The Rise of World Models: Europe Takes the Lead with AMI Labs

The newly founded European startup AMI Labs, which focuses on developing World Models, has just raised $1.03 billion. With a pre-money valuation of $3.5 billion, the company joins the ranks of the highest-funded AI projects in Europe.

The main goal of AMI Labs fundamentally differs from typical startups for applied AI. The team conducts basic research to create AI systems that comprehend the real world rather than just processing language.

What exactly are World Models?

To understand AMI Labs' vision, one must distinguish the concept of World Models from conventional language models, the Large Language Models (LLMs). While classic LLMs are trained to statistically predict which word comes next, World Models attempt to deeply understand the dynamics, physical laws, and spatial properties of our reality. A World Model creates an internal neural representation of its environment and uses this to simulate complex cause-and-effect relationships. This can be thought of as artificial imagination, allowing the system to mentally play out the consequences of actions before they are executed in the physical world.

Text alone is an insufficient basis for grasping causality and the physics of the real world. That is why language models often fail at tasks requiring physical understanding due to a lack of implicit knowledge.

Because World Models possess an anchored understanding of physical reality, they are considered a promising solution to the problem of hallucinations. Deterministic reliability is essential, especially in safety-critical areas such as healthcare, where faulty AI responses could have life-threatening consequences. For robotics and autonomous vehicles, World Models open up entirely new possibilities.

What points to the growing importance of World Models

The recently completed funding round of AMI Labs amounting to $1.03 billion at a pre-money valuation of $3.5 billion (originally only €500 million was targeted) indicates strong investor confidence in basic research beyond classic language models. This is particularly remarkable because AMI Labs does not pursue short-term revenue targets and it could take years before theoretical concepts become commercial applications.

The capital gives the company the necessary leeway to tackle the two biggest cost factors in AI development: computing power and recruiting top talent in locations like Paris, New York, Montreal, and Singapore. The involvement of high-profile investors and strategic partners such as Bezos Expeditions, Eric Schmidt, Mark Cuban, as well as NVIDIA and Samsung, illustrates that the tech industry is willing to pump significant funds into this still-nascent technology.

What else points to the imminent rise of World Models

Besides the success of AMI Labs, there are numerous other indicators that World Models could be considered the next big breakthrough in AI development:

• Other startups founded by prominent scientists in this field are also currently attracting massive capital. For instance, Fei-Fei Li's startup World Labs recently also secured $1 billion, while the European company SpAItial received an unusually high seed funding of $13 million.
• Competition is intensifying as industry giants like NVIDIA are already working on their own World Models for physical AI. Existing successes like Google DeepMind's GraphCast system, which beats traditional supercomputers at weather forecasting, are also de facto considered World Models for the Earth's climate.
• Similar to how GPT is a Foundation Model for text, research is already working on Foundation World Models for robotics and physics, such as the Octo or RT-2 projects, which are trained on thousands of different robots to acquire universal knowledge about physical manipulation.
• The technology is already being adapted in highly complex areas. Startups like Wayve and Waabi use end-to-end World Models for autonomous driving to safely train rare and dangerous traffic situations in a neural simulation. In life sciences, so-called Virtual Cells are also being developed, with which the reaction of cells to drugs can be simulated on a computer.

As long as AIs do not understand the world causally and physically like a toddler, it will remain limited in its analytical capabilities. World Models could prove to be the missing puzzle piece on the path to Artificial General Intelligence (AGI).