An explainer republished by 36Kr from the WeChat account Xianzao Ketang walks through the surging concept of “world models”: rather than replicating a full digital world, they aim to give AI an internal, runnable simulator of physical reality that internalizes gravity, collisions, occlusion and causality as底层 capabilities—here rendered in English as core abilities. The article argues that today's large language models rely on statistical patterns and “memorize answers,” knowing facts but not the world, while world models emphasize common-sense reasoning, forecasting and cross-domain transfer. It cites embodied AI and autonomous driving as early uses—generating rare edge cases and letting robots trial actions in a “dream”—while flagging extremely high training costs, weak handling of complex social interactions and a long road to large-scale commercialization. The funding and acquisition figures it cites (AMI Labs' seed round, AMD's purchase of World Labs, Tian Keyu's new startup) come only from that article and were not independently verified here.

World Labs·Fei-Fei Li·AMD·+5