The OpenAI team, supported by tech maven Elon Musk, showcased an AI bot at a tournament in Seattle that decisively beat several of the world’s best Dota 2 players in one-on-one matches. The stunning upset over pro gamer and crowd favorite Danylo “Dendi” Isutin was broadcast live from the stage at The International, a $24 million Dota 2 tournament backed by Valve.
In the first match, the machine-learning algorithm defeated Dendi in ten minutes. Dendi then resigned from the second match, and declined a third. On the OpenAI blog, developers boasted that the bot had previously conquered the top 1v1 player in the world and the top overall player in the world.
Musk’s billion-dollar OpenAI venture has a noble goal — nothing less that saving humanity from the impending apocalypse unleashed by our AI overlords. On a far less grander scale, the OpenAI algorithm for Dota 2 was developed by playing many games against itself, also known as “learned bot behavior,” and then utilizing techniques that could take human players years to master. In a new video, OpenAI detailed some of the rather esoteric strategies used in its demonstration matches, such as last hitting (scoring extra gold by dealing the last blow) and raze dodging (using spell-casting lag to their advantage).
One-on-one matches are far less involved than the standard five-on-five bouts in tournament play, which feature a much wider range of techniques and strategies. Still, it’s an impressive accomplishment, and OpenAI plans to have its bots ready for full five-on-five matches at next year’s Invitational.
Greg Brockman from OpenAI, in a video released before the match, remarked that “Dota is a great test for artificial intelligence,” due to the game’s complexity and open-ended style of play. “Our bot is trained entirely through self-play. It starts out completely random with no knowledge of the world.” The bot then plays against itself for thousands of matches, developing strategies and gaining insight as it goes.
In an interview with Business Insider, Brockman expressed his hope that their “self-playing” style of machine learning will lead to far greater advances in AI. “At OpenAI, we’re not just about publishing a paper,” he said. “It’s really about building systems and doing something that would have been impossible before.”
