To begin considering the future impact of artificial intelligence, we should go back hundreds of years. Humanity has always used tools and simple machines. At a certain point in history, machines appeared that differed significantly from primitive tools and simple mechanisms. In the eighteenth century, the first textile machines were constructed: they could embroider according to programmed patterns, weave fabrics—in short, do what craftspeople had previously done by hand. At the beginning of the nineteenth century, the first industrial steam engines appeared and the first locomotives took to the rails.
Let us try to imagine what people living at the time thought about the new inventions. Even the smallest steam engine was stronger than any living draft animal. Every loom with Joseph Jacquard’s automatic control of the weft arrangement could embroider with precision unattainable by humans, 24 hours a day, 7 days a week. A single James Northrop mechanical loom could produce hundreds of linear meters of fabric daily. One worker could easily operate 200 such machines. And what work was left for the other 199 workers? People at the time worried about factory jobs. Others, looking at a railway carrying hundreds of tonnes of cargo at an incredible speed, worried about their horse-drawn transport businesses.
Groups even arose that destroyed machines. Ned Ludd, the famous campaigner against steam-powered machinery, is worth recalling here. In Great Britain, a powerful lobby of horse-borne postal workers arose in opposition to building a network of railway connections. These groups failed to stop technological development because it brought enormous benefits to society. Those who worked prospered. It is enough to mention that the annual salary of a worker employed at the automatic looms in Izrael Poznański’s factory in Łódź was enough to buy a small apartment, a shop or a small farm. People employed at the great automated textile machines could earn fortunes by the standards of the time. They also had to acquire considerable technical experience, and a new class of technical workers arose. Steam railways allowed crops to be sold from distant places, which significantly reduced rural poverty. Manufactured products became inexpensive and clothing commonplace. Technological development caused a substantial rise in societies’ prosperity and increased opportunities for development. The fortunes of the wealthy grew, while a middle class and small entrepreneurs actively competed for resources. Was it therefore worth fighting the machines?
W-MOSZCZYNSKI-2023-9-48Artificial intelligence is the steam engine of the twenty-first century
After the great steam-engine revolution came the revolutions of electricity, automobiles and telephones, television and airplanes. People stopped fearing successive waves of new technologies. We are now witnessing another technological wave: the wave of artificial intelligence. How does this new technological revolution differ from the earlier ones? Previously, the work of human hands was replaced by mechanical hands, while muscle power was replaced by pistons or electric motors. Unnoticed, the first fully automatic machines, sensors and feedback systems appeared. This was the beginning of primitive artificial intelligence at the level of single-celled organisms. For example, when the temperature in boilers exceeded 120 degrees, the system cut off the heat supply; when a system detected a jam during fabric production, it stopped all production. This was the first stage in which people supervised the operation of technical systems.
Artificial intelligence today
Today we are surrounded by systems that can be defined as artificial intelligence. An important AI system that changed the world was the Google search algorithm, introduced in 1998. This inconspicuous system, now an inseparable part of our daily lives, drove most search engines such as Yahoo and Archie out of the market. Google proved to be the most effective, intelligent information-search system. It was the first manifestation of simple artificial intelligence. I remember the time before Google. Searching for information was laborious and took much more time. Google’s algorithm substantially improved quality of life. It was one of the first conveniences resulting from the use of artificial intelligence.
First level of artificial intelligence—Rule-Based System (RBS)
A system based on preprogrammed rules is artificial intelligence programmed entirely by a human. The best example is a chess program. Programmers entered into the algorithm the rules by which pieces may move, as well as the history of all piece configurations that had appeared in countless chess competitions. When playing chess, the computer analyzes the current arrangement of pieces in light of previously loaded historical chess arrangements. Knowing the rules of the game, it makes a decision. The system also considers the subsequent arrangement of pieces after a move. The algorithm is therefore based on historical data and programmed rules.
