Hans Moravec’s robot intelligence predictions were off. Currently, AI is slow to advance due to an inefficient approach that relies on computer technology. Only by focusing on how the human brain accumulates intelligence through learning will AI technology make a real leap forward.
In his 1988 book Children of the Mind, Hans Moravec, a world-renowned robotics theorist, predicted that in the 21st century, robots would become more intelligent every decade, with the fourth generation of robots reaching human levels by 2040. But he was wrong. He expected the first generation of robots to be developed by 2010 to have the intelligence of a lizard and be able to clean the bathtub, cut the lawn, and cook gourmet meals around the house, but here we are in 2024, well past 2010, and the robots we see in our daily lives are, at best, vacuum cleaners. Depending on your point of view, a robot vacuum cleaner is just a machine that moves according to sensors, not a robot with intelligence. While IT technologies, such as the internet and the rapid development of smartphones, have been advancing at a pace that meets or exceeds people’s expectations, the development of artificial intelligence technology has been much slower. Is this because AI is more difficult than other technologies? While the answer to this question is yes, I would argue that the current approach to AI is wrong.
Many AI advocates view the brain as a kind of well-designed computer, and study AI technology as an extension of computer technology. Rather than focusing on how the actual human brain works, they are only interested in the behaviors that appear to be the result of human or animal intelligence. They believe that AI can be realized once computers have enough memory and processing power. However, this is a very inefficient approach. For example, consider the process of simply picking up an object like a wallet. Humans can pick up a wallet without thinking about “how do I pick up a wallet?” at all. However, current robotics technology has to calculate the shape and position of the wallet, the shape and angle of the robot hand’s extension, and the grasping force in order to grab it. Picking up a flying object is even more complicated. The trajectory of the object would need to be tracked every second of its flight, and the grasping process would need to be timed to coincide with its arrival. This process is much more complicated than it sounds, and it’s not easily realized even at the current speed of technology. Of course, given the speed at which computers are evolving, robots like this may not be far off. But is it really the right direction to go when humans have to perform a huge amount of calculations every second to accomplish what they can do unconsciously based on previous experience without any calculations at all?
How does the human brain manage to do things that are very difficult for machines to do without thinking? The answer to this question is still a mystery that scientists will have to unravel in the future, but what’s certain is that it’s not because the brain is faster than a computer. The brain’s counterpart to the transistor, the basic unit that performs binary computational tasks in computers, is the neuron, which receives inputs from synapses (junctions between neurons), synthesizes them, and decides whether to send outputs to other neurons. Typically, a neuron finishes its job in about 5 milliseconds and goes back to work, which means it does this 200 times per second. That may seem fast, but modern silicon-based computers can perform such operations a billion times per second. Machines are already fast enough compared to the human brain. But in terms of intelligence, they still don’t match the human brain. This fact suggests that increasing speed is not the answer to reaching human intelligence.
Although we don’t know the specific mechanisms, the most important characteristic that distinguishes human intelligence from current artificial intelligence is that it is accumulated and developed by learning. Humans learn over the course of many years of evolution and develop their intelligence by constantly accumulating what they see, hear, and feel as they are born and grow up. Even the process of grasping an object is easy to do without complex calculations because the body remembers the many failed attempts it has made since it was a newborn baby. Even if a robot behaves like a seemingly intelligent being, if it hasn’t learned and developed its own intelligence, but is merely acting according to an algorithm instilled in it by humans, it cannot be said to have intelligence. DeepFlyz, the computer that beat the world chess champion, was able to beat humans not because it had a better understanding of chess than humans. It was only able to do so because it was able to list the millions of possible chess moves millions of times faster than humans, following an algorithm designed by humans based on the rules of chess. Just like a calculator can do math much better than a human, but it doesn’t have intelligence, DeepFlitz can beat a human at chess, but it’s just a machine that produces an output based on a given input, and it doesn’t have chess intelligence. Intelligence is the ability to understand and learn.
According to evolutionary theory, humans evolved from an initially unintelligent single cell about 3 to 4 billion years ago. From this apparently unintelligent state, we evolved to our current state through repeated evolution. Considering that lizards appeared about 100 million years ago, the development of AI is happening thousands of times faster than that of living things. However, unless we shift the focus of AI technology from “how can we process complex computations at a faster rate?” to “how does the human brain learn, remember, and utilize them?”, the pace of development of AI technology may gradually slow down and show limitations. Of course, the methods used by the human brain are not the only ways to reach human-level intelligence, but why look for other, less obvious paths when the answers are right within us?
