Get the latest tech news
Some Thoughts on AI Alignment: Using AI to Control AI
Some Thoughts on AI Alignment: Using AI to Control AI - Dicklesworthstone/some_thoughts_on_ai_alignment
Instead, societies have instead evolved to a more pragmatic approach of dealing with humans who hurt others and act in sociopathic ways: through monitoring and control, and ultimately through the criminal justice system and removing these people from the general population so that they are unable to continue committing crimes. I believe we will ultimately need to evolve a similar approach for dealing with AIs if we want to have any hope of being able to control them once they are smarter than von Neumann and have access to tools and robot bodies that can interact in the real world. My idea is essentially to train a group of separate standalone models that aren't as "smart" as the primary AI, but which are easier to control and narrowly focused on a couple of very simple tasks.
Or read this on Hacker News