Technological Advances are Becoming Scary Good
The beginning of the end is here, folks. Humans had a good run while it lasted. Google’s DeepMind AI team built a new AI, AlphaStar, specifically to beat StarCraft II professionals. And they succeeded. In a big way. After completing its training, AlphaStar beat Grzegorz “MaNa” Komincz 5-0 in their match. Halfway through its training, it beat MaNa’s teammate Dario “TLO” Wünsch 5-0 as well. Elsewhere in the world, Boston Dynamics is making robot dogs that can communicate (or at least share information) and solve problems to open doors, and bipedal robots that can “see” their environment and figure out how to navigate around it. I would like to take this time to welcome our new artificially intelligent robot overlords and state my unending loyalty.
I can’t tell you how excited I am that this movie exists
Let’s break these down and start with AlphaStar and StarCraft II. The original game was released back in 1998 and StarCraft II was released in 2010. That means people have been playing StarCraft II for almost a decade. StarCraft II is a real-time strategy game, where players move units, anything from harvesters to attacking units, to collect resources to build more units and buildings, and of course, attack each other. Matches are typically one versus one and consist of 5 games. Professional players have seen almost everything that can be done in the game, so that fact that AlphaStar won both its matches against human competition is impressive. Even more impressive, MaNa is one of the top 10 players in the world using Protoss, one of the three alien races that players choose from. So, AlphaStar, a computer intellegence, beat one of the very best players at a very challenging game.
Artificial intelligence is seen as one of the last frontiers for computer programming. We’ve developed AIs that can do many things, but they are usually most successful at tasks that have a “best” strategy. Chess is a good example of this, as there are a set number of pieces and accepted strategies to use against specific opponent moves. StarCraft II presents a much more fluid situation and a bigger challenge for AIs. A player can choose to build any number of units from a varied unit list. Additionally, changing map layouts and resource allocation present an increasingly complex environment for the AI to try and deal with.
The training method that they used to help AlphaStar deal with this complex game is seriously cool. They have a fantastic article you can check out that will get much more in depth than we will here, but let’s address some of the highlights. In the 14 days that AlphaStar was training, the system logged 200 years worth of gameplay for each iteration of its neural network. This is ridiculously impressive for two reasons. First, they beat two unbelievably good StarCraft II players in with just 7 days of training for each one. Second, they had roughly 800 iterations of neural networks and each one played 200 years worth of StarCraft II games against each other. They built their own league and played a silly amount of StarCraft.
The victory of AlphaStar over its human competition was broken down to see why it won. You might think, “Adam, it’s not fair! The computer can calculate decisions and react much faster than any human could,” and you would not be wrong. But in this case, that’s not what caused AlphaStar to win. Actions per minute, or APM, is the measurement for StarCraft II and AlphaStar averaged 277 APM, which seems pretty impressive until you stack it up against MaNa who averaged 390, and TLO who averaged 678! While AlphaStar is doing a lot of actions it’s not the reason why it won.
“But Adam, because of the reaction time, it’s doing those actions much faster, so they have more impact.” Sure, computer signals can travel faster than my eyes can see what’s happening and my brain can interpret that and tell my fingers to move units to a certain location, but reaction time wasn’t the reason AlphaStar won, either. AlphaStar’s reaction time, the time between its observation and performing an action averaged out to be 350 ms. When comparing this to human reaction time we can look at sports. According to NPR, it takes about 450 ms for a 95 mph fastball to reach home plate once it’s been thrown, and it takes about 150 ms for a batter to physically swing the bat to make contact with the ball. So, a batter has to decide if they’re going to swing in about 300 ms. AlphaStar’s reactions are fast, but they aren’t superhuman.
Which leads us to one simple and scary fact: AlphaStar won its games against some of the best human competition available because its strategy was better. Here’s what MaNa, again one of the 10 best Protoss players, had to say about AlphaStar: “I was impressed to see AlphaStar pull off advanced moves and different strategies across almost every game, using a very human style of gameplay I wouldn’t have expected.” So there’s that.
So that covers the mind of our new overlords, what about their bodies? Don’t worry, Boston Dynamics has that covered. Atlas is their latest creation and here is what they had to say about it: “Atlas does parkour. The control software uses the whole body including legs, arms and torso, to marshal the energy and strength for jumping over the log and leaping up the steps without breaking its pace. (Step height 40 cm.) Atlas uses computer vision to locate itself with respect to visible markers on the approach to hit the terrain accurately.” Yeah, they’re building robots that can scan the area around them and then calculate how to build and conserve momentum to navigate obstacles, literally, on the run.
Cool cool cool cool cool, tight tight tight tight tight. Thanks, Boston Dynamics. Now, not only are robots going to be smarter than me, they’ll also have better coordination. The Atlas not only jumps vertically, it also has to adjust its movement laterally to get up the three jumps. That’s very impressive. My only hope to avoid subjugation by robots in the future was being able to run away or lead them into obstacles. I guess I have to come up with something else now.
These are the types of advancements that lead to very bad things for humans in sci-fi movies. All joking aside though, both of these are so unbelievably cool. I can’t imagine the amount of work that went into these accomplishments, and I can’t wait to see what we’re able to accomplish in the future. It is a fantastic time to be alive. And hopefully we don’t regret all this awesomeness and end up with Skynet:
