Microsoft Research has unveiled its Phi-2 small language model (SML) with claims of exceptional performance despite its relatively small size. The Phi-2 model constitutes 2.7 billion parameters, a scale enabling it to operate on consumer-grade hardware such as laptops or mobile devices. Its performance is said to rival that of much larger models, such as Meta’s Llama 2-7B and Mistral-7B, both with 7 billion parameters.
A Scaled Approach with Reduced Bias
Phi-2 has been benchmarked to outperform even Google’s latest Gemini Nano 2 model, which has half a billion more parameters. Additionally, Microsoft Research asserts that Phi-2 demonstrates lesser instances of providing biased or ‘toxic’ responses compared to the Llama 2 model. The researchers believe that achieving such a balance of efficiency and reduced bias can significantly impact the future deployment of AI in various real-world scenarios.
Moreover, the compact nature of Phi-2 doesn’t seem to compromise its problem-solving abilities, as evidenced by the performance of Phi-2 on a physics problem that had been previously showcased by Google for its Gemini Ultra model. Despite Phi-2’s smaller size, it correctly answered and assisted in rectifying student errors on the physics question, suggesting an advanced comprehension capability within the model. Phi-2 follows just a few months after Microsoft unveiled Phi 1.5 in September.
Licensing Limitations
Despite the promising advancements introduced by Phi-2, there remains a notable barrier to its widespread adoption. The model is currently licensed solely for research purposes under the Microsoft Research License, which restricts its usage to non-commercial, non-revenue-generating research activities. Until the licensing terms are expanded, businesses aiming to utilize Phi-2 for product development or commercial endeavors will not be able to do so.
As Microsoft Research continues to push the boundaries of what small language models can achieve, it indicates a shifting paradigm where leaner AI can perform tasks previously reserved for their larger counterparts, allowing for broader application and integration in low-power environments.
Last Updated on November 7, 2024 11:26 pm CET
