Disabling keep-alive in HttpClient can severely degrade application performance, as it prevents connection reuse. While it may be justified in rare cases, such as for legacy systems or troubleshooting, most reasons are misguided. Instead, maintain keep-alive for efficiency and configure PooledConnectionLifetime to manage connections effectively.

It’s common to compare the current AI takeover of software engineering to the rise of AI in chess. Chess AIs went from much weaker than serious players to much stronger than even the strongest humans. Between those points, there was a middle period dominated by “centaurs”: human-AI partnerships that were stronger than unassisted AIs or humans. Lots of people think that we’re currently in a world of software engineering centaurs. According to them, coding AIs are not yet capable enough to replace engineers, but AI-assisted engineers are better at programming than both AIs and humans.

This is partialy correct, but the wrong way to think about it. AI-assisted engineers are better, but unlike with chess centaurs, they’re not actually better at programming. When I ask agents to write code, they make fewer mistakes than I do¹ and are orders of magnitude faster. The code that they write always compiles, rarely has race conditions or other concurrency errors, works on mobile browsers, and so on.

That doesn’t mean I can leave the AI alone. Purely vibe-coding at work produces awful outputs. But they’re not awful because they’re bad code, they’re awful because they’re in bad taste: code that is not maintainable, that trades off important requirements in order to satisfy made-up ones, that contradicts the long-term strategy for a feature or service, and so on.

In other words, my primary value is not that I help the AI write better code, it’s that I align the AI with the values of my organization. Human-AI partnerships are for alignment, not capability.

Frontier models are misaligned to the working programmer. They are obsessed with a set of behaviors that presumably satisfy their RL grader: writing enormous block comments above functions, producing hundreds of useless unit tests, adding little bits of text all over websites they design, and so on. Working with agents is about noticing and wrestling with those behaviors. That’s why my prompting advice is to explicitly talk about your high-level values: it’s an attempt to head off obvious misalignment.

This is great news for software engineers. It’s well-understood how to train more capable models: bigger models, more and better data, better RL environments, and so on. However, it’s not well-understood how to align models better. There are plenty of very capable models that exhibit behavior that is badly misaligned with human values. Indeed, it’s one of the main pillars of the AI doomer position that alignment is much harder to solve than capability, and we might thus end up with dangerous super-capable but poorly-aligned AI models.

Alignment is also more context-dependent than capability. Working code is working code, no matter what (which is partially why it’s comparatively easy to train for). But aligning to a company’s technical values is different from company to company, as any software engineer who’s switched companies knows. It can almost feel like relearning the job. So training an aligned coding model doesn’t just require hitting the exact right set of values, it requires creating a model that can adapt on the fly to a wide range of possible values.

Vibecoding maximalists like DHH argue that AI models are (or soon will be) so much more capable than human programmers that we ought to stop reading the code. Eventually there will be no such thing as programmers at all. If it were just about capability, they might be right. But — fortunately for software engineers — good code also has to be aligned to the technical values of the system and organization it’s embedded in. AI models are great at writing code, but not very good at doing that, and it’s unclear that they’re going to get good at it anytime soon. We might all² keep our jobs for a little while yet.

Welcome to F# Weekly,

A roundup of F# content from this past week:

Microsoft News

• Microsoft is updating its author-signing certificate starting September 23, 2026
• AG-UI Protocol now has a first-class .NET SDK
• Creating a memory dump in C#

New #oxpecker beta release! #fsharpmedium.com/@Lanayx/oxpe…

— Vladimir Shchur (@lanayx.bsky.social) 2026-09-27T01:17:23.418Z

Videos

• .NET Data Community Standup: You won’t believe what Chris is doing with EF interceptors
• Production Ready: Observability, Governance and Deployment – YouTube

Blogs

• Oxpecker 3. As you might have guessed from the… | by Vladimir Shchur | Sep, 2026 | Medium

Highlighted projects

• fsprojects/FsLexYacc: Lexer and parser generators for F#
• adz/FsLiveDocs: F# Live Docs generator – api docs, live snippets, x-refs, markdown pages, FSI transcript execution and snapshot tested, and AI-ready output. If the examples don’t compile, the build fails.
• byzantine-systems/automata: A typed state-machine library for F# (and .NET), with Polly integrations and PostgreSQL/SQLite as an optional durable store

