Associate Professor Tor Ole Bighton Odden from the Center for Computing in Science Education (CCSE) at UiO is visiting UiT. As part of his visit he will be giving a talk about how Large Language Models like ChatGPT work from an educational research perspective.
Title: ChatGPT as improv artist, blurry JPEG, or conceptual blender: Models for LLM "cognition"
Abstract: Over the last year and a half, generative AI tools like ChatGPT have gone from niche research projects in the field of Natural Language Processing to tools that have upended sectors like education, code development, social science, and more. This rapid development is partially due to the fact that these tools have, through gradual updates, developed capabilities that could not have been predicted in November 2022, like the ability to reason spatially, mathematically, and visually. Many users reasonably wonder how, and more importantly, why they work. In this talk, I will approach this subject from an educational research perspective, in which we use theoretical frameworks and models to try to explain the features of complex systems like human (and now machine) learning. I will use three theoretical models—the improv artist (also called the stochastic parrot), the blurry JPEG, and the conceptual blender—to highlight different ways of understanding the capabilities of large language models like ChatGPT, and contrast their implications for how we approach and use these tools.