Azure Weekly

Issue #482 - 1st September 2024

To kick off this edition, here are a few AI highlights which seem to follow a theme on security. Architecting secure Gen AI applications: Preventing Indirect Prompt Injection Attacks, and Securing your AI Apps on Azure: Recordings and Slides, and Secure APIM and Azure OpenAI with managed identity, followed by a guide highlighting how to Fine Tune GPT-4o on Azure OpenAI Service.

It's the end of the month, so there's another post covering the Microsoft Fabric August 2024 Update. Barry Smart publishes part 3 of his Microsoft Fabric Machine Learning Tutorial - Part 3 - Testing Notebooks, and Jean Hayes covers Microsoft Fabric Metadata Driven Pipelines with Mirrored Databases.

In App Dev, there are three useful guides: Migrate App Insights Instrumentation Key to Connection String in Simple Steps, and Deploy a container to Azure App Services using Azure CLI and user-assigned managed identity, and Using AZD for faster incremental Azure Static Web App deployments in GitHub Actions.

When I went to University in 1997, I started an Intelligent Systems Degree, which covered AI during the last "hype cycle". While I have a healthy scepticism of the current Generative AI hype cycle, there is a deep parallel with the Space Race of the 1960's; global superpowers spent vast percentages of their GDP on getting into space, and then to the Moon. While only 12 people have walked on the Moon, the whole of humanity benefited from "space-age materials" - the innovations required to deliver on the mission, but whose ancillary uses were myriad, impactful and created new opportunities for existing industries. AI vendors are pouring billions into AI research, languages, models, frameworks, and hardware; with the latter we're seeing improvements in GPUs, the creation of TPUs / NPUs and now custom chipsets: Inside Maia 100: Revolutionizing AI Workloads with Microsoft's Custom AI Accelerator. Like "space-age materials", hopefully these investments will deliver similar benefits to other traditional complex data and computationally expensive problems.