Inside the ‘Dark Forest’ of AI World Models: Why Leading Labs Are Keeping Their Commercial Plans Under Wraps

This week, moderating a panel on world models at the All In conference—unrelated to the popular podcast of the same name—provided a rare opportunity to dig deeply into one of the most mysterious and heavily funded corners of the artificial intelligence landscape. The premier players driving this specific technological frontier are Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs. While both ventures have accumulated massive amounts of industry buzz and staggering sums of venture capital, they currently rank remarkably low on the traditional scale of trying to make money or establishing immediate commercial revenue streams.

At their foundational core, world models are designed around automating spatial intelligence. Because of this versatile capability, the field could theoretically head in a multitude of exciting and highly lucrative directions, ranging from advanced industrial and humanoid robotics to interactive video generation and increasingly complex autonomous driving systems.

However, when pressed on where the technology will actually see real-world commercialization, the outlook quickly grows foggy. The closest thing to an authoritative voice on the matter is Michael Rabbat, a co-founder of AMI Labs and the company’s vice president of world models, who joined the panel discussion. When questioned directly about what specific products the company is actively developing, Rabbat remained notably cagey. “We’ll talk about it when we’re ready to talk about it,” he stated on stage. Later, in an email elaboration, he clarified that the organization is still firmly situated in a research and building phase, meaning they are intentionally avoiding public discussions regarding any concrete product plans or deployment timelines.

To be fair, AMI Labs is less than a year old, making a period of strategic quietness standard practice in early-stage deep tech. Yet, this sort of tight-lipped secrecy extends far beyond a single startup, permeating the entire world-modeling space. World Labs’ Marble platform stands as perhaps the most fully developed product in the arena, with demonstrations ranging from straightforward media creation and building explorable environments for video games to sophisticated CGI special effects. While there are certainly robotics use cases being explored behind the scenes, the platform’s public-facing demonstrations often feel more heavily tailored toward proving sheer technological capabilities rather than driving a focused commercial enterprise.

This pervasive secrecy even extends outward to the supply chain supporting these ambitious labs. On the sidelines of the same conference, Alex de Vigan, the CEO of Physicl—a data supplier operating within the burgeoning world model ecosystem—noted the information blackout firsthand. He acknowledged that he knows Physicl’s specialized data has proven useful for whatever these major labs are currently building, but he remains completely in the dark regarding the final destination or application of his own product. “I wish they would tell us more. We could build more useful data if we knew what they were working on,” de Vigan remarked.

A significant portion of this prevailing mystery stems from the inherent versatility of world models as an overarching concept. At its simplest iteration, a world model functions as a navigable, predictive map of the physical environment, conceptually similar to the predictive AI architectures that currently assist self-driving cars in navigating unpredictable roadways. However, that exact same underlying modeling approach—which helps an autonomous vehicle weave seamlessly through heavy urban traffic—could theoretically also teach a humanoid robot how to safely lift and carry heavy boxes in a warehouse, or transform a few minutes of standard video footage into a fully interactive, explorable virtual environment.

AMI Labs has already dipped its corporate toes into a surprisingly wide array of industries, including manufacturing, biomedicine, robotics, and even AI software tailored for doctors through its ongoing partnership with Nabia. It is practically a given that the company will not actively pursue all of those distinct verticals simultaneously, raising questions about which specific applications are ultimately rising to the top.

Despite the ambiguity, virtually no one in the tech industry doubts that there are numerous viable, highly profitable businesses waiting to be built on top of world model technology. As long as these pioneering labs can easily secure monumental rounds of fundraising, there is little to no immediate market pressure forcing them to narrow their focus onto a single, definitive product. In fact, seasoned founders argue there is good strategic reason to maintain this broad ambiguity.

Consider the alternative scenario: if AMI Labs were to announce tomorrow that it had successfully built a commercially viable humanoid robot or a next-generation Hollywood rendering engine, a multitude of competing labs would instantly pivot their resources toward that exact same market. Almost overnight, the pioneering lab would find itself facing intense, aggressive competition from rival world-model startups, well-funded neolabs, and tech giants like OpenAI and Anthropic.

In many ways, this dynamic represents the flip side of easy fundraising. The same venture capital ecosystem that allows a startup to research and build comfortably under the radar is simultaneously well-positioned to fund a dozen potential rivals the exact moment a clear path to market is illuminated. Even if that eventual market competition is ultimately inevitable, common industry logic dictates that it is best to delay that rivalry for as long as humanly possible, which naturally means keeping quiet about the exact nature of what is being forged in the lab.

Fans of speculative science fiction, particularly the work of author Cixin Liu, will readily recognize this dynamic as a classic dark forest scenario: when you are operating in uncharted territory and do not know who else might be lurking in the woods, the safest and most prudent strategy is simply not to attract any unnecessary attention.

This post was first published on September 18, 2026. When you purchase through links in our articles, we may earn a small commission, which does not affect our editorial independence.

Russell Brandom has been covering the tech industry since 2012, with a primary focus on platform policy and emerging technologies. He previously worked at The Verge and Rest of World, and has contributed written reporting to Wired, The Awl, and MIT’s Technology Review. He can be reached via email at [email protected] or securely on Signal at 412-401-5489.