Artificial Superintelligence [Audio only] | Two Minute Papers #29

Dear fellow scholars, this is two minute papers with now. I'll eat my hat if this is going to be two minutes, but I really hope you fellow scholars are going to like this little discussion. Neil Degrasse Tyson described a cool thought experiment in one of his talks. He mentioned that the difference of the human and the monkey DNA is really small, a one digit percentage. For simplicity, let's say it is 1% for this 1% difference. There is a huge difference in the intellect of humans and apes. The smartest chimpanzee you can imagine can do tasks like clapping his hands to a given simple rhythm or strike a match. Compared to the average chimpanzee, such an animal would be an equivalent of Einstein or John von Neumann. He can clap his hands. What is that for humans? Children can do that even before they start studying. They can effortlessly do something that rivals the brightest minds monkeys could ever produce. Imagine if there were a species that is the same 1% difference away from us humans in the same direction. What could they be capable of? Their small children would be composing beautiful symphonies, perfect harmonization for hundreds of instruments. Or they would be deriving everything in the history of physics, from Newton's laws to quantum electrodynamics. And their parents would be like, oh, look at what little Jimmy did. That's adorable. And they would put it on the fridge with a magnet, just like we do with the adorable little scribbles of our children. Just thinking about the possibilities gives me chills. Now let's transition into neural networks. An artificial neural network is a crude approximation of the human brain that we can simulate on a computer to recognize images, paint in the style of famous artists, or learn to play video games, and a number of different very useful things. The number of connections that we can simulate on a graphical card of our computer grows closely to what's predicted in Moore's Law, which means that the computing capacity that we have in our home computers doubles every few years. It's pretty crazy if you think about it, but most of you fellow scholars have phones in your pockets that have more computing capacity than NASA had to land on the moon. As years go by, there will be more and more connections in these artificial neural networks. They don't have to adhere to stringent constraints like our brains do, such as fitting into the human cranium. A computer can be the size of a building or even bigger. Computers also transmit data with the speed of light, which is way faster than the transfer capabilities of the human brain. Nick Bostrom asked a lot of leading AI researchers on the Speed of progress in this field. And the conclusion of the study was basically that the question is not can we achieve human level intelligence? But when we will achieve it. However, the number of connections is not everything. As an artificial neural network is by far not a 1:1 copy of the human brain. We need something more than this. A very promising possible next frontier to conquer is called recursive self improvement. Recursive self improvement means that we tell the program to instead of work on an ordinary task like do better image recognition, we would order it to work on improving its own intelligence, ask the program itself to rewrite its code to be more efficient and more general. So we have a program with a ton of computational resources working on getting smarter, and as it suddenly gets just a bit smarter, we then have a smarter machine than can again be asked to improve its own intelligence, but it is now more capable of doing that. Therefore, if we do this many times, leaps are going to get bigger and bigger, as an intelligent can do more to improve itself than an insect can. This way we may end up with an intelligence explosion, which means a possible exponential increase in capabilities. And if this is the case, talking about human level intelligence is completely irrelevant during this process. Given enough resources, the system may go from the intelligence of an insect to something way beyond the capabilities of the most intelligent person who ever lived. In about a second or less. It would come up with way better solutions in milliseconds than anything you've seen on two minute papers. And there's plenty of brilliant works out there. And of course, it could also develop never before seen superweapons to unleash an unprecedented destruction on Earth. We wouldn't know if it would do it, but it is capable of doing that, which is quite alarming. I am not surprised that Elon Musk compares creating an artificial superintelligence to, say, summoning the demon. And he offered $10 million to research a safe way to develop this technology. Which is obviously not nearly enough, but it is an excellent way to raise awareness. Now, the classical argument on how to curb such a superintelligence if one recognizes that it is up to no good. People say that, well, I'll unplug it, or maybe lock it away from the Internet. The problem is that people assume that they can do it. We can lock it up in any way we can think of, but there's only so much we can do. Because as Neil Degr, as Tyson argued, even the smartest human who ever lived would be a blabbering, drooling idiot compared to such an intelligence. How easy is it for a grown adult to fool a child? A piece of cake. The intelligence gap between us and the superintelligence is more than a thousand times that. It's even more pathetic than a child or even a dog who tries to fool us. We humans can anticipate threats like wielding weapons or locking dangerous animals into cages. And so can super intelligent beings also anticipate our threat, only way better. It can trick you by pretending to be broken. And when the engineer goes there to fix the code, the manipulation can begin. It could also communicate with gravitational waves or any kind of thing that we cannot even fathom. Just as an ant has no idea about our radio waves, and we don't need to characterize super intelligent beings as an adversary. The road to hell is paved with good intentions. It may very well be possible that we assign it a completely benign task that anyone could agree with, and it would end up in a disaster in a way we cannot anticipate. Imagine assigning it the task of maximizing the number of paper clips. Nick Bostrom argues that it would at first maybe create better blueprints and factory lines, and after some point it may run out of resources on Earth. Then, in order to maximize the number of paperclips, it would recognize that humans contain lots of useful atoms. So eradicating humanity would only be logical to maximize the number of paperclips. Think about another task. Creating the best approximation of the number PI one can approximate to the most decimals. By using more resources to have more resources, one builds more and bigger computers. At some point it runs out of space and eradicates humans because they are in the way of creating more computers. Or it may eradicate humans way before that, because it knows that they are capable of shutting you down. And if you get shut down, there's going to be less digits or paper clips. So again, it's only logical to kill them. The task will be done, but no one will be there anymore to say thank you. It is a bit like a movie where there's an intelligent car and the driver is in a car chase situation, shouting, we're too slow and fuel is running out. Please throw out all excessive useless weights and along some empty bottles. The person would be subsequently ejected from the vehicle. We don't know what is going to be the next invention of mankind, but we know what's going to be the last artificial superintelligence. It has the potential to either eradicate humanity or solve all of its problems. It is both the deadliest weapon that will ever exist and the key to eternal life. We need to be vigilant about the fact that we have tons of money invested in artificial intelligence research, but barely any to make sure we are doing it in a controlled and ethical way. This task needs some of the brightest minds of our generation, and perhaps even the next one, and this needs to happen before we get there. When we are there, it's already too late. I highly recommend an absolutely fantastic article on Wait, but why? About this or Nick Bostrom's amazing book Superintelligence? There are tons of other reading materials in the description box for the more curious fellow scholars out there. Thanks for watching and for your generous support, and I'll see you next time.