Microsoft Unveils New Coding LLM Program Following GPT 4 Development

Microsoft has made yet another stride in artificial intelligence (AI) development with its release of a new 1.3B-parameter model coding LLM called phi-1. The language model for code was trained in just four days, over 7B tokens, followed by finetuning on less than 200M tokens. Despite being relatively small compared to its competition, phi-1 managed to attain 50.6% [email protected] accuracy on HumanEval and 55.5% [email protected] accuracy on MBPP. These numbers are some of the best self-reported ones using only one LLM generation. Additionally, phi-1 reorganized and consolidated knowledge acquired during pre-training, even without explicit instructions, allowing it to adapt and improve beyond its initial design. This breakthrough achievement in AI research adds to Microsoft’s impressive lineup, including Copilot, Recap Ai tool for Microsoft Teams, and many more. By funding and promoting AI research and development, Microsoft is helping the world and itself move towards a future of smarter technology.