The study of artificial intelligence fundamentally changes the research on intelligence by providing a platform on which various theories can be tested. With programmable computers, each theory on intelligence, as far as it is concretely spelled out, can guide the building of artificial systems that do exactly what the theory specified as "intelligence".
Therefore, a theory of intelligence in general should cover intelligent systems that are either naturally exist or artificially built. On one hand, it should be a descriptive theory of natural intelligence, by providing a summary and explanation about the related observation in various domains. On the other hand, it should be a prescriptive theory of artificial intelligence, by providing principles and instructions about how to build systems that can be described as "intelligent". In this way, a general theory of intelligence involves both science and engineering. It should give us a better understanding about intelligence, as well as produce new technology to better meet our needs.
No existing theory can serve the above purpose. The existing descriptive theories of intelligence (such as psychology and education) basically taking "intelligence" to mean "human intelligence", and rarely attempt to separate the human-independent aspects of intelligence from its human-specific ones, or even think such a separation is possible. The related prescriptive theories (such as computer science and traditional artificial intelligence) basically taking "intelligence" to mean "problem-solving capability", and rarely attempt to separate the intelligent approaches of problem-solving from unintelligent ones, or even think such a separation is possible.
Based on the working definition, the theory should gradually introduce notions, principles, structures, mechanisms, and conclusions that are both consistent with our understanding of natural intelligence and instructive for the building of artificial intelligence. Whenever possible, the conclusions should be concrete enough to be formalized, though the theory as a whole is an empirical theory, not a formal theory (see Section 6.3).
The theory should be organized, and gradually introduce its contents. Its notions should be as clear and well-defined as possible, as well as be self-consistent. The theory should be as simple as possible, and without any unnecessary details. Whenever assumptions are needed, they should be explicitly introduced and justified, rather than be smuggled in.
The theory should not attempt to cover domains that are not directly related to intelligence (according to its working definition). Especially, a theory of intelligence will not replace any part of psychology or computer science, though may cooperate with them.
The theory should address the existing problems in AI, rather than only discuss the problems introduced by itself. However, given its different basic assumptions, its solutions are not necessarily in the form expected by the traditional theories. Especially, it may dismiss certain problems as beyond the scope of a theory of intelligence. On the problems it does solve, its solution should be specified to the level that can be directly implemented using existing technology.
Mainstream AI research has been too strongly influenced by computer science and mathematics. After failing to capture "intelligence" as a whole by notions like "computation" and "algorithm", the research has turned to domain-dependent and problem-specific tools. Consequently, the field of "AI" runs into an identity crisis, since it cannot indicate what it can contribute that is not already covered by computer science. At the same time, the "AI systems" built in this way are usually as rigid and brittle as conventional computer systems, though useful and powerful for certain limited purpose.
There are other, often "biologically inspired", theories of intelligence. They usually aim at the flexibility and adaptivity of the systems, rather than their built-in problem-solving capability. However, in them very often "biologically inspired" becomes "biologically justified", in that they take "the human way" as "the only way" to achieve intelligence, and consequently attempt to simulate the biological mechanism or process (the human brain or its evolution process) as closely as possible. Though such research will greatly contribute to our understanding of that biological mechanism or process, the computer models are actually "artificial human intelligence" or "artificial evolution", not "artificial intelligence" (though related to it). Such a limited vision of intelligence will not lead us to a truly general theory of intelligence.
There are many theories that lack a clear internal structure. Instead, they are more like collections of judgments, with contradiction and redundancy. They often fail to tell AI researchers what to do to build intelligent systems, or assume highly idealized situations that can be satisfied neither in natural systems nor in artificial systems.
Like in other branches of science, we cannot expect a perfect theory of intelligence, though we can expect one that is better than the existing theories, as explained in Section 6.3.
This theory proposes a working definition of intelligence, which is relatively simple and sharp, and it captures the following common beliefs:
Based on the working definition of intelligence, a concrete model, NARS, is described at the conceptual level, with its internal design, dynamic process, and external activities. This model unifies many processes and mechanisms that are traditionally studied and reproduced in isolation. In the discussions, many problems in artificial intelligence and cognitive sciences are addressed in a consistent manner.
This theory provides a new understanding of the essence of intelligence, as well as guidance on how to build thinking machines.