MAI

MAI stands for Microsoft AI. It refers to a series of computer programs that Microsoft has developed itself and that can write text or generate speech. Such programs are called models: they have learned from vast amounts of text which words typically follow one another, and use this to put together new sentences. So far, two representatives have become particularly well known. MAI-1 generates text and answers questions, while MAI-Voice reads text aloud in a natural-sounding voice. Both run within the Copilot chat assistant, which Microsoft offers in Windows, Word, and the Bing search engine.

Microsoft’s break from OpenAI

Over the years, Microsoft has invested many billions of dollars in OpenAI, the company behind ChatGPT. In return, the corporation was allowed to use OpenAI’s models in its own products. As a result, Copilot was for a long time essentially a relabeled OpenAI model at its core. That was convenient, but it made Microsoft dependent on a partner that is also a competitor.

With MAI, this situation is changing. Microsoft now owns its own models and can decide for itself how they behave, what they cost, and where they run. For a company that sells software to hundreds of millions of users, ongoing computing costs are an enormous expense. Smaller, proprietary models for simple tasks save a lot of money here.

On the stock market, MAI is therefore seen as a signal. Analysts read it as an indication of how much Microsoft will still depend on OpenAI in the future. Every announcement of a new MAI model is followed accordingly closely.

How the models are built

Behind MAI is a dedicated department led by Mustafa Suleyman, a co-founder of the AI lab DeepMind. The models are trained on large data centers equipped with thousands of specialized graphics chips. During training, the model is presented with texts and must guess the next word each time. If it gets it wrong, the internal numerical values are adjusted slightly. After billions of such rounds, a program emerges that can formulate fluent answers.

Microsoft emphasizes that MAI-1 was trained using comparatively few chips. While competitors sometimes use over a hundred thousand graphics processors, MAI-1-preview reportedly used around 15,000. The aim, then, is not to build the absolute strongest model in the world. The goal is a model with a good ratio of performance to cost.

MAI-Voice works on a different principle. Instead of predicting words, this model predicts how a sound signal continues. This produces a voice with intonation and pauses for breath. Microsoft states that one minute of audio is generated in under one second, on a single chip.

MAI in Copilot and in headlines

Most often, people encounter MAI without noticing it. Anyone who asks Copilot a question in Windows or Bing now sometimes receives an answer from an MAI model. Features such as Copilot Daily, a read-aloud news summary, also run on MAI-Voice. The name rarely appears on the surface, because Microsoft presents the Copilot brand to the outside world.

In business news, on the other hand, MAI is spelled out in full. There, the focus is on rankings comparing models, personnel changes in Suleyman’s team, or the relationship with OpenAI. A common misconception is that MAI is a replacement for ChatGPT. So far, Microsoft uses the models alongside one another and chooses depending on the task.

For context, it helps to compare it with other abbreviations. GPT belongs to OpenAI, Gemini belongs to Google, Llama belongs to Meta, Claude belongs to Anthropic. MAI is simply Microsoft’s entry on this list. Anyone who can keep the names apart will understand most headlines about the AI industry much faster.