Reinventing the Future of AI with Long Context Jamba Instruct

AI21 recently partnered with AGI House to hold its first hackathon leveraging Jamba's 256K context window. Around 100 developers from the Bay Area participated, developing a variety of use cases through 12 hours of intensive coding and collaboration.

In the session, Zapier co-founder Mike Knoop spoke about the challenges of AGI, and AI21's Or Dagan shared the process of designing Jamba through a hybrid architecture combining Transformer and Mamba. There was also a demo of a glossary sheet generator built using Jamba-Instruct.

One of the winning projects, 'FBI AGI', unveiled an AI agent called Emma designed to tackle online harassment. Emma blends naturally into a Discord server, participating in conversations while identifying and profiling the characteristics of offenders.

Emma leverages Jamba's long context to learn from vast amounts of conversation, and by collecting metadata generated during conversations, performs sophisticated monitoring—demonstrating the potential to solve real social problems.