AI Talk: How does ChatGPT work?

An introduction to the technology behind large language models

To work effectively and responsively with the new generative AI tools we need to understand how the tools work and what they are capable of.

In this meeting, we will take a look at the technology underlying large language models like ChatGPT. These models are built using deep learning, a very special form of machine learning. And they also have a very particular internal arrangement, called the transformer model. The combination of deep learning and transformer models, first introduced in 2017, has led to spectacular progress, creating systems that truly rival human intelligence on many important dimensions. We will look at the essential features of this technology, and we will put current developments in the broader perspective of AI research and the general theory of intelligence. 

These new models have already been adopted enthusiastically on a global scale, and they have the potential for productivity improvements in almost every area of human endeavor. At the same time, there are profound challenges concerning plagiarism, bias, the spread of misinformation, and indeed, the potential that humans could simply lose control of this powerful technology, with consequences that are difficult to imagine. We won’t solve these problems at this meeting — but I hope that, with an understanding of the technology, we will all be in a better position to find the way forward with AI.

This meeting is the first in a series of talks on AI in education. Please note that the meeting does not count towards continuous professional development in the HETEP.

Daniel Hardt is an associate professor at Copenhagen Business School, department of Management, Society and Communication. His research deals with language from a theoretical and practical point of view, and his primary research areas include artificial intelligence, computational linguistics, and machine learning.