DeepMind's Gato does 600 tasks with one set of weights

DeepMind published “A Generalist Agent,” describing Gato, a single 1.2-billion-parameter transformer trained on 604 distinct tasks spanning several modalities: playing Atari games, captioning images, holding text-based dialogue, and controlling a real robot arm to stack blocks. Every input — image pixels, joint torques, button presses, text — was serialised into the same token sequence format and handled by one network with one set of weights, rather than by separate models switched between tasks.

DeepMind framed the result as a step toward generalist agents rather than a finished capability: on many of the individual tasks, Gato performed below dedicated single-task models, and its context window of 1,024 tokens limited how much history it could draw on. Its contribution, in DeepMind’s account, was that a single policy could hold competence across a very wide range of embodiments and task types without catastrophically forgetting earlier ones — a property that had eluded most prior multi-task reinforcement-learning systems.

The release triggered a public argument disproportionate to the paper’s own claims after Nando de Freitas, a DeepMind research director, posted on Twitter that “the game is over” and that reaching artificial general intelligence was now “just about scale.” The remark drew wide press coverage suggesting DeepMind believed it was close to human-level AI, prompting pushback from researchers including Gary Marcus, who argued Gato remained far from anything resembling general intelligence, and MIT’s Jacob Andreas, who noted the model was still far from being able to adapt to genuinely new, untrained tasks. De Freitas later said the phrasing was a poor choice, intended as a call to stop debating symbolic-versus-connectionist approaches rather than as a claim about imminent AGI.

Gato was not released as a product and was followed by no direct successor bearing its name, but the episode became a frequently cited example of the gap between a paper’s technical claims and the public reception generated by a single line of commentary from within the lab that produced it.