Meet MAI-1: Microsoft's new AI model with 500 billion parameters that aims to "beat" the GPT

Microsoft appears to be working on its own large language model, which could potentially become the main competitor to Google AI, Anthropic, and OpenAI - despite the fact that the corporation itself invested $10 billion in the developer of ChatGPT and obtained the priority right to use its products.

According to The Information, the development of MAI-1 is led by Mustafa Suleiman - former head of Google AI, who was CEO of Inflection before Microsoft acquired the startup's intellectual property for $650 million in March and took on a larger number of employees.

Although MAI-1 may be based on methods developed at Inflection, Microsoft is positioning the product as a completely new large language model, say two unnamed employees. The model allegedly contains 500 billion parameters - significantly more than previous Microsoft open-source models (e.g. Phi-3), bringing MAI-1 closer to GPT-4 OpenAI, which, rumor has it, may boast 1 trillion parameters and significantly outperforms competitive models like Meta and Mistral with 70 billion parameters.

The development of MAI-1 involves a dual approach to AI at Microsoft: focusing on small local LLMs, as well as more advanced large models running in the cloud (similar to Apple's approach). MAI-1 is also evidence that Microsoft is ultimately willing to explore artificial intelligence independently of OpenAI, which supports the company's generative functions, including the Copilot assistant embedded in Windows.

It is reported that the exact purpose of MAI-1 has not yet been determined, but will depend on performance. Microsoft allocated a large cluster of servers with Nvidia graphics processors for model training and collected training data from various sources, including texts generated by GPT-4 and public data from the Internet.

Announcements about MAI-1 are likely to be expected later this month at the Microsoft Build developer conference.