As the artificial intelligence (AI) race accelerates toward advanced machine intelligence (AMI), tech giants are pushing boundaries to create agents that can reason more like humans. A key piece of this puzzle is physical reasoning, essential for AI systems to navigate and act within the real world.
Meta Platforms (META) has stepped into this frontier with the launch earlier in June of its Meta Video Joint Embedding Predictive Architecture 2 (V-JEPA 2), a powerful video-based world model that helps machines anticipate how the physical world behaves. Unlike language-only models, V-JEPA 2 learns from video data to predict object dynamics and human interactions using latent-space simulation. It allows robots to “think before they act,” performing tasks like picking up and placing objects in unfamiliar environments.
