Luma’s Amit Jain Says Most ‘World Models’ Lack the Intelligence to Earn the Name

The Luma AI CEO argues a true world model pairs language-level reasoning with physics, calls 3D-first approaches a fool's errand and says Luma is moving to agentic world models.
Luma’s Amit Jain Says Most ‘World Models’ Lack the Intelligence to Earn the Name
Luma’s Amit Jain Says Most ‘World Models’ Lack the Intelligence to Earn the Name

Luma AI CEO Amit Jain thinks most of the companies calling their products “world models” are building the wrong thing. A real world model, he argues, needs the reasoning of a language model plus an understanding of physics, and making a video generator interactive doesn’t get you there, The Information reported.

Jain laid out the argument in an interview with reporter Rebecca Bellan, captured in a published transcript. Asked what’s missing from current world models, he answered: “Bigger issue is they don’t have any intelligence.” He called the efforts at World Labs, Runway and similar companies “very lazy attempts at trying to make a video model interactive.”

Language plus physics

For Jain, speed and interactivity are beside the point. “What we need is world understanding and language understanding together. Even if it is slow, even if it is not interactive, that makes it a world model,” he said. An image model with a strong grasp of “physics, of architecture, of physical motion, of causality, of long range understanding” would count as more of a world model, in his view, than a navigable 3D scene.

He’s skeptical of approaches built on 3D data, even though Luma started out in 3D. “The lesson we learned, there is no 3D data in the world,” he said. Video is the opposite: billions of phones record the physical world every day. Citing Rich Sutton’s “bitter lesson,” the idea that general methods that scale with compute and data win over specialized ones, Jain called specialized 3D and 4D approaches “a fool’s errand.”

Running out of text

Data supply is a big part of his case. LLMs are being trained on 15 trillion to 20 trillion tokens, and “all of humanity’s text,” Jain said, comes to barely 30 trillion. Language models capture human logic but only “our record” of the universe. “An LLM can describe how to swim, but it cannot drive a robot that can swim,” he said. Training one model on text, audio, video and images together, instead of bolting vision onto an LLM, would produce “AI systems that are substantially more powerful than what we have now.”

Agents next

Jain said video generation on its own is effectively “done,” and Luma’s next step is “intelligent, agentic world models”: systems that take a full task, such as making a 30-second ad, then generate, reject and refine their own work. Humans stay in the loop. “These models cannot read your mind yet,” he said. “The more context you share, the better it is.”

Subscribe for Updates

AITrends Newsletter

The AITrends Email Newsletter keeps you informed on the latest developments in artificial intelligence. Perfect for business leaders, tech professionals, and AI enthusiasts looking to stay ahead of the curve.

By signing up for our newsletter you agree to receive content related to ientry.com / webpronews.com and our affiliate partners. For additional information refer to our terms of service.

Notice an error?

Help us improve our content by reporting any issues you find.

Get the WebProNews newsletter delivered to your inbox

Get the free daily newsletter read by decision makers

Subscribe
Advertise with Us

Ready to get started?

Get our media kit

Advertise with Us