Artificial intelligence gains acceptance: What differentiates weak AI from strong AI.

“The evolution of artificial intelligence can be roughly divided into three phases. The topic of AI first gained momentum around the turn of the millennium. According to Malte Zuch, “often, it was not artificial intelligence in the strict sense of the term but rather hard-coded, rule-based systems. Only since the second decade, from about 2010, have systems been increasingly used that we refer to as AI today. They are able to independently classify complex data sets or derive values from them. They are trained in advance with practice data sets, which is why people also speak of “supervised learning”. Examples include the dynamic pricing systems on popular travel platforms or personalized purchase recommendations in online shops.

According to Zuch’s diagram, the decade of “weak” AI ends this year: “The next ten years belong to “strong” AI, which is far superior to human beings with its abilities. One example is the identification of breast cancer: already today, artificial intelligence often obtains more accurate results than experienced doctors.” Reinforcement learning is the key term. With this method, intelligence systems learn independently how to process complex data and to make strategic decisions based on this. A feature that can be invaluable for companies that process data on a daily basis.

All it knows is its surroundings. In real terms, this may be the legal frameworks or the boundaries of its own company within which the AI can move. Otherwise it doesn’t follow any rules at all but decides solely on the basis of its self-learned knowledge base. Through non-stop testing, it gains experience by trial and error. It is able to differentiate between good and bad experiences and takes note of every single test that resulted in the different experiences. The program may seem completely inept at first. However, its performance becomes more precise and ambitious with every test run, until in the end, a superhuman intelligence system emerges that can independently make the right strategic decisions, even in the most complex situations.

In a demo game, Malte Zuch and his team illustrated how efficiently this kind of software can work compared with its rule-based historical predecessors. A playing field is displayed, which represents the business world of EOS. Yellow fields indicate the receivables of which the AI, the black square, should realize as many as possible within 100 steps. At first, the problem looks trivial – the challenge, however, is that the artificial intelligence has to independently learn “survival” in a randomly generated environment without any prior knowledge. After just one night in training, it is already in a position to successfully navigate over two thirds of the receivables.

Of course, this is just a highly simplified prototype. However, it impressively demonstrates what AI is capable of if not “preloaded” with any habits or previous experiences but is given space to try things out. One prominent example of this technology is the computer program AlphaGo, developed by DeepMind, a company that belongs to Google. It not only taught itself the highly complex board game Go, but was also quickly capable of competing against the best players in the world. By trying out quite a few moves beforehand, it amazed viewers with strategies that even long-time Go players had never seen before. Malte Zuch also finds it “impressive”. “It seems that innovation only comes about by questioning rules.” A quality that is indispensable for the development of artificial intelligence .