- The meaning of artificial general intelligence for the AI industry and the world.
通用人工智能对人工智能行业和世界的意义。 - Is artificial general intelligence possible? Various development approaches and predictions.
人工通用智能可能吗？各种开发方法和预测。 - Potential risks of creating strong AI that rivals human intelligence. Should we be wary of AI?
创建可与人类智能相媲美的强大人工智能的潜在风险。我们应该警惕人工智能吗？
目录
What’s AGI AI's main concept?AGI AI 的主要概念是什么？
What is AGI artificial intelligence capable of?AGI人工智能有什么能力？
AGI vs AI difference AGI 与 AI 的区别
Narrow AI vs General AI vs Super AI狭义 AI vs 通用 AI vs 超级 AI
AGI development approachesAGI 开发方法
Challenges in the development of AGI technologyAGI技术发展面临的挑战
Future of artificial general intelligence通用人工智能的未来
Risks from artificial general intelligence人工智能的风险
General AI Myth or Fact一般人工智能神话或事实
The recent blast in the development of AI has brought many thoughts and issues to the surface. After the world witnessed how capable this technology can be, all of a sudden, the science fiction plots about androids with intelligence that rival humans don’t seem impossible anymore. Some experts state that the first steps to creating the next generation of AI – artificial general intelligence – have already been taken.
最近人工智能发展的爆炸式增长带来了许多想法和问题。在全世界见证了这项技术的强大后，突然之间，科幻小说中关于拥有与人类相媲美的智能的机器人的情节似乎不再是不可能的了。一些专家表示，创建下一代人工智能——通用人工智能——的第一步已经迈出。
We’ve decided to do our own research on the topic of AGI (artificial general intelligence), the actual state of its development, characteristics, and predictions. Of course, Atlasiko shares our analysis of the AGI meaning with you to answer the popular question “What is AGI in AI?”. Read ahead not to miss the significant transformation happening in the tech industry which can impact the whole world.
我们决定对 AGI（人工智能）这一主题、其发展的实际状况、特征和预测进行自己的研究。当然，Atlasiko 与您分享我们对 AGI 含义的分析，以回答“AI 中的 AGI 是什么？”的热门问题。继续阅读，不要错过科技行业正在发生的可能影响整个世界的重大变革。
What is AGI? 什么是通用人工智能？
To start with our explanation, let’s give a comprehensive AGI definition. So, artificial general intelligence is a term used to describe an intelligent agent with human-level cognitive abilities within the software. In other words, it’s an AI that reached the level of development to be able to solve any unfamiliar issues and tasks on par with humans. Some other specialists define AGI as a system that works autonomously and exceeds ordinary people in economically valuable tasks.
为了开始我们的解释，让我们给出一个全面的 AGI 定义。因此，通用人工智能是一个术语，用于描述在软件中具有人类认知能力的智能代理。换句话说，它是一种发展到能够与人类同等地解决任何不熟悉的问题和任务的发展水平的人工智能。其他一些专家将 AGI 定义为一种能够自主工作并在具有经济价值的任务中超越普通人的系统。
Apart from two variants of artificial general intelligence definition, it also has a few names. The system can also be called general artificial intelligence as well as Strong or True AI. In some papers, you can come upon the name “real artificial intelligence”.
除了通用人工智能定义的两个变体外，它还有一些名称。该系统也可以称为通用人工智能以及强人工智能或真人工智能。在一些论文中，你可以看到“真正的人工智能”这个名字。
What’s AGI AI's main concept?
AGI AI 的主要概念是什么？
The fundamental concepts that characterize AGI meaning in AI are “intelligence” and “consciousness”. To be considered AGI, the next-level AI has to obtain artificial cognition similar to or even the same as the natural one of humans. Just like our minds create new neuron connections living through experiences, learning, and solving, artificial general intelligence has to develop new links in its system and act on them resembling a conscious thinking process. While the intelligence concept is rather clear meaning cognitive capabilities, there are different points of view on the “consciousness” statement.
