Microsoft would like to emancipate itself from OpenAI with its own ChatGPT

Microsoft is reportedly developing MAI-1, its own LLM which should compete with GPT-4. The interest would above all be to emancipate itself a little from OpenAI and not to depend on the company for its advances in artificial intelligence.

Last year, we learned that Microsoft was investing ten billion dollars in OpenAI. Enough to access the company’s advances in artificial intelligence, in particular GPT-4, the state-of-the-art large language model from OpenAI. However, Microsoft would like to depend less on it: the firm would have developed MAI-1, its own LLM.

MAY-1: an LLM from Microsoft that would like to compete with GPT-4

The information was revealed by the aptly named The Information (and relayed by Neowin). According to the newspaper’s sources, Microsoft has developed internally a large language model called MAI-1 (the M refers to Microsoft). It would have 500 billion parameters, compared to the 1,000 billion of GPT-4. A less efficient model a priori therefore.

The project would be headed internally by Microsoft’s new head of AI, Mustafa Suleyman, who arrived in May. As a reminder, the character co-founded DeepMind AI, a subsidiary dedicated to AI at Google, then created his own specialized company, Inflection. Moreover, according to The Informationsome training data would have been reused from Inflection, but no pieces of code.

Remember that this is not the only LLM that Microsoft has developed itself: several versions of Phi (most recently Phi-3) have appeared. It is a small language model with 3 billion and some parameters, designed to run locally on computers and/or smartphones. At the moment, it is not deployed in any Microsoft service.

Depend less on OpenAI: this would be Microsoft’s objective

Microsoft may have spent a lot of money on OpenAI, but the firm does not have control over it. Which means that in terms of the evolution of GPT or Dall-E, or even Sora, Microsoft depends somewhat on the goodwill of OpenAI. An affiliation which could turn against Microsoft’s projects. The latter may want more developments to improve Copilot, when OpenAI may want to delay finalizing its future versions of GPT for example.

Ultimately, Microsoft could very well do without GPT-4 altogether and upgrade Copilot to MAI-1. For the moment, it is difficult to go further than these simple assumptions. Today, Microsoft has several versions of Copilot: initially, the company could very well deploy MAI-1 (or a later version) on one version of its chatbot only. The next time to hear about MAI-1 is perhaps May 21, during Microsoft’s Build conference in Seattle.