The SP Theory of Intelligence, and Its Realisation in the SP Computer Model, as a Foundation for the Development of Artificial General Intelligence

Abstract

The theme of this paper is that the SP Theory of Intelligence (SPTI), and its realisation in the SP Computer Model, is a promising foundation for the development of artificial intelligence at the level of people or higher, also known as ‘artificial general intelligence’ (AGI). The SPTI, and alternatives to the SPTI chosen to be representative of potential foundations for the development of AGI, are considered and compared. A key principle in the SPTI and its development is the importance of information compression (IC) in human learning, perception, and cognition. More specifically, IC in the SPTI is achieved via the powerful concept of SP-multiple-alignment, the key to the versatility of the SPTI in diverse aspects of intelligence, and thus a favourable combination of Simplicity with descriptive and explanatory Power. Since there are many uncertainties between where we are now and, far into the future, anything that might qualify as an AGI, a multi-pronged attack on the problem is needed. The SPTI qualifies as the basis for one of those prongs. Although it will take time to achieve AGI, there is potential along the road for many useful benefits and applications of the research.

1. Introduction

The theme of this paper is that the SP Theory of Intelligence (SPTI), and its realisation in the SP Computer Model (SPCM), is a promising foundation for the development of artificial intelligence at the level of people, or higher, also known as ‘artificial general intelligence’ (AGI).

Readers who are not already familiar with the SPTI are urged to read the introduction to it in Appendix A, and perhaps other sources referenced there.

Because AGI is far from being achieved by any system, the focus of this paper is on foundations for the development of AGI, abbreviated as ‘FDAGIs’.

In the paper, six alternatives to the SPTI, abbreviated as ‘ALTs’, are described and compared with the SPTI in terms of their potential as FDAGIs.

Although the SPTI scores well in comparison with the ALTs, it would be wrong, and quite unrealistic, to assume that all research effort should be switched into its development. Since there are many uncertainties between where we are now and, far into the future, anything that may come close to human intelligence, a multi-pronged attack on the problem is needed. And the SPTI qualifies as the basis for one of those prongs.

Although the quest for AGI, if it succeeds, is likely to take some time, any programme of research like this is likely to produce many potential benefits and applications at points along the road.

Presentation

The main sections of the paper are these:

• In Section 2, six ALTs are described.
• Section 3 describes how the ALTs and the SPTI may be evaluated as potential FDAGIs.
• Section 4 evaluates the ALTs and the SPTI as FDAGIs in terms of the criteria in Section 3.
• Section 5 summarises how the SPTI compares with the ALTs, and how the ALTs compare with each other, in terms of the criteria in Section 3.
• The paper concludes that the SPTI is indeed relatively promising as an FDAGI (Section 6).

Abbreviations are listed with their meanings after the Conclusion (Section 6), and the meaning of each one is also made clear where it is first used.

The appendices that follow the Conclusion (Section 6) should not be regarded as part of the main substance of the paper in Section 1, Section 2, Section 3, Section 4, Section 5 and Section 6. They present background information in support of the main presentation. Although they describe some elements of previous publications in this programme of research, they are merely assisting the main presentation, and should not be regarded as self-plagiarism.

• Appendix A introduces the SPTI, with pointers to where fuller information may be found.
• Appendix B describes strengths of the SPTI in both intelligence-related and non-intelligence-related domains.
• Appendix C describes topics related to the role of information compression (IC) in biology, especially HLPC, and in the SPTI.
• Appendix D provides an entirely novel perspective on the foundations of mathematics.
• Appendix E describes the benefits of a top-down, breadth-first research strategy with wide scope.

2. Six Systems That May Serve as FDAGIs

This section describes six systems, each of which is a potential FDAGI.

2.1. ‘The Society of Mind’ by Marvin Minsky

In the book The Society of Mind [1], Marvin Minsky writes:

“I’ll call ‘Society of Mind’ this scheme in which each mind is made of many smaller processes. These we’ll call agents. Each mental agent by itself can only do some simple thing that needs no mind or thought at all. Yet when we join these agents in societies—in certain very special ways—this leads to true intelligence.”

[1] (Prolog, p. 17), emphasis in the original

Later, he writes:

“Since most of the statements in this book are speculations, it would have been too tedious to mention this on every page. … Each idea should be seen not as a firm hypothesis about the mind, but as another implement to keep inside one’s toolbox for making theories about the mind.”

[1] (Postscript, p. 323)

In short, the society-of-mind idea (SOM) is largely a council of despair: the human mind is too complicated for there to be any coherent theory of its structure and workings. Nevertheless, the SOM represents a distinct approach to the development of AI which is relevant to issues considered in this paper.