在 AI 中表征 AGI 含义的基本概念是“智能”和“意识”。要被认为是 AGI，下一级人工智能必须获得与人类自然认知相似甚至相同的人工认知。就像我们的大脑通过体验、学习和解决问题来创造新的神经元连接一样，通用人工智能必须在其系统中开发新的连接，并像有意识的思考过程一样对它们采取行动。虽然智力概念比较明确是指认知能力，但对于“意识”的说法却有不同的观点。
Naturally, the development of AI to more progressive stages gets the attention of not just computer science specialists but also philosophers who study the philosophy of mind and human existence. Thus, they present their own perspective on what the AGI system might be. The hypothesis about general AI suggested by an American philosopher, John Searle, gives us two AGI definitions that address the consciousness concept.
自然地，人工智能发展到更先进的阶段不仅引起了计算机科学专家的关注，也引起了研究心灵哲学和人类生存哲学的哲学家的关注。因此，他们对 AGI 系统可能是什么提出了自己的看法。美国哲学家 John Searle 提出的关于通用人工智能的假设为我们提供了两个针对意识概念的 AGI 定义。
Strong AI vs Weak AI
强人工智能与弱人工智能
| Strong AI 强人工智能 | Weak AI 弱人工智能 |
|---|---|
| The AI system acts upon its subjective experience producing human-level thought processes and making conscious decisions, which cannot be tested in typical ways. 人工智能系统根据其主观经验产生人类水平的思维过程并做出有意识的决定，这无法通过典型的方式进行测试。 |
The AI system only replicates the behavior of human minds, pretending to have consciousness as a major cognitive quality, but cannot actually process its subjective conscious experience. 人工智能系统只是复制了人类思维的行为，假装具有意识作为主要的认知品质，但实际上并不能处理其主观的意识体验。 |
In Searle’s “Chinese room argument”, he actually theorizes that it’s impossible for AI to really become “strong” in the sense of this particular hypothesis and obtain a human-like mind. The maximum that we’ll achieve is exactly weak AI which means just a program with generally intelligent behavior.
在 Searle 的“中文房间论证”中，他实际上推论说，AI 不可能真的在这个特定假设的意义上变得“强大”，并获得类人的思维。我们将达到的最大值恰好是弱人工智能，这意味着只是一个具有一般智能行为的程序。
At the same time, computer scientists, like Stuart Russel and Peter Norvig, set philosophical hypotheses aside saying that the main aspect that should be evaluated is the outcome. It doesn’t matter if AI just pretends to think or actually thinks like a person as long as it gives the expected results. Therefore, the debate about whether artificial general intelligence is required to have real consciousness is still going on.
与此同时，像 Stuart Russel 和 Peter Norvig 这样的计算机科学家将哲学假设放在一边，表示应该评估的主要方面是结果。 AI 只是假装思考还是真的像人一样思考都没有关系，只要它给出预期的结果即可。因此，关于人工智能是否需要具有真实意识的争论仍在继续。
Even though modern science has been developed to the point where we can make artificial organs and body parts, we still can’t replicate the mental part of our existence. So, general AI is basically an attempt to reproduce minds granting human-like intelligence to machines.
即使现代科学已经发展到我们可以制造人造器官和身体部位的地步，我们仍然无法复制我们存在的精神部分。因此，通用 AI 基本上是一种重现思维的尝试，它赋予机器类似人类的智能。
Perhaps, even from this brief answer to “What is AGI?” you can tell that the idea is rather controversial. Indeed, some find it fascinating while others say it’s outright creepy. There are many dimensions to the topic, as well as thoughts on it, which we address further in the article.
也许，甚至从对“什么是 AGI？”的这个简短回答中也可以看出这一点。你可以看出这个想法颇具争议。事实上，有些人觉得它很迷人，而另一些人则说它完全令人毛骨悚然。这个话题有很多方面，也有很多想法，我们将在本文中进一步讨论。
What is AGI artificial intelligence capable of?
AGI人工智能有什么能力？
As AGI intelligence is still a hypothetical system, there’s no way to know the full extent of its capabilities. However, there are certain characteristics that indicate true AI distinguishing it from other forms. We’ve already mentioned that one of the fundamental requirements of AGI is to be able to perform cognitive computing in a way indistinguishable from humans, but, of course, there’s more to it. As scientists developed different approaches to achieving general artificial intelligence and perspectives on the evaluation, they outline various capabilities associated with the system.
