Recently, Google hit the table by presenting Gemini, its new jewel in the field of AI, promising to be a more complete and intelligent alternative to the famous GPT-4 from OpenAI.
It should be noted that the search engine company’s generative AI has sophisticated multimodal capabilities, mastering human-style conversations, language and content.
In addition, it is capable of understanding and interpreting images, encoding effectively, among many other things. There is no doubt that it is a great Swiss army knife of functionality.
However, one of the big companies, like Microsoft, was not going to sit idly by in the AI race, and has taken us by surprise by presenting its new artificial intelligence model called Phi-2.
Microsoft’s Phi-2 can run on your mobile thanks to its compact size
Phi-2 is a language model that has 2.7 billion parameters. Despite its small size, it can compete without problems and even surpass other much larger competing models.
This technology has been trained with high-quality data, selected for its educational value. Besides, does not use reinforcement learning with people’s opinions, this makes the AI more respectful and avoids bias in responses.
Microsoft claims that Phi-2 performs better than Meta and Mistral’s Llama 2 models on parameters 7B and 13B. This is because it can solve complex reasoning tasks, such as those involving coding and mathematics.
In addition, it can compete and even surpass Google’s AI, Gemini Nano, although the latter is larger.
The best of all is that This model is capable of running on mobile phones and laptopsso it can sweep away the competition with its compact size.
The company has not skimped on features for Phi-2, since it can master everything from conversations and human content to understanding and interpreting images. This AI is a very versatile toolbox that will give you a lot to talk about.
It is worth mentioning that Microsoft has focused on the quality of training datausing carefully designed synthetic sets to teach the model common sense reasoning and general knowledge.
How was Phi-2 trained?
Over two weeks, the model processed 1.4 trillion tokens using 96 Nvidia A100 graphics cards. All learning was carried out autonomously, without subsequent human intervention. This strategy has proven to be effective, providing more robust performance.
The new one Microsoft artificial intelligence toolPhi-2, has been made available to users as part of the Azure AI Studio model catalog.
Without a doubt, it is a very promising technology, since its capabilities are amazing, and its small size makes it more accessible than other AI models.
It will be interesting to see how the competition between Microsoft Phi-2 and Google Gemini in the coming monthsthis means that we are entering the threshold of the real AI war, so 2024 will be very interesting in this area.
