AI Built By Elon Musk Deals Crushing Defeat To World's Best Players Of This Strategy Game

With multiplayer games, bot have always been the amateurs way to practice, with pros always preferring hone their skills against human players capable of strategy, deceit, and surprise.

As it turns out, artificial intelligence can do that too now, and better than humans.

OpenAI is an artificial intelligence research lab co-founded by Elon Musk. And for the past few years, they've been working on an AI that is capable of playing the popular MOBA Dota 2 at a professional level. This week, they achieved near to the highest level of ability with the AI.

At a special event this past weekend, OpenAI Five (the AI-controlled Dota 2 team) took on last years Valve 2018 International championship winners, OG, in a best of three exhibition match. Not only did OpenAI Five beat their human opponents, they positively trampled them.

The AI team used unconventionally aggressive tactics, including spending in-game currency to buy back early after a death, something well-versed players avoid as it slows down their long game. It worked however, with OpenAI Five beating back OG to take two wins.

While the AI took 30 minutes to win the first match, the second took only about half as long to clinch the second.

Of course, there are some limitations to its capabilities. Dota 2 has about 145 heroes that players can choose from. For this match though, each time was allowed to pick from 17 characters, the limit of how much OpenAI has trained so far. Summoning an illusion features were also disabled for the match, both of which require the AI to keep track of additional variables the system hasn't trained for.

However, it was able to optimally use the "Captain's Draft" mode, which lets a team strategically ban heroes the enemy might be proficient with, as well as allowing teams to counter pick each other. In fact, the AI has even mastered pro-level micro-strategies like denying and creep-blocking. However, this is apparently the last planned public demonstration of OpenAI's capability, at least likely until the AI is capable of playing with and versus all the available characters.

Talking about how Five plays the game. pic.twitter.com/OADOjPo72v

¡ª Greg Brockman (@gdb) April 13, 2019

To get to this point, the researchers have been using a deep learning reinforcement technique. They didn't code the AI with the ability to play the game, just the ability to learn from it. Then they dropped the algorithm into a match and had it learn from scratch. Through trial and error, and using a generative adversarial technique, OpenAI basically kept playing over and over in an accelerated virtual environment, and learning from its mistakes each time.

What that means is, in the 10 months it's been active, the AI has played Dota 2 for the equivalent time of more than 45,000 years .

Even though you won't see it beating pros on a stage in the near future, you can see for yourself just how good OpenAI is. The company has opened up registrations for the regular public to challenge the AI (both in co-op and versus modes) between April 19 at 6:30am to April 22 12:29pm. Don't expect to come away with a win though.

And this is just the beginning. Though this could mean a new age of games with incredibly devious NPC opponents, the learnings can also easily translate other robots and tasks.