World model companies are keeping a lot of secrets

The world models space has become the startup industry's most secretive corner, even as it attracts billions in venture capital and generates relentless hype around the next frontier of AI.

Companies building world models, systems that learn to simulate and predict how the physical world works, operate under a veil of secrecy that rivals classified government projects. Founders refuse to disclose technical details. Investors stay quiet about deal terms. Even data suppliers, the companies providing training material, sign non-disclosure agreements that prevent them from discussing their involvement.

The silence reflects several realities. First, the competitive advantage in world models depends on technical breakthroughs that remain hidden. Companies like Waymo, Tesla, and dozens of stealth-mode startups believe their world model architectures and training methods represent genuine moats. Once those details leak, competitors can copy the approach or improve upon it. Protecting IP through secrecy beats publishing papers that earn academic credibility but destroy commercial advantage.

Second, the field remains unproven. No company has shipped a world model capable of real-world robotics or autonomous vehicle deployment at scale. The technology works in controlled lab settings and simulations, but the gap between demo and product remains vast. Founders know that premature technical disclosure invites skepticism. Better to stay quiet until launch.

Third, funding dynamics reward mystery. Investors pour money into world models companies without clear revenue models or timelines to profitability. The narrative around "next-generation AI" sustains valuations. Detailed technical disclosure could puncture that narrative if results disappoint. Vagueness allows each stakeholder to project their own optimistic assumptions onto the technology.

The secrecy extends to data sources. Companies training world models need massive amounts of video, robotics footage, and sensor data. Their suppliers sign strict contracts that prevent them from confirming which companies they work with or what data they've provided. This creates an opaque supply chain where even researchers trying to understand the field cannot trace which datasets power which models.

This secrecy creates inefficiencies. The field cannot benefit from open collaboration or peer review. Researchers working on similar problems independently duplicate efforts. Data suppliers cannot learn from each other about best practices for capturing training material. Investors make bets based on incomplete information and founder claims that cannot be independently verified.

Some founders argue secrecy is temporary. They contend that once world models reach commercialization, the field will open up. Papers will get published. Benchmarks will emerge. The community will coalesce around standards. Until then, they maintain, keeping cards close to the chest protects nascent technology from premature competitive pressure and public criticism.

But the current phase of radical secrecy cannot last forever. As companies spend years and billions on world models development, pressure will mount to show progress. Employees will leave for competitors and discuss what they learned. Data breaches will expose training procedures. Regulators will demand transparency around AI systems trained on proprietary datasets. The wall of silence cracks eventually, but for now, world models remain the startup world's most closely guarded secret.