Lll

Atlas

Biosec

• modeling the world as a vulnerable and fragile entity
• dual use research

  • advancing our understanding makes us vulenrable to more diseases -> gain of function

AI Safety Cruxes:

• Why can’t complex models develop complex and ill-defined reward functions? (not a crux anymore, this is the problem)
• Why cant; we make good and measured objectives to optimize for? (literally alignment)
• Why won’t super AI be indifferent? (bad protection)

CDT (Causal Decision Theory) / Newcomb’s problem