Week 2026-21
As much as we hate to admit it, it’s time to face the fact that maybe we, as humans, just won’t be that needed anymore. @vlkodotnet
Week’s Highlight: The Dead Economy Theory
A long time ago, back in the previous millennium, came an era that would later be called the Industrial Revolution. Industrial, because machines replaced human labor. And people didn’t take it well. Everyone insisted that if machines replaced them, they’d have nothing to do and would starve. Today we know that machines made goods cheaper so more people could afford them, machines shortened distances, made agriculture more efficient, and people moved on to higher value-added work.
That’s how it went for decades, and that added value came to mean knowing how to fill in an Excel sheet correctly, build a nice presentation, answer an email, chat with someone on a helpline, write lots of related text, build the umpteenth version of a CRUD form, lay out graphics for a website, and so on. There was plenty of all this work to go around, humanity was happy, and GDP kept growing. When you asked a small child what they wanted to be, their eyes would light up with a vision of a padded leather chair, a wooden desk, and a computer monitor sitting on top of it. What more could a person want than a job in a warm office with minimal responsibility, but a good salary, a phone with an unlimited data plan, meal vouchers, a gym membership, and a company car?
But one day along came evil AI which — just imagine — was able to automate all these tasks. Sometimes so well that not even the boss could tell it came from AI. People quietly figured this out themselves and passed the AI’s work off as their own, and often nobody even noticed. That’s because the boss himself was using AI to check his subordinates’ work. This could have gone on forever — everyone secretly paying for AI and not having to do anything. But there are levels of management in a company that noticed the AI and decided to cut out the unnecessary human middleman, since they could just use the AI directly.
The layoffs began, company productivity sometimes shot through the roof, and sometimes the company went bankrupt because the AI wiped the production database here, handed out discounts that didn’t cover costs there, and so on. But those were marginal side effects that only a group of so-called AI HATERS kept pointing to.
Unlike during the Industrial Revolution, the laid-off people didn’t find higher value-added work (after all, the AI handled all of that), and where there’s no value-added work, there isn’t much money, and therefore no demand for all those goods the economy was producing…
That’s my poetic way of leading you into the next article:
What do you think — is this a realistic prediction of the future? Should we be asking what happens to people in an economy where AI replaces most of the work? What will people do all day? Or is the era when people looked for meaning in their work already behind us?
By the way, it’s no coincidence that the US issued a warning about so-called “anti-tech extremism.” A few incidents of attacks on the CEOs of AI companies are already known, and attacks against data centers are expected as well.
Coder Insights
I’ve collected quite a few interesting articles that I can’t really categorize any other way. Let’s start with one explaining how push notifications work on modern operating systems. A push notification is no longer a direct link from the service/app to the notification center of your phone, the way it used to be. Both Apple and Google now deliver notifications through centralized services that add filtering, categories, delivery timing, and AI summaries. In the future, it’s quite possible your push notification won’t even reach the user and will be handled directly by an AI agent.
A guide to migrating from Go to Rust. By some research, Go is still a more widely used language than Rust. But Rust is gaining users, thanks to its stronger type system and the elimination of nil errors via Option. On the flip side, some of Rust’s traits can put programmers off. On top of that, Rust is an excellent fit for the AI coding loop — because that’s the era we live in.
Finally, here’s something that keeps me up at night. I’ve written here several times that, under current legislation, AI-generated code isn’t a copyrightable work. Over the years we’ve built up a pile of hand-written code, but in recent months the AI-generated kind has been piling up too. That code could be covered by copyright fairly well by storing the prompts/conversation with the AI agent. We just still can’t quite figure out where to put them. Git Notes could be a pretty usable solution. The advantage is that they can be attached to a commit after the fact. The downside is that you’ll need some tool to read them, because GitHub, for example, no longer supports them natively.
