Microsoft AI introduced its first in-house AI models last month, emphasizing its complex relationship with OpenAI. Mustafa Suleyman, head of Microsoft AI, announced the company’s plans to make “significant investments” in the computational resources necessary for developing future AI models internally.
During a recent town hall event for employees, Suleyman stated, “We should have the capacity to build world-class frontier models in-house of all sizes,” while also acknowledging the need to utilize external models when necessary. He mentioned that the initial model, MAI-1-preview, was trained using a relatively small cluster of 15,000 H100s, indicating a need for expansion in computational capacity.
Suleyman indicated that Microsoft aims to develop AI models that can compete with the efforts of Meta, Google, and xAI, suggesting that their upcoming clusters would be “six to ten times larger” than those used for MAI-1-preview. He noted, “Much more to do, but it’s good to take the first steps,” reflecting on the progress made thus far.
Microsoft’s CEO, Satya Nadella, also spoke at the town hall, expressing enthusiasm about enhancing model capabilities to create model-forward products. He affirmed that Microsoft would continue to support multiple models within its products and cited GitHub Copilot as an example of this approach.
Additionally, Microsoft plans to incorporate Anthropic’s AI models into certain features of Microsoft 365. Reports suggest that the Microsoft 365 Copilot will utilize Anthropic models in part after testing revealed that these models showed better performance than OpenAI’s in applications like Excel and PowerPoint.
Source: https://www.theverge.com/report/776853/microsoft-ai-training-capacity-investments-in-house-models