Mary Phuong

Eval­u­at­ing and mon­i­tor­ing for AI scheming

Un­faith­ful Rea­son­ing Can Fool Chain-of-Thought Monitoring

• Mary Phuong 26 May 2024 10:49 UTC

  16 points

  9

  on: AI com­pa­nies aren’t re­ally us­ing ex­ter­nal evaluators

  The latest Gemini tech report has some more info on GDM external safety testing: https://​​storage.googleapis.com/​​deepmind-media/​​gemini/​​gemini_v1_5_report.pdf
  (section 9.6, p. 71)

  What links here?

  • AI com­pa­nies aren’t re­ally us­ing ex­ter­nal evaluators by Zach Stein-Perlman (24 May 2024 16:01 UTC; 242 points)
  • AI com­pa­nies aren’t re­ally us­ing ex­ter­nal evaluators by Zach Stein-Perlman (EA Forum; 26 May 2024 19:05 UTC; 88 points)

Threat Model Liter­a­ture Review

Clar­ify­ing AI X-risk