2.2. ‘Gato’ from DeepMind

A team of researchers from the DeepMind company has created “A generalist agent” called ‘Gato’, described in [2]. They say:

“In this paper, we describe the current iteration of a general-purpose agent which we call Gato, instantiated as a single, large, transformer sequence model. With a single set of weights, Gato can engage in dialogue, caption images, stack blocks with a real robot arm, outperform humans at playing Atari games, navigate in simulated 3D environments, follow instructions, and more.

“While no agent can be expected to excel in all imaginable control tasks, especially those far outside of its training distribution, we here test the hypothesis that training an agent which is generally capable on a large number of tasks is possible; and that this general agent can be adapted with little extra data to succeed at an even larger number of tasks.”

[2] (p. 2)

A “transformer sequence model” mentioned in the quote is “a transformer [deep] neural network akin to a large language model” [2] (Caption to Figure 2, p. 1), and ‘transformer’ means that different parts of the input data are given different weights, somewhat like human attention.

The paper [2] does not say explicitly that the authors are aiming for AGI, but it is clear that they see Gato as a stepping stone towards AGI: in [2] (p. 18) they reference the book by Nick Bostrom [3] with its main focus on possible dangers from the development of AGI.

2.3. ‘DALL·E 2’ from OpenAI

The OpenAI research organisation says on its website openai.com (accessed on 1 December 2022):

“OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity.

“We will attempt to directly build safe and beneficial AGI, but will also consider our mission fulfilled if our work aids others to achieve this outcome.”

The latest system in this quest is ‘DALL·E 2’, which in 2022-06-09 was described as follows on the OpenAI website openai.com/dall-e-2/ (accessed on 1 December 2022):

“DALL·E 2 is a new AI system that can create realistic images and art from a description in natural language.

“DALL·E 2 can make realistic edits to existing images from a natural language caption. It can add and remove elements while taking shadows, reflections, and textures into account.

“DALL·E 2 can take an image and create different variations of it inspired by the original.

“DALL·E 2 has learned the relationship between images and the text used to describe them. It uses a process called “diffusion,” which starts with a pattern of random dots and gradually alters that pattern towards an image when it recognizes specific aspects of that image.

“In January 2021, OpenAI introduced DALL·E. One year later, our newest system, DALL·E 2, generates more realistic and accurate images with 4x greater resolution.

“DALL·E 2 is preferred over DALL·E 1 for its caption matching and photorealism when evaluators were asked to compare 1000 image generations from each model.

“DALL·E 2 is a research project which we currently do not make available in our API. As part of our effort to develop and deploy AI responsibly, we are studying DALL·E’s limitations and capabilities with a select group of users. Safety mitigations we have already developed include: Preventing Harmful Generations … Curbing Misuse … Phased Deployment Based on Learning …

“Our hope is that DALL·E 2 will empower people to express themselves creatively. DALL·E 2 also helps us understand how advanced AI systems see and understand our world, which is critical to our mission of creating AI that benefits humanity.”.

Overall, it is clear that OpenAI is aiming for AGI, and that DALL·E 2 is seen as a stepping stone in that direction. Hence, in the terms of this paper, it has potential as an FDAGI.

2.4. ‘Soar’ from John Laird, Paul Rosenbloom, and Allen Newell

The ‘Soar’ cognitive architecture was first described in [4] and it is described at various points throughout Allen Newell’s book Unified Theories of Cognition [5]. Now, the most comprehensive description of Soar is in The Soar Cognitive Architecture by John Laird [6]. And there is a more recent introduction to Soar in [7].

Some idea of how Soar is organised may be gained from Figure 1 which shows its memories, processing modules, learning modules and their connections.

In Laird’s introduction to Soar [6], it is described like this:

“Soar is meant to be a general cognitive architecture [8] that provides the fixed computational building blocks for creating AI agents whose cognitive characteristics and capabilities approach those found in humans [5,6]. A cognitive architecture is not a single algorithm or method for solving a specific problem; rather, it is the task-independent infrastructure that learns, encodes, and applies an agent’s knowledge to produce behavior, making a cognitive architecture a software implementation of a general theory of intelligence. One of the most difficult challenges in cognitive architecture design is to create sufficient structure to support coherent and purposeful behavior, while at the same time providing sufficient flexibility so that an agent can adapt (via learning) to the specifics of its tasks and environment. The structure of Soar is inspired by the human mind and as Allen Newell (Newell, 1990) suggested over 30 years ago, it attempts to embody a unified theory of cognition.”

[7] (p. 1)

Apart from being a cognitive architecture, Soar may of course also be a potential FDAGI.

