Artificial Intelligence: Mankind's Last Invention - Technological Singularity Explained
It could be terrible and it could be great it's not clear I'm right but what one thing is for sure we will not control it go is arguably the most complex board game in existence its goal is simple surround more territory than your opponent this game has been played by humans for the past 2,500 years and is thought to be the oldest board game still being played today. However it's not only humans that are playing this game now in 2016 google deepmind's alphago beat 18 time world champion Lisa doll in four out of five games now normally a computer beating a human at a game like chess or checkers wouldn't be that impressive but go is different go cannot be solved by brute force go cannot be predicted there are over 10 to the 170 moves possible and go to put that into perspective.
There are only 10 to the 80 atoms in the observable universe alphago was trained using data from real human go games it ran through millions of games and learned the techniques used and even made up new ones that no one had ever seen and this is very impressive alone however what many people don't know is that only a year after alphago's victory over Lisa Dole a brand new AI called alphago zero beat the original alphago not in four out of five games not in five out of five games not in ten out of ten games but beat alphago 100 to zero 100 games in a row the most impressive part it learned how to play with zero human interaction. This technique is more powerful than any previous version why it isn't restricted to human knowledge no data was given no historical figures were given with just the bare-bones rules alphago zero surpassed the previous alphago and only 40 days of learning in only 40 days it surpassed over 2,500 years of strategy and knowledge it only played against itself and is now regarded as the best go player in the world even though it isn't human.
But wait, if this AI learned how to play without any human interaction made-up strategies of its own and then beat us with those strata jeez then that means there's more non-human knowledge about go than there is human and if we continue to develop artificial intelligence then that means there's gonna be more and more non-human intelligence eventually there's going to be a point where we represent the minority of intelligence maybe even a very miniscule amount that's fine though we can just turn it off right it's a thought but think when modern-day humans begin to take over the planet why didn't the chimps and the and earth all's turn us off if this artificial intelligence becomes super intelligent and learns through and is connected to the Internet well we can't just shut down the entire Internet there's no off switch so what happens if we end up stuck with AI that is constantly and exponentially getting smarter than we are what if it gets to a point that us humans get in the way and the AI hits the off switch on human.
When people think of AI they tend to think of super intelligent AI AI that serves the human race but could also end us at a moment's notice what is this really going to happen pop culture mostly movies tend to depict AI not as benevolent creations but rather as robots with a malicious intent except for Tarzan interstellar he's pretty cool there's a lot more to AI than you might think there's a lot more types that serve different purposes artificial narrow intelligence also known as weak AI is the only form of artificial intelligence that humanity has created so far. It makes this sound kind of bad but trust me it does a really good job at what it's supposed to do narrow AI is AI that was created for the sole purpose of handling one task it's the kind of AI that alphago is narrow AI is good as speech and image recognition and also at playing games like chess or go or even pretty complex games like dota 2 at the International 2017 World Championship. OpenAI artificial intelligence destroyed pro player Dendi 2-0 but much like alphago 0 it wasn't taught how to play the game it played out millions of years worth of 1 vs. 1 matches against itself and learned on its own it started out barely knowing how to walk and eventually as time went on its surpassed human level skill.
If you use Spotify you'll see that it creates a daily mixes for you based on the music you listen to Amazon learns from it teaches itself your buying habits suggest you new products but it seems like a common reoccurrence where these AI teach themselves how to do the task at hand how is that even possible well through something called machine learning. Machine learning is the science of trying to get computers to learn and think like we humans do machine learning is essentially the same way that babies learn. We've sort of has small screaming sacks of meat but over time we improve our learning we take in more data and more information from observations and interactions and most of the time we end up pretty smart the most popular technique out there to make a computer mimic a human brain is known as a neural network our brains are pretty good at solving problems but each neuron in your brain is only responsible for solving a very miniscule part of any problem think of it like an assembly line where each neuron in your brain has a certain job to do in order to completely solve a problem.
Let's make a simple example of a neural network in order to say that someone is alive they have to either have a pulse or they must be breathing but not necessarily both at the same time. This yellow dot represents a neuron in a neural network it functions just like a neuron and your brain does it takes in information and then gives an output if this neuron takes in the information that says hey this person has a pulse and is breathing then the neuron deciphers this information and says okay, this person is alive. It learns to analyze situations where the human would be declared alive if it's breathing or has a pulse in situations where the human would be dead where neither of those is true, that's essentially a bare-bones explanation of how it works of course no neural network is really this simple many have millions of parameters and are much more complex than just this one layer Network.
