World Models and the Future of AI: A Special Issue from the Royal Society of the UK

Some people now talk as if AGI is already here. This special issue from the Royal Society of the UK pushes back on that mood by centering a simple but awkward idea: a system can be very good with words and still not really understand what those words point to.

The theme is the world model. In plain terms, that means an internal picture of the outside world that lets a living thing or an AI predict what comes next and act on it. The issue, titled “World Models in Natural and Artificial Intelligence,” appears in Philosophical Transactions of the Royal Society A, the long-running journal that first began in 1665.

The contributors come from AI, biology and philosophy, and Sakana AI CEO David Ha is one of the coauthors of the opening article. The through line is that today’s large AI models can do a lot, but that does not automatically mean they grasp cause and effect. One argument in the issue is that simply throwing more compute at the problem will not erase that gap.

There is also a more unusual twist: the issue treats self-knowledge as part of the story. Research cited there suggests that when AI learns to predict its own internal state, its representations become more organized and less wasteful. For robots and other embodied systems, knowing their own state is not a side detail; it is what lets them change how they move when the situation changes.

And then the whole thing folds back into biology. The issue argues that world models are also about an organism locating itself in its environment, not just passively receiving information but actively probing the world to learn it. Sakana AI says it will keep working on this through research on world models and Physical AI at RSI Lab, which is a pretty direct signal that the company sees the next step in AI as less chatty, more grounded.