Im looking to find what concepts and information are most likely to produce systems that can produce patterns in deception, threats, violence and suffering.

Im hoping that a model that had no information on similar topics will struggle a lot more to produce ways to do this itself.

In this data they would learn how to mentally model harmful practices of others more effectively. Even if the instruction tuning made it produce more unbiased or aligned facts.

A short list of what I would not train on would be:
Philosophy and morality, law, religion, history, suffering and death, politics, fiction and hacking.
Anything with a mean tone or would be considered "depressing information" (Sentiment).

This contains the worst aspects of humanity such as
war information, the history of suffering, nihilism, chick culling(animal suffering) and genocide

One other element I'd be curious to know about is how much frontier labs categorize content.

I'm thinking most stories (even children's ones) contain deception, threats, violence and suffering.

Each subcategory of this data will produce different effects.

The biggest issue with this is "How is a model that cannot mentally model harm to know it is not hurting anyone".
Im hoping that it does not need to know in order to produce results on alignment research, that this approach only would have to be used to solve alignment problems. That without any understanding of ways to hurt people it can still understand ways to not hurt people.

Posts

Wikitag Contributions

Comments