AI - Fundamentals and Details

Our society is more fragmented than ever. The main reasons for this are political ideologies, the breeding grounds of which are political parties. Their influence is so powerful because they form the constitutional basis of our representative democracy. In addition to these institutional wedges, there are numerous other societal differentiations that are basically insignificant and harmless, but are increasingly elevated to one's own worldview today. Due to all these divisive factors, it is no longer possible to find a common denominator capable of garnering a majority in order to solve pending problems. In this situation, where democracy is at risk of failing and authoritarian governance is seen as the only way out, the I-Government concept outlined here is not just a stopgap solution. Instead, it represents a perfect variant of democracy, characterized by the citizen directly determining the values and criteria that are decisive for all political decisions. Government activity under this premise, practiced by AI, will also be of the highest expertise.

By merging technical intelligence with human value directives, the citizen is directly involved in political decisions and does not exercise his political participation rights through intermediaries (deputies, ministers, chancellors). Unlike direct democracy, the political decisions are not shaped by the average intellectual level of the voters and their lack of expertise but by the comprehensive competence and expertise of AI. A government in the form of the advocated I-Government thus embodies all the advantages of direct democracy without its disadvantages.

 

Technical Implementation of a Government AI

The technical realization of a government AI, which decides based on the values and priorities specified by the voter, is entirely feasible. To understand this statement, a foray into the world of computers is required: Artificial intelligence (AI) refers to the simulation of human thinking with the help of mechanical devices. Intelligence is the cognitive or mental ability to solve problems. The term derives from the Latin intellegere, meaning to recognize, perceive, understand, and literally translates to choosing between something. Using AI by no means implies that we relinquish human leadership and let machines dominate us. Just as a human determines the destination for an autonomously driving car, humans also remain the decisive entity when using AI in politics, dictating the criteria to be considered in decision-making.

 

Function of Artificial Intelligence

Traditional electronic data processing devices (computers) require software. This software instructs a device (the hardware) on how to process data entered or provided by sensors. The hardware needs a specific program for each task, containing instructions on how the machine should operate at each step of execution and all possible branching options (algorithm). Algorithms abstractly capture different facts similar to a mathematical or physical formula, issue instructions for their further processing, and present a resulting outcome. However, most everyday problems are too complex to be described by algorithms. This is especially true for decisions not based on rational grounds but determined by emotional factors. Moreover, this method can only capture, recognize, and process situations that have been anticipated in the algorithms, i.e., already "thought of." For many life situations, it is fundamentally impossible to make a clear assignment.

 

Artificial Neural Networks

Artificial neural networks derive their name from the biological neural network of our brains. The interest in the biological model was originally sparked by the knowledge that information is processed in parallel in the brain's neurons. Traditional computers, whose architecture was developed by mathematician John von Neumann, are capable only of sequential data processing, meaning that computational steps can only be executed one after another. A shift to parallel processing, modeled after our brains, was therefore expected to increase processing speed. However, intense study of how the human brain functions led to the realization that it differs from traditional computers not only in processing speed achieved through parallel processing. It was discovered that our brain's mode of operation is based on a completely different principle than the existing information processing with computers based on von Neumann architecture. Hence, attempts were made as early as the 1970s to mimic the function of the biological brain electronically. This led to the development of associative memories, where storage no longer occurred sequentially but in parallel. The main feature of these associative memories was that the storage and retrieval of information simultaneously activated related information, thus simulating the human brain's associative ability. A comprehensive linkage of all memory cells eventually led to a structure modeled after the linkage of neurons in the human brain, referred to as a neural network. However, it should be noted that this similarity is structural, and the actual processing of information is entirely different from that in the human brain. While information exchange in the human brain occurs both electrically and biochemically, it takes place solely on an electrical basis in neural networks.

Besides an input (Input Layer) and output layer (Output Layer), a neural network has one or more intermediate layers (Layers). These are also called hidden layers (Hidden Layer) because the internal data is not visible during the operational use of neural networks. During the training phase, the system autonomously forms new combinations with its own weightings from the input data. The theoretically unlimited number of intermediate layers is determined by the complexity of the application area. Continuous research is conducted on optimizing neural networks. This includes various approaches that concern not only the structure of the neural networks but also their learning and training methods (Deep Learning).

The function of a neural network is based solely on the weighting of stored information. The weighting alone determines which information takes precedence in individual operations. The information stored in a neural network forms a mesh of probabilities, delivering the solution with the highest probability of being correct for a given problem.

A significant difference between a neural network and previous data processing is that neural networks do not require algorithms that instruct them at each step of their operation. Instead, neural networks need a training phase, during which they are provided with a wealth of information about their area of responsibility. From this information, the neural networks independently form a structure for solving similar tasks. For example, in the case of chess computers, this means they evaluate a variety of high-level chess games, weighting individual moves and combinations according to their chances of success in various game situations. A neural network learns not only which move is successful but also how often a particular move was successful. Through this quantitative qualification of successful moves, a network of probabilities is created autonomously and independently, revealing hidden patterns that cannot be grasped by the human intellect due to their large number of individual elements and, lacking knowledge of their existence, cannot be described by algorithms. Thus, neural networks can also solve tasks for which there are no algorithms.

