Over five billion matches have been played in the video game Dota 2, which stands for “Defense of the Ancients.” Described as a mixture of chess and basketball, it’s a very competitive, fast-paced game of strategy played by two five-person teams. Each team defends their ancient base while trying to destroy the enemy’s. Nearly 120 different hero avatars possess unique abilities that can change and level up. There’s imperfect information, deception, and unknown assumptions and dependencies. Players must communicate and coordinate with their teammates and react fast. Every decision—no matter how small—has an impact. It is incredibly complex, extremely hard to learn, and even harder to master.
Humans have been programming computers to defeat other humans at games for eighty years. So, naturally, someone set out to create an artificial intelligence (AI) bot that could defeat the Dota world champions.
Based in San Francisco, OpenAI is an artificial intelligence research laboratory with a mission to develop powerful machines that help humanity. They chose to experiment with Dota 2 because of its complexity and nearly limitless potential options. To win the game, they knew they would have to create and train an AI to behave with human-like intuition, and that would further their research exponentially.
On March 18, CHM hosted OpenAI Technical Lead David Farhi and former Research Engineer Susan Zhang, now an AI researcher at Meta, in a discussion with CHM VP of Innovation Marguerite Gong Hancock. First, there was a film screening of Artificial Gamer, a documentary about OpenAI’s quest to beat a human team in Dota 2.
Watch the film below to follow the team’s daily ups and downs as they create algorithms for the AI to teach itself with reinforcement learning and scale up compute power to support it with 300,000 CPUs. Find out who wins when they take on the reigning champions in front of 17,000 spectators at The International, one of the biggest esports tournaments in the world.
In the panel discussion following the film, Susan Zhang said she remembered things being much more chaotic than how it was portrayed in the documentary. She recalls finding a bug on the day of the Finals and trying to fix it up to the last second. David Farhi noted that a lot of effort went into the systems-stability side of things—they worried about the AI not crashing, let alone winning the game. He said every day was full of little surprises. “We were bad at predicting what would be easy and what would be hard. It’s just surprising that it worked,” he said, laughing.
Susan remembered a white board that outlined all the code they were trying to merge back together. They called it “the merge of doom.” Nine out of ten times, things didn’t work. It’s difficult to come up with explanations out of opaque results, to find some method in the madness to improve the system. Regarding AI research, she says, “We’re still just confused.” It’s easy to fall into the trap of thinking we’re further along than we are.”
Clearly still shaped by their intense experience with Dota, the AI researchers focused more on what they didn’t know than their milestone achievement. But they are clear in their belief that as these powerful systems develop, it is critical to ensure that the benefits are widely distributed and truly meet OpenAI’s vision to serve humanity.
Blogs like these would not be possible without the generous support of people like you who care deeply about decoding technology. Please consider making a donation.