2.5. ‘ACT-R’ from John Anderson, Christian Lebiere, and Others

‘ACT-R’ (which stands for “Adaptive Control of Thought—Rational”) is a cognitive architecture developed by John Anderson and Christian Lebiere, and others, which, like Soar, is inspired by Newell’s writings on the need for unified theories of cognition. A relatively full account is in Anderson and Lebiere’s book The Atomic Components of Thought [9].

On the ACT-R website act-r.psy.cmu.edu (accessed on 15 December 2022), ACT-R is described as “A theory for understanding and simulating human cognition”. In an overview of the system, the authors say:

“Adaptive control of thought–rational (ACT–R …) has evolved into a theory that consists of multiple modules but also explains how these modules are integrated to produce coherent cognition. The perceptual-motor modules, the goal module, and the declarative memory module are presented as examples of specialized systems in ACT–R. These modules are associated with distinct cortical regions. These modules place chunks in buffers where they can be detected by a production system that responds to patterns of information in the buffers. At any point in time, a single production rule is selected to respond to the current pattern. Subsymbolic processes serve to guide the selection of rules to fire as well as the internal operations of some modules. Much of learning involves tuning of these subsymbolic processes.”

[10] (Abstract)

A schematic representation of the ACT-R cognitive architecture is shown in Figure 2.

As with Soar (Section 2.4), ACT-R is a cognitive architecture, but it is also a potential FDAGI.

2.6. ‘NARS’ from Pei Wang

“Artificial General Intelligence” is a group of researchers, aiming for AGI, who hold an annual conference about their research. Details of all the conferences held to date may be seen via agi-conference.org (accessed on 25 November 2022). Although AGI is the goal of this research, it is acknowledged that the problem is difficult, and no real breakthrough has yet been achieved.

One of the systems that is associated with that AGI research is Pei Wang’s Non-Axiomatic Reasoning System (NARS) (see, for example, Refs. [11,12,13], and sources referenced there).

NARS is an environment for Non-Axiomatic Logic (NAL): “This [NAL] logic is designed for the creation of general-purpose Artificial Intelligence (AI) systems, by formulating the fundamental regularity of human thinking at a general level.” [14] (Location 94). Wang recognises that, to achieve AGI, there must be integration of different aspects of intelligence. In [11] he discusses how reasoning and learning may be integrated within the NARS environment.

So NARS is a potential FDAGI, and is evaluated in this paper alongside other ALTs and the SPTI.

3. How the ALTs and the SPTI May Be Evaluated, Viewed as FDAGIs

This section describes how the ALTs and the SPTI may be evaluated, viewed as FDAGIs.

Although the ALTs are not normally seen as scientific theories, they represent ideas about the nature of AGI which seem to need the same kind of evaluation as may be applied to scientific theories. Hence, this section applies some widely-accepted conclusions from the philosophy of science about desirable or mandatory features in any scientific theory:

• Ockham’s Razor. It has long been recognised that any good scientific theory should conform to the rule that, in words attributed to the English Franciscan friar William of Ockham: “Entities should not be multiplied beyond necessity”. As described in Appendix C.1, this means that any good theory should combine conceptual Simplicity with descriptive or explanatory Power.
• Falsifiability. Karl Popper proposed that any scientific theory should be falsifiable [15,16], meaning that, for the given theory, it should be possible to conceive of evidence that would show that the theory was wrong.

3.1. Evaluation Headings

In Section 4, which follows this main section, the ALTs and the SPTI are evaluated under the four headings shown here:

• Simplicity. It may not be possible to measure the Simplicity of a system precisely in terms of bits or bytes of information, but a more informal assessment may nevertheless be useful. A large system should have a low score (0) for Simplicity, while a small system may have a high score (2), and something in between may be given a score of 1.
• Power in Modelling Aspects of Human Intelligence. As with Simplicity, it may not be possible to measure Power precisely in terms of bits or bytes of information, but a more informal measure may nevertheless be useful. A system with high Power should have a high score (2), while a system with low Power should have a low score (0), and something in between may be given a score of 1.
• Other Strengths or Weaknesses. In addition to Simplicity and Power, there may be other strengths or weaknesses of a given system that are relevant to the development of AGI. An assessment that emphasises weaknesses is scored $-1$, an assessment which is mainly about strengths is scored 1, and when strengths and weaknesses are at least roughly equal, or when there are none, the score is 0.
  Under this heading it is appropriate to consider anything that suggests that the given ALT or the SPTI, viewed as a theory of intelligence, might be unfalsifiable, as described above.
• Combined Score. To simplify comparisons amongst two or more systems, a ‘combined score’ is calculated as the sum of the scores from the three headings above.

3.2. A Definition of Intelligence

Evaluating the Power of the ALTs and the SPTI in Section 4, below, means assessing their strengths or weaknesses in modelling human intelligence.