由于 AGI 智能仍然是一个假设的系统，因此无法了解其功能的全部范围。然而，有一些特征表明真正的 AI 有别于其他形式。我们已经提到，AGI 的基本要求之一是能够以与人类无异的方式执行认知计算，但是，当然，还有更多。随着科学家们开发出不同的方法来实现通用人工智能和评估的观点，他们概述了与系统相关的各种功能。
In theory, a completed AGI system is thought to be able of:
理论上，一个完整的 AGI 系统被认为能够：
- abstract thinking; 抽象思维；
- following common sense in making decisions;
在做决定时遵循常识； - comprehension of cause and effect;
理解因果关系； - creativeness; background knowledge;
才思;背景知识; - transfer learning. 迁移学习。
Some scientists also add such typical for human cognitive qualities as sentience, imagination, motivation, social intelligence, and reasoning, but they aren’t considered fundamental
一些科学家还添加了诸如感知力、想象力、动机、社交智能和推理等典型的人类认知品质，但它们并不被认为是基本的
Apart from these abilities, there is a set of functional features that the AGI computer must have in order to operate autonomously. The essential practical side of capabilities includes sensory perception, a sufficient level of motor skills, natural language understanding and processing, and a navigation system.
除了这些能力之外，AGI 计算机还必须具备一组功能特性才能自主运行。能力的基本实践方面包括感官知觉、足够水平的运动技能、自然语言理解和处理以及导航系统。
Researchers believe that AGI systems will be able to perform higher-level tasks, such as the following:
研究人员认为 AGI 系统将能够执行更高级别的任务，例如：
- utilize multiple learning methods and algorithms;
利用多种学习方法和算法； - comprehend belief systems,
理解信仰体系， - utilize miscellaneous types of knowledge,
利用各种类型的知识， - produce definite structures for tasks;
为任务制定明确的结构； - comprehend symbol systems,
理解符号系统， - engage in metacognition, and utilize knowledge on its basis.
从事元认知，并在此基础上利用知识。
AGI vs AI difference AGI 与 AI 的区别
In order to give you more understanding of just how revolutionary achieving AGI might be, let’s compare it with the technology that we can experience now – artificial intelligence. Exactly this technology and its recent advancement urged scientists to activate the discussions and research about true artificial intelligence. Although both systems are based on similar algorithms and principles, the AI vs AGI difference is actually tremendous.
为了让您更多地了解实现 AGI 可能会有多么革命性，让我们将其与我们现在可以体验的技术——人工智能——进行比较。正是这项技术及其最近的进步促使科学家们激活了关于真正人工智能的讨论和研究。尽管两个系统都基于相似的算法和原理，但 AI 与 AGI 的差异实际上是巨大的。
Researchers refer to the artificial intelligence we know and use now as Narrow AI (and weak AI in the mainstream artificial intelligence science). The name is basically self-explanatory as the system is only capable to carry out a specific, “narrow” set of tasks.
研究人员将我们现在所了解和使用的人工智能称为狭义人工智能（Narrow AI）（主流人工智能科学中的弱人工智能）。该名称基本上是不言自明的，因为该系统只能执行一组特定的“狭窄”任务。
Contrary to narrow AI, AGI in theory doesn’t have any limitations in capabilities. It’s supposed to be able to handle any unfamiliar problem and have knowledge in various areas.
与狭义的AI相反，AGI在理论上没有任何能力限制。它应该能够处理任何不熟悉的问题并拥有各个领域的知识。
Narrow AI vs General AI vs Super AI
狭义 AI vs 通用 AI vs 超级 AI
Down below we compared the two types mentioned above, general AI vs narrow AI, as well as the superior to them stage of AI – super AI.
