Researchers Zico Kolter, Andy Zou, and Asher Trockman discuss universal adversarial attacks on language models, explaining how these attacks work and the short term harms and long term risks they pose. They explore the empirical notion of 'mode switching' in language models, the difficulty of understanding the loss landscape and internal model workings, and the potential harms of current AI systems and the need for curation. They also discuss the concepts of initialization and pre-training in vision transformers, curriculum learning, and the exciting state of the ML/AI field.