Microsoft is reportedly developing its own large-scale language model 'MAI-1,' which could compete with Google, OpenAI, and Anthropic's AI models
Microsoft is reportedly working on its own large-scale language model (LLM), MAI-1 , which could rival cutting-edge AI models such as Google's Gemini, OpenAI's GPT-4, and Anthropic's Claude.
Meet MAI-1: Microsoft Readies New AI Model to Compete With Google, OpenAI — The Information
New Microsoft AI model may challenge GPT-4 and Google Gemini | Ars Technica
https://arstechnica.com/information-technology/2024/05/microsoft-developing-mai-1-language-model-that-may-compete-with-openai-report/
Microsoft reportedly developing MAI-1 AI model with 500B parameters - SiliconANGLE
https://siliconangle.com/2024/05/06/microsoft-reportedly-developing-mai-1-llm-500b-parameters/
Report: Microsoft Working on 'Far Larger' In-House AI Model
https://www.pymnts.com/artificial-intelligence-2/2024/report-microsoft-working-on-far-larger-in-house-ai-model/
Microsoft is developing its own large language model with 500 billion parameters
https://mspoweruser.com/microsoft-is-developing-its-own-large-language-model-with-500-billion-parameters/
In order to use OpenAI's latest AI in its own products, Microsoft invested hundreds of billions of yen in the company in January 2023 and entered into a long-term partnership . The relationship between Microsoft and OpenAI has deepened over time, and the two companies are reportedly planning a $100 billion (approximately 15.4 trillion yen) data center project.
Microsoft uses OpenAI's LLM GPT-4 Turbo for its AI assistant, Microsoft Copilot, but The Information reported that the company is developing its own LLM.
The proprietary LLM that Microsoft is developing is called 'MAI-1,' and the development team is led by Mustafa Suleyman, former head of Google's AI development team. It was announced in March 2024 that he would become CEO of Microsoft AI, a newly established AI division at Microsoft.
Google DeepMind co-founder becomes CEO of Microsoft AI, the company's new AI division - GIGAZINE
While MAI-1 may be built on technology from Inflection, an AI startup founded by Suleiman that Microsoft acquired, a source familiar with the project said, 'MAI-1 is an entirely new LLM.'
According to reports, MAI-1 has about 500 billion parameters, which is considerably more than Microsoft's previous open source AI models such as Phi-3. However, technology media Ars Technica points out that it will require more computing power and training data to run. With the number of parameters reaching 500 billion, MAI-1 is expected to be able to perform on par with OpenAI's GPT-4.
OpenAI explained that the initial version of GPT-3 had 175 billion parameters, but has not disclosed the number of parameters for GPT-4. However, some reports have said that OpenAI's flagship LLM contains 1.76 trillion parameters, and Google's Gemini Ultra also has 1.6 trillion parameters. On the other hand, the number of parameters of the LLMs developed by Meta and Mistral is around 70 billion, so MAI-1 will be an LLM that far exceeds these.
It has been suggested that Microsoft's MAI-1 may focus on two approaches: 'small-scale LLMs running locally' for mobile devices and 'cutting-edge LLMs using the cloud.' In fact, SiliconANGLE points out that with 500 billion parameters, it would be 'too complex' to run locally on consumer devices. Apple, which is rumored to be developing its own LLM, is also reportedly considering two approaches, 'small-scale LLMs running locally' and 'cutting-edge LLMs using the cloud,' just like Microsoft.
Regarding this report, Ars Technica pointed out, 'OpenAI's technology currently powers some of Microsoft's most ambitious generative AI features, such as Copilot, a chatbot built into Windows. But the report that Microsoft is developing its own LLM underscores the company's ambitions in the AI field.'
An anonymous source who provided information to The Information said that the exact purpose of MAI-1 has not been decided even within Microsoft. To train MAI-1, Microsoft is using a large server cluster equipped with NVIDIA GPUs, and is using text generated by GPT-4 and various data collected from the Internet.
Microsoft may announce MAI-1 as early as its developer conference, Microsoft Build, scheduled for late June 2024.
Related Posts:
in Software, Posted by logu_ii