In a recent breakthrough, Google's DeepMind researchers have introduced a comprehensive framework for defining and evaluating Artificial General Intelligence (AGI), titled the "Levels of AGI." This initiative aims to bring clarity to the AGI landscape, similar to the widely accepted six levels of autonomous driving. The researchers believe that establishing a standardized definition is crucial for the AI community, steering away from vague interpretations and providing a common ground for evaluating AI models.
AGI, often perceived as the ultimate goal in artificial intelligence, lacks a universally agreed-upon definition. While some consider it highly autonomous systems surpassing humans in economically valuable tasks, others stress the need for flexibility, adaptability, and creativity across various domains. DeepMind's proposal seeks to address these ambiguities by focusing on the capabilities of AGI, rather than the methods it employs.
Drawing inspiration from the six-level system used in autonomous driving, the Levels of AGI range from "emerging" (comparable to an unskilled person) to "superhuman" (surpassing any human). Notably, the team emphasizes that AGI doesn't necessarily require physical embodiment in a robot to interact with the world. Instead, the focus should be on intelligent thinking tasks, with an emphasis on learning how to learn.
The researchers argue that AGI should be measured by its ability to perform tasks that matter to people, observing its development over time rather than fixating on a singular final goal. This approach diverges from the notion that AGI must be embedded in a physical entity and suggests that measuring progress should encompass a spectrum of intelligent functions.
While some experts predict the imminent arrival of AGI, others, including Meta's Chief AI Scientist Yann LeCun, challenge the very term, proposing the use of 'human-level AI' instead. Despite the varying opinions, the Levels of AGI provide a structured framework for evaluating AGI progress, from models like DeepMind's AlphaFold and advanced chatbots such as OpenAI's ChatGPT to potential future advancements.
As the AI community grapples with the possibilities and risks associated with AGI, the Levels of AGI framework offers a valuable tool for assessing performance benchmarks, generality, and autonomy. Whether AGI is on the horizon or a more distant prospect, the industry now has a standardized reference point to navigate the evolving landscape of artificial intelligence.
