Microsoft announces advancements in Phi-3 fine-tuning, novel generative AI designs, and innovative Azure AI enhancements, empowering businesses to tailor and amplify AI applications.

We are thrilled to unveil a range of enhancements designed to empower builders to swiftly craft bespoke AI solutions, characterized by increased versatility and scalability, via the Azure AI suite.

Artificial intelligence is revolutionizing every industry by generating novel opportunities for growth, transformation, and progress. While scaling AI applications demands a robust and adaptable infrastructure capable of addressing the diverse needs of modern organizations, allowing them to develop solutions rooted in their institutional expertise. We’re thrilled to unveil a series of enhancements designed to empower builders to swiftly craft bespoke AI solutions that offer greater flexibility and versatility through our robust toolchain.

• Allows builders to quickly and easily tailor designs for cloud and edge scenarios without requiring reconfiguration of compute resources.
• Together with this technology, it empowers builders to construct even more efficiently using a more performant model without incurring additional costs.
• To leverage Azure AI’s capabilities and provide customers with a more extensive range of options and flexibility.

Unleashing value through revolutionary mannequin innovations and personalized designs.  

In April, we successfully launched a series of innovative, open-source fashion designs, courtesy of Microsoft’s creative endeavors. Our Phi-3 fashion is a standout among small language models (SLMs), boasting exceptional performance and value for its size – in fact, it consistently outperforms similar-sized and larger models alike. As enterprises seek to customize AI capabilities to meet specific needs and elevate response quality, fine-tuning a smaller model emerges as a compelling alternative without compromising on performance. Starting now, developers can leverage their expertise to build AI-driven experiences that are more closely tied to their customers’ needs, while ensuring safety and efficiency throughout the process.

Considering their modest computational requirements, seamless cloud and edge integration, Phi-3 models are ideally positioned for refining base model performance across diverse scenarios, such as learning a new skill or occupation – for instance, Tutoring services for students seeking academic support, or enhancing consistency and high-quality responses to complex inquiries through collaborative learning strategies. tone or type of responses in chat/Q&A). The versatility of Phi-3 is being leveraged for innovative applications.

Microsoft and Khan Academy join forces to expand educational opportunities globally for learners of all ages? As part of its collaborative efforts, Khan Academy leverages the power of Khanmigo for Lecturers, a pilot AI-powered teaching assistant serving educators across 44 countries, while also exploring the potential of Phi-3 to elevate math tutoring capabilities. A recent study by Khan Academy published an analysis paper that explored the disparities in AI models’ performance in evaluating mathematical accuracy during tutoring scenarios, including benchmarks from a fine-tuned Phi-3 model. Discovering that when a student commits a mathematical mistake, Phi-3 surpasses the majority of prevailing main generative AI models in accurately identifying and rectifying pupil errors.

We’ve refined Phi-3 to optimize its performance within the device as well. In June, we successfully launched our innovative platform designed to empower builders by providing a robust and dependable framework for developing apps that integrate secure and trustworthy AI-powered experiences. Phi Silica leverages the esteemed Phi family of fashion styles, specifically tailored for Non-Procedural Units (NPUs) within Copilot+ PCs. Microsoft introduces its cutting-edge Small Language Model (SLM), custom-built for its Neural Processing Unit (NPU) and seamlessly integrated into the email inbox.

You may currently attempt to fine-tune Phi-3 models.

We’re thrilled to announce that our Fashion-as-a-Service (serverless endpoint) capability within Azure AI is now generally available. Furthermore, this allows builders to quickly and seamlessly initiate AI development without needing to manage complex underlying infrastructure.

Phi-3 Vision, a multi-modal mannequin within the Phi-3 household, made its debut at Microsoft Build and is now available through the Azure AI model catalog. It will soon be available through a seamless serverless endpoint as well. The Phi-3-small AI model, available in two configurations of 128K and 8K, excels at processing data with a modest 7 billion parameters. Meanwhile, the more advanced Phi-3-vision model boasts an impressive 4.2 billion parameters, optimized for comprehending charts and diagrams to generate insightful answers and respond to queries.

The phi-three project is receiving a positive response from the team. We recently launched a significant upgrade that significantly enhances our product’s core quality and instructional integrity. After retraining the mannequin, the model demonstrated a significant improvement in its ability to follow instructions accurately and provide structured support for optimal results. We additionally improved multi-turn dialog high quality, launched assist for <|system|> prompts, and considerably improved reasoning functionality.

The table below showcases improvements in instructional compliance, organized presentation, and logical thought processes.

We continue to improve upon the Phi-3 security framework, implementing additional measures to further strengthen its overall effectiveness. Microsoft’s iterative process for enhancing the security of Phi-3 products entailed multiple cycles of testing and refinement, red-teaming, and vulnerability discovery, ultimately yielding significant improvements in protection. This method significantly reduced harmful content by 75%, thereby substantially enhancing the fashion models’ performance on responsible AI metrics. 

Can the world’s largest fashion retailer unlock new possibilities by leveraging AI? With more than 1,600 customizable styles available in Microsoft Azure AI, the answer is yes.

With Azure AI, we’re committed to delivering a comprehensive portfolio of cutting-edge, open-source innovations and advanced tooling to help address customers’ diverse needs for scalability, speed, and design excellence.

Last year, we launched the Azure AI model catalog, featuring the broadest collection of models with over 1,600 models from esteemed suppliers, including AI21, Cohere, Databricks, Hugging Face, Meta, Mistral, Microsoft Research, OpenAI, Snowflake, and others.

