There is a fear of AI. Indeed, AI CAN be dangerous, but so are governments (causing wars), raising farm animals(disease), using electricity(global warming), etc... and I would argue that this will not be different. Unlike most fearing the doom of humanity, I am suggesting an optimistic future.

Moore's law reconsidered.

Generalized Moore's law suggests that the capability of several things during its growth phase double every X year. This is true for computers, biology, etc. However, it also does not go faster than that. It already takes self-improvement into consideration. If the AI is as good as a human, it will take several years for it to be several times better. Moreover, this development will likely come to aid in human potentials such as genetic modification and brain-computer-interface among other technologies, meaning humans won't quickly become obsolete like people would assume. We may eventually become AI or merge with AI.

Outer/Inner alignment reconsidered/RLHF is one of the solutions.

Reality is messy. Human value is not a simple function. Why are we expecting the AI to be this way? What people solving double alignment problems try to do is essentially create a paperclip maximizer, but the one which optimizes "human value" instead, and then argue that even making a paperclip maximizer maximize paperclip would be a big advancement. The reason we have many different goals is that the REWARD only creates learning, but is not what we're optimizing for. We want money, housing, entertainment, etc... if we only wanted rewards(dopamine, serotonin, etc), we would've all just consumed drugs all day. What this means is that something like a shut-down switch would probably work. The AI that strictly pursues reward might see this as bad, but who is to assume the AI does? The AI was not trained to avoid the shut-down switch being pressed. Also, AI forcing humans to press its reward button seems like a good strategy until you realize the AI probably learned to do something which correlates with humans pressing the reward button like helping humans. It has not been trained to force humans to give rewards. As you can see, in this example, the AI strictly solves neither the outer nor inner alignment problem, yet it works. This reward button is what RLHF is.

We're in the early day of alignment research.

The progress in RLHF and so on is despite us being only at the early...

Posts

Wiki Contributions

Comments