Arriving at a meaning for ‘human intelligence’ that will be widely accepted is difficult, witness the variety of definitions that have been advanced, many of which have been documented by Shane Legg and Marcus Hutter in [17].

As mentioned in Section 3, this section describes a definition of intelligence that may serve in the evaluation of any system, viewed as a potential FDAGI. This definition reflects the state of the SPTI as it has been developed to date. No doubt, other features will be added with further development of the SPTI. And, given the complexity of the concept of human intelligence, it should not be surprising to find that other sets of features, with at least equal claims for their validity, have been adopted by other researchers in other studies.

Some readers may object that the adoption of a definition of intelligence that reflects what the SPTI can do means the creation of an unreasonable bias in favour of the SPTI. That might be true if that definition was the only one to be adopted in the assessment of potential FDAGIs. However, as noted in the previous paragraph, alternative definitions of intelligence are recognised in this paper and given the same weight as the definition in Section 3.4, below.

3.3. Taking Account of the Distinction between the ‘Core’ of Each System and How It May Be When It Is Enhanced via Learning or Programming

In Soar, ACT-R, and the SPTI, there is a ‘core’ to the system which may be enhanced via learning or via programming.

In these cases, it is probably best for the core to be the primary focus of the evaluation. The main reasons are:

• The core is relatively well defined but the way in which the core may be enhanced via learning or programming is less well defined.
• The core comprises aspects of human intelligence, such as those described in Section 3.4, which are probably inborn and not learned.
• It seems likely that features like those described in Section 3.4 are desirable features in any FDAGI. Likewise for other intelligence-related features with comparable validity mentioned near the beginning of Section 3.2.

3.4. The SPTI Core as a Definition of Intelligence

It seems likely that, although unsupervised learning is an important part of the SPTI, the features described here are, in people, inborn and not learned. These features, which may be seen as the central ‘core’ of human intelligence (Section 3.3), are described quite fully in [18] and more briefly in [19].

What follows is an outline of the SPTI definition of intelligence. There is more detail in Appendix B.1.

• Information Compression. In view of substantial evidence for the importance of IC in HLPC [20], IC should be seen as an important feature of human intelligence.
• Natural Language Processing. Under the general heading of “Natural Language Processing” are capabilities that facilitate the learning and use of natural languages. These include:

  -

  The ability to structure syntactic and semantic knowledge into hierarchies of classes and sub-classes, and into parts and sub-parts.

  -

  The ability to integrate syntactic and semantic knowledge.

  -

  The ability to encode discontinuous dependencies in syntax such as the number dependency (singular or plural) between the subject of a sentence and its main verb, or gender dependencies (masculine or feminine) in French—where ‘discontinuous’ means that the dependencies can jump over arbitrarily large intervening structures.

  -

  Also important in this connection is that different kinds of dependency (e.g., number and gender) can co-exist without interfering with each other.

  -

  The ability to accommodate recursive structures in syntax, and perhaps also in semantics.

• Recognition and Retrieval. Capabilities that facilitate recognition of entities or retrieval of information include:

  -

  The ability to recognise something or retrieve information on the strength of a good partial match between features as well as an exact match.

  -

  Recognition or retrieval within a class-inclusion hierarchy with ‘inheritance’ of attributes, and recognition or retrieval within a hierarchy of parts and sub-parts.

• Probabilistic Reasoning. Capabilities here include: one-step ‘deductive’ reasoning; abductive reasoning; probabilistic networks and trees; reasoning with ‘rules’; nonmonotonic reasoning; ‘explaining away’ meaning ‘If A implies B, C implies B, and B is true, then finding that C is true makes A less credible.’ In other words, finding a second explanation for an item of data makes the first explanation less credible; probabilistic causal diagnosis; and reasoning which is not supported by evidence.
• Planning and Problem Solving. Capabilities here include:

  -

  The ability to plan a route, such as for example a flying route between cities A and B, given information about direct flights between pairs of cities including those that may be assembled into a route between A and B.

  -

  The ability to solve analogues of GAPs in textual form, where a ‘GAP’ is a Geometric Analogy Problem.

• Unsupervised Learning. Chapter 9 of [18] describes how the SPCM may achieve unsupervised learning from a body of ‘raw’ data, I, to create an SP-grammar, G, and an Encoding of I in terms of G, where the encoding may be referred to as E. At present the learning process has shortcomings summarised in [19] (Section 3.3) but it appears that these problems are soluble.
  In its essentials, unsupervised learning in the SPCM means searching for one or more ‘good’ SP-grammars, where a good SP-grammar is a set of SP-patterns which is relatively effective in the economical encoding of I via SP-multiple-alignment (SPMA, Appendix C.1),

4. Evaluation of the ALTs and the SPTI as FDAGIs

This section evaluates the ALTs and the SPTI in terms of the criteria described in Section 3. These evaluations have been made carefully, trying to avoid biases or distortions arising from the author’s association with the SPTI.

