Are Aligned Language Models “Adversarially Aligned”?

Channel:
United States Simons Institute for the Theory of Computing
Subscribers:
68,600
Published on ● Video Link: https://www.youtube.com/watch?v=uqOfC3KSZFc


Duration: 1:02:41
2,126 views
70

Nicholas Carlini (Google DeepMind)
https://simons.berkeley.edu/talks/nicholas-carlini-google-deepmind-2023-08-16
Large Language Models and Transformers

An "aligned" model is "helpful and harmless". In this talk I will show that while language models may be aligned under typical situations, they are not "adversarially aligned". Using standard techniques from adversarial examples, we can construct inputs to otherwise-aligned language models to coerce them into emitting harmful text and performing harmful behavior. Creating aligned models robust to adversaries will require significant advances in both alignment and adversarial machine learning.



Other Videos By Simons Institute for the Theory of Computing


2023-08-19A data-centric view on reliable generalization: From ImageNet to LAION-5B
2023-08-19Pretraining Task Diversity and the Emergence of Non-Bayesian In-Context Learning for Regression
2023-08-18In-Context Learning: A Case Study of Simple Function Classes
2023-08-18Watermarking of Large Language Models
2023-08-18Human-AI Interaction in the Age of Large Language Models
2023-08-18Are LLMs the Beginning or End of NLP?
2023-08-18Beyond Language: Scaling up Robot Ontogeny
2023-08-18Integrating Language into Intelligent Architectures
2023-08-17Meaning in the age of large language models
2023-08-17Formalizing Explanations of Neural Network Behaviors
2023-08-17Are Aligned Language Models “Adversarially Aligned”?
2023-08-17Language Models as Statisticians, and as Adapted Organisms
2023-08-17On Localization in Language Models
2023-08-17Panel Discussion
2023-08-16Large Language Models Meet Copyright Law
2023-08-16How to Use Self-Play for Language Models to Improve at Solving Programming Puzzles
2023-08-16Build an Ecosystem, Not a Monolith
2023-08-16Interpretability via Symbolic Distillation
2023-08-16A Theory for Emergence of Complex Skills in Language Models
2023-08-16Scaling Data-Constrained Language Models
2023-08-15Understanding the Origins and Taxonomy of Neural Scaling Laws


Tags:
Simons Institute
theoretical computer science
UC Berkeley
Computer Science
Theory of Computation
Theory of Computing
Large Language Models and Transformers
Nicholas Carlini