The Age of Soul Machines: When Computers Exceed Human Intelligence

Author: Ray Kurzweil
Translator: Shen Zhiyan et al. / Country: United States
Publisher:
Publish Date: 2002-06-01
Features: There are many solutions to achieving human-level intelligence for machines. We can develop and train a system that combines neural networks with large-scale parallel processing capabilities with other software, enabling computers to learn human language and knowledge, including the ability to read and understand written materials. Although computers currently have limited ability to process non-computer language files and absorb knowledge from them, their capabilities in this area are rapidly improving. It is estimated that by the late third decade of the 21st century, computers will be able to independently read various documents, understand and imitate the content they read. Then we can let computers read all written materials—books, magazines, scientific journals, and other existing materials. Finally, computers can independently absorb knowledge through various media and information services; and share information with other computers (in this regard, computers are much stronger than their creators, humans). Once computers reach human-level intelligence, their potential for development will undoubtedly be astonishing. Since the invention of computers, computers have greatly surpassed humans in both the speed of processing information and memory capacity. A single computer can store billions, even trillions, of factual data, while we find it difficult even to memorize a few phone numbers. Computers can search through databases containing billions of records in fractions of a second and can easily share knowledge. If we combine the intelligence level of the human brain with the high speed, accuracy, and data-sharing capabilities of computers, the result will be like a tiger given wings—powerful and limitless.
The nervous system of mammals is indeed a masterpiece of creation. However, we will not simply replicate it. Although the neurons in the nervous system are complex, their primary function is to maintain life processes rather than process information. Moreover, the reaction speed of neurons is extremely slow, while electronic circuits are at least a million times faster. As long as computers achieve human-level abilities in understanding abstract concepts and recognizing various patterns, and possess other characteristics of human intelligence, they can apply this ability to all the knowledge bases humans have acquired—or will acquire.
The common reaction to the terrifying claim that computers may one day become formidable rivals to human intelligence is often to dismiss it, which is largely based on people's understanding of contemporary computer functions. After all, when people interact with personal computers, they feel that even if computers have intelligence, it is still very limited and negligible. It is hard to imagine a personal computer having a sense of humor, its own opinions, or other endearing human qualities. But computer technology will never stand still. Many of the functions of computers that were once considered impossible just a few decades ago, such as speech recognition technology, which can accurately transcribe continuous natural language, understand human language, and respond intelligently, as well as interpret medical reports like electrocardiograms and blood test results, are now on par with the accuracy of professional doctors. Of course, we must also mention the outstanding performance of computers in world-class international chess tournaments. In another decade, we will see translation devices providing real-time interpretation services; intelligent personal computer assistants will rapidly search and understand the knowledge bases of the entire world; and a vast number of other machines will also possess increasingly broad and flexible intelligence. By the 2020s, the intelligence of the human brain and computers will become increasingly indistinguishable. Computers will far surpass humans in computing speed, storage capacity, and accuracy. In contrast, the advantages of human intelligence will gradually fade.
Computer software technology has long been far better than many people imagine. Take text or voice recognition systems as an example. I have personally experienced the astonishing level of achievement computers have reached in these areas. For instance, voice recognition software seen by users in recent years, due to its low technical quality and possibly being randomly free, could only recognize a limited vocabulary and required pauses between words, even then, it often made errors. Users would be amazed to discover that current systems can recognize 60,000 English words; even if you speak at a normal pace, their accuracy is on par with that of a typist.
As the pace of evolution accelerates exponentially, the development of technology also speeds up. While some animals also use tools, the invention of technology has distinguished modern humans from other animals. Technology is not just about making and using tools. It involves recording the experience of making tools and the progress in their complexity. This requires innovation and other means to continue the evolutionary process. The genetic code of the technological evolutionary process is a record passed down by species that make tools. Just as the code of early life was directly written into the chemical structure of organisms, the early records of technological tools were the tools themselves. Later, the genetic code of technological evolution adopted written language, and now it is typically stored in computer databases. Finally, technology itself will create newer technology. We are ahead of ourselves. Our evolutionary journey is measured in thousands of years. There have been several branches of modern humans that have existed on Earth. About 100,000 years ago, Neanderthals appeared in Europe and the Middle East, but they mysteriously disappeared around 35,000 to 40,000 years ago. Despite their primitive appearance, some of their customs were close to those of modern humans, such as complex burials: they would cover the deceased with decorations like flowers. We still don’t fully understand what happened to the close relatives of modern humans, Neanderthals, but they were clearly different from our ancestors from about 90,000 years ago. Several species of primates and their branches had the ability to create technology. Only the smartest and most aggressive species were spared from extinction and survived. This pattern of intellectual and physical competition will repeat throughout human history, and technologically advanced groups will remain in the dominant position. This trend may not be a good omen for humans, as by the middle of the 21st century, computers will surpass humans in intelligence and technological complexity. About 40,000 years ago, only the branch of early humans survived among the species of primates.
