World Model Companies Are Sitting on Breakthroughs But Won't Talk About Them
Key takeaways
- Multiple world model startups have raised hundreds of millions in funding while remaining almost completely opaque about their technical progress and capabilities
- Data partnerships and investor agreements are being used to justify keeping research private, limiting broader AI research community's ability to contribute or verify claims
- The secrecy strategy works as a competitive advantage only temporarily, with evidence of success eventually leaking through customer results, partnerships, or published papers
There's a peculiar phenomenon happening in the AI industry right now. A handful of startups focused on building world models, the kind of AI that can simulate complex environments and plan actions within them, have raised extraordinary amounts of funding. They've generated massive amounts of buzz in research communities and venture capital circles. Yet almost nobody outside those circles actually knows what they're building or how close they are to their goals.
This secrecy isn't accidental. It's deliberate strategy, and it's becoming increasingly frustrating for researchers, competitors, and the general public trying to understand where AI is actually heading.
World models are genuinely different from the large language models that have dominated recent AI discourse. Instead of predicting the next token in a sequence, world models aim to build internal representations of how physical systems work. Train one on robot videos and it learns the mechanics of motion. Train another on factory floor footage and it understands manufacturing processes. The idea is that once you have an accurate model of how the world works, you can use it to plan and reason about complex, multi-step problems.
The potential applications are enormous. Manufacturing, robotics, autonomous vehicles, logistics, scientific discovery, drug design. If someone cracks genuine world models, they won't just have built a cool AI research project. They'll have built something with real economic value across dozens of industries.
Which is probably why funding has absolutely flooded into this space. Multiple startups have raised hundreds of millions of dollars to work on world models. The cash is there. The hype is real. But the actual outputs remain almost completely opaque.
TechCrunch recently investigated this mystery, trying to get world model companies to explain what they're actually working on. The responses were remarkably consistent: we can't tell you, we're under non-disclosure agreements with our data partners, our investors want us to keep it quiet, the research isn't ready to share yet. One company declined to even describe their approach in general terms.
This creates a genuinely weird situation. We have companies that have raised hundreds of millions of dollars to solve one of the most important problems in AI, but the broader AI research community can't actually see their progress or contribute to it. You can't build on work you don't know exists. You can't spot mistakes if you're not allowed to look. You can't verify claims if there's no way to test them.
The funding sources for training data add another layer of secrecy. Some of these companies have made exclusive deals with data providers, meaning the training data itself is locked away. Others are working with companies that have legitimate confidentiality concerns about their proprietary processes. That confidentiality might be justified, but it also makes it almost impossible to audit whether the models are actually working as claimed or if they're overselling their capabilities.
There's also a timing incentive here. If you're racing to solve world models before your competitors, publishing your methods isn't appealing. You want to maintain your lead for as long as possible. Fair enough from a business perspective, but it's not great for the field as a whole.
The long-term problem is that this kind of secrecy only works until it doesn't. Eventually, if these models are actually working, evidence will leak. Customers will talk about results. Partnerships will become visible. Papers will be published, either by the companies themselves or by researchers they've worked with. If world models turn out to be less impressive than the hype suggests, that will also eventually become obvious.
What we're really watching is a test of how much value secrecy can actually provide in AI research. In past eras, the answer was probably quite a lot. Trade secrets mattered. But in a field where so many people are working on similar problems, where open source implementations exist, where talent is highly mobile, sustained secrecy might be harder to maintain than these companies think.
