Recently, I had the opportunity to lead a discussion on world models at the All In conference, distinct from the similarly named podcast. This experience allowed me to explore one of the more enigmatic sectors of artificial intelligence. The prominent entities in this field include Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs. While both organizations have garnered significant attention and investment, their focus on immediate revenue generation is relatively low.
World models fundamentally aim to automate spatial understanding, with immense potential for diverse and profitable applications, ranging from robotics and interactive media to advanced self-driving technologies.
However, when I sought to understand how these technologies might be commercialized, the responses were vague. Michael Rabbat, co-founder of AMI Labs and its VP of World Models, who joined me on the panel, was somewhat reserved regarding the company’s initiatives. He stated, “We’ll share details when we’re ready.” In a follow-up email, he mentioned, “We are still in a research and development stage, and as such, we aren’t ready to disclose any product plans or timelines.”
It’s worth noting that AMI is a relatively new entity, having been established less than a year ago, which justifies their discretion. Nonetheless, this level of reticence seems to be pervasive across the world-modeling sector. World Labs’ Marble stands as one of the more advanced offerings in this arena, showcasing applications from straightforward content creation to crafting interactive environments for video games and CGI effects. While there are applications in robotics, the platform appears primarily aimed at illustrating its capabilities.
This secrecy extends to the companies’ partners as well. At the same conference, I spoke with Alex de Vigan, CEO of Physicl, which provides data to the emerging world model sector. De Vigan expressed that while he is aware that Physicl’s data has been useful, he remains uninformed about the specifics of the projects in question. “I wish they would share more details. Knowing their work could help us enhance our data offerings,” he remarked.
The ambiguity largely stems from the flexible nature of world models. At its simplest, a world model serves as a navigable representation of the environment, akin to the AI models driving autonomous vehicles. The techniques utilized for navigating traffic by companies like Waymo could similarly aid humanoid robots with tasks or convert brief video clips into interactive spaces. AMI has already dipped into several sectors including manufacturing, biomedical applications, robotics, and even AI software for healthcare via a collaboration called Nabia. It’s unclear which of these avenues they will ultimately pursue — perhaps some will stand out more than others.
There is widespread consensus regarding the potential for various business opportunities stemming from world model technology. In an environment where fundraising is relatively straightforward, there’s little incentive to concentrate on a single path. In fact, there’s valid reasoning against it. If AMI were to announce a breakthrough like a humanoid robot or a cutting-edge film rendering technology, it could pique the interest of competing labs, as well as companies in the broader tech landscape, including OpenAI and Anthropic.
This situation highlights the dual-edged nature of easy funding; while it fosters innovation, it also enables competitors to pursue similar financing. The same financial resources that allow for discreet development also support potential challengers once a clear market direction is established. Even though competition may be unavoidable, it is beneficial to postpone it for as long as feasible, hence the inclination to maintain confidentiality regarding their projects.
Fans of Cixin Liu’s work might recognize this as a “dark forest” scenario: if one is unaware of others lurking in the shadows, it is wise to minimize attention.
This article was originally published on September 18, 2026.
