Microsoft has shown off the newest member of its Phi family of creative AI models.
Microsoft says that the new model, called Phi-4, is better than the ones that came before it in a number of ways, especially when it comes to answering maths problems. That’s partly because the training data was better.
As of Thursday night, Phi-4 can only be used for study under a Microsoft research license agreement and on Microsoft’s brand-new Azure AI Foundry development platform, which has very limited access.
The size of this Microsoft small language model is 14 billion parameters, and it will be up against small models like GPT-4o mini, Gemini 2.0 Flash, and Claude 3.5 Haiku. It’s often faster and cheaper to run these smaller AI models, and over the last few years, they’ve slowly gotten better.
The performance boost in this situation is due to Microsoft using “high-quality synthetic datasets” along with “high-quality datasets of human-generated content” and some unspecified improvements made after training.
These days, a lot of AI labs are focussing on what new things they can create with fake data and after training. Alexandr Wang, CEO of Scale AI, said in a tweet on Thursday that “we have reached a pre-training data wall.” This confirms what many news stories have been saying about this topic over the last few weeks.
Also Read: Now That Microsoft Has Added Openai as a Rival in Ai and Search,
Phi-4 is the first model in the Phi line to come out since Sébastien Bubeck left the company. Before joining OpenAI, Bubeck was one of Microsoft’s vice presidents of AI and a key figure in the company’s work on the Phi model. He left Microsoft in October to join OpenAI.
What do you say about this story? Visit Parhlo World For more.
