“This guy is scary,” the human player repeated during play as the bot whittled down his health.
In a surprise event Friday (Aug. 11), Elon Musk-backed research lab OpenAI announced it had built an AI bot that could beat professional e-sports players at Dota 2, an extremely popular multiplayer online battle arena (MOBA) game, in 1-on-1 matches. The full game is traditionally played 5v5.
The OpenAI bot beat its human opponent, a pro gamer called Dendi, in less than 10 minutes of play. “Dire victory,” for the OpenAI’s bot flashed across its screen after the human’s character died. A second round ended similarly, with Dendi surrendering after a few minutes.
“This guy is scary,” Dendi repeated during play as the bot whittled down his health.
OpenAI says the bot was trained on self-play—meaning it started knowing nothing about the game, slowly learning which techniques worked and which didn’t by playing a virtual version of itself. OpenAI CTO Greg Brockman said it took two weeks to train the bot, and it had previously beaten the world’s top 1v1 player and the overall top-ranked player in the world.
“We just let it play lifetimes of 1v1 against itself, and coached it on what we thought was good or bad,” Brockman said in a video played before the match. The next step, Brockman added, was training the bot to play as five players at once, and take on a full team of professionals. In the meantime, the company said it will soon let anybody play against the bot. The company released a video about how the bot learned:
It’s unclear how this compares to Facebook and DeepMind’s attempts to tackle Starcraft II, or even DeepMind’s Go victory in 2016—unlike Starcraft, there isn’t already an established community of AI researchers trying to beat the game. Dendi, however, joked that he has lost to the in-game bots before.
“It feels a little bit like a human, but at the same time it’s something else,” Dendi said of OpenAI’s bot.