HW Insights
It was starting to look like this year wouldn’t bring anything interesting for end users of portable computers. And then suddenly Nvidia unveiled the RTX Spark platform — a 20-core Grace CPU, a Blackwell GPU, and 128 GB of LPDDR5X RAM. I get the sense they shrank it down a bit more compared to the original Spark so it would fit into a laptop body. Together with Microsoft they added native Windows support, where a machine like this can handle all your AI tasks. The original DGX Spark carries a price tag of around 5,000 euros, so let’s just hope the new models on this platform won’t have the same one. If you’ve been waiting for the N1 and N1X chips to launch, those are exactly what’s used here. It’s not directly confirmed yet, but there will also be RTX Spark versions with less RAM and other CPU configurations.
Speaking of Microsoft — they unveiled the new Surface Laptop Ultra. Without announcing a price, of course.
At the opposite end of the performance spectrum should be laptops with Qualcomm’s Snapdragon C. Still Windows, with 8 GB of RAM, but with a price tag from 300 to 500 dollars.
Intel unveiled the Arc G3 and Arc G3 Extreme gaming chips, which will power gaming consoles as soon as this fall. The first announced console is the Predator Atlas 8.
For gamers, there’s the StepSense sound card, which amplifies the movement sounds of opponents and thereby gives players a competitive advantage.
BIZ Insights
Last week brought updates to Google Search. Not everyone saw it as a welcome change, so it doesn’t hurt to remember that alternative search engines exist too. To my surprise, Bing isn’t among them.
Google is replacing the Fitbit app with Google Health. The app got a new UI built around short AI reports on your activity. You can have a conversation about each of these reports. For me, that’s more useful than some of the charts I couldn’t even interpret. The bad news is that they removed badges and group challenges.
Anthropic announced that the current quarter will be its first profitable one, with projected revenue over 10 billion dollars. The reason is a big surge in users and a favorable deal with SpaceX to rent data centers, which covered the compute power they needed.
OpenAI isn’t slacking either, reporting 5.7 billion dollars in revenue last quarter. The catch is that, compared to Anthropic, for every dollar it earns it loses 1.22 dollars. The reason is that Anthropic uses cheaper chips from Google and Amazon for inference and, above all, doesn’t subsidize 900 million users, the vast majority of whom use ChatGPT for free.
OpenRouter, a system that brokers access to different AI models from one place, has raised a 133 million dollar investment.
AI Insights
Anthropic’s new top model, Opus 4.8, is out. Compared to version 4.7 the benchmarks improved, and we programmers will be glad it more often admits when it can’t solve something and asks for instructions. It also got better at holding context over long tasks and at the number of overlooked errors. The price stays the same, but a cheaper Fast mode was added, where for 2× the price you get 2.5× faster computation.
China responded with a new model, MiniMax M3, which is 3-in-1: a 1M context, native multimodality, and on top of that coding and agentic capabilities. All of it up to 9× faster than the previous version. An open-weight version of the model is also in the works.
Step 3.7 Flash is an agentic multimodal open-weight Flash model with 198B parameters that you can also get on your own local hardware with 128 GB of RAM.
We’ll close out the model parade with the smallest one, LFM 2.5, with 8 billion parameters and a 128K context, which will be able to fairly satisfyingly power your personal assistant on a mobile device.
.NET Insights
Microsoft is introducing a new memory-safety model around the use of the unsafe keyword. It’s no longer enough to mark a method as unsafe. Now you also have to wrap every use of it in an unsafe { } block.
Here’s part two of the hypothetical scenario on how to handle a state machine at Uber scale in .NET.
Links Drop
The Enhanced Games took place in the US. They had an interesting rule: athletes can use any supplement/drug/hormone/doping that isn’t banned within the territory of the US. The games were generously funded, and the reward for a world record was 1 million dollars.
Ferrari unveiled its first electric vehicle, and it stirred up more controversy than any car has in a long time. For Ferrari fans it’s a toy car, but from a technological standpoint it’s an interestingly engineered vehicle that seems to have a second shell around its body. The well-known design studio of Jony Ive was involved in its creation.
AudioMass is an open-source audio editor that runs right in your browser.
Closing Visual
This weekend everyone was caught off guard by the explosion of Blue Origin’s New Glenn rocket.
The saddest part is that it could have a negative impact on the Artemis missions, which will now have to use the still-unfinished SpaceX Starship or risk being overtaken by China.
