DeepMind's Gato AI Model Impresses Despite Exaggerated AGI Claims
DeepMind unveils Gato, a versatile AI model that performs over 600 distinct tasks simultaneously. While some researchers claim this brings us close to human-level AI, experts argue the real breakthrough is the model's ability to learn multiple skills without forgetting.
DeepMind unveils Gato, a new generalist AI model that performs an impressive 604 distinct tasks, including playing Atari games, captioning images, chatting, and controlling a robot arm. Following the release, a coauthor of the research paper sparks controversy by claiming on social media that the path to artificial general intelligence is simply a matter of scaling up this technology. This bold declaration fuels a wave of sensationalist media coverage suggesting human-level AI is just around the corner.
However, veteran researchers warn that this extreme hype overshadows the actual scientific merits of the project and distracts from other important areas of artificial intelligence. Similar exaggerated claims previously surround the launches of prominent AI systems like GPT-3 and DALL-E. Experts argue that framing Gato as an imminent stepping stone to superhuman intelligence misrepresents the true nature and current limitations of the technology.
The genuine innovation of Gato lies in its multitasking architecture rather than any supposed march toward artificial general intelligence. Unlike previous models such as AlphaZero, which forgets how to play Go before learning chess, Gato learns multiple different tasks at the exact same time. Although Gato does not execute these individual tasks as flawlessly as specialized single-task models, its ability to switch between hundreds of skills without catastrophic forgetting represents a small but highly significant advance in machine learning.