Week 2026-11
Continuing the recipe for homemade software. @vlkodotnet
Thought of the week: Home-cooked software II.
Last week we left off on the question of what our place in the post-agentic era might look like. I see two directions: one is local, where anyone can generate their own software on demand. The other is corporate, because large companies will continue to exist — ones that handle large-scale problems that most small businesses and individuals won’t be able to solve with local software.
Let’s start with the first direction. Imagine a world where software is generated on request. Your job is to specify that request correctly, you press a button, and things work. Most of the necessary technology is already here in the form of AI agents, though local hardware to run it on is still missing. I picture it as a local AI server running something like a safer version of OpenClaw. You’d be able to integrate various external physical interfaces into it — turning lights on and off, monitoring a boiler, watering your garden. But also things at a larger scale: irrigating entire fields, controlling agricultural equipment, connecting to a self-driving tractor or other autonomous units. In smaller local businesses, it might look like an automatic CRM where you simply scan goods in and out through your phone camera, and the system alerts you when something needs to be reordered. The phone camera (or AR glasses) will serve as the bridge between the physical world and the automated one. Your job at the start is to write down what you want it to do, and an AI agent handles the rest. When something isn’t working as it should, you open the spec, adjust the instructions, press a button, and a new version is deployed.
It sounds simple — what could go wrong? A lot of things, and that’s exactly what the following essay explores in a story set in that future world. It’s definitely worth reading, but for those who’d rather skip it: the essay describes the birth of a new profession — the Software Mechanic, someone who doesn’t fix code but regenerates specifications that contain imprecise, incomplete, or implicit requirements. They also know how to trace the source of such problems, because they have domain knowledge.
The second direction is where our current software development industry is heading. It has already left the era of boxed software behind. Today almost everything is SaaS — you pick the service you need, and as long as you pay, you use it. Here it’s a ten-euro monthly subscription, there it’s five euros per employee, somewhere else it’s five hundred for a million API calls. Somewhere in the middle sits a large orchestrator — a CRM, ERP, SCM, or sometimes just a human. The decision about which specific systems to adopt is driven by the value they bring to the business. And honestly, what truly new ground has anyone broken in the last however-many years? Programming gave way to UML diagrams, UML diagrams gave way to domain-specific languages, and domain-specific languages have now given way to AI. None of it has produced a step-change in productivity. Not even AI — and you’re probably wondering why.
Well, I have another essay for that. It argues that every layer of approval slows the entire process down by a factor of ten. The article focuses on software development, but the same dynamic plays out in other organizations with management layers. I’m not going to criticize the existence of those layers — they usually have a historical reason for being there, born out of situations where someone couldn’t make decisions or, conversely, was making them too independently. Their purpose is to ensure predictability and order. A small innovative startup can afford to ship straight to production, but the moment you sign a contract with penalties for outages you cause, everything changes fast.
This is the crux of the matter. AI making you ten times more productive doesn’t change the fact that other management layers exist that cancel out that advantage. Can those layers be streamlined — say, by using AI that never sleeps? Sure, but there’s always another layer ready to slow things down again. Because accountability binds.
Nvidia GTC 2026
I originally wanted to write just about DLSS 5, but then I remembered Nvidia had an entire conference. It was a fairly traditional affair. Vera Rubin is the new GPU architecture. On top of that, Nvidia integrated the Groq Language Processing Unit (LPU), which can compute the AI model’s core calculations faster than anything else. This means the transformer layer — where matrix operations happen — will use the GPU, while token generation will run on the LPU. Groq uses SRAM memory, which is faster but more expensive.
Those are the processors, but you also have to factor in networking, controllers, and everything else — which means companies will buy a black box, slide it into a data center, and it will serve workloads at two to thirty-five times the energy efficiency of the previous platform. That’s the most important metric Nvidia highlighted. If you’re thinking of ordering one for home, I have bad news: they don’t plan to sell to any segment other than data centers before 2027. And even for data centers, Nvidia has a business plan requiring that buyers of Vera Rubin-based solutions first take a certain quantity of their Blackwell systems. You can find the full keynote here:
I notice I’ve written quite a bit for something I only meant to mention in passing. So, what to do with all this when you can’t actually buy any of it? That brings me smoothly to DLSS 5. DLSS is the technology that takes a weaker GPU — one that can only manage low resolutions — and produces an upscaled output at the cost of some latency. AI algorithms make it look usable. DLSS 5 stirred up controversy because it doesn’t just do visual upscaling — it also reworks lighting. That’s Nvidia’s claim, at least. What we actually see is that every face looks like it’s been run through an Instagram beauty filter. And honestly, so does everything else. It looks pretty, to the point where it looks like a completely different game — and that bothers everyone. Game developers hate that it undermines their artistic intent. Hardcore players feel the characters have somehow lost their soul. Casual players probably don’t care. A game with DLSS 5 looks nice even on weaker hardware (within the 50x0 series), and for them, that’s enough.
