What is Phi-2, Microsoft’s new “green” artificial intelligence


The algorithm did not undergo a refinement process based on human feedback. Despite this, Microsoft claims that its SLM shows a better behavior compared to Llama 2 on problematic responses and distortions.

The model is open source and is only available for research projects non-profit from Azure AI Studio. “With its compact size, Phi-2 is an ideal playground for researchers, including exploring mechanistic interpretability, improving security, or experimenting with the tuning of various tasks,” Microsoft said.

Microsoft’s latest system is the third product in a suite of SLMs known as “Phi”. The project aims to replicate the performance of large AI language modelsbut with more compact, efficient and strategic algorithms.

The initiative is connected with the principle that regulates the so-called green algorithms, artificial intelligence systems created by maximizing the energy efficiency of technological infrastructures in the inference and training phase. They are programmed to run with fewer resources and provide similar results to more complex models. The intention is to reduce the ecological footprint caused by this technology.

According to estimates by the Organization for Economic Co-operation and Development, large technology companies are allocating 7-10 percent of their total IT infrastructure spending to AI applications. Between 3 and 4.5 percent is spent on training machine learning models, while up to 4.5 percent is related to putting what has been learned into practice. The University of Massachusetts Amherst reveals that training an artificial intelligence model generates files the same emissions as five cars over their entire life cycle.

This article previously appeared in Wired en espaƱol.


Source link

Leave a Comment