What is AGI (Artificial General Intelligence)? Are We Close?

Since the beginning of AI research, AGI has been intensively investigated. However, there is still disagreement among academics on what truly constitutes AGI and the best ways to develop it. While the general objective of human-like intelligence is somewhat simple, the specifics are complex and subjective. Therefore, the construction of a framework to comprehend machine intelligence as well as models capable of satisfying that framework constitute the quest of AGI.

In the hypothetical level of machine learning (ML) development known as artificial general intelligence (AGI), an artificial intelligence (AI) system is capable of matching or surpassing human cognitive abilities in any endeavor. It stands for the basic, abstract objective of AI development, which is to replicate human intelligence artificially in a machine or piece of software.

The problem is both technological and philosophical. Philosophically, a formal concept of “intelligence” and a consensus on how that intelligence might appear in AI are necessary for a formal definition of AGI. In terms of technology, AGI necessitates the development of AI models with previously unheard-of levels of complexity and adaptability, metrics and tests to consistently validate the model’s cognition, and the processing power required to maintain it.

From narrow AI to general AI

The best way to understand the concept of “general” intelligence, or general AI, is to contrast it with narrow AI, a word that essentially defines almost all contemporary AI, whose “intelligence” is only proven in particular fields.

The term “artificial intelligence” is thought to have originated with the 1956 Dartmouth Summer Research Project on Artificial Intelligence, which brought together mathematicians and scientists from Dartmouth, IBM, Harvard, and Bell Laboratories. “The study [was] to proceed based on the conjecture that every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it,” according to the plan.

The goal of this emerging field of “AI” was to create a road map for autonomous devices. However, the development of computers with human-like intelligence proved elusive in the ensuing decades.

The search for computing machines that can carry out particular jobs that normally demand a high level of intellect in humans—like playing chess, diagnosing medical conditions, forecasting, or driving a car—has advanced significantly. However, these models—like the ones that drive self-driving cars—only show intelligence in their particular fields.

At the suggestion of DeepMind cofounder Shane Legg, AI researcher Ben Goertzel popularized the term “artificial general intelligence” (AGI) in a seminal book of the same name in 2007. An artificial general intelligence would be a new kind of AI with, among other things, “the ability to solve general problems in a non-domain-restricted way, in the same sense that a human can,” in contrast to what he called “narrow AI.”

AGI vs. strong AI vs. artificial superintelligence

Strong AI and artificial superintelligence are frequently used interchangeably with artificial general intelligence (AGI), which is closely linked to other machine learning ideas. Although there is some overlap between these ideas, each represents a unique understanding of AI.

AGI vs. strong AI

The term “strong AI,” which refers to an AI system exhibiting consciousness and primarily acts as a contrast to weak AI, is a concept that is frequently explored in the writings of philosopher John Searle. They are not just synonyms, even though strong AI is typically compared to AGI (and weak AI is typically compared to narrow AI).

Strong AI is essentially a conscious mind in and of itself, whereas weak AI is merely a tool to be employed by a conscious mind, that is, a human being. Strong AI is not specifically concerned with relative performance on different tasks, even though it is usually inferred that this consciousness would require a comparable intelligence equivalent to or greater to that of humans. Because consciousness is typically regarded as either a requirement or an outcome of “general intelligence,” the two ideas are frequently confused.

In the end, strong AI and AGI represent complementary notions rather than the same, despite their similarities.

AGI vs. artificial superintelligence

As the term suggests, artificial superintelligence refers to an AI system whose general skills greatly surpass those of any human being in any task.

It’s important to note that we may have already attained a limited type of superhuman intelligence on several times, as evidenced by a few AI systems that purport to perform better than any human on the particular task or domain for which they were created. For instance:

• AlphaFold exceeds all human scientists in predicting a protein’s 3D structure from an amino acid sequence.

• IBM’s Deep Blue defeated world champion Garry Kasparov in chess in 1997.

• IBM Watson® defeated Jeopardy! champions Ken Jennings and Brad Rutter in 2013.

• AlphaGo (and its successor model, AlphaZero) is considered the world’s greatest player at Go.

These models may be a step toward artificial superintelligence, but they have not yet attained artificial “general” intelligence because such AI systems are unable to learn new tasks on their own or extend their problem-solving abilities beyond their limited scope. A true ASI would be superior in every task that a human might be expected to complete, not just in one particular area.

AGI does not require superintelligence. Theoretically, both AGI and strong AI—but not artificial superintelligence—would be represented by an AI system that exhibits consciousness and intelligence on par with that of a typical, unimpressive human.

Existing definitions of artificial general intelligence

Although several definitions have been put forth throughout the history of computer science, experts cannot agree on what precisely should be considered AGI. Instead than concentrating on the particular algorithms or machine learning models that ought to be employed in order to attain machine intelligence, these definitions typically emphasize the abstract concept of machine intelligence.