下面我们比较了上面提到的两种类型，通用 AI 与狭义 AI，以及比它们更高级的 AI——超级 AI。
| Artificial narrow intelligence 人工狭义智能 |
Artificial general intelligence 通用人工智能 |
Artificial super intelligence 人工超级智能 |
|---|---|---|
| A narrow range of abilities according to the algorithms written by a developer 根据开发人员编写的算法，能力范围很窄 |
Can make decisions in unknown circumstances without training (display of general intelligence) 无需训练即可在未知情况下做出决定（一般智力的展示） |
Far exceeds the capabilities of even the most gifted humans in basically everything 基本上在所有方面都远远超过即使是最有天赋的人的能力 |
| Completely dependent on the dataset it was trained on in task execution 完全依赖于它在任务执行中训练的数据集 |
Can perform any task that a human is capable of which broadens the range of capabilities 可以执行人类能够完成的任何任务，从而扩大能力范围 |
Has the capacity of perfect recall, can multitask with top-level efficiency, operates superior knowledge base, etc. 拥有完美的回忆能力，能够以顶级效率进行多任务处理，拥有卓越的知识库等。 |
| Can exceed human capabilities only in a specific task it was created for 只能在为其创建的特定任务中超越人类的能力 |
Its processes and outcomes are indistinguishable from human ones (passed the Turing test) 其过程和结果与人类没有区别（通过图灵测试） |
Basically is a new species with exceptional cognitive characteristics 基本上是一个具有特殊认知特征的新物种 |
AGI development approaches
AGI 开发方法
Although AGI artificial intelligence is still a hypothetical concept, the greatest minds of computer science have already been working on possible methods and ways to achieve this technology. After conducting a meticulous analysis, we chose the most popular approaches to AGI development.
尽管 AGI 人工智能仍然是一个假设的概念，但计算机科学界最伟大的头脑已经在研究实现这项技术的可能方法和途径。经过细致的分析，我们选择了最流行的 AGI 开发方法。
Brain emulation 大脑模拟
One of the possible and most debatable approaches is human brain emulation. It can be done by thorough scanning of the human brain, mapping, and uploading it on a capable computational device. Despite sounding quite futuristic, the appropriate hardware for simulating the brain actually exists in the present. According to Raymond Kurzweil, an American computer scientist, the sufficient volume of calculations per second to simulate our brain is 10^16, while the world’s fastest supercomputer (as of March 2023), Frontier, is able to perform 10^18 calculations in 1 second. Of course, due to the massive size and uniqueness of this computer, it’s not accessible for experiments just yet and clearly cannot be used in a commonplace. Moreover, Kurzweil’s estimates don’t include the fact that the majority of exciting artificial neural networks use simplified models of biological neurons. To fully simulate the human brain with all characteristics would require more computational capacity.
一种可能且最有争议的方法是人脑仿真。它可以通过对人脑进行彻底扫描、映射并将其上传到功能强大的计算设备上来完成。尽管听起来很有未来感，但模拟大脑的合适硬件实际上存在于当下。根据美国计算机科学家 Raymond Kurzweil 的说法，每秒足以模拟我们大脑的计算量是 10^16，而世界上最快的超级计算机（截至 2023 年 3 月）Frontier 能够在 1 分钟内执行 10^18 次计算第二。当然，由于这台计算机的庞大体积和独特性，它目前还不能用于实验，显然不能用于普通场合。此外，Kurzweil 的估计不包括大多数令人兴奋的人工神经网络使用生物神经元的简化模型这一事实。要完全模拟具有所有特征的人脑，需要更多的计算能力。
Another problem is the scanning process. The human brain remains to be fully discovered since even with centuries of research there are still dark spots for scientists. The most popular suggestion is to use special nanobots that will accumulate accurate data about brain functioning but even then scientists won’t have a guarantee that the bots were able to capture all peculiarities. Therefore, to ensure successful brain emulation for achieving general AI, researchers still have to spend at least two more decades developing the required technologies.
另一个问题是扫描过程。人类的大脑仍有待完全发现，因为即使经过几个世纪的研究，科学家们仍然存在黑点。最受欢迎的建议是使用特殊的纳米机器人来收集有关大脑功能的准确数据，但即便如此，科学家们也不能保证这些机器人能够捕捉到所有的特性。因此，为了确保成功实现通用 AI 的大脑仿真，研究人员仍需花费至少二十年的时间来开发所需的技术。
Algorithmic probability 算法概率
Another approach to achieving AGI is based on the theory of algorithmic probability introduced by Ray Solomonoff. According to the method, the intelligent agent is able to predict the environment and decide on the best action even when given unfamiliar circumstances using the smallest dataset of environmental observations (Solomonoff’s induction) and the possibility of an event based on prior knowledge of conditions related to it (Bayers’ theorem).
