Elon Musk’s OpenAI published an update on Monday about its ongoing quest to build unbeatable eSports teams powered solely by A.I. By August, the venture plans to field a team of five neural networks that will be ready for global competition.
Almost a year after crushing some of Dota 2’s best players in one-on-one matchups, OpenAI is back to master team contests. In the blog post published Monday, the research company said its A.I. plays about 180 years worth of games against itself every day in preparation for The International tournament in August.
While their one-on-one demonstration was more than impressive, there is still some early evidence that OpenAI might be biting off more than it can chew by promising a formidable five member squad. In fact, we might go so far as to say it wouldn’t stand a chance against a coordinated team of pros in a tournament setting. And that’s because it’s been training on easy mode this whole time.
Understanding Dota 2
Dota 2 is a multiplayer online battle arena, or MOBA game, where two teams of five players duke it out for control over a map. Each player selects from a roster of over 100 unique characters — known as heroes — to put together a deadly squad. The end goal: amass more power than your opponent and mount a devastating attack on their home base.
Team synergy and character composition is vital. Each hero fills a certain type of role — think of this like defenders, midfielders, and attackers in soccer. Without a delicate balance, defeat is guaranteed.
“This is the game that to me feels closest to the real world and complex decision making (combining strategy, tactics, coordinating, and real-time action) of any game [A.I.] had made real progress against so far,” wrote co-chairman of OpenAI Sam Altman in his own blog post.
But while OpenAI has indeed been able to compete against some middle-to-high caliber teams, that’s only because it’s dumbed down the rules.
Human beings who square off against OpenAI might feel like they’re playing with one hand tied behind their back. To start, human players are also forced to pick the exact same heroes as their AI opponents — a set up known as a mirror match which prevents people from being able to create a complementary team.
But the biggest reason to question OpenAI’s readiness for the big league it’s its proscription on warding, Roshan, and other various items that are often used by pros in-game.
In Dota 2, players can only see parts of the map they are near or control, everything else is cloaked in darkness by the Fog of War. Wards, for example, provide vision of these areas and are used in basic strategies to spy on and anticipate the moves of an opposing team.
Roshan is a powerful monster at the center of the map that takes a coordinated team effort to attack. Killing this monolithic beast puts the team that pulls it off at a massive advantage. When Roshan is killed it drops an item that instantly revives the hero carrying it, giving one team six players for a moment instead of five.
Both of these aspects are core to even average level games. Pro Dota 2 players make constant use of wards and take on Roshan as soon as the window of opportunity presents itself. In order to tout its dominance, OpenAI would need to incorporate these restrictions. Otherwise, it’s basically training its A.I. to master the Dota equivalent of Tee-ball.
In other words, as impressive as 180 years worth of gameplay each day may sound, OpenAI will never get good enough to beat conniving, strategizing humans until it learns how to take advantage of these key strategic elements to the game. Until then, score one for the meatbags.