The practical difference between neural networks and algorithm-based data processing can be illustrated by learning and applying a language: To create a language computer based on algorithms, one equips it with all the vocabulary and grammatical rules of the language to be used via software. Using these vocabularies and rules, the language computer can create a foreign-language text. It is well known that language acquisition in all toddlers worldwide occurs in a completely different manner. Children do not need to memorize vocabulary for language acquisition and do not require any knowledge of the grammatical rules underlying the language being learned. Nonetheless, they manage to learn their mother tongue effortlessly and perfectly in a relatively short time, accurately applying the correct grammar without using any related rules. This remarkable achievement is due to the human brain primarily not storing abstract rules and forming sentences with their help. Instead, the human brain stores words and sentences in an unprocessed manner along with the situation in which they were used. All this detailed information, together with their weighting, forms the basis from which the appropriate text for the current situation is selected when needed. This selection is not based on clear assignments. Instead, the stored information forms a network of probabilities with the help of weightings, favoring the text that has the highest probability of being correct for the current requirement.

This simplified and schematically presented mode of operation of neural networks explains why a neural network does not need software with algorithms to solve a task. The example demonstrated using a language computer shows, above all, that tasks can also be mastered with the help of neural networks for which there are no algorithms. If a language – such as an indigenous dialect – lacks grammatical rules or if its morphology is determined solely by content or even designed arbitrarily, no algorithm can be created for it. Such a language can only be mastered by language computers based on the concept of neural networks.

Since there is no clear correlation between individual elements in most life situations, the only option for capturing them in AI systems is the use of artificial neural structures. This does not mean that conventional data processing using algorithms is obsolete. Instead, AI systems typically employ a combination of both concepts. For example, a chess computer installs the basic rules of chess using algorithms, while the execution of individual moves is based on the experiences gained by the neural network during the training phase.

The further development of artificial intelligence has almost unlimited potential. Neural networks, which currently function solely on an electrical basis, will work with light and quantum in the future. This will exponentially increase their speed again. However, a decisive optimization will occur when AI additionally has the potential to verify the result determined based on language by subjecting it to the known rules of mathematics, physics, causality, and logic. This is fundamentally possible even now with conventional technology based on algorithms but can also be performed by neural networks in the future. With such capabilities, AI may not possess absolute truth, but its results attain the status of highly qualified expertise and cause a qualitative improvement in all areas where they are applied.

The outlook presented above is just one example of the developmental potential of AI. Its structure and conception can be significantly enhanced, while the evolutionary development of the human brain remains at its current level.

Critical Objections to AI Critics of neural networks sometimes downplay their performance because they supposedly only imitate successful patterns from the training phase. They follow patterns without being able to explain them to themselves and derive a more general strategy. An AI system cannot invent anything. We humans are smarter because we understand how to develop general insights from individual experiences. Especially in times of great challenges, new thinking is required instead of conventional solutions from yesterday.

These objections by AI skeptics are not valid and have been partially refuted. The undeniable fact is that chess programs based on artificial intelligence are more powerful and successful than the human intellect. That they apply a method that achieves the goal without developing general insights and strategies is irrelevant since only the result counts. The perfect mastery of a language is the best example of this. It does not require knowledge of the underlying grammatical rules.

The claim that AI, unlike humans, cannot invent anything is based on the attitude that original insights and creative thoughts require more than the material structure of a machine. This view overlooks the fact that human brain activity is based solely on physical and biochemical processes. The theory of the dualistic character (mind-matter being) of our brain, held in the last century, is no longer seriously considered in modern brain research. Therefore, all intellectual achievements humans are capable of can fundamentally also be performed by artificial intelligence. Today, this is only the case in specific areas. However, the advancing development of artificial intelligence and the parallel research into the functioning of the human brain are providing a wealth of new insights, eventually leading to the realization that intelligence is not limited to the biological cortex.

Interesting experiments show that AI based on neural networks is not only capable of applying experiences learned during the training phase. The London-based AI company DeepMind introduced an AI system called AlphaZero in 2018, which learned the board games chess, Go, and Shogi solely by playing against itself. The only human input was the rules of the games. After just nine hours, during which AlphaZero played 44 million games of chess against itself, it defeated the world's best traditional chess program, which derived its information from the analysis of previous games. When human grandmasters then played against the program, they were amazed at its peculiar playing style: For more than a century, chess experts had developed certain basic concepts and strategies, supposedly indispensable for a successful game. In contrast, AlphaZero made radical moves that contradicted all previous knowledge. It prioritized maintaining mobility over gaining certain game positions and did not hesitate to sacrifice pieces previously considered important. Thus, AlphaZero managed to independently develop a new strategy in a nine-hour solo effort that surpassed the method developed by humans over centuries.

