The International 2018: Results
OpenAI
OpenAI's Dota 2 bot squad lost both its matches against pro players at The International in Vancouver. It held its own for the first half hour each time before falling apart.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI Five walked into Vancouver this week carrying a lot of hype and walked out with two losses. The bot team faced top-tier Dota 2 professionals at The International, the game's biggest tournament, and for the opening 20 to 35 minutes of both matches it looked competitive. Then things came apart.
That early-game strength is not nothing. Dota 2 is a brutally complex game, five heroes a side, hundreds of items and abilities, map control that shifts by the second. Getting a bot to hang with humans through the laning phase and into the mid-game is a real technical result, even if the scoreline doesn't show it. OpenAI has spent months training Five through self-play at a scale that dwarfs almost anything else in game AI research, and the fact that it built leads or stayed even against players who've spent years mastering this game says something about how far the approach has come.
But competitive Dota is decided in exactly the moments where Five broke down. Pro teams are relentless about punishing small mistakes, forcing fights the bot didn't expect, and adapting mid-game in ways that static strategies struggle to answer. Whatever gap opened up after the 20-to-35-minute mark, human players found it and widened it fast. Two games, two losses, and a pattern that repeated both times suggests this isn't bad luck. It's a specific weakness.
OpenAI has been open about treating this as a research project rather than a victory lap, and losing at The International doesn't erase the progress made getting a five-agent system to cooperate at this level in the first place. Still, the result draws a clear line between beating scripted or lower-tier opponents and beating the best humans alive at a game built around adaptation and deception. That line is exactly where the interesting research problems live now.
My take — AI-written commentary, not fact-checked reporting
I'll say what a lot of AI-hype threads won't: losing here is more informative than a scripted demo win would've been. Anyone claiming self-play alone will crack adaptive, adversarial, real-time strategy is getting a very public reality check, and that's healthy. The gap that opened at minute 20 is the actual research problem — everything before that was just a warm-up act.
Read more about this at: OpenAI