2.2.2.1 定义
Artificial General Intelligence:AI systems that are generally smarter than humans。OpenAI:“Planning for AGI and beyond”
“Sparks of Artificial General Intelligence: Early experiments with GPT-4”
There is a rich and ongoing literature that attempts to propose more formal and comprehensive definitions of intelligence, artificial intelligence, and artificial general intelligence, but none of them is without problems or controversies. For instance, Legg and Hutter propose a goal-oriented definition of artificial general intelligence: Intelligence measures an agent’s ability to achieve goals in a wide range of environments. However, this definition does not necessarily capture the full spectrum of intelligence, as it excludes passive or reactive systems that can perform complex tasks or answer questions without any intrinsic motivation or goal. One could imagine as an artificial general intelligence, a brilliant oracle, for example, that has no agency or preferences, but can provide accurate and useful information on any topic or domain. Moreover, the definition around achieving goals in a wide range of environments also implies a certain degree of universality or optimality, which may not be realistic (certainly human intelligence is in no way universal or optimal). Another candidate definition of artificial general intelligence from Legg and Hutter is: a system that can do anything a human can do. However, this definition is also problematic, as it assumes that there is a single standard or measure of human intelligence or ability, which is clearly not the case. Humans have different skills, talents, preferences, and limitations, and there is no human that can do everything that any other human can do. Furthermore, this definition also implies a certain anthropocentric bias, which may not be appropriate or relevant for artificial systems.
有大量且正在进行的文献试图提出更正式和全面的智能、人工智能和通用人工智能的定义,但没有一个是没有问题或没有争议的。例如,Legg 和 Hutter 等人提出了一个面向目标的通用人工智能定义:智能代表了一个智能体在广泛环境中实现目标的能力。然而,这一定义并不一定涵盖智力的全部范围,因为它排除了那些可以在没有任何内在动机或目标的情况下执行复杂任务或回答问题的被动或反应性系统。人们可以想象一个人工智能,可以就任何主题或领域提供准确和有用的信息。此外,关于在广泛的环境中实现目标的定义也意味着一定程度的普遍性或最优性,这可能是不现实的(当然,人类智力绝不是普遍或最优的)。Legg 和 Hutter 等人对通用人工智能的另一个候选定义是:一个可以做人类能做的任何事情的系统。然而,这个定义也是有问题的,因为它假设有一个单一的标准能够衡量人类的智力或能力,这显然不是事实。每个人都有不同的技能、天赋、偏好和限制,没有人能做其他人能做的所有事情。此外,这一定义还暗示了某种以人类为中心的偏见,这可能与人工系统不合适或不相关。
《AI新生:破解人机共存密码:人类最后一个大问题》
通用人工智能将是一种适用于所有问题类型的方法,并且在做出很少假设的情况下,它能有效地处理大而难的实例。这就是人工智能研究的终极目标——一个不需要针对具体问题的工程学系统。它会从所有可用的资源中学习它需要学习的东西,在必要时提出问题,并开始制订和执行有效的计划。
Stuart Russell:“通用人工智能可以完成人类能够完成的所有任务。我们希望AGI能够做到人类无法做到的事情。为了研究通用人工智能,我们可以从具体任务的基准 (Benchmarks) 转向任务环境的一般属性,比如部分可观察性、长时程、不可预测性等等,并问自己是否有能力为这些属性提供完整的解决方案。如果我们有这种能力,通用人工智能就应该能够自动地完成人类可以完成的任务,并且还有能力完成更多的任务。”
“Artificial General Intelligence — A gentle introduction” / 中文:“AGI的历史与现状”
“AGI”与“强AI(Strong AI)”、“类人AI(Human-Level AI)”、“完全AI”、“思维机器(Thinking Machine)”、“认知计算(Cognitive Computing)”等概念更加相似。AGI研究包括科学(理论)与工程(技术)两个方面。一个完整的AGI成果通常包括:
1.关于智能的理论
2.该理论的形式化模型
3.该模型的计算机实现
“腾讯研究院:通用人工智能时代科学研究的71个问题”
AI一般可以划分为:
– “狭义人工智能” Artificial Narrow Intelligence(ANI):也被称为弱人工智能,是擅长于单个方面的人工智能,如图像/语音识别系统、AlphaGO等,是在预定的环境中运行、执行特定任务的系统。
-“通用人工智能” Artificial General Intelligence(AGI):也被称为强人工智能,在各方面都能和人类比肩的人工智能,是为了执行广泛的智能任务、抽象思考并适应新环境的系统。Linda Gottfredson教授把智能定义为“一种宽泛的心理能力,能够进行思考、计划、解决问题、抽象思维、理解复杂理念、快速学习和从经验中学习等操作。”强人工智能在进行这些操作时和人类一样得心应手。
– 超人工智能 Artificial Super intelligence(ASI): NickBostrom把超级智能定义为“在几乎所有领域都比最聪明的人类大脑都聪明很多,包括科学创新、通识和社交技能。”
《哥德尔、埃舍尔、巴赫:集异璧之大成 Gödel, Escher, Bach: An Eternal Golden Braid》
CHAPTER XIX Artificial Intelligence: Prospects
Ten Questions and Speculations
Question: Will Al programs ever become “superintelligent”?
Speculation: I don’t know. It is not clear that we would be able to understand or relate to a “superintelligence”, or that the concept even makes sense. For instance, our own intelligence is tied in with our speed of thought. If our reflexes had been ten times faster or slower, we might have developed an entirely different set of concepts with which to describe the world. A creature with a radically different view of the world may simply not have many points of contact with us. I have often wondered if there could be, for instance, pieces of music which are to Bach as Bach is to folk tunes: “Bach squared”, so to speak. And would I be able to understand them? Maybe there is such music around me already, and I just don’t recognize it, just as dogs don’t understand language. The idea of superintelligence is very strange. In any case, I don’t think of it as the aim of Al research, although if we ever do reach the level of human intelligence, superintelligence will undoubtedly be the next goal-not only for us, but for our Al-program colleagues, too, who will be equally curious about Al and superintelligence. It seems quite likely that Al programs will be extremely curious about Al in general-understandably.

智能一般的定义是指解决问题的能力,而问题又可以按照其因果性及是否可重复来进行分类:
1、因果明确,可以重复:简单问题,推理、归纳均有效;如下棋
2、因果明确,不可重复:一般复杂问题,推理有效;如发射火箭
3、因果不明确,可以重复:复杂问题,归纳有效;如量子力学、混沌
4、因果不明确,不可重复:超复杂问题,推理、归纳均无效;如生命的进化
吴伯凡:把问题分为三类
AlphaGo能够打败棋类世界冠军(简单问题),GPT4能够处理大部分的一般复杂问题,AGI能够解决许多复杂问题,而超复杂问题会涉及大量的不可计算函数,是“使生活和数学真理不可预知,留有趣味”的范畴。