Simplifying AI Model Deployment with Models-as-a-Service
Developing AI-powered applications has become more accessible with the availability of various tools and models. However, one aspect that developers often struggle with is deploying these models. Microsoft aims to address this challenge with its Models-as-a-Service (MaaS) solution, which simplifies the process of hosting AI models. MaaS is akin to cloud services, charging for access rather than infrastructure. By offering MaaS through its AI Azure Studio product, Microsoft allows developers to focus more on the creative process rather than the technicalities of model deployment.
Abstracting the Complexity
According to Seth Juarez, the principal program manager for Microsoft’s AI platform, deploying a model involves a series of complex steps and configurations. With Models-as-a-Service, Microsoft abstracts away this complexity. Developers can choose from a catalog of models, including open source and those built by organizations like OpenAI. By simply clicking a button, developers can have an endpoint to use the chosen model. This eliminates the need for virtual machines and streamlines the deployment process.
Renting Inference APIs and Fine-Tuning
Developers can rent inference APIs and perform fine-tuning through a convenient pay-as-you-go plan. Microsoft offers over 1,600 models that serve various purposes, making it easier for developers to leverage AI functionality in their software. The aim is to provide developers with seamless access to AI capabilities without the need for extensive technical knowledge or infrastructure management.
Expanding the Catalog
Since its inception in 2023, Microsoft has been gradually adding models to its Models-as-a-Service program. Initially, Mistral-7B and Meta’s Llama 2 were available. Recently, TimeGen-1 from Nixtila and Core42 JAIS were added, with more models from AI21, Bria AI, Gretel Labs, NTT Data, Stability AI, and Cohere coming soon. While this is only a small fraction of what’s available on AI Azure Studio, the selection process for MaaS models varies. Some models are included through company partnerships, while others require API work to ensure uniform function signatures. Specialized models may need to be deployed in alternative ways.
The Future of Model Ownership
Juarez envisions a future where developers can choose between owning and renting models, similar to being a homeowner or renter. In the “Models-as-a-Service” model, Microsoft takes care of the upkeep, allowing developers to focus solely on utilizing the models. However, there will always be developers with specific requirements who prefer to deploy their models in their own containers using managed inference. Microsoft aims to cater to both approaches, providing flexibility for different use cases.
The Demands of AI
The rise of AI as a prominent technology in replicating the cloud computing business is due to a shift in demand. Juarez explains that users are now demanding specific features and services from tech companies, rather than companies pushing out technologies they think users need. This shift is fueled by the close alignment between AI research and commercialization. Consumers are driving the demand for AI experiences and enterprises are playing catch-up to meet these demands.
In conclusion, Microsoft’s Models-as-a-Service offers a streamlined solution for deploying AI models, eliminating the complexities and technicalities that developers often face. By providing access to a wide range of models and handling the infrastructure upkeep, Microsoft enables developers to focus on unleashing the creative potential of AI in their applications. This approach aligns with the growing demand for AI experiences and services, where users are actively shaping the technology landscape.