The Turing Test

In his 1950 work “Computer Machinery and Intelligence,” Alan Turing—a key contributor in the development of theoretical computer science—published one of the first and most significant definitions of machine intelligence. His main contention was that conduct, not ethereal intellectual traits, can be used to determine intelligence. Recognizing the challenge of providing precise definitions for ideas like machines and thought, Turing suggested a straightforward solution based on a party game known as the Imitation Game.

The “Turing Test” is straightforward: a human observer must examine text samples to ascertain whether they were produced by a machine or by a human. According to Turing, a program can be considered to exhibit human-like intelligence if a human cannot detect the difference between its output and that of another human.

Strong AI: Systems possessing consciousness

A higher standard for AGI is set by another definition that has been put forth: an AI system with consciousness. Strong AI holds that “the computer is not merely a tool in the study of the mind; rather, the appropriately programmed computer really is a mind,” as stated by Searles.

In 1980, Searles wrote a well-known philosophical argument against the Turing Test’s capacity to demonstrate powerful AI. He tells the story of an English speaker who has no knowledge of Chinese being imprisoned in a room filled with books with Chinese symbols and English-language instructions on how to manipulate them. He contends that even though the English speaker doesn’t understand the other person’s communications or even his own responses, he may trick someone in a nearby room into believing he can speak Chinese by just following the instructions to manipulate numbers and symbols.

Analogies to the human brain

Replicating the human brain itself is an intuitive approach to artificial general intelligence (AGI), which seeks to recreate the type of intelligence that, as far as we know, has only ever been attained by the human brain.4 This insight gave rise to the first artificial neural networks, which in turn produced the deep learning models that are today at the forefront of almost every AI subfield.

Human-level performance on cognitive tasks

Simply defining AGI as an AI system capable of doing all cognitive tasks that humans can perform is a more comprehensive approach. Although this concept is intuitive and helpful, it remains unclear: which tasks? Which individuals? Its practical application as a formal framework for AGI is restricted by this ambiguity.

This framework’s most significant contribution is that it restricts the application of AGI to non-physical jobs. This ignores abilities that are frequently regarded as examples of “physical intelligence,” such as the use of physical tools, movement, or object manipulation. This removes the need for additional robotics development as a precondition for the creation of AGI.

Ability to learn new tasks

Emphasizing the capacity to learn—more especially, the capacity to acquire as wide a range of tasks and concepts as humans are capable of—is another intuitive approach to AGI and intelligence in general. This is similar to Turing’s theory in “Computing Machinery and Intelligence,” which suggests that rather than directly programming a computer system as an adult mind, it could be preferable to create a childlike AI and give it a period of instruction.

Even though cutting-edge multimodal AI models can handle a wider range of tasks, such as computer vision, speech recognition, and natural language processing (NLP), they are still restricted to a limited set of fundamental abilities that are reflected in their training data sets. For example, they are unable to learn how to drive a car. A true AGI would be able to learn from new experiences in real time, something that many animals and even human infants cannot accomplish.

Economically valuable work

AGI is defined in its charter as “highly autonomous systems that outperform humans at most economically valuable work” by Open AI, whose GPT-3 model is frequently credited with starting the current generative AI era with the debut of ChatGPT.

According to the DeepMind report, this definition leaves out aspects of human intelligence that are difficult to quantify economically, including emotional intelligence or artistic inventiveness. At most, such facets of intelligence can provide economic value indirectly, such as when creativity generates successful films or emotional intelligence drives psychotherapy equipment.

Flexible and general capabilities

AGI is “a shorthand for any intelligence…that is flexible and general, with resourcefulness and reliability comparable to (or beyond) human intelligence,” according to psychologist, cognitive scientist, and AI researcher Gary Marcus. Similar to a particular and useful application of the “learn tasks” paradigm, Marcus suggested a collection of benchmark tasks meant to show that flexibility and general competency.

“Artificial Capable Intelligence”

Mustafa Suleyman, the CEO of Microsoft AI and co-founder of DeepMind, coined the term “Artificial Capable Intelligence” (ACI) in 2023 to refer to AI systems that are capable of carrying out intricate, multistep, open-ended activities in the actual world. More precisely, he suggested a “Modern Turning Test” in which an AI would be required with developing USD 100,000 in seed money into USD 1 million.Twelve In general, this combines Marcus’s emphasis on adaptability and general intelligence with OpenAI’s concept of economic value.

In practical terms, this definition of intelligence as a certain type of economic production is excessively restrictive, even though this standard probably demonstrates true ingenuity and transdisciplinary expertise. Furthermore, there are serious alignment hazards when profit is the only consideration.