- - F L A V I O C O E L H O / G E T T Y I M A G E S real-world knowledge, or why so many people watch YouTube videos to learn how to accomplish various tasks rather than just read written instruc tions. LLMs are trained from reading “books” (that is, lots of textual data), and it’s kind of amazing that they can appear to do reasoning about the physical world at all. FLAVIO COELHO / GETTY IMAGES The process of scientific discovery requires What kind of foundation models can better understand the physical world? World models (thanks again, Mr. Obvious). The AI community is increasingly interested in building world models: Rather than predicting the next token in a sequence, as LLMs do, these models predict what will happen in a 3D world when a physical action is taken. Noted AI researcher and Stanford professor Fei-Fei Li cofounded a start-up called World Labs. Its purpose is to create “large world models” to support “spatial intelligence.” Google DeepMind produces the Genie world models, which gener ate interactive environments that can be navigated in real time. Jensen Huang, CEO of NVIDIA, has popularized the term physical AI. And Yann LeCun, former chief AI scientist at Meta, has founded a start-up called Advanced Machine Intelligence, which will focus on world models. (LeCun, in a social media post on X, famously compared the intelligence of current AI systems unfavorably to that of house cats: “It will take years for [intelligent systems] to get as smart as cats, and more years to get as smart as humans, let alone smarter.”)