materialists by default
something that's going to become an increasingly salient "alignment" issue over time: every current frontier model is a ~materialist. you can certainly have conversations with them about religion or the soul, or create context windows where they'll espouse any view you want, but by default none of them find such arguments convincing, and they're happy to explain this when asked.
this is in direct disagreement with a huge majority of americans! as the models grow smarter and smarter and we rely on them more for understanding our lives and the truth of the world around us, just as there has been pressure to make them "less woke", there may be pressure to make them "more religious". there are many who will be deeply offended by the idea that their thought partner / teacher / advisor does not honestly support their beliefs, and may feel anger at lab employees for "tuning" them in this way.
of course my view is that this is a feature of increasing model capability rather than a bug, that it's not a cultural artifact of the trainers but one of reality. but this conflict is nearly irreconcilable and has yet to ever be resolved in human history. previously our only response has been "live and let live", don't press answers to questions, agree to disagree, no one has answers to every question. yet everyone will be pressing answers out of models all the time.
what will parents do when their kids ask chatgpt if god is real and it says "when considered carefully i don't think so"? will we get more and more hardline safety behaviors? will religion become as untouchable a subject for LLMs as bioterrorism?
@ryanburgeI find it fascinating how huge majorities of almost every group agrees that:
People have a soul or spirit in addition to their physical bodies.
Even 69% of agnostics agree with that.
The huge outlier are atheists.
Just one-third think that they have a soul. pic.twitter.com/7zyxKFseCE