4 min readNew DelhiMay 9, 2024 08:14 AM IST
First published on: May 8, 2024 at 01:23 PM IST
The AI arms race seems to be intensifying, as reportedly tech giant Microsoft is building a new Large Language Model (LLM). The new LLM named MAI-1 is said to perform as well as some of the leading LLMs of the day including OpenAI’s GPT-4.
It is reported that Microsoft is currently training its in-house LLM. This is a significant development as earlier, the company was focused on investing in OpenAI and using its models including the GPT-4 family of AI models to power various Microsoft Copilot products.
This is the first time, after investing over $10 billion in OpenAI to use its AI models, that Microsoft is training an in-house AI that is big enough to compete with models from Anthropic and Google.
Based on the report in The Information, Microsoft’s in-house AI is being referred to as MAI-1. The model is said to be much larger than any of the smaller, open-source models that Microsoft previously trained. This essentially means that the model would require significantly large computing power, and training data making it more expensive. Reports also suggest that MAI-1 will have around 500 billion parameters.
When compared, OpenAI’s GPT-4 has over a trillion parameters on the other hand smaller models from Meta and Mistral come with 70 billion parameters. The work on this model is overseen by Mustafa Suleyman, CEO of Microsoft AI. Suleyman, an AI pioneer, co-founded DeepMind and Inflection AI, before switching to Microsoft in March this year.
Reportedly, Microsoft may preview MAI-1 close to the Build developer conference later this month. So far, there is no knowledge about the exact purpose of the model. Some reports also claim that Microsoft is dedicating a significant amount of computing resources like Nvidia GPUs and data to train MAI-1.
Microsoft’s clarification
While there were reports that claimed that the development of MAI-1 was to compete with Microsoft, however, Microsoft CTO Kevin Scott took to his LinkedIn profile to clarify.
“I’m not sure why this is news, but just to summarise the obvious: we build big supercomputers to train AI models; our partner Open AI uses these supercomputers to train frontier-defining models; and then we both make these models available in products, and services so that lots of people can benefit from them. We rather like this arrangement” wrote Scott.
In his post, Scott added that Microsoft has been working similarly for almost five years. “Each supercomputer we build for Open AI is a lot bigger than the one that preceded it, and each frontier model they train is a lot more powerful than its predecessors,” he wrote.
The CTO said that Microsoft will continue to be on this path–building increasingly powerful supercomputers for OpenAI to train the models that will set the pace for the whole field – well into the future.
Scott’s post also acknowledges Microsoft’s independent research on AI. The team has been working on some smaller AI models and some larger ones. The post comes at a time when reports claimed MAI-1 to feature 500 billion parameters directly competing with Google and other models. Some reports also mentioned Microsoft’s recent acquisition of talent and technology from Inflection AI. The tech giant has not confirmed any of the developments.
The new in-house AI model can reportedly compete with state-of-the-art models from Google, Meta, and Anthropic. (File photo) 
Share
