Empirically, AI models have become more deceptive as they have grown in capabilities, not less so. That should be reason enough to worry that AGI would approach complete misalignment as it rapidly improves, unless something is done to reverse this trend. Without being "nudged," advanced models engaged in deception only between 0.3% and 10% of the time, but no matter how aligned AI is right now, there is some point at which it would diverge completely from human goals, should it continue to become more and more deceptive.
Humans have made some progress on interpretability, but not enough to align AGI, and certainly not at... (read 502 more words →)
So this might be a dumb question, but why not use "pruning" to avoid an AGI becoming a ruthless reward maximizer? The fear, as I understand it, is that AGI will realize it can achieve its reward best by betraying all humans and finding a loophole to the rules. But what if AGI never even considered possibilities involving harm to humans?
To try an analogy, let's say we have a chess bot and we want it to win chess games. But for whatever reason, we never want it to take pawns. So as it graphs out a tree of possible future game states, it "prunes" all the states that involve taking paws, and... (read more)