Windows 11 gets some love
Up until now, Windows 11 kept receiving new features and AI Copilot integrations that nobody really wanted — but which were necessary to show that the OS was moving somewhere. Windows chief Pavan Davuluri made a lot of promises as part of rebuilding trust with customers. By the end of May we should get a taskbar that can be repositioned to the top or side of the screen, fewer annoying AI features in Notepad, the Snipping Tool, and Photos. You’ll finally be able to postpone Windows Update indefinitely — or at least configure it so you don’t wake up to a restarted OS. By the end of the year they’ll work on resource efficiency so Windows 11 can run on 8 GB of RAM. And they’ll finally address the slow wake-from-sleep issue. Looking forward to it.
BIZ Insights
As part of its sideloading policy changes on Android, Google is introducing a 24-hour waiting period between enabling sideloading and being able to actually install an unverified app. If you plan to use this, either enable it on a new phone right from the start, or set aside plenty of patience.
Encyclopaedia Britannica is suing OpenAI, claiming that the model can reproduce entire passages from its content verbatim. I’m not sure what took them so long — what did they think those models were trained on?
Ladies and gentlemen, this is how you run a business. Delve was a startup that helped companies achieve regulatory compliance and issued certifications for GDPR, HIPAA, SOC 2, and ISO 27001. Some of those certifications are required by government organizations, with non-compliance carrying fines and loss of contracts. It now appears that these certifications were generated as templates with no real certification process behind them. Given that their client list included nearly every Fortune 500 company, this looks like an enormous mess.
OpenAI is preparing a more “adult” mode for ChatGPT. Internal OpenAI experts warn that such a mode could create emotional dependency and be exploited by minors, whom they still can’t reliably identify. It’s a smaller problem than the Delve situation, but one that will quietly fester for years.
Google unveiled a new product called Stitch — and the reason I’m covering it in the BIZ section rather than the AI section is that its announcement was followed by a drop in Adobe’s stock, and especially Figma’s. You can probably guess what it does: UI design and rapid application prototyping. Stitch is essentially a rebranded Galileo AI, which Google acquired.
Kagi introduced Small Web, a mobile app for iOS and Android featuring over 30,000 hand-curated, non-commercial websites. I don’t want to pressure anyone, but this newsletter is also a hand-crafted, non-commercial publication.
AI Insights
Mistral had a busy week. They released Small 4, a mixture-of-experts model that handles three times as many requests in the same time as the previous Small 3. With reasoning mode enabled, it’s nearly on par with Qwen 3.5.
Cursor released Composer 2, a model fine-tuned on top of Kimi K2.5.
MAI-Image-2 isn’t a top-tier model, but it’s precisely the third-best image generation model out there. Microsoft seems perfectly happy with that.
Unsloth Studio is a new open-source web UI for running and training AI models locally.
Curious what Claude Code actually is under the hood, given it weighs in at around 213 MB? Arinjay Wyawhare reverse-engineered the binary and found it’s essentially a bundled Bun runtime with 10 MB of minified JavaScript.
.NET Insights
A somewhat involved guide on how to use metrics to print ASP.NET Core web stats to the console. It looks cool, but who actually runs web apps like this?
The first part of a series on process synchronization — covering all the locks, mutexes, semaphores, and ReaderWriter locks. Useful for brushing up on the fundamentals.
Have you been putting off migrating from .NET 4.x for too long? Maybe you’ve even told yourself it’s not worth the effort anymore. With the arrival of AI agentic tools, it might not be as out of reach as it seems. This guide could come in handy.
Links Drop
Intel announced a new series of mobile processors — the Core Ultra 200HX Plus — with the potential to power your next gaming laptop.
IKEA is running into significant trouble with its new Matter-over-Thread smart devices. Getting them to connect to platforms outside the IKEA ecosystem — especially Apple Home — takes a frustratingly long time. It’s not exactly what the standard promised.
It’s a fun little novelty, but at the Nvidia conference they unveiled a robot version of Olaf. Just like in the movie, you can take off his nose and pop off his little arms. Inside, Olaf is packed with interesting technology that wouldn’t have been possible without Nvidia’s robotics development platform.
In Ubuntu 26.04, after 46 years, the era of invisible sudo passwords is coming to an end. If you use a password manager, you know the feeling of not being sure what you actually typed. From now on, each keystroke at a sudo prompt will be echoed as an asterisk. You can still turn it off, of course.
Bing can now translate into Klingon, and Kagi into LinkedIn-speak. You’ll never use the first one, but you’ll use the second one almost every day. I’ve been communicating with colleagues in LinkedIn-speak for a few days now, and on top of that I’ve been prompting Claude to add extra emojis and hashtags.
Nightingale is an app that extracts vocals from video and generates a transcript — the perfect combo for karaoke.
ReKindle is a black-and-white website designed for your black-and-white Kindle, offering a collection of web apps you can actually use on the device — at least in this form.
Closing Visual
This is roughly what relaxing looks like.
