Microsoft president Brad Smith has posted a blog that lays out how the company plans to support software developers building artificial intelligence (AI). Among the areas he covered included providing open application programming interfaces (APIs) to foundational AI models hosted on Azure and a commitment to only use public data when training Microsoft’s AI models.
The overall goal is to maintain Microsoft Azure as an open platform, for distributing and deploying AI software.
“We are committed to enabling AI innovation and fostering competition by making our cloud computing and AI infrastructure, platforms, tools, and services broadly available and accessible to software developers around the world,” said Smith, adding that Microsoft’s goal is to make Azure the best place for developers to train, build and deploy AI models, and use those models safely and securely in applications.
Among the company’s plans going forward is to expand its AI acceleration hardware capabilities on Azure to enable the training and deployment of more foundation models, both proprietary and open source.
He announced French AI startup Mistral AI will use Microsoft’s AI infrastructure for its premium models available to customers through Microsoft’s models-as-a-service (MaaS) offering on Microsoft Azure. Mistral AI claims its foundational model is the top-performing model for maths and coding, and is second to GPT-4 in benchmarks that measure common sense and reasoning.
MaaS enables model developers to publish and monetise their AI models. “By providing a unified platform for AI model management, we aim to lower the barriers and costs of AI model development around the world for both open source and proprietary development,” said Smith.
Microsoft hosts 1,600 open source and proprietary models from companies and organisations such as Meta, Nvidia, Deci and Hugging Face. Models from Cohere and G42 are also planned. “We are committed to expanding this type of support for additional models in the months and years ahead,” he added. “We’re committed to ongoing and innovative steps to make the AI models we host and the development tools we create broadly available to AI software applications developers around the world in ways that are consistent with responsible AI principles.”
Microsoft also wants to support developers building customised AI software by enabling them to fine-tune existing models based on their own unique data sets and for their specific needs and scenarios. “With Azure Machine Learning, developers can easily access state-of-the-art, pre-trained models and customise them with their own data and parameters,” said Smith.
He claimed the approach Microsoft has taken rapidly democratises AI, and helps to foster a culture of innovation and collaboration among developers. “We are committed to supporting the AI developer community by making our AI tools and resources available on GitHub, giving developers access to the latest innovations and best practices in AI development, as well as the opportunity to collaborate with other developers and contribute to the open source community,” said Smith.
He said Microsoft would ensure the AI models the company hosts on Azure, including the Microsoft Azure OpenAI API service, are all accessible via public APIs, adding that Microsoft publishes documentation on its website to show developers how they can use these APIs and underlying models. This, said Smith, enables any application, whether it’s built and deployed on Azure or other private and public clouds, to use these APIs and access the underlying models.
From an intellectual property and data protection perspective, he said: “We respect the needs of developers by ensuring we do not use any non-public information or data from the training, building, deployment or use of developers’ AI models in Microsoft Azure to compete with those models.”
Smith stressed that trust is central to the success of Microsoft Azure. “We build this trust by serving the interests of AI developers and customers who choose Microsoft Azure to train, build and deploy foundation models,” he said. “In practice, this also means that we avoid using any non-public information or data from the training, building, deployment or use of developers’ AI models to compete against them.”