Microsoft at the moment at its annual Construct convention launched a number of updates to Azure AI, the corporate’s cloud-based platform for constructing and operating AI functions. Azure AI competes with related choices from rival cloud suppliers reminiscent of AWS, Google, and IBM.
The updates embrace the addition of recent governance options, new massive language fashions (LLMs), and Azure AI Search enhancements. Microsoft additionally introduced that it’s making Azure AI Studio typically out there.
Azure AI Studio, a generative AI software growth toolkit that competes with the likes of Amazon Bedrock and Google Vertex AI Studio, was launched in a preview in November of final yr.
In distinction to Microsoft’s Copilot Studio providing, which is a low-code instrument for customizing chatbots, Azure AI Studio is aimed toward skilled builders, permitting them to decide on generative AI fashions and floor them with retrieval augmented era (RAG) utilizing vector embeddings, vector search, and their very own information sources.
Azure AI Studio may also be used to fine-tune fashions and create AI-powered copilots or brokers.
New fashions added to Azure AI
As a part of the updates to Azure AI, Microsoft is including new fashions to the mannequin catalog inside Azure AI Studio, bringing the variety of fashions out there to greater than 1,600.
The brand new fashions embrace OpenAI’s GPT-4o, showcased this week. Earlier in Might, Microsoft enabled GPT-4 Turbo with Imaginative and prescient via Azure OpenAI Service. “With these new fashions builders can construct apps with inputs and outputs that span throughout textual content, photos, and extra,” the corporate mentioned in an announcement.
Different fashions which were added by way of Azure AI’s Fashions-as-a-Service (MaaS) providing embrace TimeGen-1 from Nixtla and Core42 JAIS, which at the moment are out there in preview. Fashions from AI21, Bria AI, Gretel Labs, NTT Knowledge, Stability AI, and Cohere Rerank are anticipated to be added quickly, Microsoft mentioned.
Additional, Microsoft is updating its Phi-3 household of small language fashions (SLMs) with the addition of Phi-3-vision, a brand new multimodal mannequin that’s anticipated to grow to be out there in preview.
In April, Microsoft had launched three Phi-3 fashions—the three.8-billion-parameter Phi-3 Mini, the 7-billion-parameter Phi-3 Small, and the 14-billion-parameter Phi-3 Medium—to assist resource-constrained environments for on-device, edge, and offline inferencing and be less expensive for enterprises.
Microsoft’s Phi-3 builds on Phi-2, which might perceive 2.7 billion parameters whereas outperforming massive language fashions as much as 25 instances bigger, Microsoft mentioned on the time of the launch. Phi-3 Mini is at present typically out there as a part of Azure AI’s Fashions-as-a-Service providing.
Different modules of Azure AI have been additionally up to date, together with Azure AI Speech, which now consists of options reminiscent of speech analytics and common translation to assist builders construct functions to be used instances requiring audio enter and output. The brand new options can be found in preview.
Again in April, Microsoft had up to date its Azure AI Search service to extend storage capability and vector index measurement at no extra price, a transfer it mentioned will make it extra economical for enterprises to run generative AI-based functions.
Azure AI will get new governance, security options
At Construct 2024 Microsoft additionally launched new governance and security options for Azure AI, with the corporate updating its mannequin output monitoring system, Azure AI Content material Security.
The brand new characteristic, named Customized Classes, is at present in preview and can enable builders to create customized filters for particular content material filtering wants. “This new characteristic additionally features a fast choice, enabling you to deploy new customized filters inside an hour to guard in opposition to rising threats and incidents,” Microsoft mentioned.
Different governance options added to Azure AI Studio and Azure OpenAI Service embrace Immediate Shields and Groundedness Detection, each of that are in preview.
Whereas Immediate Shields mitigate each oblique and jailbreak immediate injection assaults on LLMs, Groundedness Detection checks generative AI functions for ungrounded outputs or hallucinations in generated responses.
Microsoft mentioned that it at present has 20 accountable AI instruments with greater than 90 options throughout its choices and companies.
To be able to safe generative AI functions, Microsoft mentioned that it was integrating Microsoft Defender for Cloud throughout all of its AI companies. “Risk safety for AI workloads in Defender for Cloud leverages a local integration with Azure AI Content material Security to allow safety groups to watch their Azure OpenAl functions for direct and in-direct immediate injection assaults, delicate information leaks, and different threats to allow them to rapidly examine and reply,” the corporate mentioned.
To be able to bear down on safety additional, enterprise builders also can combine Microsoft Purview into their developed functions and copilots with the assistance of APIs, in keeping with Jessica Hawk, company vp of information, AI, and digital functions at Microsoft.
This may assist builders and copilot prospects to find information dangers in AI interactions, shield delicate information with encryption, and govern AI actions, Hawk added.
These capabilities can be found for Copilot Studio in public preview and will likely be out there in public preview for Azure AI Studio in July by way of the Purview SDK.
Different safety updates embrace integration of what Microsoft calls “hidden layers safety scanning” into Azure AI Studio to scan each mannequin for malware.
One other characteristic, referred to as Facial Liveness, has been added to the Azure AI Imaginative and prescient Face API. “Home windows Hi there for Enterprise makes use of Facial Liveness as a key aspect in multi-factor authentication (MFA) to forestall spoofing assaults,” Hawk defined.
Copyright © 2024 IDG Communications, .
