OpenAI interviews
OpenAI interviewsInternal agent engineering and evaluation research.
A useful answer needs permitted evidence
“High-quality answers depend on rich, accurate context.”
Bonnie Xu, Aravind Suresh, and Emma TangArticle authors, OpenAI
“users can only query tables they already have permission to access.”
Bonnie Xu, Aravind Suresh, and Emma TangArticle authors, OpenAI
Test the test
“Use human feedback to calibrate automated scoring.”
OpenAIEvaluation guidance
Check whether the test transfers
“Eval-awareness itself does not necessarily lead to a change in model behavior.”
Joy Hsu and Sam ArnesenArticle authors, OpenAI Alignment Research
Last reviewed