Microsoft seems to be working on its own large language model, which will potentially become a major competitor for AI Google, Anthropic і OpenAI — despite the fact that the corporation itself invested $10 billion from the developer ChatGPT and received the priority right to use its products.
According to the report The Information, the development of MAI-1 headed by Mustafa Suleiman — Former Google AI executive who was CEO of Inflection before Microsoft acquired the startup’s intellectual property for $650 million in March and took most of the staff.
While MAI-1 may be based on methods developed at Inflection, Microsoft is positioning the product as a brand new large-scale language model, the two unnamed employees said. The model supposedly contains 500 billion parameters — significantly more than previous Microsoft open-source models (e.g., Phi-3), which brings MAI-1 closer to to GPT-4 OpenAI which is rumored to boast as many as 1 trillion parameters and is significantly larger than the competing Meta and Mistral models with 70 billion parameters.
The development of MAI-1 implies a dual approach to AI at Microsoft: focusing on small on-premises LLMs, as well as on more modern large models running in the cloud (Apple is considering a similar approach). MAI-1 is also an indication that Microsoft is ultimately ready to explore AI independently of OpenAI, which supports the company’s generative functions with its technology, including Copilot assistant built into Windows.
The exact purpose of MAI-1 is reportedly yet to be determined, but will depend on performance. To train the model, Microsoft dedicated a large cluster of servers with Nvidia GPUs and collected training data from a variety of sources, including GPT-4-generated texts and public data from the Internet.
We should probably expect an announcement about MAI-1 this month at the Microsoft Build developer conference.