Thursday, May 23, 2024
AITech News

Microsoft could challenge GPT-4 and Google with new AI model: MAI-1


Microsoft is said to be entering another player into the AI model market, with its newest iteration titled ‘Mai-1,’ and reports suggest it could be better than the rest.

Although the technology giant has already invested over $10 billion into OpenAI, the company is said to be training a new in-house AI model that can compete.

The news comes from The Information who say the development of the large language model is being led by Mustafa Suleyman, the former Google AI leader who served as the CEO of the AI startup Inflection.

Inflection is a California-based company that creates conversational AI that you can talk to. Its team created the personal AI called Pi which was released in May 2023. Microsoft was one of its investors in a $1.3 billion funding round.

Both Suleyman and the other co-founder of Inflection, Karén Simonyan, were poached by Microsoft in March 2024. Other members of the team joined the duo at Microsoft too.

MAI-1 is expected to be created solely by Microsoft and is separate from Inflection, but it could build on training data and other technology as they paid for the start-up’s rights to its intellectual property for the sum of $650 million in March.

What we know about Microsoft’s MAI-1

Aside from this report, there hasn’t been any confirmation or news about MAI-1 officially published.

The rumor suggests it could have roughly 500 billion parameters, meaning MAI-1 will be significantly larger than Microsoft’s previous open-source models like Phi-3.

OpenAI’s GPT-4 is rumored to have over 1 trillion parameters and the French AI tool Mistral is thought to be at 70 billion parameter models. Microsoft and Mistral signed a multi-year partnership in February.

While start-ups are finding their feet and making waves in AI like Anthropic and xAI, Microsoft is a massive company that should have the huge amounts of data needed to train the MAI-1 model.

Microsoft has been allocating a large cluster of servers with Nvidia GPUs and compiling training data from various sources for MAI-1, according to the source. This includes text generated from OpenAI’s GPT-4, alongside public internet data.

The Information further reports how a preview of MAI-1 could come as early as the Microsoft Build conference later this month (May 23 – 25) if progress is made within development.

Featured Image: Via Ideogram