Microsoft has announced a new AI model called Phi-3 Mini. The Windows maker says the AI model is as powerful as GPT 3.5, the language model behind the free version of ChatGPT, despite being trained on a much smaller data set.
Microsoft has revealed details about Phi-3 Mini for the first time in a scientific paper. The model was trained on 3.3 trillion tokens and has about 3.8 billion parameters. That is many times smaller than OpenAI’s GPT 3.5, but according to Microsoft, Phi-3 Mini can nevertheless achieve comparable performance.
What is striking about Phi-3 Mini is that it is trained with a catalog of ‘children’s books’, which in turn are written by other language models. This allowed Phi-3 Mini to learn many simple concepts, just like children listening to fairy tales.
For smartphones and laptops
The fact that the model is so small makes it possible to use it locally on laptops or smartphones, without them being connected to the internet. It is not intended that such AI models will be used in place of larger ones. Rather, they are developed to perform specific tasks. They can also be useful for companies that use internal data to train their own AI model.
More companies are working on AI applications that only run locally on devices themselves and do not require a connection to the cloud. The new AI features that Apple is working on should also work on the iPhone itself.
Read more about Microsoft, or stay informed with our new app.
