Yesterday, OpenAI Five defeated Blitz, Cap, Fogged, Merlini, and MoonMeander—five players ranked in the top 0.05% of Dota players—by a score of 2:0. Four of these players had competed professionally. The match was livestreamed, reaching a peak audience of 100,000 viewers.
OpenAI released a scalable, task-agnostic system that achieved state-of-the-art results on a diverse range of language tasks at the time; the system combines the Transformer architecture with unsupervised pre-training, and the results have been open-sourced.
OpenAI has released the full version of Gym Retro, expanding the number of publicly available games from approximately 70 Atari titles and 30 Sega titles to over 1,000. It covers multiple backend emulators and simultaneously open-sourced tools for adding new games to the platform.
OpenAI created a Dota 2 1v1 competitive AI that defeats top-tier professional players under standard tournament rules. The AI learned entirely from scratch through self-play, without using imitation learning or tree search.
OpenAI has released the Baselines project, aimed at reproducing reinforcement learning algorithms to match the performance levels reported in papers. The initial release includes DQN and its three variants, with other algorithms expected to be released over the coming months.
OpenAI has released the public beta of OpenAI Gym, a toolkit for developing and comparing reinforcement learning algorithms. It includes an expanding set of environments such as simulated robots and Atari games, along with a website for comparing and reproducing results.
OpenAI is a non-profit artificial intelligence research company aiming to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return.