Technology giant Microsoft may be developing a new AI (Artificial Intelligence) model with parameters reaching 500 billion, which will directly challenge Google and OpenAI.
On May 6th local time, according to foreign media reports, Microsoft is developing a new AI big model called MAI-1, which is far larger than some of Microsoft's previously launched open source models. In terms of performance, it may be able to compete with well-known big models such as Google's Gemini 1.5, Aerospace's Claude 3, and OpenAI's GPT-4. Microsoft may demonstrate this new model at the Build Developer Conference starting on May 21st.
According to reports, the development of MAI-1 was led by Mustafa Suleyman, former Google AI leader and former CEO of AI startup Inflection. In March of this year, just after leading the latest round of financing for Inflection, Microsoft poached two co founders and some employees, including Suleiman, from the company. According to insiders, Microsoft has paid over $650 million to Infection for this.
Although MAI-1 may have borrowed some of the technology from Inflection's models, some Microsoft employees have stated that MAI-1 is a brand new large language model with approximately 500 billion parameters, far exceeding any open-source model previously trained by Microsoft. Less than a month ago, Microsoft just launched a small AI model called Phi-3, in which the Phi-3 mini was optimized to be deployable on mobile phones with 3.8 billion parameters. After training with 3.3 trillion tokens, Microsoft claimed that its performance is comparable to models such as GPT-3.5.
In contrast, the scale of MAI-1 is much larger and requires more computing power and training data. To train this model, Microsoft has configured a large number of server clusters containing NVIDIA GPUs and collected training data from various sources, including text generated by OpenAI's GPT-4 and public Internet data. This may put MAI-1 at a similar level to OpenAI's GPT-4, which is rumored to have over 1 trillion parameters.
The development of the Phi series model and MAI-1 reflects Microsoft's dual strategy in the AI field, developing small local models for mobile devices while developing larger scale models supported by the cloud. However, even within Microsoft, the exact purpose of MAI-1 has not been confirmed.
More importantly, MAI-1 reflects that Microsoft does not want to fully rely on its partner OpenAI's technology in the AI field. The technology from OpenAI is driving Microsoft's various generative AI features, including chatbots integrated into Windows. Previously, employees from Microsoft's AI platform team complained that the company's AI strategy focused too much on the partnership with OpenAI, resulting in a reduction in Microsoft's budget for original AI research.
On the 6th, Microsoft's Chief Technology Officer Kevin Scott posted on LinkedIn, indirectly responding to reports about MAI-1. He stated that OpenAI uses supercomputers built by Microsoft to train AI models, and Microsoft's research and product teams are also building AI models: "AI models appear in almost every product, service, and operational process we have, and the teams that make and operate them sometimes need to do some customization work, whether it's training a model from scratch or fine-tuning models built by others. There will be more such cases in the future, some of which are named Turing and MAI."
In order to achieve breakthroughs in AI, Microsoft is also vigorously hoarding computing resources. In early April, there were reports that Microsoft planned to stockpile 1.8 million AI chips by the end of this year, indicating that the company hopes to double its holdings of GPUs by 2024. According to insiders, Microsoft is expected to spend approximately $100 billion on GPUs and data centers from fiscal year 2024 to fiscal year 2027 (ending on June 30, 2027).