In recent updates from Microsoft’s Azure OpenAI Service, there’s been a significant emphasis on enhancing the safety, security, and responsibility in AI development and deployment. I truly think this topic will get hotter and hotter alongside generative AI adoption. Here’s a detailed exploration of the latest offerings and how they aim to bolster AI’s safe and responsible use.
Risks & Safety Monitoring in Azure OpenAI Studio
At Ignite 2023, Azure OpenAI Service announced new AI safety and Responsible AI features. These include jailbreak risk detection and protected material detection, enhancing the security of LLM deployments. The introduction of these features demonstrates Microsoft’s dedication to creating AI systems that are not only powerful but also secure and trustworthy. That was just the beginning of a journey. I’m sure we will get more new features at Build, in the mean time.
Azure OpenAI Studio has introduced a Risks & Safety dashboard for deployments utilizing a content filter configuration. This feature allows users to monitor the outcomes of content filtering activities closely. Through detailed reports, developers can observe trends over time in harmful request rates and adjust their content filter configurations accordingly. This proactive approach not only aligns with business needs but also upholds Responsible AI principles.
Enhanced Data Insights and User Behavior Analysis
Key to this update is the ability to gain insights into potentially abusive users, a feature that’s particularly groundbreaking. By analyzing request content and user behavior, Azure OpenAI Studio offers detailed reports on users whose actions have consistently resulted in blocked content. These insights enable more targeted actions to mitigate abuse and ensure the model’s responsible usage.
Connectivity and API Updates
Azure OpenAI’s new updates include the ability to connect to an Elasticsearch vector database for Azure OpenAI On Your Data. This connection, coupled with the introduction of a chunk size parameter during data ingestion, allows for more efficient data management. The general availability (GA) of the 2024-02-01 API adds support for the latest features, including Whisper, DALL-E 3, fine-tuning, and more, further enriching the Azure OpenAI offerings.
Regional Support and Model Deprecations
The updates also extend regional support for DALL-E 3, making it accessible in the East US and AustraliaEast Azure regions, alongside SwedenCentral. Moreover, Azure OpenAI Service now includes a dedicated page for tracking model deprecations and retirements, ensuring users are always informed about the availability and support status of various models.
Whisper and DALL-E 3 General Availability
The Whisper speech-to-text model and DALL-E 3 image generation model have reached general availability (GA) for REST and Python, marking a significant milestone in their development. These advancements highlight Microsoft’s commitment to continually expanding and refining its AI offerings, with client library SDKs still in public preview.
Looking Ahead: GPT-4 and API Lifecycle
In a noteworthy update, the planned upgrade of GPT-4 1106-Preview to GPT-4 0125-Preview was canceled, with upgrades now pending the release of a stable model version. This shift underscores the importance of staying informed about API version lifecycle guides to manage deployments effectively.
More references
- Risks & Safety monitoring in Azure OpenAI Studio and how to use it: [Microsoft Learn].
- 2. Introduction to Risks & Safety monitoring feature in Azure OpenAI Service: [Microsoft Community Hub]
- 3. Updates in Azure OpenAI Service, including the GA API release and Whisper & DALL-E 3 general availability: [Microsoft Learn]
- 4. Announcement of new AI Safety & Responsible AI features in Azure OpenAI Service at Ignite 2023: [Microsoft Community Hub]
Quae Nocent Docent 