OpenAI Five: Dota 2
Five neural networks coordinating in real-time, executing long-term strategies in a continuous environment to defeat world champions.
Dota 2 represents a significant step up in complexity from board games. It is played in continuous space and time, features hidden information, and requires team coordination. In 2019, OpenAI Five defeated the world champions, Team OG.
The Scale of Training
OpenAI Five was trained entirely through self-play, accumulating 45,000 years of Dota gameplay over 10 months. It experienced 80% of the game's theoretical state space, allowing it to develop entirely new meta-strategies.
The most profound emergent behavior was selfless play. The agents learned to sacrifice themselves for the team objective, something humans struggle with due to ego.
| Metric | Human Baseline | OpenAI Five |
|---|---|---|
| Reaction Time | ~250ms | Constrained to ~250ms |
| Experience | 10-20,000 hours | 45,000 years |
| Strategy | Defined roles (Carry, Support) | Fluid, context-dependent roles |