Microsoft Launches Small AI Model, Phi-3

Microsoft has introduced a mini artificial intelligence (AI) model that it says is capable of handling many tasks that had been thought to require far larger models. It also says Phi-3 can outperform models twice its size across a variety of benchmarks that evaluate language, coding and math.

Microsoft in a blogspot explained that the small AI model is cost-effective and that can perform tasks such as content creation and create social media posts while using smaller amounts of data.

It point out that Phi-3 Mini operates on 3.8 billion parameters and promises to offer a balance between capability and economy, making it suitable for devices with limited computing power like smartphones and laptops. By using fewer parameters, Microsoft says that it aims to make AI more accessible and affordable, particularly for businesses with smaller datasets or those needing to integrate AI without heavy computational demands.

Microsoft also announced in its blog post that Phi-3 Mini is readily available on platforms like Azure, Hugging Face, and Ollama, and is also optimised for Nvidia’s GPUs, ensuring broad compatibility and ease of integration.

In addition to this, Microsoft also shed light on how Phi-3 was trained after being inspired by children’s books. In a separate blog post, the company noted that the conventional approach to training large language models involved feeding them massive amounts of internet data to grasp the intricacies of language and generate intelligent responses. However, Microsoft researchers, led by Sebastien Bubeck, Vice President of Generative AI Research, proposed a novel idea. Instead of relying solely on raw web data, they sought exceptionally high-quality data.

The blog post also talked about how Microsoft’s Ronen Eldan was once reading bedtime stories to his daughter when he wondered how she learnt to connect the words written in the book. And inspired by Eldan’s nightly reading sessions with his daughter, Microsoft researchers decided to curate a specific dataset for Phi-3’s training.