Your model organisms might be fried

• Model organisms tend to lose capabilities, coherence, and stability of preference structure as we train them.
• Severe organisms for testing techniques; natural organisms for understanding motivations/behavior.
• Suggestions
  • use natural alignment post-training pipelines
  • intervene on behavior by shaping upstream motivations (don’t just train it to exhibit a behavior, train it to express the desire for this behavior & integrate into identity)
  • use their measurements of naturalness as diagnostics but don’t hillclimb them.