The world’s best Dota 2 players just got destroyed by a killer AI from Elon Musk’s startup

Tonight during Valve’s yearly Dota 2 tournament, a surprise segment introduced what could be the best new player in the world -- a bot from Elon Musk-backed startup OpenAI. Engineers from the nonprofit say the bot learned enough to beat Dota 2 pros in just two weeks of real-time learning, though in that training period they say it amassed “lifetimes” of experience, likely using a neural network judging by the company’s prior efforts. Musk is hailing the achievement as the first time artificial intelligence has been able to beat pros in competitive e-sports.

While the demonstration was highly limited to a few variables of gameplay, it was still remarkable to witness crowd-favorite Dota 2 pro Danylo “Dendi” Ishutin get crushed in a live 1-vs-1 match with the bot. Some of the bot’s maneuvers looked eerily human. After being defeated by the bot twice, Dendi forfeited future matches with it, and expressed surprise that a bot could outplay a human. He said the bot “feels a little like [a] human, but a little like something else.”

Dota 2 is an astoundingly complex game in which two teams of 5 players compete to siege and destroy the opposing team’s base. The game features 113 playable heroes who each possess unique abilities, as well as dozens of items that can enhance and extend each hero’s capabilities — meaning the full extent of the game’s possibilities are virtually incomprehensible, at least to a player with human limitations.

Elon Musk founded OpenAI as a nonprofit venture to prevent AI from destroying the world — something Musk has been beating the drum about for years. Just last month he told a group of US governors that AI represents a “fundamental risk to the existence of civilization.” Others, like Facebook’s Mark Zuckerberg, have been less than impressed by Musk’s killer robot doomsaying.

At least for now, killer AI seems limited to parlor tricks at e-sports tournaments. The OpenAI team’s bigger plan is to extend its Dota 2 bot’s capabilities into something that can compete across the full game in a 5-vs-5 match by next year’s tournament.

Comments

Valve’s own Dota 2 bots are really impressive in their own right, given how many abilities and interactions they have to juggle. I always thought that was an underrated aspect of the game, and I marvelled at the mathematical and coding complexity that underlies such a setup.

But a bot that trains itself is just magic. We’re in the territory of magic now.

Well pay of the trick is you don’t know how it’s done ! The built in bots probably just followed some clever rules and that’s all.

I don’t believe is claim that this is way harder than go. I mean in this game you have clear metrics about the game state : you know your life points and stuff. In go without being really knowledgeable you see a board and can’t tell who is winning especially early in the game. This makes the problem more subtle.

I think that’s just Musk being Musk. He obviously has no data or research to back that up. And even if he did, Dota is like 5 years old. It’s a ridiculous statement to get retweets.

Musk is really out of control with his BS. He is really shameless about it and think people are stupid.

The math required for measuring relative advantages and optimal points are much much simpler with DOTA than with chess or go. For instance, in chess and go, the strategic locations could vary greatly from game to game. In DOTA, they stay mostly the same. Even if in DOTA, you have a seemingly infinite number of positions and routes. They matter little because what matters is the optimal routes and the value locations which stay practically the same for a given map. Just the fact that you have masters of those latter games studying the game for decades and still feel they can still learn from the game and improves proves the AI required to beat them were vastly superior to the one that played the DOTA game.

You are talking about shameless? I only see half-knowledge coming from you. Have you ever read a book about AI? Actually, what you are saying doesn’t even have to do with AI… You could make a Bot even without knowing anything about the "most used routes". This topic is route agnostic. This Bot doesnt get any input about routes, he learns from 0. You are talking about Musk saying nonesense and write such a comment.

This guy is a freaking Musk troller lol. Just look at his comments. The Bull calling others, BS.

Musk being a seasoned programmer (He sold his first software 500$ at 12 years old), an expert in AI and a life time gamer, he knows EXACTLY what he’s talking about.
He surely chose this game BECAUSE it is one of the hardest to master.

I think most AI experts would disagree with you that DOTA2 is less complex than Go

Because ?
I gave an argument about the value function. This seems like a scaled up version of the Atari paper maybe a few tricks.
And it takes much longer for a human to learn go than dota2. Dota 2 just has a bigger action space, but you can probably reduce it a lot with simple heuristics.

Well know if they publish a paper…

DOTA is not simpler than Go. You can start playing Go almost right away, you’ll just suck at it. But the same holds for DOTA, you will need to spend quite literally hundreds of hours to get halfway decent.

And pro players have to play literally tens of thousands of hours to get to the top.

And go and chess players spend decades playing the game.

