The future of productivity agents with NinjaTech AI and AWS Trainium

The article discusses how the startup NinjaTech AI built their cutting-edge productivity agent NinjaLLM, the backbone of their AI assistant MyNinja.ai, using AWS Trainium chips. It highlights the advantages of using multiple specialized models fine-tuned for different tasks and how AWS Trainium enabled cost-effective and scalable training of these models.

Specifically, the article covers:

• NinjaTech AI's approach of using multiple specialized models optimized for different tasks like research, coding, and advising to build their AI assistant
• The process of creating a training dataset of 20 million tokens by generating synthetic data and incorporating user feedback
• Fine-tuning the models on AWS Trainium instances, which allowed quick and low-cost iterations for improving model accuracy
• Evaluation of the fine-tuned models on benchmarks like HotPotQA and Natural Questions, achieving notable improvements over baselines
• Future plans for further improving performance through techniques like ORPO and ensemble models