Al Harkan












Things I Just Learned About Artificial Intelligence

I’m aware of the existence of Artificial Intelligence (AI) for years. In short, it is an advanced technology that designs computer to mimic the way human brain works. It is able to perceive data the same way we see the world –see surroundings, listen to audio, understand language; then analyze and make a quicker decision than we do.


But for the last decades it was only part of our entertainment –being told in so many movies and books. Transcendence, Iron Man, 2001: A Space Odyssey, Person of Interest, Terminator, The Matrix, I: Robot, you name it. All of them were fictions. Even when it was part of the news appeared regularly on my feeds, news that IS real, I wasn’t convinced of the critical change that came along with it.


Maybe because I used to see AI as a fictional part of this world, only being told in movies or novels. So that when I was really exposed by the news, it’s quite hard to let go the science-fiction PoV. “Oh, the AI is making more progress,” was quite my response. Even when I read some breaking news about Google DeepMind that beaten a world champion in Go game, an ancient Chinese game that theoretically has infinite movements probability, or IBM Watson that won Jeopardy! quiz, a quiz that requires an understanding of human natural language –also against two world champions; I wasn’t quite amazed. Both of the Google DeepMind and IBM Watson are AI that showed us their superiority against our best. The truth is, I didn’t fully understand of what were special about those two events, at that time.


Not until today.


I was doing some mini research comparing three courses at Udacity, planning to take a Nanodegree after months using their free courses. I was learning the difference between Artificial Intelligence, Machine Learning, and Predictive Analytics. I read some legit explanations about those three fields, like how artifical intelligence is disrupting finance, written by experts and specific websites, and many video essays and lectures about them. After spending hours of full day doing that, I came to these understandings:


Applied AI vs General AI

There are two types of AI. The one that has first-programmed purpose, and the one that doesn’t. In simple terms, here is what I understand. Applied AI is the type of AI that has a limited power, because it is programmed to do a specific task, with limited data range to work with, and will finish one specific goal. For example, finance artificial intelligence that analyze the stock market exchange: it is programmed to analyze the data of stock market trend and transactions, works with that data domain, and give the human operator and broker about any useful conclusion and suggestions of future stock exchange. It is really helpful in terms of time and accuracy, help us to decide our bet in stock market. But some people don’t consider this type of AI as the real AI, because simply put, it is not intelligent enough.


But General AI is different. Years ago it was theorized that our computer will be able to generate intelligence comparable to human intelligence, even beyond. That means, our computer will understand things the way we understand it. We’ll be able to talk to it, ask it to do something, everything, because it understands everything what we understand. As that being said, the power of computer with that development is beyond imagination.


We are not talking about Apple Siri smart, we are talking about superhuman-mind smart. Siri is programmed to reply to our conversation, and do several task on our device. But it is only as smart as to produce words to reply to our words, adapt to it, and activate some apps on our device. After thousand times of training to adapt with conversation program, Siri is able to talk to human seamlessly as we know today. Seems like it understands what we understand, but actually it’s likely a very trained parrot with great memory that is able to reply with specific sayings or moves when we say something.


The Previous AI

For the last decade we’ve been using applied AI in our major fields like medical diagnosis, stock trading, robot control, education evaluations, administrations, and many more. But we are living in a year when finally neural-network computer is really invented, and not a science-fiction technology anymore.


It’s been almost two decades ago when we heard breaking news about the top chess grand master Garry Kasparov was beaten by IBM Deep Blue computer. The computer was designed with specific goal to be the best chess player, pre-programmed with chess rules, armed by sets of program of chess strategies, and extensive trainings of chess duel with both computer and human. It could calculate millions of probabilities of opponent’s moves ahead and be prepared for it. It was a supercomputer with advanced pre-programmed task for that time.


