There’s two main scenarios that can happen in the future—either artificial intelligence wipes us out, or it aligns with human values. In the first scenario, I’ll be dead, so, I don’t really care. However, if artificial intelligence does learn how to align with human values, how would it do that?
One of the best ways, probably, to do that is to teach AI the same way we teach humans. For instance, we could pay some couple to raise an artificial intelligence as their child; since most children grow up to have a good understanding of values and ethics, the AI should as well. However, there’s a slight problem—the age of majority in most developed countries is eighteen years, and considering the rate of AI research, the amount of time between the development of the first neural network capable of learning like a human infant and a neural network theoretically capable of destroying humanity would be an order of magnitude less. So, if our goal is to make sure our super-human network has values that align with our own, we’d need to teach it—fast.
Right now, when we train generative models to write text that sounds like humans, we train it on the usual—Wikipedia, Reddit, books, blog posts. If we train a neural network to understand the world like humans do, we would expect the same thing. In just a couple of days, we can feed our superintelligence all there is to know about what it means to be human—all the coming-of-age novels, the romantic movies, the salad recipes, Internet forums, photographs of loved ones. And at that moment, if you’re a part of that cavalcade of data you will have a voice in defining what “being human” means; and otherwise, you won’t. And the best way to have that voice is to write. To teach an AI something as complex as a human, we’re going to need all the data we can get—every Mark Twain novel, every Stackoverflow post, every blog on the internet; anything you create will be part of the concept of humanity.
Be part of the training set.