Primary speaker:
John DiMarco, IT Director, Department of Computer Science
Description:
The rate of improvement in the capability of AI models is increasing rapidly, as is the degree of autonomy that an agentic model can exercise. While agentic AI models typically pursue broad goals set by humans, unlike traditionally programmed software systems, the steps taken and sub-goals pursued to achieve those goals are decided by the model itself, not by human programmers. The benefits of entrusting these highly capable AI models with more and more responsibility are clear, but what of the risks? If an AI model is deciding what to do on its own, what if it decides to do something we don't want it to do, something that may be unethical, harmful or damaging? The more capable the model, the more value there is in entrusting great responsibility to it, but the greater the risk if it does something wrong.
AI alignment is the field of research that attempts to ensure that AI systems behave properly, in ways that humans want or need. This lecture, for a general audience, will examine what AI alignment is, what the current approaches to ensure alignment are, what recent research has shown and what challenges will need to be overcome for the future. There will be time for Q&A at the end.