There are two types of computational transformations: one where information is preserved, and another where information is destroyed. The former example is a number multiplied by another constant (excluding division). This transformation is reversible; as long as you divide by the constant, you can obtain the original number. In the other case, if we multiply this number by zero, the original information cannot be restored. We cannot divide by zero to get the original number because zero divided by zero is undefined. Therefore, this type of transformation destroys the input information. This is another example of the irreversibility of time (the second law of thermodynamics), because you cannot reverse the process of destroying information. The irreversibility of computational processes is often used as a reason for their usefulness: because this transformation is one-way and "purposeful." However, the reason this computational process is irreversible is based on its ability to destroy information rather than create it. The value of computation lies in its ability to selectively destroy useless and redundant information. For example, in tasks such as pattern recognition for faces and speech, preserving the informative features of the pattern while discarding a large amount of data from the original image or sound is crucial to the process. Intelligence is the process of accurately selecting relevant information and skillfully and purposefully destroying the rest. This is exactly what neural network patterns excel at.
Whether it’s the human brain or a computer, a single neuron receives thousands of continuous signals representing a vast amount of information in an instant. The neuron has only two choices for its response: to fire or not to fire. Regardless of the choice, it reduces a large amount of input data to a simple information signal. Once a neural network is well-trained, this simplification of information is purposeful, useful, and necessary. We see this practice of condensing a large amount of complex information into simple "yes" or "no" responses frequently in human behavior and society at many levels. Consider the vast amount of information that emerges in legal proceedings. The results of all these activities can essentially be reduced to a simple answer: guilty or not guilty, plaintiff or defendant. Some trials may involve two possible outcomes, but my view has not changed. This simple "yes" or "no" conclusion also appears in other decisions. Take the example of elections: each voter receives a huge amount of information (perhaps none of it relevant), but can only make one decision: the incumbent or the challenger. Then millions of voters make similar decisions, and the final result is also a simple piece of data.
Since there is simply too much data to be preserved in the world, we must continuously destroy a large portion of it and input the results into databases. This is the essence of the all-or-nothing function of neurons. Next time you do some cleaning and try to throw away some old items and expired archives, you will understand how difficult it is—purposefully destroying information is the essence of intelligent work.
An intelligent machine may have extraordinary means to execute the three working modes we discussed earlier: exhaustive recursive search, large-scale parallel processing for pattern recognition, and rapid iterative evolution. But without knowledge to initiate it, it cannot function. Even if it performs these three simple working modes straightforwardly, it still requires some knowledge to start. A chess-playing software that uses recursive analysis needs to know the rules of chess. A neural network pattern recognition system needs some templates to start with, so it can learn, and evolutionary algorithm software needs a starting point to be improved. These simple working modes all have powerful organizational functions. But initially, some knowledge is still needed as a seed to grow other knowledge. Therefore, regardless of which working mode you choose, the geometric shape of the network and the important parameters reflect a certain level of knowledge. If the total organization of connection points and feedback loops in a neural network is not set up correctly, its learning and exploration phase will never end.
Humans are born with this form of knowledge. The human brain is not a blank sheet of paper, on which our life experiences and epiphanies are recorded. Moreover, the human brain is composed of several specialized regions, each responsible for specific functions:
● Highly parallel early visual circuits, which are adept at distinguishing visual changes;
● Groups of visual neurons in the cerebral cortex, which can quickly recognize edges, straight lines, curves, various geometric shapes, familiar objects, or even the faces of specific individuals;
● Auditory neural circuits in the cerebral cortex, capable of recognizing sound waves formed by various frequency combinations;
● The hippocampus, which can store experiences from sensory organs and remember certain events;
● The amygdala, whose circuits can transform fear into a series of warning signals and transmit them to other parts of the brain and many other organs.
Because the human brain has this complex system of interconnected regions, with each region specializing in different types of information processing, humans are able to continuously cope with the various complex environments of daily life. As Marvin Minsky and Seymour Papert put it, the human brain is a "large system containing many relatively small, distributed systems, arranged according to embryological principles to form a complex society, which is partially controlled by a set of sequential symbolic systems (but only partially controlled), these symbolic systems were added later." They add, "These subsymbolic systems, which do most of the work, prevent the rest of the brain from understanding how they function. This can help explain why humans can unconsciously perform many complex tasks without fully understanding how they do them."