Mathematics Events
[PAST EVENT] CS Distinguished Talk: Prem Devambu
Access & Features
- Open to the public
Title: Bimodality in Software
Abstract: Every day, there's a new LLM, at a bigger size scale, with a new training method, and potentially high-water-mark SOTA performance on an ever-growing set of coding tasks. While innovations pour forth, many of the underlying ideas seem rather isolated, and sometimes even a bit arbitrary. But are there any general principles on how to use LLMs? We argue one, peculiar to software-related LMs: arising from the Bimodality of Software, wherein software is both formal (with well-defined computational semantics) and natural (written in repetitive ways that are easy for humans to read and write). We show how bimodality is an appealing yet powerful principle that has remained durably useful, over 3 generations of language models, and different training approaches, including very modern in-context learning.
Bio: Prem Devanbu got his B. Tech from IIT Madras, and his PhD from Rutgers University. After 20-odd years at Bell Labs, he joined UC Davis some 25 years ago, where is now a Distinguished Research Professor. His main research interests are in using software data and meta-data to make the life and work of Human Programmers more meaningful, rewarding, and fun.
Sponsored by: Computer Science Department
Contact
[[oscarch]]