停止的机器解决了人工智能对齐的不可确定性
Gabriel A Melo1, Marcos R O A Máximo2, Nei Y Soma2
1Department of Computer Science, Instituto Tecnológico de Aeronáutica, São José dos Campos, SP, Brazil. gam@ita.br.
Scientific reports
|May 4, 2025
概括
人工智能 (AI) 的内部对齐问题是无法解决的. 我们建议,人工智能对齐应成为内在的建筑属性,并纳入停止约束,以保证安全.
科学领域:
- 人工智能的人工智能
- 理论计算机科学 理论计算机科学
- 人工智能 安全 AI 安全
背景情况:
- 内部对齐问题质疑人工智能模型是否满足对齐函数.
- 这个问题通过赖斯定理被证明是不可决定的,与图灵的停止问题联系在一起.
- 目前的调整策略通常会对人工智能模型进行后期调整.
研究的目的:
- 为了证明内部对齐问题的不可决定性.
- 倡导人工智能对齐作为固有的建筑特征.
- 为人工智能法官功能引入停止约束,以确保有限的执行.
主要方法:
- 简化图纸证明图灵的停止问题.
- 对可证明对齐的AI操作的无数集的分析.
- 具有内在对齐和停止约束的AI架构的概念建模.
主要成果:
- 严格证明内部对齐问题的不可决定性.
- 识别可数量的可证明对齐的AI操作集.
- 展示具有内置对齐和停止保证的AI架构.
结论:
- 人工智能对齐应该是从一开始就被设计成人工智能架构的保证属性.
- 外部对齐评判函数需要一个停止约束来实现可靠的AI行为.
- 本质上是硬对齐的AI方法确保了安全性和可预测的终止.
相关概念视频
Introduction to Cognitive Psychology
229
Cognitive psychology is the field of psychology dedicated to examining how people think. It attempts to explain how and why we think the way we do by studying the interactions among human thinking, emotion, creativity, language, and problem-solving, as well as other cognitive processes. Cognitive psychology studies how information is processed and manipulated in remembering, thinking, and knowing.
This field emerged in the mid-20th century, following a period dominated by behaviorism, which...
This field emerged in the mid-20th century, following a period dominated by behaviorism, which...
229
Natural and Artificial Concepts
77
In psychology, concepts can be divided into two categories: natural and artificial. Natural concepts are formed through direct or indirect experiences. For example, consider the concept of snow. If you live in a place with regular snowfall, such as Essex Junction, Vermont, you know snow through direct experiences. You’ve seen it fall, touched it, shoveled it, and played in it. You recognize its texture, appearance, and even its smell. In contrast, if you live on an island like Saint...
77
Stereotype Content Model
13.9K
The Stereotype Content Model (SCM) was first proposed by Susan Fiske and her colleagues (Fiske, Cuddy, Glick & Xu, 2002; see also Fiske, 2012 and Fiske, 2017). The SCM specifies that when someone encounters a new group, they will stereotype them based on two metrics: warmth—or that group’s perceived intent, and how likely they are to provide help or inflict harm—and competence—or their ability to carry out that objective. Depending on the warmth-competence...
13.9K
Reason and Intuition
6.3K
The human brain processes information for decision-making using one of two routes: an intuitive system and a rational system (Epstein, 1994; popularized by Kahneman, 2011 as System 1 and System 2, respectively). The intuitive system is quick, impulsive, and operates with minimal effort, relying on emotions or habits to provide cues for what to do next, while the rational system is logical, analytical, deliberate, and methodical. Research in neuropsychology suggests that the...
6.3K
Hypothesis: Accept or Fail to Reject?
27.3K
The outcome of any hypothesis testing leads to rejecting or not rejecting the null hypothesis. This decision is taken based on the analysis of the data, an appropriate test statistic, an appropriate confidence level, the critical values, and P-values. However, when the evidence suggests that the null hypothesis cannot be rejected, is it right to say, 'Accept' the null hypothesis?
There are two ways to indicate that the null hypothesis is not rejected. 'Accept' the null...
There are two ways to indicate that the null hypothesis is not rejected. 'Accept' the null...
27.3K
ortho–para-Directing Deactivators: Halogens
5.2K
Halogens are ortho–para directors. They are more electronegative than carbon. Therefore, as ring substituents, they can withdraw electrons through the inductive effect and deactivate the aromatic ring towards electrophilic substitution. Halogens also have an electron-donating resonance effect on the ring, which influences the orientation of the incoming electrophile. If an electrophile attacks at the ortho or the para position, the halogen donates electrons and stabilizes the intermediate...
5.2K


