Mark Farragher

Mark Farragher

@mark.mdft.social

Microsoft Certified Trainer | Former CTO & CoFounder | Cloud, Data & AI Expert | AI FIXER | and more!

38 followers32 following69 posts

Mark Farragher reposted

Framework
@frame.work

Pre-orders are live on the new 192GB Framework Desktop! DIY Edition starts at $6,799 USD and ships in November.

Mark Farragher reposted

OpenAlternative
@openalternative.co

Apache Superset (@apachesuperset.bsky.social) just crossed 75,000 stars on GitHub. Self-serve data exploration and dashboarding: connect any SQL database, drag-and-drop or raw SQL to build charts, publish interactive dashboards without writing app code. openalternative.co/apache-supe...

openalternative.coApache Superset: Open Source Alternative to Power BI, Splunk and TableauConnect any SQL database, build charts with drag-and-drop or raw SQL, and publish interactive dashboards without writing application code.

Mark Farragher reposted

Andrew Cairns
@acairns.co.uk

First few clips for the intro cut together to get a feel. Still a couple of issues to solve. But, it's getting there! Not looking forward to lip sync stuff, though... 😅

Mark Farragher
@mark.mdft.social

In today's free lesson, we explain the cost of a BOX operation in dotNET intermediate language in terms of heap allocation, memory zeroing, type descriptor initialization, and reference placement on the stack. mdftpro.com/play/high-pe...

mdftpro.comOptimizing Boxing And Unboxing - Basic Performance OptimizationsWe explain the full cost of a box operation at the IL level including heap allocation, memory zeroing, type descriptor initialization, and reference placement on the stack. The lesson identifies no...
Mark Farragher
@mark.mdft.social

In April, Microsoft released the upgraded AI-103: Azure AI Agents Developer Associate certification. This is an upgrade for the retired AI-102 exam. Read all about it in our latest blog post: mdftpro.com/blog/microso...

mdftpro.comNew: AI-103 Azure AI Agents DeveloperJust three months ago in April, Microsoft retired their existing AI-102 certification and replaced it with the upgraded AI-103: Azure AI Agents Developer Associate. We're happy to announce that...
Mark Farragher
@mark.mdft.social

Today's free lesson is from the DP-100 Azure Data Scientist course, and explains how an inference pipeline uses a trained model to generate predictions from new data, then scores and returns predictions through a web service output. mdftpro.com/play/azure-d...

mdftpro.comInference Pipelines - The Azure ML DesignerThe lesson explains how inference pipelines use trained models to generate predictions from new data, applying the same transformations from training via the apply transformation block, then scorin...
Mark Farragher
@mark.mdft.social

Did you know you can do a lot of data science work in the C# language? There's often no need to switch to python, you can keep using your favorite language instead. As a dotNET developer you have a lot of options. Read more in our blog post: mdftpro.com/blog/courses...

mdftpro.comData Science in C#? Yes, we can!Did you know you can do a lot of data science work in pure C#? For many things there's no need to switch to python, you can just keep using your favorite programming language. Analyzing dataset...
Mark Farragher
@mark.mdft.social

In May, Microsoft released the all new AI-200: Azure AI Cloud Developer Associate certification. And we already have it online on MDFT Pro, all 115 video lessons! Read all about it in our latest blog post: mdftpro.com/blog/microso...

mdftpro.comNew: AI-200 Azure AI Cloud DeveloperJust two months ago in May, Microsoft retired their existing AZ-204 certification and replaced it with the entirely new AI-200: Azure AI Cloud Developer Associate. We're happy to announce that ...
Mark Farragher
@mark.mdft.social

In today's free lesson (free for the next 72 hours), learn about the small and large object heaps in the #dotNET runtime environment. Knowledge about heap memory management will help you become a better developer and write faster and more efficient #csharp code. mdftpro.com/play/high-pe...

mdftpro.comThe Small and Large Object Heaps - The .NET Garbage CollectorWe cover the disadvantage of generational garbage collection for large long-lived objects (multiple expensive memory copies during migration), and present the solution: the Large Object Heap, which...
Mark Farragher
@mark.mdft.social

(2/2) In today's free lesson, we compile a simple C# app to intermediate language and inspect the generated STELEM instruction. The lesson explains how STELEM provides a massive performance boost, making 1-dimensional array performance on par with C/C++ code.

Mark Farragher
@mark.mdft.social

(1/2) Today I coded for 4 hours with #OpenCode using #zAI GLM 5 instead of Claude Sonnet 4.6. I perceived no difference in capabilities. GLM5 handled edits well, made some mistakes just like Sonnet, and produced output of the same quality.

Mark Farragher
@mark.mdft.social

(2/2) But GLM 5 is 3x smaller than Sonnet! Private AI models are wasteful. They're too big, use too much power, and are too expensive to run. #openweight is the future.

Mark Farragher
@mark.mdft.social

Microsoft has introduced strict rate limits on Github Copilot this week, making the Pro service basically unusable. This is the final push I needed to switch to Open Weight models. My current setup: OpenCode --> OpenRouter --> GLM 5.

