Week 2026-03

You already know Ralph—so what comes next? @vlkodotnet

Week’s Reflection: What Comes After Ralph?

The holy grail of AI-assisted development is essentially running your own software house. You’d have something—or someone—to communicate your requirements to, with processes running in the background that deliver results. No worrying about team morale, no dealing with vacation schedules. The only things you’d care about are how many tokens you’ve burned and what you’re paying for them.

Imagine you’ve designed a system built on the Ralph loop or some similar technique. There’s still a “human in the loop,” though. So the logical next step is to remove that too. What remains are just the input requirements—without them, the system would have nothing to execute. Those input requirements become the most important fuel for the entire system, and the quality of those requirements determines whether the system works or not. But that’s more or less standard practice these days, so nothing groundbreaking so far.

What this system should change is the internal workings, which would be fully automated and managed by a whole hierarchy of AI agents, each with a specific role. The main lead would be the only point of contact, with workers beneath who receive various oversight agents. These oversight agents would monitor whether the workers are completing their tasks, test their output, and watch for inconsistent states. There’s also an AI agent role responsible for synchronizing work across individual workers. The goal? Abstract away code writing entirely. Welcome to Gas Town.

Read at your own risk, but according to the author, this is roughly what the future should look like. It’s a mental exercise—the system has many weaknesses—but the concept is interesting. And where there are interesting concepts, interesting critiques emerge that develop the concept further, find its flaws, and explore potential solutions.

You don’t need to worry just yet—these systems behave unpredictably, though that could change over time. In the meantime, you can try a simplified system like Get Shit Done. You can run it on an existing project and expect a very intensive phase of problem specification and planning, after which everything runs on its own.

Before diving into your project, read the following article so you don’t set your expectations too high. The quality of results from your agentic framework depends on factors like your technology choice—the more popular the technology, the better the results. Another challenge is taking a project from prototype stage to real product. Prototyping is easy; refinement takes time, because you get used to the speed of those first results and then end up waiting for some feature loop to finish.

Business Insights

Sony and TCL are joining forces, and the result should be a stronger brand combining TCL’s technology with Sony’s reputation. This will take some time though—possibly until next year.

Monnett is a social network with no AI, bots, or tracking. Development and data are based in Europe, and critics say there’s one more thing missing: users.

BirdyChat is another EU chat app that also supports the WhatsApp protocol, letting you send messages, photos, and files to each other over an encrypted connection.

Threads from Meta was the only place in their MetaVerse without ads—until now. After reaching 400 million active users, it’s time to monetize them.

Tesla robotaxis no longer need human oversight and have started operating completely autonomously—though only in limited numbers so far.

If you haven’t heard about the Google-Epic dispute in a while, there might be an undisclosed partnership deal behind that silence. Android will promote Fortnite, and Epic will promote game development on Android using their Unreal Engine.

AI Insights

Qwen 3 TTS is an open model supporting voice cloning, voice design, natural high-quality speech, and voice control through text instructions. It stands out from other models with its speed, though it only supports 10 languages.

Sweep Next-Edit is an AI model designed for extremely fast prediction of what comes next in your code. This is useful for various IDE interfaces.

Kona is a non-autoregressive reasoning model. This should mean it can determine whether a solution is heading in the right direction and backtrack if needed.

OpenAI Codex now has its own version of the Ralph loop.

Google is expanding AI Mode with a “Personal Intelligence” feature that will be able to see into your Gmail and photos.

Dev Insights

You may have noticed the .NET corner has been missing lately. I need to look into why no interesting .NET links have reached me for the past three weeks. As a consolation, here are some links you can use as tools. First up: GitHub Copilot support in SQL Server Management Studio (SSMS).

winapp is new—and no, it’s not a music player, but shorthand for Windows App Development CLI. It helps you install required SDKs, manifests, certificates, or create MSIX packages.

Copilot Memories is a new feature that creates specific preferences and rules based on the projects you work on, then feeds those to Copilot prompts.

Claude Code is now available as a Visual Studio Code extension.

Links Drop

The Asus Zenbook Duo 2026 looks like a solid piece of hardware. It’s both a laptop and a dual-monitor system. Panther Lake is powerful enough to handle everything, and in book mode it doesn’t look as inconsistent as the old model. The only thing missing is a 64GB RAM version.

NexPhone is a phone that can run Linux alongside Android. And if that wasn’t enough, in dual-boot mode it can run Windows 11—though it’s the Windows version designed for kiosks and cash registers.

Nintendo unveiled the Talking Flower—a strange physical device that does nothing except shout catchphrases from Nintendo games when you press a button. To be fair, it can also do this without the button at scheduled times. For $35.

It looks like this year we might see Windows PCs with Nvidia ARM processors, codenamed N1x.

OpenAI released a case study on how they use PostgreSQL for their 800 million users. A system with 50 read replicas helps make that possible.

Here are some PostgreSQL tips on unconventionally optimizing indexes and queries.

ChartGPU renders charts using WebGPU. This could come in handy if you have massive amounts of data and your charting library can’t handle it.

I’ve seen computers boot from all sorts of things, but never from a classic vinyl record.

Closing Visual

Lego and Crocs announced a collaboration.