Please register to join us June 10 at 6:15pm to hear from our speakers:
- Brandon Roy, Head of Applied Research at the MIT Center for Constructive Communication, will speak on “Designing Tools and Metrics to Support Constructive Communication”
- Itai Rolnick, Senior Software Engineer at Google Research, presents “Train Away Your LLM’s Bad Habits with Constitutional AI”
Meetup Program
Talks
Designing Tools and Metrics to Support Constructive Communication
Speaker: Brandon Roy
Toxic polarization on social media and growing intensity of in-person debates have sparked renewed energy for designing healthy venues for discourse in the public sphere. We introduce work at the MIT Center for Constructive Communication to foster healthy discourse across a variety of populations and organizations through the practice of small group, facilitated conversations. We briefly introduce our approach to scaled data capture and AI-assisted annotation, then focus on early work to operationalize a notion of “constructive” conversation. We emphasize the emergent relationships between participants in terms of their “responsivity” to each other using text embedding based semantic similarity between conversational turns. In parallel, we are exploring visualization approaches to support our intuition and evaluation. As this is early work, we conclude with some planned next steps and an invitation for further discussion. Joint work with Ph.D. candidate Maggie Hughes.
Train Away Your LLM’s Bad Habits with Constitutional AI
Speaker: Itai Rolnick
This presentation provides an overview of Reinforcement Learning with AI Feedback (RLAIF) and Constitutional AI as emerging methods for enhancing the quality and alignment of Large Language Models (LLMs). RLAIF, a successor to Reinforcement Learning from Human Feedback (RLHF) leverages AI-generated feedback instead of relying solely on human input. This approach empowers developers and architects to guide LLM behavior and output quality by specifying their core principles in natural language, eliminating the need for extensive manual data collection.
We will go over the technical details of RLAIF and Constitutional AI, showcasing their potential to scale and simplify the development of LLMs.
Brandon Roy, Head of Research, MIT Center for Constructive Communication
Brandon is a research scientist at the Center for Constructive Communication, with a background in machine learning and cognitive science. Since joining CCC, he has focused on network analysis, with a particular interest in how networks reflect and shape human communication and social structures. He has been doing this work as part of the StoryLine and HealthPULSE projects. He enjoys supporting and collaborating with students on their research. Previously, he led a data science team at Twitter focused on modeling and understanding user interests and behavior. He received a Ph.D. at the MIT Media Lab in the Cognitive Machines group, and he is also an advisor to Cortico.
Itai Rolnick, Senior Software Engineer at Google Research
Itai specializes in machine learning (ML) models and their application across Google products. With a strong foundation in AI/ML, distributed systems, and customer-driven innovation, he excels in building and motivating teams to deliver results in fast-paced environments. Previously, as VP of Engineering at Basis Technology, Itai led a team of 30+ software developers in developing a cutting-edge, scalable text analytics platform. His extensive experience in technology management, entrepreneurship, and innovation spans 15 years. Itai holds an MBA and a B.Sc. in computer science from Tel Aviv University.
