Sakana AI Highlights Royal Society Theme Issue on World Models

The concept the issue holds onto is a single one: the world model. It refers to the substrate by which a living thing or an AI takes the outside world in, predicts what happens next and acts on that prediction. The editors chose not to fix a single definition, and instead set causal, self-referential, individual goal-directed, collective and narrative forms side by side for comparison. Rather than settling the definition first and lining up the cases behind it, the editing lets the cases illuminate one another.

The author list of the lead article shows what kind of editing that is. Fourteen names appear, led by Adam Safron and Michael Levin, and they include David Ha of Sakana AI, Irina Rish of the Université de Montréal, David Krakauer and Melanie Mitchell of the Santa Fe Institute, Samuel J. Gershman of Harvard University and Joshua B. Tenenbaum of MIT. It is a list that mixes an AI company, neuroscience labs, a complexity institute and cognitive science groups in one paper. In the table of contents METAL checked, this paper sits in the introduction slot.

The paper compresses the issue's question into a single sentence. Adam Safron of Tufts University, the first author, and his co-authors write that a recurring theme is "the extent to which current AI systems trained on vast quantities of data learn the context-sensitive, temporally embedded, value-laden dimensions of world modelling that characterize diverse biological intelligences, or whether their impressive capabilities arise primarily from statistical surface regularities." They go on to say that "with this collection, we hope to clarify the conceptual landscape, identify key points of similarity and divergence between natural and artificial minds." It is a declaration to organise the language a verdict would need, rather than to deliver the verdict.

The eighteen papers do not lean one way either. Douglas Hofstadter contributes a piece asking whether there is an I in AI, and Gary Marcus co-authors a paper weighing whether a sentence is worth a thousand pictures. Stuart Russell takes on the representation complexity of model-based and model-free reinforcement learning, and Alison Gopnik examines experimentally how sensitive children and adults are to controllability and variability in their causal interventions. Skepticism, formal theory and developmental psychology experiments sit inside the same table of contents.

The issue, with Adam Safron and Michael Levin listed as guest editors, sets out three questions in its scope note. The first is "how intelligent are language models, and how might their capabilities evolve in the coming years." The second is how artificial systems might be given the adaptive intelligence of naturally evolved organisms, and the third is whether depending on learning from massive statistical associations fundamentally limits a system's ability to learn coherent models of self and world. All three aim not at capability but at what capability rests on.

Sakana AI pulled out three points from the issue. The first is that being able to do something and knowing something are different: what today's large models display may be the result of learning how words are arranged, and there are voices arguing that adding compute alone will not close that gap. The second is AI that knows itself, the finding that training a system to predict its own internal states tidies its internal representations and cuts waste. The third is that AI's hard problems lead into life's hard problems, with the prospect that AI research moves steadily closer to artificial life research.

Seen from the humanities side, a familiar distinction has come back. Someone who has memorised the whole map and someone who knows the way around the neighbourhood score the same on most tests, and they part ways when a road is blocked. That is exactly the point the issue keeps returning to, which is why it asks readers to look not at capability scores but at what capability stands on. The observation that biological intelligence does not merely receive the information it is given but reaches into its environment to learn the world is the ground for that demand.

World models are a subject METAL has followed. METAL has reported on research where adding the other party's beliefs as a variable lifted a world model's prediction score from 63 to 88. That study showed the same thing. Memorising more of the world and putting the other minds inside it into the calculation are different jobs, and when the second is missing, predictions become accurate in the wrong direction.

There are things to weigh while reading. The issue came out on May 14, 2026, and Sakana AI's introduction follows four months later. The same issue also carries one correction notice unrelated to world models, and since the editors said outright that they would not set a single definition, this collection does not end the argument. Sakana AI said it will continue its world model and physical AI research at its RSI Lab, and the fact that the company stands on one side of an academic argument is part of the picture. METAL has reported that this company was founded in Tokyo by an author of the Transformer paper.

One journal issue will not change the pace of an industry. Still, the problem remains that as capability rises the language for asking what that capability rests on grows stale, and this issue is the work of sharpening that language again. The distance between handling words and understanding the world has not closed, and an AI company, philosophers and developmental psychologists have sat down together on the side that is building the ruler for it. The yardstick that will judge the next generation of models is being made right there.