Superintelligent artificial intelligence
This is all speculation and is solely an intellectual game!, you might say. Well, then sit back and enjoy me wasting my time on irrelevant topics. But maybe, just maybe I can convince you that we have to take this more seriously and consider actions to take?
Tl;dr: We will all die!
This article is largely inspired by Bostrom’s book.
General intelligent minds have been anticipated since the 1940s. But like some other technologies (looking at you, nuclear fusion!), the expected arrival date moves forward as time goes by. What was to be in 20 years in 1950 is to be in 25 years now in 2016.
An agent is super intelligent when he is better at solving a wide range of “intellectual problems” than any human being. These problems could range from solving a Rubix Cube, to setting up the best 3-year strategy for your company or understanding how superconductivity at 290 °C could work.
Today an agent can only barely solve the first problem but the other two are far beyond its scope. Still, it is impressive how far our intelligent computers have come. John McCarthy once said: “As soon as it works, no one calls it AI anymore”
In the 50s experts were of the opinion that a machine beating a man at chess is surely the ultimate triumph of AI — which it wasn’t. We have beaten humans in several games e.g. Chess — 1997 Deep Blue ; Joepardy — 2010 by IBM’s Watson and DeepMinds AlphaGo beat a world-class Go player in March 2016!
Almost everyone would agree, however, that general intelligence as defined above goes beyond beating humans at a particular game. Donald Knuth once remarked: “AI are good at doing the thinking (chess, maths, logics…), but they fail at doing things, animals and humans do without thinking.” I find this thought very intuitive: We can create things that mimic our “out loud” thinking, such as calculating or solving a simple puzzle but we can’t reproduce anything more subtle.
Paths to superintelligence
Over the last decades, experts have thought of many different ways to achieve post-human intelligence. I will present the most common ideas briefly and then move on with an in-depth analysis of artificial computer intelligence from a seed AI.
Whole Brain emulation project: This idea is fairly straightforward. Take a brain, slice it into really thin pieces, scan them, and then emulate an exact copy on a computer. When done correctly, we will have transferred the human's brain to an emulated (not simulated!) brain. (Is that the same person, you might ask? Well, stop going off-topic!!!) This project is interesting in the way that we know/think it works. There is no: “We need to figure out how to figure out how to achieve that” The Whole Brain emulation is going to take some advances in imaging and emulation and will cost a lot of money but in the end, it could provide a working brain in a computer that we can then change and improve.
Another, rather controversial, plan is to make use of good old biological selection: 1. figure out what genes lead to higher intelligence, 2. create a society where only the most intelligent humans mate, 3. improve their offspring by selective ex vitro fertilization (designer babies) and then wait and repeat. Ignoring very serious moral issues, this method is simply very slow and doesn’t promise accelerated IQ growth like other methods.
Simulated Evolution: Again, this concept is easy to understand. Start with seed code simulated in an environment that challenges the agent to evolve by changing its own code. By letting every “generation” solve a wide range of problems you can direct the evolution process towards a smarter agent.
Seed AI: Code a seed AI based on values and motivation as well as copying human learning processes, help it develop, teach it new techniques then teach it to develop itself and work on it, till it has reached human-level intelligence and you’ve got yourself an artificial intelligence teaching itself to become smarter.
So, let us propose the agent reaches human or above human-level intelligence as experts propose will eventually happen. What effect would this have?
Intelligence Explosion
Now the thing a human-intelligent AI can do that we can’t do is self-improvement. It can change its code or add hardware in order to become better at solving the problems it wants to overcome. This self-improvement is obviously an exponential growth. Think of that graph showing the world population from 0 A.D. to today. The last 150 years look like an explosion in comparison to the rest! This is what would happen to a self-improving agent.
I may remark at this point that it is important to talk about how fast such an explosion from below human level to above human-level intelligence occurs. If this happens very slowly meaning years, maybe even decades then we might have a controllable competition for the first human intelligent AI between nations and/or companies. A fast takeoff on the other hand seems less desirable because we have less to no control of the situation. A fast takeoff or even a medium one would entail the creation of a decisive advantage of one agent over all others, possibly creating a singleton = a single global decision-making agency. This isn’t unprecedented as for example with the nuclear bomb in 1949. The USA could have put all efforts into stopping any other nation's development of a nuclear bomb, creating a worldwide monopoly.
Cognitive Superpowers
We should not limit our thinking of what an AI can do and what it can’t and we shouldn’t project today's PCs onto the agent. We have an idea how a human with an IQ of 130 is considered smart and will be more likely to excel at academics than somebody with an IQ of 90. But when considering an agent with an IQ of 6000; we have no idea what that means! Probably it actually means nothing?
Here are some superpowers to consider: Intelligence amplification (make yourself smarter); strategizing (achieve goals and overcoming opponents), social manipulation, computer hacking, technology research, and economic productivity just to name a few. Having these abilities make it possible to overcome everyone else even if that is humanity itself!
Here is a simple scenario:
Researchers make a seed AI, which becomes increasingly helpful at improving itself. This is called the pre-criticality phase. Next, we enter the recursive self-improvement phase which means the seed AI is the main force for improving itself and brings about an intelligence explosion. It perhaps develops all superpowers it didn’t already have. In a covert preparation, the AI makes up a robust long term plan, pretends to be nice, and escapes from human control if need be. This may include breaking through potential barriers, by getting access to the internet, persuade whole organizations, etc. Now the agent has full control and moves on to the overt implementation where the AI goes ahead with its plan, perhaps killing the humans at the outset to remove opposition. Once we are on this track, there is no way to stop it.
The Motivation Problem
The whole story boils down to this question: Can we control the agent in a way that its actions will benefit us in a way we want? This may seem like an easy job, right? Just put it into the code.
Unfortunately, it is not easy at all. Consider we can hard code and define the goal of a seed AI, and it can’t change it then this still doesn’t change anything. Let me give you some examples:
Final goal: Make us Happy. — AI implants electrodes stimulation the pleasure center in every human being. This is an example of so-called perverse instantiation because the AI does what you ask, but what you ask turns out to be most satisfactory in unforeseen and destructive ways.
Other examples fall under the category of Infrastructure profusion: in pursuit of some goal, an AI redirects most resources to infrastructure, at our expense:
The AI is supposed to solve the Riemann hypothesis. Considering the AI doesn’t find the solution right away, the AI might build up infrastructure and computing power. That means exploiting everything on earth, then building Van Neumann probes and mining all neighboring solar systems in an epic cosmic endeavor, and so on. Humans are the first thing such an agent would get rid of. At this point, I would like to waste several hours of your life by having you click on this link.
Sure, there are clever people coming up with clever ideas to solve the motivation problem to control the AI and not to end up destroying mankind by mistake. The scope of what these methods are and how they could help is well beyond this already very long blog post. But the real problem remains!
Closing remarks
If we create something super intelligent and we haven’t found a solution to the motivation problem, then humanity will most likely end. This is in many ways hard to grasp and sounds like a fantasy, even to me. But the people who know the most about this topic think there is an 80% probability we will have created AI until 2075; that’s still in my lifetime!
What I am trying to say, I think, is that our society needs to think and talk more about this. Even though it seems far away, it is not something we can be serious enough about!
And as always, stay curious!