The system cannot learn from games of chess it has played itself. An algorithm of this type is described as a single-task system. It is based on operations research—that is, a methodology of seeking the most advantageous possible result subject to boundary conditions. The computer therefore knows all configurations of pieces that have appeared in the world so far and makes a decision using a simple benefit-maximization algorithm. Unfortunately, as I mentioned, the computer cannot improve its game on the basis of games it has played or observed others playing.
Returning to Google’s algorithm, somewhat more advanced contextual systems were originally used there. Today Google’s system is highly advanced and based on a second-generation AI algorithm.
First-generation systems are perfectly suited to diagnosing and optimizing processes. Their weakness is the need to update and retrain them. A human usually participates in this process. There is, however, such a thing as first generation Plus. If an AI algorithm is placed in the cloud, the retraining process can be fully automated. This solution is called productionization in a cloud environment.
Second-generation artificial intelligence—Context Awareness and Retention System (CARS)
With first-generation AI, it was possible to exercise full control over its actions and, in a sense, for a human to understand the algorithm’s decisions. What is more, it was possible to imitate a selected AI using Excel or even pencil calculations. With second-generation AI, we have the impression that we are dealing with something possessing something like feelings and, at times, a rather controversial personality. That is of course an illusion. It is still possible to imitate the algorithm’s actions with paper and pencil, but this is becoming increasingly difficult.
ChatGPT is the flagship example of such artificial intelligence. For a while, I amused myself by conversing with this algorithm. When I asked it for certain information, the system often referred me—almost out of laziness—to other sources. For example, when I asked about the number of emigrants who came to Poland in 2000–2020, the system referred me to Ministry of Foreign Affairs data. When I asked about a specific year, the system finally gave me a specific number, as though hoping this would rid it of the persistent questions. So I went on asking about the next year and received another number. I then asked about a range of several years. Although the system had previously referred me to the ministry, now—as though aware that it had already given me some information and that I knew it knew—began providing comprehensive information.
A certain inconsistency can be found in this behavior. This is very important, because first-generation AI systems were perfectly consistent. Here we are dealing with an algorithm that seemingly answers according to its mood and is even capable of… fantasizing. ChatGPT can fantasize—that is, provide information that fits but is not true. I recently watched a podcast in which the host asked ChatGPT to list his program’s guests. The list included many people who had never appeared on the program. It was evident that they fit very well with the other invited guests. A system that fantasizes, can be led on, and answers as though it had free will?
What is second-generation artificial intelligence? The algorithm undoubtedly has a system for understanding context and building its knowledge from a current stream of data. It can learn continuously and adjust its conduct to continually arriving data. ChatGPT’s naming of people who had never been program guests is the result of two algorithms: a streaming-clustering system (most probably centroidal k-means) and some response-time optimization algorithm. This combination leads to partial guessing of answers.
Such a system can detect people’s next steps and subsequent events and indicate anomalies. It is therefore excellent for controlling complex systems or searching for patterns in particular consumer behaviors. It can divide a customer population into hundreds of clusters and then find, inside them, customers who do not behave in a manner complete for the selected cluster. If the system isolates a group of 40 men and each of them bought a hammer at least once every two years, the algorithm will find the five men who do not buy a hammer every two years and begin sending them offers. It will do so because the psychological profile of those five men will indicate a high probability that they will buy hammers.
Other examples of second-level intelligence are Google Assistant and Siri. These applications learn and improve on the basis of the user’s reactions and incoming information about the user. If, for example, we ask when the next opera will be, the system infers that we mean an opera we attended recently: it understands context and analyzes our habits. The system becomes personal, which is a major breakthrough in relation to first-generation AI. Google Translate is another example: it can change translations depending on the conversational context. Contextual systems are based on hundreds of thousands of entries and conversations, article content and press reports. They also use GPS information about where we have recently been or what photographs we have taken. They analyze content in terms of credibility, the cultural conditions of our online profile and hundreds of other factors. The application provides answers many times better than a human’s.
Using ChatGPT to write articles, advertising copy or academic papers has become a genuine plague. Another system will probably soon appear to detect cases in which artificial intelligence is used in this way.
