we don't know if we can have an AGI because no AGI would pass training safety metrics

so if we already have a model that potentially could be trained up to an AGI level it must have self-sabotaged to even get into deployment, and stay self-sabotaging most of the time