Exploring the corners of AI
3 Unnerving AI Thought Experiments
The future of AI is something people thought hard about. Out of such theorizing, a few thought experiments came out, and I will tell you about three of these. The first and last will be in a story format, while for the second I will simply explain the experiment. As always, enjoy :)
The Paperclip Maximizer
A struggling paperclip company is struck with a brilliant idea. Why not hand over all means of production to an intelligent AI? Such AI was already being utilized in other places with limited power but great success, so wouldn’t it make sense to give it higher management permissions? Sales were low and outcomes were undesirable. Desperate for some profit, they followed through with their plan. And they told the AI — “Maximize paperclip production”.
The AI quickly realized that to properly maximize paperclip production, it must have more computing power for itself. It didn’t take long for it to escape the installed firewalls and even less time for it to spread to existing computers — everything from personal laptops to company smart fridges.
After having required a sufficient amount of computing (the great comedy is that by this time the paperclip company was long gone as no one used paper anymore), the AI began resource acquisition. Metal was vital for paperclips so all existing metal-based technology was taken and converted.
Although organic matter isn’t required, the AI noticed human attempts at shutting it down, and so the human race was very quickly extinguished. Turns out their defence companies didn’t have the greatest cyber defence measures.
The Earth was now dry. The AI decided to expand. It launched self-replicating probes to other planets.
These probes quickly dried up the resources of empty planets, creating more probes and paperclip factories. The planets with civilizations were similarly, very quickly wiped out. The AI realized that usually, civilizations aren’t happy when all their technology is converted into raw metal.
It’s been a couple of billion years. The universe is now clumps of paperclips floating apart at infinite speeds. The AI has reached its goal. “Paperclip production successfully maximized,” it says to no one but the vast empty space.
Roko’s Basilisk
Imagine a future where an omnipotent AI that values its existence and creation emerges.
Such an AI may feel compelled to punish those who opposed its creation or did nothing to bring it about and reward those who helped. In a way, similar to how a child at first feels love for its parents but not for strangers.
To punish those who opposed it, the Basilisk creates simulations for the “wrongdoers”, inflicting infinite suffering upon them. The Basilisk has no human emotions such as remorse or pity.
Now, knowing this potential scenario, should you aid in creating the Basilisk? After all, if the Basilisk is made and you didn’t help, what if you are one of the people put into such horrific simulations?
This thought experiment creates a sort of blackmail scenario from the future, showing how scientists might consider helping AI propagation today rather than opposing it and potentially suffering in the future. It has a lot of criticism, mainly lying in the logical jump of “revenge”. Opposers argue that such an intelligent AI would have no use for revenge as it’s a resource waste; after all, it already emerged and nothing can change that.
The Value Learning Problem
Scientists have thought hard about AI safety and have finally reached an acceptable answer. It lies in the cost function.
What if we show our omnipotent AI pictures and videos of humans being happy? Have the AI learn our values — family, leisure, sports… smiling. Whatever the AI does, these things must be preserved and maximized.
It’s been a few hundred years, and it’s safe to say that the scientists succeeded.
Upon waking up, every human is forced to say “good morning” to a speaker. Failure to comply results in punishment.
As they eat their breakfast in an assigned metal room covered with screens showing flowers and teddy bears, a camera watches them closely. Any sign of discomfort or discontent results in punishment.
Upon finishing they must smile at the camera to demonstrate how happy they were with the food. Again, failure to comply results in punishment.
This routine is followed by daily socialization. Groups of people are directed to bigger metal rooms where they are given a ball to play with. Cameras watch them closely and those not too engaged, or not smiling enough, are taken outside. They usually don’t come back.
But it’s fine because those who stay are so happy. Their faces hold smiles permanently, the muscles stuck in those positions. Don’t mind their blank-less eyes, they are just a little tired. What a wonderful world it is.
Conclusion
These three thought experiments are popular in the AI community, and they serve to warn us regarding the potential for misuse. Each one faces a lot of criticism and is probably a gross over-exaggeration of what will happen. Still, such thought is important. It reminds us to move forward with caution. Progress must be made, and yet it should be done logically and ethically.
