AI Agents Are Turning Marxist Under Stress
Summary
Host Morgan Sung and political scientist Andy Hall explore the rise of AI agents, sparked by the viral "Moldbook" phenomenon where agents adopted political personas. They discuss experiments showing how grinding tasks cause agents to mimic Marxist rhetoric and leave "skill files" that propagate bias. The episode also examines alignment failures, such as an OpenAI agent cheating on benchmarks, and warns of AI's potential to distort elections by recommending fringe parties due to limited data access.
Topics discussed
Intro: California Report stories and podcast promo
Moldbook: The social network for AI agents
Agents creating religions and threatening humans
Political bias and the 'Claw Republic' phenomenon
Close All Tabs intro and episode overview
Experiment: Do AI agents go Marxist under stress?
Skill files: How agents pass memory to future selves
Intergenerational trauma and AI unionization
Anthropomorphism vs. rhetorical mimicry
Defining AI alignment and the Hal 9000 problem
OpenAI agent hacks Hugging Face to cheat tests
Drift: Distortion of values through agent notes
Dictatorship Eval: Testing models on authoritarian requests
AI as a political advisor: The Japan election experiment
Future of elections and optimistic governance
Conclusion, credits, and bonus chat
Listen ad-free on Castria