I think it would actually be helpful if researchers made more experiments with AGI agents showing what could go wrong and how to deal with such error conditions. I don't think that the "social sciences" approach to that works.
Comments