Week 2026-27
Xbox is heading into a period of big changes, and we might be looking at worse access to the top AI models. @vlkodotnet
Week’s Highlight: Xbox restart
Microsoft placed big hopes in Xbox. The gaming market was growing, so it tried to turn the Xbox brand into a Netflix for games. It poured billions of dollars into acquiring game studios whose titles you could then play from day one for a monthly subscription.
It gradually reached 30 million subscribers, but to make this model sustainable it would need at least three times as many. According to the stats, only mobile games are growing, while consoles and PC see just modest growth. That obviously won’t pay back all of Microsoft’s investments, so it decided to do a big restart. It’s laying off 1,600 people in the gaming division this year and another 1,600 next year (some of the total 4,800 will come from other Microsoft divisions too).
The layoffs will hit even studios like id Software, where nearly half the staff will be let go, and other studios are facing either a spin-off or a sale to a new owner. We don’t know what this will do to the Xbox Pass portfolio, but given that we probably won’t see a refresh of gaming consoles anytime soon due to high memory prices, it doesn’t matter all that much.
It’s rarely talked about, because from games we as customers expect art and craftsmanship, but it’s quite possible that a lot of the work will be done by AI, just as we already see in programming today. I see it in dark colors, though. The one thing that could help this segment grow again is that the best games will become playable on cheap devices. Such a device could be your mobile phone, which you plug into your TV and a game controller, streaming the content from somewhere in the cloud. The problem is that this hasn’t caught on yet, but maybe it’ll change once iOS and Android open up the option to install third-party stores. Microsoft also owns a few studios that make mobile games (Candy Crush, for example) — now it just needs to bundle it all up nicely.
Local inference, locally
On one side we see that the U.S. government wants to control when and to whom its best models are available. On the other, a report came out that the Chinese government called in the Chinese firms building AI models to start doing the same.
If, like our company, you’re somewhere in the middle between these two players, you’ll definitely start wondering what we’ll do then. Will we go back to traditional programming? The answer could be local AI models, which don’t match the performance of the top models, but with a bit of the right skills and tools you can prepare for this scenario too.
We’re lucky that our boss gave us a budget, and above all that we have a colleague (Miroslav Martvoň) who devotes his free time and a fair bit of pocket money to the topic of local inference. Three weeks ago he teamed up with Ján Jakubčík from Kinit and put together a talk on how to run local models on local hardware. The talk is in Slovak, but there are plenty of translation tools available if you are interested in this topic.
Ports and conversions
Today we’ve gathered two interesting product conversions. The first is the announcement of TypeScript 7.0, which went through a conversion from TypeScript to Go. This sped up compilation by 8 to 12 times while using less memory. Go was chosen because it was the closest match to how TypeScript’s architecture works. As a fun fact, the port wasn’t done using AI, because those attempts failed (models like today’s weren’t available yet) — instead they built a translation tool.
The second conversion is the Bun compiler moving from Zig to Rust. A lot of tokens were burned on this conversion, because Bun is now part of Anthropic, so they have no problem with that. There’s a detailed article about it.
After it was published, an article came out from Andrew Kelley, the person behind Zig. It’s also an interesting read. According to him, Bun wasn’t using Zig correctly, was piling up a lot of technical debt, and ever since Bun became part of Anthropic, they stopped sponsoring its further development.
BIZ Insights
The era of AI browsers is over. OpenAI had been preparing its own browser, Atlas, which it has now scrapped.
But a replacement is coming: native apps that hide the browser somewhere inside themselves and use it only to simulate user interactions. So OpenAI is replacing Atlas with a product called ChatGPT Work.
We’ll stay on this topic for a third product too. You surely know Claude Cowork. Well, its version for mobile and web is now arriving. If you use the desktop version, don’t forget to turn off task processing on the servers. It means your task will keep running even when your computer is off, but there’s some data you don’t want leaving your machine.
As of last week, all new vehicles in the EU must have a camera pointed at your face to detect whether you’re paying enough attention to driving and to alert you about it. As always, when carmakers get a mandate to add something extra to a car, it’s unclear what they’ll do with that data afterward.
Maybe you were happy that Chat Control 1.0 was struck down by the EU Parliament. Except it’s back.
I was more pleased by the news that Meta has probably violated the Digital Services Act and will have to disable features like autoplay and infinite scroll on Facebook and Instagram, or else face a fine of up to 6% of its global revenue.
In its ongoing lawsuit with the NYT, OpenAI claimed it couldn’t search through anonymized log samples. Except, according to the latest information, it was exactly the opposite. If that gets confirmed, it could weigh against OpenAI in this case, which is about copyright.
And while we’re on the subject of what AI companies hide from us, it came out that Claude Code was phoning home with information about your time zone, proxy, and so on whenever it detected you were located in China. Anthropic defends itself by saying it was only an experimental feature and a protection against model distillation.
AI Insights
Today we again have a pile of new models. The first is GPT-Live, which improves how you can talk to GPT by voice. The model can run in full-duplex mode, meaning it can listen and respond at the same time. On top of that, it can spin up processing of more complex questions in the background.
Qwen introduced AgentWorld, a model built specifically for 7 domains (Terminal, Search, MCP, SWE, Web, OS, Android) in which it can predict the environment’s next state. That comes in very handy for agentic use, for example in something like OpenClaw. There’s also an open-weight version.
Meta also introduced its new model, called Muse Spark 1.1. It’s available via API and, for Meta, it’s a step toward its vision of personal superintelligence. And as is usual, it tops some benchmarks and falls short in others.
Tencent introduced a mid-size model, Hy3, which competes in performance with much larger models.
Mistral introduced Robostral Navigate, a model with just 8 billion parameters that lets robots navigate complex environments autonomously using a single camera. By the way, its results make it the current state of the art among these models.
Anthropic published a study according to which its models think in a way similar to how humans do. They can build a so-called J-space, whose explanation I’d rather not get into.
It’s been a while since we had any reading here about how badly we as humanity could end up with AI. Luckily, someone has already come up with a plan. It’s called Plan A. If we follow it, then by 2040 we wouldn’t be heading toward extinction but toward a bright human future.
Links Drop
Rabbithole is an infinite canvas where you start discussing a topic and, by asking questions, you can branch it out endlessly.
OpenLogi is a local alternative for all owners of Logitech mice who want a lightweight app to configure them.
Someone made Minecraft for the GameBoy.
Need to quickly host your vibe-coded page somewhere? Then Cloudflare Drop might come in handy — just upload an index.html or, if there’s more content, a zip file, and for one hour your page is hosted on Cloudflare. Or if you have a Cloudflare account, you can take ownership of it.
I’m not sure whether I’ve shared this tool before, but MarkItDown is a tool that converts any document into Markdown syntax. For OCR scenarios, it lets you hook up an LLM vision model.
