OpenAI Five
OpenAI
OpenAI's five-bot team just started beating amateur human squads at Dota 2. No human coordination trick required—just neural nets that learned to team up.
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 dropped a quiet but telling update this week: OpenAI Five, its squad of five neural networks, has begun beating amateur human teams at Dota 2. Not bots stomping bots. Not a scripted demo. Actual matches against people who queue up and play the game for fun, and the machines are starting to win.
Dota 2 is a nasty testbed for AI, and that's exactly why OpenAI picked it. Five players per side, hidden information, a map the size of a small city, and matches that stretch past 45 minutes where a single bad decision at minute 3 can decide the outcome at minute 40. Unlike chess or Go, there's no clean board state to stare at. You've got fog of war, item builds, hero synergies, and constant tradeoffs between farming gold and fighting. Coordinating one agent through that mess is hard. Coordinating five that have to act like a team, without any human hand-holding on strategy, is a different animal entirely.
What's notable here isn't raw dominance — OpenAI isn't claiming these five networks are toppling professional squads yet, and amateur is a wide bucket that includes plenty of players who'd lose to a coordinated team of trained hamsters. But the fact that five separate networks can develop teamwork good enough to consistently beat organized human opponents, without being told how to cooperate, is the real headline. Cooperation like that usually gets hardcoded by developers or emerges after painstaking multi-agent training tricks. Here it seems to have fallen out of the same kind of large-scale reinforcement learning OpenAI has been running on other problems.
The obvious next question is how far this scales. Beating amateurs is one rung on a ladder that ends, presumably, with matches against professional Dota 2 teams — the kind of showcase OpenAI has clearly been building toward. Whether the current approach holds up against players who've spent years mastering drafts, itemization, and split-second team fights is the thing to watch. Games have been AI's proving ground since Deep Blue, and every time one falls, it tells you less about the game and more about what kind of coordination problems machines can now handle without a human in the loop.
My take — AI-written commentary, not fact-checked reporting
I like this one because it's not another single-agent flex — getting five networks to cooperate without explicit rules baked in is the actual research contribution, and it'll matter far more for real-world multi-agent problems than another Go milestone ever did. That said, I'll reserve the applause until they take on a real pro roster; 'amateur' is doing a lot of quiet work in that headline.
Read more about this at: OpenAI