Week 2026-08

Last week, Anthropic went head-to-head with the US government — and OpenAI came out the winner. @vlkodotnet

Week’s Highlight: Anthropic vs USA

It would be naive to think anything other than the obvious about CEOs of major companies: their job is to paint the best possible picture of their products. They don’t really care about the good of humanity — they care about the money they can extract from you. That’s why Dario Amodei’s actions as founder of Anthropic surprised everyone.

Anthropic was an AI model supplier for the US government, which used the model to analyze the attack on Venezuela, among other things. When Anthropic found out, they started digging for details. This sparked a conflict between the US government and Anthropic, where Anthropic refused to allow their AI model to be used for mass surveillance of American citizens and fully autonomous weapons systems. Naturally, such audacity infuriated US leadership, and the Department of War cancelled its contract with Anthropic and designated them as a dangerous partner.

This would mean all military sector suppliers have six months to remove Anthropic from their systems. Whether they can actually enforce this is debatable — Anthropic’s legal position says otherwise.

Guess who swooped in on that $200 million contract and offered a replacement. If you guessed OpenAI, you guessed right. OpenAI even claims they have the same contract guaranteeing they won’t conduct mass surveillance or autonomously kill people. The difference is in the details. Anthropic required it explicitly, while OpenAI has a clause saying “in accordance with applicable law.” And the legal landscape is all over the place right now — all it takes is an executive order, and by the time a court case proves otherwise, a lot of time will have passed. What makes this debate even more relevant is a study published by King’s College London. They gave the three best LLMs a hypothetical game simulating crises between nuclear powers. The LLM models tried to win at all costs — regardless of the fact that it would trigger nuclear war — because these models have slightly different motivations than the good of humanity.

This will get interesting for OpenAI, because a petition emerged in support of Anthropic, signed by top engineers and scientists from OpenAI and Google — specifically 741 from Google and 96 from OpenAI.

A wave of CancelGPT swept across the internet.

Anthropic published a guide on how to transfer your memory from OpenAI so you don’t have to shape your AI model from scratch.

But we know how long these controversies last. In two weeks, nobody will remember, and OpenAI will have a $200 million contract in their pocket.

Block AI Layoffs

Block CEO Jack Dorsey announced the layoff of 40% of employees. His reasoning: it’s clear that AI will replace many jobs, and instead of slowly letting people go, they’d do it all at once and be done with it.

This immediately caused a huge uproar online. Everyone started pontificating about how the AI revolution is here and we’re all going to lose our jobs and so on. I’m not saying the revolution isn’t happening, but we’re talking about Jack Dorsey here. The guy whose successor, after Dorsey sold his previous company Twitter, cut the workforce from 7,500 to 2,800 and nothing happened. He’s simply known for being unable to estimate how many employees he needs. This time he overshot again, because the headcount trajectory doesn’t match the revenue trajectory at all. But he sold it well, timed it perfectly — the stock jumped 10%, and that’s what matters.

BIZ Insights

Meta spotted an opportunity when they saw AMD close a deal with OpenAI to supply chips — and receive AMD equity on top. So Meta tried the same thing. AMD supplies AI chips, Meta pays for them, and gets a discount in the form of AMD shares.

Not everything that glitters is gold, and AWS learned that the hard way. Their 13-hour outage in December was apparently caused by their AI tool Kiro. Well, actually it was caused by an employee who ran it under their own permissions — permissions they shouldn’t have had in the first place. So they ultimately blamed human error, but that’s not stopping us from jumping to conclusions, right?

Samsung, the second-best-selling mobile phone brand, unveiled its new S26 flagships. Slightly better processor, slightly better camera, and slightly better display (they added a special discreet mode so nobody can peek at your screen from the side). But the biggest news isn’t in the hardware — it’s in AI. You get three tiers of AI: local AI that can see your screen and do all sorts of things with photos, agentic AI via Gemini that handles morning summaries, and Perplexity for everything else.

AI Insights

Google introduced Nano Banana 2, a new model that’s as fast as the original Nano Banana and as accurate as Nano Banana Pro. It can insert up to 14 of your uploaded image objects and text into a picture. And all of that for pocket change.

Google also created Lyria 3, a new music generation model that produces tracks up to 30 seconds long and can adapt to video or images you include in the prompt.

Microsoft is launching Copilot Tasks, which is essentially Claude Cowork but from Microsoft — and probably better integrated into Windows. The supposed advantage is greater security when you give it access to your computer. We’ll see.

Pi is something like Claude Code, but fully customizable to your needs. It supports multiple AI models and stores communication history in a tree structure. It has RAG support and various types of context engineering. If Claude Code feels like a tool for noobs and you don’t want to burn tokens building yet another pointless agent, Pi is a solid choice.

A review of Kimi K2.5 dropped. In short: a good model, especially for front-end and visual tasks, but it’s chatty — what you save on the per-token price, it spends in sheer volume.

We asked Claude Code across 2,430 repositories what it recommends, and surprisingly it prefers custom solutions over existing tools. When it did pick a tool, it favored specific ones: GitHub Actions – 94% (CI/CD), Stripe – 91% (payments), shadcn/ui – 90% (UI components), Vercel – 100% (JS deployment), Tailwind CSS – 68%, PostgreSQL – 58%.

If you have at least 5k stars opensource project on GitHub or 1M monthly downloads on NPM, you can get Claude Max 20x free for six months.

Claude Remote Control can be used, for example, to control your Claude Code instance from the comfort of your phone. Handy when you step away from your desk and don’t want Claude Code sitting idle because it’s waiting for a simple response from you. You do need at least a Max subscription for now, and it’s not available for enterprise customers.

.NET Insights

Today’s .NET section is a bit lighter, but something is better than nothing. It covers HybridCache, which uses a Pub/Sub architecture to handle cache invalidation across multiple nodes in a system.

Links Drop

Many people in my circle use Hetzner as cheap hosting for their projects, so it’s worth knowing that starting April 1st, they’re raising prices by 10 to 30%.

Sophia Script is a so-called debloat app that can disable or remove over 150 things from your Windows 11. Since it’s a PowerShell script, you can customize it yourself, or use SophiApp for a GUI-based approach.

If you work from home and you’re fed up with cloud services, you can unwind with the game Data Center. You become a data center administrator, running cables, swapping disks, and so on.

Finally, a science update. Researchers have likely developed a universal vaccine against respiratory diseases. It works by activating your immune system before flu season.

Closing Visual

How hiring has changed over the last five years.