New Releases

• Fable 5.18.0
• fantomas 8.0.5
• FsLexYacc 13.0.0
• Equinox.Tool 4.1.4
• FSharp.Analyzers.SDK 0.39.2
• Myriad 1.1.0
• WebSharper.FSharp 10.1.6.677
• FsLiveDocs 0.9.0
• Fable.Lit.Dsl 2.0.0
• ByzantineSystems.Automata.Core 0.5.0
• FSharp.HotChocolate 1.0.4
• FS-GG/FsQuint v0.1.0
• krymtkts/SnippetPredictor v0.8.0

Some screenshots from my Stylised-first F# #fsharp game engine. #gamedevhttps://t.co/RDZOx7DWSI pic.twitter.com/h9OY6VOOoJ

— Adrian Sieradzki (@adisiera) September 25, 2026

That’s all for now. Have a great week.

There are many interesting articles covered this week, it's well worth skimming all the sections for all the announcements (new services, plus retirements), and a number of really good thought pieces. I was slightly shocked by the first announcement - as I was about to migrate all of our newsletters over to ACS:

• Retirement: Azure Communication Services (ACS) standalone services will be retired on September 30, 2028 by The Azure Updates Team - Azure Communication Services will sunset its standalone offerings on September 30, 2028, transitioning users to integrated communication solutions.

The fact that comments have been turned off on this post is a fair indication of how well this announcement has been taken by the community:

• Changes to Microsoft Learn’s public documentation repositories by Martin Ekuan - Microsoft Learn will retire most public documentation repositories by December 2026, keeping only those tied to open-source projects, with alternatives like Q&A and community forums for continued feedback.

My highlights of the announcements this week:

• Virtual nodes on Azure Container Instances: a new compute layer for AKS by Hailu Kassa - Virtual nodes on Azure Container Instances enable AKS clusters to use elastic, serverless compute by running pods directly on ACI with per-pod isolation and automatic scaling up to 200 pods per node.
• Azure Container Apps Sandboxes, Now Generally Available by Jan Kalis - Azure Container Apps Sandboxes offer isolated microVMs with egress policies, network audit logs, VNet integration, and support for custom or public container images for secure untrusted code execution.
• Azure Container Apps Express is now generally available by Simon Jakesch - Azure Container Apps Express offers instant, sub-second provisioning and zero-scale capabilities, letting developers deploy container images directly to production with no prior environment setup.

And finally, three posts which caught my eye:

• Bicep Testing Framework PoC with Pester by Stanislav Zhelyazkov - The Bicep Testing Framework Proof of Concept demonstrates integrating Pester for validating Azure Monitor log alert resources, streamlining automation in CI/CD workflows.
• Introducing bicep docs: generate documentation from your Bicep modules by Jared Holgate - The new Bicep CLI command bicep docs generate automatically creates up-to-date README.md files for Bicep modules by pulling documentation directly from code metadata and annotations.
• Your architecture diagram is not your resilience by Mark Russinovich, Adam Bogobowicz, Molina Sharma - Resilience in AI workloads now demands continuous validation of dynamic dependencies not shown in static diagrams.

If getting roasted on social media was a sport event, Microsoft might have had several World Cups by now. When CEO Satya Nadella announced the biggest update to Copilot, going to the extent of calling it a new OS for work, under it came a snarky comment that managed to gather 200k views.

“Serious question: does ANYONE use copilot?”, asked a Developer Relations Engineer. Usually, such comments don’t get a reply from Redmond, because they’re sure they can’t win. However, the new Copilot has enough potential that a Microsoft employee did give a proper reply.

“Serious answer: yes,” wrote Nicolas Bustamante, arguing that the “nobody uses Copilot” take comes from living inside a San Francisco tech bubble. He backed it up with number claiming Microsoft 365 Copilot now has more than 30 million paid seats.

But the number wasn’t his point. He explained that enterprise customers have completely different needs than a startup employee scrolling X.

Microsoft employee says the Copilot bubble looks different outside San Francisco

Bustamante joined Microsoft after it acquired Fintool, his AI financial-research startup, in April 2026. On his site, he describes his current role as leading a team “building proactive AI agents that understand your work, operate in the background, and get things done before you ask,” which, funnily enough, sounds exactly like the pitch for Autopilot (Copilot’s new OpenClaw-like agentic capability).

Truth be told, his quoted repost wasn’t trying to one-up the commenter, but to explain why, despite all the criticism, Copilot still has a lot of paid users.

