While much of the artificial intelligence sector is focused on achieving and labelling its creations as 'AGI' (Artificial General Intelligence) or 'superintelligence', Alexandre LeBrun, CEO of AMI Labs, is taking a starkly different approach. LeBrun, who leads Yann LeCun's world model startup, has stated that his company deliberately avoids these terms, finding them poorly defined and ultimately unhelpful in the pursuit of advanced AI.
LeBrun, speaking in Seoul where he was scouting for industrial partners, global companies, and researchers, highlighted his company's focus on 'world models'. Unlike Large Language Models (LLMs) that predict the next word or text, world models are designed to predict the next physical state of the world. He used the analogy of nudging a glass off a table: the intuition that it will tip and spill is what a world model aims to capture. This understanding of physics and real-world interaction is crucial for applications beyond language processing.
AMI Labs, currently in its pre-product phase, is actively engaging with players in robotics, manufacturing, and electronics. LeBrun emphasised that world models must prove their capabilities outside laboratory settings. He pointed out that current robots often operate on fixed routines, lacking true awareness of their surroundings. Integrating context-aware AI, even at a basic level, could bring about significant improvements in robot safety and functionality, preventing incidents like a robot accidentally harming a child at a public event.
LeBrun clarified that world models are not intended to replace LLMs but rather to complement them. He drew a parallel to the human brain, where distinct functions handle language and reasoning. LLMs remain highly efficient for language processing, while world models are positioned to provide the essential context and understanding of the physical world that LLMs currently lack. This complementary approach suggests a future where AI systems combine both linguistic and physical intelligence.
The potential impact of world models extends across virtually any industry that interacts with the physical world. LeBrun noted that while factory robots performing repetitive tasks are effective today, the real challenge arises when robots need to operate in dynamic, open environments such as homes or streets. Ensuring safety in these settings is a major hurdle, and LeBrun believes world models offer a crucial part of the solution. His previous experience with Nabla, an AI health startup, also informs his view that current AI only scratches the surface of healthcare, with real-world experience being paramount – an area where world models could contribute significantly.
To train these sophisticated world models, AMI Labs requires access to real environments and close collaboration with partners. LeBrun indicated that this need for real-world data and industrial infrastructure is a key factor in his engagement with regions like Asia, particularly South Korea, which boasts advanced industries in robotics, semiconductors, and manufacturing – sectors that were less impacted by the initial wave of AI development and offer the speed necessary for rapid innovation.