Hector Liu, Director of the Institute of Foundation Models‘ Silicon Valley Lab, discussed the rapid evolution of foundation models, the growing importance of world models and physical AI, and why open source transparency is central to the future of artificial intelligence.
Throughout the conversation, Liu explained how IFM differentiates itself by simultaneously developing both language models and world models, with a focus on production-grade AI systems rather than purely academic experimentation. He also outlined the organization’s belief that fully open source AI development improves both innovation and security by allowing the broader scientific community to inspect, improve, and validate AI systems collaboratively.
Liu also spoke about IFM’s affiliation with the MBZ University of Artificial Intelligence, and how the two organizations partner to support AI learning and production.
Core Takeaways
Building Both Language and World Models: IFM develops both traditional language models and physical world models, a combination Liu says is necessary for creating AI systems that truly understand and interact with the real world.
Production-Focused AI Research: Unlike many university research groups, IFM is structured around full-time researchers and engineers focused specifically on building deployable, production-ready foundation models.
Open Source as a Security Advantage: Liu argued that fully transparent open source AI development improves safety and reliability because the broader research community can inspect training data, identify risks, and contribute fixes.
The Future of Physical AI: Liu believes world models and physical intelligence will become one of the most important frontiers in AI, enabling systems to understand real-world space, motion, and physical interactions beyond what language alone can describe.
Most Important Quotes
Foundation Models Represent a Historic Shift in AI
“Foundation models are probably one of the most important breakthroughs in AI. They solve a lot of problems that prior models could not solve. They are pretty general and they show a path toward the future of how AI can be further improved. It almost evolves weekly or even daily constantly. So for researchers, the curiosity and the willingness to take challenges is basically all fulfilled in this era.”
Why IFM Focuses on Both Language and World Models
“At IFM, we build two main types of models: language models and world models. Language models are good at solving test questions and coding tasks, while world models are designed for physical AI to understand the physical world and the space around us instead of only understanding books. What we really want to do is make AI understand the world, and by design we realized that focusing on only one area is not sufficient. You have to understand both.”
Open Source Improves AI Security
“One opinion our lab holds is that transparency and open source are actually the right approach for improving security and safety. By open source, we mean being transparent about the entire production pipeline of how the model is made. The community can see what data goes into the model, comment on risks, and help improve it. We actually rely on the community to make security even better.”
The Future of AI Will Depend on Physical Intelligence
“There are always certain things you cannot fully describe in language. For example, how to ride a bicycle. You can ride a bicycle yourself, but you will never be able to verbally teach a kid how to ride until they physically experience it.” Similarly “world models are about solving and understanding those real-world problems outside of the book. That is where we see major advances happening in AI.”
