Week 2026-22

Jun 11, 2026

Microsoft held its Build conference, and as usual it was all about AI — plus ARM on Windows this time. @vlkodotnet

Week’s Highlight: Microsoft Build 2026

You might expect a conference to bring a pile of different announcements these days, but what we actually get is companies showing how well they can integrate AI — how they’ve finally found the right way to make it useful. A day before the event, Nvidia introduced its new processor platform, RTX Spark. RTX Spark is an Nvidia ARM processor paired with a Blackwell GPU, and Microsoft already has two products built on it: the Microsoft Surface Laptop Ultra and the RTX Spark Dev Box. One is a portable device, the other a developer workstation for your desk. What matters most for Microsoft is that both fully support Windows. After Qualcomm Snapdragon, that makes Nvidia the second manufacturer enabling full-fledged Windows on ARM. Older programs run through emulation, and as Jensen Huang put it, you’ll be able to run every Windows program from the past 40 years on it. Does that include games? We don’t know yet, but it’s been confirmed that all anti-cheat and DRM software will run on it.

Nvidia plans to keep investing in these chips, with next generations already in the works under the codenames N2X and N3X — just expect a premium price tag.

To give you the easiest possible path to making your Windows app ARM-compatible, Microsoft will provide AI agents to help with the conversion.

WinUI 3 app development is getting similar AI support.

Microsoft also introduced a new local model, Aion 1.0 Instruct, which replaces Phi-4-mini and runs even on weaker hardware. It’s meant to handle summarization and other basic browser tasks locally, without burning tokens in the cloud.

The most important announcement, though, was a collection of new MAI models designed entirely by Microsoft, set to replace the OpenAI models. This should be the definitive end of their partnership. You can still buy GPT models in Azure, but Microsoft will use its own MAI models across its Copilot products. The lineup includes a new Thinking model, speech recognition models (covering Czech and, most recently, Slovak), and voice output models.

For us developers, MAI Code 1 Flash is the interesting one — they benchmark it against Haiku 4.5, and I hope it becomes the model in the GitHub Copilot ecosystem that doesn’t eat through all your Copilot credits in a few days.

Microsoft also unveiled Project Solara — devices meant to bring Ambient Computing to the corporate world. They’re supposed to replace your desktop or laptop with a set of devices you interact with by voice or simple gestures. Behind the scenes, they’ll call various AI models and tap into knowledge bases and cloud databases. Two devices have been shown so far: an Amazon Echo/Google Home clone, and something that looks like a smartphone but is really just a screen wired up to AI agents.

For the topics I didn’t cover, check here:

or here:

BIZ Insights

TSMC let out a sigh — they no longer have the capacity to keep up with AI chip demand. They’re not giving up and are building new fabs, but wherever demand outstrips supply, expect prices to climb.

The UK CMA is imposing new obligations on Google, which will have to let publishers opt out of having their content used in Google’s AI features. The setting isn’t supposed to affect rankings in traditional search. Nice idea, but the way I see it, give it a few more months and nobody will be going back to traditional search anyway.

GOV.UK is replacing Stripe with the Dutch provider Adyen. Adyen will enable direct bank transfers.

Meta may soon join the ranks of companies selling cloud compute. That is, if everything goes well and badly at the same time: well, meaning they manage to build out all their planned AI infrastructure — and badly, meaning it sits unused because nobody wants their AI models.

Meta also launched its Meta Support AI chatbot this year to simplify and speed up its support processes. They may have simplified things a bit too much: hackers discovered that to take over an account, all it takes is connecting through a VPN as close as possible to the account’s location and then asking for a password reset to a new email, claiming the original email account was hacked. The AI happily obliged — and ignored the 2FA settings too.

You’ve surely heard of token-maxxing — the trend of internal company leaderboards ranking who burns the most tokens. Well, Uber found out they had blown their entire annual AI tools budget in just 4 months this way, so they introduced a monthly cap.

VoidZero is becoming part of Cloudflare. You probably know VoidZero from Vite. Their projects will remain open-source and platform-agnostic, but we can expect easier integration with Cloudflare services.

Anthropic confidentially filed a Form S-1, starting its journey toward an IPO. The timing is ideal for Anthropic: projected revenue growth, an expected first profitable quarter, and Mythos about to launch. A confidential filing means the documents aren’t public, so there’s not much more I can say about it.

AI Insights

It looks like Google Spark — the agentic system that’s basically Google’s own take on OpenClaw — can handle a lot. If you live in the Google ecosystem, it can take a fairly vague task and automatically pull in information from your emails, calendar, and location history. Tell it to send something to your wife, and it figures out on its own who your current wife is, what her preferences are, and schedules everything around your calendar. It pulls things into a task you’d never expect it to.

Codex also wants to be more than a developer tool — a universal work assistant. It recently added plugin support that can build you things like interactive websites, apps, design drafts, data analytics, and so on.

Bonsai Image 4B is an extremely compact image generation model that fits even on your phone.

Gemma 4 12B can process multimodal content without a vision or audio encoder — it goes straight in as model input. On top of that, it has Multi-Token Prediction (MTP) and a permissive Apache 2.0 license.

Nemotron 3 Ultra 550B is Nvidia’s contribution to open models. It’s aimed more at data centers.

Anthropic released a reference implementation of how vulnerability discovery should work using various skills and autonomous pipelines.

And finally, another article from Anthropic, describing how the era of AI agents may be followed by an era of AI training and improving itself. Anthropic uses itself as the example: AI has sped up their development and processes, and Claude’s output is already at a human level. Humans are still needed to pick the goals, decide what to actually do, and set priorities.

Links Drop

Dell introduced the XPS 13, meant to be that $600 computer for students. It’s a machine that could compete with the Apple Neo, though Dell claims the idea was born long before the Neo hit the market.

Cyberdecks are a charming trend of small computers built on Raspberry Pi and other single-board computers with a tiny screen, packed into little enclosures. It probably won’t become a global trend, but I find it adorable.

Mouseless is a paid product, but an interesting app that lets you replace your mouse with the keyboard. Can’t imagine how? Definitely watch the product video. It’s a pretty interesting app — I’m trying out the trial myself.