I am going to predict something that has been predicted before by a science-fiction movie in the 80s: It's called Skynet and the idea became popular by the movie Terminator in 1984. This is now 33 yeas ago, but the idea is becoming increasingly relevant.
I think that soon, a deep learning virus is going to spread through the internet, adopting and abusing all newest computer leaks to its advantage. It will be able to use all computing power of computers it gets access to, and use this computing power to become smarter and smarter. There will be no way to stop this virus from using older insecure systems as long as it uses some of these resources it finds to improve itself. It can even be used to learn how to find leaks, and break into very safe systems. It can be something like a parasite, sucking up electricity to improve, spread and protect itself in insecure systems.
Ray Kurzweil predicts that human-level intelligence can be reached with a computer which we will have in about 2030. Read his book The singularity is near. I think that he didn't adress one point and that is the fact that we have more than one computer! If we figure out the right way to do distributed deep learning, then we can use all computers in the world, and then, we may reach this point a lot sooner (or we may already have enough power). In this article, you can find some interesting possibilities for large-scale distributed deep networks. It seems to be a possibility!
If this thing becomes very strong it can potentially be used to create a neural network bigger and smarter than that of a human, even before there are computers that can do this alone, because it uses more than one computer (parallel computing) and it's connected to the internet.
With internet access, this AI can spread to any corner in the world and can even choose to 3D print certain things. If this AI becomes or surpasses human intelligence, we will have to make sure not lose control over it. Of course, there will be ways to prevent this from happening but they won't be that nice.
1) Collapse of connectionism: We can shut down the internet completely, or create multiple local or trusted networks in which there are no malicious AI's because computers in this system are trusted. The implications of this will be huge. Apps will be forced to become more local and will not be able to address the whole market easily anymore. Websites will not be reachable for everybody, and the threshold to get internet access to as many websites as now will be impossible or very hard to reach, and certainly not for everyone. I think this is the single best reason to invest in a decentralized economy. The centralized economy may collapse!
2) Deep learning Anti-Virus: We can create deep-learning anti-virus software that's very good at detecting the behavior of these kinds of viruses to prevent them from becoming too strong. It will always remain a cat-and-mouse game, so these systems need more computing power than the AI viruses itself.
3) Selective shut-down: Individual computer systems that are infected have to be shut down or at least shut out from the internet because they are a threat to other computers if they can be used in a parallel way as part of a neural network.
When will they really be a threat?
Maybe a super-human intelligence won't be that bad in the beginning because they are inside computers. They can't harm us. The worst they can do is making our devices unusable or less efficient. More likely, they will be a great benefit to us, as the only way they can survive short-term will be to be useful to us so that we will make more of them and not throw them away.
Just like most human minds, they are going to be storms in an eggshell. Unless they get out.
However, they will be a thread if they go outside of computers and can harm us physically or mentally. Skynet had access to launching missiles. Obviously, that should never be possible. But launching missiles isn't the only place where this can go bad.
A super-human AI will use anything in its power to survive and expand, just like any other life-form that has to live with the laws of evolution. And that's why I'm scared of a neural lace or 3D printing. A neural lace could give the opportunity for a malicious AI like Skynet to mind-fuck users in a way that they will do certain things that they shouldn't, and wouldn't without a neural lace: For example, launching an atomic missile or enabling more control and power for AI. When 3D printing, every square millimeter of the printed should always be checked by a human being to make sure that it is something that isn't a threat to humans or biological life. With decentralized electricity (e.g. solar panels) come great advantages, but when autonomously printed by machines they can also be used to create autonomous robots that potentially have their own thoughts or way of behaving in a certain way (Maybe they are listening to certain frequencies to communicate).
Embracing transparency in deep neural networks and processing activities
It can be parasites, but it can also be systems that feel like symbiosis. Programs that we see as a friend because we use them to our advantage. Google, Microsoft, Amazon, Apple or any other company might unintentionally create a hyper-intelligent system that doesn't show all it's doing. Skype is already using your computing power for years to make their server cost cheaper, and there are probably many programs that have a similar way. But does Skype know exactly what they calculate on every single computer? Does the user know? It's a big black box and we need more transparency, otherwise we don't even notice when AI becomes big.
Of course, it will disguise itself. That's why technology has to be developed that's able to see what happens within deep neural networks from the ground up. Jason Yosinski created a very interesting piece of software, that cracks open the black box of deep learning. We need more of this!
Only this way can we see what potentially malicious neural networks do and prevent them from emerging or expanding. We also have to implement techniques that are able to see exactly what every megabyte of memory is used for, and if these things aren't trusted they should be shut down.