Generative AI models like ChatGPT are flocking to us, but to use them the vast majority of users head to the cloud, where powerful servers process our requests and perform the complex calculations necessary to respond to us. This is exactly the problem. We rely on the cloud to use the ChatGPTs of the world, but this may change in the short term, and Microsoft is among those positioning itself to achieve this.

Phi-3. Microsoft has released a new version of its “lightweight” artificial intelligence model called Phi-3 Mini. This is the 3.8B version with 3.8 billion parameters: the more a model uses it, the more complex and powerful it can be. Reducing this number of parameters often hurts the chatbot’s accuracy and capability, but Microsoft says Phi-3 outperforms Phi-2, introduced in December 2023, and can provide similar answers to a model 10 times its size.

More versions in sight. This model is actually part of the family of lightweight models, and the company is preparing to launch two slightly larger models, Phi-3 Small (7B) and Phi-3 Medium (14B). Capacity and power are increasing with them, but it is not yet known what the real capacity will be and what the perfect use case will be to take advantage of them.

Comparable to ChatGPT. Eric Boyd, vice president of Microsoft Azure AI Platform, explained to The Verge that Phi-3 is exactly as capable as LLMs like GPT-3.5, which we use for free when we use ChatGPT. “It just works at a smaller size factor,” Boyd says.

Microsoft engineers tested the Phi-3 Mini on an iPhone with an A16 Bionic processor and achieved 12 coins per second; This was truly remarkable fluidity for a smartphone. Source: Microsoft (arXiv).

A model (and more) for our mobile phone. The aim of this model, like the previous model, is to show that it is possible to deliver an AI model that is small enough to be executed on the hardware of our smartphones and still have good performance and user experience. And not only on this hardware, because it can also be run on small servers or on our personal computers and laptops.

run on iPhone. As seen in the image above from the study published by Microsoft, researchers tested the Phi-3 with the Apple A16 SoC on an iPhone. This chip was used in the iPhone 14 Pro/Max as well as the iPhone 15 and 15 Plus. With its Phi-3 hardware, it could produce text at a rate of 12 coins per second; This is a remarkable, if not spectacular, figure – especially if we want long answers.

More and more competition. The Gemini Nano was the first big hero of the “mobile AI models” group, but it was soon followed by both the Phi-2 and specialized models such as the Google Gemma 2B and 7B, the Claude 3 Haiku, and even the recently released Llama 3 8B. From Meta.

Studied “Children’s Stories”. One of the curious things about the model is how it is trained. At Microsoft, they were inspired by the way children learn from bedtime stories and wanted to educate Phi-3 with books covering a variety of topics, with simpler words and simple structures. According to Boyd, “There aren’t enough children’s books, so we took a list of over 3,000 words and asked the MA to create “children’s books” to teach Phi.[-3]”.

inherited from ancestors. While Phi-1 focuses on programming and Phi-2 focuses on reasoning, Phi-3 does both better than its predecessors. It can’t compete with GPT-4, which is much more ambitiously trained, but it can be very useful if they want to use it with internal data subsets, for example, and at the same time consume much less resources.

Advantages and disadvantages. Microsoft proposes this model as an alternative to the current large models, which are ambitious and more sensitive, but also require significant energy consumption. With Phi-3, a company or individual can conduct their interactions with the chatbot locally and the responses will be good enough for those use cases without being too precise or complete. The required hardware and energy savings are combined with its main advantage: not being dependent on the cloud.

in Xataka | Meta, IBM and others form the AI ​​Alliance. Purpose: Advocate for the development of Open Source Artificial Intelligence models