“The "Utility Engineering" paper found that as LLMs scale, they develop more coherent preferences, exhibit measurable self-preservation instincts, and form political and demographic biases that can be modeled as utility functions.”