Atari

Contribution to AI

Thirty-five years after the hardware stopped mattering commercially, the 2600 became the standard against which learning agents were measured. The Arcade Learning Environment, released in 2012 by Bellemare, Naddaf, Veness and Bowling on top of the hobbyist Stella emulator, wrapped hundreds of cartridges in a single interface of joystick actions, screen pixels and score. Its argument was that the games were useful precisely because game designers rather than researchers had built them: varied, genuinely hard, and free of any tilt towards the method under test. DQN used it — seven games in the 2013 preprint, forty-nine in the 2015 Nature paper — to show a single architecture and one set of hyperparameters learning from raw pixels, and for most of a decade afterwards progress in reinforcement learning was reported as a score table over those titles. Montezuma's Revenge, a 1984 cartridge on which early agents scored nothing at all, became the canonical demonstration that reward alone will not drive exploration, and the attempts to beat it produced much of the intrinsic-motivation literature. That the console was cheap, held to 128 bytes of memory, and kept running by enthusiasts long after its commercial life is what made any of this practical. The second inheritance is economic: arcade and living-room demand for chips that draw pictures quickly sustained a graphics-hardware industry whose later programmable descendants turned out to be what neural networks wanted.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).