Amazon Web Services (AWS) shared that the large language model Falcon 40B by Abu Dhabi’s Technology Innovation Institute (TII), used AWS for its training. AWS also partnered with TII to enhance resilience via SageMaker, ensuring streamlined training processes and minimized interruptions.
FalconLM Performance
TII‘s FalconLM has gained the reputation as the most potent open-source language model currently available. With two variants, Falcon 40B Instruct and Falcon 40B, it outperforms all others on the Hugging Face OpenLLM Leaderboard, relegating Meta’s LLaMA to third place.
SageMaker: The Preferred Choice for Training
Several organizations, including Stability AI, AI 21 Labs, Hugging Face, and LG AI, rely on Amazon SageMaker for developing, training, and deploying their Large Language Models (LLMs), seeking improved performance and cost efficiencies. This preference for SageMaker prompted TII to utilize the service for developing the Falcon 40B model.
SageMaker’s Managed Services and the TII Customisation
Being a fully managed service, SageMaker enabled TII to prioritize the development of custom training methodologies and optimizations over managing its machine learning infrastructure. To decrease training costs and expedite market entry, TII introduced multiple optimizations, inclusive of a custom matrix multiplication for faster training. AWS also collaborated with TII to enhance SageMaker’s resilience, leading to uninterrupted training and minimal requirement for developer intervention.
Amazon Security Lake Announcement
In addition to the TII collaboration, AWS also unveiled the general availability of Amazon Security Lake. This service automatically consolidates a company’s security data from various sources. This includes AWS environments, leading SaaS providers, on-premises setups, and cloud sources, into a purpose-built data lake. The objective is to enable customers to act on security data faster and streamline security data management across hybrid and multi-cloud settings.
Availability
Amazon Security Lake is currently only available in selected regions, with plans for expansion to additional AWS regions in the pipeline.
AI NewsTime