AJ Weeks

If alignment evaluations are phrased differently, does model behaviour change? In this post, we describe our attempt to answer this question. This project was completed as a 5-day project sprint during the Capstone Week of ARENA 8.0. TL;DR: The data we collected were noisy! Almost every eval, model, and prompt...

The current race towards producing general artificial intelligence systems brings with it severe risks, yet no AI company developing frontier models is addressing these risks at a level proportional to the pace of development. The rapid integration of this poorly-understood technology into nearly all aspects of society is precarious at...