Simon Angus

Simon explains how large language models work and how they differ from familiar kinds of thinking and understanding. In discussing how models are trained, he highlights some important strengths and limitations and how this affects how we should be using and interpreting the results.

Simon’s key messages:

1. State of the art Generative Large language models (LLMs) (e.g. ChatGPT/pro, GPT-4) are statistical completion engines trained on largely Western, English, (and US-centric) web text at heart.
2. Recent conditioning of LLMs to generate more 'aligned' outputs for public use (e.g. Reinforcement Learning from Human Feedback and ‎Proximal Policy Optimization) have made a big leap forward in their safety and appropriateness.
3. However, all users (students and staff) need to develop a healthy scepticism and general awareness of what falls in, and outside of, generative AI tools' current scopes of application.