Week 2025-51

What seemed impossible yesterday is just another item on China’s to-do list. @vlkodotnet

Week’s Highlight: China’s EUV Dream

For a long time, we believed that extreme ultraviolet lithography machines (EUV) were such a perfect invention of Western civilization that developing them elsewhere would require enormous investment and many years of research. That’s why they became a tool of political control—anyone without access to this technology couldn’t manufacture processors better than 14 nm. The older deep ultraviolet lithography (DUV) is used for that, though China managed to push its limits to 7 nm. They essentially squeezed every last drop from this technology, since not a single EUV system has ever been sold to China.

But this is China we’re talking about. In 2020, they launched a secret program to reverse-engineer and build their own EUV chip manufacturing system. Today, they have a working prototype that doesn’t yet produce functional chips but should within two to three years. How they achieved such massive progress makes for fascinating reading. Huawei oversees the entire project, coordinating the supply chain, development, and manufacturing of necessary equipment. Workers are divided into isolated teams and even sleep in the factories (the Manhattan Project analogy fits perfectly here). They aggressively recruited people working directly at ASML, some of whom were given false identities. This led to situations where two acquaintances met during development but had to address each other by fake names. On top of that, they bought decommissioned components from older ASML machines or acquired them through intermediaries.

Business Insights

TikTok in the US will finally get proper American ownership. The new American investor consortium will hold 50%, existing US-based ByteDance investors get 30.1%, and the original ByteDance retains 19.9%. TikTok will also receive an American version of the algorithm trained exclusively on Amaerican data. That’s a lot of America in one paragraph.

Google is opening up its store, but if you buy an app or game through a link outside the Play Store, it’ll cost you $2 to $4. The fees won’t kick in immediately—only when Google decides the time is right.

Google is suing SerpApi, a company that bypasses all the technical protections Google has implemented to prevent scraping of search results. When technical measures fall short, legal ones can step in.

Roomba maker iRobot has filed for bankruptcy. The company will be acquired by Chinese firm Picea Robotics, which was their contract manufacturer.

Ford is discontinuing its massive electric pickup, the F-150 Lightning. Turns out fans of huge pickups with huge engines don’t really want huge vehicles without huge engines. This decision also kills the partnership with SK On, with whom Ford had plans to build battery factories for electric vehicles. All that remains is their own LFP cell manufacturing plant.

Wondering who else besides Nvidia is cashing in on the AI boom? It’s expert data suppliers. This data is the most valuable commodity in reinforcement learning—the training phase where models are optimized through rewards. Subject matter experts who can evaluate response accuracy are in extremely high demand.

AI Alliances Are Sprouting Like Mushrooms

If there’s anything interesting about this new AI era, it’s how quickly companies can unite around a common goal. Enter the Agentic AI Foundation, featuring Anthropic, Google, Microsoft, OpenAI, AWS, and others. Their aim is to build an open ecosystem for AI agent development.

The Agentic Advertising Foundation, meanwhile, will focus on developing agentic AI in advertising. The AdCP (Agentic Communication Protocol) standard should help with that.

AI Insights

Google managed to respond to GPT 5.2 with a new fast model for everyday use: Gemini 3 Flash. It will power the Gemini app, and you can also access it via their API and agentic platform Antigravity.

OpenAI countered Banana Pro with ChatGPT Images, their new image generation model. It handles editing existing photos, creating new ones based on references, and manages longer text. Unlike typical diffusion models, this one represents every part of the image with tokens.

Apple’s SHARP model generates a 3D Gaussian representation of a scene from an image in about one second. I haven’t figured out what it’s actually good for, but probably sharpening images.

OpenAI’s latest GPT 5.2 now has a Codex version designed for developers.

For those who care about response speed, Nvidia introduces the Nemotron 3 model family. They support up to 1M context and run 2–3× faster than comparable models with similar or better accuracy.

Mistral released the third version of their OCR model. Converting about 1,000 pages will cost you one to two dollars.

Microsoft’s TRELLIS model won’t make 3D artists happy—it can generate 3D models from either text or images. Though if you look at the generated models up close, they don’t need to panic just yet.

SAM 3 (Segment Anything Model 3) from Meta can not only isolate specific objects in images but also separate the sounds they make.

Cursor got a visual editor so we don’t have to prompt simple changes—we can just modify things the easiest way: visually.

And finally, a lament from a Kenyan native who complains that how they’re taught, how they write and summarize text—that’s exactly how AI models do it.

.NET Insights

If you’re frustrated by the messy old .sln format, maybe it’s time to try the new .slnx. Of course, you’ll need to update your build stack since not every CI/CD system supports this new format yet.

What’s it like developing a .NET project in the new Zed editor? Someone tried it for you. The verdict: it’s fast, but the .NET tooling still doesn’t match VS Code quality.

Link Drop

Samsung unveiled the new Exynos 2600, the first announced chip with 2 nm technology. It promises better performance than the older Exynos 2500 in both CPU and AI, plus it features AMD’s RDNA GPU technology. Benchmarks will tell us whether it’s enough to compete.

Cloudflare’s annual report analyzes global internet traffic trends. There was more AI and AI bots, Starlink expanded, mobile devices now dominate traffic in 117 countries, and high-volume DDoS attacks increased tenfold.

Quill OS is an open-source system for Kobo e-readers.

Closing Visual

This is an older ad, but what else would you share before Christmas?