AI Safety Reading Group: CODA

This week, we will be discussing the paper: "Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data". Dinner, boba provided for the first 20 who RSVP weekly (before Friday morning!) To avoid food wastage, please only RSVP if you are going to show up. If you do not have access to CODA, please arrive by 4:55pm at the latest. The AI Safety Initiative at Georgia Tech is a community of technical and policy researchers at Georgia Tech aimed at reducing risks from advanced artificial intelligence. As such, we focus on topics including: Our recent work includes AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors and An Independent Safety Evaluation of Kimi K2.5‍. An in-depth overview of our research can be found here. The goal of this reading group is to introduce GT’s technical research talent to the above problems, current approaches/solutions, and birth & execute impactful research. Past meeting details here.