advertisement

AI learns from pro gamers - then crushes them

For decades, the world's smartest game-playing humans have been racking up losses to increasingly sophisticated forms of artificial intelligence.

The defeats began in the 1990s when Deep Blue conquered chess master Garry Kasparov. In May, Ke Jie - until then the world's best player of the ancient Chinese board game "Go" - was defeated by a Google computer program.

Now the AI supergamers have moved into the world of e-sports. Last week, an artificial intelligence bot created by the Elon Musk-backed start-up OpenAI defeated some of the world's most talented players of Dota 2, a fast-paced, highly complex, multiplayer online video game that draws fierce competition from all over the globe.

OpenAI unveiled its bot at an annual Dota 2 tournament where players walk away with millions in prize money. It was a pivotal moment in gaming and in AI research largely because of how the bot developed its skills and how long it took to refine them enough to defeat the world's most talented pros, according to Greg Brockman, co-founder and chief technology officer of OpenAI.

The somewhat frightening reality: It only took the bot two weeks to go from laughable novice to world-class competitor, a period in which Brockman said the bot gathered "lifetimes" of experience by playing itself.

During that period, players said, the bot went from behaving like a bot to behaving in a way that felt more alive.

Danylo "Dendi" Ishutin, one of the game's top players, was defeated twice by his AI competition, which felt "a little like human, but a little like something else," he said, according to the Verge.

Brockman agreed with that perspective: "You kind of see that this thing is super fast and no human can execute its moves as well, but it was also strategic, and it kind of knows what you're going to do," he said. "When you go off screen, for example, it would predict what you were going to do next. That's not something we expected."

Brockman said games are a great testing ground for AI because they offer a defined set of rules with "baked-in complexity" that allow developers to measure a bot's changing skill level. He said one of the major revelations of the Dota 2 bot's success was that it was achieved via "self-play" - a form of training in which the bot would continuously play against a copy of itself until it amassed more and more knowledge while improving incrementally.

For a game as complicated as Dota 2 - which incorporates more than 100 playable roles and thousands of moves - self-play proved more organic and comprehensive than having a human preprogram the bot's behavior.

"If you're a novice playing against someone who is awesome - playing tennis against Serena Williams, for example - you're going to be crushed, and you won't realize there are slightly better techniques or ways of doing something," Brockman said. "The magic happens when your opponent is exactly balanced with you so that if you ... explore and find a slightly better strategy it is then reflected in your performance in the game."

Tesla chief executive Elon Musk hailed the bot's achievement in historic fashion on Twitter before going on to once again highlight the risk posed by AI, which he said poses "vastly more risk than North Korea."

Musk unleashed a debate about the danger of AI last month when he tweeted that Facebook chief executive Mark Zuckerberg's understanding of the threat posed by AI "is limited."

Article Comments
Guidelines: Keep it civil and on topic; no profanity, vulgarity, slurs or personal attacks. People who harass others or joke about tragedies will be blocked. If a comment violates these standards or our terms of service, click the "flag" link in the lower-right corner of the comment box. To find our more, read our FAQ.