For special purpose models, fully curated training data and thought patterns are the obvious way to go, but models that are meant to seriously help you think will either be able to infer dangerous ideas from incomplete data or they will be useless

→ View original post on X — @plinz