AMI Labs bets on world models as the next step after LLMs, with VLA as its path to products

After the LLM boom, AMI Labs wants to shift AI's center of gravity from language to understanding the physical environment. Continuing a text is not enough for a machine that has to act safely in the real world, plan its steps, and assess the consequences of its decisions in advance.

AMI Labs is the research company of Yann LeCun, one of the key pioneers of deep learning. The project raised $1.03 billion at a $3.5 billion pre-money valuation, a sign that interest in world models has moved beyond academic debate.

The company starts from a simple premise: data from cameras, sensors, and other instruments is structured differently from text. It is continuous, noisy, and high-dimensional, and it does not fit the "predict the next token" logic well. So instead of adapting LLMs to every scenario, AMI is betting on a different foundational layer: the world model.

Here, a world model is not a video generator, nor simply a multimodal system that takes images, text, and actions as input. It is a model that builds a latent internal representation of an environment, picks out stable relationships, and discards incidental details. What matters is not every pixel of the next frame but the structure of what is happening: where objects are, how they move, what constraints the environment imposes, and what will change after the agent acts. Such an architecture has to answer not only "what do I see?" but also "what happens if I do this?"

That is why JEPA, the Joint Embedding Predictive Architecture, sits at the core of the approach. In this framework, the model predicts neither raw data nor a sequence of tokens, but a state in representation space. That lets it avoid spending compute on noise and random variation and learn from the features of a scene that actually matter.

The V-JEPA 2 research already offers a practical argument for this approach. The system was first pretrained on more than a million hours of internet video, and then an action-conditioned version was fine-tuned on less than 62 hours of unlabeled robotics video. After that, the model was able to operate Franka robotic arms zero-shot in new labs, picking up and moving objects without collecting environment-specific data and without a reward function.

But a world model on its own is not yet a finished agent. It can predict how a situation will unfold, but something still has to turn that understanding into a concrete action. That is where VLA comes in: a visual-language-action layer that connects perception, the user's goal, a language command, and the actions the system is allowed to take.

A key claim from AMI and related work is that VLA and world models do not compete. On the contrary, without an internal forecast, a VLA remains too reactive. It may produce the right action "for now," but it handles long, fragile, and physically sensitive scenarios less well, where the system needs to mentally play out the consequences of a touch, a movement, a collision, or a mistake.

That makes the most obvious markets for this approach not chat interfaces but industries where errors are costly: industrial automation, robotics, wearables, and healthcare. If a text model botches an article summary, the damage is limited. If an intelligent system misreads the state of equipment, misjudges risks in medicine, or miscalculates a robot's trajectory, the consequences are physical.

Tellingly, AMI's first named partner is Nabla, a digital health company. That does not mean AMI has already solved reliable AI for clinical settings, but it does show the direction: less focus on flashy demos and more on controllability, predictability, and internally simulating the environment before acting.

The main takeaway is that after the LLM era, the AI conversation is gradually moving from describing the world in language to modeling it internally. AMI's approach is still a research program, not a ready replacement for large language models: the term "world model" is already getting blurry, and transfer to new environments has yet to be proven. But if this line of work pays off, the next practical breakthrough in AI may come not from yet another chatbot, but from systems that first understand physical reality and only then act in it.