Abstract
At present,two distinct camps,namely effective accelerationism and super alignment,have emerged.The drive for development and the presence of security risks serve as the two wheels propelling the evolution of AI technology.Grounded in the social ecosystem theory,this paper contends that effective accelerationism and super alignment represent two integral aspects of the development and governance of large models.They are not antagonistic but rather engage in a dynamic interplay.Alignment is pursued to foster better development,and conversely,technological advancement provides more resourceful pathways and scope for exploring value to facilitate the realization of alignment.
This research commences by examining the current state and challenges of value alignment in AI large models.It posits that traditional governance paradigms are ill-equipped to address the intricate,uncertain,and systemic risk profiles inherent in AI large models.Resilient alignment,characterized by flexibility,fluidity,and adaptability,is capable of achieving more agile broadband,extensive-range,and dynamic control.Secondly,in practice,alignment constitutes a complex systems engineering endeavor,with a clear demarcation between the minimum threshold and the aspirational benchmarks.There exist two practical approaches:the top-down and the bottom-up methodologies.Fundamental value alignment pertains to the bedrock issues,and all specific regulations must be in strict compliance with this underlying tenet.Value framework and structural alignment relate to matters of principle,which can be tailored to suit the specific circumstances of different countries,ethnic groups,and cultures.Functional value alignment emphasizes the need for flexible adaptation based on the intended users and specific contexts,discerningly focusing on the crucial elements to accommodate the diverse alignment scenarios.Finally,alignment objectives are subject to change in response to the complex flux of time and space.There are multiple levels of control,including goal-based control,process-based control,and detail-based control.Goal-based control is aimed at delineating the boundaries between artificial intelligence and human intelligence from a macroscopic perspective,thereby establishing the baseline principles andspecific rules.Process-based control is concerned with overseeing the potential risks that may arise throughout the entire value chain of large-model technology,encompassing training,generation,application,and dissemination.Detail-based control allows for flexible adjustments according to specific scenarios,enabling the fulfillment of individual intentions and values within a diverse array of human-machine interaction settings.
In contrast to previous research,this paper transcends the conventional fixed frameworks of aligning large models with particular types or specific values.It introduces a concept of resilience,advocating the substitution of single static alignment with broadband,extensive-range,and dynamic control.On the basis of this concept,the paper deliberates on the standards,pathways,and objectives of value alignment.
In summary,this paper puts forward the concept of value alignment within a complex paradigm.The nucleus of this concept lies in clarifying the demarcation between human intelligence and artificial intelligence.By placing human beings at the center,it endeavors to identify the greatest common denominator for development amidst the dynamic interaction among humans,technology,and society,thereby attaining a state of dialogue,adjustment,and control.This not only enriches the connotation and expands the scope of the alignment issue but also contributes certain theoretical innovation to the discourse on the alignment issue from a developmental perspective.关键词
人工智能大模型/价值对齐/超级对齐/有效加速主义/韧性对齐/生成式AI治理/复杂性范式Key words
AI large model/value alignment/super alignment/effective acceleratism/resilient alignment/governance of generative AI/complexity paradigm分类
信息技术与安全科学
