SafetyPage

Evaluators and safety organisations · Article · Apollo Research

Embedded Evaluators are necessary for meaningful external testing

Read at apolloresearch.ai ↗Read at sourceNo licence statement found

Outline · 45 paragraphs

  1. Current external access is insufficient for meaningful safety assessments¶3
  2. Embedded evaluations are necessary¶13
  3. We might not get more warning shots¶18
  4. Embedded third party evaluations are feasible¶22
  5. Requirements to make Embedded Evaluations a success¶27
  6. Deeper access has already produced important safety findings¶37
  7. Embedded evaluations are necessary but not sufficient for safety¶44