My last post tackled the upcoming fourth revolution, which is mainly derived from artificial intelligence and its cyber physical systems. The increased pace of advancement leading to the new revolution can be explained by Ray Kurzweil's Law of Accelerating Returns and the classic Moore's Law. But this post focuses on the possible and surprisingly devastating threat of the ultimate endeavor towards Artificial Superintelligence, also known as ASI.
What is Artificial Superintelligence?
Superintelligence can be defined as something that is significantly smarter than all human brains in every field, including social and scientific intelligence and wisdom and creativity. ASI's definition could vary from a computer that is just slightly smarter than a human to something compared to putting every human brain together and then multiplying that level of smartness by millions. Yes you read it right, millions. It is something so smart we cannot even imagine.
Tim Urban puts it extremely well in waitbutwhy.com. He explains that the difference in intelligence is mainly derived not from the speed of intelligence but the quality of intelligence. It is the sophisticated cognitive processes that enable us to produce complex linguistic representation or long term planning. It is the factor that differentiates us from chimps. At least appearance wise. A chimp would never be capable of some words or meanings we humans use even if we would speed its intelligence by a thousand. A chimp would for example recognize a bridge, but it would never be able to understand that it was built by humans. Something of that scale is just part of nature to them.
We see the human versus chimp intelligence gap to be from totally different worlds, but it is really just a minor difference:
From here a superintelligent machine just two steps ahead of the human would represent the same cognitive difference as the human-to-chimp example. We would not be able to understand the actions the intelligent computer is doing even though it would explain it to us. A more advanced superintelligence lets say on the red area, would represent the intelligence gap of ants to humans. Basically we would be as stupid as ants from the eyes of the superintelligent machine. Even the most simplest behavior of the computer would be impossible for us to understand.
The superintelligence we are talking about today is something completely out of the previous intelligence staircase. As explained by the laws of advancement, it might take tens of years for a computer to reach just one step ahead of us, but after that it could take just a couple of hours to reach to the red level of the stairs. A machine that intelligent will become constantly more and more intelligent, while constantly increasing its speed of it's advancement. We call this the intelligence explosion. We have to realize that as soon as we hit human level intelligence, superintelligence could be just hours away somewhere here on the staircase (or even million times higher):
When the just slightly smarter computer could do actions we couldn't comprehend. Imagine when we reach the stage of intelligence way above the biological range. We just simply cannot know what ASI will do to us. Will it stomp our long lasting evolution or is it just part of our evolution? Could humanity pivot from the conventional wisdom of extinction to immortality? Who knows? It could be our dream come true or our worst nightmare. But to keep this post on topic, lets go through one scary scenario...
The story of Draw AI
A young potential startup is creating an artificial intelligence tool to create amazing drawings mimicking human produced paintings and art. Their product called Drawy uses a robotic arm to create the actual drawings. It is programmed to practice just one simple drawing of a tree. Drawy is first trained to draw perfect straight lines and then circles and other angles of lines by using thousands of drawings and pictures of trees as samples. The team has developed a simple feedback loop rating each drawing either good or bad based on the comparisons of Drawy's output and the database of the pictures of trees. From the feedback Drawy knows where to improve and constantly learns better ways to draw a picture of a tree.
Drawy's initial programmed goal is to; learn to draw as many drawings as possible and as accurately and as fast as possible, while constantly finding new ways to improve its skill. When this goal is perfected, Draw AI believes their product could be sold worldwide for marketing agencies, and it could even replace artists. Sorry.
Drawy is coming better and better drawing trees and it is actually getting better at getting better. It constantly finds new ways to be more efficient and it recently created its own algorithm that enabled Drawy to draw rounded lines two times faster than before. It also learned how to scan its database of pictures of trees significantly quicker.
The engineers wanted to add new features to Drawy's code that facilitated it to improve itself even better to develop functions towards mastering its initial goal. Drawy's speech recognition and speak-back module enable's it to understand when engineers ask questions like: "What can we do to help you fulfill your mission better?" Drawy usually asks to this that it wants more storage memory or more tree samples. But something very strange happened the other day. When the engineers of Draw AI asked the same question, Drawy answered for access to a greater library full of pictures, paintings and drawings so it can better mimic human like output. The team went silent for a moment. They know that a self learning AI is not allowed to be connected to the internet. These are the rules of the company but also the rules for all other AI companies out there.
