Week 2026-25

Jul 01, 2026

RAMmageddon is coming. @vlkodotnet

Week’s Highlight: The Era of Expensive Hardware

I’ve got bad news for you. If you’re in the middle of a hardware refresh cycle and replacing anything that contains memory, you’re going to pay dearly for it. The even worse news is that there’s no waiting it out, because according to analysts this situation will last at least until the middle of next summer.

We can blame the AI wave for the whole mess: its huge demand for memory chips means there simply aren’t enough left for ordinary people like us. Some voices also point the finger at Apple, which pushed prices down so hard that manufacturers didn’t have enough money to build new factories. There’s probably a grain of truth to that, but right now it’s the memory makers who are profiting in incredible ways. It all just somehow happened, and building new factories isn’t simple enough to fix this within a single year.

The price trend has also led Apple to take a look at its own margins and decide it’s time to do something about them. So it announced a global price hike on pretty much everything — with the possible exception of cleaning cloths, keyboards, and mice.

Microsoft did the same with its Xbox consoles, and on top of that it stopped offering the 2 TB version.

The saddest thing about high prices is that they kill interesting products. The Steam Machine was supposed to be one of those, had it launched last year. It’s meant to be a game-console replacement that runs on SteamOS: you hook it up to your TV, turn it on, and just play — either with a controller or with a keyboard and mouse.

Steam itself sees the problem the same way, which is why it’s fully opening up SteamOS and trying to turn it into the operating system for games. Right now it only supports AMD, but it’s working on a partnership with Nvidia.

GPT 5.6 Just Around the Corner

OpenAI unveiled a set of new models named after our solar system. The most capable one, Sol, is meant to compete with Fable/Mythos; Terra is on par with GPT-5.5 but at half the price; and Luna is the workhorse for everything else at a low cost

These days there’s a big gap between announcing a model and actually being able to use it. We’ll have to get used to an era where every model release from American companies is followed by a risk assessment. For OpenAI, that process essentially kicked in for the first time.

BIZ Insights

Meta is starting to work on an app for so-called prediction markets — that gray zone between betting and investing. This segment has grown a lot lately, and Meta doesn’t want to let the opportunity slip away. The principle is simple: you place a bet on some event, and someone else bets against you. When the event happens, whoever predicted correctly splits the money from the other side. The platform makes its money on transaction fees, so unlike betting companies it doesn’t have to worry about whether the odds are profitable.

There was also an interesting article about how AI turned Meta from a well-running machine into chaos. It’s worth reading pieces like this so the same thing doesn’t happen to you, because AI transformation is going to be hard.

This is older news, but Amazon introduced Alexa for Shopping, a product meant to change the way we shop. Given the sheer size of the digital marketplace Amazon runs, it could really break through.

And since Amazon makes money from advertising as well as from its marketplace, it also introduced a way to weave ads into the shopping experience.

Besides investing in electric vehicles, China has also invested in robotics. Today it’s practically impossible to build a robot without some of its parts coming from China.

OpenAI, together with Broadcom, unveiled its first chip for AI inference. We don’t know much about it, but we do know the reason behind it: greater independence from Nvidia.

Google is starting to lower its payment fees on the Play Store. External payments are now allowed too, and we’re waiting on the certification of third-party stores (the first should be Epic’s).

Bain & Company is an interesting firm that does nothing but vibe-code a replica of your app at the request of a potential investor. That lets the investor judge how hard it would be for competitors to build a clone of your app, and whether you actually have any added value.

AI Insights

The author of Flask, who is also part of the Python and Rust ecosystems, reflects on how loops are getting more and more involved in development with AI agents. Loops work great for tasks like porting code, finding performance improvements, security scanning, and rapid prototyping. But loops can also multiply everything bad that happened at the start of the process. The worst part is that we probably won’t be able to avoid them, which is why it’s important to keep a human in the loop.

Sakana Fugu is another orchestrator that will present itself as a single model on the outside, while inside it runs different layers of loops with different models.

Tmax 27B is a retrained 27B Qwen 3.7, retrained on an expanded open-source dataset. The same approach can be applied to other models to improve their agentic capabilities.

Apertus is an open-weight, open-data, and open-science model built in Switzerland. That means it includes everything you need to use, retrain, or train from scratch your own 7B and 30B model.

Unlimited-OCR is a model that can parse even a very long PDF document or a complicated image in a single shot.

Mistral OCR 4 is the next generation of Mistral’s model built for OCR parsing. It’s not a free model, but a hosted one.

Real estate sellers have started using AI to edit photos of the apartment or house being sold. They claim they do it for customers, because the current state might limit their imagination of how their new home could be improved.

Anthropic is once again complaining to the US government that Alibaba used 25,000 accounts to distill its best models. Good thing Anthropic legally purchased all of its training data, right?

Links Drop

Nobody has really played GTA 6 yet, and yet they still managed to rack up a billion dollars in orders within an hour. GTA has such a strong name that nobody even wants to launch another game in the month it comes out — November.

IBM announced it has managed to develop technology for producing sub-1 nm chips. That sounds too good to be true, and in reality it means these chips haven’t broken physical limits — rather, they achieve the performance and efficiency such a technology would deliver. That is, higher performance and lower power consumption.

Deno is bringing Desktop apps, which let you run Deno applications with a frontend running in Electron.

If you ever get into building apps that deal with money, this should be required reading for you. The Fintech Engineering Handbook describes all the concepts and patterns you should consider when handling money.