Face it, we’re doomed if Skynet ever happens.
Over the last weekend, OpenAI’s Arena challenge saw artificial intelligence-powered bots beating humans consistently over the course of 7,257 games, with humans only managing to win just 42 games, a mere 0.6%.
To be fair, these were the same bots that had crushed TI8 winners OG in a 2-0 best-of-three sweep with wins decided in less than 40 minutes. Of course, when OpenAI first made its debut Dota 2 1v1 match against Danil “Dendi” Ishutin back in 2017, everyone was pretty sure the writing was on the wall for humans.
The first human team to beat OpenAI’s system during Arena was Alpha Red, a professional Dota 2 team from Thailand, placing a 34-24 kill scoreboard over a 45 minute win. While there were allegations of cheating from the human players’ side, OpenAI’s system went on to maintain the 99% win rate throughout the weekend.
Only one team managed to achieve a 10-streak win, using a consistent strategy of split-pushing, where players would take one of the three lanes on the map and avoid team fights with the AI opponents. They would also cut creeps, preventing creeps from approaching their base, so that the AI would not be able to break through the backdoor protection (which makes buildings a lot tankier).
The 10-streak team also abused invisibility, which the AI had yet to properly figure out, to escape ganks and attack buildings before running away. The AI was also not reactive enough to understand that it had to protect its Ancient, allowing its human opponents, who bought items designed for building damage, to quickly destroy the Ancient once it was exposed.
It’s not all doom and gloom though, besides trouncing humans at Dota 2, OpenAI bots also partnered humans in a co-operative mode, playing a total of 34,466 games, which also shows the potential of human-AI partnerships.
Greg Brockman said in a tweet that OpenAI’s algorithmic advances did have one drawback: “scripted components are a clear weakness that won’t change with training,” but then it was also “[…] just a handicap. There are still *many* aspects of the game, like invisibility, which are learnable.”
Avid Dota 2 players also discussed ways to exploit these perceived weaknesses through Reddit threads and Discord servers, all dedicated to the efforts against OpenAI’s bot. The dev team behind OpenAI’s Dota 2 bot also opened a Reddit AMA to address questions from curious players.
After their livestream was done, OpenAI’s Jonas Schneider thanked everyone who participated and said that their team hopes to be back with more in the future.