Research Progress Meeting
Date: October 19, 2023
Time: 4:00- 5:00 pm
Location: Sessler Conference Room- 50A-5132 [In-Person and HYBRID]
Speaker: Joshua Batson (Anthropic AI)
Title: More is Different: Generalization in Large (Language) Models
Abstract: Specialized machine learning models have been successfully applied in science and industry for decades. In recent years, a new paradigm has emerged: very large models trained on highly diverse training data have demonstrated remarkable capabilities across hundreds of tasks. Many billions of dollars have since been invested in training and deploying such models. In this talk, I will review some of these developments with a focus on the phenomenon of generalization: as models scale, what changes? What do we know about the internal functioning of these models and how that emerges during training? What does this portend for the future? I will finally speculate, with audience participation, on three potential relationships with physics: the ‘physics’ of model training, the use of models as scientific assistants, and the direct use of models to study physical phenomena.
Join Zoom Meeting
https://lbnl.zoom.us/j/98854322464?pwd=K2tKUm1VZjRlV1J5RHE3cXdHQzRxdz09
Meeting ID: 988 5432 2464
Passcode: 142239