World model companies are keeping a lot of secrets

Russell Brandom’s TechCrunch dispatch from the All In conference puts a spotlight on world models—the push to automate spatial intelligence for robotics, interactive video, and more ambitious driving systems. The headline names are Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs: heavy buzz, serious fundraising, and almost no public product roadmap.

When Brandom pressed AMI co-founder and VP of World Models Michael Rabbat on commercialization, the answer stayed fog: talk when ready; still research and building; no public product plans or timeline. World Labs’ Marble is the most visible artifact in the space, with demos spanning media creation, game-ready environments, CGI, and robotics—yet Brandom reads the platform as capability theater more than a clear go-to-market bet.

We’re covering this because spatial models sit on the path to useful machines in the physical world. Secrecy can protect a young lab. It can also starve partners, delay scrutiny, and turn the whole category into a fundraising dark forest.

Our read: the quiet isn’t just early-stage manners. It’s strategy. AMI is less than a year old, so Rabbat’s cageyness is defensible on paper. The pattern Brandom maps across the category is less defensible as a public norm. If world models are as versatile as the pitch—self-driving-style maps, humanoids moving boxes, minutes of video turned into explorable spaces—then every concrete product claim invites copycats with the same easy capital. Delay the reveal, delay the herd.

We’re for the useful AI here when it ships. Better spatial intelligence could mean safer warehouses, better simulation for science and entertainment, and robots that don’t smash into walls. Marble’s demos gesture at that. AMI’s toe-dips into manufacturing, biomedicine, robotics, and doctor-facing software via a Nabia partnership gesture at range. Range without a declared priority is also a tell: as long as money is cheap, focus is optional.

Name the downside. Physicl CEO Alex de Vigan told Brandom he knows his company’s data helps—and still doesn’t know what it’s helping. “I wish they would tell us more,” he said; better data follows clearer goals. That’s a real cost of the dark-forest play Brandom invokes via Cixin Liu: suppliers guess, customers wait, outsiders can’t audit claims, and the first lab to name a humanoid OpenClaw or a Hollywood rendering stack risks waking OpenAI, Anthropic, the neolabs, and peer world-model shops at once.

I’m watching whether AMI ever leaves the “when we’re ready” loop with a product and a date—and whether Marble graduates from capability demos to something buyers can buy without a science-fair framing. Useful spatial AI earns our benefit of the doubt. Permanent secrecy while fundraising does not.

Brandom moderated a world-models panel at All In (unrelated to the podcast) and followed up with Rabbat by email. The piece was first published September 18, 2026, and carried on TechCrunch dated September 20. The competitive pressure he describes is straightforward: world-model funding that buys stealth also funds rivals once a path to market is public.