DeepMind Unveils Gato: A Single AI Model for Hundreds of Diverse Tasks
DeepMind introduces Gato, a general-purpose transformer that performs over 600 tasks like playing Atari, controlling robots, and captioning images using a single set of network weights. This multi-modal approach represents a significant step toward artificial general intelligence.
DeepMind introduces Gato, a single general-purpose transformer sequence model that functions like an AI Swiss Army knife. This innovative system performs over 600 diverse tasks—ranging from engaging in dialogue and captioning images to stacking blocks with a real robot arm and navigating simulated 3D environments—all while using the exact same network and weights.
The researchers hypothesize that training an agent on a massive variety of tasks enables it to adapt to new challenges with very little additional data. By serializing different types of multi-modal data into a flat sequence of tokens, Gato successfully handles varying modalities, observations, and action specifications without requiring hand-crafted policy models for each specific field.
Gato trains on a wide array of datasets, including agent experiences in simulated and real-world environments, along with vision and language data from sources like MassiveText and COCO captions. In evaluations, the generalist model even beats human players at Atari games, leading the DeepMind team to view this multi-task, multi-embodiment system as a meaningful step toward artificial general intelligence.