OpenAI Five has a near perfect record against humans in Dota. But the success of AI in gaming hasn’t brought dismay to its community, as it did when machines defeated the best players in Go and Chess.
Last week, OpenAI Five crushed the TI8 champions, OG. In the post-game comments, captain NoTail predicted that his team could adapt and take at least one game out of five. There was no encore performance at the event, but now in a limited release, called OpenAI Five Arena, the rest of the public has had their shot at Open AI, with as many attempts as humanly possible. So far, OpenAI Five bots has a 99.3% win record against humans, but Dota is far from a solved game.
That’s for the competitive mode, where you can recruit your own team of human members to compete against five bots. For a more serene experience, in cooperative mode, you can play with bots, with or without other humans.
When playing against the bots, humans have described it as a futile effort. They machines micro perfectly. They understand the pixel perfect ranges of spells and attacks, the microsecond differences between cast and attack animations, and they can dance around these parameters with uncanny precision.
Crystal Maiden’s AI thoroughly outplayed Topson, one of the game’s top mid players, by manipulating the parameters of Shadow Amulet, spellcasting, highground vision, night vision, fog of war, and Sniper's cast and attack animations. Throughout game 1 against OG, the AI was also able to micro their heroes to turn between Riki’s Blink Strike animation, effectively negating the Backstab damage.
It seems reasonable that the community has retreated to mechanical skill as a way to rationalize our inferiority. We can understand being slower than a machine. Or else we would have to admit that these machines are also besting us in Dota with traits that are innately human: insight, intuition, and creativity.
Humans have had some success at abusing our unpredictably, but this could easily be due to a lapse in its programmers rather than the ability of machine learning to adapt. Bots, so far, have been poor at detecting invisibility, keeping their courier alive, and chasing a hero through the trees. They can be preemptive with their ultimates and BKB activations, leaving an exploitable window down the line.
And yet, the bots still win, almost all of the time. Our ability to make snap decisions without thinking, from intuition and our gut, is still based on our experience. Highly skilled players innately understand good and poor matchups, the likely locations of heroes off the map, or whether a fight is good to take or not—without having to rationalize the chain of events.
OpenAI has learned from their mistakes as well, but they have 45,000 years on us, and they remember every moment of it. The result of a teamfight is determined before it even begins. They’ve weighed risk and reward, and they understand the result of a teamfight before it even begins. If they initiate on you, the bots understand the odds are in their favor, once they unload all their spells, in proper succession. Can your team do the same?
““All of this has happened before, and it will all happen again.””
Like a perfect blackjack player, they know they odds, and they make the most optimal decision, all the time. And each decision, from what fights to take to what lane to pressure, forges a path towards winning. It’s easy to surrender that bots are just mechanically better in inhuman ways, but through Dota it seems that they might also be better at making human-like decisions.
Playing with OpenAI, as the only human in cooperative mode, is like playing with a stack of your better skilled friends. They have exponentially more experience than you, and their decisions are infallible. If your team is failing, there's only yourself to blame. So why make any decisions at all? It's better to just surrender to the will of the team, move along with the machine, feeling as though it might be better off without us in the end.