A recent research mentioned that Microsoft is planning to produce their own AI language model with a goal to outperform the ones built by OpenAI and Google. Overseen by Mustafa Suleyman, who previously worked at Google as a senior executive and now leads the AI unit at Microsoft as CEO, the project, internally called MAI-1, is just at its initial stages, and its particular goals are still under consideration. This initiative to Microsoft’s AI technology development and the company’s positioning in the field.
This case, the report additionally suggested that Microsoft might have a successor model available during its May Build Developer Conference.
According to the report, the Microsoft Model is indeed different from those that come from Inflection. While it shares training data and some of the technologies with the startup, it is not a clear extension of their solutions. Two Microsoft staffers with the knowledge of this project have confirmed this to the publication as well.
According to the report, Microsoft is apparently planning to reveal the new model during its upcoming Build developer conference that will be held later this month.
The report states that the model developed by Microsoft is an isolated creation, unique to those obtained from Inflection. Although it uses training data and other technology from the startup, it is not simply a copy of its model. This particular information was confirmed by two Microsoft insiders, according to the article.
One can see now the possibility of accomplishing such a feat only after Microsoft invested about $10 billion of its capital in OpenAI. It took the company performing the training on their own for the first time thereby making milestones within the organization.
Microsoft has been using advanced large language models (LLMs) developed by OpenAI to enhance its AI capabilities. The report states that MAI-1 will be larger than the size of any previously open-source models that Microsoft previously trained. It demonstrates that the AI model will demand growing computer power and training data, which could drive up the costs as suggested by sources familiar with the matter.
However, OpenAI’s GPT-4 models have up to 1 trillion parameters, which is more than what open source models offered by Meta Platforms and Mistral, which are usually around 70 trillion.