4.1. The SOM

Since no computational core is specified for the SOM, the evaluation necessarily focuses on the nature of the system when it is populated by many small agents:

• Simplicity, Score $=0$. At first sight, the SOM looks simple: intelligence is merely lots of little agents. However, a little reflection suggests that in the SOM as Minsky describes it:

  -

  There would be large numbers of agents, which collectively would be far from simple.

  -

  Since IC is not mentioned, we may assume that there would be no corresponding benefit via the simplification of those many agents.

  In short, the SOM is weak in terms of Simplicity and it has been assigned a corresponding score of 0.
• Power, Score $=0$. Regarding the (intelligence-related) descriptive/explanatory Power of the SOM:

  -

  In the second of the quotes in Section 2.1, Minsky says: “… most of the statements in this book are speculations” [1] (p. 17). In other words, there is little evidence in support of the proposals in the book.

  -

  The SOM concept of human intelligence is little better than a theory that merely redescribes what it is meant to explain (Appendix C.1).

  An example may be seen in [1] (p. 20) where Minsky suggests that the picking up of a cup of tea by a person may be analysed into such agents as one for grasping the cup, an agent for balancing the cup to avoid spills, an agent for the thirst of the person picking up the cup, and an agent for moving the cup to the person’s lips.

  In short, the SOM is weak in terms of its descriptive/explanatory Power and it has been assigned a corresponding score of 0.
• Other Strengths or Weaknesses, Score $=-1$. Apart from the weakness of the SOM in terms of Simplicity and Power, it is weak because, in terms of Popper’s ideas: the theory is unfalsifiable. This is because any proposed falsification of the theory could be met by the addition, omission, or substitution of agents within the theory. The SOM is too malleable to be plausible as a theory of human intelligence. Because of this additional weakness in the SOM, it has been assigned a score of $-1$ for ‘other strengths or weaknesses’.
• Combined Score $=0+0-1=-1$.

4.2. Gato

Since the case for Gato depends critically on the idea that intelligence may be built up via the learning of many more-or-less simple skills, it should be evaluated in terms of how it will be when many such skills have been learned.

• Simplicity, Score$=1$. For Gato to do many things, as indicated in the quote at the beginning of Section 2.2, it must be trained in all of them (apart from the ability to learn required to learn those many skills). And training across many such tasks means adding a substantial amount of information to the core system. From the perspective of Simplicity, things do not look so good for Gato. However, since Gato can achieve quite a lot with a small computational core, it seems reasonable to give it a middling score of 1.
• Power, Score$=2$. As noted in [2] (p. 2), Gato, with appropriate training, “… can… engage in dialogue, caption images, stack blocks with a real robot arm, outperform humans at playing Atari games, navigate in simulated 3D environments, follow instructions, and more.”
  Although the researchers’ strategy is not entirely clear, it seems that the Gato research is based on the assumption in [2] that any artificial system that can learn the kinds of things that people can learn may be seen to have achieved AGI, or it will have done after it has been scaled up. This assumption has been adopted by at least one of the authors of the [2] paper:

  “According to Doctor Nando de Freitas, a lead researcher at Google’s DeepMind, humanity is apparently on the verge of solving artificial general intelligence (AGI) within our lifetimes.

  “In response to an opinion piece penned by [the author of the article in which this quote appears], the scientist posted a thread on Twitter that began with what’s perhaps the boldest statement [seen by writers at The Guardian] from anyone at DeepMind concerning its current progress toward AGI:

  ‘… My opinion: It’s all about scale now! The Game is Over!’ (Tristan Green, The Guardian, “DeepMind researcher claims new ‘Gato’ AI could lead to AGI, says ‘the game is over!’ ”, 16 May 2022).

  There are (at least) two difficulties with the assumption that AGI is merely the ability to learn the variety of skills that can be learned by people:

  -

  There is an implicit assumption that the learning of those skills is achieved via supervised learning or reinforcement learning, (“For simplicity Gato was trained offline in a purely supervised manner; however, in principle, there is no reason it could not also be trained with either offline or online reinforcement learning (RL)” [2] (p. 3] but it appears that: most human learning is unsupervised, meaning that it is learning without predefined associations between, for example, words and pictures, and without rewards or punishments [18] (Chapter 9).

