Where did it all go wrong for us humans?
We thought we were so smart. We invented fire, the wheel, ice cream, those little googly eyes you stick on stuff to make it look like a face. But then we went and created AI and now we can’t even enjoy video games properly anymore.
In our last article, we covered the evolution of OpenAI. Now we take things up a level and see how OpenAI Five is literally beating humans at our own game.
Follow us as we take a look the landmark Dota 2 series featuring OpenAI Five vs OG. And remember, when the robots eventually do arrive, they don’t like stairs.
OpenAI has come a long way since beating Danil “Dendi” Ishutin at The International 17. Where previously they were limited to 1v1 mid-only games as Shadowfiend, now they can play a much-expanded roster. At the time of writing, OpenAI picks from a pool of 17 heroes. It was previously 18, but due to extensive changes to the way Lich plays in patch 7.20, the development team cut him from the pool.
OpenAI Five has won big games against caster teams featuring William ‘Blitz’ Lee, Kevin ‘Purge’ Godec and Jake ‘SirActionSlacks’ Kanner. Great results for OpenAI Five, but what would happen when they took on the pros?
The game against OG was not the first game OpenAI Five had played against professional Dota 2 teams. Before taking on the TI18 winners, OpenAI played Alliance, another TI winning team, in a best of 3 series. In a private series played behind closed doors, OpenAI Five would go on to crush Alliance 2–0.
But the OpenAI Five vs OG game was different. The OpenAI Five vs OG game was live.
Among the rules in place for the OpenAI Five vs OG game is the ‘No illusions or controllable units’ rule. This is particularly interesting, mainly due to the reason this rule is in place.
In a panel discussion with the OpenAI Five developers before the game, the devs explained that illusions and controlled units would give OpenAI an unfair advantage. This is because it can control these units simultaneously with zero inefficiencies. By contrast, a human player, no matter how efficient at micro-management, will always experience a delay.
Rather than being something too complicated for OpenAI to master, the ‘No illusions’ rule is there to stop the AI from having an unfair advantage.
The series would be a best of three, with the aforementioned 17 hero roster to choose from. In previous games, OpenAI Five had been used to playing with five invincible couriers. This was to stop the human players from exploiting the AI by attacking something not directly controlled by it. For the game against OG, however, it would be a single, vulnerable courier just like a regular game of Dota 2.
Finally, the game would be played using the current Dota 2 patch, bringing the heroes and abilities right up to date.
Before game 1 had even begun it seemed to favor OpenAI Five. OG’s first pick in game 1, an Earthshaker, immediately set the win percentage to 64.3% in OpenAI Five’s favor. By the end of the drafting phase that percentage had climbed to 67.6%. This remained a running theme throughout the drafting phases of the series.
Brooke Chan, an OpenAI developer on the host panel gave some interesting insight into OpenAI Five’s drafting mechanics. When asked by Jorien ‘Sheever’ van der Heijden why OpenAI waited between hero picks, as if it’s thinking about what to pick, Brooke’s reply was excellent. Laughing before she even answers the question, Brooke tells the panel that the pause is entirely artificial and is there so as not to rush the human players. OpenAI Five knows what it wants to pick instantly, it just artificially adds time to help the human players out.
In practice games the day prior to the OpenAI Five vs OG match, OpenAI had begun experimenting with a tri-lane setup. For the uninitiated, a tri-lane typically involves three heroes on your safe lane to secure farm and early kills for your carry. This is a setup they will only employ if the draft setup favors such a tactic.
For this series, however, OpenAI opts for a standard 2-1-2 setup. And with that, the games begin.
The Games – OpenAI Five vs OG
Spoiler alert, OpenAI Five beat OG 2–0 in the series.
We could go through a lengthy play-by-play here, describing how the OpenAI Crystal Maiden secured a 3 – 1 trade with a beautiful ultimate, or how OG taunted the AI and then got immediately punished.
We’re not going to do that.
Instead, we’re going to cover some of the things that set OpenAI Five apart from the human teams.
We alluded to it in our previous article, but OpenAI Five operates with absolute numerical precision. The AI Crystal Maiden securing a 3 – 1 trade is a great example. Despite being on sub-10% health and surrounded, the Crystal Maiden still deems it viable to cast her ultimate, Crystal Nova. This decision, combined with the arrival of the OpenAI Five team into the fight (something the Crystal Maiden obviously knew was about to happen) secured a big net worth swing in OpenAI’s favor.