实现 AGI 的另一种方法是基于 Ray Solomonoff 引入的算法概率理论。根据该方法，即使在给定不熟悉的情况下，智能代理也能够预测环境并决定最佳行动它（拜耳定理）。
As a continuation of this approach, a DeepMind senior scientist, Marcus Hutter created a mathematical theory of artificial general intelligence – AIXI. It’s a theoretical reinforcement learning agent that also uses Solomonoff’s induction to choose the best possible action based on observations and rewards from the environment.
作为这种方法的延续，DeepMind 资深科学家 Marcus Hutter 创建了通用人工智能的数学理论——AIXI。它是一种理论上的强化学习代理，它还使用所罗门诺夫的归纳法，根据对环境的观察和奖励来选择可能的最佳行动。
Despite the sound theoretical proof of both models, they are believed to be incomputable in practice, which means it’s impossible to create an accurate algorithm to always solve the problem correctly. Currently, there are a few approximate to artificial general intelligence examples like AIXItl and UCAI, however, they have a major drawback in terms of computation time which makes the models inefficient in practice. Many researchers now consider the AIXI model a benchmark for artificial intelligence AGI capabilities as it’s a mathematically proven functioning AGI.
尽管这两种模型都有可靠的理论证明，但它们在实践中被认为是不可计算的，这意味着不可能创建一个准确的算法来始终正确地解决问题。目前，有一些近似人工智能的例子，如 AIXItl 和 UCAI，但是，它们在计算时间方面有一个主要缺点，这使得模型在实践中效率低下。许多研究人员现在将 AIXI 模型视为人工智能 AGI 功能的基准，因为它是一种经过数学验证的功能性 AGI。
Integrative cognitive architecture
整合认知架构
This method of AGI development is based on the idea of replicating identified central cognitive processes of the human brain individually within AGI technology. The approach to AGI software named CogPrime was first introduced by Ben Goertzel (OpenCog).
这种 AGI 开发方法基于在 AGI 技术中单独复制人脑的已识别中央认知过程的想法。名为 CogPrime 的 AGI 软件方法首先由 Ben Goertzel (OpenCog) 引入。
The CogPrime system uses an action-selection module to determine the best course of action in a scenario while simulating the cognitive processes of the brain to detect information about its surroundings. This enables it to produce an intelligent model and subsequently an AGI program. The disadvantages of this paradigm include the requirement for proper memory type separation as well as the need for system-wide synergy in order to produce an efficient computing environment. In comparison with previous approaches, CogPrime was able to overcome the incomputability issue as most technologies for its implementation are available now, but the system's capabilities are much below human brains. Thus, at this stage of development, it can’t be considered true artificial intelligence.
CogPrime 系统使用动作选择模块来确定场景中的最佳动作过程，同时模拟大脑的认知过程以检测有关其周围环境的信息。这使它能够生成智能模型，随后生成 AGI 程序。这种范式的缺点包括需要适当的内存类型分离以及需要系统范围的协同作用以产生高效的计算环境。与以前的方法相比，CogPrime 能够克服不可计算性问题，因为现在大多数实现它的技术都可用，但系统的能力远低于人脑。因此，在这个发展阶段，还不能算是真正的人工智能。
Challenges in the development of AGI technology
AGI技术发展面临的挑战
Insufficient technology and great energy consumption levels
技术不足和能源消耗水平高
We’ve already mentioned that present-day technologies can’t execute cognitive operations on a human level. Even the most powerful existing supercomputers could provide just the sufficient capacity to replicate human mind calculations, not to mention multitasking and other complex processes our brain is capable of. Moreover, even recent AI releases came across the problem of enormous energy consumption. Therefore, to create an efficient real artificial intelligence people need to solve many other technologies and resource-related challenges.
