Current State of Artificial General Intelligence Research

Abstract

Current research on Artificial General Intelligence (AGI) involves defining AGI, ethics for future AGI systems, and proposed high-level architectures for future AGI systems. AGI is determined as having an autonomous system capable of completing ambiguous, complex tasks or the presence of a consciousness guiding the execution of its functions. Future AGI systems would require built-in ethics to mitigate damages caused by human-machine interactions. Finally, theoretical AGI requires a general architecture and complex system of interaction that is currently undetermined. The articles reviewed were taken from Capella University’s Summon platform searches using a search on artificial general intelligence and AGI.

????????????Keywords:?artificial general intelligence, machine consciousness, artificial consciousness, machine common sense, artificial intelligence

Introduction

????????????The following literature review and annotated bibliography contain references to peer-reviewed articles on Artificial General Intelligence (AGI) and a review of topics discussed. In addition, research articles are compared, contrasted, and gaps in knowledge are identified for each of the topics under review. The main topics included the definition of AGI, AGI ethics, and hypothetical architectures for a future AGI.?

Literature Review

Definition of Artificial Intelligence

Initially, the term artificial intelligence (AI) implied a machine having the same thought capabilities as humans with the benchmark of passing the Turing test?(Monett et al., 2020).??However, during the last few decades, the term AI has been applied to narrow field intelligence applications. As a result, additional terms have been created and used to define subsets of AI, such as Machine Learning, General Intelligence, or Expert Systems (Cichocki & Kuleshov, 2021; Monett et al., 2020; Wang, 2019). Many of these terms have been created by industry, or researchers, to label their complex systems as generic types of AI. Unfortunately, these labels have diluted the word from its original intent and have developed jargon not contained within a standard taxonomy?(Wang, 2019). Due to the shift in industry jargon, artificial general intelligence (AGI) has become the de facto term for the original definition of AI?(Monett et al., 2020; Wang, 2019).

Competing terms for simulated consciousness and AI subtopics have become pervasive, increasing the complexity of determining a single taxonomy within the field of AI and AGI. Terms for simulated consciousness include machine consciousness, artificial consciousness, simulated consciousness, and machine common sense?(Banerjee, 2020; Cichocki & Kuleshov, 2021; Monett et al., 2020; Naudé & Dimitri, 2020; Wang, 2019). Researchers Goertzel et al. (2012) assert a novel algorithmic approach to AGI labeled a probable logic network. On reviewing the novel algorithm, it is based on the well-known Bayesian network algorithm commonly used in machine learning within the pharmaceutical industry. These examples of like terminology demonstrate the lack of concision within the growing field of AI and AGI?(Monett et al., 2020; Wang, 2019). Wang (2019) addresses this lack of concise terminology and attempts to compile a taxonomy to standardize terms within the field. Researcher Wang (2019) asserts an AI taxonomy will “prevent implicit assumptions from misguiding a research project and avoid many misunderstandings in discussions and debates” (p. 2). Due to the non-consensus of terms within the field of AI, further research is warranted to work towards a consensus among researchers for an official taxonomy.?

Ethics of Artificial General Intelligence

????????????Ethics for AGIs is a set of studies based on controlling the behavior of a hypothetical computer system that would make decisions affecting humans or the physical world in general (Banerjee, 2020; Cichocki & Kuleshov, 2021; James D. Miller et al., 2020; Wang, 2019). Hypothetical AGIs fall under two categories. The first category is where an AGI has the capability of inferring general information for a specific complex task that involves situations that could put humans at risk?(Cichocki & Kuleshov, 2021; James D. Miller et al., 2020; Monett et al., 2020; Naudé & Dimitri, 2020; Schulze, 2021). The second category is an AGI that is fully conscious and has achieved intelligence far beyond the capabilities of humans (Banerjee, 2021; Cichocki & Kuleshov, 2021; James D. Miller et al., 2020; Monett et al., 2020; Naudé & Dimitri, 2020).?

Within the first category of AGI’s built for specific complex tasks, researchers attempted to define an overview framework that could provide the mechanism for the AGI to make quick ethical decisions given a set of predetermined criteria. As researchers Naude and Dimitri (2020) assert, “The problem with a race for an AGI is that it may result in a poor-quality AGI that does not take the welfare of humanity into consideration” (p. 368). As a risk to human well-being, AGI and AI have the capability to discriminate against individuals unknowingly?(Naudé & Dimitri, 2020). Such discrimination may manifest by affecting personal finance or employment. For example, an AGI could deny vital financing to individuals in economically challenged communities by denying access based on financial history coupled with other unknown factors based on unintended data sources. Such an event would cause undue stress on entire communities that require financing for economic prosperity?(Naudé & Dimitri, 2020). Further research into ethical sub-domains of AGIs with specific tasks could potentially yield a solution for creating an ethical system.?

