Notes on Mathematics for Engineers (1959) [pdf]
Summary
The article is a LessWrong post by Dean Valentine discussing evaluation awareness vs. evaluation paranoia in frontier AI models. It examines how alignment evals are crafted, why models may or may not exhibit true eval awareness, and the challenges of designing reliable tests, with examples like honeypots and GDPx benchmarks.