Ninjatech AI recently unveiled the world’s first multi-agent personal artificial intelligence (AI) system – MyNinja.ai – with the aim to tackle time-consuming tasks to increase productivity. The AI is designed to competently handle a variety of tasks independently, such as scheduling meetings, conducting online deep research, writing assistance, and generating code. This is achieved using specialized agents capable of breaking down complex tasks into simpler solutions while continuing to evaluate and learn from previous tasked operations.
Recognizing that no single large language model (LLM) is perfect for every task, MyNinja.ai was built with multiple LLMs specified for different tasks. It relied on the partnership of these different models and cost-effective methods for training them – an otherwise expensive undertaking.
Building a dataset involved multiple task-optimized models as well as fine-tuning the model for each task. By implementing the Lima approach and using a training sample size of roughly 20 million tokens, much attention was put on improving the output structure and tone. For effective evaluation, crowd-sourced user feedback was employed after generating an initial synthetic dataset.
For a pre-trained base model, Llama models were used, specifically the Llama 2 model in different sizes (7B, 13B, and 70B). Trainium chips were utilized to minimize costs and expedite training, allowing for quick tuning and assessment of the models. When newer Llama models were launched leading up to the release of MyNinja.ai, it was easy to transition to using the more accurate Llama 3 for another round of fine-tuning.
To evaluate the performance of our AI model, two primary factors were considered – the model’s ability to answer user questions and the system’s ability to answer questions with references from its sources. Two benchmarking datasets – the HotPotQA and Natural Questions (NQ) Open datasets – were used in this evaluation process. Generally, the model showed considerable improvements compared to baseline models.
Plans for the future include the application of ORPO to better align models for better user results, and the construction of a custom ensemble model from various models utilizing a routing layer. By maintaining quality in various tasks, we hope this effort will radically simplify our model serving and scaling architecture.
In conclusion, empowering access to this transformative technology requires the provision of purpose-built AI chips, top open source models, and a training architecture – a need that AWS has helped fulfill. Ninjatech AI’s ambition to make everyone more productive by relieving them of routine tasks is progressively becoming a reality with agents like MyNinja.ai.
The authors of this piece, Arash Sadrieh, Tahir Azim, and Tengfui Xue, are leading figures at NinjaTech AI. They have wide-ranging backgrounds in computer science, large language models, deep learning, healthcare, and 3D computer vision, among other areas of expertise.