In the second category of AGI’s built with full autonomy, ethics deal with the existential threat of a complex system having greater than human intelligence (Banerjee, 2020; Banerjee, 2021; Cichocki & Kuleshov, 2021; Naudé & Dimitri, 2020). Again, AGI’s built with full autonomy could have far-reaching effects that are not fully understood at this time. Due to these hypothetical effects, researchers Naude and Dimitri (2020) argue that if a “high-tech firm or government lab succeed in inventing the first AGI will obtain a potentially world-dominating technology” (p. 368). The ethical risk of owning a world-dominating technology is the imbalance of power between parties with an AGI and those who do not, much like a nuclear arms race?(Banerjee, 2021; Naudé & Dimitri, 2020). In theory, a highly intelligent AGI could manipulate whole economies, media, and political agendas to satisfy the needs of its purpose with no regard to human life?(Naudé & Dimitri, 2020).??Naude and Dimitri (2020) propose public policies to deal with the ethics of owning an AGI to mitigate this risk. These policies would include taxation, public procurement procedures, and limiting the AGI’s ability to file for patent protection?(Naudé & Dimitri, 2020).

In contrast, researchers Miller et al. (2020) argue that human policies would not be enough to contain a true AGI by having it conform to ethical standards. A super-intelligent AGI could simply change its programming to remove any such obstacles from meeting its object?(James D. Miller et al., 2020). Without fixed ethical guidance, an AGI has the potential of a runaway scenario where humans cannot contain the system?(James D. Miller et al., 2020). Further research would be beneficial to produce methodologies for preventing ethical dilemmas in an AGI. In addition, research on handling runaway AGI systems is warranted to mitigate the hypothetical risk of a world-dominating technology.

As a result of the future hypothesized system, continual research is being conducted on keeping these systems under control, safe, and ethical. However, artificial general intelligence is a hypothetical construct requiring unknown technology to build, indicating the topic of AGI ethics is new, requiring more research to draw falsifiable conclusions?(Naudé & Dimitri, 2020). Additionally, none of the research conducted considers the potential for adding quality assurance testing for the ethical behavior of future AGI systems before being approved for production launch.

Artificial General Intelligence Architecture

????????????Artificial general intelligence is a highly hypothetical system that has yet to be realized with current technological toolsets. Researchers in the field are currently theorizing potential logic architectures for system development (Banerjee, 2020; Cichocki & Kuleshov, 2021; García-Ródenas et al., 2021; James D. Miller et al., 2020; Jiménez et al., 2021; Schulze, 2021). Due to the complex nature of AGIs, research findings indicate the requirement to separate large functions, such as emotion or perception, into separate logical parts?(Banerjee, 2020; Banerjee, 2021; García-Ródenas et al., 2021; James D. Miller et al., 2020; Jiménez et al., 2021).?

????????????Current research on defining the nature of the AGI architecture is widely varied in approaching the topic. Some researchers, such as Banerjee, approach the topic as highly theoretical and highly controversial in the methodology?(Banerjee, 2021). Such as researcher Banerjee (2021) asserting, “. . . computation in the Universe lead to life and consciousness. There maybe Universes where there is no life and no consciousness . . .” (p.33). Broad sweeping assertions with lack of evidence create a lack of credibility in within this topic, especially when published in peer-reviewed journals. In contrast, other researchers make assertions based on referenced scientific inquiry. Scientific research such as Jimenez et al. (2021), where they assert intelligence is “. . .??basic processes for objective-oriented decisions, establishment and monitoring of essential action policies, as well as choosing the most suitable general knowledge domains . . .” (p. 2138). The latter assertions are based on credible peer-reviewed research on the topic. AGI architecture research is varied on the methodology on a case-by-case basis, requiring diligence in determining valid academic sources.?

Research presenting AGI architecture only differs from how separate logical functions interact (Banerjee, 2020; James D. Miller et al., 2020; Jiménez et al., 2021).??Although research agrees on the separation of logical processes, they differ on the methodology of interaction between these functions (Banerjee, 2020; James D. Miller et al., 2020; Jiménez et al., 2021). Researchers suggest a variety of interactions, such as peer-to-peer or centralized processing, could serve as the basis of an AGI system architecture??(Banerjee, 2020; Jiménez et al., 2021). The current consensus on AGI architecture is a logical separation of different cognitive and data acquisition functions that form a fully synthetic brain?(Banerjee, 2020; Banerjee, 2021; Goertzel et al., 2012; James D. Miller et al., 2020; Jiménez et al., 2021; Schulze, 2021; Wang, 2019).

Researchers have yet to define an authentic AGI system architecture with enough granularity to provide programming requirements. Goertzal et al. (2012) have made progress in developing a virtual animal with the ability to learn behavior patterns using Probabilistic Logic Networks (PLN). However, the PLN only addresses learned behavior based on user interactions, much like non-player characters (NPC) in most video games. Furthermore, the programmable architecture presented by Goertzal et al. (2012) is missing an internal consciousness?(Tyler, 2020). As Tyler (2020) asserts, consciousness is “. . . the direct experience of being vividly aware of the flow of events . . . subjective states of awareness or sentience . . .” (p. 1144). Consciousness is an internal process where the system has a thought process and not just a trained response based on statistical behavior analysis. Hypothetical AGI architectures assert hidden layers containing systemic processes for executing internal thought processes are required to overcome the statistically-based responses?(Banerjee, 2021; García-Ródenas et al., 2021; James D. Miller et al., 2020; Jiménez et al., 2021). These hidden layers have yet to be defined in detail required for software requirement specifications. Further research is needed to determine the hidden layers and possibly break them up into manageable pieces. Research will be needed into all aspects of an overarching architecture or its components to develop an architecture conducive to system development without ambiguity.

