The AI industry is quietly pivoting. After three years of selling large language models (LLMs) as the path to general intelligence, a new consensus is emerging: these systems don't actually understand the world. They are masters of fluency, not comprehension.
To fix this, a new category of AI called 'world models' has attracted billions in funding. The goal is to build systems that don't just predict the next word, but internally represent how reality works-how objects persist, how actions have consequences, and how time passes.
The biggest signal came in March 2026, when startup Advanced Machine Intelligence, co-founded by Yann LeCun, raised over $1 billion. LeCun has long argued that LLMs lack true reasoning and planning, and that human-level intelligence requires grounding in the physical world.
This is not about abandoning LLMs. Companies like Meta, Google DeepMind, and Fei-Fei Li's World Labs are building world models trained on video, robotics, and simulation. Their goal is to move AI from generating plausible sentences to making reliable predictions about the physical world.
If successful, these systems could replace parts of human judgment-not just by writing better, but by understanding cause, effect, and context. For journalism and high-stakes decision-making, this represents a profound shift.
The key test will be in robotics, long-horizon planning, and the ability to verify what these models actually 'know.' Until then, the industry's claim of understanding remains a promise, not a reality.