In the middle of game one between the human team OG (world champions at the game Dota 2) and the bots of OpenAI Five, an AI player sent a message to the in-game chat:

"We estimate the probability of winning to be above 95%."

The crowd laughed in surprise.

"From our human perspective, this is an even game, Owen," one announcer said.

His partner agreed.

"This game does not look one-sided."

OpenAI Five, though, went on to win twice in a row, becoming the first AI to beat world champions in esports.

OpenAI Five, a group of five bots developed by the company OpenAI, had done something significant-- unlike previous human-machine competitions, this one involved teamwork among AI bots, against a team of close-knit human players that worked together.

"We use Dota as a testbed for general-purpose AI systems," OpenAI wrote in a blog post, referring to Dota 2, the game played at the tournament.

Complicated, real-time strategy games, the blog post says, "Start to capture the messiness and continuous nature of the real world, such as teamwork, long time horizons, and hidden information.""

The AIs played an average of 180 years per day training under a model of machine learning known as Reinforcement Learning, which rewards positive outcomes. OpenAI Five observes every forth screen frame (the game runs at 30 frames per second, similar to a movie) as a list of 20,000 numbers, and sends back an 8-digit string in response.

The AIs themselves are totally self-taught, meaning that they play against versions of themselves.

"In the first games, the heroes walk aimlessly around the map," OpenAI wrote in a blog post. "After several hours of training, concepts such as laning, farming, or fighting over mid emerge. After several days, they consistently adopt basic human strategies."

"Once trained, these neural networks are creatures of pure instinct—their neural networks implement memory but do not otherwise learn further," according to OpenAI.