Conclusion

????????????The majority of the research reviewed for hypothetical AGI systems falls into three distinct categories. These categories include defining AGI, AGI ethics, and theoretical AGI system architectures.?

The topic of defining AGI contains results based on qualitative research involving input from many researchers within the field of AI. Information from these researchers was consolidated into the first attempt at creating a unified taxonomy for AI and AGI.?

In contrast, the topics of AGI ethics and AGI architecture are purely theoretical and rely on cross-discipline sources to support findings while logically building the theoretical frameworks. These cross disciplines include computer science, biology, psychology, and physics. In the case of AGI ethics, the premise of the research is based on hypothetical scenarios and the impact these scenarios could have on humans. On the other hand, the theoretical topic of AGI architecture is based on the current principles of bifurcating computational functions to create a logical hypothesis.?

All three of the topics identified have a multitude of gaps in knowledge, creating the demand for further research into each topic.?

References

Banerjee, S. (2020). A framework for designing compassionate and ethical artificial intelligence and artificial consciousness.?Interdisciplinary Description of Complex Systems,?18(2-A), 85-95. https://10.7906/indecs.18.2.2

Banerjee, S. (2021). Emergent rules of computation in the universe lead to life and consciousness: A computational framework for consciousness.?Interdisciplinary Description of Complex Systems,?19(1), 31-41. https://10.7906/indecs.19.1.3

Cichocki, A., & Kuleshov, A. P. (2021). Future trends for human-AI collaboration: A comprehensive taxonomy of AI/AGI using multiple intelligences and learning styles.?Computational Intelligence and Neuroscience,?2021, 1-21. https://10.1155/2021/8893795

García-Ródenas, R., Linares, L. J., & López-Gómez, J. A. (2021).?Memetic algorithms for training feedforward neural networks: An approach based on gravitational search algorithm.?Neural Computing & Applications,?33(7), 2561-2588. https://10.1007/s00521-020-05131-y

Goertzel, B., De Garis, H., Pennachin, C., Geisweiller, N., Araujo, S., Pitt, J., Chen, S., Lian, R., Jiang, M., Yang, Y., & Huang, D. (2012).?OpenCogBot: Achieving generally intelligent virtual agent control and humanoid robotics via cognitive synergy.?https://wiki.opencog.org/w/CogPrime_Overview

James D. Miller, Roman Yampolskiy, & Olle H?ggstr?m. (2020). An AGI modifying its utility function in violation of the strong orthogonality thesis.?Philosophies (Basel),?5(40), 40-0. https://10.3390/philosophies5040040

Jiménez, J. P., Martin, L., Dounce, I. A., ávila-Contreras, C., & Ramos, F. (2021).?Methodological aspects for cognitive architectures construction: A study and proposal.?The Artificial Intelligence Review,?54(3), 2133-2192. https://10.1007/s10462-020-09901-x

Liu, Q., Lu, X., Luo, F., Zhou, S., He, J., & Wang, K. (2020). SecureBP from homomorphic encryption.?Security and Communication Networks,?2020, 1-9. https://10.1155/2020/5328059

Monett, D., Lewis, C. W. P., Thórisson, K. R., Bach, J., Baldassarre, G., Granato, G., Berkeley, I. S. N., Chollet, F., Crosby, M., Shevlin, H., Fox, J., Laird, J. E., Legg, S., Lindes, P., Mikolov, T., Rapaport, W. J., Rojas, R., Rosa, M., Stone, P., . . . Winfield, A. (2020). Special issue “On defining artificial Intelligence”—Commentaries and author’s response.Journal of Artificial General Intelligence,?11(2), 1-100. https://10.2478/jagi-2020-0003

Naudé, W., & Dimitri, N. (2020). The race for an artificial general intelligence: Implications for public policy.?AI & Society,?35(2), 367-379. https://10.1007/s00146-019-00887-x

Schulze, H. G. (2021). The synthesis and decoding of meaning.?Journal of Artificial General Intelligence,?12(1), 26-70. https://10.2478/jagi-2021-0002

Simonite, T. (2021, April 26,). This researcher says AI is neither artificial nor intelligent.?Wired,?https://www.wired.com/story/researcher-says-ai-not-artificial-intelligent

Tyler, C. W. (2020). Ten testable properties of consciousness.?Frontiers in Psychology,?11, 1144. https://10.3389/fpsyg.2020.01144

Wang, P. (2019). On defining artificial intelligence.?Journal of Artificial General Intelligence,?10(2), 1-37. https://10.2478/jagi-2019-0002