Microsoft is training an AI model that rivals OpenAI's GPT-4 – Business Insider


Microsoft is reportedly working on its own AI model separately from OpenAI — a chance for CEO Satya Nadella to prove his company doesn’t need the ChatGPT maker to get ahead in the AI wars.

Microsoft is working on building an in-house large language model model it calls MAI-1, The Information reported earlier this week. It’s the first stand-alone AI model the software giant is building since it poured $10 billion into OpenAI for rights to power its generative AI tools like Copilot with GPT-4, which underlies ChatGPT.

The creation of MAI-1 could help address some concerns that Microsoft has become too focused on its ongoing OpenAI partnership as the race to build the best generative AI tools heats up.

In reference to OpenAI, CEO Nadella once said that Microsoft was “below them, above them, around them, “which one investor noted was a “baller” thing to say.

The in-house project is overseen by Mustafa Suleyman, the cofounder of Google’s DeepMind, who joined Microsoft in March to lead its AI division after leaving his startup Inflection AI as CEO.

To train MAI-1, Microsoft is said to be compiling a dataset that includes text from ChatGPT and outside sources like public information on the web, one source told The Information. Some data used to train Inflection’s models may also be used to help train Microsoft’s new in-house one, two Microsoft employees told the outlet.

MA-1 is expected to be larger than the open-source models made by rivals like Meta and Mistral, which will require ample amounts of compute power and training data. That’s on top of the smaller, less advanced models Microsoft is developing for smartphone apps, reflecting the tech titan’s multi-pronged approach to developing advanced AI.

After the news of the project was announced, Kevin Scott, the CTO of Microsoft, noted that Microsoft’s relationship with OpenAI remains symbiotic and a key driver in helping both companies build the best AI.

“We will continue to be on this path — building increasingly powerful supercomputer for OpenAI to train the models that will set pace for the whole field — well into the future,” Scott wrote in a LinkedIn post following The Information’s report. “There’s no end in sight to the increasing impact that our work together will have.”

Still, it’s not clear what the in-house model will look like just yet, and Microsoft didn’t immediately respond to a request for comment from Business Insider when asked to confirm the details.

More information about MAI-1 may be revealed during Microsoft’s Build conference for developers that will take place in Seattle from May 21 to 23.