Capability or Alignment? Respect the LLM Base Model’s Capability During Alignment — LessWrong