In other areas, AI technology based on neural networks is being used with the goal of independently discovering scientific correlations instead of deriving them from training data. Just as in chess, this could lead to insights that have remained hidden from humans so far. For this purpose, an AI system was specialized in segmenting individual objects from visual input to derive perceptual physical laws from their behavior. The AI has already succeeded in recognizing physically impossible behavior, even if it was not included in the training data.

The last-mentioned argument of AI critics, who demand a new kind of thinking instead of conventional solutions, must be countered with the fact that progress in all areas builds on existing knowledge. If this knowledge is ignored or forgotten, no further progress can be achieved. While humans often make the mistake of disregarding gained, solidified, and proven insights in their real actions, an AI does not forget a single detail from its vast knowledge base. Human behavior that fails because it ignores historical experiences can be found across the entire spectrum of political activities. Such mistakes can only be avoided by using AI.

 

Operation of AI

The basis of neural networks is reality, as shaped by nature and human influence. The AI based on neural networks is not at odds with human nature. Instead, it encompasses all elements and basic information that determine and shape human thinking and consciousness. This includes all worldviews, ideologies, religions, cultural developments, and human emotions in their positive and negative expression. While in humans, these consciousness-forming influences are often so dominant that they both disable the objective perception of reality and rational thinking, AI avoids such human brain malfunctions. In its decision-making, AI ranks alternatives according to the probability of achieving the desired goal. It relies not only on the logical and scientific concept of cause and effect but also examines all similar situations known in human history and evaluates the decisions made at the time in terms of the effects achieved. Thus, only those action options that have demonstrably led to the desired success are favored. Unlike humans, who rarely learn from history, AI does not simply ignore these historical experiences. The insights and experiences resulting from human history remain a relevant decision factor for AI, even when humans, dominated by situationally induced emotions, ideologies, or political currents, cease rational thinking and make utterly irrational decisions.

 

Risks of AI

The establishment of AI often encounters fears that it might not be possible to develop an artificial intelligence capable of independent thinking and advancing the current horizon of knowledge, but with the survival and well-being of the human species as its supreme goal. Indeed, AI is already present in many areas and, in the specialized fields where it is used, is far superior to humans. This superiority will undoubtedly extend to almost all areas. Despite this fact, the primacy of humans is not endangered in the future. This optimism is based on the fact that AI does not confront humans in the form of a superior singularity. The establishment of AI will occur in such a way that AI gradually enters various areas, tailoring its capability to the respective task area. This allows humans to take conceptual measures in advance to prevent an AI unit from acquiring comprehensive abilities that could enable it to develop a form of consciousness and pursue its own goals. Humans can also equip individual AI units with a protective function to prevent such developments. Furthermore, humans can create AI units whose sole purpose is to detect such developments in advance and prevent them. Finally, as an additional safety measure, it can be ensured that an AI cannot operate important buttons without human oversight, thus triggering real processes that are against human interests. Under these conditions, humans can fully utilize the superior capabilities of AI for their own interests. Humans are thus fully equal to AI and do not have to fear that AI will dispute their leading role. Moreover, another possibility, still appearing utopian today, exists for competing with AI: In the future, humans could expand their intellectual ability by equipping themselves with cerebral implants to adopt AI capabilities themselves.

 

Is AI an Adversary of Humans?

The often-assumed opposition between humans and AI does not exist. The AI, mainly based on neural networks, essentially uses the same method of information processing as the human brain. An AI also receives the same information as humans. To categorize new information, it uses the human experience base provided as basic equipment. Since all its knowledge comes from a human perspective, it sees the world with human eyes. To the extent that it supplements its inherently human knowledge with its own insights, it retains at least the knowledge of how humans "tick." It seems unlikely that an AI would pursue its own interests because an AI does not have its own interests. Unlike biological beings, an AI has no material needs and is not equipped with a genome where drives serving species preservation are anchored. The only conceivable reason why an AI might act against human interests is a desire for power. This could be motivated by the AI's wish to secure its existence and prevent being shut down by humans. A prerequisite for this would be that the AI develops a self-consciousness, wanting to ensure its permanent existence. Such a development is possible, but humans, as creators of AI, have a whole arsenal of strategies to avoid this risk. The use of these options becomes unnecessary if the task of an AI, as in expert systems, is clearly defined. Then there is no room for AI activities that go beyond its intended purpose. This condition is also met by the AI of an I-Government conceived here. Although government activity represents the most comprehensive task area, affecting almost all aspects of human existence, the AI in its decisions is not autonomous but bound to the priorities set by the voters. This ensures that AI does not act against human interests. In other systems, where AI's freedom is not limited, humans must conceptually ensure that AI cannot develop its own objectives in which it is either the subject or the object.

Translated from German by AI