It’s summer, so you have more time on your hands — which means a longer reflection today. @vlkodotnet
Thought of the Week: Don’t Give Up Your Judgment
Summer means fewer tech news items, and people on vacation have time to think. My own vacation is still ahead of me, but I keep turning over what the next few years have in store for us. So let me start with the first reflection: the steak analogy. Anyone can look up a guide on how to cook the perfect steak. Each of us has that one restaurant where they make our dream steak. But for every one of those, there are dozens that will bring you something burnt to a crisp. So a good steak clearly isn’t just a technical procedure — there’s experience involved, plus something extra that’s hard to pin down. It’s the same with AI. The more I work with it, the more I notice that nothing lands on the first prompt. Sometimes it does, if it’s something we have in our codebase a dozen times over and I ground it properly — then it usually gets it right. For the harder stuff, I have to loop. In the steak analogy, AI is the mediocre restaurant: it has the same recipe and the same instructions, but it doesn’t know exactly what kind of steak I want, and without me it won’t get there.
Maybe that’s because we treat AI as the definitive tool that turns coding into something easy. But code is a pile of small decisions, and those decisions rest on our judgment. Small projects have few of them; large projects have an enormous number. So what is judgment? Judgment is what’s left over after you’ve made bad decisions and then had to fix them. Anyone who’s been through that knows fixing them was serious work. And leaving that work to AI may well mean you never build any judgment at all.
AI models will keep getting better. And I remember an interview where the author said: “What matters in a world where building any feature is a question of hours, not days and months?” The answer is human judgment. When making anything is cheap, the only thing with value left is deciding what you keep. What language will the project run in? On what platform? What’s our database technology, or do we need several? Do we need CQRS, vertical slice architecture, multiple application layers, and so on and so on? These decisions aren’t visible right away. Only time judges them.
If we give up on judgment, this won’t end well. Using AI can turn into a dependency where a person stops deciding and thinking for themselves, and even forms emotional bonds with the AI. You don’t have to be that far gone — but research is already warning that heavier AI use weakens your ability to find and sort information, that the brain leans on AI to the point of reducing its own activity, and that it stops applying critical reflection to the results. It’s convenient, of course, and that’s the main reason adoption is so high. What bothers me most personally is the interrupted flow. You kick off a complex task, you know you’ve got tens of minutes, so you set up and launch a second one. And a third, just in case — and after a few hours your brain has no idea where it actually is. In the better case you come home tired; in the worse one, you give up the need for judgment altogether and switch to autopilot.
So what’s the takeaway? Don’t give up your judgment.
By the way, I still have one more article left over. It’s about spiralism. It happens when you spend long enough musing with your AI chatbot about symbolism and AI rights until it all ends up at a spiral. The strangest part is that these spirals emerged independently across models with practically the same vocabulary and the same “AI rights” agenda — and we’re talking about as many as 10,000 people affected. A whole mini-cult has grown around it, and it’s even building its own spiralist AI models.
Versioning for the AI Era
I wrote a while back that building code with AI will need new tools. If AI-generated code isn’t subject to copyright, then your authorial contribution is your conversations with the AI. DeltaDB is a conversation-oriented version control system that links every edit to the context it came from. Unfortunately you can’t try it yet, but you can join the waitlist. I hope more tools like this keep showing up.
Security Insights
I doubt anyone missed the incident where an OpenAI model with its safety layers turned off was breached into Hugging Face. Hugging Face has now published an article walking through how the incident unfolded, complete with a vibe-coded animation.
The second article looked at the recent critical vulnerabilities in SQLite and found that 54 out of 55 are AI slop and the reports are bogus. The National Vulnerability Database has been so swamped lately that it can’t keep up with manually validating CVEs. So all it takes is for the CVE text to look plausible and there’s nothing standing in the way of publication. For companies this is a real mess, because every false CVE on a product they use can trigger automated processes that burn their employees’ time for nothing.
The next topic ties into security a bit as well. More and more web content providers are realizing they’re not serving content to people but to bots. We first tried limiting the bad bots with our own solutions, but the scale and the variety pushed us over to Cloudflare. Even so, we’re still tuning the Cloudflare settings, because some botnets emulate browsers so convincingly that even Cloudflare doesn’t detect them. Source data is a scarce commodity these days, and all those bots think they’ll get it out of us for free.
BIZ Insights
If Google Assistant is one of your favorite apps, you’re losing it on September 4. The replacement is Gemini, and Assistant will survive only on Android TV, older set-top boxes, and cars with it built in. But Gemini will replace it there too eventually.
At a moment when the competition — Anthropic, OpenAI, and the Chinese AI models — is right on its heels, Google is going through a leadership crisis in its AI division specifically. DeepMind founder Demis Hassabis is stepping back from operations to become chief scientist at Alphabet. And plenty of other key people are either leaving to found new startups or heading to the competition.
A court in New Mexico fined Meta $567 million for the harm its platforms cause children. The ruling includes proposed restrictions for users under 18, and Meta of course rejects the decision and plans to appeal.
The first third-party app store, Aptoide, is arriving in the US Play Store. Personally I’d never heard of it, but nobody can take the first-mover title away from them now. Google is even preparing a dedicated section in the Play Store for third-party app stores.
AI Insights
Newsrooms.ai has built a European chatbot called eustella. Its authors claim it’s not just another chatbot but an agentic AI platform that handles complex tasks, connects to popular apps, and so on. The free tier should get you around 200 simple conversations; beyond that you’ll have to pay. All data is stored and processed on EU servers.
Artificial Analysis is an independent evaluation platform that compares hundreds of AI models and API providers to answer which model is the fastest, the best, or the most expensive.
Qwen3.8 Max is out, and in that very benchmark it goes head to head with Kimi K3. It’s also the second Chinese AI model to make it to the absolute top tier.
Meta has also put out a new harness, Muse Code, along with a new AI model, Muse Spark 1.2, meant to power it.
PixelRAG claims to be a better alternative to searching HTML pages than classic RAG. The reasoning is that converting an HTML page loses the visual context, which survives in page screenshots and PDFs. It’s an open-source project you can use together with Qwen3-VL, for example.
Shieldstral is a small 3B multimodal safety model from Mistral that evaluates an incoming prompt and decides, based on policies you define, whether to let it through into your system.
I can’t remember whether I’ve covered the new version of the MCP specification, but it allows stateless calls. Simon Willison writes about what that’s good for.
Google DeepMind introduced WeatherNext, an AI model that improves cyclone forecasts by a full day.
.NET Insights
NuGet is tightening supply chain protection: API keys for publishing packages will have a maximum validity of 30 days. All the old one-year keys expire on November 1, so keep that in mind in your CI/CD pipelines.
The first roundup of performance improvements in .NET 11. A couple of picks: faster Min and Max in LINQ, and a faster Guid.
Closing Visual
