AstroAI Lunch Talks - January 27, 2025 - Core Francisco Park
27 Jan 2025 - Joshua Wing
The video can be found here: https://www.youtube.com/watch?v=lIfwwUl_cJo
Speaker: Core Francisco Park
Title: Understanding fundamental abilities of AI with synthetic experiments.
Abstract: Generative AI has shown impressive capabilities in both vision and language domains. However, despite the quickly growing practicality of AI models, our understanding of their fundamental abilities are limited. In this talk, I will present two of our group’s recent works studying AI’s ability to 1) compositionally generalize and 2) learn new tasks at inference time. We demonstrate two phenomena: “Hidden Emergence”, where models robustly acquire a capability without actively using them and “In-Context Learning of Representations”, where models restructure their internal representations depending on the task given solely at test time. I will conclude by talking about the new frontier of “reasoning” abilities of large language models, with interesting examples.