What is AGI anyway?

What is AGI anyway?

Introduction

Artificial General Intelligence (AGI), refers to a hypothetical type of artificial intelligence with the ability to surpass human cognitive abilities. Unlike today’s AI, which excels at specific tasks like playing chess or recognizing faces, AGI aspires to a broader kind of intelligence.

Imagine a machine that can not only master complex games but also learn new languages, write poetry, solve scientific problems, and adapt to unforeseen situations — that’s the essence of AGI.

Creating AGI is one of the primary goals of companies such as OpenAI, Google and Anthropic. These companies are hoping to unlock the secrets of human-level intelligence in machines.

Currently, there’s no universally accepted benchmark for proving the existence of AGI. The ongoing debate centers around various proposed tests. One of the most well-known is the Turing Test, devised by Alan Turing in 1950. This test proposes that a machine can be considered intelligent if it can hold a conversation indistinguishable from a human. Another interesting proposition is the coffee test by legendary Steve Wozniak. This test is more anecdotal, suggesting that an AGI should be able to seamlessly interact with humans in everyday situations, like having a conversation over coffee, without revealing its artificial nature. More recent proposals include the modern Turing Test, which expands on the original concept by incorporating judges who can evaluate the machine’s ability to reason, solve problems, and adapt to new situations. The debate continues to evolve, with researchers proposing new tests that search deeper into different aspects of human intelligence, such as creativity, common sense reasoning, and the ability to learn and transfer knowledge across domains.

Academic Definition

There isn’t one single, universally accepted academic definition of AGI. The field is still young, and there’s ongoing debate about what exactly constitutes general intelligence, even in humans. Here are some attempts of definitions found in academic discussions of AGI.

Human-Level Performance

AGI is often defined by its ability to perform a wide range of cognitive tasks at a level comparable to or exceeding humans. This goes beyond just mimicking human behavior; it involves genuine understanding and the ability to adapt to new situations.

In the article Proceedings of the 3d Conference on Artificial General Intelligence (2010) (pdf) Laird and Wray outlined different characteristics with multiple requirements for human-level intelligence. Their conclusion is the list is not yet sufficient to guarantee the achievement of human-level intelligence when met.

Generality

Unlike today’s AI that excels at specific tasks, AGI wouldn’t be limited to a single domain. It would be able to learn and apply its knowledge across various fields, demonstrating flexibility and adaptability.

In the article Levels of AGI: Operationalizing Progress on the Path to AGI (pdf), a framework for classifying AGI was proposed and it included performance, generality and autonomy but explained that for now “if
you were to ask 100 AI experts to define what they mean by AGI you would likely get 100 related but different definitions”.

Learning and Reasoning

True intelligence involves more than just accessing and processing information. An AGI system would be expected to learn new things, reason through problems, and draw sound conclusions based on incomplete or uncertain information.

In the article When brain-inspired AI meets AGI (link) the evolution if AGI systems is investigated. The author’s conclusion was that “While significant progress has been made in the development of AGI and brain-inspired AI, there are still several limitations that need to be overcome before we can achieve true human-level intelligence in machines”.

Common Sense Reasoning

A key aspect of human intelligence is the ability to apply common sense in everyday situations. An AGI might be expected to understand social cues, make inferences, and navigate the complexities of the real world.

In the article Artificial Intelligence and Common Sense: The Shady Future of AI (link), the authors concluded that “The experiment results reassert the weakness of AI and the requirement of AGI.”

Self-Awareness and Consciousness

Some definitions of AGI might encompass a level of self-awareness or consciousness similar to humans.

In the article Consciousness for AGI (pdf) the author Boltuc Piotr emphasises that “If Artificial General Intelligence (AGI) is to dwell in the world of meaningful existences, not just their shadows, as the case of Church-Turing Lovers highlights, it requires full epistemic subjectivity, meeting the standards of the Engineering Thesis in Machine Consciousness”.

Is AGI a Buzzword?

There is a head-to-head battle among AI companies to create the most advanced model, with the term AGI often used as a benchmark to claim superiority.

Sam Altman, CEO of OpenAI in the article Planning for AGI and beyond explains how AGI can benefit humanity, but does not give a clear definition of how OpenAI’s model will achieve AGI.

Google DeepMind’s researchers take a more modest approach. They came up with a framework to define the different levels of AGI and why is important (pdf).

I couldn’t find any references of Anthropic regarding AGI besides this youtube video where the CEO Jared Kaplan is asked if we will achieve AGI and he avoided directly addressing the term.

What do the Twitter Experts Think

In the previous paragraphs I analysed what the researchers and the main companies views about AGI, it’s only fair to share some tweets of “AI Twitter Experts”.

Generality

Someone claims that AGI is already here.

Someone else claiming that the reason Sam Altman was removed from OpenAI as CEO last year (2023) was because they achieved AGI.

Self-awareness and Consciousness

Someone, took a screenshot of the chat prompt and then asked the AI about the screenshot claiming that this is his version of self-awareness test.

Common Sense

Someone argues that AI is superior to most humans.

Do we Really Need AGI?

While AI offers significant benefits in automating tasks and improving lives, the concept of AGI remains an ambiguous term, at least in the academic world. Researchers haven’t established a clear definition or method for measuring its achievement. This lack of clarity allows AI companies to liberally use AGI as a marketing term, often inflating their progress to appear ahead of the competition. Perhaps, the true focus should be on developing robust and practical AI solutions for real-world problems, rather than chasing an ill-defined benchmark.