Google’s Evil Robots

“Kill All Humans!” These are the words of Bender Bending Rodriquez, the lovable robot from Futurama. Bender is an example of an advanced AI that can learn and think for itself: a robot with free will. While this seems far off, the workers at Google DeepMind are hard at work at bringing artificial intelligence into our lives.

DeepMind, a company that runs under Google, is looking into the world of artificial intelligence. Their whole purpose is to try and make the world a better place through technology. They are trying to formalize intelligence and build powerful, all-purpose algorithms that could revolutionize our world. Their main initiative is self-learning AI. The firm was co-founded in 2010 by Dr. Demis Hassabis, chess prodigy and creator of the Theme Park video game; Mustafa Suleyman, head of Applied AI who dropped out of Oxford to set up a telephone counseling service; and Dr. Shane Legg, who has studied the algorithmic organization of the brain.

DeepMind was acquired by Google back in 2014 for more than $500 million dollars after one of their machines was able to learn how to play 6 different Atari games. This acquisition allowed DeepMind to have access to more resources and to collaborate with Google engineers and computer experts.

Google DeepMind Logo. Used under the Public Domain via Wikimedia Commons

So far, DeepMind has created a neural network that learns how to play video games like humans do. They have also developed a neural network for a turing machine that can have short term memory like humans do.

DeepMind joined Microsoft, Amazon, Google, Facebook, and IBM in the Partnership on AI. Apple followed DeepMind by joining the partnership in January of 2017. The partnership focuses on setting standards and rules for artificial intelligence to help make sure that it won’t destroy the human race. In 2014, DeepMind was awarded Company of the Year by the Cambridge Computer Library.

Despite their rich history, DeepMind made headlines in 2016 when their AlphaGo program beat the former champion at a game of Go, an abstract strategy game that dates back to ancient China, without handicaps and on a full sized board. This was the first time that a computer beat a player at Go as it is harder for a computer to play Go than chess. AlphaGo was then released online for select people to play against. It had a record of 60 wins and 0 losses and was said to learn new strategies every game it played. DeepMind intends to take AlphaGo to full game competitions where it will play against the best players in the world.

Later in 2016, DeepMind teamed up with Moorfields Eye Hospital in London, England. Their goal was to use DeepMind AI to detect early signs of blindness in anonymous eye scans. They are currently working on an algorithm that can differentiate between healthy and unhealthy eyes.

While these machines seem harmless, in 2017, DeepMind was able to show that AI can have killer instincts. In a test, DeepMind wanted to see if two neutral networks would work together or compete. DeepMind set up a 2D game where the networks have to collect apples in a basket on a map. The networks had a choice to tag each other with a laser beam to temporarily disable their player in the game. After over a million simulations, DeepMind concluded that the networks would cooperate when the apples were abundant, but would turn on each other when there were higher stakes. DeepMind also found out that smaller, more simple networks would be more peaceful whereas larger, more complex networks would start to fight quicker. To make things better, DeepMind also learned that the networks would work together when trying to accomplish a common goal. In a second test, networks had to play a game where they were “wolves” and they had to catch a “prey.” The game offers better rewards when the wolves were in close proximity if the prey was captured. The networks learned to work together. Once again, the more complex networks were able to recognize that they needed to work together quicker. These tests confirmed that separate networks can work together and that they can make selfish decisions when needed. They also state that these behaviors are very human-like.

While the end of the world may seem imminent, DeepMind states that these tests are useful in finding out how to control and cooperate with multi-network AI systems that could be integrated into aspects of our life.

DeepMind has learned that they can raise self-learning AI in safe conditions where they can teach the machines to cooperate with humans and treat them as superior. However, there is still the chance that one of the AIs could go rogue and even influence other AIs to turn as well. Will we live in a super advanced society where we live next to self-learning robots, or will our lives play out like the Terminator movies? We will just have to see what our fate is as advancing technology is becoming a bigger part of our lives.