我们已经提到，当今的技术无法在人类层面上执行认知操作。即使是现有的最强大的超级计算机也只能提供足够的能力来复制人类的思维计算，更不用说我们大脑能够处理的多任务处理和其他复杂过程了。此外，即使是最近的 AI 版本也遇到了巨大的能源消耗问题。因此，要创造一个高效的真正的人工智能，人们需要解决许多其他技术和资源相关的挑战。
Replicating Transfer Learning
复制迁移学习
Applying information gained in one domain to another is referred to as transfer learning. Humans regularly engage in this, and it is a significant aspect of society. For instance, we can learn how to use a foreign language word in class and apply this knowledge to make a sentence with it at home. The main aim of replicating transfer learning is to prevent retraining, so a capable AGI artificial intelligence could use one skill for solving tasks in different fields
将在一个领域中获得的信息应用到另一个领域称为迁移学习。人类经常参与其中，这是社会的一个重要方面。例如，我们可以在课堂上学习如何使用外语单词，然后在家里运用这些知识造句。复制迁移学习的主要目的是防止再训练，因此一个有能力的 AGI 人工智能可以使用一种技能来解决不同领域的任务
Facilitating collaboration and common sense
促进协作和常识
Human functioning depends on both common sense and teamwork with other human beings to complete tasks. Since today's algorithms are so limited in scope, dependable teamwork and common sense are yet far off in the future. The system must be endowed with these qualities to ensure that it is a true general artificial intelligence and not just another niche AI.
人类的功能取决于常识和与其他人的团队合作来完成任务。由于今天的算法范围如此有限，因此可靠的团队合作和常识在未来还很遥远。该系统必须具备这些品质，以确保它是真正的通用人工智能，而不仅仅是另一种小众人工智能。
Understanding Mind and Consciousness
了解思想和意识
As we defined above, consciousness is one of the main concepts and most reliable ways to prove the existence of general intelligence as it’s an essential component of human existence. However, even we, humans, can’t fully grasp all the secrets and peculiarities behind our minds. Thus, it continues to be a substantial barrier to the development and realization of general artificial intelligence.
正如我们上面所定义的，意识是证明通用智能存在的主要概念和最可靠的方法之一，因为它是人类存在的重要组成部分。然而，即使是我们人类，也无法完全掌握我们思想背后的所有秘密和特性。因此，它仍然是通用人工智能发展和实现的重大障碍。
Future of artificial general intelligence
通用人工智能的未来
After getting to know more about AGI, we can state that the question “Is artificial general intelligence possible?” isn’t a matter of doubt anymore. The answer is clearly positive as the scientists dedicate their full attention to the development of true AI. Now researchers pose another question – “When will we have artificial general intelligence?”, and let’s admit, the predictions are ambiguous.
在进一步了解 AGI 之后，我们可以提出“通用人工智能是否可能？”这个问题。不再是怀疑的问题。答案显然是肯定的，因为科学家们全神贯注于开发真正的人工智能。现在研究人员提出了另一个问题——“我们什么时候会拥有通用人工智能？”，让我们承认，这些预测是模棱两可的。
For example, a famous Australian roboticist, Rodney Brooks, concluded that a functional AGI system won’t be implemented till 2300 saying that present-day science is far from understanding “the true promise and dangers of AI”.
例如，澳大利亚著名机器人专家罗德尼·布鲁克斯 (Rodney Brooks) 得出的结论是，功能性 AGI 系统要到 2300 年才能实现，并表示当今的科学远未理解“人工智能的真正前景和危险”。
His statement was supported by remarkable researchers, such as Geoffrey Hinton and Demis Hassabis, who said that general artificial intelligence is nowhere close to being implemented.
他的声明得到了杰出研究人员的支持，例如 Geoffrey Hinton 和 Demis Hassabis，他们表示通用人工智能离实现还很远。
However, there’s also another point of view expressed by a Canadian computer scientist, Richard Sutton, who evaluated the possibilities of developing general intelligence AI in a span of the next two decades. He specified a 25% possibility of understanding AGI technology by 2030, a 50% chance that it’d happen by 2040, and only 10% – never.
