Artificial intelligence (AI) has been around for some decades. However, the intelligence of AI has thus far remained shrouded in mystery, and AI went through several phases of hype and disappointment as a result. Distinguishing between different levels of AI provides insight into ourselves and what a future with smart humans and machines might look like.
- The Dartmouth Summer Research Project on Artificial Intelligence in 1956 is generally considered the founding moment of AI as a scientific discipline. The workshop was essentially a large brainstorming session of multiple weeks on how to integrate the nascent fields of automated thinking (e.g. cybernetics, automata theory) and set up a research agenda for AI. The next couple of years were full of high hopes and optimism, only to be followed by the “AI winter” in the 1970s and 1980s, which saw reduced interest and research funding. In recent years, interest in AI has increased steadily, with AI systems recently beating humans at chess, Go, poker, and Jeopardy, and Russian President Putin saying that the nation that develops general AI first will be the ruler of the world.
- Our term algorithm is derived from the name of 9th century Persian scholar Mohammed ibn Musa al-Khwarizmi, Latinized as Algoritmi. He coined the term to refer to arithmetic techniques for solving problems, such as how to equally divide land or how to write music. In this sense, he can be considered the ancient grandfather of computer science and the philosophy of “computationalism”.
- Researchers at Google’s DeepMind have developed a deep-learning algorithm that can recognize objects in images and other things from a single example – something known as “one-shot learning”. The improvements were realized by adding a memory component to a deep-learning system, so it functions more like the connected faculties of the human mind.
- A Turing test was developed by computer scientists and philosopher Alan Turing, to test whether machines can express intelligence. The crux is that the machine’s expression and behavior should be indistinguishable for objective human observers. When the observer cannot distinguish between the machine’s and human expression and behavior, the machine passed the Turing test and is said to have a significant degree of intelligence.
- Google’s Duplex, an AI voice assistant, now identifies itself as a machine when someone talks to it. That is because many people became frightened when at its first demonstration, users didn’t realize that they were talking to an AI system (it passed the Turing test very convincingly).
- Philosopher John Searle distinguishes between two sorts of AI: weak and strong AI. Weak AI is not real intelligence, as it only simulates human intelligence but doesn’t intentionally think the content of its processes, as strong AI does. He explains this by means of his famous “Chinese room argument”: in a room sits an AI that produces Chinese symbols and passes the Turing test (it is indistinguishable from a Chinese speaker). But the AI only formally applies computer rules to formal symbols, but does not “understand” Chinese in the sense that a native speaker understands the meaning of the symbols and spoken language.
Intelligence is generally defined as the skill to achieve complex goals in complex environments. And because there are many different types and sorts of goals and environments, there are many different types of intelligence. Human intelligence is special because it is general: man can teach itself the capabilities to solve many problems and achieve a large variety of goals. Most AI is, currently, very narrow, as it can be applied to very specific, context-dependent cases (e.g. playing a game of chess, recognizing images on social media).
And because there are many different types and sorts of goals and environments, there are many different types of intelligence.
How does intelligence generally work? At first, memory capacity is essential for boosting intelligence. Humans use a wide variety of mediums to transfer and store information, such as books, symbols, language, or internet websites. These mediums have in common that they can load and store information for an extended period, and that the information can change when their referents change (e.g. we can change Wikipedia when someone wins an election). The remarkable thing is that his is not bound to a specific type of medium or material substrate: the medium only needs to be able to take on a specific state for an extended period, long enough to contain the data until this is needed and requested. The smallest unit that can do this is a bit (it can only take 0 or 1 as a value). Bits are, in this sense, the atoms of computing and digital communication.
Other examples of information media are vibrating molecules (for spoken communication through air), RNA and pen and paper. This substrate-independency of information is why software does not need an update to make computers quicker, but its computing power and memory-processing power does. As almost everything can contain information, it needs to be stored in a stable system: a system in which it costs energy to change its contents and outcome. Some systems are much easier to use to change or transfer information: hard drives are easy to change using magnetism, while carving messages in a tree costs more energy. Computation is then converting and transferring information via the rules of a specific function: taking up information (input) and processing (function) this into new information (output). Possible functions are specific mathematical functions, a cooking recipe to make pasta or a chess strategy you came up with to win the game. In general, we call these algorithms, and as such we can create algorithms for all kinds of activities and processes.
The problem is that what seems easy for computers is still difficult for humans: high-level reasoning, such as solving very large equations. That also works the other way around, as low-level sensimotor skills, such as doing the dishes or walking, is often hard for AIs. This is known as Moravec’s paradox, and goes to the core of developing and applying AI in most social and economic contexts as going from specialized or narrow AI and limited applicability to general AI remains very difficult. Many human activities, which are also essential for most work, such as having a general discussion or walking through a building, are hard to automate. Interestingly, jobs at the lower and upper end of our traditional occupation ladder can most easily be substituted by AI. The first, such as call center work or cashiers, because their jobs entail predictable physical work that can be automated. The latter are jobs, such as data processing or certain types of advocacy, as they require overview and strategic analysis of a huge range of data, something in which AI and smart algorithms have made huge progress in recent years.
Thus, for many activities and occupations it is not relevant whether AI can pass the Turing test; what matters is how applicable it is to a wider range of activities. Therefore, intelligence requires a definition less narrow than passing a Turing test, but one that takes into account the multidimensionality of activities (e.g. some activities, such as building Lego with your kids, are not defined by finding efficient solutions) and how to interpret them. As such, the scope for work that is not qualified by cost-benefit analyses or fast problem-solving is expanding, for example social work like teaching, service jobs and care, as well as craftsmanship and cooperative forms of production. As the world around us becomes ever artificially smarter, intentional and meaningful work becomes ever more important.
- Intelligence generally has positive connotations, but is in itself a value-neutral concept. In his book Superintelligence, philosopher Nick Bostrom discusses the dangers of creating general AI that surpasses human capabilities and intelligence. Installing human-friendly values in both special (e.g. weapons) and general AI will require philosophical insight to translate value-neutral data into beneficial output (i.e. creating ethical algorithms).
- We have written before how we can consider brain-computer interfaces (BCI) the final stage in the process of enhancing the value of information and making information exchange more efficient. Combining AI with BCIs will then render the holy grail of cognitive intelligence, and will expand our current capabilities beyond imagination.
- As information is substrate-independent, finding the most dense and efficient computational matter is very valuable as it can serve as a substrate for virtually any real object. This matter has been hypothesized as “computronium”. This might become a reality soon, as the computer power of quantum computers is set to boost findings in the field of programmable matter.