This month, we introduced three advanced language models: OpenAI’s GPT-4, Meta Llama 3.1 (405B), and Mistral Massive 2, all accessible via Azure OpenAI Service.

With the initial success still building momentum, we’re thrilled to announce that Cohere Rerank is now available on Azure, marking a significant milestone in our journey. By leveraging Cohere’s advanced language models on Azure AI’s robust infrastructure, businesses can effortlessly integrate cutting-edge semantic search capabilities into their applications, ensuring seamless, reliable, and secure operations. This integration enables customers to harness the flexibility and scalability of Azure, combined with Cohere’s highly performant and environmentally sustainable language models, to deliver exceptional search results in manufacturing, ultimately empowering businesses to make data-driven decisions and drive innovation.

TDBank, a leading financial institution group in North America, has recently partnered with Cohere to leverage the entire range of massive language models (LLMs) and Cohere Rerank for its operations.

As a leading financial institution, TD is leveraging the power of AI to deliver highly personalized and intuitive experiences for customers, colleagues, and communities. We’re thrilled to collaborate with Cohere, exploring how its language models can be deployed on Microsoft Azure to accelerate our innovation journey.

Kirsti Racine, Vice President and Lead of Artificial Intelligence Expertise at TD.

AtomicWork, a digital office expertise platform and seasoned Microsoft Azure customer, has significantly bolstered its IT service management platform through the integration of Cohere Rerank technology. By seamlessly integrating the mannequin into its AI-powered digital assistant, Atom AI, Atomicwork has significantly enhanced search accuracy and relevance, providing rapid, more precise answers to complex IT support inquiries. The seamless integration has significantly enhanced IT efficiency and amplified productivity across the entire organization. 

The driving force behind Atomicwork’s digital office capabilities lies in the synergy between Cohere’s Rerank model and Microsoft Azure AI Studio, enabling Atom AI, our cutting-edge digital assistant, to deliver precise and efficient results that drive tangible outcomes. This strategic partnership underscores our commitment to delivering exceptional, secure, and reliable enterprise AI solutions.

Vijay Rayapati, CEO of Atomicwork

Cohere’s flagship generative model, Command R+, optimizes performance when integrated with Cohere Rerank within a Retrieval-Augmented Generation (RAG) framework, exclusively available on the Azure AI platform. Together, they are capable of supporting some of the most exacting enterprise workloads found in manufacturing settings. 

Earlier this week, we announced the availability of Meta Llama 3.1-405B alongside its newest fine-tuned variants, including 8B and 70B, which can now be accessed through a serverless endpoint on Azure AI. The Llama 3.1 model, designated as 405B, is capable of facilitating the creation of superior artificial knowledge and distillation, with its 405B-Instruct variant serving as a trainer prototype and the 8B-Instruct and 70B-Instruct models functioning as pupil prototypes. .

Azure has successfully launched Mistral Massive 2, solidifying its position as the go-to primary cloud provider for this cutting-edge model. Mistral Massive 2 surpasses its predecessors in terms of coding efficiency, logical reasoning, and autonomous decision-making capabilities, rivalling the performance of other prominent styles. What’s more, Mistral Nemo – a cutting-edge AI model developed jointly with NVIDIA – boasts an impressive 12 billion parameter count, further propelling the frontiers of natural language comprehension and era-specific insight. .

During the final week, Microsoft made significant updates to its Azure OpenAI Service, allowing developers to create a wider range of AI applications at a lower cost and latency while enhancing security and data deployment options. We will soon be announcing additional features for our GPT-4o Mini model. We’re thrilled to introduce a new feature to Microsoft Teams.  

Safe and Responsible Innovation with Artificial Intelligence  

Constructing AI options responsibly lies at the core of AI improvement within Microsoft. Our team offers a robust suite of solutions designed to help organisations effectively manage risks associated with AI development throughout its entire lifecycle, covering both traditional machine learning and generative AI applications, enabling them to measure, mitigate, and respond to potential threats in real-time. Azure AI evaluations enable developers to continuously evaluate the quality and security of models and deployments by leveraging built-in and customizable metrics, thereby informing mitigation strategies. Further options, including instant shielding and protected material detection, are enabled by default in the Azure OpenAI Service. The versatility of these capabilities enables them to be utilized as content filtering mechanisms, seamlessly integrated with any model from our extensive catalog, including Phi-3, Llama, and Mistral. Furthermore, developers can effortlessly merge these functionalities into their application via a unified API interface. In the realm of manufacturing, builders can leverage cutting-edge technology to ensure high-quality production, robust security, and uncompromising data integrity, thereby enabling swift and effective interventions through the utilization of real-time alerts.

Azure AI leverages sophisticated algorithms to proactively monitor third-party and open-source software libraries, identifying potential security risks such as emerging vulnerabilities, malware, and other suspicious activity, prior to integrating them into its proprietary model repository. The verification results from the Mannequin Scanner, embedded in each model card, will provide development teams with greater assurance as they select, refine, and deploy open-source models for their applications. 

As we venture into the Azure AI ecosystem, we strive to seamlessly integrate cutting-edge technology with our customers’ needs, empowering them to build, deploy, and scale their AI solutions with unwavering reliability and trust. What we eagerly anticipate witnessing next will surely be impressive.

Azure AI, a cloud-based artificial intelligence platform, continues to evolve and expand its capabilities.

• Explore the latest Azure AI model catalog to gain in-depth knowledge and stay ahead of the curve in machine learning innovation.
• Meet Sebastien Bubeck, a leading researcher at Microsoft.