  -

  Gato suffers from a weakness that is similar to that in the SOM and described in Section 4.1 (item ‘Power’): the belief that human intelligence is merely a collection of skills, somewhat like the discredited theory that human cognition may be understood as a collection of instincts including such implausible instincts as a putting-on-of-shoes instinct, a planting-of-seeds-instinct, a car-driving instinct, and so on.

  With regard to the second point, an alternative view, adopted with varying degrees of confidence by many people working in AI, is that some capabilities are more central to the concept of intelligence than others (Section 3.2). It seems, for example, that skills like those outlined in Section 3.2 are likely to be inborn and fundamental in human intelligence, whereas others such as how to drive a car, how to play cricket, how to play the piano, and so on, are learned and not inborn, and only tangentially relevant to our concept of intelligence.
  Thus, in brief, with regard to the Power of Gato: the model of learning that has been adopted in Gato is unlikely to conform to the fundamentals of learning in people; and the emphasis in Gato on the variety of skills that the system may learn says little or nothing about the fundamentals of intelligence. It is possible that, like the SOM, Gato is too malleable to be plausible as a theory of human intelligence.
  However, since it is not impossible that human intelligence is merely a knowledge of many skills, and since Gato clearly has the potential to model many of them, the Power of Gato is hereby assigned a score of 2 (in the range 0 to 2).
• Other strengths or weaknesses, Score$=-1$. Although Gato, with appropriate training, can demonstrate a variety of capabilities, we may say in a similar way that, with the installation of appropriate apps, a smartphone or any ordinary computer can demonstrate a variety of capabilities, and these may include AI capabilities. Does this make the smartphone or ordinary computer a good FDAGI? Certainly not. This is an additional reason to doubt the potential of Gato as an FDAGI.
  Apart from that, the reliance of Gato on a variant of the DNN model is a weakness that stems from the well-known shortcomings of DNNs, most of which are summarised in Appendix B.2.2. Hence Gato has been assigned a score of $-1$ under the heading ‘other strengths or weaknesses’.
• Combined Score$=1+2-1=2$.

4.3. DALL·E 2

From the description of DALL·E 2 in Section 2.3, it is clear that in attempting to integrate two aspects of intelligence—the processing of images and the processing of natural language—it is being developed within a bottom-up strategy with its associated problems, described in Appendix E. This has a bearing on judgements both of its Simplicity and of its descriptive/explanatory Power:

• Simplicity, Score $=1$. Because DALL·E 2 is being developed via a bottom-up strategy, and because it is at a relatively early stage in that process, it may be seen to be relatively strong in terms of Simplicity.
  Although a mature version of DALL·E 2 that embraces several aspects of intelligence is likely to be relatively large, it seems best to stick with the assessment of the Simplicity of DALL·E 2 as it is now. Hence, the Simplicity of DALL·E 2 is hereby assigned a middling Score of 1.
• Power, Score $=2$. In some respects, DALL·E 2 has shortcomings in its Power to model aspects of human intelligence:

  -

  Although the bottom-up strategy for the development of DALL·E 2 seems plausible and is popular, it has apparently never proved successful (Appendix E.2). For this reason, and because DALL·E 2 is still at an early stage of that bottom-up process, its Power with respect to the development of AGI may be judged to be weak.

  -

  While some parts of the project are relevant to the development of AGI (e.g., The creation of “a new AI system that can create realistic images and art from a description in natural language”, see Section 2.3.), other parts seem to be more focussed on meeting the needs of potential users of the system (e.g., “Safety mitigations we have already developed include: Preventing Harmful Generations … Curbing Misuse … Phased Deployment Based on Learning …”), see Section 2.3. More generally, the project is not tightly focussed on modelling aspects of human intelligence.

  -

  One study concludes that “DALL·E 2 is unable reliably to infer meanings that are consistent with the syntax. These results challenge recent claims concerning the capacity of such systems to understand of human language.” [21] (Abstract).

  -

  DALL·E is a version of GPT-3 trained on both images and text [22] (p. 4, footnote). However, despite the impressive capabilities of GPT-3 with natural language, it has a tendency to produce ‘tortured phrases’ such as ‘colossal information’ instead of ‘big data’, ‘counterfeit consciousness’ instead of ‘artificial intelligence’, and more [23].

  However, it would be perverse to give DALL·E 2 a score of 0 for Power because the kinds of things it can do are undoubtedly impressive. Hence it has seemed most appropriate to give it a Power score of 2, on the scale 0 to 2.
• Other strengths or weaknesses, Score $=-1$. DALL·E 2 is a ‘transformer’ model [24], and a transformer model is a kind of DNN (Section 2.2), and DNNs have well known shortcomings compared with people and the SPTI, most of which are summarised in Appendix B.2.2. There is more detail about the shortcomings of DNNs in [25].
  As with Gato, these shortcomings are why DALL·E 2 has been assigned a score of $-1$ for ‘other strengths or weaknesses’.
• Combined Score $=1+2-1=2$.

