AI was the very first thing mentioned in Satya Nadella’s opening keynote of Microsoft Ignite 2023, dominating over 50% of the sessions, and with Gen-AI at its peak, that has been no surprise. In this blog, we will highlight some of our favourite announcements on Azure AI Platforms.
At last, the public preview of Azure AI Studio has been announced.
Azure AI Studio is a game changer for Gen-AI development in Azure, allowing users to quickly explore, build, and deploy models at scale. The platform integrates pre-built services, content safety features, and responsible AI tools, streamlining the complexities of generative AI with a focus on privacy, security, and compliance. While still not recommended for production workloads, this is still a big step forward for practical, usable Gen-AI.
This isn’t the first we have heard from Azure AI Studio but it is the first look we’ve had of the brand new features include more affordable pricing, structure JSON formats and extended prompts. Other prominent updates include Prompt Flow and MaaS – which deserved their own sections!
Prompt flow is now generally available in Azure Machine Learning (AML) and in public preview in Azure AI Studio.
AML prompt flow is designed for enterprise LLMOps by offering support for version control, collaboration using any source code management tool, connection to diverse foundation models, vector databases, and comprehensive prompt management. Since its initial announcement in May, Prompt Flow has undergone significant enhancements, adding numerous features based on community feedback.
Prompt flow will be an essential piece of Gen-AI process in Azure by faciliting easy debugging, sharing, and iteration of flows through seamless team collaboration. Additionally, users can create prompt variants and assess their performance through large-scale testing. The capability to deploy real-time endpoints is a standout feature, unlocking the full potential of LLMs for application use.
The new Azure AI model catalog will be a hub for discovering essential foundational models, now including 40 new models across a range of modalities. This new Model-as-a-Service offering will include NVIDIA Nemotron, Stable Diffusion models and Code Llama as well as Jais, a 13 billion parameter model for the Arabic language; Mistral, an LLM with 7.3 billion parameters, known for it’s fast inference features and Phi models.
The most powerful feature of the model catalog is the ease which users can either fine tune their foundational models of choice or host them for inference. MaaS streamlines the process for developers, particularly during the development and testing phase, by eliminating the necessity for dedicated VMs to host models. With cost-effective, token-based billing for inference APIs, the catalog provides an appealing and straightforward starting point for generative AI projects. The model catalog is anticipated to be generally available in the near future.. The model catalog is due to be generally available soon.
Ignite 2023 announced that the model benchmarks feature in Azure Machine and Azure AI studio is now in public preview.
This new feature plays a crucial role in evaluating and comparing the performance of different foundational models, simplifying the model selection process. By providing quality metrics, users can make informed decisions tailored to their project requirements, optimizing the performance of their AI solutions.
Model benchmarks enable users to compare models based on metrics such as accuracy, empowering them to make data-driven decisions in selecting the most suitable model for their specific task. This approach ensures that AI solutions are optimized for peak performance.
Previously, assessing model quality demanded significant time and resources. However, with the prebuilt metrics in model benchmarks, users can swiftly identify the most fitting model for their project, reducing development time and minimizing infrastructure costs. Within Azure AI Studio, users can access benchmark comparisons in the same environment where they build, train, and deploy their AI solutions, enhancing workflow efficiency and promoting collaboration among team members.
The general availability of managed feature store in Azure Machine Learning was also announced. Feature stores are essential in MLOps workflows facilitating feature reuse, faster experimentation, consistency and time travel. Azure Managed Feature Store will free Data Scientists and ML Engineers from the overhead of setting up and managing the underlying feature engineering pipelines. This integration across the machine learning lifecycle accelerates model experimentation, enhances model reliability, and reduces operational costs, significantly simplifying the MLOps experience.
Previously, Azure Machine Learning Batch endpoints have been able to take a machine learning model and deployed it for batch inference. With the latest update, these endpoints can deploy both models and Pipeline components. Pipeline components are an individual step of an AML pipeline that are reusable and movable across workspaces including across different environments.
This new deployment option now gives Data Scientists the capability of creating a component out of an entire pipeline and hence given them the versatility any component has facilitating nested pipelines and more complicated deployments.
Last year, at Ignite 2022, Microsoft Fabric garnered significant attention, and its impact continues to strongly resonate this year. For professionals like Data Scientists and AI practitioners, the introduction of OneLake's integration with Azure Machine Learning and Azure AI Studio proves to be a huge development in constructing Azure AI with Fabric workloads. A key element of Microsoft Fabric, OneLake streamlines data management by consolidating data into a single copy, effectively dismantling data silos. This fresh integration with Azure AI tools represents a pivotal leap forward in effectively managing and processing intricate AI and ML datasets.
Azure AI Video Indexer introduces it’s new Bring-Your-Own-AI feature to merge Ai models with your VI AI pipeline. You can now use models from Hugging Face, Azure Florence or any other model and connect it to VI insights to perform licence plate detection, tattoo detection, special signs in video, uniforms, masks and more.
Not to mention the many, many AI announcements from the event:
New AI Chips - Maia and Cobalt: Meeting the demand for efficient, scalable compute power.
Text-to-Speech Avatar Creation: A new service in AI Speech Studio
AI Safety Features: Jailbreak detection and protected material detection in preview.
Microsoft Copilot Studio: Customize Copilot for Microsoft 365 and create your own standalone copilots.
"Azure Cognitive Search" becomes "Azure AI Search" and now includes vector search and semantic ranker.
Fabric in GA
GPT-4 V models