China's Wu Dao 2.0 Challenges US AI Dominance With Record Size
Chinese researchers unveil Wu Dao 2.0, a 1.75-trillion-parameter multimodal AI model that dwarfs GPT-3 and highlights a growing global race to build frontier systems.
Researchers at the Beijing Academy of Artificial Intelligence (BAAI) introduce Wu Dao 2.0, a massive multimodal AI model containing 1.75 trillion parameters. This system is ten times larger than OpenAI's GPT-3 and demonstrates an ability to generate text that is indiscernible from human writing. The release underscores a significant shift in the global AI landscape as international players rapidly close the research gap.
The debut of Wu Dao 2.0 exemplifies a broader trend of model diffusion, where multiple state and private actors worldwide develop their own GPT-3-style systems. Countries like Russia, France, and South Korea are actively training similar models to ensure their native cultures and languages receive proper representation in these powerful algorithms. This movement highlights a rising wave of AI nationalism as nations assert their technological sovereignty.
Built on the open source FastMoE system, Wu Dao 2.0 trains on both supercomputers and conventional GPUs without requiring proprietary hardware. The model handles a diverse range of tasks, including natural language processing, image generation, and even predicting 3D protein structures. BAAI leaders describe this mega-model as a power plant for the future of AI, transforming massive amounts of data into fuel for next-generation applications.