The world model ambition growing inside a SG AI video startup

Wei Liu wants to predict the future. Not by being a fortune teller, but by teaching AI models to predict what happens next in a video.

His ambitions stretch beyond just creating another AI video generation platform. Video Rebirth, his Singapore-based startup, wants to build a world model, which refers to an AI system that simulates and predicts how the real world works.

“Without video generation, we cannot build a world model,” Liu tells Tech in Asia.

Video Rebirth co-founder and CEO Wei Liu / Photo credit: Video Rebirth

Startups like Yann LeCun’s AMI Labs and Fei-Fei Li’s World Labs have secured billions of dollars in their pursuit of world models, which are often described as the foundation for the next generation of AI systems that can reason, plan, and act in the real world.

For Video Rebirth, which has only raised US$80 million from AMD Ventures and Hyundai, the thesis revolves around building a world model through video, rather than through 3D reconstruction (World Labs) or abstract reasoning (AMI Labs).

Liu believes this is the path to building foundational infrastructure that video game companies, robotics firms, and autonomous vehicle manufacturers could rely on to simulate the real world at scale. But whether video generation can lead to genuine world understanding remains unproven.

A broader take on world models

Before co-founding Video Rebirth in late 2024, Liu spent more than eight years at Tencent, where he led the team that built the Hunyuan foundation model series, which has since been integrated into products like WeChat.

When Video Rebirth was established, world models were still a niche research concept. Google had only formed a new team dedicated to developing world models several months later.

Difu Li, Video Rebirth’s co-founder and chief strategy officer, notes that Singapore offered easier access compute resources. It currently has 32 people split between the city-state and Hong Kong, roughly one-third of whom are AI researchers.

Earlier this month, Video Rebirth launched its first product, Bach, targeting producers, filmmakers, and game studios. Unlike most platforms that produce single, isolated clips, Bach can create a multishot video of up to 30 seconds in a single run.

Bach allows users to create a multishot video of up to 30 seconds in 1080p resolution / Image credit: Video Rebirth

Bach is also used in-house so the team can test its proprietary architecture for the future world model product.

Becoming the “Anthropic of video generation”

Real-world messiness