Image by OpenAI
“We predict the probability of winning to be 95%” echoes the OpenAI bots in the opening minutes of the match. This weekend in an exhibition match, the OpenAI Five team dominated five pro level players in an exhibition match.
Blitz, former caster and pro player, said that the match felt like no opponent he had played for. It felt like playing a strong team, where each small error compounded over and over. This opinion was a testament to how far the OpenAI project has come since it’s one on one tournament at TI last year.
The comparison of OpenAI to a really good opponent was enlightening. More than being an oppressive opponent with mechanical advantages, removed from limits like the rate of human action and thinking, OpenAI now feels more and more like a thinking being. They react in realtime. They seamlessly intermesh their spells, weave between farming and pushing strategies, and group at the most opportune times to push.
Watch OpenAI Five Benchmark from OpenAI on www.twitch.tv
Playing against five audience members, the OpenAI Five decimated their opponents outer towers and raxes in about thirteen minutes. In their opening 2-0 against Team Human, the games averaged 23.5 minutes, but the outcome was determined long before.
Even Team Human’s small victories—a first blood here, an early Shadowblade on Shadow Fiend—felt like it was little recourse. It felt fatalistic, as if OpenAI let them have those crumbs, because in the end it didn’t matter. The win probability would nonetheless rise, rise from 90%, to 95%, to 99%.
To be fair, the game wasn’t without its restrictions. The pool was limited to 18 heroes, expanded from 5 in the initial announcement. Each team had 5, indestructible couriers—just a layover from the OpenAI 1v1 code. And the OpenAI bots had an artificial 200ms delay, which was partly an artifact to allow the team to scale their project, but it also happens to be close to a human’s reaction time, even if it seemed as though Fogged’s Earthshaker was instantly hexed on his blink initiation.
The bots are learning from playing 180 years of matches, per day. They recalibrate on the spot, dissecting which actions lead to the most advantageous path to victory. One audience member asked if the OpenAI bots would learn how to drop certain items while consuming regeneration items to maximize health or mana recovery–one of the numerous tricks that high level players do in order to extract every advantage they could. The team simply deferred, saying that yes, if the OpenAI bot felt like it increased their chances to win.
What players might learn from watching the bots, rather, is how good a team can be when it’s deliberate in its decision making. To borrow from basketball’s parlance, it’s to play with “thrust.” Make quick decisions, and then execute. It’s possible to push towers without pushing heroes, if you just get your team to attack the tower together. It’s possible to dominate a lane with three supports, as OpenAI did when they trilaned with Lion, Crystal Maiden, and Lich. Perhaps one of OpenAI’s greatest strengths is they haven’t yet learned how to flame each other.
The OpenAI bots didn’t even last hit that well, but they were near perfect in their spellcasting. They never overlapped stuns and they always found a way to nail AoE spells at their maximum possible ranges. They maximized their mana pool every fight, even spamming Assassinates and nukes before a team fight broke out. Capitalist’s Crystal Maiden couldn’t even stand in the lane, even far behind his own tower.
And then there’s the importance of the draft. In game 1, OpenAI calculated a 95% win probability solely from the draft, 76.2% for game 2, and in game 3, the OpenAI devs, in a spell of overconfidence, allowed Twitch chat to draft their final lineup. The crowd cobbled together a team that the OpenAI Five predicted would give them a 2.9% chance of winning.
The next step for the OpenAI team is removing some of their imposed restrictions, smoothing out some quirks like warding and item buying, and finally playing a pro team at the upcoming TI. The question may not be how relevant Dota can be if computers can best humans at it, in real time. Chess and Go have been able to thrive despite Kasparaov and Lee Sedol falling victim to Deep Blue and Deepmind, respectively. Already, we’re seeing glimpses of the project’s real world applications. The foundation of OpenAI Dota helped the team program a robotic hand to mimic the dexterity of a human hand. Or to be more accurate, the team trained OpenAI to do so. What more can this machine learn?
