从“世界模拟器”可以走向通用人工智能吗?  

Can the“World Simulator”Evolve into General Artificial Intelligence?

作　　者：吴静 WU Jing

机构地区：[1]南京师范大学公共管理学院 [2]南京师范大学数字与人文研究中心

出　　处：《阅江学刊》2024年第5期90-100,173,174,共13页Yuejiang Academic Journal

摘　　要：OpenAI发布的通用视觉大模型Sora以其强大的文生视频能力拥有了对当下世界进行描绘的话语权力,但其背后存在的大语言模型对世界表征的合法性问题以及带来的认知风险和社会性后果,需要将哲学和技术批判结合起来进行反思与预防性研究。本文对多模态大语言模型技术进行分析,探究文生视频的技术本质和应用困境。同时,基于大语言模型的通用性,分析了其背后蕴藏的技术理性及对用户价值认知的重塑。研究表明,在从技术通用性到知识公理性的过程中,人工智能技术出现了认识论上的悖论。基于深度学习的生成式人工智能无法真正理解世界运作方式,只是处理信息单元和语义的逻辑关系,且现有的生成式大模型在反事实推理和合规性方面仍存在缺陷,还远未达到全方位理解世界、表征世界的程度。因此,在虚实共生、人机融合成为新的时代命题时,如何把握“虚”“实”之间的关系成为一个需要审视的维度。这不仅关系到社会资源的投入,而且更关系到人类自身的重塑和对于世界的理解。Open AI's universal visual model,Sora,with its powerful text-to-video capabilities,has gained significant authority in depicting the contemporary world.However,the underlying large language model raises issues regarding the legitimacy of world representation,as well as cognitive risks and societal consequences that require a combined philosophical and technological critique.The paper analyzes multimodal large language model technologies,exploring the technical essence and application challenges of text-to-video technology.Additionally,it examines the technical rationality behind large language models and their impact on reshaping user value perceptions.The study reveals an epistemological paradox in the transition from technological generality to knowledge axioms.Generative artificial intelligence based on deep learning cannot truly comprehend the operational mechanisms of the world.It only processes logical relationships between information units and semantics.Moreover,existing generative large models still have shortcomings in counterfactual reasoning and compliance,falling short of achieving comprehensive understanding and representation of the world.Thus,as virtual-reality integration and human-machine fusion become new era challenges,understanding the relationship between virtual and real becomes a critical dimension to examine.This not only concerns the allocation of social resources but also affects human self-reconstruction and our understanding of the world.

关 键 词：大模型 通用人工智能 模态 数据 

分 类 号：F49[经济管理—产业经济] TP18[自动化与计算机技术—控制理论与控制工程]