DS Meetup 2025-05: LLM interpretability, and discussion

To all data enthusiasts in Bergen, let’s meet up, learn, and connect!

17:00-17:20—mingling and pizza
17:20-18:15—"Interpretability of Large Language Models" with William Wale from UiB, followed by discussion
Abstract: Large Language Models are becoming increasingly ubiquitous. People are using them for mundane every-day tasks, and they are becoming widely deployed in industry. However, they are trained end-to-end using gradient descent on large datasets, and this means that they end up being black-boxes. This is a problem. If we don't know how LLMs work "on the inside", we are severely limited in what guarantees we can make about their behavior.

The last three years has seen the rapid development of a field called mechanistic interpretability which seeks to address this problem by developing a mechanistic understanding of the learned behavior of neural networks. This talk will survey some of the development in the field, talk about some applications, and then afterwards there will be discussion.

We will discuss the future of AI, this time focusing on potential dangers. So take some time to think about what you think might happen! Let's share perspectives and concerns about the trajectory of AI development.
18:15-??:??—mingling

Location: Capgemini
Krinkelkroken 1, 5014 Bergen

If your company or school would be interested in sponsoring an upcoming event, please let us know. We'd like this to be a community-led event, and we rely on venue and food sponsors to make that happen. Venue sponsors will have priority for a presenter slot. Share with us the cool stuff that's happening at your organization!