We’ll kick off the OpenAI Five Benchmark with some warmup games against audience members; please let us know on the invite request⁠(opens in a new window) form if you’d like to volunteer.

Because our training system Rapid⁠ is very general, we were able to teach OpenAI Five many complex skills since June simply by integrating new features and randomizations. Many people pointed out that wards and Roshan were particularly important to include—and now we’ve done so. We’ve also increased the hero pool to 18 heroes. Many commenters thought these improvements would take⁠(opens in a new window) another⁠(opens in a new window) year⁠(opens in a new window). We’ll see how well these new game mechanics work on August 5th!

We will play with the following restrictions (the crossed out restrictions are those lifted since the original OpenAI Five blog post), which correspond to the last bits of the game we haven’t integrated:

We’ve increased the reaction time of OpenAI Five from 80ms to 200ms. This reaction time is much closer to human level, though we haven’t seen evidence of changes in gameplay as OpenAI Five’s strength comes more from teamwork and coordination than reflexes.