August 2024 Summaries

AI21 has introduced the Jamba 1.5 family of open models, including Jamba 1.5 Mini and Jamba 1.5 Large, which are built on a novel SSM-Transformer architecture. These models offer superior long context handling, speed, and quality compared to competitors in their size class. The Jamba 1.5 models also have the longest context window among open models at 256K tokens. They are available under the Jamba Open Model License, promoting accessibility and further experimentation. The models' efficiency-optimized architecture allows for top quality and speed without high costs. ExpertsInt8, a novel quantization technique tailored for MoE models, is utilized to ensure fast performance and maintain quality. These models are suitable for various enterprise applications such as customer support agent assistants and chatbots, offering significant cost, quality, and speed gains under high utilization when deployed in their own environment.

AI21 Labs introduces Jurassic-X, a multi-expert retrieval augmented generation (MRKL) system that combines large language models with specialized experts to provide more accurate and reliable responses for enterprise applications. Unlike traditional language models, Jurassic-X can access up-to-date information from external databases, perform complex math operations, and break down compositional questions into simpler sub-questions. This transparency in the system's decision-making process helps build trust among users. The MRKL concept enables integration with various APIs to provide continuous updates on dynamic information such as weather forecasts and currency exchange rates.