alt Hacker NewsProbabilistic safety is too low a bar. We need to project the model outputs onto a safe subspace; lobotomize them, if you will. It may make them dumber, but that's a fine price to pay.
I don't work in this space so I don't know the latest, but here's an example: Provably safe systems: the only path to controllable AGI (https://news.ycombinator.com/item?id=37619285)
I don't think the neuro-symbolic model is going to dumb anything down at all. Cyc didn't end up producing anything useful, but the only reason an llm is able to do math at all is that it can keep throwing stuff at lean all day. personally I think that that synthesis will turn out to be more powerful than just llm with constraints, but I don't really work work in the field either.