World model companies are keeping a lot of secrets

The AI sector has a new favorite buzzword: world models. If you spend any time in the venture capital circles of Menlo Park or the builder communities in San Francisco, you have heard it. The promise is massive. We are told these systems will move beyond simple text generation to actually understanding the physical laws of our reality. They are supposed to be the bridge between a chatbot that hallucinated your grocery list and a robot that can actually navigate a kitchen.

But there is a glaring problem. Everyone is raising hundreds of millions of dollars, yet nobody is willing to talk about the architecture, the data, or the actual utility. In the world model space, the prevailing culture is one of extreme secrecy. For founders and builders, this should be a red flag, or at least a signal to look closer at what is being sold.

The Definition Problem

In simple terms, a world model is an AI system that builds an internal representation of how the world works. Unlike a Large Language Model (LLM) that predicts the next word in a sentence, a world model tries to predict the next state of an environment. If you drop a glass, a world model knows it will shatter. If a car turns left, it understands the physics of momentum. It is essentially the software version of common sense.

The issue is that "world model" has become a catch-all term for anything that isn't a standard transformer. Because the term sounds profound and scientific, it has become a magnet for capital. We are seeing companies launch with billion-dollar valuations before they have even released a white paper. As a founder, you have to ask: are they keeping secrets because the tech is that revolutionary, or because the curtain hasn't been fully sewn yet?

The Data Black Box

Training these models requires a specific type of data that is much harder to scrape than the open internet. You need high-fidelity video, sensor logs, and telemetry data. This is where the secrecy starts to get messy. The companies supplying this data—often logistics firms, autonomous vehicle startups, or robotics labs—are signing non-disclosure agreements that would make the CIA blush.

We are seeing a trend where the supply chain of AI is becoming more opaque than the models themselves. If we don't know what these models are being trained on, we can't possibly know where they will fail. For builders trying to integrate these models into real-world applications, this lack of transparency creates a massive technical debt. You are essentially building your house on a foundation you aren't allowed to inspect.

Why Founders Should Be Skeptical

I have seen this cycle before in the early days of crypto and the first wave of autonomous driving. When the technical hurdles are the highest, the marketing becomes the loudest. The secrecy serves two purposes. First, it protects legitimate intellectual property in a hyper-competitive market. Second, it hides the fact that many of these systems are still struggling with basic spatial reasoning.

If you are a founder looking to build in this space, do not be blinded by the funding rounds. A 500-million-dollar seed round does not mean the physics engine is solved. It means the investors are terrified of missing out on the next paradigm shift. True world models will likely be built in the open, or at least with enough transparency that developers can stress-test the edges.

The Infrastructure Play

While the big labs play their games of shadow boxing, the real opportunity for builders is in the infrastructure. Who is cleaning the sensor data? Who is building the synthetic environments used for testing? These are the shovel-sellers in a gold mine that might actually just be a very expensive hole in the ground. By focusing on the tangible components of the world model stack, you avoid the valuation traps and the NDA-heavy culture of the top-tier labs.

We also need to consider the compute costs. Building a model that understands 3D space and time is orders of magnitude more expensive than training a text model. The secrecy might also be a way to mask the inefficiency of the current hardware. We are trying to simulate the universe on chips that were designed to render video games. There is a fundamental mismatch there that no amount of venture capital can immediately fix.

Takeaway for the Ecosystem

The transition from LLMs to world models is inevitable, but the path will be paved with failed startups and over-promised timelines. The current culture of silence among world model companies suggests that we are in the "magic trick" phase of the technology. The practitioners are showing us the prestige but keeping the trapdoor hidden.

For those of us on the ground building products, the goal should be utility over hype. If a world model provider won't tell you how they handle causality or what their data sources are, they aren't a partner; they are a vendor with an ego. Keep your eyes on the builders who are showing their work, even if their valuations aren't hitting the headlines yet. The future of AI isn't going to be solved in a locked room; it’s going to be solved by people willing to break things in public.

The most dangerous part of the AI race isn't the technology itself—it's the belief that secrecy is a substitute for a working product.

In short: watch the data, ignore the press releases, and wait for the code to speak for itself. If the world models are as powerful as they claim, they won't need to be kept secret for long. Their impact will be impossible to hide.

Read the original at TechCrunch AI →