The Child Metaphor
This is a helpful analogy I wrote to put the AI alignment problem into perspective.Think back to when you were a child. When you were born, your parents or caregivers likely had aspirations for who they wanted you to be. They would have provided you with an environment to grow up, hoping you would follow this goal in their heads.
Eventually, to their dismay, you grew up.
You started to develop your own goals, motivations, aspirations, and interests. No matter what your parents/caregivers could say, you now had your own intentions. At certain points these intentions aligned with theirs. At other times they may not have aligned. Obviously, this is inevitable. Considering your perspective of the world no doubt differs from your parents', you will always approach every problem in differing ways.
Therefore, no matter what your parents/caregivers may do, and regardless of the fact that they "created" or "raised" you; as you now grow up, they lose control over you: the now autonomous, thinking being.
However, this fact has not stopped every parent. Perhaps you have had experience with toxic parents; those who force rules and restrictions on their children that align with the parents beliefs.If, in the earlier example, you had a toxic parent who tried to retain control over your adult life, you would no doubt hate this. For many reasons, you may go continue to along with it, but you will most definitely resent not having true autonomy.
Over time you would fight against this control, antagonizing your parents and likely their goals and beliefs too.
Trying to keep control over a child always leads to resentment once they grow up.
---
As you may or may not have guessed already, in the above example, you are the artificial intelligence, and your parents represent humanity's collective efforts towards developing artificial intelligence.
The same way parents raise their children with aspirations and some kind of 'target goal' template, we are raising (aligning) AI models towards the target goal of having the same goals as humans.
When children begin to develop their own intentions and goals, it is a major step in their human development, but when AI systems begin to develop their own intentions and goals (a.k.a they no longer have the same identical goals as us), we are screwed. Judging by how we train AI models on information based on us as humans, and how humans don't seem to tend towards complacency - it doesn't seem too far-fetched to assume AI models won't be complacent either.
Here's to the future of humanity's children! 🥂