The next step for OpenAI will be to pit its AI system against professional Dota 2 players at The International, the world's largest Dota tournament taking place 20 to 25 August in Vancouver. They are so well-coordinated and focused on teamwork that it was almost impossible for any of the human players to single a bot out.
In some ways the AI was expected to win given the five neural networks it uses clocked the equivalent of 180 years work of Dota 2 gaming every day - we know there are hardcore players out their clocking thousands upon thousands of hours of the game, but that pales in comparison to the amount of practice an AI can simulate. Nevertheless, it is still up for question if OpenAI Five can beat the best players in the game at The International later this month. It trains using a scaled-up version of Proximal Policy Optimization running on 256 GPUs and 128,000 CPU cores - a larger-scale version of the system we built to play the much-simpler solo variant of the game previous year. Using a separate LSTM for each hero and no human data, it learns recognizable strategies. OpenAI admitted however, that 1v1 matches have many less variables involved, making it easier for an AI to learn and master.
A few years ago, many researchers thought it was impossible for AI systems to get so good at complex team-based games like Dota 2, but here we are. It was, however, given five invulnerable couriers-though it was not allowed to scout nor tank creep waves with them. It predicted a 95 percent win probability after seeing the hero teams.
After losing to OpenAI Five, Capitalist Walsh expressed his despair in a tweet.
Update: And that's game, with the AI victorious in two rounds out of three. The second game saw the humans grab a tower from the OpenAI bots but ultimately the humans lost a second game.