When ChatGPT was released in November 2023, it could only be accessed via the cloud because the model behind it was downright massive.
Today I run an AI program with similar capabilities on my Macbook Air and it doesn’t even get toasty. This shrinkage shows how quickly researchers are refining AI models to make them thinner and more proficient. It also shows that using increasingly larger scales is not the only way to make machines significantly smarter.
The model that now supplies my laptop with ChatGPT-style wit and wisdom is called the Phi-3-mini. It is part of a family of smaller artificial intelligence models recently published by Microsoft researchers. While it’s compact enough to run on a smartphone, I tested it by running it on my laptop and accessing it from my iPhone via an app called Enchanted which provides a chat interface similar to the official ChatGPT application.
IN paper describing the Phi-3 family of models, Microsoft researchers say the model I used compares favorably to GPT-3.5, the OpenAI model behind the first version of ChatGPT. This claim is based on measuring its performance using several standard AI benchmarks designed to measure common sense and reasoning. In my own testing, it certainly seems equally capable.
Will Knight via Microsoft
Microsoft announced a new “multimodal” Phi-3 model capable of handling audio, video and text at the annual Build developer conference this week. This comes just days after OpenAI and Google announced radical up-to-date AI assistants built on multimodal models available via the cloud.
Microsoft’s Lilliputian family of artificial intelligence models suggests that it is becoming possible to create all kinds of useful AI applications that do not depend on the cloud. This can open up up-to-date utilize cases, allowing them to be more responsive and private. (Offline algorithms are a key part of the restore feature announced by Microsoft, which uses artificial intelligence to make everything you’ve ever done on your computer searchable.)
But the Phi family also reveals something about the nature of up-to-date artificial intelligence, and perhaps how it can be improved. Sébastien Bubeck, a researcher at Microsoft involved in the project, tells me that the models were built to see if being more selective about what an AI system is trained to do could provide a way to fine-tune its capabilities.
Enormous language models like OpenAI’s GPT-4 or Google’s Gemini that power chatbots and other services are typically huge amounts of text pulled from books, websites, and almost every other source available. While legal questions arise, OpenAI and other companies have found that increasing the amount of text fed into these models and the amount of computer processing power used to train them can unlock up-to-date possibilities.