Episode Details

Is political bias in AI actually improving? This episode examines the surprising trajectory of large language models and political neutrality. We explore the technical reasons behind the shift—from better RLHF pipelines to distributional preference learning—and unpack the fascinating paradox of DeepSeek: a Chinese model that often feels more open on Western political topics than American models do. We discuss how concentrated censorship under Chinese regulations creates space for broader engagement on other issues, and what mixture-of-experts architecture and chain-of-thought reasoning mean for political balance.