The world is full of sounds and visuals and just data in general and we take in all of this to form our view of reality however as more and more complex topics show up with more and more data it becomes harder and harder for humans to do this analysis on their own this is where machine learning comes in handy. Machines can not only analyze data given to it but also learn from it and adapt its own view of it. Let's go back to alphago zero in only 40 days it surpassed thousands of years of strategy and knowledge and even made up some of its own strategies but how did it do all of this so quickly biological neurons in your brain operate at about 200 Hertz that's proved to be fine for us but modern transistors operate at over 2 gigahertz. An entire order of magnitude faster those neurons in your brain traveled to what are known as axons and they travel at about a hundred meters per second which is pretty fast and gives us pretty good reaction time but it's only about a third as fast as the speed of sound computers however transmit information at the speed of light or 300 million meters per second so there's quite a big difference between our brains capabilities and our computers.
In just one week a computer can do 20,000 years worth of human level research or simulations or anything that is trained to do a brain has to fit inside your head there's a limit to how much space they can take up but a computer could fill an entire room or even an entire building now obviously we can't require an entire server room to run like we saw with open AI it only took a USB stick but for more intelligent AI it may require much more power artificial general intelligence or AGI is AI with more than a single purpose AGI is almost at or equal to human level intelligence and this is where we're trying to get to but there's a problem.
The more we search into it the harder and harder it seems to be able to achieve think about how you perceive things when someone asks you a complicated question you have to sort through a ton of unrelated thoughts and observations and articulate a concise response to that question this isn't exactly the easiest thing for a computer to achieve see humans aren't able to process information at the speed of light like computers can but we can plant things we can think of smart ways to solve problems without having to brute-force through every option getting you a computer to human level thinking is hard we humans can create things we invent things we create societies and play games and laugh these are all very hard things to teach a computer how can you teach a computer to create something that doesn't exist or hasn't even been thought of and what would be its incentive to do so I believe AGI or strong AI is the most important artificial intelligence to be created and here's why machine learning is exponential meaning that it starts off rather slow but there's a certain tipping point where things start to speed up drastically.
The difference between weak AI and strong AI is millions of times larger than the difference between strong AI and super intelligent AI once we have the artificial general intelligence that can function like a human being for the most part at least it may help us reach super intelligence level in only a few months or perhaps even weeks. But here comes another big problem see many people tend to see intelligence on a graph like this we have maybe an ant here a mouse at about here the average human here and maybe Einstein right here just above if you asked most people where a super intelligent AI would lie on this graph most would probably put it somewhere around here. But this just isn't the case although AI might not be at human level intelligence yet it will be one day and it won't stop at human level it'll most likely just zoom past and continue getting on more and more advanced until eventually the graph looks something like this this is what is known as the technological singularity where artificial intelligence becomes so advanced to the point where there's an extremely explosion of new knowledge and information, some that might not even be able to be understood by humans.
If we make a super intelligent AI that AI would be able to improve upon itself and it turn gets smarter in a shorter amount of time which means that this new and improved AI could do the same thing it continues to repeat this process doing it faster and faster each time the first recreation may take a month the second a week a third a day and this keeps going until it's the billions of times smarter than all of humanity compared to chimps. We share about 96% of their DNA we are 96% chimp but in that little four percent we went from being extremely hairy and mediocre primates to his B C's that has left the planet we're a species that plans to colonize Mars a species that made missiles and millions of inventions. All compacted down into that four percent DNA difference the number of genetic differences between a human and a chimp is ten times smaller than the genetic difference between a mouse and a rat, but yet only being four percent different from them their entire fate seems to lie in our hands.
Super intelligent AI is different from software we know today we tend to think our software is something that we program in the computer will follow our rules but with highly advanced AI and through machine learning the AI teaches itself how and what life is as well as how to improve it after a while there is no need for any human inter perhaps humans would even slow it down as opposed to speeding it up. We have to be careful about how we make it as Sam Harris has stated I mean the first thing you want to do is not give it access to the Internet right, your what you just cage this thing right so because you don't want to get out right so but you want to tempt it you want to see that it see if it's trying to get out how do you know whether it is so this is called a honeypot strategy where you tempt it to make certain moves in the direction of acquiring more power for example if we're somehow able to give a super intelligent AI orders and it follows those orders it may just take the quickest and easiest route to solve them.
Just because we make a super intelligent AI that doesn't mean that it's going to be wise see there's a difference between intelligence and wisdom intelligence is more about making mistakes and acquiring knowledge and being able to solve problems through that wisdom on the other hand is about applying the correct knowledge in the most efficient way wisdom is being able to see beyond the intelligence gained and being able to apply that to other things in hopefully a productive way if we give AI in order to solve world hunger well the easiest way to solve world hunger is just to kill all life on the planet and then nothing would ever be hungry again. But obviously that isn't what we want we would have to somehow teach the AI to have human like values a moral code to follow and somehow work around that in a way we'd want it to be like a super intelligent human cyborg you may not even realize it but the majority of humans are already cyborgs humans are already becoming less biological and more technical we already put our minds on to non-biological.