Amazon unveils the next generation of AI models
December 05, 2024
At AWS re:Invent, Amazon introduced the next generation of AI models, Amazon Nova, featuring six neural networks for processing text, images, and videos.
Micro is optimized for text tasks with minimal costs. Lite and Pro support multimodal processing, differing in speed and accuracy. Premier is designed for advanced analytics. Canvas generates images, while Reel creates videos from images.
The highlight of these new solutions is the increased context window size. For instance, Micro supports 128,000 tokens, while Lite and Pro handle 300,000 tokens, significantly expanding the models' capabilities.
Neural networks and chips
One of the key events of the conference was Amazon's collaboration with Apple in utilizing Trainium2 chips. Apple confirmed its adoption of these chips to enhance services like Siri, Apple Maps, and Apple Music, improving efficiency by 40%. The chips will also support the startup Anthropic in training large language models, boosting computational power by 500%.
New cooling technologies and efficient solutions
The company announced enhancements to its data center infrastructure, including the clustering of semiconductors for AI training and the introduction of liquid chip cooling. Backup power systems will also be optimized, and energy efficiency improved. Over the coming years, the company plans to invest $75 billion in modernization.
A new era of accuracy
Amazon also unveiled the Automated Reasoning tool to combat AI hallucinations. It verifies neural network responses for factual accuracy, enhancing the precision and quality of AI performance in real-world applications. The system employs "logically precise" reasoning and focuses on improving analytical quality, distinguishing it from similar solutions by Microsoft and Google.