OpenAI, a non-profit research organization co-founded by Elon Musk, first used machine learning to teach a bot how to play Dota 2 last year. It was able to best various pro players in one-on-one games and took down the legendary Danil "Dendi" Ishutin in a live match at The International 2017. The bot wasn't unbeatable, however, and was still very one-dimensional. It had mastered just one of over 100 heroes and Dota 2 is a complex, team game -- one-on-one matches test a very specific skill set that isn't the be all and end all of a proper match.
OpenAI Five, a full team of bots working together, was the natural next step. The squad trains exclusively by sparring against itself, racking up 180 years of playtime every day. In mirror matches, where both teams play the same heroes, OpenAI Five has taken games off amateurs and semi-professional players alike. But this past weekend was the all-bot team's toughest test yet. While the games were still played with a special rule set that eliminates some of the more advanced mechanics, both teams were able to pick from a pool of 18 heroes and make use of a few strategic items and objectives that have previously been off-limits. OpenAI Five's reaction time was increased from 80ms to a more human-like 200ms, though, to even the field for the meatsack team.
OpenAI Five steamrolled the first two games, displaying excellent teamwork, but when the audience was asked to pick the bots' heroes for the final game, it all fell apart. Dota 2 has a lot to with how different heroes complement each other and counter the enemy, so the human team were able to secure the win with their heavily favored lineup. Still, OpenAI Five won the best of three series against some highly skilled players. The all-bot team's next test is only a few weeks away, when they will take on a professional squad at the top of the game during The International main event. First it was Go, now it's Dota 2. The machines are taking over.