Recent News in Analytics and AI: January 2025 Edition

This month's summary of recent news in analytics and AI covers DeepSeek's impressive performance and its major weakness, Microsoft Fabric's introduction of new features to ease some well known pain points in the platform, the arrival of a lighter variation of Microsoft 365 Copilot, getting surprisingly good results from data processing benchmarks run on mobile phones, and more. Each section presents three analytics and AI news highlights.

Read on and get up to speed.

Power BI

• The Tabular Model Definition Language (TMDL) has been well received, making it much easier to collaborate on Power BI semantic models. With the new addition of a TMDL view in Power BI Desktop, you can perform productivity-boosting actions like search-and-replace and multi-line edits. In addition, the TMDL view enables you to edit every property in a semantic model, something that was previously only possible through external tools. Learn more.

• Power BI semantic models have gained a new version history feature, and it's currently in preview. The behaviour is similar to the version history experiences in Microsoft Word, Excel, and other Office apps. It's aimed at self-service users who are unlikely to be familiar or comfortable with the more 'pro' experiences like Git integration. Learn more.

• It's easy to forget that Power BI supports Python and R visuals. If you're a user of these advanced visualisation options, you'll be happy to know that the runtime and library versions are getting a much-needed update. Python and R will be upgraded to 3.11 and 4.3.3, respectively. The Power BI Service will also use up to 2 vCores when rendering the visuals. Learn more.

Microsoft Fabric

• The preview for CI/CD and Git support in Dataflow Gen2 is finally here (it was originally slated for November last year). There are two pleasant surprises that come with this: (1) the new version of the Dataflow Gen2 item allows you to save changes without automatically triggering a refresh, and (2) the new takeover experience (more on this below) can be used to take ownership of the CI/CD enabled dataflows. Learn more.

• Until recently, limited item ownership transfer options in Fabric meant major issues could arise when a person who owns a Fabric item leaves an organisation. Some items supported ownership transfer through the UI (e.g. Dataflows Gen2), some required PowerShell scripting (e.g. Warehouses), and others offered no option other than to raise a support request (e.g. Lakehouses). In a worst-case scenario, this could cause many of your data pipelines to start failing as soon as the owning user's account is deactivated. Thankfully, this issue is partly remedied by the new item takeover UI experience that enables you to take over an item if you have the appropriate permissions. You can't use the new method to assign ownership to a service principal, but API support is coming soon and this could make that possible. Learn more.

• A common pain point with running background workloads (e.g. data pipelines) and interactive workloads (e.g. Power BI reports) on a single Fabric capacity is that the former can cause severe throttling in the latter. One way to protect your interactive workloads is to place them in separate capacities, but this is not always an option due to the associated costs. Fortunately, Microsoft introduced a new feature called Surge Protection. It's currently in preview and allows you to limit how much of a capacity can be consumed by background processes, giving you the ability to ring-fencing some capacity for your interactive workloads. Learn more.

Microsoft 365 Copilot and Copilot Studio

• In this age of generative AI and agentic AI, knowing how to instruct AI through prompts (I.e. prompt engineering) is an invaluable skill. Although AI models are improving each month, that's useless if we don't know how to structure and articulate how we ask them to perform a task. As good as they are, they're not mind readers. A recent article from the Microsoft 365 Copilot team uses a light-hearted project to help you learn how to write effective prompts. The mini-project aims to create your dream destination using step-by-step instructions that include using prompts that provide a context, define the tone, purpose, inclusions and exclusions, and much more. Learn more.

• Microsoft is serious about enabling every organisation to do more with AI and has doubled down on this vision with the recent launch of Microsoft 365 Copilot Chat. With Microsoft 365 Copilot Chat, you get access to some Microsoft 365 Copilot features for free with the option to use more advanced features, like Copilot Agents, on a pay-as-you-go basis which lowers the barrier to entry. Learn more.

• Imagine if Copilot could see what you see on your web browser and help you complete tasks based on that context. That's what Copilot Vision is, a new vision-based AI service that can help you do your holiday shopping, play online games, and support you in other online tasks and experiences. It takes a privacy-first approach and is only enabled on an opt-in basis for a single web browsing session. Once that session ends, it deletes all the data it captured during the session. The experimental feature works with a few websites, which will grow over time. It's also gradually rolling out, starting with the US. Learn more.

Azure Analytics and AI

• With the rapid rate of advancement in AI models, you might get the best results in your AI solution by combining multiple models. Although this approach often increases the complexity of an AI solution's complexity, it can significantly improve the overall performance if done well. A recent article from the Azure AI team covers the key considerations for taking this abroad and then explores some implementation scenarios. Learn more.

• DeepSeek emerged from nowhere, with AI models rivalling those from OpenAI, Google, Anthropric, and others leading in the generative AI space. The most impressive part is that it does this at a fraction of the cost. The Azure AI team wasted no time in making the DeepSeek R1 model available in Azure AI Foundry, where you can quickly and securely test whether it lives up to the hype for your use cases. It's worth noting that although the DeepSeek models may match models from OpenAI in performance, they don't currently have the same safeguards (more on this towards the end of this post). Learn more.

• The o3-mini model is available in Azure OpenAI services. Being a Large Reasoning Model (LRM), it excels at code, math, scientific reason problem and, compared to its predecessor, o1-mini, it's faster, supports structure outputs, and is a better fit for AI automation thanks to its support for functions and external tools. Learn more.

Open-Source Analytics and AI

• AI security researchers explored DeepSeek R1's vulnerabilities and compared these with OpenAI's o1-preview and GPT-4o, Meta's Llama 3.1-405B, Google Gemini 1.5-pro, and Claude 3.5-Sonnet. The most shocking result was that DeepSeek R1 had a 100% attack success rate, whereas OpenAI o1-preview had just 26%. You may wish to increase the performance and cost-effectiveness of your AI solutions but that usually comes with an implied assumption of equivalent or better security. Learn more.

• In case you missed it, Apache Superset 4.1 landed towards the end of last year. It's a minor release that includes a new Big Number visualisation, tables with time comparison functionality, and updated Histogram, Heatmap, and Sankey chart types. It may be a minor release, but these are all welcome additions to what is arguably the closest thing to a Power BI or Tableau when it comes to report and dashboard creation in the open-source software community. Learn more.

• Mobile phones aren't the first thing that comes to mind when you think of the hardware that runs an OLAP database. But that hasn't stopped the team over at DuckDB from running TPCH SF100 benchmarks on an iPhone 16 Pro and a Samsung Galaxy S4 Ultra. The article compares their performance to an AWS EC2 r6id.large instance and explains the motivation behind conducting these unusual experiments. Learn more.

Did You Find This Useful?

Get notified when we post something new by following us on X and LinkedIn.