And they did it by adapting and understanding how OpenAI plays
ainodehna stack abused these:
- invis (riki, shadow blades)
- split push, avoiding fights
- playing conservatively until late game
- abused roshan
- focused on buildings with BKB
and of course, good skill and farming
Learning from human amounts of data would be a big breakthrough. Hard to say what that breakthrough would require since no one's done it.
If we just wanted to make Five stronger though, there's an incremental path:
We trained a single AI for the past 10 months, something we haven't seen before in reinforcement learning.
OpenAI Five at The International was 1.5 months old.
OpenAI Five at Finals was 10 months old.
*Huge* difference in performance. And the curves still haven't leveled off.
After 2000 games, the first 2-win steak by humans! Five’s win record is 99.5%.
(Incidentlalt, it had a 99.9% win rate versus the Five version that played at The International — puts that number into perspective!)
“The bots are locked,” the player said, as transcribed on Reddit. “They are not learning, but we humans are. We will win.”
One of the reasons that an AI system which can learn from human amounts of data (vs 45k years of experience) will be a big deal!
6 human wins so far, out of 1288 games. No one’s yet beaten Five twice in a row!
Over 1000 concurrent games running, with more than 2,500 concurrent human players. Pretty cool to see Five scale to play all those games at once.
OpenAI Five wins the first 500 games versus humans. But a few groups are systematically searching for every weakness — and making real progress. There is hope yet that Five can be beaten!