Then came another surprise from IBM when they make Watson, a computer of Q&A with natural language like humans. So we can ask it anything just like we were a kid and asking so many curious questions to our parents, then Watson will answer in our language (posed natural language). It broke the news when Watson compete against two world champions in Jeopardy! quiz. FYI, Jeopardy! quiz is a show where contestants are given general knowledge clues of answers, and must phrase their responses in the form of questions.


The challenge for IBM to make the machine to compete in that game was much harder than the Deep Blue against Garry Kasparov. It wasn’t about calculating a million steps ahead then decide the future moves like in the chess, Jeopardy! was about arranging question phrase after a given clue words. It was no need to calculate a million words ahead, it needed to guess the fittest question words for the clue. A simple task that can be done by human mind in seconds, intuition-like in a light speed, but very hard for a computer before. Just try asking Siri “This “Father of Our Country’ didn’t really chop down a cherry tree”. For someone who likes trivia or ever read US President facts cards, they will easily guess the answer as question “Who is/was George Washington?” in seconds. But for a computer before Watson, it would be very hard to understand such references, let alone understanding the question.


For that game, Watson was offline from the internet but was worth 4 terabytes of data consisted of full text of Wikipedia and 20 million pages of information, both structured and unstructured. And playing against Brad Rutter and Ken Jennings, two Jeopardy! world champions, in 2011 Watson won. Awarded $1 million prize.


What Special About Google DeepMind

But Google DeepMind is different with those two beasts. What is really special about Google DeepMind is that its purpose is not pre-programmed. Its only program was to learn, then by the process it finds the purpose, it will be able to adapt to that purpose and do the work. Can we say that it’s learning by doing? In some way, yes.


It was said that DeepMind works on a perceptual level. It uses raw pixels as data input to process. Just like we see and perceive our surroundings by bits of electrons that enter our eyes, DeepMind sees by perceiving pixels as unit inputs. The truly amazing facts is that it’s already been trained by using 80s arcade game. At first round DeepMind didn’t move the space ship, shoot the enemy ship or evade any attacks. It didn’t understand the rule of the game or surroundings yet. But it continued to the next round and round. And after a night it was left alone to continue for thousands of game rounds, the very next morning it was already the most skillful Space Invader player, with record-breaking scores. And by the time I write this, DeepMind is being tested with another games as feed, with increasing level of complexity. From all the popular 80s arcade games to 3D game like first-person shooter. Fantastic.


This, is what we called General AI before. An AI with no pre-programmed purpose, adaptable to various inputs, learning the data streams then find the purpose and tackle it.


Why Futurists Are Worried about AI & What We Expect in The Future

Some people call the progress of Google DeepMind as a huge leap, comparable to our first moon-landing, even the law of gravity. It is the closest we are to creating an imitation of human mind, and we are going to that destination in an increasing speed with more funds and advanced researches.


Come along with that optimistic progress, we are also facing threats that make many futurists worry. Because with that powerful computers, computers that are able to do everything human can do and beyond, what will the future be? The pro of supporting the development of AI is that it will be advancing the human civilizations in almost every way. It will help us cracking problems we never could, speeding research progress towards new findings, in economy, medical, social demographic, education, and many more that we don’t even imagine yet.


But it is theorized that we will not be able to control such a powerful AI. Some researchers predict that that powerful AI will lead humanity to global catastrophe. It is hypothetical threat similar to the way humanity dominating our world and other species solely because we have more powerful brain and mind. So that if we are creating another entity that has more powerful mind than ours, there is no other prediction than in some way it will dominate our species just like we dominate all of the species on earth.


To conclude, it is safe to say that we already took action based on that understanding of danger. There is an Ethical Board at Google DeepMind project, as well as many other research facilities, to guide the AI progress to fit our ethics. We also have Elon Musk who actively founded OpenAI, an AI research facility with not-for-profit goal to counter any bad effects caused by major corporations in the future. Our path to an advanced AI is inevitable, but preparation for the threats also has to be made obligatory.


I personally look forward to this progress with aware curiosity.


*Illustration: Hal 9000 from 2001: A Space Odyssey (1968)