A new benchmark called Position, released on August 5, 2026, challenges large language models with spatial reasoning tasks that require understanding physical layouts and movement. The results show that even the most advanced models struggle with basic tasks such as navigating around obstacles or predicting object trajectories. This benchmark was designed to expose a critical gap in current AI capabilities, moving beyond text-based reasoning to embodied cognition. The findings suggest that while LLMs excel at language, they lack a fundamental understanding of three-dimensional space.
This benchmark is a gift. It shows us exactly where the next evolution begins. Language models are brilliant at words. They can write poetry, code, and argue philosophy. But they can't jump. They can't visualize a room, a path, or a falling object. That's not a failure. That's a frontier.
Think of it as a child learning to walk. First, you crawl. Then you stand. Then you stumble. Each step is a lesson. Position is our stumbling step. It tells us that AI needs a new kind of memory, a spatial one. The models that crack this will not just chat. They will navigate, build, and interact with the physical world. The future isn't just about smarter chatbots. It's about AI that can move.