Elon Musk's OpenAI Bot Defeats Top 'Dota 2' Player at The International

True, it was a test with a limited scope, but the results were still shockingly lopsided.

This is a companion discussion topic for the original entry at https://waypoint.vice.com/en_us/article/gyygb7/elon-musks-openai-bot-defeats-top-dota-2-player-at-the-international

the article points it out already but man it is extremely Marketing to say an ai defeated an esports player in a game “far more complex than chess and go” when 1v1 mid games are almost entirely about mechanical skill rather than decision making based on limited knowledge and many variables. like the reason ridiculous ais (this, toolasisted in fighting games) 1v1 games handily is due to perfect execution and not perfect reasoning. still, it’d be interedting to see a neural network ai try to play in a team context.


I’d be way more interested in bots that are programmed to effectively replicate human reaction speed, pattern recognition and multitasking ability. A bot that reacts with inhuman speed and perfection is a funny gag and might be an interesting engineering challenge but doesn’t really provide any value to a player. Losing against effective bots isn’t any more interesting than winning against ones that play poorly. Neither will capture the same thrill as playing against another human mind making decisions and trying to execute on them.

The 1v1 format also could be incredibly easy to abuse if an exploit is found. It’s be more impressive and harder to exploit if there were more bots on the larger scale.


Yep, although the actual implementation initially left rather a lot to be desired, the Drivatars in Forza are passable for providing something of a model (??? very technical implementation talk that counters this impression plz…) for how teaching AIs can go (and they did get refined with the option to disable overly aggressive moves from the range of behaviours). We can absolutely quantify preferred approaches to corners, reaction times (in many situations, providing a probability curve of responses), choices about gear shifting and a hundred other detailed measurements with which to try and create AIs that are somewhat like an actual player. See also some fighting games that try to model typical reaction times and move selection based on specific pro players or even the player of the game. Of course, it becomes rather more difficult to do that once you’ve got the complexity of something like a Dota (with free movement concerns and rather more varied intentionality vs a circuit racer or even a fighting game).

It sounds like this attempt was a lot more about using the reaction time advantage to make an AI that can win (marvelling at the neural net self-learning process that builds an imperfect black box) rather than making one that is a good copy of an actual realistic play style (from which you can tweak reaction times and quality of decision making in order to build an AI that either competes with the best players or offers a fair challenge to more typical players). There would presumably never be a point where it could be used to explore hero on hero strats or other such insights that a more authored approach would provide (from the “we just sim it playing itself” description of the training data used/generated - maybe that simplification hides that they are doing something extremely interesting and valuable for general video game AI that can emulate an online experience offline - would be really cool if a company called OpenAI had some detailed technical papers about their approach published).

honestly i would be really into a dota tournament that was just 5v5 neural network AI matches, especially stuff on the extreme low and extreme high end of performance

1 Like

My brother, some time ago when someone was showing off an AI with 20k APM or so, micro-managing zerglings to avoid siege-tank splash damage against a pro player said that watching two of these AIs would be awesome, and might be the future for esport.

It isn’t.

Even now, this sort of thing seems to be most interesting in that Paul Bunyan-esque struggle for relevance of human labor way.

You’re never cutting to the shot of the crying parent, who is so proud of their child. Gone would be the very human stories of a family being doubtful, discouraging of a pro because they didn’t think games were a thing. They can’t be a career. Nor that of the supportive family, or the kid who used games to escape, who now found themselves on the world stage, with thousands of screaming fans.

It continues to amaze me that people doubt what AI is capable of at this point. It’s… almost funny hearing people talk about a simplified game of dota, without team communication/planning, and that it’d take more than one AI to control all 5 units, manage economy, communicate instantly, if not just hyper effectively, if needed, and they aren’t just watching the map, with one-frame snapshots of every piece of vision the team has. Gross is a good word to describe how one-sided a full blown game of DotA against pros would look.

As a curiosity, sure. As… a replacement for players? Nah. Not yet. Give those AIs aspirations, shortcomings, flaws, drive, rivalries, we’ll talk again.

I get that the human element is a huge part of sports but I feel that still exists with AI fights and can still be very fun to watch even when we are not aware of them.

Take the Student StarCraft AI Tournament as an example. The human element has been shifted off of the players themselves and on to the coaches completely. Instead of it being a story about two people going at it it is instead a story about the people who built these bots and that can be a very exciting thing to read about or watch.


man all i said was i think it would be fun to watch a goofy learning AI come up with some nonsense hyper-optimized way to play the game by itself not that ti8 should be all hard bots

1 Like

The problem with the bot is that it’s forced into such a specific environment into which it can essentially perform frame-perfect movement into that it seems more impressive than it is. The skill build and itemisations were predetermined, the creep blocking was programmed into it, and Shadow Fiend’s Razes are skillshots that the bot can a) easily calculate, and b) cancel out of perfectly if the opponent moves out of range, since they do instant damage and can be cancelled at any point in the animation. The “machine learning” aspect is just the bot learning to do everything else perfectly, which as it turns out is just last-hitting which the DotA bots can do already.

Also you can beat the bot easily by just baiting the waves out of lane because apparently the bot never learned to attack the tower, just get CS. Your wave pushes in and attacks the tower and since the bot doesn’t have anything to “learn” how to defend against this it just loses.