The Future of AI: Toward Truly Intelligent Artificial Intelligences

This article contains some reflections about artificial intelligence (AI). First, the distinction between strong and weak AI and the related concepts of general and specific AI is made, making it clear that all existing manifestations of AI are weak and specific. The main existing models are briefly described, insisting on the importance of corporality as a key aspect to achieve AI of a general nature. Also discussed is the need to provide common-sense knowledge to the machines in order to move toward the ambitious goal of building general AI. The paper also looks at recent trends in AI based on the analysis of large amounts of data that have made it possible to achieve spectacular progress very recently, also mentioning the current difficulties of this approach to AI. The final part of the article discusses other issues that are and will continue to be vital in AI and closes with a brief reflection on the risks of AI.

The final goal of artificial intelligence (AI)—that a machine can have a type of general intelligence similar to a human’s—is one of the most ambitious ever proposed by science. In terms of difficulty, it is comparable to other great scientific goals, such as explaining the origin of life or the Universe, or discovering the structure of matter. In recent centuries, this interest in building intelligent machines has led to the invention of models or metaphors of the human brain. In the seventeenth century, for example, Descartes wondered whether a complex mechanical system of gears, pulleys, and tubes could possibly emulate thought. Two centuries later, the metaphor had become telephone systems, as it seemed possible that their connections could be likened to a neural network. Today, the dominant model is computational and is based on the digital computer. Therefore, that is the model we will address in the present article.

THE PHYSICAL SYMBOL SYSTEM HYPOTHESIS: WEAK AI VERSUS STRONG AI

In a lecture that coincided with their reception of the prestigious Turing Prize in 1975, Allen Newell and Herbert Simon (Newell and Simon, 1976) formulated the “Physical Symbol System” hypothesis, according to which “a physical symbol system has the necessary and sufficient means for general intelligent action.” In that sense, given that human beings are able to display intelligent behavior in a general way, we, too, would be physical symbol systems. Let us clarify what Newell and Simon mean when they refer to a Physical Symbol System (PSS). A PSS consists of a set of entities called symbols that, through relations, can be combined to form larger structures—just as atoms combine to form molecules—and can be transformed by applying a set of processes. Those processes can create new symbols, create or modify relations among symbols, store symbols, detect whether two are the same or different, and so on. These symbols are physical in the sense that they have an underlying physical-electronic layer (in the case of computers) or a physical-biological one (in the case of human beings). In fact, in the case of computers, symbols are established through digital electronic circuits, whereas humans do so with neural networks. So, according to the PSS hypothesis, the nature of the underlying layer (electronic circuits or neural networks) is unimportant as long as it allows symbols to be processed. Keep in mind that this is a hypothesis, and should, therefore, be neither accepted nor rejected a priori. Either way, its validity or refutation must be verified according to the scientific method, with experimental testing. AI is precisely the scientific field dedicated to attempts to verify this hypothesis in the context of digital computers, that is, verifying whether a properly programmed computer is capable of general intelligent behavior.

Specifying that this must be general intelligence rather than specific intelligence is important, as human intelligence is also general. It is quite a different matter to exhibit specific intelligence. For example, computer programs capable of playing chess at Grand-Master levels are incapable of playing checkers, which is actually a much simpler game. In order for the same computer to play checkers, a different, independent program must be designed and executed. In other words, the computer cannot draw on its capacity to play chess as a means of adapting to the game of checkers. This is not the case, however, with humans, as any human chess player can take advantage of his knowledge of that game to play checkers perfectly in a matter of minutes. The design and application of artificial intelligences that can only behave intelligently in a very specific setting is related to what is known as weak AI, as opposed to strong AI. Newell, Simon, and the other founding fathers of AI refer to the latter. Strictly speaking, the PSS hypothesis was formulated in 1975, but, in fact, it was implicit in the thinking of AI pioneers in the 1950s and even in Alan Turing’s groundbreaking texts (Turing, 1948, 1950) on intelligent machines.

This distinction between weak and strong AI was first introduced by philosopher John Searle in an article criticizing AI in 1980 (Searle, 1980), which provoked considerable discussion at the time, and still does today. Strong AI would imply that a properly designed computer does not simulate a mind but actually is one, and should, therefore, be capable of an intelligence equal, or even superior to human beings. In his article, Searle sought to demonstrate that strong AI is impossible, and, at this point, we should clarify that general AI is not the same as strong AI. Obviously they are connected, but only in one sense: all strong AI will necessarily be general, but there can be general AIs capable of multitasking but not strong in the sense that, while they can emulate the capacity to exhibit general intelligence similar to humans, they do not experience states of mind.

