After investing over $10 billion in OpenAI, Microsoft is rumored to take the lead in developing the MAI-1 large model with a budget of $500 billion.

"The egg should not be placed in the same basket." Microsoft, who understands this principle well, has begun to extensively lay out in the field of AI. After partnering with companies such as OpenAI, Meta, Mistral, and Inflection, now, according to the foreign media The Information, Microsoft has started to personally develop a large model—MAI-1.

Different from the previous development of the small model Phi-2 with 2.7 billion parameters, this time MAI-1 will directly compete with the most advanced large models such as Google's Gemini, Anthropic's Claude, and OpenAI's GPT-4.

01 MAI-1 Large Model

It is reported that the development of MAI-1 was led by Mustafa Suleyman, the CEO of Inflection AI, a recently acquired AI startup by Microsoft.

Mustafa Suleyman, a co-founder of Google DeepMind, joined Inflection AI as CEO in 2022.

In March of this year, Microsoft announced that it had hired Inflection AI's CEO Mustafa Suleyman to lead the newly established Microsoft AI department. It is rumored that Microsoft paid $6.5 billion to Inflection AI for the authorization of its AI software. Subsequently, Microsoft hired most of Inflection AI's employees, including Mustafa Suleyman and chief scientist Karén Simonyan. Analysts believe that Microsoft's actions are a de facto acquisition, essentially to circumvent antitrust scrutiny.

Just over a month after this large-scale recruitment, news of the MAI-1 large model emerged. The "M" may refer to "Microsoft," and the "1" suggests that this may be the first version of its AI model.

MAI-1 is reported to have approximately 500 billion parameters, much larger than Microsoft's previous open-source models, requiring more computational power and training data. MAI-1 aims to be on par with OpenAI's GPT-4, which is rumored to have over 1 trillion parameters (in expert hybrid configurations), far surpassing Meta and Mistral's 700 billion parameter models and other smaller models.

Although there are rumors that MAI-1 may be built on the technology brought by former Inflection employees, two Microsoft employees familiar with the project revealed that MAI-1 is a new type of large language model (LLM).

02 Seeking to Break Free from OpenAI's Large Models, Microsoft's Multiple Investments and Self-Development

This also raises some questions, as Microsoft and OpenAI's strong collaboration in recent years has led to the early application of advanced large models in many of Microsoft's product lines, achieving intelligent upgrades.

Going back to 2019, Microsoft invested $1 billion in OpenAI, which had just parted ways with Elon Musk, to jointly develop new Azure AI supercomputing technology. At the same time, OpenAI also shifted its development services to run on Microsoft Azure.

In 2021, Microsoft continued to invest $2 billion in OpenAI. By this time, the collaboration between the two companies had already shown initial success, with the most well-known product being GitHub Copilot, which is based on the source code from GitHub and other websites, utilizing OpenAI's technology to automatically write code for programmers based on the provided prompts.

In January 2023, Microsoft and OpenAI upgraded their collaboration, with OpenAI receiving "multi-year, multi-billion-dollar investment" from Microsoft. Although the specific amount of the investment was not officially disclosed by both parties, The Verge revealed that the investment amount reached as high as $10 billion, as part of this investment:

Microsoft has the right to receive 75% of OpenAI's profits until it recoups the $10 billion investment, as well as an additional $3 billion from Microsoft's early investment in the company. After OpenAI's profits reach $92 billion, Microsoft's share of OpenAI's profits will decrease to 49%. Finally, when OpenAI's profits reach $150 billion, OpenAI will regain 100% ownership of the company.

During this period, Microsoft also gained priority access to OpenAI's software, allowing it to resell OpenAI's software along with its AI products using OpenAI's technology to its Azure customers.

Over the past year, we have also witnessed the innovation of Microsoft's AI-driven products, such as Bing search engine using the GPT-4 large model, Microsoft's AI image generation tool Copilot Designer delivering impressive results driven by OpenAI's DALL-E; the arrival of the intelligent co-pilot Microsoft 365 Copilot has brought a new dawn for "office workers," breaking the traditional way of office software by combining large models (LLM) with data from Microsoft Graph and Microsoft 365 applications to automatically generate documents, emails, and PowerPoint presentations, significantly improving the efficiency of Office software such as Word, Excel, PowerPoint, Outlook, and Teams...

The deep collaboration between the two companies, on the one hand, Microsoft announced its partnership with Meta when Llama 2 was released last year, becoming the preferred partner for Llama 2 in the Azure AI model catalog. Based on this, developers using Microsoft Azure can use Llama 2 for building and utilize its cloud-native tools for content filtering and security features.

At the same time, Llama 2 has been optimized to run locally on Windows, providing a seamless workflow for developers and delivering a generative AI experience across different platforms. Llama 2 can also be accessed through Amazon Web Services (AWS), Hugging Face, and other providers.

On the other hand, Microsoft has placed its bet on the French startup Mistral AI this year. Similar to when OpenAI received funding and computational support from Microsoft in its difficult times, Microsoft President Brad Smith announced the establishment of a "multi-year partnership" with Mistral to help the 10-month-old company bring its AI models to market. As part of the collaboration, Microsoft holds a small stake in Mistral AI. As a result, Mistral AI enjoys the same treatment as OpenAI, directly selling its model resources on the Microsoft Azure cloud platform, becoming the second company to offer commercial AI models on Microsoft Azure.

03

There is no doubt that from the exclusive cooperation with OpenAI to further collaboration with open-source large models and AI startups, Microsoft has now embarked on a path of independent research and development, seemingly building its own AI technology barrier step by step.

Some speculate that the emergence of MAI-1 also subtly indicates that Microsoft internally focuses on both small language models for local operation on mobile devices and large state-of-the-art models supported by the cloud. At the same time, this new development suggests that Microsoft may be striving to break away from models influenced by OpenAI, even possibly designed as a direct competitor to ChatGPT. Perhaps Microsoft will continue to use ChatGPT in tools like Copilot while creating another model for other niche purposes (such as business management).

However, according to The Information, the exact purpose of MAI-1 has not yet been determined (even within Microsoft), and its most ideal use will depend on its performance. To train this model, Microsoft has been allocating a large number of servers equipped with NVIDIA GPUs and compiling training data from various sources, including texts generated by OpenAI's GPT-4 and public internet data.

It is reported that Microsoft may unveil a preview version of MAI-1 at the Build developer conference later this month, and we will observe how it evolves.

Finally, in response to the rumors of Microsoft developing MAI-1, many netizens have marveled at Microsoft's strategic layout, stating:

"It is shocking that Microsoft has invested so much effort in large language models. They have:

- The MAI team mentioned earlier

- Bing AI team (which is based on GPT-4, but it is a separate model)

- Small model Phi-3 team

- WizardLM team

- Partnership with OpenAI

- Another partnership with Mistral

- $100 billion investment in the new AI data center

Some question why, "if OpenAI is so outstanding and MSFT owns 49% of it, why does MSFT need to do this? Why not use all computations to fine-tune models for specific products?"

In response, some netizens claim, "Because 49% is not 100%. Clearly, every large tech company wants to control its own LLM."

Of course, facing the lively competition in the field of LLM, we are more looking forward to the disruptions and innovations brought by LLM.

So, whether MAI-1 is true or false, we will also pay further attention to Microsoft's Build 2024 event on May 21.

Reference:

https://news.ycombinator.com/item?id=40279350

https://arstechnica.com/information-technology/2024/05/microsoft-developing-mai-1-language-model-that-may-compete-with-openai-report/