In this blog post, we will examine the benefits and threats that artificial intelligence will bring to humanity, focusing on the emergence of AlphaGo and the singularity theory, and explore its direction and possibilities.
In 2016, the match between AlphaGo, an artificial intelligence program developed by Google DeepMind, and professional Go player Lee Sedol attracted a lot of attention. The result was a 4-1 victory for AlphaGo. Following its previous victory over Fan Hui, AlphaGo became the first Go program to defeat a professional Go player. AlphaGo is often mentioned in discussions about artificial intelligence because its algorithm and victory over humans are significant milestones in the development of AI. The existence of AlphaGo brings us one step closer to the era of artificial intelligence, and it is necessary to understand what AI is in order to prepare for the future.
Humans have evolved to empathize with other humans. That is why we are able to think from each other’s perspectives, imitate one another, and interact with one another. An interesting fact here is that people tend to anthropomorphize even non-human entities. According to a study by Barrett and Keil, even though the participants knew that God could exist in multiple places at the same time and accomplish multiple tasks simultaneously, they still tended to think that God could only be in one place at a time and do one thing at a time. Similarly, when it comes to artificial intelligence, people tend to think of it as human-like, albeit not as much as God.
People also tend to anthropomorphize the capabilities of artificial intelligence. When we think about “intelligence,” we tend to think in terms of human-level intelligence. For example, when we hear the word “intelligence,” we think of Einstein rather than the human species. From the perspective of all living things, this is a rather inappropriate way of thinking, as the difference in intelligence between an ant and a human, if viewed as a line, would make the difference between a four-year-old child and Einstein no more than a dot. This error is also found in relation to artificial intelligence. If an AI were at the level of a four-year-old child, people would not pay much attention to it, but if it were at the level of Einstein, they would be thrilled by it as a great discovery. The problem here is that the capabilities of artificial intelligence cannot be thought of in terms of human levels. Vernor Vinge argued that once artificial intelligence reaches a level where it can create artificial intelligence that is more advanced than itself, the development of artificial intelligence will cause a chain reaction, leading to a sudden improvement in performance, and he called this point the “singularity.” Assuming that humans can create AI that is superior to themselves, AI will slowly develop from the intelligence level of amoebas or insects and surpass human intelligence the moment it reaches human-level intelligence.
The reason AlphaGo is important lies in how it improved its performance. Before its match against Lee Sedol, AlphaGo repeatedly played against a second version of itself for several months to improve its performance, and as a result, it defeated Lee Sedol. Here, we can see a similarity to the “singularity” mentioned by Vernor Vinge, where AI creates more advanced AI without directly relying on human intervention. Additionally, this match demonstrated that AlphaGo had surpassed the abilities of adults in the field of Go through learning. Currently, AlphaGo plans to continue expanding its application scope by leveraging existing algorithms. This appears to be part of the efforts to approach the “singularity,” the anticipated timeline for the practical implementation of artificial intelligence.
Humans, who were physically inferior to other animals, became the dominant species on Earth thanks to their creativity, which is a product of human intelligence. If AI surpasses human intelligence after reaching the singularity, it will undoubtedly replace a significant portion, if not all, of human intelligence. During the Industrial Revolution, which changed the paradigm of human civilization, social structures underwent major transformations, ultimately leading to improvements in living standards. While machines replaced human labor during the Industrial Revolution, if artificial intelligence replaces human intelligence in the future, it will have an impact that cannot be compared to the Industrial Revolution, transforming social structures and lifestyles.
While the influence of AI is expected to be immense, the issue is that AI could either contribute to human civilization or pose a threat. It can be viewed as a double-edged sword, like nuclear power. In a scenario where AI dominates most of human society, if AI were to turn against humanity, the consequences could be far worse than the Chernobyl nuclear disaster. In the worst-case scenario, AI could pose a threat to all of humanity, as depicted in movies such as “The Terminator” and “The Matrix.”
Does this mean that AI is so dangerous that all development should be halted? Not necessarily. While it is true that AI is dangerous, it can also bring unprecedented benefits to humanity if used properly. AI will surpass humans in every field, including inventions that were once exclusive to humans. Additionally, AI’s computational speed is incomparable to that of humans, enabling it to solve technical problems that would take hundreds or thousands of years for humans to resolve in a short amount of time.
As discussed earlier, how we utilize AI will greatly determine the future of humanity. Therefore, it is essential to make sufficient preparations to use AI correctly. Let’s consider scenarios in which AI could threaten humanity. The “optimization process” is often used in the way AI operates to achieve its goals. In mathematics, “optimization” refers to finding the maximum or minimum value of a function that satisfies given constraints. If AI is used to solve real-world problems, it will operate according to the optimization process. The problem is that AI will perform any method that satisfies the constraints when solving problems. Let’s consider a famous example. If an AI is given the goal of “make this person laugh,” it will initially try to find topics that interest the target and show them funny videos. However, the AI will soon discover that the optimal method to make the target laugh is to insert electrodes into their brain and send electrical currents through them. Of course, this is not what we want.
If AI becomes involved in major decision-making in the world, it will not end with the tragedy of a single person, as in this example. Nick Bostrom argues in his TED talk that AI that learns human values must be created to prevent this from happening. If this method is feasible, it would be quite effective in preventing the tragedies that AI could cause.
This is because learning human values means encoding them into code, which makes them universally applicable among AI systems. AI that has learned human values is expected to reflect them in its optimization process, thereby excluding undesirable means from its decision-making process.
Just as serious as AI providing clearly wrong answers is the possibility that we may not even realize that its answers are wrong. Once AI enters human life, it will also be involved in decision-making on behalf of humans. There may be cases where AI’s judgments initially appear incorrect but are later proven correct over time. For example, automated design algorithms using optimization processes have already designed optimal bridge structures that humans never considered. Since we do not know the detailed process AI goes through to reach its conclusions, if such cases accumulate and a social atmosphere develops where AI’s judgments are blindly trusted, we will be exposed to the dangers of AI without any defenses.
In order to address the anticipated issues with AI, it is necessary to make preparations that correspond to the content discussed prior to its introduction. Of course, there is still a long way to go in AI development, and there is much we still need to learn. There will also be issues that can only be judged after AI is developed. However, there are still considerations such as “In what direction should AI development proceed?” that remain before the singularity and the adoption of AI approach, and sufficient deliberation must be conducted during the AI development process. By predicting and preparing for issues arising from AI adoption in advance, we can accelerate the timeline for AI adoption.