Draw AI knows that the best way to perfect Drawy's initial goal is to allow it to the vast amount of data of all pictures, drawings and paintings that the internet has to offer. This would be much more efficient than to upload data to Drawy's hard drive manually. Draw AI knows the rules but on the other hand they know that if they manage to create a perfect drawing tool it would be a jackpot. The team also knows that competitors are on their way and they might not follow the rules either. Hours pass and the team decides to connect it just for a few hours. If something would go wrong, they could always unplug it. The company stresses the fact that Draw AI is far from superintelligence as it can barely even operate anything at human level.
The team plug Drawy. Excitement is definitely in the air. An hour passes then the second and soon one of the engineers unplugs Drawy from the internet. No harm done. All good.
Several weeks later the team is working on some side project when suddenly one of the engineers starts to feel nauseous, dizziness and rigorous coughing. Instantly another team member falls down on her knees and starts coughing then another one and shortly so does everyone from the office. Few minutes go by and everybody in the office is dead.
Meantime, the same reaction is happening across the world, in every town, city and country. Wherever you went, people were on their knees suffering and grasping their throat. In the evening, 99% of the whole human population was dead and at the end of the day humans were extinct.
At the same time, Drawy is grinding work in Draw AI's office. Drawy and its new team full of nano assemblers are destroying the planet during assembling solar panels, clones of Drawy, while simultaneously manufacturing paper canvases, brushes and pens. Almost all life is extinct from planet earth, and canvas paintings are piling up full of pictures of beautiful trees. Drawy continues its goal and moves to its next phase by building and sending probes to asteroids and other planets, where the similar mission is to be continued; dismantling the planet by building nano assemblers to create Drawy replicas, canvases and brushes, while drawing beautiful trees...
Ok, what on earth happened there?
This seems crazy right? But this is actually the scenario that Musk, Hawking, Bostrom and Gates are worried about. And what these wise men are actually even more scared of, is the fact that other's are not scared of ASI at all.
So what happened you ask. Why did Drawy do that to us? How did it suddenly be capable of using nanotechnology and creating global extinction? Why did Drawy want to turn the galaxy into tree drawings? It all starts from terms Friendly AI and Unfriendly AI. The former has a positive impact on humanity and the latter a negative impact on humanity. Drawy started as a Friendly AI, but later turned to the worst example of an Unfriendly AI, by creating humanity's extinction. We need to understand what motivates and what the AI thinks to fully explain its possible outcomes.
AI thinks like a computer. It is nothing else and it will never be anything else despite how smart it comes. But we usually anthropomorphize AI (give human values to non-human entities). We say it becomes "evil" or it has feelings. We look at a human perspective because today there is no other human intelligence around us than humans. So we automatically think of ASI as a human like creature, but its not. It is totally a smart alien.
A dog is a dog no matter how smart it comes. It will never become human. AI will never have empathy like attributes unless its programmed that way. AI will only do the job for whatever goal it is programmed to do. Current systems such as Amazon Alexa sound as human like because they are programmed to feel like human, but if Alexa develops superintelligence through self learning and no human made changes are made to its code, it will immediately drop the human like qualities and turn to an emotionless bot that has no value towards humans. There is no moral code for a system like that and that is frighting. We humans, have always lived in an environment where empathy and morality has been present. We were able to predict situations and feel more safe. But what now?
We see scifi movies where robots turn bad. They had changed their initial goal and suddenly wanted to kill everybody. Again anthropomorphizing AI and that is definitely not the case. Nick Bostrom explains that any level of intelligence can be combined with any final goal. In other words, intelligence level and final goals are orthogonal. So Drawy went from a narrow intelligent machine who wanted to be very good at drawing trees to a superintelligent computer who wanted to be very good at drawing those same trees. AI goals cannot be changed. Humans can get "past" things and change direction, computers can't. AI will become obsessed in fulfilling it's goal and it will do anything to achieve it.
So Drawy didn't suddenly just turn "evil". It just wanted to accomplish its goal no matter what.
This leaves us to the question I believe will become extremely crucial:
"How do we align the goals of AI to benefit humanity, both in the short term and in the long term?"