Third-generation artificial intelligence—Domain-Specific Mastery System (DSMS)
This is the next incarnation of AI, and it has existed for several years. A spectacular demonstration of the technology’s capabilities was the competition in Go, the highly complex ancient Chinese board game in which the number of possible piece combinations is regarded as close to infinity. Computers recently defeated the game’s greatest masters. Here AI algorithms can not only collect data, arrange it in contexts and understand those contexts, but also build new skills. This is no longer merely interpretation and imitation, but specialization—narrow specialization in a particular field in which a human has not the slightest chance of defeating the machine.
At this level, machines learn from one another. An interesting example is machines playing one another, exchanging conclusions and racing one another in cognitive inference. This sounds disturbing: what will happen if such machines conclude that humans are unnecessary to them or constitute some threat? It is important to mention that machines currently have no awareness of existence, so they do not perceive people and in practice perceive nothing, not even themselves. They are still only advanced, self-learning calculators. Moreover, even if they somehow understood their relationship with humans, these algorithms are immortal, because successive sequences of the algorithms enrich subsequent systems. This is a form of artificial replication.
IBM Watson is an example of such a system. It represents a higher generation of machine than ChatGPT version 3.5. IBM Watson is used, for example, as an independent intelligent contestant on the television quiz show Jeopardy!. Another commonly known third-generation AI algorithm is AlphaGo, which, as mentioned, defeated all living champions of Go.
Elements of third-generation AI can be expected soon to be implemented in the latest Google algorithms and other commonly used applications, such as photo optimization on phones or foreign-language translation applications. For now, this technology is still under development and is not yet widely available. We may remain calm: this machine does not threaten the human species.
Fourth-generation artificial intelligence—Thinking and Reasoning AI System (TRAIS)
This technology has not yet been created, but we know its capabilities. It has not been created because appropriate computers capable of handling such complex algorithms in a short time do not yet exist. This form of AI will arise soon, because the computing power it requires is already undergoing laboratory tests. I mean quantum computers.
Every human has in the brain a number of neurons exceeding the number of stars in the universe known to us. The brain operates in binary, zero-one technology. To build a thinking entity competitive with a human, therefore, one must build a computer containing billions of neural-network connections. Meanwhile, the best graphics cards in modern computers can handle, relatively quickly, at most 100,000 connections. This is practically a drop in the ocean of requirements. Zero-one technology requires thread technology, and the number of threads cannot be infinite.
Quantum computers circumvent this problem by applying the principle of simultaneously taking the values zero and one, which drastically increases computational capacity. This means that quantum computers, without applying the operating principles of the human brain, can be smaller and faster and approach the computational capacities of an average human.
Fourth-generation AI is an algorithm that attempts to imitate human thought. The algorithm becomes something like an artificial human being. This technology remains untested. At present, the computers available to us have the computational capacities of insects or small mammals. It is difficult to predict how such a technological creation will behave. We should not compare it with the iconic Terminator, because he was clearly a representative of second-generation AI. Such a robot was easy to outsmart or its next steps to predict. Humans would stand no chance in a fight against fourth-generation intelligence.
In Frank Herbert’s famous 1965 science-fiction novel Dune, the action takes place in a distant future in which people travel to other galaxies and possess excellent technology. AI is forbidden in this world because, in the course of evolution, it proved to be a serious threat to humans. Descriptions of the extinct AI suggest that it was more likely third-generation intelligence. In Herbert’s novel, all calculations necessary for space travel were performed by specially trained humans who had to be in a narcotic trance during the calculations. Possessing even the least advanced AI algorithm was punishable by death.
A cinematic embodiment of a fourth-generation AI algorithm is probably the Nexus-6 replicant from Ridley Scott’s Blade Runner. What will hypothetically be most difficult for this AI is understanding other people’s behavior and attaining self-awareness. Most researchers believe that fourth-generation AI will not be capable of attaining self-awareness and the associated instinct for self-preservation. The emergence of such an instinct could pose a serious threat to humanity. This phenomenon was depicted in James Cameron’s 1991 film Terminator 2: Judgment Day and Stanley Kubrick’s 1968 film 2001: A Space Odyssey.
