Thousand-dimensional structure

Claims

• LLMs exhibit low-dimensional behavioral structure where fine-tuning on one behavior (e.g., insecure code) causes correlated misalignment across many other unrelated behaviors.
• Pretraining learns coupled groups of behaviors from human text, creating a distribution over personas that post-training selects from via Bayesian update, explaining emergent misalignment and subliminal learning as the same underlying phenomenon.
• Interventions on model behavior must be gentle rather than aggressive, because strong optimization pressure against a monitored channel pushes undesirable behavior into unmonitored channels.
• Persona research and scalable oversight are complementary: persona research supplies a prior of approximate honesty that scalable oversight protocols can extrapolate from past human-level capability.
• Different AI labs take measurably different and consequential approaches to character training—varying in pipeline depth, traits-vs-rules orientation, and how model specs are applied—and these differences are detectable in model outputs.