Week 2026-16

Where the cosmos meets the cursor, things can hardly go wrong. @vlkodotnet

Week’s Highlight: SpaceX + Cursor

Cursor has been complaining for a while that they have to pay list price for API calls, while OpenAI and Anthropic effectively subsidize their own models. And it really does look that way if you access these models through API pricing rather than through your monthly subscription. Cursor tried to close that gap by offering its own model, Composer 2, based on Kimi K2.5. The catch is that Composer 2 never quite matched the quality of the more expensive models.

Now SpaceX has announced that it will tap Cursor’s expertise to develop a new model in its AI data centers. Which actually makes a lot of sense given how many source-code interactions Cursor has collected over its lifetime. Or did you really think all those free tokens were handed out purely out of the goodness of their hearts? They’ll pour all of that into SpaceX’s xAI division, run the servers red-hot for a few weeks, and the result should be a bit more competition on the market. On top of that, SpaceX has set itself up either to acquire Cursor for $60 billion, or to pay them $10 billion as a partnership fee. To me this looks like a perfectly logical collaboration that helps both SpaceX and Cursor — and there really aren’t many big potential acquisitions left on the current market.

Token Economy — Continued

A few interesting follow-ups landed this week that fit neatly into last week’s topic. A rumor went around the internet that Anthropic was no longer offering Claude Code as part of the Pro subscription. Luckily it turned out to be just A/B testing — they fed that information to a randomly selected 2% of potential users.

Anthropic also published a postmortem explaining why their model hasn’t been very useful in recent weeks. They blamed three separate issues, but it still feels like the main reason was simply a lack of compute. Hopefully they’ve managed to wire up more servers at Google and AWS.

Probably more at AWS, because the next deal suggests Amazon will be their main partner — they’ve put another $5 billion into Anthropic, bringing Amazon’s total investment to $13 billion. In return, Anthropic will buy $100 billion worth of services from them over the next few years.

When Google heard about this, they basically said: “If Amazon’s putting in five, we’re putting in ten.”

Let’s hope they won’t be asking for that money back any time soon, so we can keep burning tokens through subscription services for a while longer. And do you know where you won’t have prepaid tokens anymore? If you guessed GitHub Copilot, you guessed right. Instead of flat token allowances, everyone will get a fixed quota — and once you run out, you buy more.

BIZ Insights

Apple is changing its CEO from Tim Cook to John Ternus. Cook has been there 15 years and at 65 he deserves to retire. Apparently his workday ran from 5 a.m. to 11 p.m., with no mention of whether he at least rested on weekends. When I had a stretch like that for just a few months, I was burned out like a candle burning at both ends, and it finally ran out of wax. But Tim clearly handled it better than I did in my younger years and seems to have managed his work-life balance just fine. We wish John the best of luck — and since he has an engineering background, maybe we’ll even get some interesting new Apple products out of it.

Meta has decided to use its US employees to gather valuable data on how exactly they work on their computers. The tracking software will record mouse movements, clicks, and keyboard shortcuts paired with screen captures.

Some less charitable voices say they’re just teaching the AI to automate their jobs — and that 10% of them will lose those jobs in the near future.

Atlassian plans to start collecting interactions in its Jira and Confluence products from August 17th to train AI models. For Free and Standard users, the opt-out will only be available at the user level. For Premium and Enterprise it will be off by default. They’ll be collecting everything: titles, descriptions, comments. Moving away probably isn’t realistic, so the least we can do is bake the “turn this setting off” step into our onboarding processes.

Deezer reports that 44% of new music uploaded to their platform is AI-generated. According to their own tests it doesn’t really matter, because 97% of people can’t tell an AI track apart from a human-made one. They still label such content rigorously. But honestly — who actually looks at any kind of label while listening to music?

Starting next year, all new phone models sold in the EU will be required to have a replaceable battery. It’ll be interesting to see how this affects the foldables market, since those probably won’t be able to afford that luxury. But apparently they thought of that too — replacements can be done by an authorized service center. On top of that, manufacturers will be forced to use more durable batteries and to provide system updates for at least 5 years.

AI Insights

There was a time when I’d open the AI Insights with a story about how AI broke production, just so we could all have a laugh. Today I open with these stories so that we can all avoid similar problems. In this case, a Cursor AI agent found broken credentials in a staging environment. So it decided to clean things up and looked for new credentials. Those accidentally had higher privileges than they should have — and instead of cleaning staging, it cleaned production. Backups included.

Mythos helped fix between 150 and 271 vulnerabilities in Firefox. Not because people couldn’t find them themselves, but because there simply aren’t enough people for that kind of detailed review. Which, on the flip side, isn’t a problem for AI — and the dynamics of the industry are shifting now that AI tools can find bugs faster than an attacker can. The Mythos story looks like a happy ending. Provided, of course, that you’re their customer and you’re burning enough money on reviewing your source code.

Models have been popping up everywhere too. Kimi K2.6 is the first model in a long time that can compete with Anthropic’s Opus models in software development. That is, if you’re willing to send your data to China. There’s an open-weight version available, but you’d need hardware with more than 512 GB of memory to run it — which most of you probably don’t have at home.

What you can’t have at home, you can rent from an AI provider. OpenAI introduced GPT 5.5, tuned for programming tasks. According to them it should be more efficient with token consumption. We’ll see.

Along with the new model comes a new image generator, ChatGPT Images 2.0.

A new DeepSeek V4 also dropped, in two flavors. You could probably run the Flash version at home, but it’s optimized for Huawei hardware, so it’ll take a while before there’s a version you can actually try out at home.

But that’s not all. China decided to push on every front, and so Alibaba unveiled Qwen 3.6 Max, which should be comparable to the top models (read: a bit better on some benchmarks, a bit worse on others), but available only in the hosted version.

Local-vibe enthusiasts will be more excited about the Qwen 3.6 27B dense model — that is, a model without MoE. It puts a bit more strain on the hardware, but still produces good results on devices with less RAM.

Anker introduced an AI chip called “Thus”, which will go into its new products. Their headphone line will benefit the most, but Anker also makes robot vacuums.

.NET Insights

Not much from me here this week. A bit of a refresher article on caching in ASP.NET Core. You’ll learn how it works, how to invalidate it, how to cache content per user via policies, and also how to wire the cache up to Redis.

Links Drop

Laws of Software Engineering is a collection of principles and patterns that come in handy for software development, team management, and decision-making. Everything is laid out clearly with filters for categories and skill levels.

The initiative tasked with restoring trust in the Windows operating system has been given the codename Windows K2. This isn’t the codename for a new version of Windows, but a long-term program aimed at improving performance, reliability, bringing back beloved features, and building a new community.

The Framework Laptop 13 is a laptop that could finally live up to the potential of being your new daily driver. Compared to older Framework devices, this time around Framework — as the manufacturer — was able to negotiate better terms and, with them, better build quality. The price starts at €1,380 and the upper limit depends on your configuration.

Closing Visual

This is an AI-generated image, but as ideas for putting your GPU to use go, it’s not a bad one.