然而，加拿大计算机科学家 Richard Sutton 也表达了另一种观点，他评估了在未来二十年内开发通用智能 AI 的可能性。他指出到 2030 年理解 AGI 技术的可能性为 25%，到 2040 年有 50% 的可能性，只有 10% 的可能性——永远不会。
According to our research, software development specialists in Atlasiko also tend to think that artificial general intelligence won’t arrive sooner than at the end of this century or even the next one. Although we have great theoretical advancements, modern science still has too many obstacles to overcome to implement AI with general intelligence in real life.
根据我们的研究，Atlasiko 的软件开发专家也倾向于认为，人工智能不会比本世纪末甚至下个世纪更早到来。尽管我们在理论上取得了很大的进步，但现代科学要在现实生活中实现具有通用智能的 AI 仍然有太多障碍需要克服。
Risks from artificial general intelligence
人工智能的风险
Reading this article you probably remembered some fictional scenarios from popular movies where intelligent robots take over humanity. Well, it’s pretty logical as those plots are based on real scientific concerns. Evaluation of existential risks takes a great place in the general AI dispute.
阅读本文时，您可能还记得流行电影中一些虚构的智能机器人接管人类的场景。嗯，这很合乎逻辑，因为这些情节是基于真正的科学问题。存在风险的评估在一般的 AI 争论中占有重要地位。
Even now, the rapid advancement of artificial intelligence causes many discussions and controversies as it impacts various industries and a global workforce marketplace. The development of artificial general intelligence will alter the whole world tremendously. If we’ll manage to achieve true AI with human-like consciousness, there’s no guarantee this technology will be willing to be managed by humans. To put it simply, scientists now can’t tell if we’ll get a friendly R2-D2 or an android rebellion. Exaggerations aside, let’s take a look at some risks most discussed among experts.
即使是现在，人工智能的快速发展也引发了许多讨论和争议，因为它影响着各个行业和全球劳动力市场。通用人工智能的发展将极大地改变整个世界。如果我们设法实现具有类人意识的真正人工智能，则无法保证这项技术会愿意由人类管理。简而言之，科学家们现在无法判断我们会得到一个友好的 R2-D2 还是一个机器人叛乱。撇开夸张不谈，让我们来看看专家们讨论最多的一些风险。
- Laking control. The brightest minds of the scientific community such as Stephen Hawking, Stuart J. Russel, Frank Wilczek, Geoffrey Hinton, OpenAI’s CEO Sam Altman, and others addressed the lack of attention to the control over artificial intelligence. Without proper management and monitoring strong AI can simply be misused causing major disruptions and damage to society.
缺乏控制。史蒂芬·霍金、斯图尔特·拉塞尔、弗兰克·威尔切克、杰弗里·辛顿、OpenAI 的首席执行官萨姆·奥尔特曼等科学界最聪明的头脑解决了对人工智能控制缺乏关注的问题。如果没有适当的管理和监控，强大的 AI 可能会被滥用，从而对社会造成重大破坏和破坏。 - The AI alignment problem. The more advanced the AI system is, the more challenging it can be to align its goals with human ethics. With the development of stronger cognitive abilities, true AI may be able to build strategies misaligned with intended goals and principles, for example, power-seeking. Such behavior has already been noticed in some reinforcement learning agents when they displayed instrumental convergence (more capable agents used their bigger capacity of power to achieve their goals, which is similar to what humans do). Therefore, before deploying any AI it’s vital to ensure the alignment of objectives.
AI对齐问题。人工智能系统越先进，使其目标与人类道德相一致的挑战就越大。随着更强的认知能力的发展，真正的人工智能可能会制定与预期目标和原则不一致的策略，例如权力寻求。这种行为已经在一些强化学习智能体中被注意到，当它们表现出工具收敛时（更有能力的智能体使用他们更大的能力来实现他们的目标，这与人类所做的相似）。因此，在部署任何人工智能之前，确保目标一致至关重要。 - An issue with specifying goals. For each intelligent agent, the utility function is specified by the human developer. Writing this function correctly is utterly important as it defines the set of values which would be the basis for AI’s decisions. So, if some important values happened to be not added to the utility function description, the general intelligence would act upon its own assigned tasks despite possibilities of harm or damage.
