Symon Perriman

VoiceRAG: retrieval-augmented generation (RAG) apps with a real-time voice interface. The new GPT-4o realtime API for audio is now available in Azure OpenAI. This model allows developers to build applications with user interfaces like ChatGPT’s advanced voice mode, where users have a real-time, full-duplex audio channel with the model. If the model uses audio for input and output, how do you integrate a retrieval system to ground responses, along the lines of what typical RAG applications do? In this short write up we present a simple pattern for this we called VoiceRAG. The blog post includes a pointer to a repo with a sample implementation of the VoiceRAG pattern you can clone and try with our data or your own. https://aka.ms/voicerag

867

17 Comments

Happy to share a set of announcements relating to Gemini and VertexAI. -- Model Updates: * New versions of Gemini 1.5 Pro and Flash with improved math, long context understanding, and vision. * Gemini 1.5 Flash is nearly 2.5x faster than GPT-4o mini. * Gemini 1.5 Pro's 2M context window is now generally available. -- Pricing: * Gemini 1.5 Pro pricing reduced by 50%. -- Imagen 3: * Generally available for Vertex AI customers on the allowlist. * Improved prompt understanding, instruction-following, photorealistic quality, and control over text within images. -- Open AI Ecosystem: * Support for new models from Meta, Mistral AI, and AI21 Labs. -- Vertex AI Platform Enhancements: * Controlled Generation for fine-grained control over AI outputs. * Batch API for faster developer workflows. * Supervised fine tuning for Gemini 1.5 Flash and Pro. * Distillation techniques for smaller, specialized models. * Prompt Optimizer for automatic prompt optimization. * Prompt Management SDK for versioning, restoring, and generating prompt suggestions. * Grounding with Google Search and dynamic retrieval. * Multimodal function calling. * Gen AI Evaluation Service for defining custom evaluation criteria. * Expanded data residency commitments.

235

3 Comments

Sharing here the 3rd and final episode of the podcast (and accompanying write up) that Robert Lutton and I had the opportunity to do on navigating data transformation journeys. In this installment, we discuss the notion of adopting and practicing a "Progression over Perfection" approach to navigating the multi-dimensional change management facets of data journeys. Hope you find it useful!

190

3 Comments

Just before the holidays, we decided to try something new and hosted our first ever virtual conference. As a series A startup, when you start planning you never know how big of a pull you'll have for speakers or audience size when you start planning an event. After all, we definitely aren't at Google anymore. SmallCon turned out better than we could've hoped. Thousands of engineers from several hundred companies attended, and the speaker quality was incredible. Some of my highlights were: 1. "Building an LLM or Agentic application to get to 70% accuracy is easy. Getting above 90% is the interesting part" - Manjeet Singh speaking about the AgentForce platform. 2. "The hang-up I had [with fine-tuning] was that if every model required it's own piece of infrastructure you'd be spending tens of thousands to hundreds of thousands of dollars for every use case...What I saw with Predibase and the LoRAX serving architecture suddenly completed changed that paradigm" - Paul Beswick, Global CIO at Marsh McClenan 3. "I think we started by seeing this new movement as a threat. We had been in the business for a long time, and it felt like everyone could do what we were doing...When we started looking at Small Language Models we were really pessimistic and we set let's set the bar really high for these guys and see how it goes" (spoiler, it went well) - Mo Asady, CTO at Convirza All our recordings are now available online, if you're looking to procrastinate on coming back to the work week from the holidays I'd definitely recommend watching.

29

2 Comments

New Relic and NVIDIA released the first observability integration making it easy for companies to monitor the health and performance of their AI applications built with NVIDIA NIM.

