“Current language models like GPT-4 lack situational awareness and do not understand they are AI models being trained, a capability that will present a fundamentally different alignment challenge once achieved.”