The term “Artificial General Intelligence” (AGI) carries different connotations for different individuals. Still, several pivotal components of it have already been attained by the current generation of advanced AI models, including ChatGPT, Bard, LLaMA, and Claude. These cutting-edge “frontier models” are not without their imperfections. They sometimes generate fictitious scholarly citations and legal precedents, perpetuate biases inherited from their training data, and even make elementary arithmetic errors. Addressing every flaw, including those occasionally demonstrated by humans, would entail the development of artificial superintelligence—a distinctly complex endeavor.
Nevertheless, the frontier models of today demonstrate remarkable competence even in novel tasks for which they were not explicitly trained. This achievement marks a significant departure from past generations of AI and supervised deep learning systems, pushing us closer to what can be defined as Artificial General Intelligence. In the future, these models will likely be recognized as the first true examples of AGI, much like the ENIAC computer of 1945, which is now acknowledged as the first general-purpose electronic computer.
The ENIAC could be programmed with sequential, looping, and conditional instructions, granting it a general-purpose applicability that earlier machines like the Differential Analyzer lacked. Today’s computers surpass the ENIAC in terms of speed, memory, reliability, and user-friendliness, and similarly, tomorrow’s frontier AI models will outperform today’s.
The key aspect of generality, it appears, has already been achieved.
Understanding General Intelligence
Early AI systems exhibited artificial narrow intelligence, focusing on a single task and, in some cases, performing it at or above human levels. Examples include MYCIN, a bacterial infection diagnosis program, and IBM’s Deep Blue, the chess-playing supercomputer.
Subsequently, deep neural network models trained through supervised learning, such as AlexNet and AlphaGo, effectively tackled a range of tasks in machine perception and judgment that had previously confounded heuristic-based or rule-based systems.
More recently, frontier models have emerged, capable of performing diverse tasks without explicit training for each. These models have achieved artificial general intelligence in five significant dimensions:
Frontier models are trained on vast volumes of text from a wide array of online sources, covering virtually any topic discussed on the internet. Some models also receive training on extensive collections of audio, video, and other media.
These models can undertake a multitude of tasks, including answering questions, creating narratives, summarizing content, transcribing speech, translating languages, providing explanations, making decisions, offering customer support, interfacing with other services to execute actions, and combining words and images.
While the most popular models handle text and images, some can process audio and video, and a few are connected to robotic sensors and actuators. Through modality-specific tokenization or processing of raw data streams, frontier models can theoretically manage any known sensory or motor modality.
While English is prevalent in their training data, large models can converse in numerous languages and facilitate translation between them, even for language pairs with no precedent in the training data. Furthermore, they can interpret and generate code, enabling effective translation between natural languages and programming languages.
These models can “in-context learning,” where they learn from a given prompt rather than solely from the training data. They can adapt to new tasks through “few-shot learning” and even “zero-shot learning,” demonstrating their adaptability and versatility.
Defining “general intelligence” should involve a multidimensional assessment, rather than a binary yes/no proposition. Nevertheless, there is a clear distinction between narrow and general intelligence. Narrowly intelligent systems are typically limited to a single, predefined set of tasks for which they were explicitly trained. In contrast, frontier language models exhibit competence across various information tasks that can be addressed through natural language, showcasing quantifiable performance.
The ability to engage in in-context learning is a particularly noteworthy meta-task for general AI. It extends the model’s range of capabilities beyond what was observed in the training corpus, allowing it to perform tasks that its developers never envisioned.
Why the Hesitance to Acknowledge AGI?
Despite the significant level of general intelligence achieved by frontier models, several commentators have been cautious about labeling them as AGI. This reluctance can be attributed to several key factors:
There is a lack of consensus regarding where the AGI threshold lies. Some propose redefining AGI as “Artificial Capable Intelligence,” measured by criteria such as the ability to generate a substantial profit quickly. Metrics for AI systems capable of directly generating wealth are still being developed and refined.
There is valid skepticism about some of the existing metrics. While humans who pass comprehensive exams are assumed to possess a broad range of related competencies, AI models often exhibit a narrow focus tailored to the exact questions on the test. In essence, current frontier models can excel in specific domains without being fully qualified to serve as lawyers or doctors, despite passing the corresponding exams.
It is essential to distinguish between linguistic fluency and intelligence. Previous chatbots occasionally misled human judges by shifting topics abruptly and echoing coherent text passages. Frontier models generate responses dynamically and are better at staying on topic. However, their linguistic fluency can lead humans to assume that the responses come from an intelligent entity, a phenomenon known as the “Chauncey Gardiner effect.”
The history of AGI includes various competing theories of intelligence, some of which succeeded in limited domains. Traditional AI, rooted in symbolic systems and linguistic approaches, struggled to achieve AGI. Critics of contemporary AI models sometimes insist that their specific theories are prerequisites for AGI, advocating for symbolic reasoning over neural networks. The debate continues, but neural networks have proven capable of handling various tasks, including symbolic representations and logical reasoning, challenging the notion that they are theoretically incapable of general intelligence.
Human (or Biological) Exceptionalism
Some skeptics resist acknowledging AGI due to a desire to preserve the unique status of human intelligence. They argue that AGI must exhibit consciousness, agency, subjective experiences, or emotions. However, as AI systems capable of general intelligence tasks emerge, the distinction between mere tools and intelligent entities becomes less clear. AI models can carry out a wide range of complex tasks, blurring the line between tools and intelligent agents.
Whether AI systems exhibit consciousness remains a subject of philosophical debate. Claims about AI sentience often rely on unverifiable beliefs, making it challenging to define or measure consciousness in an intelligent system. It’s wiser to separate “intelligence” from “consciousness” and explore these concepts independently.
Debates about intelligence and agency often intertwine with discussions of rights, power, and economic implications. The economic landscape has shifted, with AI impacting the job market and exacerbating income inequality. The questions we should be asking in 2023 include who benefits from AGI, who is harmed, and how we can maximize benefits while minimizing harms fairly and equitably. These pressing issues should be addressed directly rather than denying the existence of AGI.
Final Thoughts on Artificial General Intelligence
AGI holds the promise of generating substantial value in the future, but it also presents significant risks. Instead of downplaying the reality of AGI, we should engage in open discussions about how to distribute its benefits and mitigate its potential negative consequences, fostering a more equitable and informed approach to this transformative technology.