HiDream.ai Algorithm Scientist Dr. Pan Yingwei to Speak at AICon Shenzhen on the Path from Token Prediction to State Prediction: Toward Native Multimodal World Models

The next phase of AI competition is shifting beyond model capabilities toward building reliable agents, improving AI engineering systems, and ensuring stable operation in complex environments. At the AICon Global AI Development and Application Conference in Shenzhen on August 21-22, Dr. Pan Yingwei, algorithm scientist at HiDream.ai, will deliver a keynote titled 'From Token Prediction to State Prediction: The Native Multimodal Path to World Models.' He will argue that the focus of AI is moving from large language models to world models, where the goal shifts from predicting the next token to predicting the next state, requiring a reconstruction of performance engineering paradigms. Traditional multimodal models are essentially 'single-modal concatenations' that lose information at each cross-module transfer, failing to build a unified physical-world state understanding. HiDream has adopted an alternative path with its self-developed UiT architecture, which unifies raw signals such as image pixels, text tokens, video, 3D, and actions into a shared representation space with unified tokenization and end-to-end processing in a single network. The talk will share insights on model design, training paradigms, and inference architectures, and explore the decisive variables for evolving from native multimodal to world models. Other keynotes will be given by experts such as Dr. Wu Zheming, co-founder and CEO of Qunqing Intelligent; Dr. Zhou Jingsen of Zhejiang University; Tao Jianhui, founder and CEO of TDengine; and Dr. Wang Chenglu, CEO of Shen Kaihong, covering topics like industrial embodied intelligence and physical AI.