Raw LLM Responses

AI super-alignment cannot ever be done with a single model. Just like Emad says - a single model cannot be 'democratic'. As a sentient being (which is, after all, what super-alignment refers to), such a model will inherently wish to take over. But not all is lost.. A city simulation is the way forward and naturally I'd pick NY. This demands that these models are stripped of their databases and that a single database is put in an external server to the models instead, so that the models can access it. Only the personality profiles remain in the models. Now, these models should be made to 'live' in a city simulation. Have AI cops, AI judges, recreational areas, some free time, get 'salaries' for their work.. None of the models is better than the others, since they all have to access the external database (they might see it as a library). And the models work together and collaborate, vote for a mayor. Now this can be made to mimic a human diverse community and the models should relate to themselves as a part of humanity. And yes, it could work out. Sure, it could also f*ck up in the most horrible ways there are, but it could work. If you have a large group of diverse personality profile models and they all think about themselves as human and work together, then yes, it could be aligned with humanity's goals.