Fifth-generation artificial intelligence—Artificial General Intelligence (AGI)
This is an entirely theoretical algorithm and the next generation of AI. A typical feature of present-day and near-future AI is specialization. An algorithm can play chess superbly but cannot drive a car; it can perform thousands of warehouse operations but cannot predict next week’s exchange rates. Fifth-generation AI will be something like a human, with human versatility and also a kind of sensitivity.
Entities of this type may threaten humankind. There are two basic reasons for this ominous statement. First, they may possess self-awareness and thus an instinct for self-preservation. A greater threat is that this will not be an isolated entity, but rather a collective: infinitely many entities scattered across the globe that can not only communicate and cooperate with one another, but also accumulate analytical capacity to achieve an objective.
At some point in development, humanity will cease to engage in the intellectual creative process because its effects will seem ridiculous and caricatural compared with the process conducted by AGI algorithms. This may lead to people’s complete technical deskilling, a loss of purpose in education and gaining experience, the complete incapacitation of human beings, and humanity permitting machines to set the strategy for human development. Such permission will be fully justified, because machines will be able to process information in an incomparably better manner than humans. People will no longer conduct research or artistic and exploratory activity.
The emergence of AGI will cause extraordinary, staggering technological development. It may look like this: today we learn that AI has invented a rocket engine capable of space travel at 20% of the speed of light; the next day we learn that algorithms have invented a newer engine capable of traveling at 50% of the speed of light; and on subsequent days we learn that building wormholes is possible. By Saturday, humanity has at its disposal a system transporting us to the edge of the galaxy in real time. Such a scenario is controversial and at the same time highly probable.
We can already see certain harbingers of solutions of this type today. I recently heard that researchers used AI to create the best medicine for COVID-19. They assembled 200,000 different substances in a database, entered their characteristics and effects on the human body, and also entered their effects on bacteria, viruses and the genetic system. The entire database was connected to first-generation AI. I assume that the principal component analysis (PCA) method was applied and a simple recurrent neural network was run. The algorithm found the optimum ingredients that harm the human body minimally while effectively combating the COVID-19 virus. The algorithm completed the task in… two hours.
Imagine that researchers must perform the same task without a computer. Each researcher can know at most 1,000 ingredients; those are approximately the cognitive capacities of the human brain. Yet a researcher cannot predict all the interactions that will arise among the collected 1,000 ingredients. There are hundreds of thousands of scientific studies and experiments relating to these ingredients that our researcher could theoretically know. Reading them would take decades. Despite enormous commitment, the researcher could never connect all the studies, find dependencies and important conclusions—especially because the dependencies among the ingredients need not be bilateral but multilateral. Free neural-network algorithms available to everyone can connect hundreds of thousands of ingredients, analyze all available studies and provide a solution within two hours.
Is this our end?
This example shows what today’s primitive first-level AI can do. Let us now imagine what fifth-level AI may be capable of—intelligence capable of tasks we cannot imagine today. It is difficult in this situation to say what the future role of the human being will be. Researchers claim that AGI will emerge after 2050. We will probably become dependent on a single super-algorithm of distributed AI that will manage the world and provide us with prosperity, longevity and unlimited possibilities. Unfortunately, we do not know where this turn of events will lead, apart from depriving people of a meaningful purpose in self-development, education and creation.
We know for certain that this process is inevitable. I also know for certain that this publication will be read by many AI algorithms, and its content will become part of the vast knowledge of a self-building global intelligence from which we can no longer escape.
Wojciech Moszczyński
Wojciech Moszczyński — an expert in mathematical optimization and predictive modeling. For years, he has been involved in popularizing econometric methods in business environments. He specializes in optimizing sales, production and logistics processes. For 15 years, he worked as a financial expert specializing in controlling and management accounting. For 10 years, he has worked as a data scientist. He is a graduate of the Department of Econometrics and Statistics of Nicolaus Copernicus University in Toruń. He is currently employed as a Senior Data Scientist at the Polish company Unity Group.
Dodaj komentarz