5

With MongoDB, we help engineering leaders get off the modernization hamster wheel and embrace predictable, optimised spending that fuels their most ambitious initiatives. With powerful features like automated cluster sizing, on-demand pausing, and granular spend visibility, MongoDB Atlas helps you right-size your data infrastructure and take complete control of your IT costs. Perfect way to break free from the constraints of economic uncertainty and get easier budget approvals. This blog helps you understand how MongoDB Atlas can help you take better control of your data costs. #MongoDB #cloud #innovation

29

1 Comment

🥳🏆Announcing the Winners of "The Microsoft Fabric and AI Learning Hackathon!"🏆🥳 🌐 Learn all about the winning entries and honorable mentions => https://lnkd.in/gp4gWjYP 🌍 We've widened the scope even further for database and developer use cases within the product, making it more accessible and versatile than ever before! 🔍 Dive in, innovate, and stay tuned for the groundbreaking projects that emerged from this exciting event! #MicrosoftFabric #AIHackathon #Innovation #DataAnalytics 🚀📊🤖

44

4 Comments

New Episode: 268: Long Time Show Host is CloudPod’s first Casualty to AI (For This Week, at Least) https://bit.ly/4bURA59 Episode 268 of The Cloud Pod podcast, bringing you all the latest news from Cloud and AI with AWS, Azure, Google, OCI, OpenSSH, CloudFlare, and Lambda #aws #gcp #azure #cloud #cloudpodcast #podcast

5

1 Comment

Check out this excellent snippet, then head over to listen to the full episode. #SADA ##GeminiPro

2

It has been a week since Amazon Web Services (AWS) announced our AWS Generative AI Accelerator and applications have been flying in 🚀🚀🚀 I can't share how many startups have applied, but it definitely has been the most active and exciting of our accelerator launches 🤩 What is awesome about this accelerator is that it's truly global, so startups from anywhere around the global can apply 🌐 As long as your startup is building a tech or tech-enabled product, has a tech lead, a working product, and is still early stage (not raised a Series A), then you are eligible for the program ✍️ Learn more about the program and apply using the link in the post below and remember to submit your application before 11:59 PM ET July 19 📌 #startups #AWS #GenerativeAI #accelerator

9

1 Comment

Rocky Linux is now available for Amazon WorkSpaces! Customers can now choose Rocky Linux alongside other popular OS options in Amazon WorkSpaces. This addition empowers businesses to maintain their preferred Linux environment while leveraging the scalability and security of Amazon Web Services (AWS)'s cloud-based virtual desktops. #aws #euc #vdi #linux CIQ #rockylinux #daas #remotedesktop #amazonworkspaces https://lnkd.in/dU5MxrDk

49

2 Comments

"We just signed an agreement with Meta – for them to use Oracle's AI Cloud Infrastructure – and collaborate with Oracle on the development of AI Agents based on Meta's Llama models," CTO Larry Ellison https://lnkd.in/gMbPy8eK

29

🚀 Exciting News! Azure AI now features over 1,550 Hugging Face models, marking a significant milestone! 🎉 With these models, you can deploy on MaaP (Model as a Platform), which means you simply need to bring your own GPU-enabled VM. Azure handles the provisioning, so no extensive GPU infrastructure knowledge is required. Just select the VM size and apply for the necessary quota. You have many options for deployment: - Use AI Studio to deploy online endpoints manually. - Use Terraform or Azure Developer CLI for infrastructure as code - A CLI first approach with the Azure ML CLI - A code-first approach with the Azure ML Python SDK to deploy your endpoints with application logic Happy deploying! #AzureAI #HuggingFace #MachineLearning #AI #CloudComputing #PythonSDK #AIStudio

10

1 Comment

Just watched a fascinating interview with  Google Cloud CEO Thomas Kurian on CNBC, where he discussed the exciting ways Google Cloud is helping businesses monetize AI and how Google Cloud differentiates versus others. #AI, #GoogleCloud, #digitaltransformation, #business #futureoftech

18

Incredibly proud of the latest Kinesis Client Library version (3.0) which, when compared to previous KCL versions, reduces compute costs by up to 33%. KCL 3.0 introduces an enhanced load balancing algorithm that continuously monitors resource utilization of the stream processing workers and automatically redistributes the load from over-utilized workers to other underutilized workers. This ensures even CPU utilization across workers and removes the need to over-provision the stream processing compute workers which reduces cost. Additionally, KCL 3.0 is built with the AWS SDK for Java 2.x for improved performance and security features, fully removing the dependency on the AWS SDK for Java 1.x. #Kinesis #DataStreaming #KCL

94

Check out this great article by Kara Saucerman! It talks about CETO, a high-performance gateway serving as the front-end for Microsoft AI services. It leverages YARP (Yet Another Reverse Proxy) and the latest version of .NET.

9