I completely agree with you. Saying Dota 2 is harder than Go without hard data is not a very honest thing to say. For start, The amount of probability for Go is already astronomical. Most importantly Go is a very balanced game. Only player who goes first gain a small advantage that is balanced by giving the player who goes second a few points at the end, and there is virtually no exploit for Go. This make Go is very hard to master without AI coming up with its own sophisticated rules sets. On the other hand, Data 2 is an asymmetrically balanced game which mean it is impossible for human to balance it perfectly. This mean every characters has different potential strength level. They are only balanced to the level that are not too obviously broken to the human (Including the Developers). The AI can discover these imbalance and exploit them against human players. Secondly, Due to its turn base nature every move in Go is a carefully measured response from both players. This ensure both player are playing as optimally as they could. A top human Go professional often make single digit so call "mistakes" during a match. On the other hand, real time game like Data 2 even "professional" human player will constantly making less than optimal moves due to brief lose of concentration, fatigue, and other human conditions. This also tie with another point about how does AI control the game? Unless it is strap to a robotic arm with similar speed of human hand. It is gaining unfair advantage by directly input commands into the game. This not only mean it will have faster reaction and respond time, but also it will have higher APM (Actions per minutes). If they limit AI’s APM comparable to the top pro player it will still have an advantage since human can’t always perform at their peak performance for long duration. Finally, all of my previous points only address why its is very easy for an AI to reach the level to beat a human player compare to Go. My final point will address why game like Dota 2 might in many ways easier to learn computationally. Unlike Go, game like Data 2 also don’t actually require full awareness of the entire game (board). Parts of it due to impossibility (fog of war), so instead the AI will focus on creating a good algorithm for dealing with its immediately nearby area (like when to move forward, retreat, go around, etc), and have a secondary system to give it a brief awareness of the whole situation. Like dEcmir already mentioned those stats are very easily to extract and interpret form the game. This mean unlike Go which will require AI to pay attention to everything on the board, and constantly reevaluate and adjust according to the situation. The AI for Data 2 probably already decided a predefined build when it pick it characters due to the games inherent imbalanced, and only adjust it slightly (get item A instead B) when something out of ordinary happened once in a long while. To be honest this seams to me more like a beef up version of the traditional game bot than anything else. Also those match are all 1v1, so the bot don’t even have to worry about high level teamwork and such.

To slightly recap what I said, and give more thought. (Also please ignore that I type Data instead of Dota in my previous post damm auto-correct.) Dota 2 allow much higher margin of error for the human player. AI can gain an upper hand by simply making less errors than human or perform better than human in many ways (e.g. APM). Second, Dota 2 seems to have a lot bigger game space. In an quick glazed it might make people believed it create more possibilities for AI to chew though thus making it more computationally difficult; however, this actually push AI to optimized a far smaller game space around itself as it is the most rational and efficient thing to do (This is a very popular strategy in many other attempts to make AI play computer games even for games that allow players to clearly see the entire map). As long as it can perform better than human in this local area when human appears on the scene it will gain a massive upper hand in an 1v1 situation. In a 1v1 game there is no flanking and team work. The other components such as schedule when to go to which lanes or when to go to jungle are less important as long as the decision it made are reasonable enough since not only it doesn’t matter so much in a 1v1 game, but also any non-optimal moves are drown out in a game like Dota 2 with huge margin of error for players.

WallOfText.

You can not play DOTA against halfway decent players by just doing local fights, it requires strategic thinking (where you apply pressure, how do you estimate the position of the enemy, whether to try for Roshan, etc) just as any board game. Pro players can easily win against perfect-playing local bots that have superhuman reaction speeds. But AI can definitely be limited to human-level reaction speeds, just to make it more interesting.

Of course it requires strategic thinking but this strategic thinking is 1) relative shallow and short-term and 2) limited in scope (e.g. the strategically important locations where to apply pressure is very small), given the real-time nature of the game.

Excellent points!!! Elon Musk should read this but I suspect he already knows this — he just has this habit to BS and hype things up for the crowd.

Any sufficiently advanced tech is indistinguishable from magic…

…but I’m not calling AI ‘magic’ until it’s grown smart enough to solve a few more complex systemic problems with the help of nanotech agents: poverty, misery, disease, death, energy, a sustainable earth vs nonstop growth, and helping man transcend our limited biology before we kill ourselves with increasingly powerful tech. In our lifetimes.

You forgot war and corporate greed.

funny how we think that the requirements for a bot to beat us in Dota or Go are a parlour trick compared to a bot killing a human or winning a live war.

what is the difference? data? that we have not read an article about a neural net being fed all the known data on historical military conflict?

humans interacting with landscapes in conflict are simply visual data, so the greyish square filled landscape makes the human dots move differently to the greenish noisy landscape; more dots, or bigger dots, or dots moving in certain patterns, dots sneaking up on dots. assign some variables to the dots.

all human conflict – a group of dots with variables moving across a picture and reducing other dots.

the current neural nets can already out manoeuvre us, we just have not given them the opp.

The only winning move is not to play. How about a nice game of chess?

Is it sad that I feel like future generations will hold that movie up as a warning that wasn’t taken seriously until it was too late?

all the known data on historical military conflict

History is written by the winners, and all is fair in love and war, i.e. NO RULES to follow, unlike in a closed-system such as a video game.

Thank God.

it is not about rules

it is about patterns

there are rules in war, tanks cannot cross water, men without food die. even deception offers rules, ammo transported as medical supplies gets a free pass.

we get tied up in the rules.

but on a more real level how these rules express is as visual patterns. it is the pattern that matters, not the rule.

if you feed the visual patterns into a neural net with one rule, be the last blue dot standing…

the system will determine its own rules of engagement. read the articles on how neural nets played go or better yet played poker.

i just need to say, i am not alarmist or worried about squads of AI robot death squads hunting down humans.

rather what is going to happen is that the next level of warfare will not be air superiority, but strategic superiority – whereby we use neural nets to direct troops in ways no human would.

sounds great, until you read how they play.

we follow rules, men die without food. the neural nets will follow patterns, 10,000 blue dots die from no food after they end 15,000 red dots. there will be some food left from the red dots, 10 blue dots live, win!

View All Comments
Back to top ↑