TechCrunch is reporting that companies working on so-called world models — AI systems designed to simulate and predict the behavior of physical and digital environments — are operating with an unusual degree of secrecy, even by the standards of a notoriously tight-lipped industry. According to the outlet, founders are deflecting questions, and even the companies supplying these startups with data appear reluctant or unable to describe what the end product actually looks like.
To understand why this matters, it helps to understand what world models are and why the technology has attracted so much attention so quickly. A world model, in the sense that researchers and investors now use the term, is an AI system that builds an internal representation of how the world works — not just pattern-matching on text or images, but learning the underlying rules of physical systems, cause and effect, and the way events unfold over time. The ambition is enormous. If a system can accurately simulate reality, or some meaningful slice of it, the downstream applications range from robotics and autonomous vehicles to drug discovery, climate modeling, and beyond. Yann LeCun at Meta has been the most prominent public advocate for this architectural approach, arguing for years that it represents the path toward something closer to general machine intelligence than large language models alone can achieve. That kind of intellectual sponsorship, even when it comes from outside the startups themselves, has a way of drawing capital.
And capital has arrived. The world-model space has accumulated substantial funding in a short period, with several companies commanding significant valuations before releasing any product that outside observers can meaningfully evaluate. That combination — serious money, serious hype, and a near-total absence of verifiable technical claims — is precisely what makes TechCrunch's reporting worth sitting with.
Secrecy in AI is not new, of course. OpenAI's name has become something of a running industry joke given how little of its research now appears in the open literature. Google DeepMind, Anthropic, and others have all been criticized for publishing less than they once did as competitive stakes have risen. But the secrecy TechCrunch describes here has a different texture. It is not simply that these companies are declining to publish research papers or withhold model weights. It is that even the people doing business with them — the data suppliers, the infrastructure partners — apparently cannot or will not characterize what is being built. That suggests either that the technology is at a stage where its nature is genuinely unclear even to close observers, or that the companies have constructed unusually aggressive confidentiality arrangements, or both.
The likely reading is that both dynamics are at play simultaneously. World models, as a technical category, are still loosely defined enough that two companies could describe themselves as building one while pursuing quite different architectures with quite different goals. That ambiguity is not purely a PR strategy — it reflects genuine uncertainty in the research community about the best approaches. At the same time, the competitive incentive to say nothing is obvious. If the core insight of your company is a novel way to train a simulation of physical reality, describing it in enough detail for outsiders to evaluate it is also describing it in enough detail for a well-resourced competitor to replicate it.
The consequences of this opacity fall on several groups in different ways. Investors face the most immediate exposure. Funding a company whose technical approach cannot be independently assessed is a bet on founders rather than on any verifiable progress, and the history of AI hype cycles offers no shortage of cautionary examples. Enterprise customers who might eventually deploy these systems will struggle to conduct meaningful due diligence. Policymakers and researchers trying to anticipate what capable world-modeling systems might do — and what risks they might introduce — are working largely blind. And the broader AI research community, which has historically benefited from the rough norm that significant capability advances get written up and shared even when the code does not, is being cut out of a technical conversation that could turn out to be important.
There is also a subtler risk for the companies themselves. Secrecy works as a strategy when it delays competitors. It becomes a liability when it delays trust. If and when world-model companies need to persuade regulators, enterprise buyers, or the public that their systems behave as claimed and that safety has been taken seriously, a track record of offering nothing verifiable to anyone will not help them make that case.
What to watch for is fairly clear. The moment any of these companies releases a product, a benchmark result, or a research artifact that outside parties can actually evaluate, the gap between the accumulated hype and the demonstrable reality will become visible. Watch also for whether data suppliers or infrastructure partners eventually speak on the record — when commercial relationships end or change, confidentiality obligations often weaken. And watch the fundraising rounds. The terms and valuations attached to the next wave of investment will signal whether the people with the most information, the lead investors doing technical due diligence, have found something underneath the secrecy that justifies the noise.
