Language Models as Statisticians, and as Adapted Organisms

Explore the complex behavior and diverse skills of large language models in this illuminating talk by Jacob Steinhardt from UC Berkeley. Discover two innovative approaches to understanding and improving language models that keep pace with modern advancements. Learn how language models can function as statisticians, processing vast amounts of information to generate useful statistical hypotheses and audit other models for failures. Delve into the internal structure of language models, comparing their adaptation to complex data with DNA in biology. Examine case studies that demonstrate how leveraging internal activations can enhance language model performance and honesty. Gain insights from collaborative research on reverse-engineering transformer models and steering them towards desired behaviors.

Language Models as Statisticians, and as Adapted Organisms