THE PRINCIPAL ARTIFICIAL INTELLIGENCE MODELS: SYMBOLIC, CONNECTIONIST, EVOLUTIONARY, AND CORPOREAL

The symbolic model that has dominated AI is rooted in the PSS model and, while it continues to be very important, is now considered classic (it is also known as GOFAI, that is, Good Old-Fashioned AI). This top-down model is based on logical reasoning and heuristic searching as the pillars of problem solving. At the same time that symbolic AI was being developed, a biologically based approach called connectionist AI arose. Connectionist systems are not incompatible with the PSS hypothesis but, unlike symbolic AI, they are modeled from the bottom up, as their underlying hypothesis is that intelligence emerges from the distributed activity of a large number of interconnected units whose models closely resemble the electrical activity of biological neurons. In 1943, McCulloch and Pitts (1943) proposed a simplified model of the neuron based in the idea that it is essentially a logic unit. This model is a mathematical abstraction with inputs (dendrites) and outputs (axons). The output value is calculated according to the result of a weighted sum of the entries in such a way that if that sum surpasses a preestablished threshold, it functions as a “1,” otherwise it will be considered a “0.” Connecting the output of each neuron to the inputs of other neurons creates an artificial neural network. Based on what was then known about the reinforcement of synapses among biological neurons, scientists found that these artificial neural networks could be trained to learn functions that related inputs to outputs by adjusting the weights used to determine connections between neurons. These models were hence considered more conducive to learning, cognition, and memory than those based on symbolic AI. Nonetheless, like their symbolic counterparts, intelligent systems based on connectionism do not need to be part of a body, or situated in real surroundings. In that sense, they have the same limitations as symbolic systems. Moreover, real neurons have complex dendritic branching with truly significant electrical and chemical properties. They can contain ionic conductance that produces nonlinear effects. They can receive tens of thousands of synapses with varied positions, polarities, and magnitudes. Furthermore, most brain cells are not neurons, but rather glial cells that not only regulate neural functions but also possess electrical potentials, generate calcium waves, and communicate with others. This would seem to indicate that they play a very important role in cognitive processes, but no existing connectionist models include glial cells so they are, at best, extremely incomplete and, at worst, erroneous. In short, the enormous complexity of the brain is very far indeed from current models. And that very complexity also raises the idea of what has come to be known as singularity, that is, future artificial superintelligences based on replicas of the brain but capable, in the coming twenty-five years, of far surpassing human intelligence. Such predictions have little scientific merit.

Another biologically inspired but non-corporeal model that is also compatible with the PSS hypothesis is evolutionary computation (Holland, 1975). Biology’s success at evolving complex organisms led some researchers from the early 1960s to consider the possibility of imitating evolution. Specifically, they wanted computer programs that could evolve, automatically improving solutions to the problems for which they had been programmed. The idea being that, thanks to mutation operators and crossed “chromosomes” modeled by those programs, they would produce new generations of modified programs whose solutions would be better than those offered by the previous ones. Since we can define AI’s goal as the search for programs capable of producing intelligent behavior, researchers thought that evolutionary programming might be used to find those programs among all possible programs. The reality is much more complex, and this approach has many limitations although it has produced excellent results in the resolution of optimization problems. The human brain is very far removed indeed from AI models, which suggests that so-called singularity—artificial superintelligences based on replicas of the brain that far surpass human intelligence—are a prediction with very little scientific merit

One of the strongest critiques of these non-corporeal models is based on the idea that an intelligent agent needs a body in order to have direct experiences of its surroundings (we would say that the agent is “situated” in its surroundings) rather than working from a programmer’s abstract descriptions of those surroundings, codified in a language for representing that knowledge. Without a body, those abstract representations have no semantic content for the machine, whereas direct interaction with its surroundings allows the agent to relate signals perceived by its sensors to symbolic representations generated on the basis of what has been perceived. Some AI experts, particularly Rodney Brooks (1991), went so far as to affirm that it was not even necessary to generate those internal representations, that is, that an agent does not even need an internal representation of the world around it because the world itself is the best possible model of itself, and most intelligent behavior does not require reasoning, as it emerged directly from interaction between the agent and its surroundings. This idea generated considerable argument, and some years later, Brooks himself admitted that there are many situations in which an agent requires an internal representation of the world in order to make rational decisions.

In 1965, philosopher Hubert Dreyfus affirmed that AI’s ultimate objective—strong AI of a general kind—was as unattainable as the seventeenth-century alchemists’ goal of transforming lead into gold (Dreyfus, 1965). Dreyfus argued that the brain processes information in a global and continuous manner, while a computer uses a finite and discreet set of deterministic operations, that is, it applies rules to a finite body of data. In that sense, his argument resembles Searle’s, but in later articles and books (Dreyfus, 1992), Dreyfus argued that the body plays a crucial role in intelligence. He was thus one of the first to advocate the need for intelligence to be part of a body that would allow it to interact with the world. The main idea is that living beings’ intelligence derives from their situation in surroundings with which they can interact through their bodies. In fact, this need for corporeality is based on Heidegger’s phenomenology and its emphasis on the importance of the body, its needs, desires, pleasures, suffering, ways of moving and acting, and so on. According to Dreyfus, AI must model all of those aspects if it is to reach its ultimate objective of strong AI. So Dreyfus does not completely rule out the possibility of strong AI, but he does state that it is not possible with the classic methods of symbolic, non-corporeal AI. In other words, he considers the Physical Symbol System hypothesis incorrect. This is undoubtedly an interesting idea and today it is shared by many AI researchers. As a result, the corporeal approach with internal representation has been gaining ground in AI and many now consider it essential for advancing toward general intelligences. In fact, we base much of our intelligence on our sensory and motor capacities. That is, the body shapes intelligence and therefore, without a body general intelligence cannot exist. This is so because the body as hardware, especially the mechanisms of the sensory and motor systems, determines the type of interactions that an agent can carry out. At the same time, those interactions shape the agent’s cognitive abilities, leading to what is known as situated cognition. In other words, as occurs with human beings, the machine is situated in real surroundings so that it can have interactive experiences that will eventually allow it to carry out something similar to what is proposed in Piaget’s cognitive development theory (Inhelder and Piaget, 1958): a human being follows a process of mental maturity in stages and the different steps in this process may possibly work as a guide for designing intelligent machines. These ideas have led to a new sub-area of AI called development robotics