Gato

Gato is DeepMind's 2022 attempt at a generalist agent: a single 1.18-billion-parameter decoder-only transformer trained to perform 604 distinct tasks with one set of weights. Announced in May 2022, it ingests images, text, proprioception, joint torques, button presses and other modalities, all serialized into token sequences, and outputs whatever the task demands: natural language, motor control signals, game actions. The task mix spans dialogue, image captioning, Atari games, simulated control and stacking blocks with a real robot arm.

The headline result was breadth rather than depth: Gato exceeded 50 percent of expert score on more than 450 of the 604 tasks, though it trailed task-specific systems everywhere. DeepMind presented it as evidence that a single architecture could span embodiments and modalities, a step toward general-purpose agency. Critics noted it was essentially many individually trained policies distilled into one network rather than an agent that learns new tasks from scratch. Gato was a research system with no public release, so openWeights is false. Its conceptual legacy lives on in generalist agent research and vision-language-action models.