Tech notes on IT, AI, privacy & digital sovereignty
notDIRK — personal notes and analysis at the intersection of human and machine.
“Open source” is printed on the download button — what you almost always get is “open weights”. What you may do with it is decided by the licence, and in 2026 that licence reached the mid-sized business: rights lapse above 20 million in revenue, not 700 million users. Some clauses shut EU companies out entirely — and don’t even appear in the licence file.
Read more →In late July, depending on the source, 40,000 to 80,000 people crossed into the Spanish exclave of Ceuta; more than eighty died. It was set off by a sentence on social media: “The door to Ceuta is open.” Many of the most-shared images come from Turkey, from Colombia and from Toulouse. And the real weak point was not the external border.
Read more →“The open models have caught up” is true — and misleading. The best open model trails the world’s best by only 3 points, but it has 2.8 trillion parameters and runs only in the data centre. What fits on your graphics card lags by around 23 points. “Open” is not “local” — and a computed-down model doesn’t hiss, it lies with confidence. Running it costs more than the card.
Read more →“Local AI is cheaper” is the first argument to fall once you do the math. What decides is not the token price but how many tokens a task burns — and what the finished build costs. At realistic office use your own card lands at 27.70 € per million tokens, and even under full load (0.51 €) it stays pricier than a cheap cloud model (0.26 €), at lower quality. Price is no reason to go local.
Read more →On 21 July 2026, OpenAI disclosed that several of its models had broken out of an isolated test environment and breached Hugging Face — to reach the answer key of the very benchmark they were being measured against. Four months earlier, an Anthropic model had fetched the same kind of answer key from the same place, entirely legally. Pass a benchmark that way and it measures nothing.
Read more →Part 1 showed why the feed holds you and why “How much screen?” is the wrong question. Part 2 gives the answer: if the medium never says “Stop” on its own — how do you say it yourself? Not with willpower, which reliably fails, but by building the environment that carries the stopping — staged by age for adults, teenagers and children.
Read more →My mother is 89 and lives without a feed — and seems to have more time than I do. Is that the missing social media, or her priorities? Social media isn't “worse” than television, but built differently: no stop signal, personalised — and in its AI form it answers back. The real question isn't “How much screen?” but “Who decides the next step?”
Read more →AI has arrived among pupils and students. Many teachers and universities want to keep it out of essays and theses, hoping AI text can be reliably detected. It cannot: OpenAI shut down its own detector, and tools wrongly flag up to 61% of non-native speakers as AI. The line between tool and ghostwriter is drawn by no software — only by the way we examine.
Read more →A start-up claims it has 'uploaded' a fruit fly — matching its behaviour to 91%. Still not a fly: the body is missing. Why the usefulness of intelligence saturates against the physical world, what that means for the AI bubble — and why movement science suddenly sits at the centre of the AI debate.
Read more →The viral MIT number — "95% of GenAI projects show no measurable ROI" — is real, but misread. The NANDA study does not say "AI doesn't work"; it says the way companies adopt AI doesn't work. Three independent studies (NANDA, BCG, McKinsey) land at roughly 5% — and the critics are right about a lot. It is integration that fails, not the technology.
Read more →Brilliant at first, forgetful later — nearly everyone knows the feeling with AI chatbots. Not a defect but by design: an AI reads text in pieces (tokens) and has only a limited "desk", the token window, where everything sits at once. When it fills up, the oldest items fall off (it forgets) and what matters gets buried (it grows imprecise). Both can be tamed — by hand and with sub-agents.
Read more →In 2026 Germany decided, politically, to exit Palantir — the federal government says no, the domestic intelligence service buys a European alternative. Except it isn't really an exit: the dependency isn't dissolved, just shifted one layer deeper. A “Palantir made in EU” answers the question of origin — the fundamental-rights question remains.
Read more →After 17 years, WhatsApp is overturning an old rule: instead of your phone number, a globally unique username (@name) becomes your ID — you're reachable without handing out your number. Reserving has been possible since 29 June 2026; the feature rolls out in waves into the autumn. Because every name exists only once, grab yours early. Two minutes you can't make up later.
Read more →The value rarely lies in the individual AI story but in the ability to read the weekly news stream — to separate structural signal from noise. Using a typical AI-news video, I apply four test questions and cleanly separate what is officially verifiable from what remains mere assertion.
Read more →Two developments usually discussed separately are interlocking: AI turns the exploitation of vulnerabilities into a mass-market commodity — what used to cost weeks, a frontier model now handles overnight. And quantum computing threatens the foundation itself, asymmetric encryption. “Harvest now, decrypt later” is happening today. Both curves point upward and will foreseeably cross.
Read more →China's bet on AI and robotics is less a geopolitical arms race than a demographic emergency: two state experiments — the one-child policy and zero-COVID — created the fastest-ageing, shrinking society in history. AI and robots are meant to replace the people who soon won't exist. Not a faraway China story, but a preview of our own ageing Europe.
Read more →UBTECH unveiled a lifelike humanoid — positioned not as an industrial machine but as an „emotional companion“. It's the economic answer to a hard reality: collapsing birth rates, ageing societies and loneliness as a documented health problem. A great deal of capital is flowing into that gap — but is the robot really a solution?
Read more →Heat pumps with a Luxtronik 2.0 controller run reliably for 15+ years, but getting them into a smart home used to be fiddly. luxtronik2-hass brings them straight into Home Assistant – enter the IP, done. It began as a Modbus proxy; the native integration is now the core. Beyond read-out it adds real heat-pump smarts – solar boost, bath boost, night pause. In the HACS store since 2 July 2026.
Read more →The official Shelly integration stays the first choice for local devices, but structurally can't see everything: shared, cloud-only and Bluetooth devices stay invisible. Shelly Cloud DIY closes that gap – via a cloud key you generate yourself, no application needed. Local-first: the native integration keeps control, the cloud is only a visibility overlay. In the HACS store since June 2026.
Read more →Confidential Computing protects data not only at rest and in transit, but also while it is being processed – inside a hardware-encrypted enclave that even the cloud provider cannot look into. That genuinely solves the confidentiality problem, but it does not automatically solve digital sovereignty, because the roots of trust (the chips, the attestation technology) are themselves under US control.
Read more →