words alone,” he adds.
In March 2026, one startup made European tech history by raising a $1bn seed round.
The company was AMI Labs, a startup building ‘world models’ that, unlike large language models, can understand the physical world. Rather than simply generating text, these models can retain context, reason about how situations might evolve, and plan actions safely and reliably.
Cofounded by former Meta chief AI scientist and serial AI entrepreneur Alex LeBrun, AMI Labs is betting that world models are the next step on the path to artificial general intelligence (AGI).
“LLMs predict the next word, or token, based on patterns they have learned from written sources like the internet, books and research papers. They work autoregressively, meaning they use the previous words to estimate what is most likely to come next,” says Alex Joël-Carbonell, partner at HV Capital.
World models, on the other hand, learn as humans do: through vision, touch and observation. “It’s a much richer set of inputs than words alone,” he adds.
The pathto WorldModels
The technology underpinning AMI Labs is based on LeCun's Joint Embedding Predictive Architecture, or JEPA, which he has been developing since 2022.
LeCun often compares the concept to how a baby learns. A child doesn't understand gravity by reading a book on physics. It learns by dropping objects and watching what happens. Researchers hope AI systems can learn in a similar way.
The potential applications are already becoming clear. Today’s robots, for example, work well in industrial settings where they are programmed to perform specific tasks like lifting objects, but struggle in unfamiliar surroundings.
World models could help robots understand their environment and respond to unforeseen events.
Self-driving cars could also benefit from world models. While they can operate in predefined conditions, they struggle to adapt to entirely new environments, such as different countries with unfamiliar roads and traffic rules. World models could provide the real-world context to help the vehicles adapt accordingly.
:format(webp))
“You couldn’t leave it in the street and tell it to walk, because it only knows this particular setting.”
Alexander Joël-Carbonell
Partner, HV
The future of world models
Part of the excitement around AMI Labs comes from its team of researchers who are people of “high integrity,” says Alex.
“The researchers they’ve hired all had excellent compensation packages at OpenAI, DeepMind and Facebook — they’re not doing this for the money. They’re doing it because, together, they have the chance to create something bigger than themselves, something they felt they couldn’t do at their previous companies.”
LeCun’s first hire was French entrepreneur Alex LeBrun — who successfully sold two AI startups to Nasdaq-listed companies and was scaling his third business, Nabla, an AI assistant for doctors at the time he was called up to become CEO of AMI Labs.
The company’s leadership team also includes Saining Xie as chief science officer, Pascale Fund as chief research and innovation officer, Michael Rabbat as VP of world models, and Laurent Solly as COO.
With Europe at the cutting edge of research into world models, Alex thinks the technology could become Europe's answer to the LLMs currently dominated by Silicon Valley Big Techs.
European AI research talent who moved to work for big US companies are returning to Europe to try their hand at something beyond LLM architecture.
“There’s a wave of frontier AI being built in places like Berlin, London and Paris, and I think this is Europe’s chance to play a global role as a counterbalance to the US and Asia,” says Alex.
:format(webp))
“World models, in particular, feel like a predominantly European phenomenon.”
Alexander Joël-Carbonell
Partner, HV