Report: Microsoft has plans for its own AI language model

Microsoft is working on a new, in-house AI language model that is powerful enough to compete with those of Google and OpenAI. This was reported by the tech portal The Information on Monday.

According to the report, which cites two Microsoft employees familiar with the venture, the project is being led by the recently hired Mustafa Suleyman, the co-founder of Google DeepMind and former CEO of AI startup Inflection. The exact purpose of the AI model, known internally as MAI-1, has not yet been determined and will depend on how well it performs. Microsoft could unveil the new model at its Build developer conference later this month, the report adds.

The Information also reports that MAI-1 will be "much larger" than the smaller open-source models that Microsoft has previously trained, meaning it will require more computing power and training data and will therefore be more expensive. According to the report, Microsoft has provided large server clusters with Nvidia's graphics processors as well as large amounts of data to improve the model. MAI-1 will have around 500 billion parameters, it is said. For comparison: Chat-GPT4 from OpenAI has more than a trillion parameters, while Phi-3 mini measures 3.8 billion parameters. Phi-3-mini is a smaller AI model that Microsoft launched last month to appeal to a wider range of users with low-cost options.

Microsoft has invested billions of dollars in OpenAI and uses the ChatGPT maker's technology in its productivity software. However, the relationship with OpenAI has not been free of turbulence in the past. In November, a dispute arose at OpenAI and CEO Sam Altman was briefly dismissed. Microsoft then spontaneously offered all OpenAI employees, including Altman, jobs in its own company. This did not happen. Altman remained with OpenAI and Microsoft secured a place as a "non-voting observer with no control" on OpenAI's non-profit board.

(akn)