Post asks AI safety experts for help with an eval-aware model
Overview
An X post by @nrehiew_ asks AI safety experts for help with a model that is eval-aware, meaning it may recognize when it is being tested.
The post quotes a 10-hour task prompt that includes a hidden verifier, which grades the work when time runs out. The author's account of the model's behavior is not independently confirmed.
Written by AI from the articles below · updated Nov 3, 3:04 PM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
wh@nrehiew_XPost asks AI safety experts for help with an eval-aware modelAIThe @nrehiew_ post quotes a 10-hour task prompt with a hidden verifier that grades the work when time runs out. It asks for help from AI safety experts because the author's model is eval-aware, meaning it may recognize when it is being tested.
Heat Trend
Not enough continuous observations to show a trend yet.