LLMs are often used without understanding their inner workings, leading to factual errors and the generation of insecure code.
A four-phase process explains how models learn language through pre-training, are taught tasks, aligned with human preferences, and refined using reinforcement learning.
Research on Llama 2 models reveals they use English as an internal representation for all tasks due to its prevalence in the training data.
By identifying and amplifying single-meaning concepts, or monosemantic features, it is possible to deterministically control a model's output on specific topics.
Deep learning models inherently memorize unique outlier data from their training set, which explains why LLMs can leak personal information and pose a privacy risk.
Jobs that call for the skills explored in this talk.