AI21 + AWS : Inside the Engine of a Large Language Model
AI21 + AWS : Inside the Engine of a Large Language Model
Address:
525 Market St, Courtyard Entrance, San Francisco, California 94105, US
Level(s): 300 - Advanced
Join AI21 Labs and AWS for a night of generative AI knowledge sharing, as we look under the hood of Jamba, the first production grade Transformer+Mamba model; and we explore techniques for AI-assisted code generation.
All AI enthusiasts and learners welcome!
Agenda :
5:30 PM PDT - Networking and light snacks
6:00 PM PDT - "Jamba - The Benefits of a Hybrid SSM-Transformer Model" with Yuval Belfer, Member of the Technical Staff - AI21 Labs
In this talk, we briefly dive into technical components of AI21’s Jamba model. Jamba, the first production grade Transformer+Mamba model, combines the quality of Transformers with the speed of Mamba. Jamba is built on top of an SSM-Transformer mixture-of-experts (MoE) architecture. It is based on hybrid interleaving Transformer & SSM layers, enjoying the benefits of both architectures. We describe
7:00 PM PDT - "AI Code Generation and Evaluation" with Anila Joshi, Applied Science Manager - AWS and Kamran Razi, Data Scientist - AWS
Explore AI-assisted code generation, focusing on the integration of retrieval-augmented generation (RAG) with generative AI services on AWS. Learn best practices for optimizing code repositories, setting up rapid prototyping environments using Bedrock, and leveraging agentic workflows with Langgraph. We'll also cover the RAGAS framework for evaluating code, using custom metrics like CodeBleu.
8:00 PM PDT - Q&A and Networking
Find out more?
AI21 Studio - https://studio.ai21.com/v2/chat
Hugging Face (Mini & Large)- https://huggingface.co/collections/ai21labs/jamba-15-66c44befa474a917fcf55251
Jamba 1.5 release blog - https://www.ai21.com/blog/announcing-jamba-model-family
(PDF) Jamba 1.5 paper - https://arxiv.org/pdf/2408.12570