4.4. Soar

Soar is one of the systems mentioned in Section 3.3 which is designed as an intelligence-related ‘core’ which needs to be programmed to create one or more specific systems. Hence, it is evaluated in this and following subsections in terms of its intelligence-related core features rather than the features of any of the systems that may be derived from the Soar architecture via programming.

• Simplicity, Score $=1$. In keeping with what is said about ACT-R, and the SPTI, in Section 3.3, it seems best to evaluate the Simplicity of Soar in terms of the size of its important computational ‘core’. Since that core expresses several aspects of intelligence via mechanisms for achieving those aspects of intelligence (next bullet point), and since, overall, there is little simplification via integration (as suggested by the structures shown in Figure 1), it seems reasonable to say that Soar has a middling score of 1 for Simplicity, that it is neither very complex nor very succinct.
• Power, Score $=2$. In John Laird’s book, The Soar Cognitive Architecture [6], the way in which Soar expresses aspects of human intelligence is described largely via descriptions of the mechanisms within Soar which relate to each aspect. Nevertheless, the descriptive/explanatory Power of Soar with respect to each aspect of intelligence comes over reasonably clearly. For example:

  -

  Chapter 6 describes how Soar achieves ‘chunking’, a widely-recognised aspect of human learning which became prominent largely because of George Miller’s paper about “The magical number seven, plus or minus two: …” [26].

  -

  Chapter 7 describes how reinforcement learning may be achieved via the modification or tuning of existing rules. Although unsupervised learning is probably more fundamental, there is no doubt that reinforcement can play a part in learning.

  -

  Chapter 8 (co-authored with Yongjia Wang and Nate Derbinsky) describes Soar’s semantic memory, “… a repository for long-term declarative knowledge that supplements what is contained in short-term working memory (and production memory).” [6] (p. 203).

  -

  And so on.

  Clearly, Soar embodies a definition of intelligence that is different from that described in Section 3.2. However, as noted in that section, there may be other definitions of intelligence with equal validity. Overall, it seems reasonable to say that the Power of Soar in modelling human intelligence is good, so it has been assigned a score of 2.
• Other strengths or weaknesses, Score $=0$. There seem to be no other notable strengths or weaknesses of Soar.
• Combined Score $=1+2+0=3$.

4.5. ACT-R

The background to ACT-R is similar to that of Soar, and the evaluation here is similar.

• Simplicity, Score $=1$. In keeping with what is said about Soar, ACT-R, and the SPTI, in Section 3.3, it seems best to evaluate the Simplicity of ACT-R in terms of the size of its important computational ‘core’. Since that core expresses several aspects of intelligence via mechanisms for achieving those aspects of intelligence (next bullet point), and since, overall, there is little simplification via integration (as suggested by the structures shown in Figure 2), it seems reasonable to say that the Simplicity of ACT-R is moderate, and to assign it a score of 1.
• Power, Score $=2$. ACT-R embodies a fairly detailed model of intelligence. It is different from the definition that is specified in Section 3.2 and is implicit in the structure of ACT-R (Section 2.5, second bullet point). However, as with Soar, there may be definitions of intelligence that are different from that in Section 3.2 but with equal validity. Overall, it seems reasonable to say that the Power of ACT-R in modelling human intelligence is good, and to assign it a score of 2.
• Other strengths or weaknesses, Score $=0$. There seem to be no other notable strengths or weaknesses of ACT-R.
• Combined Score $=1+2+0=3$.

4.6. NARS

Like DALL·E 2, NARS is being developed via a bottom-up strategy (Appendix E), and, so far, only two aspects of intelligence have been considered: Non-Axiomatic Logic and learning (Section 2.6). As will be described, this has a bearing on estimates of the Simplicity and Power of NARS.

• Simplicity, Score $=1$. Although a mature version of NARS that embraces several aspects of intelligence is likely to be relatively large, it seems best to stick with the assessment of the Simplicity of NARS as it is now. Hence, although the Simplicity of NARS is merely because it is at an early stage of a bottom-up process, the Simplicity of NARS is assessed as moderate, and it has been assigned a score of 1.
• Power, Score $=1$. Since NARS is still at an early stage in the integration of different aspects of intelligence, it seems reasonable to judge its descriptive/explanatory Power to be moderate, and to assign it a score of 1.
• Other strengths or weaknesses, Score $=0$. There seem to be no other notable strengths or weaknesses of NARS.
• Combined Score $=1+1+0=2$.