指定目标的问题。对于每个智能代理，效用函数由人类开发人员指定。正确编写此函数非常重要，因为它定义了一组值，这些值将成为 AI 决策的基础。因此，如果一些重要的值碰巧没有被添加到效用函数描述中，通用智能将执行它自己分配的任务，尽管有可能造成伤害或损害。 - Challenging goal modification in AI AGI. More advanced technologies such as artificial general intelligence might resist changes in their goal structure to ensure their continued existence and even oppose being shut down.
AI AGI 中具有挑战性的目标修改。人工智能等更先进的技术可能会抵制改变其目标结构以确保其继续存在，甚至反对被关闭。
Undoubtedly, to ensure the safety and stability of human society, scientists have to think through all risks and preventive mechanisms.
毫无疑问，要确保人类社会的安全与稳定，科学家们不得不想透所有的风险和防范机制。
General AI Myth or Fact
一般人工智能神话或事实
| Myth 神话 | Fact 事实 |
|---|---|
| The development of real artificial intelligence is impossible. 真正的人工智能的发展是不可能的。 |
Artificial general intelligence is still a hypothetical intelligent agent. Indeed, it’s predicted to be implemented in the future. 通用人工智能仍然是一种假设的智能体。事实上，它预计将在未来实施。 |
| General artificial intelligence has already been developed. 通用人工智能已经发展起来。 |
Although there are some approximations, none of the modern AI technologies can’t be considered generally intelligent as they don’t possess the required qualities. 尽管有一些近似值，但没有任何现代人工智能技术不能被认为是普遍智能的，因为它们不具备所需的品质。 |
| Threats from artificial intelligence aren’t real. They are just plots from science fiction. 来自人工智能的威胁不是真实的。它们只是科幻小说中的情节。 |
Many greatest scientists and inventors express concerns about the lack of control over general AI and the possible dangers it might bring to humanity. 许多最伟大的科学家和发明家表达了对通用人工智能缺乏控制及其可能给人类带来的危险的担忧。 |
| Strong AI with consciousness can become evil. 具有意识的强人工智能会变得邪恶。 |
AI can’t “turn evil” in the same meaning as humans. Whether conscious or not, the real problem is probable misalignment with our objectives. 人工智能不能像人类一样“转恶”。无论有意与否，真正的问题可能与我们的目标不一致。 |
| Goals of general artificial intelligence can only be determined by humans. 通用人工智能的目标只能由人类决定。 |
Contrary to narrow AI, general AI with advanced cognitive qualities can display behaviors different from determined goals based on subjective experience. 与狭义人工智能相反，具有高级认知品质的通用人工智能可以表现出不同于基于主观经验确定的目标的行为。 |
| The main threats are robots and androids. 主要威胁是机器人和机器人。 |
A misaligned AGI AI doesn't need to have a movable body (or even any body) to be able to cause damage. The only requirement is an Internet connection. 未对准的 AGI AI 不需要具有可移动的身体（或什至任何身体）就能造成损坏。唯一的要求是互联网连接。 |
| AI development will inevitably lead to technology surpassing humans and the downfall of our civilization. 人工智能的发展必然导致技术超越人类，导致我们文明的没落。 |
With strong regulations and a well-thought risk strategy, the development of an advanced AGI system will cause no harm and benefit the overall technology development. 凭借强有力的法规和深思熟虑的风险策略，先进的 AGI 系统的开发将不会造成损害并有利于整体技术发展。 |
Conclusion 结论
We hope that this article helped you to gain a better understanding and find a comprehensive answer to the question “What is artificial general intelligence?”. Achieving AGI will be an exceptional accomplishment in computer science and other related industries. However, we can’t bring down our cautiousness with such a powerful technology. Without proper control, it might bring negative changes and danger to society.
我们希望本文能帮助您更好地理解并找到“什么是通用人工智能？”这个问题的全面答案。实现 AGI 将是计算机科学和其他相关行业的一项非凡成就。然而，如此强大的技术并不能降低我们的谨慎。如果没有适当的控制，它可能会给社会带来负面的变化和危险。
If you want to find out more about the positive aspects of present-day AI assistance, read our blog where we post regular updates from the world of artificial intelligence and other useful insights.
如果您想了解有关当今 AI 帮助的积极方面的更多信息，请阅读我们的博客，我们会在其中发布来自人工智能世界的定期更新和其他有用的见解。