Very few times I have been more excited about a AI/ML thing/event/paper/whatever as I was about the recent OpenAI Five benchmark vs a team of Dota Casters/Pro Players/Ex Pro players.
I have played Dota 2 for about 3000 hours and even if I have never made it above 3k MMR I find the game fascinating, the complexity, the skill, the balance and patches, among so many factors make it an amazing game (I also stopped playing it since finding good games/teammates became almost impossible but thats another story). I am also involved into all this Data Science/Machine Learning/AI community so this event was quite special for me.
I still remember watching TI7 last year while making some dinner when they starting to talk about OpenAI, I thought my twitch stream changed or something but nope… They showed their version of OpenAI bot in a 1 vs 1 match and wining that quite convincingly against Dendi, who is the most famous pro-player in the Dota scene for sure.
OpenAI said they were going to try for a 5 vs 5 the next year. I was kinda of a hater on this since optimizing the 1 vs 1, while it was impressive, it’s orders of magnitude easier than the full 5 vs 5. They did it, with some restrictions but they did it.
As I watched the first games against the attendees to the event my immediate reaction was that the bots were playing as a great team: going from objective to objective, targeting the right players in the human team, using spells in a great way and also had a bit of inhuman reactions (imo). OpenAI Five had found (as good players have) that winning lanes and snowballing from there is a great way to win a game if you don’t make mistakes.
Honestly this is plain and basic old Dota, OpenAI Five optimized this to an amazing degree. They also found the best heroes for their 18 hero meta (full Dota is 100+) and as we know from pro Dota games, drafting is a huge part of the game.
From the first games against the attendees I also assumed that most of the games OpenAI played ended very early, 20-25 mins. I made a prediction that probably the way to win is going late game, as players know, Dota late game is a different game and not as studied as the early game and probably not as studied by OpenAI Five?
In game 1 and 2 the initial probabilities prediction of OpenAI Five were 95%+ which in my mind it was very optimistic, based on my knowledge of the game and the drafts I did not think the difference was that big, but it was and OpenAI won very convincingly.
My initial theory kinda got proven at the game 3 when OpenAI Five lost vs the Humans. While it was interesting to see them “desperate” to distract the players and kinda rat their way into the game the Humans ended up taking it with a very favorable draft and going a little bit late game. Exploiting the fact that the bots didn’t really knew about items (just buying from a static guide) and other of the restrictions thanks to Twitch chat the game was kinda easy for the Humans.
One of the things that OpenAI Five proved is that early game is the most important aspect of the game, and also that items kinda don’t even matter. They were able to get some easy wins playing good Dota and just buying tons salves :)
I would love to see OpenAI Five in a super late game fight and what they think they should optimize there. Also what do they think the best build for each hero are, this is fundamental part of the game and one that is being discussed so much all the time, having a “perfect” answer for this would be so interesting.
My reaction after the games is that after a couple of more attempts the Humans will for sure find a strat to beat OpenAI Five. They didn’t see any of their games before and 3 games is not enough but I bet that changing a little bit the draft and playing a little bit more safe in the first 5 minutes will go a long way. Also having more heroes give you more counters to the strat that OpenAI Five uses the most, so a little bit of an advantage for them, but it’s also a matter of time for OpenAI Five to remove all the remaining restrictions and be unbeatable for real.
One of the things that OpenAI has said before is that OpenAI Five and other ML algorithms is not how humans learn. Players have not played 180 years of Dota every year but they have understood the game to a very similar level. AI will get there for sure but in my mind this is an accomplishment of scaling more than ML. OpenAI proved (and said) that the same algorithms we have right now can go farther than we thought but we are not in what we would call real/human learning. More than impressive anyways.
Can’t wait for TI8 this and next week to see how the Pros do and next years version :)