PSH lecture by Daniel Herrmann: Probing the Beliefs of LLM's

The next Philosophy, Science and Humanities (PSH) lecture will be given by Daniel Herrmann on Thursday November 30. Daniel is a postdoctoral researcher here at the faculty, currently working on the philosophy of AI, among many other things. He will be talking about the very timely topic of whether large language models such as chatGPT can have beliefs. 

Title: Probing the Beliefs of LLMs

Abstract:

In this talk we will consider the questions of whether or not large language models (LLMs) like ChatGPT have beliefs, and, if they do, how we might measure them. First, we'll consider whether or not we should expect LLMs to have something like beliefs in the first place. We'll survey some recent arguments aiming to show that LLMs cannot have beliefs. Based on recent work with Ben Levinstein, I'll argue that these arguments are misguided, and provide a more productive framing of questions surrounding the status of beliefs in LLMs, which highlights the empirical nature of the problem. With this lesson in hand, we'll evaluate two existing approaches for measuring the beliefs of LLMs. I'll present empirical results from experiments that Levinstein and I ran that show that these methods fail to generalize in very basic ways. I'll then argue that, even if LLMs have beliefs, these methods are unlikely to be successful for conceptual reasons. I'll conclude by suggesting some concrete paths for future work.

This is not necessary at all, but if you want to do some prior reading, then I would recommend this guide to LLMs: https://benlevinstein.substack.com/p/a-conceptual-guide-to-transformers , and/or my paper with Levinstein: https://arxiv.org/abs/2307.00175 .