OpenAI Flaws and Deceptions

Artificial Intelligence has already conquered complex games like Go and chess, defeating the very best competitors in those games without batting an eye. But when the news came out that Elon Musk’s OpenAI company was building a bot capable of defeating the top Dota 2 players, the community was skeptical.

Dendi’s inability to deal with the Terminator Shadow Fiend bot at The International 2017 left people feeling like humanity was doomed. It was only a matter of time now. The end was nigh. The atmosphere reminded me of Garry Kasparov’s defeat against Deep Blue, in 1997.

This year, as they had already announced at The International 7, OpenAI are building an entire bot team capable of crushing the best Dota 2 rosters in the world. And given the speed at which they’re moving towards this goal, the chance of that happening at The International 2018 is almost 100%. So there you have it. Bold prediction: OpenAI vs. TI 8 champions (most likely Virtus.pro, PSG.LGD or Team Liquid): 3 – 0. But would this result be relevant or even fair?
Well, I think not. And here’s why.

A sleight of hand

When a magician performs a trick in front of you, they’re exploiting the weaknesses of human perception and employing a highly practiced set of moves to create the impression that they’re doing something impossible.

Of course, this is not The Prestige, so you’re obviously being tricked somehow. But, just as Michael Caine’s character says in the film, you’re not really looking for the secret, because you enjoy the state of awe so much that you just want to be fooled. In a very similar way, OpenAI has played its trick to perfection, mesmerising the Dota 2 community in 2017 and being well on its way to doing it again.

The TI 7 deception

Let’s remember the TI 7 moment more clearly. OpenAI had developed a very strong bot by having it play against itself until it had acquired “lifetimes of experience”. Then the bot was presented to the community and the results were impressive: the bot destroyed everyone it played against, including the best analysts and midlaners in the world. Then, during the Main Event of The International 2017, it showed Dendi how he should have played in the Regional Qualifiers in order for Na’Vi to have a chance of making it to the event.

There’s one problem though: the bot could only play Shadow Fiend, in a 1v1 mirror matchup. And Shadow Fiend is a hero that quickly snowballs out of control in the laning phase, due to the bonus damage it gets from Necromastery. So as soon as one of the Shadow Fiends gets a tiny advantage, it can use that advantage to build an even bigger advantage for itself. Unless a significant mistake is made by the side that has the initial advantage, the entire match is practically over in less than a minute.

Computers are far better than humans at raw computation, being capable of performing billions of operations per second. So the Shadow Fiend choice makes perfect sense, given that the hero has only 1 active skill in the laning phase (which he can use at 3 different ranges) and the matchup between two Shadow Fiends is decided through nothing more than simple decisions that the AI can calculate vastly better than a human. But try complicating matters, asking the bot to make much tougher decisions and having an Invoker vs. Invoker matchup. I’d really like to see how the OpenAI bot beats Miracle or someone like Sumiya in the midlane, even after having lifetimes of experience.

The upcoming TI 8 deception

For The International 2018, OpenAI has already prepared a team of 5 AI bots with the goal of beating the strongest rosters in the world. But, just like in 2017, there’s a catch. As lawyers and people in other professions use to say, he who controls the frame, controls the game. In other words, set up the most favorable conditions under which you’re likely to succeed and you’ll mostly likely emerge victorious. Choose the battleground and how the battle will be fought, play to your own strenghts and your opponent’s weaknesses, and victory is guaranteed.

The teams who will face OpenAI’s system at The International 2018 will not actually play a real game of Dota. They will play using 5 predetermined heroes against 5 predetermined heroes, in a mirror matchup. And just like in 2017, the 5 predetermined heroes have obviously been selected to minimize complex decision making and maximize what the AI is great at: raw computational power. I mean, just look at these 5 heroes and tell me this isn’t a joke: Viper, Lich, Crystal Maiden, Sniper and Necrophos.

Clearly, the context has been set so that the human advantage can be minimized. Complex decision making regarding drafts, item choices, hero playstyles and so on are reduced so much that raw computational power (basically, small mechanics and battle plan optimizations when every possibility is known ahead of time) is almost guaranteed to triumph.

It’s all a matter of who’s making better calculations for the most part, and, lo and behold, the AI will prove to be better at it than humans. Wow! What a surprise!

And as if this wasn’t enough, humanity’s team will likely play its first match ever using these 5 preselected heroes, while the OpenAI system will have already played that draft a million times before. The whole thing reminds me of that final scene from the Gladiator movie, when Commodus stabs Maximus in the back before the start of the fight. The good guy still wins, but I doubt our guys can do the same under these conditions.

I don’t know about you, but I seriously have a hard time giving the bots any credit in such a case. I mean, I was expecting to see some real, human-like intelligence being displayed by the Machine, and I’m not talking about the TI 7 host. He’s doing alright in that department.

Whether you agree or disagree (about my opinion as a whole, not about the last sentence), please leave a comment and let me know what you think.

Comments are closed.