Can we Trust ChatGPT?

Speaker: Dr. Julie Weeds, Senior lecturer in Computer Science and Artificial Intelligence

The media has everyone excited about pre-trained large language models (LLMs), such as OpenAI’s ChatGPT and Google’s Bard, which can generate human-like responses to questions and requests for information. In this talk, I will describe the evolution of language modelling over the last 20 years including n-gram models, word-vector models, deep neural networks and the transformer architecture underlying many LLMs. I will discuss strengths, weaknesses and pitfalls as well as current research at Sussex evaluating the accuracy of the information in text generated by LLMs such as ChatGPT.

There has been a lot of media coverage around AI recently, especially with the arrival of large language models (LLMs), such as OpenAI’s ChatGPT and Google’s Bard, which can generate human-like responses to questions and requests for information. In this talk, I will describe the evolution of language modelling over the last 20 years from n-gram models and word-vector models to deep neural networks and the transformer architecture which underly the pre-trained LLMs including BERT and, more recently, ChatGPT and Bard. I will talk about what LLMs are trained to do and therefore what we can expect them to be very good at. I will also highlight some of the potential ways LLMs might get things wrong and why. I will discuss current research at Sussex which is looking at how we evaluate the accuracy of the information in text generated by LLMs and conclude with an assessment of the extent to which we can trust LLMs.

***

Dr Julie Weeds is a senior lecturer in Computer Science and Artificial Intelligence at the University of Sussex. She specialises in Natural Language Processing and has been carrying out research in the field of distributional semantics, how meaning can be captured by statistical models of large volumes of text data, for over 20 years.

***

More details of our activities can be found at IET Sussex Network
Refreshments are available from 6:30 and the talk starts at 7pm!
Easy, free, parking available.