Have you ever played the game breakout? You know, the video game where you bounce a ball into the blocks at the top of the screen to break them. Well, for beating that game like you did as a child Google paid $500 million. Yep, a company that built an AI to beat breakout is worth half a billion.
Okay…so this company, DeepMind, is not just about winning video games. It has applied its AI to strengthen Google in more practical ways, but the idea of playing games is central to AI development at Google and its peers. That raises an interesting question.
What happens when AI wins every game?
AI has beaten the best of us at Chess (1996), Jeopardy! (2011), Go (2016), and most recently Poker (2017). The exciting (or scary) thing is many experts did not think AI would defeat a Go champion for 10 more years. I repeat: people who have devoted their lives to advancing AI did not believe this could be accomplished for 10 years. That should give us pause when pundits question how quickly AI will change the world.
But certainly there are plenty of games AI has yet to touch. Well, it recently got a lot easier to play those games. Open AI, a non-profit focused on democratizing the use of Artificial Intelligence, released an open source platform in late 2016 for anyone to easily build and test AI in 1000+ video games hosted in the cloud.
And DeepMind, the creator of Go champion AlphaGo, has set its sights on winning at StarCraft II next. StarCraft is an incomplete information strategy game — balancing resource management, scouting, and battle tactics. This is significant, because incomplete information games resemble the messiness of the real world. This is in contrast with Chess and Go, which have rigidly defined structure, where both players have complete information about each player’s pieces and past actions.
At this point you might be wondering why companies are spending billions of dollars to be the best at playing video games? A valid question. As much as Twitch viewers may disagree, video game triumphs are not an end in themselves. But let us answer that question with another: why do human beings play games? Why do we want to win?
The answer, I conjecture, is one and the same.
A simple response is that people play games because they are fun. But if that were the only reason then I have wasted my time writing this much, so let’s try again. We play games for enjoyment, but also to learn. This learning can be as simple as how to socialize with peers or as complex as high-level strategy needed to outwit an opponent. Games provide a low-stakes environment for learning both physical and mental skills necessary to succeed in life — however you define success — at least that is what Wikipedia tells me. After the game is put away, we take these newly minted skills and apply them in our efforts to set and achieve our goals. The key part here is that games inherently have lower stakes. This allows us to try and fail without the fear of long-term failure. Lose and we just play another game. But fail in life, and we worry there may not be a second chance. So to prepare, we play.
So to prepare for real-world challenges AI, like us, chooses to play.
And play they will until all our games have been beaten. Soon enough our games will just be too easy, too simplistic for them. It may be hard to imagine, but as a next step AI will most likely create its own games to play so they can keep learning in a low-stakes environment. And in the same way a dog cannot comprehend Monopoly, we will not be able to understand these new AI games. For a very in-depth analysis of this analogy visit waitbutwhy.com.
But that is a digression from the opening question: what comes after games?
After AI figures out how to repeatedly achieve a success scenario in games like StarCraft it seems a fairly straightforward transition to the real world. Why would an AI that understands how to expertly balance resource consumption and military strategy not be able to adequately handle a company’s supply chain and logistics strategy? Both have data inputs with defined goals and limitations with incomplete information.
No, this is not a perfect copy and paste situation. People are involved in our reality so it will be clunky and halting at first as we figure out how to talk with AI. But given that learning curve, it is the logical progression.
In the post-playing era of AI, there will be little choice but to let AI into the world if we want them to keep learning. There simply is no other outcome. AI has so much potential power to increase all forms of human output that someone will choose to wield it. If this is done by an oppressive regime or corrupt business, immense suffering could follow. The only way to ensure that this transition has a net positive impact on humanity is to create competing AIs. Why? Because at the base of all our ‘kumbaya’ laws and philosophies, the only thing that backs them up is power. And the only thing that can check power is power. And for better or worse, advanced AI will be the most powerful thing man has ever invented. I leave you with this:
“Liberty consists in the division of power. Absolutism, in concentration of power.” — Lord Acton
So for the sake of liberty, let us hope that many more people start playing breakout.