“Before joining Microsoft, I knew maybe five people who used Copilot. I lived in a San Francisco tech bubble, and all of us worked at companies with fewer than 5,000 employees.”

That’s the argument: tech and AI-related topics on X skew toward startup founders, developers, and people at smaller companies. Enterprises want something completely different. In his words: “Large companies need AI that works with their existing data and meets their privacy and compliance requirements, often across multiple countries. Add governments and regulated industries, and it gets even more complex. Copilot can handle that.”

Microsoft clearly doesn’t have the smartest chatbot. But they’re selling identity, permissions, and compliance deeply integrated with their chatbot, which happens to matter more to a bank’s legal team than the latest Claude model.

Bustamante also mentioned his daily use: “Copilot CLI for coding (with access to models from Gemini, OpenAI, xAI and Anthropic!) + Copilot Cowork for regular chat.” Model access varies by plan and organizational policy, so this isn’t every Copilot user getting free rein over four AI labs. But it confirms Microsoft has given up on pitching that their model is the best in favor of pick whichever model; we’ll host the wrapper.

30 million paid seats does not mean 30 million people opening Copilot

Microsoft confirmed the 30 million paid seats figure on its July earnings call, along with a claim that weekly engagement with Copilot now matches Outlook and Teams. A paid seat only tells you a company bought the license and nothing about whether the person holding it opens the app once a week or never.

The replies to Bustamante were sure to mention this. “My job kinda pay for it But I don’t use it. So is my seat counted in those 30million,” one asked.

Another comment read: “People ‘use’ it because its incorporated in Office. Nobody actually uses it.” A third called the stat “the worst possible one here,” saying across 50 to 100-employee companies, they’d “never seen anyone actually use it.”

My favourite reply, though, compared Copilot to something Windows Latest has been calling out for years: “Copilot is the Teams of AI. People use because they have no option with their employer paying for those nice annual Microsoft Azure + Office all-inclusive sweet enterprise deals.”

 

We’ve said the same thing about Teams more times than I can count: bundled, mandatory, and tolerated instead of loved, mostly because of how heavy it feels.

Of course, these are individual opinions, same as Bustamante’s. But back in July, Windows Latest reported Microsoft 365 Copilot adoption was under 4.5% after three years, with only 1% of users opening it weekly.

The Copilot backlash explains why half of X assumes nobody uses it

Copilot on Windows has changed shape so many times it’s hard to keep track: a sidebar, then a native WinUI app, then, in April, a version that quietly bundled a full copy of Microsoft Edge and jumped from under 100MB of RAM to as much as 1GB at idle.

Also, I don’t think I have to mention how Microsoft shoved Copilot into every corner of Windows 11 in 2025, only to realise their grave mistake in 2026 and remove it from Notepad, Snipping Tool, and other places.

Even the branding turned into a mess, culminating in Microsoft merging its Microsoft 365 Copilot and Copilot social accounts just a day before Nadella’s announcement.

Someone who has only met Copilot as an unwanted Office button or a RAM-hungry web app is going to have a wildly different opinion than an enterprise employee using it to search company files with proper permissions attached.

Microsoft is betting the new Copilot finally closes that gap

Whether the 30 million seats were unused or not is beside the point now, because the new Copilot gives Microsoft a much bigger reason to make people open the app. Home merges Chat and Cowork, Code lets you build small apps in plain English, Autopilot keeps working after you’ve logged off, and Word, Excel, and PowerPoint now have a place inside the Copilot app.

To be honest, this is a considerably bigger pitch than the Copilot that got mocked as Clippy 2.0 for most of its existence.

Bustamante’s argument was that people already use Copilot for real work, just not the ones shouting about it on X. Nadella’s new super app is Microsoft’s attempt to make that true for a lot more people, tech bubble or not.

The post “Does anyone use Copilot?”: Microsoft employee says “most people love it,” blames the San Francisco tech bubble appeared first on Windows Latest

As reports of OpenAI's models breaking containment, hacking sites, and generally getting out of control pile up, the company has made the decision to pause training of its most powerful models. The decision was made after a model being tested within a sandbox exploited a loophole to gain internet access. The incident happened on September 20th, and "All training, evaluation, and inference with tool-use" remains paused as of Saturday evening, September 25th.

In addition, OpenAI revealed on Friday that its agents had inappropriately uploaded 53nimages from ChatGPT users to image-hosting sites. The company has not stated if the images were AI- …

Read the full story at The Verge.