Microsoft has announced the release of Phi-2, a new smaller and more nimble artificial intelligence (AI) model geared towards more specific use cases[1%5E]. This announcement follows the unveiling of Phi-1, the first of Microsoft’s small language models (SLMs), and Phi-1.5, designed to offer increased efficiency and performance in business applications.
Phi-2: Competing with the Giants
Phi-2, a 2.7 billion-parameter language model, is designed to outperform large language models (LLMs) up to 25 times larger[1%5E]. As Microsoft is a significant stockholder and partner with OpenAI, developers of ChatGPT and the massive GPT-4 LLM with 1.7 trillion parameters, its shift towards smaller models is a significant development in the AI space.
Generative AI applications, like ChatGPT and Bard, often require extensive processing power and can be time-consuming and expensive to train for specific use cases[1%5E]. Small, targeted industry- or business-focused models like Phi-2 can often produce better results customized to businesses’ unique needs.
SLMs: The Future of AI?
Experts speculate that the emergence of smaller models could challenge the dominance of current leading LLMs like OpenAI’s GPT-4, Meta AI’s LLaMA 2, and Google’s PaLM 2[1%5E]. Gartner Research’s vice president distinguished analyst Avivah Litan points out that eventually, the scaling of GPU chips will fail to keep up with increases in model sizes[1%5E]. The chip shortage not only affects the creation of LLMs but also impacts user companies looking to tweak models or build their own proprietary LLMs.
The development of more domain-specific language models trained on targeted data to handle tasks like online chatbots for financial services clients or generative AI applications that can summarize electronic healthcare records is on the rise[1%5E]. Microsoft’s Phi-2 is being pitched as an “ideal playground for researchers”, focusing on topics like mechanistic interpretability, safety improvements, or fine-tuning experimentation for various tasks[1%5E]. Phi-2 is available in the Azure AI Studio model catalog.
A More Efficient Approach
According to Victor Botev, former AI research engineer at Chalmers University and CTO and co-founder at the AI-powered startup Iris.ai, Microsoft’s release of Phi-2 is a testament to the notion that there’s more to AI than merely increasing the size of models[1%5E]. He mentions that carefully structured, domain-specific knowledge can help ensure language models process and reproduce information factually and accurately[1%5E].
Smaller models with high performance like Phi-2 represent the way forward in an increasingly AI-driven world, as companies of all sizes seek efficient, cost-effective AI solutions tailored to their specific needs[1%5E].