Lang Xianpeng talks: Embodied AI will have its own 'Wei-Xiao-Li', physical AGI can't be achieved on financing alone

Lang Xianpeng, former head of intelligent driving at Li Auto, founded embodied AI company Kunlun Xing in March this year with Ren Geng, raising several billion yuan in three rounds within 90 days and becoming a unicorn. In his first public appearance since founding the company, he discussed the huge costs of embodied AI, saying training one embodied foundation model could cost tens or even hundreds of billions of yuan, and that relying only on financing cannot sustain any company until the realization of embodied AI. He compared the current 'battle of a hundred machines' to the new energy vehicle industry, predicting that only a few players will survive and that embodied AI will have its own 'Wei-Xiao-Li' (referring to NIO, Xpeng, Li Auto). He emphasized that current financing amounts are insufficient, and companies must achieve self-generated revenue from mass production and delivery to fund R&D. Lang outlined his company's paradigm of 'understanding' instead of 'imitation', with a focus on data efficiency and a 'data compilation' method. His team uses a Mixture-of-Experts (MoT) architecture with a joint causal attention mechanism, aiming for a 'one brain, multiple forms' capability. He also stressed the importance of tactile sensing over vision for dexterous hands, and predicted that robots entering households would take about five to ten years.