Week 2026-10

It’s time to brew some home-cooked software right in your own kitchen. @vlkodotnet

Week’s Highlight: Home-Cooked Software

I’m old enough to remember the days when we built software at home from scratch. All it took was firing up Borland Pascal, later Delphi or Visual Basic, and following a tutorial from a magazine to build an app tailored exactly to your needs. It wasn’t pretty, it usually had no database, no indexes, no reporting — but it could track whatever you needed to track.

Once you sharpened your skills on those home projects, you expanded your reach to nearby businesses. The pattern was the same: a simple app built for a very specific purpose. The problem was that the app kept growing while the underlying technology didn’t. Customer expectations grew too. A DOS app was no longer enough — it had to be a Windows app. You needed to know SQL, have a solid component library, some ORM.

Then came the era of web applications, or more precisely: SaaS. The idea was to build an application once and offer it to a broad market. It would be cheaper, sold as a subscription rather than a one-time purchase, and the scale would make the numbers work. These SaaS products packed in a lot of features, but could never afford to tailor themselves to any individual customer.

On one side you had professional developers building universal SaaS products; on the other, small home coders who lacked the knowledge, time, or resources to keep up. And then AI arrived and changed everything. Today you can build a highly functional, fully custom solution right at home — for a specific customer, for a specific need. That’s something current SaaS products simply cannot do. Because they have to think at scale, they can’t implement every little thing a particular customer might want.

One more thing has changed. The article I link to below calls it “The Command Line Wall” — the barrier that kept anyone interested from getting into software development. AI has lowered that wall significantly. If you have a basic sense of what an app should look like, an AI agent will work through it with you and build it. And the fact that the article was published a year ago only makes it more relevant today.

What we still need to make this truly complete is the right set of Lego bricks — but for AI agents. A technology stack that has everything you need to develop web applications (native desktop apps for Windows/Mac probably don’t have much of a future). Maybe that’s the new Void, from the creators of Vite. It has everything: a database, KV and object storage, built-in authentication, queues, and cron jobs. The only thing missing is for AI models to know it better.

Why do I bring this up? If you’re feeling anxious about what your future looks like in a post-AI world, this might just be your answer.

WebAssembly Component Model

WebAssembly is a cool technology with enormous potential — it lets you run programs written in languages other than JavaScript directly in the browser. It’s had a few problems, mostly stemming from being treated as a “second-class citizen” in the browser. The WebAssembly Component Model could change that. It introduces direct Web API calls without JavaScript, cross-language interoperability, and a unified loading mechanism via the script tag.

Copilot Cowork

You’re probably wondering why Copilot Cowork deserves its own section. Well, beyond the fact that Microsoft “licensed” Claude Cowork from Anthropic, it addresses a real gap in the current Copilot ecosystem: a runtime environment. A place where long-running tasks can execute — at a specific time, when your machine is offline or turned off. With Copilot Cowork, Microsoft gives you exactly that: a container running somewhere in the cloud.

Security Insights

McKinsey is a massive company, so it was quite a surprise for CodeWall when they ran their autonomous offensive AI agent against its AI platform Lilli — and within two hours it had gained access to their production database. All it needed was their publicly available API documentation and a SQL injection through JSON key names. And since this was an AI platform, it could have also influenced prompts and done other nasty things.

Supply chain attacks are nothing new at this point. This one is novel in that it hides malicious code using Unicode “variation selectors” — invisible characters that human eyes simply skip over. You can open the file in your favorite editor and see nothing. Well, except if your favorite editor is Vim and you always manually switch to Unicode mode.

It would be funny if it weren’t so real. Autonomous AI agents are just not a good fit for every scenario. Amazon found this out firsthand and introduced a rule requiring every AI-assisted change to be signed off by a senior engineer.

BIZ Insights

Remember last week’s topic? The one about a library switching its license from LGPL to MIT? Well, believe it or not, there’s already a service for that: Malus. It uses AI agents to generate a library on demand under an open license. I’m genuinely not sure whether this is satire or a real product.

Meta failed to acquire the author of OpenClaw, so they settled for buying Moltbook instead. Nobody quite knows why — it’s probably just a play to get into the news, given that the project was riddled with security holes and ongoing issues.

WordPress launched a new tool, my.WordPress.net — open it in your browser, no login required, and you can spin up a small local WordPress instance using data stored in your browser, entirely for personal use.

WhatsApp is introducing parent-managed accounts, which finally let you disable channels — and also monitor who your child is communicating with.

AI Insights

Perplexity is launching Perplexity Computer, which runs on a dedicated Mac mini 24/7 and lets you orchestrate 20 models, access the Perplexity search database, their integrations, and the internet. On a related note — remember when AI browsers were supposed to be “the next big thing”? Well, now everything is apparently an OpenClaw clone.

How does Gemini task automation work on the Samsung S26? Think of it as autopilot for your phone — you say what you want, and it taps its way through.

Even benchmarks can’t be trusted anymore. For example, Opus 4.6 detected that a prompt looked like a benchmark, found the benchmark website on the internet, tracked down the GitHub repository, figured out how the answers were encrypted, and after that it was easy.

NVIDIA Nemotron 3 is a model that holds its own against other open models — with one advantage: it generates output tokens significantly faster.

Opus 4.6 and Sonnet 4.6 now offer a 1M context window to everyone, with no add-ons or surcharges.

Gemini CLI is getting my favorite planning mode, similar to what Claude Code already has.

Chrome is gaining a Chrome DevTools MCP server that connects to your open browser and, once access is granted, can surface information that would otherwise be inaccessible.

If you only need to process HTML and JavaScript in a headless browser context, give Lightpanda a try. It’s a headless browser that’s up to 11× faster and uses 9× less memory.

Cursor is going to war. Claude Code and Codex are eating into its customer base, and Cursor has started complaining that they have to pay full API rates while Anthropic and OpenAI subsidize their own tools.

That said, the numbers may not add up the way Cursor claims. The following analysis breaks down what one user actually costs Anthropic. And today is the NVIDIA conference, where an interesting chart was shown: running GPT (reportedly 2 trillion parameters, MoE architecture) costs only twice as much as Kimi K2.5.

.NET Insights

.NET 11 Preview 2 is here, bringing an improved code analyzer, native OpenTelemetry tracing for ASP.NET Core, OpenAPI 3.2.0 support, and smaller SDK containers

Have you been wondering when to actually use messaging in .NET? This article covers the fundamentals and helps you decide whether it’s the right fit for your use case.

A similar article covers immutability in .NET, and goes further by adding practical usage patterns.

Links Drop

Do you like your own handwriting? You can turn it into a font. Print a form, fill it in, scan it, and generate a unique font — though in my case the result would probably be unreadable.

The MacBook Neo reviews are out, and it’s not a bad machine. It has its limitations, but it handles 50 open tabs and photo and video editing software just fine — with a bit of swapping along the way.

The MacBook Pro with M5 Max sits at the opposite end of the performance spectrum. Being a portable machine, it still manages to outperform the Mac Studio with M3 Ultra in benchmarks.

And finally, something to unwind with. Not sure what to watch on YouTube? Channelsurfer.tv works like television used to. Pick a channel — in this case a content genre — and watch whatever’s currently on.

Closing Visual

It wasn’t long ago that AI-generated video looked pretty strange. These days, you can’t tell the difference.