4.7. SPTI

• Simplicity, Score $=2$. An important feature of the SPTI (including the SPCM) is that all the intelligence-related Power of the system, summarised in the next bullet point, flows from the computational ‘core’ of the SPCM, without the need for any kind of additional learning or programming.
  Viewed as a model of natural intelligence, all of the SPCM’s intelligence-related ‘core’ may be seen as inborn capabilities, present at the system’s ‘birth’.
  That computational core is remarkably simple: it is largely the SPMA concept (Appendix A.2) and the SPTI’s procedures for unsupervised learning (Appendix A.4), much of which is the repeated application of the SPMA concept. In other words, the SPCM is largely the SPMA concept.
  In short, the SPTI with the SPCM is remarkably small, meaning that its Simplicity is strong. Accordingly, it has been given the highest available score of 2.
  A point that deserves emphasis here is that, in the SPTI, Simplicity may be combined with repetition of information, as described in Appendix C.3. In case this sounds like nonsense, the point is simply that the SPMA concept may be applied in several different aspects of intelligence: in hearing, in vision, in touch, and so on. That a powerful technique for compression of information may be applied in several different areas of brain function is a point that appears to have been missed in the ALTs described in this paper, and probably in other systems as well.
• Power, Score $=2$. The intelligence-related capabilities of the SPTI (including the SPCM), which are substantial, are described in: Appendices Appendix B.1 and Appendix B.2.1, with indications of a few exceptions not yet demonstrated in the SPCM; and Appendix B.2.2.
  In short, the SPTI is strong in its Power to model aspects of human intelligence, so it has accordingly been assigned the highest available score of 2.
• Other Strengths or Weaknesses, Score $=1$. Largely because of substantial evidence for the importance of IC in HLPC [20], IC is central in the structure and workings of the SPCM (Appendix C.5).
  In addition, a major discovery in the SP programme of research is the powerful concept of SP-multiple-alignment (Appendix A.2). This is largely responsible for the versatility of the SPCM across several aspects of intelligence, summarised above, and for the potential of the SPCM in other areas (Appendix B.2.3 and Appendix D).
  This versatility is itself due to the way in which the SPMA concept is a generalisation of six other methods for compression of information via ICMUP [27].
  Thus for two related reasons—the versatility of the SPMA in modelling human intelligence and beyond; and more generally the central role for IC in the SPTI—a score of 1 has been assigned to the SPTI under the heading ‘other strengths or weaknesses’.
• Combined Score $=2+2+1=5$.

5. Comparison of the SPTI with the ALTs

Table 1. This table summarises the evaluation scores assigned to the ALTs and the SPTI in Section 4. “Other S/W” is short for “Other strengths or weaknesses”.

System	Simplicity	Power	Other S/W	Combined Score
SOM	0	0	$-1$	$0+0-1=-1$
Gato	1	2	$-1$	$1+2-1=2$
DALL·E 2	1	2	$-1$	$1+2-1=2$
Soar	1	2	0	$1+2+0=3$
ACT-R	1	2	0	$1+2+0=3$
NARS	1	1	0	$1+1+0=2$
SPTI	2	2	1	$2+2+1=5$

summarises the evaluation scores assigned to the ALTs and the SPTI in Section 4.

Assuming that the evaluations are fair, and care has been taken to ensure that they are, the SPTI, with a combined score of 5, stands well above the ALTs, the best of which—Soar and ACT-R—each have a combined score of 3.

The main reason for the relative strength of the SPTI is the concept of SP-multiple-alignment, which is largely responsible for the versatility of the SPCM, both within AI and beyond, and for its small size.

6. Conclusions

This paper argues that the SP Theory of Intelligence and its realisation as the SP Computer Model is a promising foundation for the development of human-like broad AI, at the level of humans or beyond, also known as Artificial General Intelligence (AGI).

In that connection, the main intelligence-related strengths of the SPTI, and other strengths, are summarised in Appendix B. It appears that the SPTI has significant advantages over six other systems, chosen to be representative of potential foundations for the development of AGI.

The Simplicity of the SPTI, and its Power in modelling aspects of human intelligence, is almost entirely due to the powerful concept of SP-multiple-alignment (Appendix A.2), itself part of the ICMUP approach to the achievement of information compression (Appendix C.2).

As noted in the Introduction, it would be a mistake for all available research eggs to be put into one basket. Given the uncertainties in any vision of how AGI might be achieved, it would make more sense to hedge our bets with parallel streams of research. The SPTI qualifies as the basis for one of those streams of research.

Again, as noted in the Introduction, achieving AGI is likely to be far into the future, but research that is aiming for AGI is likely to produce many potential benefits and applications at points along the road to AGI. Some of those potential benefits and applications from the SPTI are described in Appendix B.2.