Why World Model Companies Are Keeping So Many Secrets

World models — systems trained to simulate environments, physics and consequences rather than just predict the next token — have become the most talked-about frontier in artificial intelligence. They power interactive video generation, robotics simulation and agents that plan before acting. But for a field attracting billions in funding, remarkably little is publicly known about how these systems actually work.

Compare the situation to the early days of large language models. Even as labs grew more guarded, researchers could still read architecture papers, inspect benchmark scores and reproduce approximate results. World model companies have skipped that phase almost entirely. Demos arrive as polished video reels; technical reports, when they exist, often omit parameter counts, training corpora, compute budgets and evaluation methodology.

What exactly is being withheld

The gaps cluster into a few consistent areas:

• Training data provenance. Interactive world models require enormous volumes of video, gameplay footage and sensor logs. Few companies specify where that material came from or whether it was licensed.
• Evaluation standards. There is no agreed benchmark for whether a model has learned physics, object permanence or long-horizon consistency. Firms grade themselves using internal metrics that outsiders cannot replicate.
• Failure modes. Cherry-picked clips rarely show what happens when a simulation drifts after thirty seconds, or when an agent trained in the model transfers poorly to the real world.
• Compute and cost. Inference for real-time interactive generation is expensive, and disclosure would reveal how far current products are from commercial viability.

The reasons behind the silence

Some secrecy is straightforwardly commercial. World models sit close to several lucrative markets — gaming, film production, autonomous vehicles, industrial robotics — and a rival with architectural details could close a lead quickly. Investors backing these companies at high valuations have little incentive to encourage openness.

Legal exposure matters too. Video data raises sharper copyright and privacy questions than text scraped from the open web, and identifying sources invites litigation. Several ongoing cases against generative media firms have made legal teams considerably more conservative about what engineers may publish.

There is also a safety argument, offered sincerely by some researchers and opportunistically by others. Systems that model the physical world and plan within it could, in principle, be misused. Whether that justifies withholding evaluation results — as opposed to model weights — is contested.

Why the opacity is a problem

Without shared benchmarks, claims about capability cannot be checked. A company can assert that its model "understands physics" while offering no way to test the statement, which makes it difficult for customers, regulators and even rival researchers to distinguish genuine progress from impressive rendering. Robotics teams deciding whether to train policies in a simulated environment are effectively buying on faith.

The academic cost is real as well. Universities cannot match the compute budgets of frontier labs, so their contribution depends on access to published methods. When those dry up, independent scrutiny weakens and the field narrows to a handful of well-funded actors.

A few groups have begun releasing smaller open world models and evaluation suites, and pressure from procurement teams and emerging disclosure rules may push others toward partial transparency. For now, though, the most consequential claims in the field remain largely unverifiable — and the companies making them appear comfortable with that.