Andy Zou – Top-Down Interpretability for AI Safety [Alignment Workshop] скачать в хорошем качестве

Andy Zou – Top-Down Interpretability for AI Safety [Alignment Workshop]

Andy Zou from Gray Swan AI presents “Improving AI Safety with Top-Down Interpretability,” showcasing how high-level representational analysis improves honesty and jailbreak robustness. By shifting focus from individual neurons to broad neural patterns, this approach helps AI align outputs more closely with its internal beliefs, boosting reliability. Highlights: 🔹 Representation Engineering - Focuses on models' internal beliefs and emergent properties 🔹 Honesty & Truth Detection - Uses internal activity scans to detect consistency with beliefs 🔹 Circuit Breaking - Prevents harmful information generation by redirecting unsafe behaviors 🔹 Real-World Testing - Models robust against jailbreaks and harmful queries The Alignment Workshop is a series of events convening top ML researchers from industry and academia, along with experts in the government and nonprofit sectors, to discuss and debate topics related to AI alignment. The goal is to enable researchers and policymakers to better understand potential risks from advanced AI, and strategies for solving them. If you are interested in attending future workshops, please fill out the following expression of interest form to get notified about future events: https://far.ai/futures-eoi Find more talks on this YouTube channel, and at https://www.alignment-workshop.com/ #AlignmentWorkshop