MAI-1: Microsoft’s new AI to compete with OpenAI

Microsoft is reportedly working on its own large language model (LLM) called MAI-1, which could rival the likes of GPT-4 and Claude. This comes as the tech giant continues to invest heavily in AI, following its multi-billion dollar investment in OpenAI.

According to The Information, Microsoft is training the new AI model under the leadership of Mustafa Sulayman. The co-founder of DeepMind and new head of Microsoft AI is reportedly in charge of a complex project that departs from previous experiments. The model, known as MAI-1, will have 500 billion parameters (half that of GPT-4) and will be able to fine-tune during training.

MAI-1 is reportedly a proprietary Microsoft model being trained in-house

The company is said to be leveraging training data and technologies from Inflection AI, the company founded by Sulayman. After hiring the British researcher in March, Microsoft acquired all of the staff and agreed to pay $650 million to Inflection to license its AI software.

MAI-1 is expected to inherit training data and other technologies from Inflection AI to develop its own language model. This would not only reduce the company’s reliance on OpenAI but also allow it to work on AI to solve specific problems with its services.

Microsoft is going all-in on AI with MAI-1

It is no secret that Microsoft is betting big on AI. Its investment in OpenAI and the implementation of this technology in most of its services are the perfect example. Unlike other market segments – such as mobile – the Redmond company has known how to react in time.

Prior to the announcement of MAI-1, the company presented Phi-3 Mini, an AI that is capable of matching GP-3.5 at a fraction of its cost. Phi-3 Mini can handle around 3.8 billion parameters and was trained on children’s stories. According to Microsoft, Phi-3 Mini used the same learning mode that a child does when reading children’s games.

Microsoft wants to take the leap to the next level with the idea of making MAI-1 a profitable investment. The problem is that 500 billion parameters require an impressive amount of processing power. Like Meta, those in Redmond will need to talk to “Uncle Jensen” to get the necessary hardware.

For now, there are no details about MAI-1 and the rest of Microsoft’s AI projects. The company is expected to announce some news during a special event on May 20. While the stars of the event are the Surface with Snapdragon X Elite, the tech company will reveal its short-term AI strategy.

The initial collaboration between Microsoft and OpenAI was a milestone that allowed the tech giant to integrate the advanced capabilities of GPT-4 into a variety of its products and services. This alliance made it easier for OpenAI to obtain the financial and technological support needed to advance its AI research and development.

MAI-1: Microsoft’s bet on technological independence

Under the leadership of Mustafa Suleyman, who brings his previous experience at Google and as CEO of Inflection, Microsoft is developing MAI-1. This model promises to be a tool of enormous potential with 500 billion parameters, which indicates a commitment to exceeding the current limits imposed by previous models such as those of Meta and Google and, of course, those of Microsoft’s own Phi-3 Mini.

The “500 billion parameters” in a language model like Microsoft’s MAI-1 refer to the number of trainable weights the model has in its neural network. In terms of AI, a parameter is a component of the neural network that is adjusted during the machine learning process. Each parameter stores an amount of information learned from the training dataset, and its adjustment is essential for the model to make predictions or generate text effectively.

Potential of a model with 500 billion parameters

• Greater understanding and context: A higher number of parameters allows the model to have a better capacity to understand and process natural language, resulting in more accurate responses and analysis. This is crucial for complex tasks such as answering questions, summarizing texts, translating languages, or generating coherent and relevant content.
• Versatility in applications: A model with so many parameters can be effective in a wide range of applications, from intelligent personal assistants and customer support tools to advanced data analysis systems and automatic content generation.
• Learning language nuances: With so many parameters, the model can capture nuances in language that smaller models might miss. This includes ironies, metaphors, and other forms of figurative language, making it more effective in interactions that require a high level of linguistic understanding.
• Personalization: The ability of a model of this magnitude to learn from large volumes of data allows it to personalize responses for individual users based on language patterns and personal preferences.

• Improvements in accuracy and reduction of bias: Although a larger model doesn’t automatically guarantee less bias, the ability to train with a vast and diverse dataset can help improve accuracy and reduce bias, as long as the training is managed carefully.

Suleyman’s selection to lead this project is part of a broader Microsoft strategy to strengthen its position in the AI market. By integrating Suleyman and the Inflection team, Microsoft gains not only experience but also intellectual property, crucial for developing new technologies.

MAI-1 is a model under development, and its exact applications and final capabilities are yet to be defined. Depending on its results in preliminary testing, Microsoft will adjust its implementation strategy. The model is expected to be one of the central themes at the upcoming BUILD 2024 event, where more technical details and practical applications are anticipated to be revealed.

The impact of MAI-1 on the alliance with OpenAI

The development of MAI-1 signals a possible shift in the relationship between Microsoft and OpenAI. While the collaboration has been fruitful, the creation of a proprietary model by Microsoft could potentially redefine the terms of their future cooperation. This reflects a two-pronged approach where Microsoft seeks to secure its leadership and autonomy in creating AI solutions.

The question of hardware remains a hurdle. While the specifics of MAI-1’s capabilities are unknown, its 500 billion parameters suggest a significant need for processing power. Like Meta, Microsoft may need to partner with a company like Nvidia (“Uncle Jensen” being a playful reference to Nvidia’s CEO Jensen Huang) to acquire the necessary hardware infrastructure to train and run MAI-1 effectively.

Microsoft’s AI strategy: A multi-pronged approach

The development of MAI-1 alongside continued investment in OpenAI suggests a multi-pronged approach by Microsoft. Here’s a breakdown of the potential benefits:

• Reduced reliance on external solutions: MAI-1 offers Microsoft a path towards greater independence in the AI domain. It reduces reliance on OpenAI and allows Microsoft to tailor AI solutions to its specific needs.
• Diversity and innovation: Having both internal and external partnerships fosters a more diverse AI ecosystem for Microsoft. This can lead to faster innovation and a wider range of AI-powered solutions across its products and services.
• Competitive edge: A strong in-house AI capability like MAI-1 positions Microsoft favorably in the competitive AI landscape. It allows the company to compete more effectively with other tech giants like Google and Meta.

The Future of AI at Microsoft

The announcement of MAI-1 is a significant development for Microsoft and the AI industry as a whole. It signals Microsoft’s commitment to becoming a leader in AI development and its increasing focus on in-house solutions. With the upcoming BUILD 2024 event, we can expect to learn more about MAI-1’s functionalities and how it will integrate with Microsoft’s existing AI strategy. The coming years will likely see a continued race for AI dominance, with Microsoft’s multi-pronged approach positioning it as a major player in this ever-evolving field.