Mark Farragher
@mark.mdft.social

(1/2) Yesterday I was building a web app with OpenCode, GitHub Copilot and Deno. A web page didn't look right, and I asked Copilot to check it out and tell me what's wrong. The agent loaded the page, got an access denied error, and decided that it needed admin privileges to debug this properly.

Mark Farragher
@mark.mdft.social

(2/2) It then tried to hack into my browser session to steal my login token. I was watching in real time and managed to pull the plug in time. Did I mention you should never run an AI agent unattended?

Mark Farragher
@mark.mdft.social

(1/2) The real problem with #genAI is that the software industry has operated for decades with no regard for quality at all. Today #AI coding agents produce low quality code at scale. And some developers are fine with running these tools day and night with no human oversight.

Mark Farragher
@mark.mdft.social

(2/2) In a sane world, we would have static type systems, code contracts, static analyzers, compilers and proof checkers in our toolchain and #codeslop wouldn't stand a chance. It's time for the industry to finally grow up.

Mark Farragher
@mark.mdft.social

(3/3) My generated code must run in a compiled language, with strong type checking, and within a security sandbox. This is why I stay away from #NodeJS and embrace #Deno instead.

Mark Farragher
@mark.mdft.social

(2/3) I create highly modular app architectures and try to implement each module with off-the-shelf #opensource components. I only use generated code to fill in the gaps, and I strictly separate code and design so that I can easily refactor the design later if needed.

Mark Farragher
@mark.mdft.social

(1/3) After a busy week of #agentic #AI #coding, I've come up with the following strategy: I treat my coding #agent as junior programmers. I run a tight loop where I prompt, and immediately check its work afterwards and refactor if I don't agree with the code. I never let agents run unattended.

Mark Farragher
@mark.mdft.social

During the migration from #OpenAI to #Anthropic, I'm going against the grain by migrating from Anthropic to #GitHub #Copilot Pro 🤪 I'm building an app with #OpenCode, Copilot and #Gemini Pro 3, and it works great. Gemini is great with #typescript and #bootstrap. This is my default agent stack now.

Mark Farragher
@mark.mdft.social

According to the MIT NANDA Report of 2025, 95% of generative AI pilot projects fail to deliver a measurable impact on the P&L. AI FIXER puts you in the 5% that succeeds!

Mark Farragher
@mark.mdft.social

I'm very excited to announce AI FIXER, a new collaboration between Composition8 | Training & Business Consulting and me. 𝐒𝐲𝐬𝐭𝐞𝐦𝐬 𝐈𝐧𝐬𝐩𝐢𝐫𝐞𝐝 𝐀𝐈: we help you design AI that makes sense. More at AIFIXER dot NL: www.aifixer.nl #aifixer

Mark Farragher reposted

James
@nonzerosum.games

Just read this great piece by Thomas Johannsen, a perfect resource for exiting the problematic big tech monopolies (and inspiring you to do so). If you've already made the shift to Bluesky for ethical reasons, here are the next steps in your journey toward digital sainthood.

geneticchoir.wordpress.comOn Swarms, Davos and the Power of the PowerlessLet’s take a look today at the power of individuals against world leaders and what swarm improvisation principles suggest is best to do. This is a rather unusual post for this blog, but hey, …

Mark Farragher reposted

Ed Zitron
@edzitron.com

My final newsletter of 2025: We're in The Enshittifinancial Crisis, the fourth stage of enshittification, where companies turn on their shareholders. Unprofitable, unsustainable AI threatens future of venture capital, private equity and the markets themselves. www.wheresyoured.at/the-enshitti...

wheresyoured.atThe Enshittifinancial CrisisSoundtrack: Lynyrd Skynyrd — Free Bird This piece is over 19,000 words, and took me a great deal of writing and research. If you liked it, please subscribe to my premium newsletter. It’s $70 a year, ...
Mark Farragher
@mark.mdft.social

I just linked my first Innr Lighting light bulb to my z-wave powered #homeassistant, and now the light in my study is adaptive and follows the color and brightness of the sun! #homeautomation #happiness

Mark Farragher
@mark.mdft.social

From the trenches of AI coding: Prompt: "Create a plot with ScottPlot" AI: [writes ScottPlot 4.x code, tries to compile] AI: "I see the problem. You are using v5 and I wrote code for v4. Let me fix that." AI: [writes more ScottPlot 4.x code] AI: "I see the problem. You are using v5 and I wrote ...

Mark Farragher
@mark.mdft.social

Software development has now become 'voodoo-coding'. Instead of making a plan and writing short, elegant and performant code to implement the plan, we have been reduced to casting complex magic spells to compel fickle gods to write the code for us.

Mark Farragher
@mark.mdft.social

From the trenches of AI-coding: I'm building an F# machine learning app. I put the cursor at the bottom of the file and tell Copilot to "add a regression pipeline to train a model on the data". Copilot inserts the new code at the top of the file, before the code that actually loads the data.