Large language models: A Statistician’s perspective

Abstract: Large Language Models (LLMs) such as OpenAI's ChatGPT (Chatbot Generative Pre-trained Transformer) series, including GPT-3.5 and GPT-4, and Google's Bard (similar in approach methodologically to GPT) have excited many people this year but how do they work exactly and do they really represent artificial intelligence? In this seminar, Richard will attempt to reveal the mystery behind such models. It turns out that LLMs are inherently statistical, essentially word prediction with a little random variation thrown in. That somewhat simplistic explanation rather hides the fact that the output from these models is often rather remarkable. It’s worth, however, delving into the good, the bad and the downright deceitful in terms of LLM output.

Bio: Richard Saldanha is a Visiting Lecturer in Statistical Machine Learning in the School of Economics and Finance at Queen Mary University of London. He is also a Guest Lecturer in Machine Learning and Finance at the Saïd Business School, University of Oxford and an MSc Supervisor for Operations Research and Analytics projects in the Department of Mathematics at the London School of Economics. In terms of research, Richard is involved in the AI for Control Problems Project at The Alan Turing Institute. Prior to this, Richard worked in the City of London for many years involved in risk, trading and fund management activities at various institutions. He is still engaged in quantitative finance via Oxquant, a small Oxford-based consultancy that he manages with Dr Drago Indjic.

Richard holds a doctorate (DPhil) in Statistics from the University of Oxford and is a Fellow and Chartered Statistician (CStat) of the Royal Statistical Society.

This event is organised by the Royal Statistical Society Oxford Local Group.