AMI Labs is quietly building a world model designed to bridge simulation and the messy physical world, and its founder refuses to use AGI rhetoric.

While the rest of the artificial intelligence industry is rushing to use labels like AGI or superintelligence, Alexandre LeBrun, the CEO of AMI Labs – a startup developing a world model based on Yann LeCun’s concept – avoids these terms altogether. In an interview with TechCrunch, he said the company does not use such words.

“We never used the word AGI. And I noticed that no one uses it anymore; they moved on to superintelligence.”

– TechCrunch, interview

“Next time we will move on to something else.”

– TechCrunch, interview

“There is no clear definition. What is superintelligence? I don’t know. It’s not a very useful word.”

– TechCrunch, interview

This stance reflects the founder’s purposeful outlook as he sits at the heart of the latest AI race.

TechCrunch spoke with LeBrun during his time in Seoul at the International Conference on Machine Learning, where he was seeking local industrial partners, global companies, and researchers. AMI Labs currently does not have a product, but is actively engaging players from robotics, manufacturing, and electronics. The world model, which integrates physics for forecasting and interacting with the real world, must prove itself outside the lab.

According to LeBrun, the world model has the potential to significantly influence robotics. Currently, robots operate according to pre-programmed schemes, “completely static,” and artificial intelligence in the physical world remains “very primitive.”

“Currently, robots operate according to pre-programmed schemes, “completely static,” and artificial intelligence in the physical world is “very primitive”.”

– Alexandre LeBrun, interview

LeBrun emphasizes that world models do not replace LLMs; they complement one another. LLMs will remain the most effective for language processing, while world models will provide context and real understanding of the world.

The majority of industries that interact with the real environment may eventually adopt robotics based on world models; LLMs are currently weaker in understanding the physical context of the real world.

Current factory robotics handle repetitive actions well, but the real problem arises when a robot goes out into open space – home or on the street – and must understand its surroundings and act safely. “Robots aren’t safe right now,” he stressed; “there is no solution for this today.”

Asian Focus and Partnerships

LeBrun does not disclose the full Asian strategy, but notes the region’s appeal due to its strong industries in robotics, semiconductors, and manufacturing – sectors that AI has not yet fully captured.

The second reason is speed: Korea has an ambitious plan to invest in AI, data centers, and physical AI, and is also known for its rapid adoption of technologies – “Korea was the fastest at deploying the Internet 25 years ago,” says LeBrun. The combination of a strong industrial base and a willingness to quickly adopt AI makes the region unique for AMI.

“Korea was the fastest at deploying the Internet 25 years ago.”

– JP Lee, CEO of SBVA, one of AMI’s investors in Asia

According to data, the government of South Korea is actively funding local LLM models and plans to develop “physical AI,” but the local ecosystem will quickly overcome barriers thanks to the initiative of developers, including local players Naver and Kakao. Despite the name and scale of support, AMI currently has no product. The startup, founded jointly with Turing Award laureate Yann LeCun after his departure from Meta, has raised over $1.03 billion with an implied valuation of around $3.5 billion, but there have been no concrete releases yet. “We’ll surprise when we’re ready,” sums up Alexandre LeBrun.

Ultimately, LeBrun’s stance demonstrates AMI Labs’ pragmatic approach: the company focuses on real-world applications, not on scientific buzzwords. Their Asian strategy centers on partnerships with industry and developers, enabling rapid bringing-to-market of solutions that can integrate world models with real processes, while remaining flexible to future technological changes.