The numerical precision also applies to win percentages. OpenAI seems to know something the human players don’t. OG was in a dominant position at the 18-minute mark. A kill against the AI Sven and the destruction of the tier 1 top tower prompts OG players to ask for a new win percentage, to the delight of the casting team.
In less than a minute, after turning a team fight around in their favor, OpenAI oblige. They state the win percentage to be above 95%, to the confusion of everyone involved. The game up to that point was incredibly even, but OpenAI seems to know something no one else does. Less than two minutes later, OpenAI destroyed OG’s tier 2 and 3 towers and both barracks on the mid lane.
Net Worth Distribution
When the game ends at the 38-minute mark, the difference between the top and bottom net worth heroes on OpenAI’s team is just over 7,000. By contrast, the difference in OG’s team is over 18,000.
Some of this may be down to a more successful slew of team fights for OpenAI Five. This will result in the support heroes obtaining more gold. However, a large portion of it is down to the general playstyle.
OpenAI will always take a kill if one is available, punishing even the most minor misstep. The decision on who gets that kill seems to come down to a question of time.
An OpenAI support hero will not pull their punches to allow a core or carry to claim the lions share of the bounty. Instead, they will nuke the enemy down in the most time efficient way possible and let the kill go to whichever hero gets the last hit. If you’ve ever played a pub game of Dota 2 you know this isn’t a popular tactic with other players, but OpenAI makes it work and makes it work well.
Unlike human players, an AI system cannot get tilted. This doesn’t stop them from throwing out the occasional bit of trash talk throughout the game, however. At certain, seemingly random intervals, a threshold would be passed in the AI’s algorithm that would prompt them to share the probability of winning in all chat. This started off at over 80% after just a couple of minutes and continued to rise up to over 99%.
We’ve all been in that situation where certain expletives on the other team are sharing their super insightful thoughts in all chat. It tends to get frustrated if not muted and it can actually alter your play style. You take the time to type back, making you less efficient, or you prioritize the trash talking player in team fights over potentially more valuable targets.
OpenAI does not suffer from this problem. They never change their playstyle based on what you say, they never do anything sub-optimal, they never, ever get tilted and they absolutely will not stop until you are dead.
OpenAI Five vs The World
If OpenAI Five can beat the best team in the world, what chance do the rest of us have? Well, OpenAI decided to give us the opportunity to find out. Between April 18th–21st 2019 OpenAI was offered to the Internet. Players could play as either a competitor or teammate.
OpenAI Five took part in 7,257 competitive games, and the results were pretty alarming. By the time the three-day event was over, OpenAI boasted a huge 99.4% win rate. This means that human players only won 42 games against it. Open AI won the other 7,215.
Only 4 teams managed to beat OpenAI more than one. Only 2 teams managed to beat it more than twice.
What Does This All Mean?
The assumption from some of the panel (specifically William ‘Blitz’ Lee) prior to the OpenAI Five vs OG game was that OpenAI would be able to play at a maximum of 4k MMR (nice one, Blitz). OpenAI proved the panel and any doubters wrong.
If OpenAI Five can surprise the casting team, and surpass even top-tier professionals at the game of Dota 2, what implications could this hold for AI in the future? We’re not AI specialists here at The Game Haus. We’re just sports and esports enthusiasts who love a good story.
With that said, it’s fair to say that this achievement marks a huge lead forwards in the capabilities and the perception of artificial intelligence in general. Even high-profile industry leaders like Bill Gates are taking notice. Gates tweeted the following:
“AI bots just beat humans at the video game Dota 2. That’s a big deal, because their victory required teamwork and collaboration – a huge milestone in advancing artificial intelligence.”
As much as we’ve joked about the Terminator during this series, and as potentially dangerous as it could be in the wrong hands, AI has almost unlimited potential as a tool for human advancement. Its capacity for learning and evolving is unparalleled and it has already shown itself to be capable of seamless reactions and adaptations.
Where AI goes from here is anyone’s guess, but we’re definitely excited to follow it’s progress and even more excited that OpenAI has chosen Dota as its proving ground.
What aspect of life would you like to see AI control or replace? Would you like to see an expanded roster of heroes for OpenAI Five to use? Let us know in the comments below.
Follow Matt on Twitter @MattyMead2006.
From Our Haus to Yours.